Job Displacement and Intragenerational Mobility. Nicholas A. Jolly Department of Economics Central Michigan University

Transcription

1 Job Dsplacement and Intrageneratonal Moblty Ncholas A. Jolly Department of Economcs Central Mchgan Unversty E-mal: September 2009 Abstract: The analyss presented here uses the 1968 through 1993 waves of the Panel Study of Income Dynamcs to examne how job dsplacement nfluences ntrageneratonal earnngs and ncome moblty. Usng ndvdual labor earnngs, ths study shows dsplacement ncreases the probablty of downward moblty for several years after separaton occurs. Furthermore, the probablty of beng n the bottom half of the labor earnngs dstrbuton ncreases sgnfcantly, not only n the year of job loss, but also for several years followng dsplacement. However, ncome from other famly members and government transfer payments mtgates dsplacement s adverse effect. After consderng these addtonal measures of fnancal well beng, the shortterm mpact of dsplacement on movements throughout the ncome dstrbuton s reduced, and the long-term effect s elmnated. JEL Codes: J63; J65 Keywords: job dsplacement; earnngs losses; ntrageneratonal moblty

2 I. Introducton Researchers are aware of the long-term effects job dsplacement has on ndvdual workers earnngs. One area that has receved relatvely lttle attenton s how ths type of nvoluntary job loss nfluences the nter-temporal movement of workers through the earnngs and ncome dstrbutons. Ths study uses data from the Panel Study of Income Dynamcs (PSID) and shows that dsplacement sgnfcantly ncreases downward earnngs moblty and decreases upward movements wthn the labor earnngs dstrbuton for several years after job loss occurs. Understandng that workers have access to resources that may buffer the negatve consequences of job dsplacement, the analyss also ncorporates earnngs and ncome from other famly members and government transfer payments. When consderng these other resources, dsplacement s short-run mpact on moblty s reduced, and the long-term mpact s elmnated. Studyng dsplacement s effect on moblty s mportant snce movements wthn the ncome dstrbuton have mplcatons for polces desgned to combat nequalty. 1 Dsplacement may rase the probablty of ncreased ncome nequalty snce ths type of nvoluntary job loss permanently reduces workers earnngs relatve to non-dsplaced ndvduals. 2 Berry, Gottschalk, and Wssoker (1988) and Stevens (2001) fnd the transtory varance of dsplaced workers earnngs ncreases upon job loss. Ths varance not only shows ncreases n statc, year-to-year measures, but also shows an ncreasng trend over tme (Stevens 2001). The ncreased volatlty of dsplaced workers earnngs s naturally a polcy concern, but the more relevant queston s whether short-term volatlty s permanent or offset by long-term 1 Economsts have questoned why nequalty has changed over tme n the Unted States (Gottschalk and Mofftt 1994). Furthermore, researchers have conducted cross-natonal comparsons n order to judge the relatve sze of nequalty n the Unted States (Burkhauser, Holtz-Eakn, and Rhody 1997). Fnally, Burkhauser et al. (1999) and Burkhauser et al. (2004) examne the shape of the earnngs dstrbuton and provde emprcal tests to show how the dstrbuton has changed over tme. 2 See Ruhm (1991), Jacobson, LaLonde, and Sullvan (1993a, 1993b), Stevens (1997), and Couch and Placzek (forthcomng) for dscussons of dsplacement s negatve mpact on earnngs. 1

3 upward moblty as earnngs recover. Lttle emprcal research exsts on dsplacement and moblty (Berry et al. 1988; DPrete 2002). The papers that do nvestgate ths topc fnd that dsplacement not only ncreases the probablty of workers recevng low levels of labor earnngs, but also ncreases the probablty of fallng nto poverty. Other research has shown that all types of nvoluntary job loss (ncludng dsplacement) reduce the probablty of movng from the bottom quntle and remanng n the top quntle of the ncome dstrbuton (Gttleman and Joyce 1999). Ths study uses a methodology that extends the prevous research. Usng the 1968 through 1993 waves of the PSID, the analyss begns by usng transton probabltes to compare the moblty patterns of dsplaced workers to a comparson group of never-dsplaced ndvduals. The study then uses a standard earnngs equaton n a latent varable model to provde estmates of the long-term earnngs losses of dsplaced workers. These estmates are then used to calculate the probablty of a dsplaced ndvdual beng n any decle of the earnngs and ncome dstrbutons relatve to non-dsplaced workers. Fnally, nonparametrc kernel densty estmators are used to analyze vsually the movements of the earnngs and ncome dstrbutons of dsplaced workers over tme relatve to the year of job loss. The rest of ths paper proceeds by dscussng the lterature on earnngs losses and the ncome dstrbuton of dsplaced workers. Secton III dscusses the data and emprcal methodology. Secton IV presents the emprcal results, and Secton V concludes. II. Prevous Lterature Emprcal fndngs n the lterature suggest that job dsplacement should affect earnngs and ncome moblty. Researchers have found that the average level of dsplaced workers earnngs falls sgnfcantly mmedately followng job loss (Ruhm 1991; Jacobson, LaLonde, and 2

4 Sullvan 1993a, 1993b; Stevens 1997; Couch and Placzek forthcomng). Several years after the event occurs, earnngs are stll below where they would be had dsplacement not occurred. These fndngs mply two results. Frst, snce labor earnngs sgnfcantly decrease n the year of job loss, the probablty of downward earnngs and ncome moblty should ncrease. Second, snce earnngs are stll below those of non-dsplaced workers even several years after job loss occurs, the probablty, and amount, of relatve upward moblty should decrease. Economsts have several theores as to why workers lose substantal earnngs upon dsplacement, and these theores have mplcatons for ncome moblty after job loss. 3 Once dsplacement occurs, ndvduals not only lose frm, ndustry, and unon wage premums, but also hgh qualty matches wth ther former employers. If frms mantan promoton from wthn polces, re-employed dsplaced workers wll have dffculty ncreasng ther earnngs by movng up the organzatonal ladder. Addtonally, workers may lose any frm/ndustry-specfc human captal after dsplacement. 4 There are other reasons why dsplacement should affect ntrageneratonal earnngs moblty. Indvduals dffer n ther ablty to adjust to job loss, and they may accept volatle earnngs n an attempt to mantan the same expected level of ncome (Berry et al. 1988). In addton, t may take tme for workers to establsh a good match wth a new employer, whch could lead to subsequent dsplacements (Stevens 1997). Farber (1999) shows dsplaced workers are more lkely to be n temporary and nvoluntary part-tme work after separaton. He notes dsplaced workers use these types of employment relatonshps as transtons nto full-tme occupatons. 3 See Fallck (1996) and Jacobson et al. (1993a, 1993b) for dscussons of these topcs. 4 Carrngton (1993) and Neal (1995) emprcally show ths type of human captal s mportant n determnng the recovery of dsplaced workers lost earnngs. They show that those workers who fnd re-employment wthn the same ndustry have smaller earnngs losses than those who swtch ndustres after job loss occurs. 3

5 Whle the above results mply a drectonal mpact of dsplacement on moblty, lttle research exsts that calculates the magntude of the effect (Berry et al. 1988; DPrete 2002). Usng an errors components model and the PSID, Berry et al. (1988) fnd that the probablty of dsplaced workers earnng less than $10,000 n a gven year ncreases from two years before separaton to the year after job loss occurs. Four years after separaton, the authors fnd the proporton of dsplaced workers wth earnngs below ths threshold s In a cross-natonal comparson of Sweden, Germany, and the Unted States (US), DPrete (2002) fnds the probablty of a US household experencng a dsplacement and enterng poverty s between and In hs study, poverty s defned as beng less than 50 percent of the medan-adjusted household dsposable ncome, adjusted for famly sze. Whle not focusng exclusvely on job dsplacement, Gttleman and Joyce (1999) use the PSID and probt models to nvestgate how voluntary and nvoluntary job separatons occurrng over a fve-year perod affect moblty. Ther defnton of nvoluntary job loss ncludes not only dsplacement, but also job was completed, temporary work, and seasonal occupaton. The authors fnd nvoluntary job loss reduces the probablty of movng from the bottom quntle of the ncome dstrbuton by fve percentage ponts. Involuntary separatons also reduce the probablty of stayng n the top quntle by 20 ponts. Ths study contrbutes to the exstng lterature by analyzng three dfferent measures of fnancal well beng: annual labor earnngs; the combned earnngs of the husband and wfe; and pre-tax, post-transfer total famly ncome, whch ncludes earnngs and ncome from all famly members and government transfer payments. The latter two are measured on a per capta bass. Prevous papers examned only one of these measures. Berry et al. (1988) used annual labor 4

6 earnngs; DPrete (2002) analyzed sze-adjusted household dsposable ncome, and Gttleman and Joyce (1999) focused on sze-adjusted famly ncome. It s mportant for any study of dsplacement and moblty to consder these varous measures of fnancal well beng. Labor earnngs are the reward an ndvdual receves for partcpatng n the workforce. Dsplacement drectly alters ths reward by possbly reducng hours worked, causng spells of unemployment, and destroyng frm/ndustry-specfc human captal. However, ndvduals may have access to other sources of ncome that protect aganst dsplacement s negatve nfluence on earnngs. Therefore, t s mportant to consder the possblty that the dsplaced worker has access to earnngs and ncome from other famly members and government transfer payments. III. Data and Emprcal Methodology Data Ths study uses data from the 1968 through 1993 waves of the Panel Study of Income Dynamcs (PSID). The PSID s a natonally representatve survey conducted annually between 1968 and 1997, and bennally thereafter. 5 It ncludes an over-sample of low-ncome households, and the results reported n the next secton come from usng both the low-ncome and natonally representatve samples. 6 To avod any potental labor market adjustments made by females such 5 Researchers also use the Dsplaced Workers Survey (DWS) to study dsplaced workers. Whle ths data source has hgh qualty nformaton on the ncdence of dsplacement, t has three major shortcomngs. Frst, there s no natural comparson group avalable because the DWS only surveys ndvduals who experence dsplacement. Madden (1988) shows the mportance of usng a comparson group when studyng dsplacement s effect on earnngs. Second, the DWS asks respondents about the most recent job loss that occurs between three and fve years before the actual survey date. Therefore, the DWS may have more recall bas and measurement error compared to the PSID. Fnally, the DWS only nqures about the most recent pre-dsplacement job. Therefore, a long earnngs hstory s not avalable. 6 Appendx B contans the results from the majorty of the analyss usng the natonally representatve sample. The qualtatve results stll hold. 5

7 as marrage, dvorce, and chld rearng, the unt of analyss s male household heads. 7 However, the analyss ncludes the labor earnngs and ncome of wves and other famly members, along wth government transfer payments. In each wave of the PSID, the ncome varables refer to the prevous calendar year. Therefore, the estmaton occurs from 1968 to Instead of relyng on the orgnal 1968 sample, ndvduals can enter the PSID samplng frame as tme progresses. The only restrcton placed on these ndvduals s that when they enter the PSID, they report beng a head of household untl The estmaton s on all ndvduals n the years when they are between the ages of 25 and 61 and report non-zero labor earnngs. 8 A natural concern wth restrctng the estmaton to those wth postve labor earnngs s mssng some potentally nterestng analyss of movements out of and nto the labor force. For ths reason, Appendx C contans the majorty of the results when ncludng observatons of zero labor earnngs. The qualtatve results stll hold. The analyss uses three dfferent measures of fnancal well beng, whch are converted to real dollars usng the approprate year s CPI-U wth =100. The frst measure s annual labor earnngs, whch nclude total wage and salary ncome, earnngs from overtme, bonuses, and commssons, and the labor porton of farm, busness, and roomers and boarders ncome. The analyss begns wth ths measure because labor earnngs are the drect reward for an ndvdual s nvolvement n the workforce. Dsplacement negatvely alters ths nvolvement by reducng hours worked, causng spells of unemployment, or destroyng frm/ndustry-specfc human captal. 7 The results are smlar when ncludng female-headed households. 8 The age restrcton avods potental retrement decsons. Borjas (2005) notes two-thrds of men retre between ages 62 and 65. When defnng the sample of dsplaced workers, ndvduals must be no older than 56 at the tme of job loss so they have the ablty to be present durng the follow-up perod. 6

8 The second measure of well beng s the summaton of head and wfe labor earnngs, whch s referred to throughout ths paper as annual parental labor earnngs. Fnally, the analyss uses total famly ncome, whch equals the sum of labor earnngs and unearned ncome from all members n the famly unt, ncludng government transfer payments. The analyss uses these two measures of well beng because earnngs and ncome from other famly members and government transfer payments may offer protecton aganst negatve ncome shocks such as dsplacement (Setchk 1991; Stephens 2002). These last two measures of well beng are adjusted for famly sze by dvdng them by the number of members n the famly unt. 9 All earnngs are pre-tax, and per capta famly ncome s pre-tax, post-transfer (Berry et al. 1988; Gttleman and Joyce 1999). From the group of males meetng the above restrctons, dsplacement s dentfed from a queston asked of those workers who have been wth ther current job/employer for less than 12 months or snce January of the prevous year. The queston asks why the worker changed jobs/employers. If the respondent states the reason s plant closure or lay-off/fre, then he s dentfed as experencng a dsplacement. Ths s consstent wth prevous research on dsplacement usng the PSID (Stevens 1997; Stevens 2001; Stephens 2002; Charles and Stephens 2004). Dsplacement s tmed as occurrng n the calendar year before the survey wave (Stephens 2002; Charles and Stephens 2004). Fnally, n the 1968 survey, the queston refers to dsplacements occurrng over the prevous ten years. Snce dsplacements reported n the 1968 survey cannot be tmed, anyone reportng dsplacement durng that wave s removed from the analyss. 9 Researchers have documented the ncreased probablty of dvorce assocated wth job dsplacement (Charles and Stephens 2004). Snce the analyss adjusts parental earnngs and famly ncome for famly sze, t may be the case that a husband wth a lower earnng wfe wll appear to have hgher famly earnngs and ncome upon dvorce. Because of ths possblty, senstvty checks usng a sample of contnuously marred couples are conducted. The qualtatve results are unchanged and presented n Secton IV. 7

9 Ths defnton of dsplacement has two potental problems. Frst, the PSID does not delneate between frng for cause and mass layoff. If those workers who are fred for cause have below-average productvty, then ths may bas the parameter estmates downwards. However, Bosjoly, Duncan, and Smeedng (1994) note that only 16 percent of those who report ladoff/fred are actually fred for cause (Stevens 1997). Second, ths defnton of dsplacement does not specfcally conform to that of the Bureau of Labor Statstcs (BLS). The BLS defnes a dsplaced ndvdual as someone who s at least 20 years old, has at least three years of tenure, and lost a job due to plant closure, abolton of a poston or shft, or slack work. The tenure porton of the BLS defnton s dffcult for researchers to mplement wth the PSID for two reasons. Frst, the codng of tenure changes from an nterval to the actual number of months startng n the 1976 survey wave. Second, the type of tenure asked of respondents changes, varyng from tenure on the current job, poston, and employer. These are dfferent concepts. Because of these changes, and snce t s mportant to follow workers wth some attachment to the labor market, the dsplaced workers dentfed above must have three consecutve years of postve labor earnngs before dsplacement occurs. Ths restrcton requres those who experence a dsplacement between 1968 and 1970 to be removed from the analyss. Upon mplementng the selecton rules, 3,410 ndvduals meet all of the sample selecton crtera. Of these, 584 experenced a dsplacement between 1971 and Emprcal Methodology The focus of ths paper s not on how an ndvdual s earnngs vary wth dsplacement, per se. Instead, nterest les n how ths type of job loss affects an ndvdual s movement and rankng n the earnngs and ncome dstrbutons. To study these concepts, ths paper employs three dfferent methodologes: transton matrces, correlated random-effects nterval regressons, 8

10 and kernel densty estmates. Each of these technques provdes a dfferent perspectve on how dsplacement alters a worker s ablty to move throughout the ncome dstrbuton over tme, and each s dscussed n turn. Transton Matrces Transton matrces provde useful summary measures of the probablty of workers movng throughout the earnngs and ncome dstrbutons over relatve tme changes. Furthermore, these matrces provde nsght nto the persstence of ncome shocks. If these shocks tend to be transtory, then the probablty of changng earnngs or ncome decles s the same over a three-year perod as over a one-year perod (Burkhauser, Holtz-Eakn, and Rhody 1997). However, f shocks tend to persst, then the probablty of changng decles grows over tme. Therefore, usng transton matrces to study dsplacement s effect on moblty wll yeld evdence as to how persstent ths negatve shock s to earnngs and ncome. To calculate the transton probabltes, the decles of the earnngs and ncome dstrbutons need to be determned. The decles are generated usng dstrbutons over the entre 25-year perod. Ths method s dfferent from generatng the dstrbutons n each year. By calculatng the decles n that manner, there are nstances when the upper and lower bounds from adjacent decles overlap between years. These bounds must reman fxed when usng the nterval regresson (descrbed below) to calculate the probablty of an ndvdual beng n any one decle. Therefore, to be consstent across methodologes, the dstrbuton s generated over the entre sample perod. After creatng the decles, the sample s broken nto two subsets. The frst contans never-dsplaced workers, and the second contans workers who experenced a dsplacement at some pont between 1971 and Usng a smlar methodology and notaton as Burkhauser et 9

11 al. (1997), ndcator varables are created for each ndvdual n each subset that capture the t, t+ r movement from one decle to another. These ndcators, ρ, equal 1 f ndvdual moves from decle d to decle l between perods t and, d, l transtonng between decles s gven by the followng equaton: t+ r. For each subgroup, the probablty of P n t, t+ r wρ, d, l = 1 = n = 1 w (1), where w s the weght assgned to ndvdual n perod t+ r. Weghts are used because of the presence of the over-sample of low-ncome households. The use of weghts s complcated by the fact that transton matrces nherently examne movements between multple perods. In the PSID, the ndvdual weghts are not comparable between survey waves. Weghts n the termnal year of the transton are used for the calculatons. For example, f the movement of an ndvdual occurs between 1990 and 1991, the 1991 weghts are used. The reference perod (tme t ) s dfferent for each subset. For those workers who experence a dsplacement, the reference pont s the perod three years before the job loss. Choosng a startng pont for non-dsplaced workers s more complcated. Ths dffculty arses because dsplacement may have occurred any tme between 1971 and Ths feature of the data makes t dffcult to algn temporally those who experence a dsplacement and those who do not. For ths reason, a random reference perod s chosen for each ndvdual n the group of never-dsplaced workers For comparablty, the actual date of entry nto the sample was also used. The qualtatve results are smlar. 10

12 Correlated Random-Effects Interval Regresson The second methodology used s a correlated random-effects nterval regresson. The nterval regresson s a latent varable model that treats earnngs and ncome as unobserved varables that fall wthn a pre-determned range. Ths range s the lower and upper bounds of the decles of the earnngs and ncome dstrbutons. The structure of ths model smultaneously captures two components. Frst, t uses a standard earnngs equaton to model the wage determnaton process and to provde estmates of the long-term mpact of dsplacement on earnngs and ncome. Second, by treatng earnngs and ncome as latent varables, the parameter estmates can be used to predct the probablty of a dsplaced worker beng n any decle of the earnngs and ncome dstrbutons. The model begns wth the standard earnngs equaton appled to panel data for the th ndvdual, y = β + µ * t z t t (2), where µ = ν + ε (3). t t Here, * y t s annual labor earnngs, per capta parental earnngs, or per capta famly ncome of person n year t. The zt contans human captal characterstcs thought to affect earnngs, and µ t s the error term. In (3), ν s a tme-nvarant, unobserved, ndvdual-specfc effect assumed ndependent of z t ; ε t s a tme-varyng error. Both are ndependently and randomly 2 2 dstrbuted as normal wth mean zero and varances σ ν and σ ε, respectvely. The model treats the earnngs and ncome measures n (2) as latent, unobserved varables that fall wthn a pre-determned, observable range. Ths range equals the lower and upper 11

13 bounds of each decle of the earnngs and ncome dstrbutons calculated from the data. Therefore, for the th ndvdual, defne y t to equal one of the decles as follows: * 1 f yt α1 * y t = 2,...,9 f α j 1 < yt α j j = 2,...,9 (4). 10 f * yt > α9 Here, theα s are the lower and upper bounds of the ncome decles, whch are fxed, known j parameters taken from the data. By pluggng (3) nto (2) and (2) nto (4), wth rearrangng, (4) becomes y t 1 2,...,9 10 f = f α j 1 f ε α ( z β + ν ) t t 1 ( z β + ν ) < ε 9 t ε > α ( z β + ν ) t t t α ( z β + ν ) j = 2,...,9. j t The probablty that y t takes on the value of any one of the decles s Pr( y t = 1,...,10 z t α1 ( ztβ + ν ) Φ, j= 1 σε α j ( ztβ + ν ) α j 1 ( ztβ + ν ), ν ) = Φ Φ, j = 2,...,9 (5). σε σε α 9 ( ztβ + ν ) 1 Φ, j= 10 σε where Φ (.) s the standard normal cumulatve densty functon. Assumng condtonal ndependence over tme, the jont densty of y s where f ( y,..., y z,..., z, β, ν, σ ) = F( y z, β, ν, σ ), 1 T 1 T ε t t ε t= 1 T 12

14 F( y t z t α1 ( ztβ+ ν ) Φ, σε α j ( ztβ+ ν ) α j 1 ( ztβ+ ν ), β, ν, σε ) = Φ Φ, σε σε α9 ( ztβ+ ν ) 1 Φ, σε The lkelhood functon for the th ndvdual s f f f y t y y = 1 = 2,...,9 t t = 10 L = σ ν 2 ν 1 2 e 2π 2σ T ν F( y z, β, σ, ν ) dν (6). t= 1 t t ε The analyss estmates equaton (6) usng adaptve Gauss-Hermte quadrature wth 36 ntegraton ponts. 11 The parameter estmates obtaned from ths model are then used n equaton (5) to estmate the probablty of a dsplaced worker beng n any one decle of the earnngs and ncome dstrbutons relatve to ther non-dsplaced counterparts. The random-effects model assumes the observed covarates are uncorrelated wth the unobserved heterogenety. If ths assumpton s true, then the random-effects model produces consstent estmates. Ths s not the case f the assumpton s volated. In fact, Gbbons and Katz (1991) provde emprcal evdence showng dsplaced workers are nherently dfferent from other workers, and these dfferences are productvty-related. Because of ths fndng, researchers control for the potental correlaton that may exst between the observed covarates and the unobserved heterogenety. Snce the standard normal dstrbuton s a sngle ndex functon, t s not possble to factor out ν from the model (Cameron and Trved 2005). Furthermore, 11 When solvng ths model, a trade-off exsts between computaton tme and precson of the ntegral s estmate (Butler and Mofftt 1982). Increasng the number of ntegraton ponts adds to computaton tme whle ncreasng the precson of the estmate. To speed computaton tme, the analyss uses parameter estmates from the pooled verson of the model as startng values. To show the number of ntegraton ponts does not sgnfcantly alter the parameter estmates, Appendx Table A-1 presents estmates when runnng ths model usng the labor earnngs dstrbuton and a range of ntegraton ponts from four to 36. As shown n the table, alterng the number of quadratures does not affect the results n any meanngful way. 13

15 ncluson of ndvdual dummy varables may lead to the ncdental parameters problem. Therefore, t s not possble to use standard fxed-effects technques here. To control for ths possble correlaton, the analyss follows Mundlak (1978) and proposes to parameterze the relatonshp between ν and the observed covarates. 12 Here, ν s lnearly related to the observed covarates asν = z λ+ u, where z s the covarate s average 2 for ndvdual over tme, and u z ~ N(0, σ ). 13 Ths s known as Mundlak s verson of the u correlated random-effects model. 14 The coeffcents assocated wth the z are nterpreted as the effect unobserved heterogenety has on the dependent varable, whereas those assocated wth z t provde the true effect of the varable of nterest. In order to estmate the lkelhood functon, (2) becomes: y * t k = x θ + xπ + D δ + D ξ + γ + γ + η. t k k 3 s k k 3 k t t Ths equaton s smlar to models used by Jacobson et al. (1993a, 1993b) and Couch and Placzek (forthcomng). Here, the structure of the model remans the same wth k D δ + k 3 s k k ztβ = xtθ + γ t, zλ = xπ + 3 D ξ k + γ, and η t = u + ε t. Theγ t s are year k dummy varables. The k D s s a dummy varable equalng one n year s f ndvdual suffers dsplacement, and k ndexes these varables startng three years before job loss. Fnally, the x t contans a quartc n potental experence. Potental experence equals age mnus educaton mnus sx. If the ndvdual has less than 12 years of educaton, then potental experence equals 12 See Wooldrdge (2002) and Cameron and Trved (2005) for a dscusson. 13 For a recent applcaton usng ths technque to account for the potental correlaton between the tme-nvarant, unobserved heterogenety and the observed covarates n a non-lnear panel data model, see Lorgelly and Lndley (2008). 14 Chamberlan (1984) dscusses a dfferent verson of the correlated random-effects model. He uses each observaton of every covarate as an explanatory varable. However, ths methodology requres a balanced panel (Stephens 2002; Sahm 2007). The analyss presented below comes from an unbalanced panel. When completely balancng the data, the sample sze falls from 3,410 to 243 male household heads. 14

16 age mnus 18 so not to overcompensate less educated workers by assgnng them larger values of experence (Stephens 2002). Educaton s defned to be the same throughout tme. Ths s done by assgnng each ndvdual hs educaton as reported n the most recent survey wave for whch that person reported educaton. Kernel Densty Estmates The fnal methodology used s nonparametrc kernel densty estmaton. Ths methodology provdes estmates of the dstrbuton of a varable wthout placng any pror assumptons on the data. Therefore, the dstrbuton s not assumed to follow a specfc functonal form. These estmators provde vsual representatons of the earnngs and ncome dstrbutons. When graphng these estmators for dsplaced workers over tme relatve to job loss, researchers can see the potental mpact dsplacement has on the earnngs and ncome dstrbutons. The estmators are constructed usng the followng formula: n 1 κ t κt p( κ t ) = K (7), nb = 1 b where κ equals annual labor earnngs, per capta parental earnngs, or per capta famly ncome, n s the sample sze, b s the bandwdth, and K (.) s the kernel functon. The analyss uses the Epanechnkov kernel snce t s the most effcent (Pagan and Ullah 1999). Ths kernel functon equals the followng: / K( ) = , < 5. otherwse The bandwdth s chosen such that the mean ntegrated squared error of the estmate s mnmzed assumng the data follow a Gaussan dstrbuton and a Gaussan kernel were used. 15

17 IV. Results Tables 1 and 2 show some descrptve measures of the sample. Table 1 presents the number of dsplaced and non-dsplaced ndvduals n each year. The number of those experencng a job dsplacement ncreases over tme. Ths s consstent wth prevous research (Stevens 2001). Table 2 shows descrptve statstcs of the sample by dsplacement status. Those who experence dsplacement tend to have relatvely lower ncomes of all types. In fact, the lower ncomes are the only notceable dfferences between the groups. These lower ncomes are expected snce the averages are calculated over the entre 25-year perod. Calculatng the average n ths manner automatcally consders any effect dsplacement has on earnngs and ncome. Table 3 presents parameter estmates from three fxed-effects regressons. The dependent varables n the regressons are annual labor earnngs, per capta parental earnngs, and per capta famly ncome; the rght-hand-sde varables nclude a quartc n potental experence, year dummy varables, and the dsplacement dummy varables. Snce the parameter estmates assocated wth the dsplacement varables are the ones of nterest, they are the ones shown n the table. The purpose of presentng these results s to show the extent of dsplacement s effect on the ntal drop and subsequent recovery of earnngs and ncome over tme. Table 3 shows job dsplacement reduces annual labor earnngs by $3,891 the year of job loss. Fve years followng the event, annual labor earnngs are stll $2,769 below where they would have been had dsplacement not occurred. These estmates are sgnfcantly dfferent from zero at the one percent level. Dsplacement also negatvely nfluences per capta parental earnngs and famly ncome. Table 3 shows per capta parental earnngs and famly ncome fall by $1,543 and $1,232, respectvely, the year of job loss. Fve years after dsplacement, the losses 16

18 n parental earnngs and famly ncome equal $853 and $878 and are sgnfcantly dfferent from zero at the fve percent level. The results shown n Table 3 ndcate earnngs and ncome fall sgnfcantly the year of job loss and stll do not recover to pre-dsplacement levels fve years after the event occurs. When examnng the parameter estmates n Table 3 as a percentage of earnngs and ncome the year before job loss, the analyss shows that the declnes n per capta parental earnngs and famly ncome are lower than the declne n annual labor earnngs. Table 3 shows that the loss n annual labor earnngs the year of dsplacement s 16.7 percent of pre-dsplacement earnngs. For per capta parental earnngs and famly ncome, the losses are 14.7 and 10.3 percent, respectvely. Fve years followng the event, the loss n annual labor earnngs s 11.9 percent of pre-dsplacement earnngs; for parental earnngs and famly ncome, the losses are 8.1 and 7.4 percent. Ths result shows that once workers have access to other sources of ncome, the short- and long-term negatve effects of dsplacement on fnancal well beng are reduced. Table 4 contans three panels showng transton probabltes by dsplacement status and ncome type. Panel A uses labor earnngs, Panel B uses per capta parental earnngs, and Panel C uses per capta famly ncome. The columns labeled t+ r ndcate relatve tme changes for the dsplaced and non-dsplaced workers. For example, concentratng on dsplaced workers, the column labeled t + 1 ndcates a change from three years before job loss to two years before job loss. For the group of non-dsplaced workers, t + 1 shows the movement from the random startng date to one year afterwards. The rows n Table 4 show the assocated movement wthn the varous dstrbutons. For example, the row labeled Down 9 ndcates a downward movement of nne decles n the dstrbuton durng the perod ndcated n the column. As n Burkhauser et al. (1997), the entres 17

19 shown n each cell are derved from equaton (1), and they show the proporton of ndvduals who make the assocated transton relatve to the number of people elgble to make the move. Column totals do not sum to 100 percent as a result. These transton matrces are useful n showng moblty patterns over relatve tme changes. The cells n bold ndcate statstcal dfferences between dsplaced and non-dsplaced workers at the fve percent sgnfcance level. These dfferences provde some evdence as to how dsplacement affects worker moblty relatve to a group of never-dsplaced ndvduals. Each panel n Table 4 shows there are two general patterns exhbted by both dsplaced and non-dsplaced ndvduals. Frst, the proporton of mmoble workers declnes as tme progresses. For example, n Panel A durng perod t + 1, the proporton of non-dsplaced workers who are mmoble equals approxmately 50 percent. By perod t + 8, ths number declnes to 25 percent. The same fgures for dsplaced workers are 47 percent and 23 percent, respectvely. The second pattern, whch s related to the frst, s the general ncrease n the probabltes of changng decles over tme. For example, agan concentratng on Panel A, the proporton of non-dsplaced workers movng down four decles n perod t + 1 equals 0.6 percent. The proporton s 2.1 percent n perod t + 8. The same numbers for dsplaced workers are 2.3 percent and 9.5 percent, respectvely. Snce the probablty of changng decles generally ncreases over tme, shocks to earnngs and ncome tend to persst. Even though dsplaced and non-dsplaced workers share the two general moblty patterns mentoned above, Panel A shows there are many statstcal dfferences between the two groups begnnng n perod t + 3. Recall Panel A uses annual labor earnngs, and perod t + 3 s a change from three years before to the year of job loss for dsplaced workers. Startng n ths perod, dsplaced ndvduals have much larger probabltes of droppng n the earnngs 18

20 dstrbuton. For example, the proporton of dsplaced workers declnng three decles s 11.7 percent. For non-dsplaced workers, the proporton s three percent. Workers experencng a dsplacement also have lower probabltes of movng up the ncome dstrbuton startng n perod t + 3. Ths s partcularly the case for movng up one and two decles wthn the dstrbuton over tme. When movng from Panel A to Panels B and C, the number of statstcal dfferences between dsplaced and non-dsplaced workers greatly dmnshes. Panel B, whch uses per capta parental earnngs, shows there are almost no statstcal dfferences between these groups of workers startng n perod t + 7. Ths fndng ndcates that four years after job loss, dsplaced workers exhbt the same moblty patterns as ther non-dsplaced counterparts. Panel C examnes the per capta famly ncome dstrbuton and shows that any major statstcal dfferences are erased n perod t + 7. Whle the transton matrces n Table 4 provde good summary measures of moblty, they do not condton on factors affectng an ndvdual s wage. Furthermore, they do not control for any possble selecton nto dsplacement. Table 5 shows the results from estmatng the lkelhood functon of the correlated random-effects nterval regresson, equaton (6). These regressons control for a quartc n potental experence, year effects, dsplacement, and tme averages of all of these varables. The estmated coeffcents found n Table 5 should be nterpreted as the effect the covarates have on the latent varable (Wooldrdge 2002). Table 5 ndcates all three measures of well beng fall the year of dsplacement. Furthermore, each ncome measure shows some recovery thereafter. Annual labor earnngs fall $4,484 the year of job loss and are $1,758 below expectatons fve years afterwards. Per capta parental earnngs declne $1,806 the year of job loss, and per capta famly ncome falls $1,

21 All four of these estmates are statstcally sgnfcant at the one percent level. However, fve years after dsplacement occurs, the coeffcents assocated wth parental earnngs and famly ncome are not statstcally dfferent from zero. Ths fndng ndcates that once dsplaced workers access other sources of ncome, the long-term negatve consequences of job dsplacement are elmnated. Table 6 presents the margnal effects of dsplacement on the probablty of beng n any one of the decles of the earnngs and ncome dstrbutons. 15 The margnal effects are calculated from equaton (5) usng the parameter estmates from the lkelhood functon, equaton (6). Panel A shows the effects usng the labor earnngs dstrbuton. Panels B and C show the effects usng the per capta parental earnngs and famly ncome dstrbutons, respectvely. Concentratng on Panel A, the results show the probablty of a dsplaced worker beng n the bottom decle the year of job loss s almost eght percentage ponts larger than for a nondsplaced worker. Fve years after dsplacement, the ncreased probablty s stll over two percentage ponts. The magntude of dsplacement s effect decreases when movng from the tals to the center of the dstrbuton. Panel A shows the year of job loss, a dsplaced worker s 0.41 percentage ponts more lkely to be n the ffth decle when compared to a non-dsplaced worker. Horzontally summng the rows shows the cumulatve effect dsplacement has on beng n partcular portons of the dstrbuton. For example, Panel A shows the probablty of a dsplaced worker beng n the bottom fve decles ncreases by almost 14 percentage ponts the year of job 15 The margnal effects are calculated wth all of the varables set at ther means except for the dsplacement varables. The dsplacement varables are set to zero when calculatng the margnal effects. 20

22 loss. Ths ncreased probablty s over fve percentage ponts fve years afterwards. 16 Ths result ndcates job dsplacement not only ncreases the probablty of downward earnngs moblty the year of job loss, but also decreases the probablty of upward moblty after dsplacement occurs. Panel B shows that the ncreased probablty of beng n the bottom fve decles the year of dsplacement s over nne percentage ponts; t s almost seven ponts n Panel C. Fve years followng dsplacement, the ncreased probablty of an ndvdual beng n the bottom fve decles s 1.9 and 1.2 percentage ponts n Panels B and C, respectvely. Furthermore, n Panels B and C, these ncreased percentages are not sgnfcantly dfferent from zero startng four years followng dsplacement. Ths fndng ndcates that dsplaced workers have the same probablty of beng n any one decle as ther non-dsplaced counterparts begnnng four years after job loss. The percentages n Panels B and C are smaller than the numbers n Panel A. Ths shows that earnngs and ncome from other famly members and government transfer payments offer protecton aganst the adverse effects of dsplacement. 17 The analyss presented above uses famly-sze adjusted parental earnngs and famly ncome. Research shows that the probablty of dvorce ncreases upon dsplacement (Charles and Stephens 2004). If a husband has a lower earnng wfe, and f the couple dvorces after job loss occurs, then the husband wll appear to have larger per capta parental earnngs and famly ncome after dsplacement and dvorce. To see f dvorce s drvng the results, the correlated random effects nterval regresson was run on the analyss sample after selectng only those males who remaned contnuously marred to the same person. Table 7 presents the results. 16 Because of the symmetrcal nature of the normal dstrbuton, ths mples the probablty of a dsplaced worker beng n the top fve decles decreases by 14 percentage ponts the year of dsplacement and sx percentage ponts fve years afterwards. 17 Tables 5 and 6 were recreated usng the square root equvalence scale for parental earnngs and famly ncome. The qualtatve results stll hold. 21

23 Table 7 shows that annual labor earnngs declne $4,783 the year of dsplacement and are $2,423 below expectatons fve years afterwards. Both estmates are statstcally dfferent from zero at the one-percent level. Per capta parental earnngs and famly ncome declne $1,534 and $1,165 the year of job loss, respectvely. Fve years after dsplacement occurs, the losses are $787 and $572, and the estmates are nsgnfcantly dfferent from zero begnnng three years after dsplacement. These results suggest that changes n martal status are not drvng the qualtatve fndngs presented above. Fgure 1 presents kernel densty estmates of the dstrbuton of logged labor earnngs for dsplaced ndvduals for selected years relatve to the year of job loss. These estmates provde good vsual representatons of dsplacement s effect on the earnngs dstrbuton. As Fgure 1 shows, the shape and locaton of the earnngs dstrbuton changes sgnfcantly over tme. The year of job loss s assocated wth a leftward shft and a flattenng of the dstrbuton when compared to the kernel three years before dsplacement. A Kolmogorov-Smrnov (K-S) test rejects the null hypothess at the one-percent level that the two dstrbutons are equal. 18 Fve years after job loss, the earnngs dstrbuton s smlar to the one three years before dsplacement. A K-S test cannot reject the null that the dstrbutons are equal. Fgures 2 and 3 show kernel estmates for the log of per capta parental earnngs and famly ncome, respectvely. The denstes shown n Fgures 2 and 3 follow the same general trends as those shown n Fgure 1. The dstrbutons the year of job loss are assocated wth a leftward shft when compared to the dstrbutons three years before dsplacement. The dstrbutons fve years followng dsplacement are smlar to the ones three years before the event. Fgure 2 shows that the kernel denstes are closer together than they are n Fgure 1. The dstrbutons n Fgure 3 are nearly ndstngushable from one another. When usng the K-S 18 The sgnfcance tests are conducted on the emprcal dstrbutons of the data as n Burkhauser et al. (1999). 22

24 statstc to test for sgnfcant dfferences between the dstrbutons over tme, smlar results found for the dstrbutons of logged labor earnngs are found here. For parental earnngs, the dstrbuton the year of job loss s sgnfcantly dfferent from the one three years before dsplacement at the one-percent level; for famly ncome, they are dfferent at the fve-percent level. The dstrbuton fve years after dsplacement s not statstcally dfferent from the one three years before job loss for both parental earnngs and famly ncome. V. Conclusons Ths study uses data drawn from the 1968 through 1993 waves of the PSID to examne how job dsplacement nfluences the nter-temporal movement of workers through the earnngs and ncome dstrbutons. The results show that job dsplacement sgnfcantly reduces annual labor earnngs, per capta parental earnngs, and per capta famly ncome n the year of separaton. Even several years after job loss occurs, these earnngs and ncome measures are stll below where they would have been had dsplacement not occurred. However, when examnng these losses as a percentage of pre-dsplacement earnngs and ncome, the losses found for per capta parental earnngs and famly ncome are lower than the percentage losses of annual labor earnngs. Ths result shows that once workers access other sources of ncome, the short- and long-term negatve mpact of dsplacement s reduced. The analyss s extended to show how dsplacement affects earnngs and ncome moblty. The results ndcate that the deep earnngs losses caused by dsplacement ncrease downward earnngs moblty not only durng the year of job loss, but also for fve years afterwards. Furthermore, upward moblty s decreased for several years followng the event. Results also show that durng the year of job loss, dsplaced workers experence an ncreased probablty of 14 percentage ponts of beng n the bottom half of the labor earnngs dstrbuton. 23

25 Dsplacement s negatve nfluence on earnngs moblty dsspates as tme progresses. Fve years followng job loss, the ncreased probablty reduces to less than sx percentage ponts. After accountng for other sources of ncome, dsplacement s effect on moblty s reduced n the short-term and elmnated n the long run. The results show that the moblty patterns of dsplaced and non-dsplaced workers are equal four years after job loss occurs. Furthermore, results ndcate that four years followng job loss, dsplaced workers have the same probablty of beng n any one decle of the ncome dstrbuton as ther non-dsplaced counterparts. The results mply that polcy ntatves desgned to assst dslocated workers should consder the potental benefts provded by the earnngs of other famly members and the exstng ncome avalable from government transfer payments. 24

26 References Berry, Steve, Peter Gottschalk, and Doug Wssoker An Error Components Model of the Impact of Plant Closng on Earnngs. The Revew of Economcs and Statstcs 70(4): Borjas, George rd ed. Labor Economcs. Boston, MA: McGraw-Hll Irwn. Burkhauser, Rchard V., Kenneth A. Couch, Andrew Houtenvlle, and Ludmla Rovba Income Inequalty n the 1990s: Re-Forgng a Lost Relatonshp? Journal of Income Dstrbuton 12(3, 4): Burkhauser, Rchard V., Amy Crews Cutts, Mary C. Daly, and Stephen P. Jenkns Testng the Sgnfcance of Income Dstrbuton Changes over the 1980s Busness Cycle: a Cross-Natonal Comparson. Journal of Appled Econometrcs 14(3): Burkhauser, Rchard V., Douglas Holtz-Eakn, and Stephen E. Rhody Labor Earnngs Moblty and Inequalty n the Unted States and Germany durng the Growth Years of the 1980s. Internatonal Economc Revew 38(4): Butler, J.S., and Robert Mofftt A Computatonally Effcent Quadrature Procedure for the One-Factor Multnomal Probt Model. Econometrca 50(3): Cameron, A. Coln, and Pravn K. Trved Mcroeconometrcs: Methods and Applcatons. New York, NY: Cambrdge Unversty Press. Carrngton, Wllam J Wage Losses for Dsplaced Workers: Is t Really the Frm that Matters? The Journal of Human Resources 28(3): Chamberlan, Gary Panel Data, n Handbook of Econometrcs, eds. Z. Grlches and M.D. Intrlgator, Elsever Scence Publshers BV Amsterdam:

27 Charles, Kerwn Kof, and Melvn Stephens Jr Job Dsplacement, Dsablty, and Dvorce. Journal of Labor Economcs 22(2): Couch, Kenneth A., and Dana W. Placzek. forthcomng. Earnngs Losses of Dsplaced Workers Revsted. Amercan Economc Revew. DPrete, Thomas A Lfe Course Rsks, Moblty Regmes, and Moblty Consequences: A Comparson of Sweden, Germany, and the Unted States. The Amercan Journal of Socology 108(2): Fallck, Bruce A Revew of the Recent Emprcal Lterature on Dsplaced Workers. Industral and Labor Relatons Revew 50(1): Farber, Henry S Alternatve and Part-tme Employment Arrangements as a Response to Job Loss. Journal of Labor Economcs 17(4): S142-S169. Gbbons, Robert and Lawrence F. Katz Layoffs and Lemons. Journal of Labor Economcs 9(4): Gttleman, Maury, and Mary Joyce Have Famly Income Moblty Patterns Changed? Demography 36(3): Gottschalk, Peter, and Robert Mofftt The Growth of Earnngs Instablty n the Unted States. Brookngs Papers on Economc Actvty 1994(2): Jacobson, Lous S., Robert J. LaLonde, and Danel G. Sullvan. 1993a. Earnngs Losses of Dsplaced Workers. The Amercan Economc Revew 83(4): b. The Costs of Worker Dslocaton. W.E. Upjohn Insttute for Employment Research, Kalamazoo, Mchgan. Lorgelly, Paula K, and Joanne Lndley What s the Relatonshp between Income Inequalty and Health? Evdence from the BHPS. Health Economcs 17:

28 Madden, Jance Fannng The Dstrbuton of Economc Losses among Dsplaced Workers. The Journal of Human Resources 23(1): Mundlak, Yar On the Poolng of Tme Seres and Cross Secton Data. Econometrca 46(1): Neal, Derek Industry-Specfc Human Captal: Evdence from Dsplaced Workers. Journal of Labor Economcs 13(4): Pagan, Adran, and Aman Ullah Nonparametrc Econometrcs. New York, NY: Cambrdge Unversty Press. Panel Study of Income Dynamcs, Packaged Core Data, Documentaton, and Questonnares. Produced and dstrbuted by the Unversty of Mchgan wth prmary fundng from the Natonal Scence Foundaton, the Natonal Insttute of Agng, and the Natonal Insttute of Chld Health and Human Development. Ann Arbor, MI, Ruhm, Chrstopher J Are Workers Permanently Scarred by Job Dsplacements? The Amercan Economc Revew 81(1): Sahm, Clauda R Stablty of Rsk Preference. Fnance and Economcs Dscusson Seres, Federal Reserve Board. Workng Paper No Setchk, Adam D When Marred Men Lose Jobs: Income Replacement wthn the Famly. Industral and Labor Relatons Revew 44(4): Stephens, Melvn Jr Worker Dsplacement and the Added Worker Effect. Journal of Labor Economcs 20(3): Stevens, Ann Huff Persstent Effects of Job Dsplacement: The Importance of Multple Job Losses. Journal of Labor Economcs 15(1):

29 Stevens, Ann Huff Changes n Earnngs Instablty and Job Loss. Industral and Labor Relatons Revew 55(1): Wooldrdge, Jeffrey M Econometrc Analyss of Cross Secton and Panel Data. Cambrdge, MA: MIT Press. 28

30 Tables and Fgures Table 1: Sample Sze by Dsplacement Status Year Dsplaced Dsplaced , , , , , , , , , , , , , , , , , , , , , Unweghted sample szes. Source: waves of the PSID 29

31 Table 2: Summary Statstcs by Dsplacement Status Varable Dsplaced Dsplaced Annual Labor Income $23, $27, Annual Total Famly Income* 12, , Annual Parental Earnngs* 10, , Age Educaton Black 23.58% 23.75% Marred 85.82% 88.09% # Chldren Experence Unweghted averages and proportons. Calculatons use all person-year observatons. * Adjusted for famly sze assumng constant returns to scale n the famly. Source: 1968 through 1992 waves of the PSID. Table 3: Fxed Effects Results Annual Labor Earnngs Annual Parent Earnngs Annual Famly Income Dependent Varable 3 Years Before (2.31)* (0.74) (0.85) 2 Years Before (0.84) (0.03) (0.16) 1 Year Before (1.41) (0.85) (0.35) Year Of (6.56)** (3.97)** (2.93)** 1 Year After (6.64)** (4.14)** (3.59)** 2 Years After (4.63)** (2.94)** (3.10)** 3 Years After (4.39)** (2.75)** (3.10)** 4 Years After (3.89)** (1.65) (1.49) 5 Years After (4.99)** (2.47)* (2.35)* Observatons Number of d R-squared Robust t statstcs n parentheses * sgnfcant at 5%; ** sgnfcant at 1% All regressons nclude a quartc n potental experence and calendar dummes. Source: 1968 through 1992 waves of the PSID. 30