Ruhr Economic Papers RWI ESSEN. Torge Middendorf. Detailed Results from a Harmonized Survey #65

Transcription

1 Torge Mddendorf Detaled Results from a Harmonzed Survey #65 Ruhr Economc Papers RWI ESSEN

2 Ruhr Economc Papers Publshed by Ruhr-Unverstät Bochum (RUB), Department of Economcs Unverstätsstraße 150, Bochum, Germany Technsche Unverstät Dortmund, Department of Economc and Socal Scences Vogelpothsweg 87, Dortmund, Germany Unverstät Dusburg-Essen, Department of Economcs Unverstätsstraße 12, Essen, Germany Rhensch-Westfälsches Insttut für Wrtschaftsforschung (RWI Essen) Hohenzollernstrasse 1/3, Essen, Germany Edtors: Prof. Dr. Thomas K. Bauer RUB, Department of Economcs Emprcal Economcs Phone: +49 (0) 234/ , e-mal: Prof. Dr. Wolfgang Lennger Technsche Unverstät Dortmund, Department of Economc and Socal Scences Economcs Mcroeconomcs Phone: +49 (0) 231 / , emal: Prof. Dr. Volker Clausen Unversty of Dusburg-Essen, Department of Economcs Internatonal Economcs Phone: +49 (0) 201/ , e-mal: Prof. Dr. Chrstoph M. Schmdt RWI Essen Phone: +49 (0) 201/ , e-mal: Edtoral Offce: Joachm Schmdt RWI Essen, Phone: +49 (0) 201/ , e-mal: Ruhr Economc Papers #65 Responsble Edtor: Chrstoph M. Schmdt All rghts reserved. Bochum, Dortmund, Dusburg, Essen, Germany, 2008 ISSN (onlne) ISBN The workng papers publshed n the Seres consttute work n progress crculated to stmulate dscusson and crtcal comments. Vews expressed represent exclusvely the authors own opnons and do not necessarly reflect those of the edtors.

3 Ruhr Economc Papers #65 Torge Mddendorf RWI ESSEN

4 Bblografsche Informaton der Deutschen Natonalbblothek De Deutsche Natonalbblothek verzechnet dese Publkaton n der Deutschen Natonalbblografe; detallerte bblografsche Daten snd m Internet über abrufbar. ISSN (onlne) ISBN

5 Torge Mddendorf* Returns to Educaton n Europe Detaled Results from a Harmonzed Survey Abstract We use the European Communty Household Panel, a harmonzed data set coverng the countres of the European Unon, to provde detaled estmates of the returns to educaton. Our results can be summarzed as follows. Frstly, average returns to educaton have been mostly stable durng the second half of the 1990s and are hghest n Portugal and Ireland and lowest n the UK and Italy. Secondly, returns to schoolng are sgnfcantly negatvely related to the educatonal attanment of the populaton. Thrdly, for most countres we fnd sgnfcant cohort effects and these are n general unform across countres mplyng lower returns to educaton for younger cohorts. Fourthly, n most countres schoolng exerts a sgnfcantly stronger mpact on wages at the top of the wage dstrbuton, aggravatng wthn-group nequalty. Fnally, we provde evdence that the more pronounced the dfference n returns to educaton along the wage dstrbuton, the hgher the average return to educaton. JEL Classfcaton: I21, J24, J31 Keywords: Returns to schoolng, cohort effects, quantle regresson August 2008 * RWI Essen. The author thanks Thomas Bauer, Mchael Fertg, Katja Görltz, John Hasken- DeNew, Wm Kösters, Chrstoph M. Schmdt and Mathas Snnng for helpful comments and suggestons. All correspondence to Torge Mddendorf, Rhensch-Westfälsches Insttut für Wrtschaftsforschung (RWI Essen), Hohenzollernstr. 1-3, Essen, Germany, e-mal: rw@rw-essen.de.

6 1 Introducton In the last couple of years t has been recognzed that human captal may be one of the key drvers of economc growth (Mddendorf 2006; OECD 2003). The channels through whch t mght work are manfold but hgher educaton s also lkely to operate ndrectly for example through better health (Berger and Legh 1989; Kenkel 1991; Groot and van den Brnk 2004) or lower crme (Lochner and Morett 2004; Buonanno and Leonda 2006). Thus, educaton yelds socal benefts and polcy-makers should strve for a hgher educatonal attanment of the populaton. Today, however, the major European countres are losng ground n provdng a better educated workforce. In Germany, the share of the populaton who has attaned tertary educaton has declned for the 25 to 34-year cohort whle Italy dd not accomplsh to ncrease t and s now below the level of Chle (OECD 2006: 32). 1 Thus, the goal of the Lsbon agenda, to make the EU the most compettve and dynamc knowledge-drven economy by the year 2010 seems far out of reach, whch has already been conceded by the European Commsson (European Commsson 2004). For the ndvdual, the decson to nvest n human captal, lke other nvestment decsons, hnges on the nternal rate of return of that nvestment. If these returns turn out to be dsproportonately hgh relatve to other nvestments, ths may be a sgn of a market falure (Harmon, Oosterbeck and Walker 2003) and thus, besdes the externaltes outlned above, another reason for nterventon. Whle there are numerous studes dealng wth estmates of the economc return to educaton n European countres, cross-natonal comparsons are flawed as they are based on dfferent datasets, dfferent estmaton procedures, or both. Ths study therefore uses a unque standardzed European survey, the ECHP, to provde detaled estmates of the returns to educaton n Europe. Frstly, we estmate the standard Mnceran wage equaton and delver comparable estmates for the EU-member countres over tme. We then explore f cross-natonal dfferences n the levels of returns can be attrbuted to the supply of educated workers. 1 Fgures are for

7 Secondly, we analyze f the returns to educaton have been stable across brth cohorts. Ths research queston s especally relevant n a European context for two reasons. Frst of all, Europe has experenced a large declne n lve brths. Snce the md 1960s these dropped by roughly 35 % untl Furthermore, educaton polcy durng the 1960s and 1970s provded for a huge upswng n educatonal attanment. If the demand for sklls has however not kept pace wth ths ncrease, t s lkely that educaton prema have fallen for younger cohorts. Despte these strong reasons for changes n the returns to educaton, so far there s only scarce emprcal evdence for European countres (see Boockmann and Stener (2006) for an analyss for West-Germany). Fnally, we use quantle regresson technques to assess the relatonshp between educaton and wage nequalty. If returns to schoolng are dfferent along the wage dstrbuton, schoolng would exert an mpact upon wage nequalty. Prevous estmaton results from Martns and Perera (2004) from the md 1990s for most of the countres ncluded n our own analyss pont at returns beng hgher at the upper part of the wage dstrbuton. Thus, contrary to common belefs, schoolng would have a postve mpact upon wthn-levels wage nequalty. Thus far, ths queston has not been addressed wth comparable data, though. The remander of the paper s organzed as follows. Secton 2 provdes an overvew of recent estmates of the economc returns to schoolng n Europe. Secton 3 explans the estmaton approach as well as the utlzed data. In secton 4 the estmaton results are presented and secton 5 concludes. 2 Recent evdence on returns to educaton n Europe The project Publc Fundng and Prvate Returns to Educaton (PURE), funded by the European Commsson, nvolved researchers from 15 European countres and amed at provdng comparable estmates of the prvate returns to educaton (for the followng results see Harmon, Walker and Westergaard-Nelsen 2001). As all of the researchers analyzed ther respectve natonal datasets, the results of ther study are per se not comparable across countres. Yet, the project partners agreed on estmatng a common specfcaton of the wageequaton, whch ncreases comparablty. 5

8 Accordng to ther results, Ireland, the UK and Portugal feature the hghest returns to educaton n 1995, whereas Norway, Denmark and Sweden are characterzed by the lowest returns. Furthermore, for the UK, Ireland, Germany, Greece and Italy they observe a substantal dfference n returns between genders as females receve a sgnfcantly hgher educaton premum than men. As the respectve datasets are longtudnal n nature, the PURE project had also the opportunty to provde estmates of the economc return to schoolng over tme. Across countres, however, they do not dentfy a common trend n returns. For men, they dentfy a fallng trend n returns from the md 1980s untl the md 1990s n Austra and Portugal, whereas Denmark, Portugal, the UK and Italy seem to exhbt an ncreasng trend. Martns and Perera (2004) rely on these estmates and address the relaton between schoolng and wage-nequalty. Utlzng quantle regresson technques, they show that for nearly all EU countres, returns to schoolng are sgnfcantly hgher at the top of the wage dstrbuton. Indvduals, who are at the top of the condtonal wage dstrbuton, are there because of ther unobserved characterstcs and the results suggest that ths group receves a hgher educaton ncrement. Ther results therefore mply that schoolng aggravates wthn-group nequalty. Henrch and Hldebrand (2005) provde comparable estmates of the return to schoolng for the countres of the EU-15 by utlzng wave 3 (1996) of the European Communty Household Panel. Ther results show, that the returns are hghest n Portugal, Span and Luxembourg and lowest n the Nordc countres Denmark and Fnland as well as n the Netherlands. Correctng these estmates by the drect costs of educaton, whch n the orgnal Mncer-framework are assumed to be neglgble, ther estmates change only slghtly by no more than 10%, leavng the overall country-rankng unaffected. The authors further examne the lnk between educatonal attanment and labor force partcpaton as well as between educatonal attanment and unemployment. Ther results show, that the postve effect of educaton on labor force partcpaton s especally marked for women whereas the probablty of unemployment decreases strongly wth male educatonal attanment. Garca-Manar and Montuenga-Gomez (2005) also make use of the ECHP and estmate returns to educaton n Portugal and Span both for employees and self-employed workers, 6

9 respectvely. Schoolng, however, has to be consdered endogenous n the Mncer-framework and the authors try to overcome ths problem by usng an Instrumental-Varables (IV) approach. In partcular, they apply the Effcent Generalzed Instrumental Varables estmator (EGIV) by Hausman and Taylor (1981). Ths makes use of the panel data structure and uses all exogenous varables as well as means of the tme-varyng regressors as nstruments to obtan a consstent estmate of the return to educaton. Ther estmaton results show, that the returns to educaton n Span are hgher for wageearners than for self-employed workers whle the opposte holds n Portugal. In the latter country, both wage earners (9.5 %) and self-employed workers (10.0 %) exhbt a hgher return to educaton than ther counterparts n Span (8.8 % and 4.9 %, respectvely). The analyss of the relatonshp between demography and earnngs was motvated by the entry of the baby-boom cohorts n the US labor market. If workers are mperfect substtutes across ages, ndvduals born n a large brth cohort wll face lower wages (Welch 1979). Analogcal, exogenous shfts n educatonal attanment across cohorts wll change the relatve supply of hghly educated workers and thus the return to educaton (Card and Lemeux 2001). At frst glance less clearly, the pure sze of a brth cohort may matter for educaton prema, too. If jobs for more educated workers also requre more tranng, the substtutablty between young and old workers dmnshes wth educaton (Stapleton and Young 1988). Thus, ndvduals born n a large cohort wll face lower returns to educaton. Europe s facng two dstnct crcumstances, whch suggest a rather huge change n educaton prema across brth cohorts. Frstly, the number of lve brths has fallen dramatcally snce the md 1960s whch may postvely affect the return to educaton. Secondly, educaton reforms n Europe takng place n the 1960s to 1970s provded for a huge ncrease n educatonal attanment whch may abate educaton prema. Despte these reasons for a change n the returns to educaton n European countres, however, emprcal evdence s scarce. Boockmann and Stener (2006) analyze cohort effects n the returns to educaton for West German workers. For women ther results show a large contnuous declne n the wage premum and the cohort born n the early 1970s features a 6 percentage ponts lower return to educaton compared to the cohort born n the second half of the 1920s. For men of the same brth cohort, they fnd a rather small declne of the educaton premum by about one percentage pont. 7

10 3 Data descrpton and Estmaton Approach 3.1 Emprcal Approach The benchmark model for emprcal estmates of the economc return to schoolng s the Mncer-equaton (Mncer 1974) relatng earnngs ( y ) of the ndvdual to hs schoolng ( s ), 2 work experence ( e ) and squared work experence ( e ). The latter can be explaned by the proposton of human captal theory that returns to experence result from on-the-job tranng. Thus, the earnngs-experence profle s expected to exhbt concavty. X s a set of other varables affectng earnngs, for example the sector of the economy the ndvdual s workng n, whle u s the error term. Ths results n the followng model: 2 ln y = α + βs + χe + δe + γx + u. (3.1) In ths framework, β can be regarded as the prvate fnancal return to schoolng. Yet, the error term u n the wage equaton captures ndvdual unobservable effects. These may have an mpact on schoolng so that ths s correlated wth the error term. A prme canddate s the ablty of the ndvdual. Ths s unobservable but very lkely postvely correlated wth schoolng as more able ndvduals choose a hgher level of schoolng. Estmatng the Mncer equaton by OLS, part of the mpact of ablty s ascrbed to educaton whch mples an upward bas of β (Card 1999). 2 Grlches (1977) emphaszes another source of bas, though. Measurement error n the schoolng varable leads to a downward bas n the OLS estmator and ths may partally or fully offset the postve bas caused by the omsson of ablty. As adequate measures of cogntve ablty lke IQ-test scores are unfortunately not avalable n our data set, we cannot correct ths endogenety bas but wll provde OLS results. As we argued before, there are reasons to beleve that cohort effects are relevant for the returns to schoolng. To evaluate ths ssue, the Mncer-equaton s extended by cohort dummes whch yelds: ln y K K 2 = α + β s + χe + δe + γx + η k Dk + k k= 2 k = 2 = 1,..., n; t = 1994,...,2000. µ D s + νt k t + u, (3.2) 2 Augurzky (2001), utlzng data from the US Natonal Longtudnal Survey of Youth (NLSY), provdes evdence that the omsson of ablty mght even bas the functonal form of the Mncer-equaton, mplyng that returns to educaton ncrease wth schoolng, whereas the opposte seems to hold. 8

11 To ncrease the sample sze, equaton (3.2) s estmated on an unbalanced panel of at most 8 years and robust estmaton technques are utlzed that account for clusterng on ndvduals. T t denotes dummy varables depctng f the observaton s from the survey year t and denotes dummy varables whch take on the value one f the ndvdual s a member of the brth cohort k, whch may span several years. Thus, n equaton (3.2) we not only account for dfferng returns to educaton across brth cohorts, but also for brth cohort man effects on wages. Dk There s, however, a notable problem whch complcates the analyss. Due to the short tmeframe of the survey, we are unable to observe dfferent brth cohorts at the same stages of ther workng lfe. The youngest cohort, for nstance, s only observed n ther early workng years. It mght be though that educaton s rewarded dfferently across the lfe cycle. Emprcally, t s mpossble to separate tme-effects, cohort effects and lfe cycle-effects perfectly from one another because age, brth year and survey year are an exact lnear relatonshp (Heckman and Robb 1985). We wll subsequently account for these lfe-cycle effects by ncludng nteracton terms between educaton and experence (see Boockmann and Stener 2006). Ths yelds: ln y K K 2 2 = α + βs + χe + δe + γx + ϑes + ρe s + η k Dk + k k= 2 k = 2 = 1,..., n; t = 1994,...,2000. µ D s + νt k t + u, (3.3) Applyng OLS to equaton (3.1), the regresson lne passes through the mean of the sample. Thereby one mplctly assumes that a dfferng mpact of the exogenous varables along the condtonal dstrbuton s neglgble. If however, the earnngs gan from addtonal schoolng s not the same across the wage dstrbuton, schoolng would have an mpact upon wage nequalty. We therefore addtonally apply quantle regresson technques (Koenker and Bassett 1978), whch allow us to estmate the return to schoolng wthn dfferent quantles of the wage dstrbuton. In partcular, wage equaton (3.1) can be rewrtten as (see Martns and Perera 2004): (ln wth 0 < θ < 1. ln y = cλ θ + uθ and Quant θ y c ) = c λθ 9

12 For smplcty, c represents the vector of all exogenous varables and λ θ the vector of parameters. Quant θ (ln y c ) s the θ th condtonal quantle of ln y gven c. The θth regresson quantle solves the problem: mn ρ (ln y c k θ λθ ) λ R where ρ ( ) s the tlted absolute value functon that yelds the θ th sample quantle as the θ soluton. The mnmzaton problem can be solved by lnear programmng methods whereas the standard errors are obtaned by bootstrappng. The frst quartle s obtaned by settng θ = 0.25, the medan by settng θ = 0. 5 and so forth. Thus, by ncreasng θ from 0 to 1 we follow the whole dstrbuton of y, gven c. 3.2 Data set In ths paper data from the ECHP s utlzed. 3 Ths s a longtudnal data set whch covers eght waves from 1994 to It started wth households from 12 member states 4 and has thereafter been extended wth Austra, Fnland and Sweden jonng n 1995, 1996 and 1997, respectvely. For most of the countres, a harmonzed ECHP questonnare (the Communty questonnare ) has been used, whch s desgned by Eurostat. Exceptons are the UK, Germany, Luxembourg and Sweden for whch data was converted from natonal surveys. For the former two countres, converted data from natonal surveys s provded through 1997 whereas for the latter two countres the ECHP questonnare has been replaced by natonal surveys n Changes n the member states populaton s reflected n the data through brths to household members and the formaton of new households from exstng ones. Non-responses and attrton rates are comparable wth other longtudnal household surveys (Peracch 2002). The ECHP covers topcs as demographcs, labor force behavor, health, educaton, mgraton and ncome, especally earnngs and transfers. The ncome measure used n ths study s gross monthly wages. The survey further reports weakly workng hours, whch are used to calculate 3 For a descrpton of the ECHP see Eurostat (2003) as well as the webste of the EuroPanel Users Network (EPUNet) 4 Belgum, Denmark, Germany, Greece, Span, France, Italy, Ireland, Luxembourg, The Netherlands, Portugal and the UK. 10

13 the gross hourly wages, our dependent varable. To compare the estmates across countres as well as over tme, gross wages are adjusted by Purchasng Power Partes. In the ECHP, as n most European surveys, nformaton on the educaton of the ndvdual s only avalable as a coded varable denotng the hghest educatonal level completed. The correspondng three categores are less than second stage of secondary educaton (ISCED levels 0-2), second stage of secondary educaton (ISCED level 3) and tertary educaton (ISCED levels 5-7). In order to make the estmates comparable to other studes, these have been transformed to a contnuous varable years of schoolng by usng nformaton from the OECD s Educaton at a Glance on the usual tme-frame necessary to obtan a certan educatonal level n the varous countres (see table 7 n the appendx). However, the estmates do therefore only contan nformaton on the return to an addtonal year of educaton between schoolng levels. The experence measure used n the analyss s workng experence as stated n the questonnare, yet due to data lmtatons t s not possble to correct ths for unemployment spells. Addtonal control varables nclude dummes depctng f one s a natve ctzen and f the ndvdual s marred, sector dummes as well as regonal dummes on the NUTS-2 level. Generally, data avalablty across countres prevented us from estmatng a more detaled regresson. For the analyss, data from all waves of the ECHP s used. Estmatons are carred out for all male wage earners, snce results for females are lkely to be dstorted by selecton nto employment. Yet, Sweden and Luxembourg are not ncluded n the analyss, snce ncome s only stated net of taxes. The Netherlands are left out, snce apparently there seemed to be codng problem wth the Dutch educaton data (see also Henrch and Hldebrand 2005: 14). The mean characterstcs of the 1994 and 2001 samples are presented n table 1. Throughout the waves, there has been a notable drop n sample sze for most of the countres as ndvduals have fallen out of the sample. Ths problem s most severe n Ireland and the Nordc countres, where the hgh moblty of ndvduals s lkely to be the reason, whereas n Portugal there has even been a slght ncrease n sample sze. In 2001, mean hourly gross wages are hghest n Denmark and Belgum and lowest n Greece and Portugal. The latter two countres, though, have seen the largest rse n wages durng the perod under examnaton. 11

14 Table 1: Summary Statstcs Country Sample Sze Log hourly wages Schoolng years Experence Natve Marred Industry Sector Servce Sector (PPP-adjusted) Austra 1,950 1, (0.56) 2.44 (0.48) (2.32) (2.26) (11.94) (11.87) 0.95 (0.21) 0.97 (0.16) Belgum 1, (0.43) 2.65 (0.45) (3.18) (3.19) (10.60) (11.05) 0.91 (0.28) 0.98 (0.16) 0.74 (0.44) 0.70 (0.46) (0.48) (0.48) Denmark 1, (0.48) 2.75 (0.45) (3.38) (3.08) (13.12) (12.54) 0.98 (0.14) 0.99 (0.11) (0.48) 0.37 (0.48) Fnland 1, (0.46) (3.06) (2.75) (11.05) (12.41) 0.99 (0.89) 0.99 (0.10) 0.66 (0.47) France 2,273 1, (2.44) (3.08) (10.91) (13.11) 0.94 (0.24) 0.97 (0.17) 0.70 (0.46) Germany 3,017 2, (0.46) 2.55 (0.51) (3.01) (3.12) (11.71) (11.26) 0.81 (0.39) 0.86 (0.35) 0.73 (0.44) 0.71 (0.45) Greece 1,906 1, (0.47) 1.96 (0.48) (2.84) (2.65) (11.82) (11.66) 0.98 (0.13) 0.99 (0.11) 0.71 (0.45) 0.68 (0.47) 0.37 (0.48) 0.34 (0.47) (0.48) Ireland 2, (0.63) 2.40 (0.59) (2.24) (2.22) (13.08) (14.16) 0.99 (0.11) 0.99 (0.11) Italy 3,447 2, (0.41) 2.29 (0.42) (3,56) 11,48 (3.61) (11.67) (11.65) 0.99 (0.03) 0.99 (0.05) 0.72 (0.45) 0.69 (0.46) (0.48) Portugal 2,362 2, (0.58) 1.65 (0.52) 8.86 (2.11) 9.23 (2.47) (14.21) (14.28) 0.99 (0.07) 0.99 (0.04) 0.68 (0.47) 0.67 (0.47) Span 3,431 2, (0.54) 2.21 (0.51) (2.89) (3.02) (13.23) (12.90) 0.99 (0.07) 0.99 (0.07) 0.70 (0.46) 0.63 (0.48) UK 2,101 2, (0.55) 2.52 (0.53) (3.21) (2.88) (13.32) (13.76) (0.17) (0.48) (0.48) Notes: Fgures are means for the sample of employees n wave 1 and wave 8 of the ECHP. Standard devaton n parentheses. 1 : Fgures for Austra and Fnland from wave 2 (1995) and wave 3 (1996), respectvely. 2 : nformaton not avalable n wave 1. 12

15 In 2001, educatonal attanment of the workforce, as measured by schoolng years, s hghest n Germany and Denmark wth about 14 years of schoolng. Portugal, n contrast, stll shows by far the lowest schoolng of the workforce wth 9.2 years. Ths s notably below the OECD average of 11.4 years (fgure for 2004; OECD 2006: 41) and reflects that the educatonal expanson n Portugal started comparatvely late from the md 1970s on after the dctatorshp (Melo and Benavente 1978). In all but three countres the educatonal attanment of the workforce has rsen snce the begnnng of the survey. Exceptons are Fnland, France and Greece, where we observe a small declne. Germany shows the hghest share of foregn employees as well as, accordng to the mportance of the German ndustry sector, the hghest share of ndustry workers. The southern European countres as well as the UK nstead show a hgh share of employees n the servce sector. 4. Estmaton Results 4.1 Estmates of the average return to educaton We begn by estmatng the earnngs functon by OLS for the latest year of the survey, Estmaton results are reported n table 2. The estmated coeffcents on years of schoolng as well as experence are sgnfcant at the 1%-level and the model explans from 28 % (UK) to 43 % (Ireland) of the wage varaton n the sample countres. Results show that the average return for males to an addtonal year of schoolng s hghest n Ireland and Portugal wth rates of return of around 10.0 %, respectvely, followed by Austra and Greece wth rates of return of roughly 8 %. In contrast, returns to educaton are lowest n the UK, Italy and Germany wth rates of return of around 4.8 % to 5.5 %. Returns to experence are hghest n Austra, Germany and the UK. In the former two countres, ths mght be explaned by the rather hgh degree of unonzed wage settng wth determnated allowances for age or experence. Furthermore, n all countres, returns to experence exhbt the presumed concave pattern, that s earnngs ncrease wth on-the-job experence but at a dmnshng rate. Natves earn statstcally sgnfcant more only n fve countres of the sample and ths dfference s the largest n Italy. Furthermore, marred males receve a szeable wage premum n all countres except Denmark. 13

16 Table 2: Returns to educaton n Europe Estmaton Results by OLS for Male Wage Earners (Dependent varable: logarthm of gross hourly wage) Varable Austra Belgum Denmark Fnland France Germany Greece Ireland Italy Portugal Span UK Years of schoolng (16.60)*** (15.13)*** (13.54)*** (10.10)*** (17.60)*** (18.15)*** (18.86)*** (13.68)*** (22.46)*** (21.32)*** (21.71)*** (13.07)*** Experence (12.27)*** Experence 2 * (9.32)*** Natve (3.63)*** Marred (3.06)*** Industry Sector (2.03)** Servce Sector (1.38) Constant (3.64)*** (6.19)*** (3.94)*** (2.59)*** (2.26)** (1.96)** (1.92)* (4.72)*** (7.82)*** (6.73)*** (1.83)* (1.54) (0.58) (0.93) (9.63)*** (5.48)*** (3.35)*** (0.40) (2.28)** (1.27) (0.47) (4.51)*** (11.67)*** (8.29)*** (4.09)*** (8.09)*** (4.10)*** (3.26)*** (9.14)*** (9.40)*** (8.22)*** (0.77) (4.57)*** (4.02)*** (3.32)*** (6.59)*** (10.41)*** (7.10)*** (0.45) (4.70)*** (2.26)** (2.42)** (0.22) (4.97)*** (3.68)*** (0.15) (5.32)*** (6.33)*** (6.00)*** (1.91)* (13.53)*** (8.68)*** (2.64)*** (5.69)*** (4.51)*** (6.42)*** (5.80)*** (10.02)*** (9.06)*** (0.12) (6.17)*** (7.28)*** (9.43)*** (0.19) (11.96)*** (8.66)*** (1.11) (5.50)*** (6.99)*** (7.06)*** (2.87)*** (16.06)*** (15.36)*** (0.45) (5.99)*** (4.29)*** (3.98)*** (7.77)*** R Obs. 1, ,773 2,460 1, ,321 2,434 2,618 2,048 Notes: Data from wave 8 (2001) of the ECHP. Regonal dummes omtted. Robust estmates. t-ratos n parentheses. *: sgnfcant at 10% level; **: sgnfcant at 5% level; ***: sgnfcant at 1% level. 1 : Indvduals workng n the agrcultural sector form the bass. 14

17 As there are multple countres wth economc returns to schoolng that are smlar n sze, we further test f these pont estmates are sgnfcantly dfferent. For Germany, Italy, Denmark and Fnland, returns to educaton for men of equal value could not be rejected at conventonal sgnfcance levels. The same holds for the two countres wth the hghest returns n the sample, Portugal and Ireland. It s nterestng to see f the economc downturn that occurred n 2001 mght have caused a drop n the returns to schoolng. We nvestgate ths ssue further and allow for varyng coeffcents over tme. Ths means that the wage equatons are estmated separately for each wave of the ECHP, that s from 1994 to From the years 1995 to 1997 however, the results may be slghtly upward based as only the hghest educatonal attanment of new entrants to the survey has been recorded. Fgure 1: Returns to Educaton n Europe Austra Fnland Denmark Germany France Belgum Greece Ireland Italy Span Portugal UK Notes: Estmaton results by OLS based on the ECHP. 15

18 As can be seen from fgure 1, for the large porton of countres the returns to educaton have been very stable durng the observaton perod. In these countres, returns have changed by at most 1 percentage pont and there s no clear tendency across countres that they have ether ncreased or decreased. In contrast, three countres have experenced a rather huge change n educaton prema. In Portugal, the returns to schoolng have fallen by around 3 percentage ponts from a peak of 13.5% n 1996 to 10.4 % n 2001 whle n France returns dropped from 8.4% n 1994 to 6 % n In Greece, though, returns to educaton have ncreased by around 2 percentage ponts from 1994 to Explanng the cross-country dfferences n returns to educaton s a dffcult task as they reflect multple factors such as the demand for sklls, the dstrbuton of workers across occupatons, employment legslaton, the strength of unons, educatonal attanment and nternatonal trade (OECD 2006: 122). Despte the huge number of emprcal estmates of the Mnceran return to educaton, there are only very few attempts (Denny, Harmon and Lydon 2002, Psacharopulos 1994; Trostel et al. 2001) to explan ther cross-country varance. Acemoglu (1999), n a more general framework, shows that the mpact of the relatve supply of sklled workers on the skll prema depends on the skll bas of technology. When countres are technologcal followers, an ncrease n the supply of sklls wll not nduce skll-based technologcal change and the return to educaton decreases strongly wth the supply of sklls. In contrast, when the supply of sklls has a huge mpact on technology, the demand for sklls rses wth supply and may even nduce an ncrease n the educaton premum. In a smple model we nvestgate ths further by regressng the estmated returns to schoolng n our sample countres on the educatonal attanment of the populaton, as measured by average schoolng years from the data set of Barro and Lee (2001). As the Barro and Lee data s only avalable for 5-year ntervals, ths leaves 23 observatons (estmated returns n 1995 for 11 countres and estmated returns n 2000 for 12 countres), and our analyss should rather be regarded exploratve. Estmaton results (table 3) confrm prevous fndngs, that there s a negatve relatonshp between the educatonal attanment of the workforce and the returns to schoolng and thus varatons n returns to educaton across countres are closely related to scarcty (Psacharopoulos and Patrnos 2004). A rse n the years of schoolng of the workforce by one standard devaton (1.5 years n 2000) yelds a decrease n the return to educaton of 1.5 percentage ponts. Yet, as our dependent varable s tself an estmate, we 16

19 have to allow for the uncertanty assocated wth t. Ths s done n a further regresson by weghtng the estmates accordng to ther standard errors (varance-weghted least squares). Results show, that the pont estmate gets rather small, mplyng a drop n the return to schoolng by 0.75 percentage ponts f the years of schoolng of the workforce rse by 1.5 years. Although we cannot rule out changes n demand for sklls by usng ths smple framework, these were obvously not strong enough to balance the rse n supply of educated workers. Table 3: Returns to educaton across countres and educatonal attanment of the populaton OLS Varance-weghted LS Average years of schoolng (3.88) (8.95) n the populaton aged 15 and over Constant (7.01) (22.45) Obs Notes: Based on estmaton results from wave 2 (1995) and wave 7 (2000) of the ECHP and data from Barro and Lee (2001). t-ratos n parentheses. 4.2 Cohort effects and returns to educaton In the prevous secton, we already consdered that there mght by tme effects whch exert an mpact on the average return to schoolng. Although the latter remaned mostly stable n most of the sample countres over tme, the returns to educaton mght dffer dependng on the cohort an ndvdual s born n. Frstly, there may exst exogenous cohort specfc factors whch have an mpact on the relatve supply of educated workers and thus the return to schoolng. Secondly, even the absolute sze of the brth cohort mght have an mpact on educatonal prema f the elastcty of substtuton between young and old workers vares wth ther educatonal level. There s evdence that elastctes of substtuton ndeed decrease wth educaton, whch can be explaned by the fact that jobs for hgher educated workers also requre more on-the-job tranng (Stapleton and Young 1988). To analyze f cohort effects are present n our sample of European countres, we subsequently extend the Mncer-equaton wth brth cohort dummes as well as nteractons between schoolng and these dummy varables. To have a suffcent number of observatons, frstly we 17

20 Table 4: Returns to educaton n Europe and cohort effects Estmaton Results by OLS for Male Wage Earners (Dependent varable: logarthm of gross hourly wage) Varable Austra Belgum Denmark France Germany Greece Ireland Italy Portugal Span UK (2.08)** (0.64) (0.20) (0.83) (0.37) (1.14) (1.49) (0.79) (1.53) (0.71) Cohort (0.73) (1.65)* (0.31) (0.64) (0.87) (0.05) (1.91)* (1.07) (0.52) (1.94)* (0.06) Cohort (0.44) (0.94) (1.09) (0.15) (1.52) (0.65) (1.76)* (0.88) (0.74) (0.54) (0.47) Cohort (2.25)** (2.53)*** (0.94) (1.98)** (1.69)* (2.10)** (0.31) (2.00)** (1.44) (0.54) Cohort > (2.32)** (2.70)*** (7.16)*** (5.63)*** (6.77)*** (3.82)*** (1.97)** (7.26)*** (6.09)*** (7.71)*** (4.92)*** (8.47)*** Years of schoolng (1.95)** (1.10) (0.16) (0.41) (0.44) (1.43) (4.57)*** (0.88) (8.16)*** (0.43) Years of schoolng*experence (6.69)*** (0.53) (0.39) (0.36) (0.38) (0.13) (0.37) (3.52)*** (2.23)** (6.12)*** (0.45) Years of schoolng*experence (4.54)*** (1.73)* (0.04) (0.73) (1.68)* (0.02) (0.55) (0.19) (1.97)** (0.40) (1.63) Years of schoolng (1.16) (1.15) (0.95) (1.72)* (2.61)*** (0.88) (0.96) (2.21)** (0.68) (1.82)* Years of schoolng (1.72)* (0.45) (3.06)*** (1.76)* (3.72)*** (1.09) (0.15) (2.75)*** (2.83)*** (1.87)* (2.36)** Years of schoolng (3.64)*** (0.56) (4.90)*** (2.86)*** (4.72)*** (2.05)** (0.13) (1.58)*** (4.78)*** (3.01)*** (2.52)** Years of schoolng > (3.48)*** (4.98)*** (3.73)*** (1.56) (2.87)*** (2.23)** (0.26) (6.96)*** (1.81)* (10.35)*** (2.19)** Experence (9.15)*** (3.95)*** (2.08)** (1.67)* (2.28)** (1.18) (1.28) (6.41)*** (0.23) (8.74)*** (1.37) Experence 2 *10-2 (6.64)*** (4.03)*** (1.04) (0.68) (9.77)*** (1.11) (3.97)*** (0.57) (2.51)** (1.61) (4.27)*** (2.97)*** Constant ,671 4,627 2,995 3,259 5,221 4,219 5,921 3,142 15,095 22,187 12,413 11,710 22,989 19,681 22,983 10,694 F-test years of schoolng-cohort nteractons (p-value) R Indvduals 2,617 2,995 2,321 Obs. 11,377 12,413 9,573 Notes: Data from all waves of the ECHP. Regonal dummes, survey-year dummes, natve dummy, marred dummy and sector dummes omtted. Standard errors adjusted for clusterng on ndvduals. t-ratos n parentheses. *: sgnfcant at 10% level; **: sgnfcant at 5% level; ***: sgnfcant at 1% level.

21 pool the data and use robust estmaton technques that account for clusterng on ndvduals. Secondly we only dstngush fve dfferent cohorts: those born between 1940 and 1949, those born between 1950 and 1959, those born between 1960 and 1969, and those born 1970 or after whereas the oldest cohort, that s those born before 1940, s our base category. The Fnsh data, however, dd not provde enough observatons even for ths rather rough classfcaton so that Fnland had to be excluded from the sample. The estmaton results of the extended Mncer-equaton are presented n table 4. In all countres, except for Belgum and Greece, we fnd sgnfcant cohort effects n the returns to educaton as the schoolng-cohort nteracton terms are jontly sgnfcant at the 1%- sgnfcance level. Most nterestngly, the pattern of educatonal prema across adjacent brth cohorts s the same for nearly all countres wth sgnfcant cohort effects. In these countres, returns to educaton decrease contnuously over cohorts born after Fgure 2 shows ths graphcally. To compare the returns across cohorts, t s assumed that workers of all brth cohorts possess a work experence of 15 years, whch certanly does not hold for each cohort. Therefore, the absolute values of returns cannot be nterpreted but only the evoluton of returns across brth cohorts. The fgure shows, that n some countres the declne of returns to educaton s qute marked. For nstance, the youngest cohort n France suffers from a 10 percentage pont lower educaton prema compared to the oldest cohort whereas n Austra the dfference s 9 percentage ponts. Yet, t has to be bourn n mnd that the youngest cohort, who s born after 1969, may not have totally entered the labor market by the tme of the survey, whch mples that ther returns may be based downwards. Furthermore, the returns to educaton of the oldest cohort may be based upwards as there mght exst a postve selecton of older workers nto employment. Those ndvduals, whose returns to educaton are relatvely hgh, mght choose to stay n employment, possbly beyond ther retrement age. Dsregardng the youngest as well as the oldest cohort, the declne n educaton prema s less pronounced. Accordng to these estmates, returns to educaton n France and Austra have fallen by 6 percentage ponts for the cohort born n the 1960s compared to the cohort born n the 1940s. 5 Calculatons wth a slghtly dfferent classfcaton of brth cohorts showed, that ths general tendency also holds for Fnland. Results are avalable from the author upon request. 19

22 Fgure 2: Returns to educaton n Europe for dfferent brth cohorts Cohort <1940 Cohort Cohort Cohort Cohort >1969 Austra Denmark France Germany Ireland Italy Portugal Span UK Notes: Calculatons based on data from the ECHP under the assumpton that workers of all brth cohorts possess 15 years of work experence. One excepton to the general trend of fallng educaton prema s Germany, where we observe a u-shaped pattern. The returns do not change for the cohort born n the 1940s then decrease for the two followng cohorts before the youngest cohort experences a rse n the return of about 1 percentage pont. In the UK, the returns to educaton frst rse for ndvduals of the 1940s cohort before they decrease contnuously over all cohorts born after In the latter country, the varaton n returns over cohorts though s rather small. The youngest cohort exhbts a return whch s only 1.2 percentage ponts lower than that of the oldest cohort. As we could only control for experence, there may stll be an autonomous mpact of age on the return to educaton, whch s wrongly ascrbed to cohort effects. Ths problem, however, may be much more prevalent n cases where ndvduals work n the publc sector because there age, and not workng experence, s mostly relevant for the pay scale groupng (Boockmann and Stener 2006). We nvestgate ths further by estmatng our model just for employees workng n the prvate sector. Estmaton results (table 5) show, that n some countres the cohort effects n returns to educaton are even more pronounced for ths group of 20

23 ndvduals. Ths may be explaned by the more flexble wage-settng n the prvate compared to the publc sector. An excepton s Ireland, where the cohort effects are no longer sgnfcant. In all other countres our prevous result, namely that ndvduals born after 1940 are contnuously facng lower returns to educaton, s confrmed. Admttedly, n Austra, the UK and Portugal, the declne n returns apples to later cohorts born n the 1950s or 1960s, respectvely. The results are however at odds wth the presumed mpact of demography on educaton prema. From a peak n the md 1960s lve brths declned steadly all over Europe (Bag and Lucfora 2005) and ths should accordngly rase the returns to educaton for subsequent cohorts. A dfferent explanaton may be that educatonal attanment has rsen so strong n most European countres, that ths has lowered the return on nvestments n educaton. As explcated n the precedng secton, the overall scarcty of educated workers seems to be postvely related to the educaton prema. Drawng agan on the data set by Barro and Lee (2001), ths may partly explan our results. In Austra, for nstance, the educatonal attanment of the workforce (those aged between 15 and 65) ncreased n the 1980s (see fgure 3 n the appendx). The cohort born n the 1960s, whch entered the labor market about 20 years later, should therefore exhbt the largest declne n returns, whch s reflected n our results. Furthermore, Span and Portugal experenced the strongest ncrease n educatonal attanment of all sample countres and these are two countres where the returns to educaton fall off very strong. It has to be noted, however, that the general tendency of a drop n educaton prema across cohorts experenced n many of the sample countres mght be attrbuted to developments n certan felds of study. College enrollments mght have rsen n a partcular feld of study, outpacng demand for that specfc labor market group, whle n other felds of study there s stll a lack of graduates. In Germany, for nstance, there s a reported endurng lack of engneers. Unfortunately, t s not possble to nvestgate ths ssue further. Nether s the feld of study part of the questonnare of the ECHP, nor s ths nformaton avalable at the country level. 6 6 Data from the World Bank on the feld of study of college graduates just starts n

24 Table 5: Returns to educaton n Europe and cohort effects Estmaton Results by OLS for Male Wage Earners n the prvate sector (Dependent varable: logarthm of gross hourly wage) Varable Austra Belgum Denmark France Germany Greece Ireland Italy Portugal Span UK (2.09)** (0.15) (1.25) (0.07) (0.31) (1.52) (0.13) (1.10) (0.62) (1.35) Cohort (0.58) (2.03)** (0.41) (1.14) (0.13) (0.14) (1.22) (0.21) (1.12) (0.83) (1.03) Cohort (0.59) (0.99) (1.71)* (0.82) (0.40) (0.01) (1.10) (1.47) (1.28) (1.75)* (1.34) Cohort (1.14) (0.63) (2.64)*** (0.03) (0.95) (0.44) (0.89) (0.46) (2.11)** (2.59)*** (1.20) Cohort > (1.51) Years of schoolng (2.31)** (4.87)*** (2.75)*** (3.70)*** (1.48) (1.64)* (6.81)*** (4.69)*** (8.19)*** (5.25)*** (5.80)*** (2.04)** (0.08) (0.17) (0.65) (0.37) (0.27) (3.61)*** (2.53)*** (7.49)*** (0.18) Years of schoolng*experence (7.09)*** (0.62) (1.35) (0.20) (0.57) (0.20) (0.69) (2.27)** (3.36)*** (4.74)*** (1.04) Years of schoolng*experence (4.88)*** (1.79)* (0.47) (0.88) (0.54) (0.12) (1.30) (1.09) (1.61) (0.63) (2.15)** Years of schoolng (0.68) (1.50) (1.03) (0.45) (0.94) (0.08) (0.56) (1.66)* (2.07)** (0.64) (2.74)*** Years of schoolng (0.02) (0.58) (2.68)*** (0.09) (1.76)* (0.17) (0.11) (3.09)*** (2.55)*** (3.35)*** (3.15)*** Years of schoolng (1.68)* (0.34) (3.71)*** (1.09) (2.63)*** (0.54) (2.11)** (3.85)*** (4.38)*** (3.15)*** Years of schoolng > (1.99)** (4.81)*** (2.56)*** (1.54) (2.07)** (1.72)* (1.44) (5.93)*** (0.12) (9.90)*** (1.85)* Experence (9.43)*** (3.70)*** (0.60) (1.19) (1.43) (0.97) (1.88)* (5.03)*** (1.42) (7.64)*** (0.83) Experence 2 *10-2 (6.58)*** (3.17)*** (0.51) (0.42) (6.86)*** (0.24) (2.42)** (1.24) (1.46) (0.11) (2.76)*** (2.40)** Constant F-test years of schoolng-cohort nteractons (p-value) R Indvduals 2,059 1,510 1,815 2,859 3,909 2,168 2,586 3,872 3,607 5,039 2,750 Obs. 8,401 5,050 6,859 11,235 17,317 7,623 8,228 15,273 15,599 18,056 8,896 Notes: Data from all waves of the ECHP. Regonal dummes, survey-year dummes, natve dummy, marred dummy and sector dummes omtted. Standard errors adjusted for clusterng on ndvduals. t-ratos n parentheses. *: sgnfcant at 10% level; **: sgnfcant at 5% level; ***: sgnfcant at 1% level.

25 Furthermore, the observed declne n educaton prema n most of the sample countres may be drven by decreasng educaton qualty. Studes for the U.S. (Murnane et al. 2000; Lazear 2003), Canada (Fnne and Meng 2002; Green and Rddell 2003) and the Unted Kngdom (McIntosh and Vgnoles 2001) suggest that cogntve sklls have an mpact on ndvdual earnngs. Meanwhle, Wößmann (2002: 106) compares students scores on varous tests of student performance n some of our sample countres, namely Belgum, France, Germany and the UK, over tme and shows that these have actually fallen between 1970 and 1995, despte the huge ncrease n spendng on educaton. 7 Thus, the apparent declne n educaton qualty may be an addtonal factor causng the drop n educaton prema across subsequent cohorts. 4.3 Quantle Regressons In order to evaluate what masks behnd the average return to educaton across the wage dstrbuton, we wll subsequently apply quantle regresson (QR) technques whch allow estmates of the returns wthn dfferent quantles of the wage dstrbuton. Most nterestng s the mpact of schoolng on wages at the very top and at the very end of the wage dstrbuton. Table 6 therefore shows estmates of the returns to schoolng for the top (9th percentle) and the bottom (1st percentle) of the wage dstrbuton together wth the OLS estmates for the survey years 1994 (1995 and 1996 n the case of Austra and Fnland, respectvely) and Estmaton results show, that n all countres and both survey years, except Denmark n 2001, the return to educaton for the 9th decle s hgher than the return for the 1st decle. The case of Portugal n 2001 may llustrate ths general fndng. Whle the OLS estmate of the return to schoolng s 10.4 %, the bottom decle features a return to educaton of just 7.0 % whereas the top decle exhbts a return of 11.7 %. Thus, returns to schoolng ncrease over the wage dstrbuton. Indvduals, who are at the top of the condtonal wage dstrbuton, are there because of ther unobserved characterstcs and the results suggest that ths group receves a hgher educaton ncrement. We have tested f the dfferences n pont estmates between the decles are statstcally sgnfcant. 8 In all but four countres (Austra, Denmark and Fnland n 1994 and 2001 as well as Belgum n 2001) the null of no dfference n the coeffcents could 7 One excepton beng Italy, where student performance has slghtly mproved. 8 Results are avalable from the author upon request.

26 be rejected. Thereby, our results n general confrm those by Martns and Perera (2004) for a somewhat dfferent country sample. 9 Table 6: Estmates of the Return to Schoolng (n %) by quantle regressons and OLS Year 1st dec. 9th dec. OLS Year 1st dec. 9th dec. OLS Austra (7.08) Belgum (5.16) Denmark (5.59) Fnland (6.72) France (10.65) Germany (6.46) Greece (6.32) Ireland (7.14) Italy (9.31) Portugal (6.82) Span (12.49) UK (5.04) 8.7 (11.54) 6.7 (20.89) 5.0 (12.65) 5.7 (13.82) 9.7 (15.52) 5.4 (11.38) 7.8 (16.45) 11.0 (11.71) 6.6 (25.99) 13.8 (17.28) 9.5 (13.12) 8.3 (15.31) 5.5 (16.0) 4.5 (14.44) 6.1 (17.59) 8.4 (20.42) 4.7 (16.43) 6.3 (16.69) 10.4 (18.70) 4.8 (20.05) 12.5 (19.40) 8.3 (25.31) (8.52) (8.95) (9.29) (4.95) (8.65) (13.16) (7.53) (7.39) (10.54) (8.35) (10.40) 6.3 (10.34) 5.7 (15.80) (4.42) Notes: Own calculatons based on the ECHP. t-ratos n parentheses. 8.6 (13.67) 6.6 (17.38) 5.8 (8.75) 6.7 (6.92) 7.1 (11.75) 5.9 (13.48) 10.3 (17.94) 11.9 (8.59) 7.0 (17.20) 11.7 (22.90) 8.6 (13.37) 5.9 (8.77) 8.3 (16.60) 6.3 (15.13) 5.0 (13.54) 5.9 (10.10) 6.0 (17.60) 5.5 (18.15) 8.4 (18.86) 10.6 (13.68) 5.3 (21.55) 10.4 (21.32) 7.2 (21.71) 4.8 (13.07) Aras et al. (2001) show, that once controllng for ablty dfferences, there reman sgnfcant dfferences n returns to educaton along the wage dstrbuton. Thus, the omsson of ablty n the wage equaton s unlkely to be the man reason for our results. Martns and Perera (2004: 365 f.) offer dfferent explanatons for the observed pattern. In general, schoolng 9 The sample of Martns and Perera (2004) ncludes the Netherlands, Norway, Sweden, Swtzerland and the US but does not contan Belgum. 24