The impact of labor market entry conditions on initial job assignment and wages

Transcription

1 J Popul Econ (2014) 27: DOI /s ORIGINAL PAPER The mpact of labor market entry condtons on ntal job assgnment and wages Beatrce Brunner Andreas Kuhn Receved: 23 February 2012 / Accepted: 25 September 2013 / Publshed onlne: 25 October 2013 Sprnger-Verlag Berln Hedelberg 2013 Abstract We estmate the effects of labor market entry condtons on wages for male ndvduals frst enterng the Austran labor market between 1978 and We fnd a large negatve effect of unfavorable entry condtons on startng wages and a szable negatve long-run effect. Our preferred estmates mply a decrease n startng wages by about 0.9 % and a lfetme loss n wages of about 1.3 % for an ncrease n the ntal local unemployment rate by one percentage pont. We show that poor entry condtons are assocated wth lower qualty of a worker s frst employer and that the qualty of workers frst employer explans as much as three-quarters of the observed long-run wage effects resultng from poor entry condtons. Moreover, wage effects are much more persstent for blue-collar workers because some of them appear to be permanently locked n nto low-payng jobs/tasks. Keywords Intal labor market condtons Endogenous labor market entry Intal job assgnment Responsble edtor: Junsen Zhang Electronc supplementary materal The onlne verson of ths artcle (do: /s ) contans supplementary materal, whch s avalable to authorzed users. B. Brunner ZHAW Zurch Unversty of Appled Scences, Gertrudstrasse 15, 8401 Wnterthur, Swtzerland A. Kuhn ( ) Swss Federal Insttute for Vocatonal Educaton and Tranng, Krchlndachstrasse 79, 3052 Zollkofen, Swtzerland e-mal: andreas.kuhn@econ.uzh.ch B. Brunner A. Kuhn Department of Economcs, Unversty of Zurch, Zurch, Swtzerland A. Kuhn IZA, Bonn, Germany

2 706 B. Brunner, A. Kuhn JEL Classfcatons E32 J31 J62 1 Introducton The recent economc crss has renewed academc nterest n the potental mpact of busness cycle fluctuatons on labor markets (e.g., Elsby et al. 2010). However, whle labor economsts have studed on the short-run assocaton between local labor market condtons and real wages extensvely for qute some tme (e.g., Blanchflower and Oswald 1990), longer-run effects of busness cycle fluctuatons on ndvdual s wages have only more recently caught the attenton of emprcal research. 1 Clearly, even small ntal wage shortfalls may, n the longer run, eventually accrue to substantal overall losses n lfetme earnngs f ntal wage losses resultng from poor entry condtons persst. 2 Indeed, recent emprcal evdence suggests that substantal losses n lfetme earnngs result from enterng the labor market durng an economc downturn, as opposed to enterng durng an expanson. 3 Oreopoulos et al. (2012) explore the effects of enterng the labor market durng a recesson on ndvduals earnngs, usng data on Canadan college graduates who entered the labor market between 1982 and They fnd a substantal ntal wage penalty of about 9 % that only fades to zero after the frst decade of a worker s career. A smlar result s reported by Kahn (2010), who focuses on male college graduates n the USA graduatng sometme between 1979 and She fnds that the group graduatng n the worst economc stuaton ncurs a wage loss of up to 13 % each year, relatve to those graduatng n the best ntal condtons, and that ths ntal wage loss perssts over the frst 20 years of a workers labor market career. Smlar results are reported by Oyer (2006), who shows that PhD students n economcs are consderably more lkely to fnd a poston at one of the top unverstes n the USA f they graduate n tmes when the demand for economsts s hgh. In a related study, he fnds that those MBA students who complete ther tranng durng a recesson suffer from negatve effects on wages (Oyer 2008). In both studes, the long-term effects on ncome appear to stem from the fact that dverse 1 Most studes estmatng the short-run assocaton between fluctuatons n local unemployment rates and wages fnd that wages vary negatvely wth local unemployment. Ths negatve assocaton s a very robust emprcal pattern; t has been shown to exst for a wde range of dfferent countres, usng very dfferent sources of data and dverse emprcal specfcatons. See Njkamp and Poot (2005) for a comprehensve survey of ths lterature. 2 Prevous research has shown that the early years n a worker s labor market career are of specal mportance (Gardeck and Neumark 1998; Neumark 2002). In terms of wages, Murphy and Welch (1990) estmate that almost 80 % of all (.e., lfetme) wage ncreases accrue wthn the frst 10 years of labor market experence. Moreover, movements across jobs are consderably more lkely at the begnnng of a worker s career than later on (Topel and Ward 1992). 3 Varous dmensons other than wages may be nfluenced by condtons at labor market entry. For example, Gulano and Splmbergo (2009) show that entry condtons have long-lastng effects on ndvduals belefs and preferences, whle Kondo (2012) fnds that the tmng of marrage of both men and women s nfluenced by entry condtons.

3 The mpact of labor market entry condtons on ntal job assgnment and wages 707 employers hre workers enterng the labor market under dfferent condtons, therefore gvng them access to dstnct jobs. The entry job s thus sgnfcant for the future career, and ths appears to be of partcular mportance for hghly educated ndvduals for whom the transton n and out of attractve postons s very low. 4 Mansour (2009) also focuses on college graduates n the USA and agan fnds negatve and persstent wage effects from entry nto the labor force durng a recesson. Moreover, he shows that falure to account for endogenous sample composton underestmates both the mmedate wage effects as well as the persstence of wage effects resultng from ntal labor market shocks. Emprcal evdence for countres outsde the USA and Canada yelds broadly smlar fndngs. Kwon et al. (2010) manly use data from the Swedsh labor market. 5 They fnd that workers who enter the labor market durng a boom are not only pad hgher wages, but that they are also promoted more quckly to hgher ranks than those who enter durng an economc downturn. Stevens (2007) fnds sgnfcant negatve, albet small, effects of ntal condtons on wages n Germany (much smaller than those found n the USA and Canada). In contrast to all other studes, however, she fnds that wage losses from poor entry condtons do not fade away, but actually ncrease over tme. 6 The avalable emprcal evdence also underlnes the fact that negatve wage effects of ntal labor market condtons are not confned to hghly sklled workers. Genda et al. (2010) focus on a separate comparson between men n Japan and the USA wth more or less educaton wth respect to the effects of ntal condtons. They fnd negatve effects of ntal condtons for more hghly sklled workers n both countres. However, they only fnd negatve wage effects for workers n Japan wth fewer sklls. They argue that the specfc hrng system and employment protecton drve the persstence of the effects for Japanese, whle the market for less-sklled workers n the USA may ndeed be qute close to a compettve market. Consstent wth ths fndng, Kondo (2008) reports that the ntal effect of enterng the labor market durng a recesson on wages s less persstent for less-sklled workers and for workers wth weak labor market attachment n the USA. 4 One mportant concern regardng the valdty of these results s that schoolng and frst entry nto the labor force may be endogenous both because ndvduals may choose to stay n school or contnue further tranng when faced wth hgh unemployment and low startng wages. Indeed, several studes fnd that enrollment rates are hgh when unemployment s hgh and the opportunty costs of schoolng are low (Clark 2011). In lne wth these fndngs, both Kahn (2010) and Oreopoulos et al. (2012) fnd the duraton of schoolng to be endogenous. Both tackle the endogenety problem by nstrumentng the unemployment rate at the tme of labor market entry wth ether the prevalng unemployment rate at a lower age or that n the predcted year of graduaton. Mansour (2009) presents drect evdence on sample selecton over the busness cycle based on AFQT scores. 5 Prevous studes for European countres have manly focused on the long-run effects of ntal condtons on employment rather than wages (e.g., Burgess et al. 2003; Raaum and Røed 2006). 6 A smlar analyss of wage effects for frm entry cohorts n the German manufacturng sector s gven by von Wachter and Bender (2008). However, ther analyss s not confned to new labor market entrants, but covers workers of all experence levels; ther results are therefore not drectly comparable to the other studes mentoned.

4 708 B. Brunner, A. Kuhn In ths paper, we present estmates of the long-run effects of busness cycle fluctuatons on young males wage profles n the Austran labor market and derve an emprcal estmate of the assocated loss n lfetme earnngs due to enterng the labor force durng a recesson, as opposed to entry durng average aggregate condtons or durng a boom. We do so usng Austran socal securty records that contan detaled ndvdual earnngs and employment hstores for the unverse of prvatesector employees from 1972 untl We complement the avalable emprcal evdence on the long-run wage effects of labor market entry condtons wth an analyss for Austra, a labor market characterzed by a hgh level of employment protecton and a centralzed wage barganng structure. We focus on low- and medum-sklled workers, whle most of the studes mentoned above focus on hgher or even hghestsklled workers (n terms of formal educaton). 7 In the second part of the analyss, we focus on changes n the qualty of workers frst employer over the busness cycle and the mportance of frst employers for ndvduals entry and subsequent wages. Indeed, several prevous studes have shown that workers ntal placements may have mportant effects not only on ther entry wages but also on ther subsequent wages (e.g., von Wachter and Bender 2006; Oreopoulos et al. 2012). More specfcally, ntal job or task assgnment may be mportant n the longer run f employers assgn otherwse dentcal workers to lower-qualty jobs or tasks n recesson, and f jobs or tasks offer dfferent opportuntes for the accrual of human captal (e.g., Gbbons and Waldman 2006). 8 Alternatvely, workers ntal job or task assgnment may have long-lastng effects on wages f workers accumulate human captal whle on the job that s not fully transferable to other jobs or tasks because t s specfc to a worker s task, occupaton, or ndustry. The remander of ths paper s organzed as follows. The next secton presents our data source, detals the sample selecton process, and dscusses the constructon of our key measures. Secton 3 presents the econometrc approach for estmatng longrun wage effects of ntal labor market condtons along wth our man results. We also present some robustness tests as well as some evdence on the mportance of endogenous labor market entry over the busness cycle. In Secton 4, we focus on the mpact of ntal labor market condtons on the qualty of workers frst employer and the mportance of frst employer on ndvduals subsequent wages. Secton 5 concludes. 7 Note, however, that workers n the Austran labor market typcally have some, potentally very specalzed, vocatonal tranng, as a sgnfcant part of the ntal vocatonal tranng n Austra s provded by dual apprentceshp tranng schemes,.e., practcal tranng provded by frms coupled wth part-tme compulsory attendance at a vocatonal school. Apprentceshps last from 2 to 4 years, dependng on occupaton. Full-tme vocatonal and techncal schools provde an mportant alternatve to apprentceshp tranng and also last up to 4 years. Detals are avalable from the report by the Federal Mnstry for Educaton, the Arts and Culture (2008). 8 Consstent wth ths lne of argument, Kwon et al. (2010) fnd that workers who enter the labor market durng a boom are promoted more quckly and to hgher ranks than those who enter durng a recesson, and Mansour (2009) shows that workers enterng n a recesson are ntally assgned to lowerpayng jobs.

5 The mpact of labor market entry condtons on ntal job assgnment and wages Data and sample 2.1 Data source We use ndvdual-level socal securty records from the Austran Socal Securty Database (ASSD), a data source descrbed n more detal by Zwemüller et al. (2009). The ASSD bascally covers the unverse of Austran prvate sector workers from January 1972 untl December 2005 and contans complete and precse nformaton about ndvduals annual earnngs and daly employment hstores. The data are therefore deally suted for studyng on the mpact of labor market shocks on long-run wage profles because they allow us to construct ndvdual wage profles for a large number of labor market entrants over a relatvely long perod of tme. 2.2 Key measures Our dependent varable for most of the analyss s the real daly wage,.e., the real wage per actual day of work, adjusted to 2007 prces. Wages are deflated wth the consumer prce ndex and nclude addtonal/specal payments such as a 13th month s salary or holday pay. 9 Real daly wages are computed as the average earnngs over all employers n a gven year. Ths means that we frst sum the total annual earnngs over all employers for any ndvdual. We then dvde overall earnngs by the total number of days worked n a gven year, also summed over all employers for a gven ndvdual and takng overlappng employment spells nto account. The regressor of man nterest s the annual male unemployment rate, our measure for external labor market condtons at the tme ndvduals frst enter the labor force. We compute annual male unemployment rates from the ndvdual-level employment hstores contaned n the ASSD raw data. Ths procedure has the advantage that we can calculate unemployment rates back untl 1972 (compared to publshed statstcs, whch only reach back untl 1978) and at dfferent levels of cross-sectonal aggregaton. 10 Unless noted otherwse, we use the male unemployment rate for all workers aged between 16 and 65 at the state level as our man regressor. 9 The Austran Central Socal Securty Admnstraton collects these data wth the purpose of admnsterng and calculatng enttlements to old-age penson benefts. For ths reason, the ASSD ncludes precse and comprehensve nformaton on annual earnngs and daly employment hstores. However, contrbutons to the old-age penson system are capped because old-age penson benefts are lmted to a maxmum level. As a consequence, annual earnngs are only recorded up to the threshold whch guarantees the maxmum beneft level ( Höchstbemessungsgrundlage, HBGr). Smlarly, there s a base threshold below whch no (otherwse mandatory) socal securty payments accrue ( Gerngfüggketsgrenze, GfGr). The two censorng ponts vary over tme n real terms: The lower censorng pont ncreased from about 14 e n 1978 to about 26 e n 2005 (per day worked); the upper censorng pont ncreased from about 78 e to 126 e per workday over the same perod of tme. 10 We decded to extract yearly male unemployment rates for the age groups 16 to 65 and 16 to 25, both at the state ( Bundesland ) level and at the common classfcaton of terrtoral unts for statstcs (NUTS) level. At the NUTS level, we use the most dsaggregated level avalable (NUTS-3), whch corresponds to one or more poltcal dstrcts n Austra. There are total of nne dfferent states and 35 dfferent NUTS-3 regons n Austra. Yearly unemployment rates are wthn-year averages of monthly unemployment rates.

6 710 B. Brunner, A. Kuhn 2.3 Sample selecton Manly for conceptual reasons, but also due to some data lmtatons, we do not work wth the unverse of all labor market entrants but only wth a specfcally selected sample. Frst, we restrct our attenton to male entrants only. On the one hand, female labor supply behavor over the lfe cycle s much more dffcult to model than male labor supply. On the other hand, we beleve that the fact that most men work full tme allows us to largely crcumvent the problem that the ASSD does not contan nformaton on workng hours. Second, we select those workers who start ther frst regular employment spell sometme between 1978 and 2000, allowng us to observe at least fve addtonal years of earnngs for each worker because the data run untl the end of 2005 (see also the Appendx). As a fnal restrcton, we focus on workers aged between 16 and 21 at the tme they frst enter the labor force (.e., start ther frst regular employment spell). 11 Essentally, ths restrcton excludes ndvduals wth hgher educaton (most mportantly, ndvduals wth a unversty degree), but t should nclude all or most ndvduals wth an apprentceshp tranng or an educaton of smlar length and scope, such as full-tme vocatonal school. 12 Our fnal sample thus conssts of male low- and medum-sklled labor market entrants who started ther frst regular employment between 1978 and We can observe these workers full labor market career from the year they frst enter nto the labor force untl the year Sample descrpton Because we can follow all ndvduals from the year of ther frst regular employment untl the end of the data n the year 2005, the resultng data set would have been too large from a practcal pont of vew. In the followng, we therefore work wth a 30 % random sample of all labor market entrants aged between 16 and 21 when frst enterng nto the labor force. Ths sample contans 217,587 unque ndvduals and about 3.35 mllon ndvdual wage observatons (.e., observatons at the level of ndvdual year). Table 1 shows descrptve statstcs for our fnal analyss sample. The frst panel shows ndvdual-level characterstcs. The average labor market entrant n our 11 Because the ASSD does not contan a comprehensve measure for schoolng, we use age at entry nto the labor force as proxy for educaton n the regressons below. To mtgate potental collnearty wth year of brth and year of entry, we use a slghtly dfferent varable as proxy n the regressons. Specfcally, we use the smaller of age at start of frst regular employment and age at start of frst regstered unemployment spell. 12 Several arguments motvate the restrcton on schoolng. Frst, the tmng of frst labor market entry, and thus the duraton of schoolng, may be endogenous. However, less-sklled workers are presumably less lkely to manpulate the duraton of schoolng. Furthermore, unobserved heterogenety resultng from, say, unobserved dfferences n nherent ablty,s arguably a more urgent problem for workers wth hgher sklls. Moreover, we beleve that our proxy for schoolng works best for workers wth low educaton levels. Fnally, only ncludng less-sklled workers n the sample s an effectve way of dealng wth rght-censored wages (see also the Appendx).

7 The mpact of labor market entry condtons on ntal job assgnment and wages 711 Table 1 Summary statstcs, male labor market entrants Mean Standard devaton Indvdual-level characterstcs Age at start of frst regular job (1.007) Age at frst entry nto the labor force (1.431) Duraton of frst regular job (years) (4.335) Any unemployment before frst job (0.461) Unemployment days before frst job ( ) Blue collar (0.498) Whte collar (0.380) Aggregate-level varables State-level unemployment rate (2.897) NUTS-3 level unemployment rate (3.144) State-level youth unemployment rate (1.790) Number of entrants aged , (1, ) State of frst entry nto the labor market Venna (0.383) Lower Austra (0.378) Burgenland (0.156) Upper Austra (0.395) Styra (0.354) Carntha (0.251) Salzburg (0.256) Tyrol (0.289) Vorarlberg (0.223) Year of frst entry nto the labor market (0.178) (0.194) (0.208) (0.207) (0.202) (0.206) (0.213) (0.212) (0.210) (0.210) (0.214) (0.217) (0.216) (0.209) (0.201) (0.190)

8 712 B. Brunner, A. Kuhn Table 1 (contnued) Mean Standard devaton (0.198) (0.196) (0.192) (0.195) (0.197) (0.204) (0.214) Number of unque ndvduals 217,587 All aggregate-level varables are computed from the ndvdual-level raw data of the ASSD (see footnote 10). See also notes of Table 10 sample s about 19 years old when startng hs frst regular employment spell, and he holds hs frst job for almost 3 years. The average age at the start of the frst job dovetals wth the fact that mandatory schoolng ends n the year when ndvduals attan the age of 16 and that apprentceshps usually last for 2 to 4 years. The hgh fracton of blue-collar occupatons s consstent wth our ntenton of only (or manly) ncludng ndvduals who receved some knd of vocatonal tranng. Interestngly, a substantal fracton of the sample (about a thrd) experences some unemployment before startng the frst regular employment spell (these ndvduals are regstered for unemployment benefts on average for somewhat more than 1 month). Consstent wth ths, we fnd that age at frst entry (our proxy for schoolng) s about half a year lower than age at the start of the frst regular job, reflectng the fact that the transton from educaton to work often nvolves short perods of nonemployment. The second panel shows varous unemployment rates as well as the number of labor market entrants. The dfferent unemployment rates use dfferent aggregaton levels (cf. footnote 10) and refer to dfferent age groups, but all of them are lmted to the male populaton only. The unemployment rate n the year of labor market entry averages about 6.6 %, rrespectve of the chosen aggregaton level (states or NUTS- 3 unts). Youth unemployment rates at the tme of entry are somewhat lower than overall unemployment rates and equal about 5 % on average. Fnally, about 3,750 male ndvduals aged between 16 and 21 enter the labor market n any gven state and year n the perod The remanng part of the table shows the dstrbuton of our sample of labor market entrants across the nne dfferent states and across years. The dstrbuton across states manly reflects dfferences n populaton sze. It may, however, also reflect other dfferences between states (e.g., dfferences n the age dstrbuton of women). The dstrbuton of entrants across years on the other hand looks farly unform, mplyng that the aggregate number of entrants aged between 16 and 21 has not changed much over tme.

9 The mpact of labor market entry condtons on ntal job assgnment and wages The persstence of ntal labor market shocks 3.1 Graphcal evdence on the evoluton of wages and ntal condtons We start wth a smple graphcal depcton of our two key measures (.e., cohorts wage profles and the ntal local unemployment rate). Frst, Fg. 1 shows wage profles by entry cohort for all labor market cohorts who frst entered the labor force between 1978 and The black dots therefore represent average startng wages for each entry cohort, and the dashed gray lne thus shows how startng wages evolve over tme. Clearly, real startng wages have ncreased sgnfcantly over the perod of analyss, from about 38 e n 1978 to about 50 e n Also note that there s some cyclcal movement n startng wages over tme whch we expect to be related to economc condtons prevalng n that year. The sold colored lnes, on the other hand, represent long-run wage profles of cohorts enterng the labor market n dfferent years. Cohorts wages clearly follow an approxmate concave path over tme, mplyng that wage growth s hghest n earler workng years and then strongly flattens later on. The fgure shows, for example, that the 1978 entry cohort starts wth a real wage of about 38 e per workday and experences a rase n real wage up to about 97 e by the year On average, ths cohort s compensaton therefore more than doubles n real terms wthn the frst 27 years of labor market experence. Focusng agan on the 1978 entry cohort, we see that ths cohort s average wage grows by approxmately 146 % (= [exp(0.9) 1] 100 %) Log real daly wage Year Fg. 1 Long-run wage profles, by labor market entry cohort. The fgure shows average log real daly wages by calendar year for each labor market entry cohort from 1978 to The dots thus show average log startngwages for eachlabor market entry cohort, and the dotted lne hghlghts the evoluton of startng wages. The flled lnes show how entry cohorts wages evolve wth ncreasng labor market experence

10 714 B. Brunner, A. Kuhn n the frst 27 years of experence. Evdently, most of ths wage ncrease happens at the early stage of the labor market career (.e., the wage ncrease n the frst 10 years equals about 86 % (= [exp(0.62) 1] 100 %)). 13 Panel (a) of Fg. 2 shows the evoluton of state-level unemployment rates whch we use as our ndcator for external labor market condtons at the tme ndvduals frst enter the labor force. Ths fgure shows that the perod from 1978 to 2000 covers several perods of both boom and downturn, and that the dentfyng varaton n ntal labor market condtons therefore does not only stem from a few neghborng labor market cohorts. The fgure also llustrates that states not only dffer n the level of unemployment but also wth respect to cyclcal varatons around a longer-run trend: although all states see an ncrease over the whole observaton perod n general, there are marked cyclcal dfferences across states. 14 The lower panel of Fg. 2 shows that our observaton perod spans several ups and downs n the busness cycle, and that there s consderable dfferentaton n the strength of these varatons across states. We thus have both suffcent cross-sectonal and longtudnal varaton n our key regressor that we can use to pn down the effect of local busness cycle fluctuatons on wages. 3.2 Econometrc framework Because we prmarly am to estmate the long-run mpact of economc shocks at the tme ndvduals started ther frst jobs, we must take care to allow the assocaton between ntal condtons and wage to become weaker or stronger as labor market experence ncreases whle also usng a generally flexble functonal form of wage profles. Takng these ssues nto account, our basc and most parsmonous econometrc model s the followng: ] ln(y t ) = ur 0 j[] α 1 +κ(exp t )α 2 + [ur 0 j[] κ(exp t ) α 3 +ψ j(t 0) +φ t 0 +θ t +ɛ t, (1) where y t denotes the real daly wage of ndvdual n calendar year t, exp t the potental labor market experence of n year t, and ur 0 j[] the unemployment rate as prevalng n state j at the tme ndvdual frst entered the labor market. Functon κ( ) denotes that we allow for a flexble functonal form wth respect to labor market 13 Wage profles of dfferent entry cohorts have somewhat dstnct overall shapes. More specfcally, the fgure shows that returns to experence generally decrease over tme, meanng that younger entry cohorts have consderably lower returns to labor market experence than older cohorts. For example, the 1995 cohort only realzes an average wage ncrease of about 58 % (= [exp(.46) 1] 100 %)n the frst 10 years, thus less smaller than that of the correspondng ncrease of the 1978 entry cohort. 14 For example, and as hghlghted n the fgure, Burgenland (located n southeastern Austra) experenced a huge ncrease n the unemployment rate from about 3 % n the late 1970s to about 8 % n the frst half of the 1980s, and then to about 9 % n the second half of the 1980s. Vorarlberg (stuated n western Austra), n contrast, experenced only a modest ncrease from about 1 % on the 1970s to about 3% n the 1980s. In 1992, however, Vorarlberg underwent a sharp declne n the local labor market condtons, when unemployment jumped from about 3 % to about 7 8 %.

11 The mpact of labor market entry condtons on ntal job assgnment and wages Year Burgenland Vorarlberg (a) Unemployment rate (n %) Year Burgenland Vorarlberg (b) Resdual unemployment rate (n %) Fg. 2 a Fluctuatons n state-level unemployment rates, The fgure shows male unemployment rates for workers aged between 16 and 65, aggregated at the level of the state ( Bundesland ) and computed from the ndvdual-level raw data of the ASSD. Two of the nne states are graphcally hghlghted. Vorarlberg s stuated n the western part of Austra (borderng Germany and Swtzerland). Burgenland s located at the eastern border of Austra (borderng Hungary n the east and Slovena n the south). b Resduals from a regresson of the ntal unemployment rate on a full set of year and state of entry dummes

12 716 B. Brunner, A. Kuhn experence. 15 Note that we allow the effect of the ntal unemployment rate on current wages to vary as potental labor market experence ncreases by ncludng the nteracton terms between the experence polynomals and the ntal unemployment rate. We also nclude a full set of dummes for state at entry and year of entry, denoted by ψ j(t 0) and φ t 0, respectvely, throughout the analyss. Fnally, we also control for the aggregate busness cycle θ t whch we parameterze usng the aggregate yearly unemployment rate, the log aggregate number of workers, as well as log aggregate annual earnngs. As we wll dscuss n more detal below, our baselne (.e., our preferred) specfcaton wll nclude some addtonal regressors: ] ln(y t ) = ur 0 j[] α 1 + κ(exp t )α 2 + [ur 0 j[] κ(exp t ) α 3 + ψ j(t 0) + φ t 0 + θ t [ ] +x 0 β 1 + φ t 0 κ(exp t ) β 2 + ln(n 0 j[] )β 3 + ɛ t, (2) where the second row contans these addtonal varables. Frst, x 0 denotes a small set of ndvdual-level controls,.e., our proxy for schoolng and two ndcator varables for workers ntal occupaton (blue or whte collar). These varables are predetermned n the sense that they relate to an ndvdual s frst regular employment spell or to the tme before havng started to work (.e., there s no tme ndex for these varables). We also nclude the full set of nteracton terms between the year of entry and the polynomal n potental experence as well as the log number of labor market entrants aged and aged at the state level. The latter two varables are ncluded to control for changes n the relatve number of entrants aged (relatve to all entrants aged n the age bracket 16 30); the former set of controls allows for cohort-specfc dfferences n the wage profle that are unrelated to dfferences n local ntal unemployment rates but due to, say, changes n the producton of educaton across cohorts. In ether case, parameters α 1 to α 3 descrbe ndvduals wage experence profles as a functon of the ntal unemployment rate and are the parameters of man nterest. Specfcally, α 1 s the elastcty of wages wth respect to the ntal unemployment rate n the year of frst entry (.e., n the year where labor market experence s equal to 0), whle α 3 tells us how the effect of ntal condtons changes as labor market experence ncreases. One mportant complcaton of both specfcatons relates to the fact that the local ntal unemployment rate does not vary over tme for any gven ndvdual. For ths reason, we cannot use standard panel data estmators such as the fxed-effects or frst-dfferences estmator because these methods not only elmnate all unobserved tme-nvarant heterogenety but also all varaton n the key regressor. We therefore rely on estmaton methods that use the untransformed data. We also have to consder that our key regressor s observed at a hgher level of aggregaton than the dependent 15 Specfcally, we nclude the frst three polynomal terms of potental labor market experence. We chose the number of polynomal terms on the bass of a nonparametrc, and therefore fully flexble, wage experence model. The frst three polynomal terms appear suffcent for reproducng the wage experence profle that a correspondng nonparametrc specfcaton predcts.

13 The mpact of labor market entry condtons on ntal job assgnment and wages 717 varable, a stuaton that may lead to grossly msleadng statstcal nference (Moulton 1986). All standard errors we report are therefore clustered by cells defned by year at frst entry state of frst entry (there are 9 states and 23 entry years, resultng n 207 dstnct cells) Man results: ntal labor market condtons and wages Table 2 shows our man results, buldng up from our most parsmonous specfcaton represented by Eq. 1 to our preferred specfcaton as gven by Eq. 2. The top panel of the table shows, for each specfcaton, pont estmates of the parameters descrbng the effect of ntal condtons on wages (.e., α 1, the coeffcent on the local ntal unemployment rate, and α 3, whch represents the three nteracton terms between the local ntal unemployment rate and years of labor market experence, ts square, and ts cube, respectvely), and the panel n the mddle of the table shows estmated semelastctes of wages wth respect to the ntal local unemployment rate at specfc values of potental labor market experence (.e., potental labor market experence of 0, 5, 10, 15, and 20 years, respectvely). For example, ε y ur(5) denotes the estmated sem-elastcty of the real daly wage wth respect to the ntal unemployment rate at 5 years of potental labor market experence. It thus corresponds to the estmated relatve change n wages resultng from a one percentage pont ncrease n the ntal unemployment rate. The frst column shows pont estmates for our most parsmonous specfcaton as spelled out n Eq. 1. As expected, there s a negatve effect of the ntal local unemployment on wages n the year of entry. Specfcally, the sem-elastcty of wages wth respect to the ntal unemployment rate equals n the year of entry. A one percentage pont ncrease n the local ntal unemployment rate s thus predcted to lower startng wages by 0.9 %. The correspondng standard error equals 0.002, and thus the mmedate wage effect s statstcally hghly sgnfcant. Moreover, the mddle panel of the table shows that there s substantal persstence of ths negatve wage effect, and a negatve and sgnfcant effect of ntal labor market condtons remans as much as 20 years after the frst entry nto the labor market. In fact, the pont estmate of the sem-elastcty at 20 years of experence s even larger than the mmedate wage penalty. We add our small set of ndvdual-level controls (.e., our proxy for schoolng and two dummy varables ndcatng a worker s ntal occupaton) n the second specfcaton. Ths has some mpact on the estmated wage effects, especally n the year of entry nto the labor market (the mmedate sem-elastcty of wages decreases from to 0.007), but otherwse, the coeffcents are remarkably smlar, and a negatve and strongly persstent wage effect of bad entry condtons remans, as the mddle panel of Table 2 reflects. We replace our set of longtudnal controls wth a set of bannual year dummes n the thrd column. Ths yelds estmates that are very smlar to those from the 16 We also computed standard errors that smultaneously account for clusterng at both levels for our man estmates. Ths yelds standard errors that are vrtually ndstngushable from those actually reported.

14 718 B. Brunner, A. Kuhn Table 2 The long-run wage effects of ntal labor market condtons Dependent varable ln(real daly wage) Mean Standard devaton ur (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) exp ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) exp 2 ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) exp 3 ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) εur(0) y (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) εur(5) y (0.002) (0.002) (0.002) (0.002) (0.001) (0.001) εur(10) y (0.002) (0.002) (0.002) (0.001) (0.001) (0.001) εur(15) y (0.002) (0.002) (0.002) (0.002) (0.002) (0.001) εur(20) y (0.003) (0.002) (0.003) (0.003) (0.002) (0.002) Potental experence Yes Yes Yes Yes Yes Yes Longtudnal controls Yes Yes No Yes Yes Yes Year of entry Yes Yes Yes Yes Yes Yes Regon of entry Yes Yes Yes Yes Yes Yes Indvdual controls No Yes Yes Yes Yes Yes Tme dummes (bannual) No No Yes Yes No No Year of entry κ(exp) No No No No Yes Yes Number of entrants No No No No No Yes Number of observatons 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 Adjusted R-squared Lfetme loss Robust standard errors are gven n parentheses and are clustered by state at entry year of entry. exp and ur 0 denote potental labor market experence (n years) and the ntal unemployment rate (n percentages), respectvely. ε y ur(k) denotes the estmated sem-elastctyof wages wth respect to the ntal unemployment rate, evaluated at k years of potental labor market experence p<0.1; p<0.05; p<0.01 (statstcal sgnfcance)

15 The mpact of labor market entry condtons on ntal job assgnment and wages 719 specfcaton wth parametrc busness cycle controls. The fourth column ncludes both the set of bannual year dummes and our set of parametrc longtudnal controls. Agan, the estmates hardlychange, althoughthe estmated sem-elastctes at hgher values of labor market experence are somewhat smaller than n the specfcaton wth only a parameterzed trend. We next add the full set of nteracton terms between the year of entry and the experence terms. Ths allows for changes n the wage profle across entry cohorts that are unrelated to ntal condtons n a fully flexble way. The estmates agan reman more or less stable. We fnally add the log number of labor market entrants aged and n the sxth column because there was a substantve drop n the relatve number of labor market entrants aged relatve to all entrants aged below 30. The specfcaton underlyng the estmates shown n column 6 thus corresponds to the specfcaton of Eq. 2. Our baselne specfcaton stll yelds an mmedate negatve and statstcally sgnfcant wage effect of about Moreover, we fnd the wage penalty resultng from enterng the labor market n tmes of hgh local unemployment to be hghly persstent. In fact, the wage penalty s even slghtly ncreasng wth labor market experence, smlar to what s found by Stevens (2007). 3.4 Robustness Table 3 presents some dfferent robustness checks. For the ease of comparson, the frst column smply reproduces our baselne estmates from column 6 of Table 2. A frst robustness check relates to the fact that a nonneglgble number of wage observatons s censored. 17 Column 2 thus shows quantle (.e., medan) regresson estmates of our baselne specfcaton. The short-run estmates are dentcal, but the estmated wage losses turn out to be moderately larger at hgher values of labor market experence. More specfcally, εur 0 (5) s already slghtly larger when estmated usng a quantle regresson, but the quantle regresson estmate of εur 0 (20) s as much as 35 % larger than the correspondng ordnary least squares (OLS) estmate. If anythng, censored wages wll thus lead us to underestmate the wage effects resultng from labor market entry n tmes of hgh local unemployment. We add the nteracton terms between our longtudnal controls and the experence terms n the thrd column to allow for the possblty that varatons n the busness cycle affect workers dfferentally at dfferent stages of ther labor market career. Ths, however, has no dscernble effect on the estmated coeffcents of our key parameters. Most of the estmated sem-elastctes are vrtually dentcal to our baselne estmates. We use dfferent unemployment rates n columns 4 and 5 as another robustness check. We use unemployment rates at the NUTS-3 rather than at the state level n the fourth column (cf. footnote 10). Usng the unemployment rate at the NUTS-3 level results n a smaller ntal and a somewhat less persstent wage effect. Ths 17 Appendx Table 10 shows that only few wage observatons are censored n the year of entry. However, top-censored wages become much more frequent as workers accumulate labor market experence.

16 720 B. Brunner, A. Kuhn Table 3 Senstvty analyss Dependent varable ln(real daly wage) ln(earnngs) Mean Standard devaton ur (0.002) (0.000) (0.002) (0.001) (0.002) (0.003) exp ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.001) exp 2 ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) exp 3 ur (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) εur(0) y (0.002) (0.000) (0.002) (0.001) (0.002) (0.003) εur(5) y (0.001) (0.000) (0.001) (0.001) (0.002) (0.002) εur(10) y (0.001) (0.000) (0.001) (0.001) (0.002) (0.002) εur(15) y (0.001) (0.000) (0.001) (0.001) (0.002) (0.002) εur(20) y (0.002) (0.000) (0.002) (0.001) (0.002) (0.003) Estmaton method OLS LAD OLS OLS OLS OLS Unemployment rate State State State NUTS-3 Youth State Baselne controls Yes Yes Yes Yes Yes Yes Longtudnal controls κ(exp) No No Yes No No No Number of observatons 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 Adjusted R-squared Lfetme loss Robust standard errors are gven n parentheses and are clustered by state at entry year of entry. The specfcaton n column 1 s the same as n column 6 of Table 2. All specfcatons nclude the same set of controls as n column 6 of Table 2 ( baselne controls ). The dependent varable s the log real daly wage n columns 1 5 and log real annual earnngs n column 6. We use the state unemployment rate for men aged ( state ) n columns 1 3 and 6, the NUTS-3 level unemployment rate for men aged ( NUTS-3 ) n column 4, and the state-level unemployment rate for men aged ( youth ) n column 5 p<0.1; p<0.05; p<0.01 (statstcal sgnfcance) s most lkely due to the fact that we also nclude entry-regon fxed effects at a fner aggregaton level whch tends to pck up more of the varaton n wages.

17 The mpact of labor market entry condtons on ntal job assgnment and wages The specfcaton n the ffth column uses state-level youth unemployment rates nstead of overall unemployment rates. The resultng wage effects are somewhat larger but stll reasonably close to our man estmates based on overall unemployment rates. Fnally, the specfcaton n the sxth column shows results when we use log annual earnngs nstead of log daly wages as the dependent varable. In ths case, we get somewhat larger, but stll very smlar estmates to those n the baselne case. The comparson wth the effect on wages also mples that workers enterng durng tmes of hgh local unemployment not only suffer from lower subsequent wages, but also from fewer employment days throughout ther labor market career (see also our Electronc Supplementary Materal (ESM) where we dscuss the effect of ntal condtons on job moblty and unemployment). 3.5 Endogenous labor market entry Instrumental varable estmates One mportant ssue that we have not yet consdered s the fact that the composton of the sample of labor market entrants may be endogenous wth respect to varaton n the local busness cycle, whch may lead to nconsstent estmates of the effect of ntal condtons on wages. 19 Varous studes have tred to tackle ths ssue by nstrumentng the unemployment rate at labor market entry wth the unemployment rate at some earler age or wth the unemployment rate at the predcted date of graduaton. 20 We have thus tred a smlar approach, nstrumentng the local unemployment rate at entry wth the local unemployment rate at age 16, the age at whch mandatory schoolng ends n Austra. However, we have some serous reservatons, at least n our context, regardng the valdty of ths nstrumental varable strategy. Most mportantly, the nstrument may have a drect effect on subsequent wages, n whch case IV estmates wll be based (e.g., Angrst and Krueger 1994). To see why ths may occur, note that the majorty 18 It may also reflect endogenety of the local unemployment rate at lower aggregaton levels, due to the fact that workers move from regons wth hgh unemployment to those wth lower levels of unemployment (Woznak 2010). 19 Bls (1985) and more recently Solon et al. (1994) and Blundell et al. (2003) have put forth ths lne of argument. In fact, the tmng of labor force entry and thus the composton of labor market cohorts may be endogenous for several dstnct reasons. Frst, some potental labor market entrants may refran from enterng the labor market altogether. Second, both the choce of educaton as well as the duraton of schoolng may be endogenous, as both job prospects are weak and opportunty costs of schoolng are low n tmes of hgh unemployment. Thrd, some workers may smply delay ther entry when faced wth unfavorable entry condtons, ether by regsterng for unemployment benefts or by stayng out of the labor force untl they fnd a job. Whatever the underlyng reason, f those workers who do not mmedately get a job are a selected group of all workers who ntend to enter employment n a gven year, then the composton of the actual entrants changes along wth correspondng changes n the unemployment rate and thus potentally bases the estmated effect of the ntal unemployment rate on wages. 20 Kahn (2010), Kondo (2008) and Oreopoulos et al. (2012) use a smlar nstrumental varable strategy. OLS and IV estmates are smlar to those of Oreopoulos et al. (2012), but IV are substantally larger to those of both Kondo (2008) and Kahn (2010). Two European studes, by Kwon et al. (2010) and Stevens (2007), only show OLS estmates.

18 722 B. Brunner, A. Kuhn of ndvduals, at least n our sample, start an apprentceshp tranng at the age of 16, as ths s the age when mandatory schoolng ends. In Austra, moreover, most of the practcal tranng s drectly provded by frms whch are free to decde whether they want to offer apprentceshps. It s thus qute obvous that labor market condtons may have an mpact on frms decson to offer apprentceshps. 21 Moreover, economc condtons may not only mpact the number, but also the type (.e., qualty) of apprentceshp frms offer. Table 4 shows OLS and nstrumental varable estmates. As before, the frst column replcates our baselne estmates from column 6 of Table 2. We frst compare our baselne estmates to the correspondng 2SLS estmates that nstrument the local unemployment rate at frst entry wth the local unemployment rate at age 16, shown n the ffth column. The 2SLS estmate of the sem-elastcty n the year of entry nto the labor force s more than sx tmes larger than the correspondng OLS estmate ( versus 0.009), and the lfetme loss n the case of 2SLS s about four tmes larger than n the baselne case (0.052 versus 0.013). Whle t appears reasonable that 2SLS estmates are larger than OLS estmates, sgnfyng postve sample selecton n economc downturns, the sze of the dfference between the estmates s almost suspcously large. Thus, let us examne the reduced form estmates as well, shown n column 3. Note that the mmedate and medum-run wage losses are consderably larger than those n column 1, whle the estmates appear to converge at hgher values of labor market experence. It seems hard to ratonalze the dfference between columns 1 and 3 wthout consderng the possblty of a drect effect of ur 16 on the dependent varable. Fnally, snce both our proxy for the completed duraton of schoolng and/or tranng as well as ndvduals ntal occupaton may be endogenous wth respect to the unemployment rate age 16, we also show estmates that do not control for these characterstcs. Our baselne estmates arehardlyany dfferent,butnote thatthe estmates usng the unemployment rate at age 16 ncrease consderably, resultng n 2SLS estmates that are as much as ten tmes larger than the correspondng OLS estmates (compare columns 6 and 2). These estmates are presumably too large to be plausble, but we thnk that they can be easly ratonalzed by acknowledgng that the nstrument has a drect mpact on wages (through ts mpact on ndvduals schoolng/tranng) n whch case 2SLS estmates are upwards based. Drect evdence on endogenous labor market entry We complement the evdence from our IV estmates wth some more drect evdence on compostonal effects from labor market entry condtons. A frst pece of evdence s provded n Table 5 that drectly examnes the assocaton of entry condtons 21 To the best of our knowledge, there s no relevant emprcal evdence for Austra. However, Muehlemann et al. (2009) provde evdence consstent wth ths lne of argument for Swtzerland, whch has an apprentceshp system very smlar to that of Austra. Ther results thus presumably carry over to Austra.

19 The mpact of labor market entry condtons on ntal job assgnment and wages 723 Table 4 OLS versus 2SLS estmates Dependent varable ln(real daly wage) Mean Standard devaton ur (0.002) (0.002) (0.006) (0.017) exp ur (0.000) (0.000) (0.001) (0.001) exp 2 ur (0.000) (0.000) (0.000) (0.000) exp 3 ur (0.000) (0.000) (0.000) (0.000) ur (0.002) (0.003) exp ur (0.001) (0.001) exp 2 ur (0.000) (0.000) exp 3 ur (0.000) (0.000) εur(0) y (0.002) (0.002) (0.002) (0.003) (0.006) (0.017) εur(5) y (0.001) (0.002) (0.002) (0.002) (0.005) (0.016) εur(10) y (0.001) (0.001) (0.002) (0.002) (0.005) (0.016) εur(15) y (0.001) (0.002) (0.002) (0.002) (0.005) (0.017) εur(20) y (0.002) (0.002) (0.002) (0.003) (0.006) (0.017) Estmaton method OLS OLS OLS OLS 2SLS 2SLS Baselne controls Yes Yes Yes Yes Yes Yes Indvdual controls Yes No Yes No Yes No Number of observatons 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 3,349,075 Adjusted R-squared Lfetme loss Robust standard errors are gven n parentheses and are clustered by state at entry year of entry. The frst column replcates the baselne estmates from column 6 of Table 2. All specfcatons nclude controls for potental labor market experence, year of entry, state at entry, longtudnal controls, the nteractons between year of entry and κ(exp), as well as controls for the number of labor market entrants ( baselne controls ). 2SLS estmates nstrument the local unemployment rate at entry (ur 0 ) wth the local unemployment rate at age 16 (ur 16 ) p<0.1; p<0.05; p<0.01 (statstcal sgnfcance)