Barriers to Entry, Deregulation and Workplace Training

Transcription

1 DISCUSSION PAPER SERIES IZA DP No Barrers to Entry, Deregulaton and Workplace Tranng Andrea Bassann Gorgo Brunello Aprl 2007 Forschungsnsttut zur Zukunft der Arbet Insttute for the Study of Labor

2 Barrers to Entry, Deregulaton and Workplace Tranng Andrea Bassann OECD, CEPN, Unversty of Pars 13 and ERMES, Unversty of Pars 2 Gorgo Brunello Unversty of Padova, KIER Kyoto, CESfo and IZA Dscusson Paper No Aprl 2007 IZA P.O. Box Bonn Germany Phone: Fax: E-mal: za@za.org Any opnons expressed here are those of the author(s) and not those of the nsttute. Research dssemnated by IZA may nclude vews on polcy, but the nsttute tself takes no nsttutonal polcy postons. The Insttute for the Study of Labor (IZA) n Bonn s a local and vrtual nternatonal research center and a place of communcaton between scence, poltcs and busness. IZA s an ndependent nonproft company supported by Deutsche Post World Net. The center s assocated wth the Unversty of Bonn and offers a stmulatng research envronment through ts research networks, research support, and vstors and doctoral programs. IZA engages n () orgnal and nternatonally compettve research n all felds of labor economcs, () development of polcy concepts, and () dssemnaton of research results and concepts to the nterested publc. IZA Dscusson Papers often represent prelmnary work and are crculated to encourage dscusson. Ctaton of such a paper should account for ts provsonal character. A revsed verson may be avalable drectly from the author.

3 IZA Dscusson Paper No Aprl 2007 ABSTRACT Barrers to Entry, Deregulaton and Workplace Tranng * We develop a theoretcal and emprcal analyss of the mpact of barrers to entry on workplace tranng. Our theoretcal model yelds ambguous predctons on the sgn of ths relatonshp. On the one hand, gven the number of frms, a deregulaton reduces profts per unt of output, and thereby reduces tranng. On the other hand, the number of frms ncreases, and so does the output gan from tranng, whch facltates the nvestment n tranng. Our numercal smulaton shows that for reasonable values of the parameters a negatve relatonshp prevals. We use repeated cross secton data from the European Labour Force Survey to nvestgate emprcally the relatonshp between product market regulaton and tranng ncdence n a sample of 15 European countres and 13 ndustral sectors, whch we follow for about 7 years. Our emprcal results are unambguous and show that an ncrease n product market deregulaton generates a szeable ncrease n tranng ncdence. JEL Classfcaton: J24, L11 Keywords: tranng, product market competton, Europe Correspondng author: Gorgo Brunello Department of Economcs Unversty of Padova va del Santo Padova Italy E-mal: gorgo.brunello@unpd.t * The vews expressed here cannot be attrbuted to the OECD or ts member countres. We are grateful to Kenn Arga, Mara De Paola, Alfonso Rosola, Francsco Serranto, Sad Souam and to the audences n Mlan Cattolca and Pars 13 (CEPN) for comments. The usual dsclamer apples.

4 Introducton Product and labour market deregulaton have attracted consderable attenton by economsts and polcy makers, and there s a broad concern that these regulatons can hamper growth and ncrease unemployment (see Blanchard and Gavazz, 2003, Alesna et al, 2005). Emprcal research on the economc effects of product market regulaton and deregulaton has focused so far on employment (Kugler and Pca, 2004), productvty growth (Ncolett and Scarpetta, 2003, Aghon et al. 2006), nvestment n captal stock (Alesna et al, 2005), nnovaton (Aghon et al., 2005), and easness to start a busness (Djankov et al, 2001). Ths paper nvestgates the relatonshp between product market regulaton / deregulaton and tranng, whch we defne as the accumulaton of human captal takng place n frms after school completon. Lookng at tranng s mportant because the accumulaton of sklls s perceved as one of the man factors affectng productvty growth (see e.g. OECD, 2007). The lterature n ths area s small but wth controversal results. Some theoretcal research suggests that deregulaton can be bad for tranng. Gersbach and Schmutzler, 2003 and 2005, for nstance, argue that more ntensve competton n the product market reduces tranng ncdence, especally when the country s large. The dea s that a larger number of compettors ncreases the rsk of poachng, whch dscourages employers from tranng. However, when the ncrease n competton affects the physcal proxmty of workers and frms, the negatve mpact of the rsk of poachng on tranng can be compensated by the postve local agglomeraton effects. Brunello and De Paola, 2007, show that n ths case the relatonshp between product market competton and tranng cannot be sgned a pror. The work by Acemoglu and Pschke, 1999, on tranng n mperfect labour markets can also have mplcatons for the relatonshp of nterest: snce deregulaton reduces rents, t mght reduce the employer's ncentve to nvest n general tranng. In contrast, Autor, 2001, provdes a model showng that tranng can ncrease wth product market competton. By focusng on temporary help frms, he shows that tranng s a tool to mprove the qualty of servce provded by frms n response to fercer competton. Emprcal work on ths ssue s scarce but less controversal and ponts to a postve relatonshp between competton and tranng. Autor, 2001, for nstance, shows that the Herfndhal ndex and tranng are negatvely correlated n US temporary help frms, whch suggests that less competton reduces tranng. In a mult-country study of Europe, whch explots cross country and tme seres 2

5 varatons n product market and labour market nsttutons, Bassann et al, 2007, fnd evdence of a negatve and statstcally sgnfcant correlaton between the ndex of product market regulaton developed by the OECD and tranng ntensty, thereby supportng and generalzng Autor's fndngs. One potental drawback of mult-country studes s that the combnaton of the country and tme dmensons does not allow to fully control for confoundng factors whch mght affect tranng and that vary across country and over tme. One way to overcome ths lmt s to add an addtonal dmenson to the data, as we do n the current paper. We buld on the approach followed by Bassann et al, 2007, by addng the sector as such dmenson. Wth data whch vary by country, year and sector, we can compare a treatment group, whch conssts of the sectors drectly affected by our measure of deregulaton, wth a control group, composed of those sectors whch are not affected drectly by such measure, whle at the same tme controllng for country and country by tme specfc effects. The assgnment to sectors s random nsofar as we can safely rule out the possblty that frms antcpatng polcy reform can swtch sector of producton. Gven the costs of swtchng sector, we beleve that ths source of endogenety s of second order of relevance. We use the OECD database on product market deregulaton developed by Ncolett and Scarpetta, 2003, whch has been desgned to pck up regulatory reforms n tradtonally heavly regulated sectors, such as transport, communcaton and publc utltes. We consder these sectors as our treatment group, and contrast ther behavour wth respect to tranng wth the control group, whch conssts of the manufacturng sector, lmtng our analyss to European countres. After the mplementaton of the sngle market programme n Europe n the early 1990s, we can assume that no sector-specfc regulatory reform affected the manufacturng sector, whch therefore qualfes as a genune control. 1 By comparng tranng ncdence n these sectors, we try to dsentangle the effects of deregulaton on tranng from other confoundng factors, whch affect both groups. To do so we match regulaton data wth tranng data from the European Labour Force Survey as well as several other sector-level databases, to obtan a rch sector-level database of tranng co-varates. Our evdence clearly suggests that product market deregulaton ncreases tranng ncdence. The estmated effects are sgnfcant, because a one pont reducton n the selected measure of regulaton whch ranges between 0 and 6 s estmated to trgger a 10 percent ncrease n tranng ncdence. To motvate our emprcal nvestgaton, we develop a model whch casts the tranng decson n an economy characterzed by mperfectly compettve product and labour markets, as n 1 The manufacturng sector, however, could be affected ndrectly, because of the close lnkage among sectors typcal of modern economes. 3

6 Blanchard and Gavazz (2003). We show that n equlbrum the mpact of deregulaton on tranng s ambguous. On the one hand, a reducton n the barrers to entry for a gven number of frms compresses profts per unt of output, and thereby reduces tranng. On the other hand, and condtonal on profts per unt of output, addtonal frm entry due to deregulaton ncreases the output gans from tranng, whch facltates nvestment. We nterpret our emprcal results as suggestve of the fact that the latter effect prevals on the former. The paper s organzed as follows. In the frst secton we develop the theoretcal model. The subsequent sectons ntroduce the emprcal strategy and the data and present our estmates of the relatonshp between tranng and product market deregulaton. Conclusons follow The Model Followng Blanchard and Gavazz, 2003, and Stevens, 1996, consder a two - stage model economy where each frm produces a dfferentated product usng labour. The number of frms m s determned by an entry condton, whch s affected by product market regulaton. The logcal sequence of the model s as follows: frst of all, frms decde entry. In the frst preparatory stage after entry, each frm nvests n tranng and pays the tranng costs. The wllngness to pay s justfed ether because the mparted sklls are frm specfc or because labour market frctons ncludng search and nformatonal asymmetres substantally reduce the transferablty of general sklls from a frm to another, as dscussed by Acemoglu and Pschke, In both the decson to enter and n the choce of how many workers to tran, the employer can perfectly foresee the wages she wll pay n the fnal stage, the level of employment L, the prces she wll be able to sell her goods at and the number of frms operatng n the market. In the second and fnal stage, and condtonal on tranng, frms and workers bargan over wages and employment, prces are set and producton occurs. Frms n ths economy share the same producton and tranng technology, and the same elastcty of product demand. Rsk neutral workers have the same reservaton wage, and there are no exogenous separatons of workers from frms. We show that the equlbrum s symmetrc: all frms choose the same tranng ncdence, each product prce s equal to the average prce, and each traned worker receves the same wage. Wth symmetry, n equlbrum there s no ex-post labour moblty and 2 Bassann et al, 2007, revew the evdence suggestng that frms bear most of tranng costs. In an economy where frms repeatedly nteract and sklls are only partally transferable, the gans from poachng can be more than offset by the costs, whch nclude addtonal tranng as well as reputaton costs. 4

7 ex-post poachng. Frms operate the technology e Y = AL [1] where Y s output, e L s employment n effcency unts, A s the productvty of labour and the subscrpt s for the sngle frm. Labour conssts of traned ( L T ) and untraned ( LU ) workers. Traned workers are sklled and untraned workers are unsklled. Snce traned workers are more productve than untraned workers, employment n effcency unts s L e = L + γl [2] U T where γ > 1 s the relatve effcency of sklled labour 3. Lettng T be tranng ncdence and L employment, L = T L and the producton functon can be wrtten as 4 T [ 1+ ( ) T ] Y = AL γ 1 [3] Snce employment s set by barganng n the second stage of the game, the decson on how many workers to tran n the ntal perod s equvalent to the choce of tranng ncdence T. Defne net and gross profts as net and gross of tranng costs. The characterzaton of the equlbrum proceeds by backward nducton and starts wth the decson concernng wages and prces. We assume that untraned workers receve the common reservaton wage V. Traned workers, however, bargan over wages and prces wth the frm. Ex-post barganng can be justfed by the fact that, when frms nvest n tranng, workers cannot commt on ex-ante wages but re-contract after the tranng nvestment has taken place (Malcolmson, 1997). Let the wage of untraned workers be W U = V and the wage of traned workers be W T = W, wth W V. We characterze the barganng as a cooperatve Nash game, whch we solve by maxmzng [( W V ) T L ] + (1 β )ln{ PY L ( W T + V (1 T ))} β ln [4] 3 Total employment s nstead defned as L = LU + LT. 4 See Dearden, Reed and Van Reenen,

8 wth respect to wages and employment. Ths s equvalent to assumng that the partes are nvolved n effcent barganng. An alternatve characterzaton would be the rght to manage model, where the employer retans authorty over employment determnaton. We follow Blanchard and Gavazz n the preference for effcent barganng, whch they argue has the advantage of capturng the possblty that frms are not operatng on ther demand for labour. However, we show n the Appendx that our key qualtatve results are not affected f we were to choose the rght to manage model. In [4] the element n brackets s the gan from a postve settlement accrung to traned workers, who have lnear utlty functons. Ths gan s weghted wth the relatve barganng power of sklled labour, β. By defnton, unsklled workers do not gan from the bargan, and are bound to ther reservaton wages. The element n braces s the gross nomnal proft from a postve settlement accrung to the employer 5. We assume that profts n the event of falure to settle are equal to zero. Notce that employment settng s equvalent to prce settng, because labour demand s a derved demand and each frm faces the followng product demand functon θ Y P Y = [5] m P ' where θ = σg( m) > 1, σ s a sutable constant and g > 0, Y s aggregate output, and P and P are the prce of product and the average prce. Followng Blanchard and Gavazz, 2003, the absolute value of the elastcty of demand wth respect to the relatve prce θ s ncreasng n the number of frms. Ths follows from the fact that an ncrease n the number of products assocated to entry rases the elastcty of substtuton between products. The outcome of the bargan s [ 1+ ( γ 1) T ] + V ( T β ) T W = βp A [6a] P A[ 1+ ( γ 1) T ] θ = [ W T + V (1 T )] θ + β 1 [6b] 5 Profts are gross because tranng costs n the second stage are bygones. 6

9 7 The former condton can also be wrtten n a more famlar way as + = VT L PY T V W β. The rght hand sde of the latter condton s the prce cost margn Λ, where the costs are net of tranng costs. Smple manpulaton of these two condtons yelds 6 ( ) [ ] T PA V P P = γ θ θ [7] We next turn to the ntal preparatory perod, when employers decde how many workers to tran. Let the cost of tranng per employee be 2 ) ( 2 T T c μ = [8] a convex ncreasng functon of tranng ncdence. Ths assumpton captures decreasng returns n the producton of human captal and rules out corner solutons for optmal tranng ncdence. Usng [5] and [6a] n the defnton of net real profts Π, 7 we obtan ( ) ( ) ( ) ( ) ( ) ( ) [ ] θ θ θ γ θ θ γ μ γ θ β T A P V m Y T A T T PA V = Π [9] The term n braces on the rght-hand sde represents net real profts per unt of output π, whle the product of all the other terms on the rght-hand sde represents output Y. The margnal varaton of real profts as a result of addtonal tranng s gven by ( ) ( ) ( ) ( ) ( ) ( ) ( ) T Y T T A T Y T Y T Y T 1 1 1) ( 1 1 1) ( = + = Π γ γ θ π γ π γ γ μ π π [10] 6 Employment s gven by [ ] [ ] θ γ γ + = + = P P T m A Y T A Y L 1) ( 1 1 1) ( 1 7 That s, nomnal profts dvded by the average prce.

10 The frst-order condton for a maxmum wth respect to T ( Π / T = 0) yelds 8 ( θ ) A( γ 1) π = μt 1 [11] Tranng s ncreasng n net real profts per unt of output π and, condtonal on π, s decreasng n tranng costs μ and ncreasng n the dfference between the productvty of sklled and unsklled workers A ( γ 1) as well as n the elastcty of substtuton θ. The latter effect comes from the fact that output gans from tranng are ncreasng n θ - see the last term on the rght hand sde of Eq.10. Substtutng out π n [11] wth ts expresson n terms of T, V and other parameters we obtan 2 2 V 1+ θ μ ± μ + 4μ( γ 1) (1 β ) * T = P 2 > 0 [12] μ( γ 1)(1 + θ ) Notce that [12] yelds a unque postve soluton. Snce both V and the parameters γ, θ, μ and β do not vary across frms, t must be that * * T = T, mplyng that all frms nvest n the same tranng ncdence. By mplcaton, equaton [7] suggests that at the optmal level of tranng frms have the same relatve prce. Snce n general equlbrum we cannot have that all frms have a relatve prce above or below 1 P (see Blanchard and Gavazz, 2003), t must be that P s equal to 1, whch mples V P θ 1 = A γ θ [ 1+ ( 1) T ] [13] where we drop hereafter the subscrpt. Usng [13], [5] and [6a] n the defnton of profts, and mposng the general equlbrum condton on prces, the equlbrum number of frms s determned by the condton that net profts per unt of output π must be equal to the cost of entry per unt of output ρ 8 The second order condtons for a maxmum hold. 8

11 1 β μt π = = ρ θ 2A 2 [ 1+ ( γ 1) T ] [14] 1 β Λ 1 Λ where = s gross real profts per unt of output, and > 0. A reducton n ρ corresponds θ Λ ρ to product market deregulaton and lower barrers to entry. Condtonal on tranng, a deregulaton ncreases the number of frms and θ. As a consequence, profts per unt of output fall untl the arbtrage condton [14] s satsfed. Blanchard and Gavazz, 2003, use a smlar convenent specfcaton for the cost of entry. In the Appendx we show that our results stll hold when we specfy more conventonally the cost of entry as a fxed cost, rather than as a cost proportonal to output. Usng [14] nto [11] to elmnate profts per unt of output we obtan ( θ 1) ( γ 1 ρ μ T = A ) [15] Equatons [14] and [15] descrbe two schedules n the ( T,θ ) plane: the schedule TT assocated to [15] mples a postve relatonshp between tranng ncdence and the number of frms 9 (see Fgure 1). The reason s that, condtonal on net profts per unt of output, the output gans from tranng are larger the larger the elastcty of substtuton between products (see eq. [11]). On the other hand, the schedule MM assocated to [14] suggests a negatve relatonshp. Condtonal on net profts per unt of output, the greater the tranng, the greater the gross profts that are requred to avod losses. In turn, ths mples that, f net profts are fxed by entry barrers, the greater the tranng the smaller the number of frms that can survve n equlbrum wthout makng losses. θ 9 ' Recall that = ( m) > 0 m σg. 9

12 T TT MM θ Fgure 1: The relatonshp between θ and T We establsh the followng Lemma. The equlbrum exsts and s unque f ρ 1 β. Proof: equatons [14] and [15] descrbe two schedules n the ( T,θ ) plane: the schedule TT assocated to [15] has a postve slope and the schedule MM assocated to [14] has a negatve slope. Moreover, tranng ncdence tends to zero as θ 1 n [15] and f ρ 1 β to a non-negatve number n [14]. Therefore, the two schedules ntersect n the relevant doman of the two varables, and ther ntersecton s unque. QED Proposton. A deregulaton of the product market whch corresponds to a reducton n the parameter ρ ncreases tranng ncdence f 10

13 1 β θ ( ρ) 1 1 β < ρ θ ( ρ) θ ( ρ) θ ( ρ) and reduces tranng ncdence f 1 β θ ( ρ) 1 0 < ρ < θ ( ρ) θ ( ρ) [16] Proof: total dfferentaton of [14] and [15] yelds Σ 1dm + Σ 2dT = Σ 5dρ Σ 3dm + Σ 4dT = Σ6dρ where ' 1 β ' Σ 1 = σ g ( m)( γ 1) ρa Σ 2 = μ Σ 3 = g ( m) 2 σg( m) Σ = ( σ g ( m) 1)( γ 1) A Σ = Σ 4 2μT + μt = 2A 2 ( γ 1) [ 1+ ( γ 1) T ] 2 Snce the determnant of the Jacoban s negatve when θ < + and ρ > 0, dt dρ < 0 f Σ 1 Σ6 Σ3Σ5 > 0, or 1 β θ ( ρ) 1 < ρ θ ( ρ) θ ( ρ) 1 β Non-negatve profts also mply see [15] that ρ. Therefore dρ < 0 θ ( ρ) 1 β θ ( ρ) 1 1 β < ρ θ ( ρ) θ ( ρ) θ ( ρ) Clearly, dt > 0 dρ f 1 β θ ( ρ) 1 > θ ( ρ) θ ( ρ) ρ. QED We also establsh the followng corollary dt f Corollary. A deregulaton of the product market ncreases the number of frms m. Proof. Follows from the proof above and from the fact that Σ Σ Σ Σ > In other words, condton [16] s less lkely to hold when deregulaton starts from a very low value of ρ or the product market s close to perfect competton ( θ ). However, for reasonable values of the parameters condton [16] holds even for relatvely low values of ρ. To llustrate ths 11

14 numercally, we set the confguraton of parameters so as to obtan that when the rato of the measure of barrers to entry to ts maxmum possble value ( 1 β ) s equal to the observed medan rato, n the sample for whch we have tranng and regulaton data (see next secton), we obtan approxmately an optmum tranng ncdence equal to the observed medan tranng ncdence. 10 Ths rato s equal to 0.62 accordng to our ndcators. Assumng μ = 0. 1, γ = 1. 2, β = 0. 5 and A = 1 (see the Appendx for the justfcaton of these choces), for ρ = (0.62*0.5) we obtan that optmal tranng ncdence s equal to 36,1%, not far from the medan level that we observe n our data (38,9%). Wth these parameters, a deregulaton whch cuts ρ ncreases tranng so far as ρ > (see Fgure 1), a low threshold, whch has rarely been attaned n hstorcal perspectve, and corresponds only to about 7% of T the observatons for whch we have data. Even more, 0 n a relatvely large regon, and t starts gettng sgnfcantly greater than zero only for ρ < 0. 03, a value that corresponds to less than 3% of the observatons for whch we have data. Conversely, for values above ρ = 0. 2, whch corresponds approxmately to the frst quartle n our data, the relatonshp between barrers to entry and tranng stocks s negatve and approxmately lnear. ρ T Rho Fgure 2: The relatonshp between ρ and T wth the selected confguraton of parameter values. 10 Where tranng ncdence s defned n terms of the stock of tranng (see next secton for the method of calculaton). 12

15 The effect of product market deregulaton on tranng ncdence s llustrated n Fgure 3, where contnuous lnes refer to the ntal equlbrum and dashed lnes to the new equlbrum. Whle the effect of deregulaton on the number of frms s not ambguous, the effect on tranng depends on the relatve shft of the two curves. Takng as gven the amount of tranng of each frm, a reducton n the barrers to entry ncreases the number of frms, thereby shftng the MM schedule outwards see equaton [14]. Ths, n turn, ncreases tranng ncdence along the TT schedule snce, condtonal on net profts, output gans from tranng are greater see equatons [10], [11] and [15]. However, for a gven number of frms, lower entry barrers also rotate the TT schedule downwards, 11 because they reduce net profts per unt of output and the net gans from tranng see equatons [10] and [15]. The overall effect on T s ambguous, but s postve f condton [16] holds. T TT TT MM MM θ Fgure 3: The effect of reducng barrers to entry 11 The barycentre of the rotaton s the pont (1,0) 13

16 So far, we have assumed that deregulaton does not nfluence total factor productvty A. However, n models of prce competton such as ours, theory tends to predct that, f frms are symmetrc, deregulaton ncreases the ncentve to nnovate (see Aghon et al., 2001). In addton, accordng to Schantarell (2005), the balance of emprcal evdence n the US and n Europe suggests that productvty ncreases wth deregulaton. If we relax our assumpton and let A = A(ρ), wth A / ρ < 0, the overall mpact of deregulaton on tranng s more lkely to be postve. However, we hasten to stress that postulatng a negatve relatonshp between regulaton and total factor productvty s not necessary for the negatve relatonshp between regulaton and tranng to hold. The Emprcal Model In the emprcal applcaton, we are nterested n nvestgatng whether changes n regulaton affect tranng ncdence, measured as the proporton of workers recevng tranng. Equatons [14] and [15] generate n mplct form a map from deregulaton ρ to tranng ncdence T, whch s plotted n Fgure 1 for a selected confguraton of parameters. As descrbed n detal n the data secton below, our measures of regulaton vary by sector, country and over tme. Therefore, we collapse our data on tranng and addtonal controls at the same level of aggregaton. As dscussed n the prevous secton, for most of our sample a monotone (negatve) relatonshp between regulatory barrers and tranng prevals. We can therefore estmate the followng emprcal counterpart of the theoretcal model: 12 T ln = λ 0 + λ1 X ct + λ2rct + ε ct [17] 1 T ct where the vector X ncludes a vector of controls, such as average age, educaton and frm sze, R s the measure of regulaton, the subscrpt s for the ndustry, c s for the country and t s for tme. We 12 From an emprcal pont of vew, the logt specfcaton has the advantage of focusng on percentage rather than percentage pont changes n tranng. Ths advantage s consderable when the cross-country varaton n tranng rates s T large, as s the case n our data. Specfcaton [17] mples that = ( T( ρ) /[ 1 T( ρ) ]) λ2. Yet, f T s relatvely small (or has ρ relatvely lttle varaton) t s easy to show that the relatonshp between regulaton and tranng s close to be lnear. However, we consder also an alternatve lnear specfcaton of the model. 14

17 postulate that the error term s as follows ε = ξ + ξ + ξ + ω [18] ct c t ct ct where ξ c s a country by ndustry effect, ξ t s an ndustry by tme effect, ξ ct s a country by tme effect, and ω ct s a standard dsturbance. We control for these unobserved effects by ncludng n the regresson country by sector, country by year, and sector by year dummes. The country by sector dummes capture cross country dfferences n the structure of each ndustry, ncludng dfferences n the parameters σ, γ, β and A; the sector by year effects capture the tme varyng dfferences n trend growth between affected ndustres (treated group) and ndustres not affected drectly by deregulaton (control group); the country by year dummes absorb country-specfc macroeconomc effects, countrywde changes n polcy (notably tranng polcy and naton-wde regulaton, on whch we have no data - see below) as well as changes n the routng of the questonnare and/or the exact formulaton of the tranng queston. Our key nterest les n estmatng the coeffcent λ 2, whch measures the relatonshp between product market regulaton and tranng. A negatve estmate of λ 2 would confrm that a negatve relatonshp between regulaton and tranng prevals. Strctly speakng, however, our theoretcal model suggests that ths coeffcent s postve for low levels of regulaton and negatve for hgher values. We can therefore try to capture ths bell-shaped pattern by fttng a quadratc n our regulatory ndcators. Yet, snce we do not have data on naton-wde aspects of regulaton affectng all sectors, we wll have to do a number of arbtrary assumptons to estmate these quadratc specfcatons, and results obtaned from them must be vewed as a tentatve extenson. Notce that the model s expressed n terms of tranng stocks. However, we only have data on tranng flows (see next secton). However, recallng that the tranng flow τ s equal to T ( δ + g) /(1 + g), where δ s the tranng deprecaton rate and g s the steady state growth rate of the tranng stock, we can, re-specfy our emprcal model n terms of the tranng flow τ, by smply substtutng t for T n equaton [17]. Alternatvely, we can construct tranng stocks from tranng flows by followng a methodology smlar to Cont (2005) and Dearden et al. (2006). In partcular, we assume a common deprecaton rate ( δ = ) and a steady state rate of growth equal to the sample average 15

18 growth rate of tranng flows 13, and reconstruct ntal condtons under the assumpton that steady state growth occurs at the begnnng of the sample, whch mples τ 1 /( δ + g ) = T0, where τ 1 s the tranng flow n the frst perod and T 0 the tranng stock n the ntal perod. Fnally, tranng stocks after mssng years are constructed by assumng steady state growth n those years. Insofar as these data are reconstructed, we prefer to specfy our model n terms of flows and use tranng stock data only n a senstvty analyss. Specfcaton [17] assumes that, condtonal on the vector X, only product market regulaton varables can vary by country, year and sector, and that confoundng factors are fully accounted for by the combnaton of country by year, country by sector and sector by year dummes. Whle ths s plausble, we cannot rule out the possblty that varables measurng labour market nsttutons at the same level of detal as regulaton varables could affect tranng ncdence. If ths s so, falure to account for these effects could erroneously attrbute them to changes n product market regulaton. Therefore, we also experment wth specfcatons that augment the vector X wth avalable measures of labour market nsttutons. Data We use three man sources of data: a) an OECD database on tranng and other labour market varables; b) OECD regulatory ndcators for seven non-manufacturng ndustres (electrcty, gas, ar transport, road transport, ralways, post and telecommuncatons); c) sector-level nformaton on output, mports and labour and captal nputs avalable n the OECD STAN database and the companon 60- ndustry database of the Gronngen Growth and Development Centre. The OECD database on tranng s drawn from the EU Labour Force Surveys. It contans nformaton on tranng and a number labour market varables (namely age, gender, educaton, parttme/full-tme status, occupaton, ndustry, frm-sze, tenure, whether the contract s temporary or permanent, whether the actvty s n the country of resdence, partcpaton n tranng, type of tranng and tranng duraton) for employed workers of 23 European countres from 1992 to 2002 (wth many mssng values, correspondng to countres and years where questons on tranng were not admnstered or data on tranng are unrelable). Data have been collected n the second quarter of each 13 In the steady state flows and stocks grow at the same rate. In country/sector unts where a decrease of tranng flows s observed, the steady state growth rate s set to zero. 16

19 year (March n most countres). Quanttatve varables (such as tenure or frm sze) are dvded nto categores (see the appendx for more detals). As regards to age, we dspose of categores coverng fve-year ntervals. We then reconstruct an ordnal varable by applyng to all observatons n each category the mean age of the category. Data are sem-aggregated nsofar as organzed by cells. Each cell corresponds to a combnaton of categores. Avalable cells cover all non-empty combnatons wth one category for each varable. Populaton weghts are reported for each cell. Tranng data refer to partcpaton n any educaton or tranng course n the 4 weeks precedng the ntervew (1 week for France). Data on the type of tranng and ts length are often mssng. For ths reason, we do not use ths nformaton. In order to avod that ntal and close-to-retrement educaton confound the nformaton on workplace tranng, we lmt our analyss to full-tme employees wth at least 1 month of tenure, aged between 25 and 54 years, and workng n ther country of resdence. Data comparablty s better wthn cells belongng to the same country and year, snce the exact formulaton of the tranng queston as well as the routng of the questonnare may change from one country to another and from one year to another. Descrptve statstcs on tranng are avalable n the Appendx. We collapse our data on tranng and selected other labour market varables (educaton, age, gender and frm sze) at the level of sectors. Industres are avalable at the 2 dgt level of the NACE rev.1 classfcaton (that, at the 2 dgt level, corresponds to the ISIC rev. 3, wth extremely few exceptons) for most countres and years n manufacturng. In servces, they are avalable at a slghtly more dsaggregated level than 1 dgt of NACE rev. 1. However, snce only NACE 1970 s avalable for 1992 and for certan country-year pars, we are oblged to aggregate a few ndustres n order to construct our sector-level database. The fnal lst of ndustres s avalable n the Appendx. We have access to detaled OECD ndcators on ant-compettve product market regulaton for seven 2 or 3 dgt non-manufacturng ndustres. Data are avalable on an annual bass and span from 1975 to 2003 for 21 OECD countres (Australa, Austra, Belgum, Canada, Swtzerland, Germany, Denmark, Span, Fnland, France, UK, Greece, Ireland, Italy, Japan, the Netherlands, Norway, Portugal, Sweden, US, New Zealand). Followng Alesna et al. (2005) we use these data to construct tme-seres ndcators of regulatory barrers for three more aggregate ndustres (utltes, transport, and communcaton servces), for whch we have tranng data (see above). Detaled regulatory ndcators concern sector-specfc entry barrers, publc ownershp, the market share of the domnant player(s) when relevant (n the telephone, gas and ralroad sectors), vertcal ntegraton n network ndustres and prce controls when relevant (n the road freght ndustry). Outsde the scope of these ndcators are naton-wde aspects of regulatons applyng to all ndustres, such as admnstratve barrers to 17

20 entrepreneurshp (admnstratve barrers on start-ups, general features of the lcensng and permt system, etc...), snce these data are not avalable n tme-seres n the OECD database. Avalable ndcators vary between 0 and 6 from the least to the most regulated. Entry barrers cover legal lmtatons on the number of companes and rules on vertcal ntegraton of network ndustres. A value of 0 corresponds to free entry. By contrast, a value of 6 apples when entry s severely restrcted. Publc ownershp measures the share of equty owned by central or muncpal governments, and takes a value of 0 n the case of no equty and a value of 6 n the case of full ownershp. All other ndcators are smlarly defned. More detals on these ndcators are avalable from Ncolett et al. (2001), Alesna et al. (2005) and Conway and Ncolett(2006). We construct our aggregate ndcators followng the same methodology of Alesna et al. (2005). Ths nvolves two steps. Frst, separate ndcators of barrers to entry, publc ownershp, market structure, vertcal ntegraton and prce controls for each of the seven ndustres are averaged to obtan four coarser (and partally alternatve) ndcators: REGOL, the overall ndcator ncludng all the regulaton dmensons; BEVI, whch summarzes barrers to entry (comprsng legal restrctons and vertcal ntegraton); REGPO, whch summarzes the degree of publc ownershp; and REGNO, whch ncludes all dmensons except publc ownershp. As our model apples explctly to barrers to entry we expect BEVI to be more related to tranng than REGPO. The same mght apply to REGNO nsofar as the presence of prce controls and a very concentrated market can, to a certan extent, be seen as addtonal barrers to start-ups. Second, the same ndcators for the three more aggregated ndustres are obtaned by smple averagng the values of the correspondng sub-ndustres. Once the four ndcators of regulaton are matched to our tranng data we obtan 402 country by sector by tme non-mssng observatons 14 concernng three (typcally regulated) servce ndustres for 15 European countres and a maxmum of 11 years. Yet, these ndustres account for a small share of total employment (about 7.5% n 2002). Moreover, n the event that reforms n these three sectors have occurred almost smultaneously, the effect of the varaton n regulaton on tranng ncdence rsks to be swept away once country per year dummes are ncluded n the emprcal analyss. To crcumvent ths problem, we add manufacturng ndustres to our dataset. For these ndustres, regulaton concerns essentally admnstratve burdens, lmtatons to foregn drect nvestment as well as barrers to trade, at least n European countres. Now, only the last two barrers can be consdered to be sector-specfc. But for them, due to the comng nto acton of the Sngle 14 Ths number drop further to 398 n logt specfcatons such as [17] snce tranng ncdence s zero for some observatons. 18

21 Market Programme (SMP) n 1992, t can be assumed that ther tme profle s flat snce at least 1994 for the 12 countres that were EU members n 1992 (see e.g. Bottasso and Sembenell, 2001). 15 The same argument can be appled to Austra, Fnland, Norway and Sweden from 1995 (see e.g. Baldwn et al., 1996, and Gullstrand and Johansson, 2005). 16 Ths s equvalent to assume that snce 1994 for the majorty of countres and snce 1995 for a few countres - regulaton n manufacturng has been equal to an arbtrary constant, whch we control for wth country by ndustry dummes. In sum, we construct an extended dataset startng from the three non-manufacturng ndustres by addng manufacturng and by droppng pre-smp years. In practce, ths s equvalent to usng a dfference n dfference estmator, where manufacturng sectors are used as the control group and servces are the treated group. We end up wth a sample composed of 15 countres and 13 ndustres for a maxmum of 7 years for each country-ndustry par, and for a total of 1339 observatons once observatons wth mssng nformaton on tranng s excluded. 17 As dscussed above, deally we would lke to estmate a bell-shaped relatonshp between regulaton and tranng. However, strctly speakng ths s mpossble n our data, snce we do not have a tme-seres for those naton-wde aspects of regulaton that apply to all sectors (such as admnstratve barrers to start-ups). In fact these ndcators exst n the OECD Database only for two years (1998 and 2003). In specfcatons such as [17] ths component of regulaton s controlled for by country by tme dummes. However, ths s no longer the case n specfcatons wth a quadratc. On a tentatve bass we can add the naton-wde ndcator of admnstratve barrers to start-ups n 1998 to our sector-specfc ndcators of regulaton, assumng the latter to be 0 n manufacturng, to obtan a reconstructed ndcator of regulatory barrers that s more comparable across countres. Yet ths ndcator wll mss the effect of admnstratve reforms that took place n the 1990s. The other addtonal relevant co-varates are taken from the OECD STAN database and the Gronngen's 60-Industry database. Notably, one mght expect that the effects of deregulaton are fully realsed wth a certan delay and/or that, partcularly n certan manufacturng ndustres, globalsaton s ncreasng the compettve pressure on busnesses ndependently of regulaton. To account for ths 15 Bottasso and Sembenell (2001) report that, on average, 75% of the measures mpled by the SMP agreement were already transposed nto natonal legal systems at the tme when the SMP came nto acton, and that vrtually all measures were transposed shortly after. 16 Accordng to the frst edtons of the Sngle Market Scoreboard (EC, 1997, 1998), by 1997 Fnland, Norway and Sweden were among the best performng countres as far as transposton of EC drectves s concerned. Only Austra appeared to lag behnd, but ts gap wth other EU countres was closed n For ths reason we check that our results are robust to excluson of Austra pror to Ths number drop further to 1294 n logt specfcatons such as eq. [17] snce the tranng ncdence s zero for certan observatons. 19

22 possblty, we nclude n [17] the logarthm of mport penetraton. Furthermore, one can magne that growng busnesses wll have a greater propensty to tran than downszng busnesses, nsofar as n the former the proporton of new hres n need of nducton tranng s lkely to be greater, 18 whle n the latter the proporton of dsmssals, upon whch employers wll be unable to recoup the cost of tranng, s lkely to be greater. Therefore we nclude the logarthm of employment growth as a further control. Fnally, there are good reasons to thnk that tranng mght vary over the busness-cycle. For nstance, accordng to Hall (2000), re-organsatons take place durng slack perods when the cost of foregong producton to re-allocate resources s smaller. The case studed by Hall concerns creaton/destructon of job matches and search. However, t can well apply to nternal re-organzaton, whch usually requres long perods of adaptaton, learnng and tranng before becomng agan fully effcent (see also Jovanovc, 2006). In support of such a vew, Sepulveda (2002) fnds that on-the-job tranng s countercyclcal usng data from the US Natonal Longtudnal Survey of Youth. To control for sector-specfc busness cycles we construct log employment and log worked hours gaps by subtractng to each of these varables ther fltered tme-seres obtaned applyng an Hodrck-Prescott flter wth standard parameters. 19 When the mpact of naton-wde nsttutons s homogeneous across sectors, ths effect s controlled n equaton [17] by country per year dummes. However, certan labour market nsttutons mght not have the same mpact on tranng n all sectors. More specfcally, Haltwanger et al. (2006) and Mcco and Pages (2006), convncngly suggest that the mpact of lay-off regulatons on job turnover vares accordng to the natural propensty of ndustres to adjust ther labour nput. They show that almost all the varaton n the cross-country/cross-sector dstrbuton of job turnover can be explaned by the dstrbuton of job turnover n the Unted States (that s the OECD country wth the least regulated labour market) and country dummes, and that the remanng varaton can be explaned by an nteracton between a country-specfc ndcator of regulatory strngency and US job turnover rates by sector. Bassann and Venn (2007) use the same methodology to explore the mpact of lay-off regulatons on productvty growth. We use varous US job turnover ndcators by sector from Haltwanger et al. (2006) and nteract them wth the OECD aggregate measure of employment 18 Yet, gven that tranng s lkely to occur n the very few weeks after hrng and we exclude workers wth less than one month of tenure, we probably already control for part of ths effect. 19 We explot here the advantage of havng relable nformaton on employment and hours. In fact, one can expect the countercyclcal pattern of tranng to be more mportant when labour hoardng domnates employment adjustments, partcularly durng slowdowns. However, we check that our results are robust to substtutng a more classcal output gap for employment and hours gaps. In a senstvty analyss, we also consder the nvestment rate as well as the level and growth of productvty to control for possble endogenety of regulaton wth respect to productvty (see e.g. Duso and Röller, 2003). 20

23 protecton legslaton (EPL), so as to obtan an ndcator of EPL mpact that vares by country, sector and tme. Unon power mght also vary along these three dmensons. In order to capture ths effect we use data on unon densty from Ebbnghaus and Vsser (2000) and data on total days lost for strkes and ndustral actons from the ILO LABORSTA Database that we dvde by the number of employed workers usng data from the Gronngen's 60-Industry database. However, these data are avalable only for macro-sectors. Therefore we cannot do anythng better than attrbutng ther macro-sector averages to each sub-sectors. Exact varable defntons, sources and sample statstcs for the sub-sample of nonmanufacturng and the post-smp sample of non-manufacturng and manufacturng are provded n the Appendx. Snce the exact month of each regulatory reform s not known and mght well be subsequent to the second quarter of the correspondng year, each regulatory ndcator s lagged one year. The same apples to all other ndcators, whch do not come from the ELFS, nsofar as they generally refer to yearly averages. Emprcal results We start our analyss by examnng the assocaton between regulaton and tranng ncdence at the sector level. We estmate equaton [17] by OLS, wth the dependent varable expressed n terms of tranng flows. Table 1 shows, n three dfferent panels, our results for each regulaton measure when only (b-dmensonal) fxed effects are ncluded. In the frst panel, dfferent measures of regulaton are ncluded one at a tme. In the second and thrd panel, the barrers to entry ndcators REGNO and BEVI and the publc ownershp ndcator REGPO are ncluded together. In each panel, Column 1 refers to transport, communcatons and energy only, and Column 2 extends the sample to cover manufacturng but drops pre-smp years. The last two columns replcate results by omttng the few observatons wth very low tranng rates, for whch our dependent varable assumes very large negatve values. On the one hand, the assocaton between regulaton and tranng s not statstcally sgnfcant when we consder only publc ownershp (REGPO). Ths s also reflected n the assocaton of the overall ndex (REGOL) and tranng, whch becomes nsgnfcant once low-tranng observatons are excluded. On the other hand, both REGNO - the most comprehensve measure whch excludes publc ownershp - and BEVI, whch consders only strctly defned barrers to entry and vertcal ntegraton, attract a statstcally sgnfcant coeffcent and are negatvely correlated wth tranng ncdence. Condtonal on ether BEVI or REGNO, the publc ownershp ndcator REGPO does never 21

24 attract a statstcally sgnfcant coeffcent. Therefore, n the ensung dscusson we shall focus on the ndcator of barrers to entry and vertcal ntegraton (BEVI). We wll also drop low-tranng observatons. 20 Table 2 presents the results obtaned by addng to the specfcatons n Table 1 addtonal controls: the mean age of employees n the ndustry, the share of women n the ndustry, the shares of employees wth upper secondary educaton or less n the ndustry, the employment share of frms wth less than 10 employees n the ndustry, the logarthm of mport penetraton (whch s defned only for manufacturng ndustres), employment and hours gaps and employment growth. 21 Panel A presents the results for the servce plus energy sample, and Panel B for the extended non-smp sample, whch ncludes manufacturng. Table 1. Estmates of logt specfcatons of tranng as a functon of product market regulaton. Dependent varable: nverse logstc transformaton of tranng partcpaton rates. Panel A: One regulatory varable per model, OLS (1) (2) (3) (4) Servces + Energy only Servces, Energy and Manufacturng Servces + Energy only (no outlers) Servces, Energy and Manufacturng (no outlers) Overall ndex (REGOL) [0.061]** [0.065]* [0.061]** [0.065] Excludng publc ownershp (REGNO) [0.047]** [0.045]** [0.046]*** [0.045]** Barrers to entry only (BEVI) [0.037]** [0.036]** [0.035]*** [0.035]*** Publc ownershp only (REGPO) [0.046] [0.060] [0.045] [0.058] Country by sector dummes yes yes yes Yes Country by year dummes yes yes yes Yes Sector by year dummes yes yes yes Yes Number of observatons Results are qute smlar f we use REGNO and/or we keep low-tranng observatons. 21 We also consdered varous specfcatons ncludng the nvestment rate as well as the level and growth of productvty. Yet, these controls never turn out as statstcally sgnfcant and ther ncluson does not change the senstvty of tranng to regulaton (results avalable from authors upon request). 22

25 Panel B: Separatng out publc ownershp, OLS (1) (2) (3) (4) Servces + Energy only Servces, Energy and Manufacturng Servces + Energy only (no outlers) Servces, Energy and Manufacturng (no outlers) Excludng publc ownershp (REGNO) [0.051]** [0.046]** [0.049]*** [0.046]** Publc ownershp only (REGPO) [0.049] [0.059] [0.046] [0.057] Country by sector dummes yes yes yes Yes Country by year dummes yes yes yes Yes Sector by year dummes yes yes yes Yes Number of observatons R-squared Panel C: Barrers to entry and publc ownershp, OLS (1) (2) (3) (4) Servces + Energy only Servces, Energy and Manufacturng Servces + Energy only (no outlers) Servces, Energy and Manufacturng (no outlers) Barrers to entry only (BEVI) [0.041]** [0.036]** [0.038]*** [0.035]*** Publc ownershp only (REGPO) [0.049] [0.058] [0.045] [0.056] Country by sector dummes yes yes yes Yes Country by year dummes yes yes yes Yes Sector by year dummes yes yes yes Yes Number of observatons R-squared Note: robust standard errors wthn brackets. All regressons nclude country by year, sector by year and country by sector dummes. In Panel A each cell refers to a dfferent specfcaton. In Panel B each column refers to a dfferent specfcaton. Observatons wth tranng partcpaton rate smaller than 0.5% are excluded n Columns 3 and 4. **, ***: sgnfcant at the 5% and 1% level, respectvely. It turns out that tranng ncdence s sgnfcantly lower when the share of low educated ndvduals wth ISCED less than 3 s hgher, a standard result n the tranng lterature (see Bassann et al, 2007). Consstently wth the fndng of Sepulveda (2002) we also fnd that tranng ncdence s countercyclcal. However, the negatve relatonshp s not statstcally sgnfcant n the case of the employment gap. Employment growth appears to be negatvely assocated to tranng n the servces plus energy sample. Yet, ths relatonshp has the wrong sgn (theory would rather predct a postve relatonshp) and becomes nsgnfcant n the post-smp sample wth all ndustres. Fnally, the logarthm of mport penetraton s postvely related to tranng n a (weakly) sgnfcant way, yeldng some support to the dea that, for the tradeable sector, globalsaton can have an addtonal effect on 23

26 tranng ndependently of deregulaton. Table 2. Estmates of logt specfcatons of tranng as a functon of product market regulaton and addtonal controls. Dependent varable: nverse logstc transformaton of tranng partcpaton rates. Panel A: Servces + Energy only, OLS (1) (2) (3) (4) All controls All sgnfcant controls Sgnfcant controls, no employment growth (3) plus employment gap Barrers to entry only (BEVI) [0.040]*** [0.039]*** [0.039]*** [0.039]*** Share males [0.804] Average age [0.039] Share frms wth > 50 employees [0.524] Share low educaton [0.851]** [0.560]** [0.563]** [0.573]*** Share medum educaton [0.718] Hours gap [3.023]** [2.838]* [2.893]* [2.894]** Employment gap [0.698] [0.550] Employment growth [0.645]** [0.517]*** Country by sector dummes yes yes yes Yes Country by year dummes yes yes yes Yes Sector by year dummes yes yes yes Yes Number of observatons R-squared