Arntz (2014) Allocation Human Capital in Space

7/26/2019 Arntz (2014) Allocation Human Capital in Space

1/21

Full Terms & Conditions of access and use can be found athttp://www.tandfonline.com/action/journalInformation?journalCode=cres20

Download by:[Andre M. Marques] Date:16 December 2015, At: 02:44

Regional Studies

ISSN: 0034-3404 (Print) 1360-0591 (Online) Journal homepage: http://www.tandfonline.com/loi/cres20

Can Regional Employment Disparities Explain theAllocation of Human Capital Across Space?

Melanie Arntz, Terry Gregory & Florian Lehmer

To cite this article:Melanie Arntz, Terry Gregory & Florian Lehmer (2014) Can Regional

Employment Disparities Explain the Allocation of Human Capital Across Space?, RegionalStudies, 48:10, 1719-1738, DOI: 10.1080/00343404.2014.882500

To link to this article: http://dx.doi.org/10.1080/00343404.2014.882500

Published online: 10 Mar 2014.

Submit your article to this journal

Article views: 368

View related articles

View Crossmark data

Citing articles: 1 View citing articles
http://www.tandfonline.com/doi/citedby/10.1080/00343404.2014.882500#tabModulehttp://www.tandfonline.com/doi/citedby/10.1080/00343404.2014.882500#tabModulehttp://crossmark.crossref.org/dialog/?doi=10.1080/00343404.2014.882500&domain=pdf&date_stamp=2014-03-10http://crossmark.crossref.org/dialog/?doi=10.1080/00343404.2014.882500&domain=pdf&date_stamp=2014-03-10http://www.tandfonline.com/doi/mlt/10.1080/00343404.2014.882500http://www.tandfonline.com/doi/mlt/10.1080/00343404.2014.882500http://www.tandfonline.com/action/authorSubmission?journalCode=cres20&page=instructionshttp://www.tandfonline.com/action/authorSubmission?journalCode=cres20&page=instructionshttp://dx.doi.org/10.1080/00343404.2014.882500http://www.tandfonline.com/action/showCitFormats?doi=10.1080/00343404.2014.882500http://www.tandfonline.com/loi/cres20http://www.tandfonline.com/action/journalInformation?journalCode=cres20


2/21

Can Regional Employment Disparities Explain the

Allocation of Human Capital Across Space?MELANIE ARNTZ*, TERRY GREGORY*and FLORIAN LEHMER

*Labour Markets, Human Resources and Social Policy, ZEW Centre for European Economic Research, Mannheim, Germany.Email:[email protected]

Department of Economics, University of Heidelberg, Heidelberg, GermanyOccupational Labour Markets, IAB Institute for Employment Research, Nuremberg, Germany

(Received July 2013: in revised form December 2013)

ARNTZM., GREGORY T. and LEHMER F. Can regional employment disparities explain the allocation of human capital acrossspace,Regional Studies. This paper examines the determinants of skill-selective regional migration in a context where modellingthe migration decision as a wage-maximizing process may be insufcient due to persistent employment disparities. Based on aBorjas-type framework it is shown that high-skilled workers are disproportionately attracted to regions with higher meanwages and employment chances as well as higher regional wage and employment inequalities. Estimates from a labour owxed-effects model and a general methods of moments (GMM) estimator show that these predictions hold, but only employmentdisparities induce a robust and signicant skill sorting. The paper thus establishes a missing link about why employment disparitiesmay actually be self-reinforcing.

Gross migration Migration selectivity Wage inequality Employment inequality Regional disparities

ARNTZM., GREGORYT. and LEHMERF.Borjas

(GMM)

ARNTZM., GREGORYT.etLEHMERF.Lesdisparitsenmatiredemploirgional,expliquent-t-elles la distributiondu capital humaintraverslespace, Regional Studies.Dansuncadreolamodlisationdeladcisiondemigrercommeunprocessusquimaximiselessalairespourrait savrer insufsante cause de profondes disparits en matire demploi, cet article examine les dterminants de la migrationslective en fonction des comptences requises. partir dun cadre du type Borjas, on montre que les travailleurs hautement qualissont attirs de faon disproportionne vers les rgions o les salaires moyens sont plus levs et sont dotes de meilleures possibilitsdemplois, ainsi que de plus grandes ingalits dans les salaires et lemploi rgionaux. Des estimations provenant dun modle effetsxes des ux de main-doeuvre et un estimateur de la mthode des moments gnralise montrent que ces prvisions sont valables,

mais ce sont seulement les disparits en matire d emploi qui entranent un important triage robuste des comptences. Ainsi, l articletablit un chanon manquant concernant les raisons pour lesquelles les disparits en matire demploi pourraient en effet se renforcer.

Migration brute Slectivit de la migration Ingalit des salaires Ingalit en matire demploi Disparits rgionales

ARNTZM., GREGORYT.und LEHMER F. Knnen regionale Ungleichheitenim Beschftigungsniveau die rumliche Allokation vonHumankapital erklren?,Regional Studies. In diesem Beitrag untersuchen die Autoren die Determinanten bildungsselektiver Arbeit-skrftemigration in einem Kontext rigider Lhne und ausgeprgten Beschftigungsunterschieden. In einer solchen Situation ist dieModellierung der Migrationsentscheidung als reines Lohn-Maximierungsproblem unzureichend. Im Rahmen von Schtzungeneines erweiterten Borjas-Modells zeigen die Autoren, dass Regionen eine umso qualiziertere Zuwanderung erfahren, je hhersowohl das regionale Lohn- und Beschftigungsniveau als auch die Lohn- und Beschftigungsungleichheit sind. Robustheitsanalysenmit Fixen Effektenfr jeden Bruttomigrationsstrom und Schtzungenmit der verallgemeinerten Momentenmethode (GMM) bestti-gen die Ergebnisse, zeigen sichjedoch nur frdie Beschftigungsunterschiede robust und signikant. Die Resultate legen nahe, dassdierumliche Allokation vonHumankapitalsomit strkerberdie Beschftigungsseitedeterminiert wird. Dieser Beitrag liefert somit einen

wichtigen Beitrag zu der Frage warum sich regionale Ungleichheiten mglicherweise selbst verstrken knnen.

Bruttomigration Selektive Migration Lohnungleichheit Beschftigungsungleichheit Regionale Disparitten

Regional Studies, 2014

Vol. 48, No. 10, 17191738, http://dx.doi.org/10.1080/00343404.2014.882500

2014 Regional Studies Associationhttp://www.regionalstudies.org
mailto:[email protected]:[email protected]


3/21

ARNTZM., GREGORYT. y LEHMERF. Pueden las desigualdades laborales regionales explicar la distribucin espacial del capitalhumano?,Regional Studies. En este artculo analizamos los determinantes de la migracin regional seleccionada por cualicacionesen un contexto donde modelar la decisin de emigrar como un proceso para maximizar salarios podra ser insuciente debido alas persistentes desigualdades de empleo. Basndonos en un esquema tipo Borjas, mostramos que los trabajadores altamentecualicados son atrados desproporcionadamente a regiones con salarios medios y oportunidades laborales ms altos as comodesigualdades ms altas de salarios y empleo. Segn los clculos de un modelo de efectos jos de ujos laborales y el clculo

del mtodo generalizado de momentos (GMM) mostramos que estas predicciones se cumplen, pero solamente las desigualdadesde empleo llevan a organizar las habilidades de una forma slida y signicativa. De esta manera encontramos el eslabn perdidoque explica por qu las desigualdades laborales podran en realidad autopotenciarse.

Migracin bruta Selectividad migratoria Desigualdad salarial Desigualdad laboral Desigualdades regionales

JEL classications: J31, J61, R23

INTRODUCTION

In contrast to the United States, employment ratherthan wages responds to labour demand shocks in

Europe (BLANCHARD and KAT Z, 1992; ABRAHAM,1996; MERTENS, 2002). For Germany, NIEBUHRet al. (2012) show that the average wage level barelyvaries across regions and remained nearly unchangedacross a period of economic shocks, while the unem-ployment rate turned out to be much more volatileacross regions. Similarly, employment disparities haveincreased in the aftermath of the recent nancial crisiswith regional unemployment rates for EuropeanNUTS-3 regions ranging between 0.6% and 33.2% in2010 (EUROSTAT, 2013). Due to low interregionalmobility rates, such disparities tend to be quite persistent

(DECRESSINand FATS,1995; BADDELEYet al.,2000).Even worse, such disparities may be even self-reinfor-cing if high employment regions attract a predominantlyhigh-skilled workforce since the local concentration ofhuman capital may trigger a process of cumulative cau-sation due to skill complementarities and local skillexternalities (ROMER, 1986; LUCAS, 1988; KANBURand RAPOPORT,2005; FRATESIand RIGGI,2007).

Yet, the only existing theoretical framework thatwas proposed by BORJAS et al. (1992) links selectivemigration to wage disparities only. In particular,high-skilled workers ceteris paribus should be attractedto regions that best reward their abilities by payinghigh wage returns to their skills as reected in a highwage inequality. Whereas BORJAS et al. (1992) andHUN T and MUELLER (2004) demonstrate the rel-evance of a wage-based selection mechanism forinternal migration in the United States, correspondingevidence in the European context such as Germany hasbeen surprisingly weak (BRCKER and TRBSWET-TE R, 2007; ARNTZ, 2010).1 In a context with strongemployment disparities and strong employmentresponses to economic shocks, this framework maynot sufce to explain skill selective migration. In par-ticular, it is argued here that employment rates are

increasing in worker ability and may therefore beseen as a return to investments in skills as also theoreti-cally argued by HELPMAN et al.(2010).

For this reason, this paper suggests that regionalemployment disparities may be a missing link toexplain skill-selective migration. In particular, it

extends the Borjas framework to allow for a selectionmechanism based on both wage and employmentdifferentials and shows that the average skill level of amigration ow is a positive function of the wage andemployment inequality in the destination as comparedwith the origin region. Moreover, unlike the Borjasframework, the model suggests that mean wage andemployment differentials also induce a positive skillsorting. As a second contribution, the paper teststhese predictions for the average skill level of grosslabour ows between 27 German regions. For thispurpose, it makes use of administrative data which

cover nearly 80% of the German workforce and deter-mine the skill content of each ow with regard toobservable skills. The paper then regresses this skillmeasure on the mean and the dispersion of the regionalwage and employment distribution. Instead of onlyconditioning on the regional unemployment rate as isdone by Pissarides and MCMASTER (1990), PARIKHand LEUVENSTEIJN (2003), ETZ O (2011) and others,this paper thus captures not only the average chances,but also allows regions to differ in how these chancesare spread among the local workforce. As a third con-tribution, the panel dimension of the data is exploitedin order to condition average time-constant utility

differentials between regions (e.g. amenity differentials)that may otherwise bias the estimation results. In orderto control for the endogeneity, the model is also esti-mated with the difference generalized method ofmoments (GMM) estimator proposed by ARELLANOand BON D (1991).

The ndings conrm the relevance of regionalemployment disparities for skill-selective migration,while regional wage differentials have no robust and sig-nicant impact. This paper thus lls an important gap inthe understanding of the self-reinforcing nature of inter-regional employment disparities. Although the focus is

on interregional migration, the main ndings shouldapply to cross-country migration as well, althougheffects may be weaker due to higher migration costs.

1720 Melanie Arntzet al.


4/21

Still, the ndings suggest that the recent divergence ofintra-European migration ows from SouthernEurope towards high-employment countries such asGermany that has been found by BERTOLI et al.(2013) is likely skill biased, thus potentially aggravating

the current NorthSouth divide.The paper is structured as follows. The secondsection presents an extended theoretical framework forthe skill composition of migrants. The third sectionintroduces the database, while the fourth section pre-sents descriptive evidence on the proposed selectionmechanism. The fth section describes the estimationstrategy and presents the ndings that are then subjectto additional robustness checks in the sixth section.The seventh section concludes.

THEORETICAL FRAMEWORKThe theoretical framework builds upon the work ofBORJAS et al.(1992), who formalize the self-selectionof interstate migrants and test their model in the UScontext.2 While they focus on the selectivity ofinternal migrants with respect to both observableskills and unobservable abilities, the analysis isrestricted to observable skills since the aim is tounderstand the spatial allocation of observable skillssuch as formal education, experience, but also otherrelevant skills as proxied by occupations which areconsidered the most important for the skill endow-

ment of a region.Let Y be a continuous random variable of a workersobservable skills in the population with mean zero; and yits realizations. High-skilled workers rank in the upperpart of the skill distribution with y. 0 and low-skilled workers rank in the lower part of the skill distri-bution withy, 0. The distribution is considered to beregion-invariant, i.e. it is assumed that observable skillsare perfectly transferable between all regions, anassumption considered to be justied for internalmigration.3

Consider two regions jand k that differ with respectto their (labour) income distribution. As a consequence,migration decisions do not depend on differentials inregional wage distributions alone but also hinge onthe probability of receiving this wage, i.e. the prob-ability of being employed as has already been discussedby TODARO (1969) and FIELDS (1976). For ease ofexposition, the theoretical framework abstracts fromother utility differentials between regions such asregional amenities or disamenities including regionalprice differentials as well as from migration costs.4 Anincome-maximizing individual chooses to live inregion j if the expected income in region j exceedsthe income in regionk, i.e.:

wj(y)ej(y) . wk(y)ek(y) (1)

where ejis the individuals chance of being employedin region jon any particular workday; and wj is thewage paid in this region if employed. Note thatemployment chances ej are not only capturing aninitial job nding chance in region j, but also should

be thought of as measuring the expected probabilitythat an individual is employed on any workdaygiven the region-specic chances of nding, keepingand losing a job.5 Also note that both wages andemployment chances are considered to depend onskills.

In particular, the wage distribution can be decom-posed into a part reecting the mean wage wthat isindependent of an individuals observable skills and apart that measures skill-specic deviations from thismean wage that depend on the returns to skills paidin region j. Following BORJAS et al.(1992), the popu-lation wage distribution in region jcan then be written

as:

wj(y) = mwj + swjy with mwj,swj. 0 (2)

Hence, an individuals potential wage is determined bytheir position in the skill distribution and the region-specic returns to these skills swj. Note that even forthe least skilled individual, the wage is still positive,thus implying mwj . swjyto hold for every y. An ana-logous decomposition of the population employmentdistribution is further assumed, which can be writtenas:

ej(y) = mej + sejy with mej,sej. 0 (3)

where mej is the average probability of employmenton any workday; and y is dened as above. Hence,an individuals employment probability is determinedby the average employment probability and theregion-specic returns to their skill level in terms ofemployment sej. Employment chances are thusassumed to be increasing in worker skills, since low-

skilled individuals are more likely to be atypicallyemployed (GIESECKE,2009), face a higher probabilityof becoming unemployed during an adverse regionalshock (MAURO and SPILIMBERGO, 1999), and aremore prone to repeated and prolonged unemploymentperiods (JUH N et al., 1991; WILKE, 2005; RIDDELLand SON G, 2011). In contrast, high-skilled workersare less affected from an economic shock as they aremore likely to be hoarded during economic downturnsand may even become more demanded (NICKELL andBEL L,1995; MORRISON,2005). Hence, just as regionswith high wage inequality, regions with a highinequality in employment chances penalize low-skilled workers and reward high-skilled workers. Theemployment dispersion thus measures how the

Can Regional Employment Disparities Explain the Allocation of Human Capital Across Space? 1721


5/21

employment chances are spread across the workforce.Since all individuals of the labour force are assumedto have positive employment chances, it must alsohold that mej. sejy for all relevant values of y fromthe interval [ymin,ymax].

If the two decompositions in equations (2) and (3) areapplied, the income in region jcan be written as:

pj(y) = (mwj + swjy) (mej + sejy)

= mwjmej+ (mejswj+ mwjsej+ swjsejy)y

: = Mj+ Rj(y)

(4)

where the rst termMjcorresponds to the income of anaverage individual with y= 0 in region j; and thesecondterm Rj(y) reects all region-specic returns to

skills.6 This second income component is an increasingfunction of yand induces a sorting of individuals intoregions that best reward their skills. In particular, theincome differential Dp between region j and kdepends on the parameters of the wage and employ-ment distribution in both regions and on individualskills. It can be written as:

Dp(y) = pj(y) pk(y)

= Mj Mk + Rj(y) Rk(y)

: = DMjk + DRjk(y)

(5)

Let Skjdenote the average skill level of the migrantsmoving from region k to j, i.e. the skill transferbetween both regions. This skill transfer depends onthe relative income differentials for workers of differ-ent skill levels, i.e. if the income differential increasesfor high-skilled workers relative to the income differ-ential for low-skilled workers, the ow of migrantswould be expected to be more skilled on average,i.e.:

Skj=f(Dy) =f(Dp|y.0 Dp|y,0) with

f(Dy) . 0 (6)

Note that the skill content of the labour ow is onlyclaimed to be a monotonously increasing function ofthe income differential, but the exact functional link isnot specied. For the derivation of partial effects, itnow sufces to examine how changes in the employ-ment and wage distribution in one region relative tothe other affect the income differential Dy between

high- and low-skilled workers. Before doing so, Dy issimplied for the case that high-skilled workers have askill level of yH= y|y.0 and low-skilled workers have

a skill level ofyL= y|y,0:

Dy= DMjk + DRjk(yH) (DMjk + DRjk(yL))

= DRjk(yH) DRjk(yL)

= (Dmejksw

jk

+ Dmwjkse

jk

)(yH yL)

+ Dswjksejky2H Dswjksejky

2L

= (Dmejkswjk + Dmwjksejk )(yH yL)

+ Dswjksejk (yH+ yL)(yH yL)

(7)

whereyH yL. 0 and yH+ yL_ 0.For the partial effects of increasing employment and

wage levels in regionj:

Dy

mwj= sej(yH yL) . 0,

Dy

mej= swj(yH yL) . 0

(8)

According to equation (8) the partial effect of an increas-ing wage differential (mwj) on migration selectivity (Dy)depends on how employment chances are distributedamong individuals in the destination region (sej).Phrased differently, the wage-based selection mechan-ism depends on the skill-specic chances of nding a

job in the destination region in the rst place. An

increasing employment dispersion thereby reectsincreasing employment chances for high-skilledworkers while at the same time posing a penalty forlow-skilled workers from lacking these skills. As aresult, the migration ow from k to jshould becomemore skilled on average. The same argument can beapplied to an increase in the mean employment chances.

The partial effects for an increasing employment andwage inequality in region jrelative to region kcan bewritten as:

Dy

swj= [mej + sej(yH+ yL)](yH yL) . 0,

Dy

sej= [mwj+ swj(yH+ yL)](yH yL) . 0

(9)

Note that the sign of these effects is always positivebecause the values of (yH+ yL) fall in the interval[ymin,ymax] for which it must hold that mej . sejyandmwj. swjy. Hence, an increasing wage inequalityrewards high-skilled workers relative to low-skilledworkers to the extent that the destination regionrewards them relative to low-skilled workers in termsof better employment chances. The strength of this

effect, however, depends on the sign of (yH+ yL).One way to interpret this nding is to think of(yH+ yL) as the total skill endowment of the home



6/21

region. If the region is skilled relative to the national skilldistribution, i.e. (yH+ yL) . 0, an increasing wageinequality in the destination region will induce a stron-ger increase in the skill level of the migration ow than ifthe region had a below average skill endowment. The

same arguments apply to the effect of an increasingemployment inequality. The skill content of thelabour ow from region k to j should consequentlybecome more skilled on average if either wage inequal-ity or employment inequality increases.

Several issues warrant a short discussion. First of all,whereas the theoretical model predicts both wage andemployment differentials to inuence the selectivity ofmigration ows, in practice, employment is expectedto be more relevant in a German setting with regionalwage rigidities. In particular, collective wage agree-ments in Germany are typically met by industry-specictrade unions and employer associations. Although such

agreements are made at the regional level, the agree-ments obtained in the south-western part of Germanyfor the metal industry provide an important benchmarkfor the rest of the economy resulting in regional wagerigidities (OCHEL, 2005). In fact, MERTENS (2002)nds wages to be rigid across regions compared withthe United States. According to SCHBand WILDASIN(2007) one reason for such regional wage rigidities maybe the low level of regional mobility in Europe as com-pared with the United States. Their theoretical modelimplies that regional labour demand shocks result instrong regional wage disparities if mobility costs are

high. This gives rise to wage-bargaining regimes thatintroduce regional wage rigidities in order to mitigatewage risks that are associated with regional labourdemand shocks. Given the theoretical and empiricalevidence in favour of regional wage rigidities inGermany, employment rather than wage differentialsis thus expected to drive skill-selective migration inGermany.

Secondly, the empirical analysis deviates from thetheoretical model in that ows are examined in amulti-region context. Still, the mechanisms derivedbased on the two-region model should be applicableto this broader context. Moreover, it may be helpfulto discuss the reasons why ows in opposing directionsare expected to exist. This may be the case becausethere are new cohorts entering the labour market,each period among which a certain share is likely tobe mismatched to its origin region in terms of theirskills. Thus, while the most able individuals will leavetheir region for regions with higher returns to theirskills, low-skilled individuals may prefer the oppositedirection in order to minimize the penalty fromlacking such skills. Moreover, individuals for whom aparticular region once offered the optimal return totheir skills need not be optimally matched forever if

individuals shift their position in the skills distributiondue to training effects or due to the depreciation ofskills.

Finally, the model abstracts from a number of poten-tial complications such as regional amenity differentials,price differentials7 as well as from migration costs. Aslong as these components are not correlated to theskill level, the key results remain unchanged.

However, there are reasons to believe that migrationcosts decrease in abilities if abilities facilitate the gather-ing of information and reduce the psychological costs ofmigration. If this is the case, the key results of the theor-etical model remain unchanged only conditional onsuch costs. Similarly, if individuals differ in how theyvalue certain regional amenities and disamenitiesdepending on their skill level, as has been argued byBERRY and GLAESER (2005), the key results alsoremain intact only conditional on regional differentialsin amenities and disamenities. Finally, a recent studyby MORETTI(2013) suggests that real wage differentialsmay vary across skill groups due to skill-specic differ-

ences in the cost of living. The present estimationapproach thus needs to take account of these complicat-ing factors.

DATA

The paper uses employment register data (BeH) of theGerman Federal Employment Agency, an administra-tive dataset that contains information on the populationworking in jobs that are subject to social insurance pay-ments, thus excluding civil servants and self-employed

individuals. The data allow individual employment his-tories to be reconstructed, including periods of employ-ment and of unemployment benet receipt on a dailybasis. For each employment period the data containindividual and rm-level characteristics including dailygross wage, educational attainment as well as micro-census region of the workplace. Thus gross labourows are identied by comparing workplaces beforeand after an interregional job transition.

The sample is restricted to the time period between1995 and 2004 because the labour ows betweenEastern and Western Germany are severely underesti-mated in the years right after reunication due to thefact that many individuals did not show up in the databefore taking up employment in Western Germany.From the mid-1990s onwards the observed labourows correspond to migration patterns that are ofciallyreported by the Federal Statistical Ofce. This is in linewith other studies that nd population and labourmigration to yield similar results in migration modelsusing German data (LEUVENSTEIJN and PARIKH,2002). The authors are therefore condent that thelabourows can be considered to proxy quite well forpopulation ows. Furthermore, a focus is made onmen between the ages of 16 and 65 years because

women exhibit a lower labour force attachment thanmen and move for different reasons since they areoften tied to the migration decisions of their spouses.8



7/21

For all subsequent analyses, the paper distinguishesbetween 27 aggregated planning districts. Theseregions lump together 97 German planning districts(Raumordnungsregionen) that are dened according tocommuting ranges and already comprise labour

market regions that are relatively self-contained. Inorder to ensure a sufcient number of job movesbetween each region for different skill levels, these plan-ning districts had to be aggregated into 27 larger regions.This was done based on an algorithm that minimizes theremaining external commuting linkages subject tomerging only up to four adjacent regions, therebyensuring that the regional division yields relativelyequally sized and self-contained labour markets.9 Foreach year between 1995 and 2004, the employmentand wage distribution were estimated for the 27regions as well as the size and composition of the 702gross labourows between these regions. The following

subsections discuss the corresponding details.

Data on interregional labourows

For the computation of interregional labour ows,information on almost the entire working populationis exploited, i.e. the full employment register data(BeH) is used. Yearly cross-sections to the cut-off dateof 30 June are thereby used and the workplace locationbetween two consecutive years is compared. The grosslabour ows were therefore calculated by identifyingthe origin and the destination region for all interregional

job moves. Note that the identication of an interregio-nal job move necessitates an individual to be employedon 30 June of two consecutive years. Hence, long-termunemployed persons are underrepresented in the databut one should be aware that the sample includes indi-viduals who have been unemployed between these twocut-off days. In total, almost 137 million individualswere observed between 1995 and 2004, of which 3.6million (2.6%) experienced an interregional job movebetween two consecutive years.

Based on these data, the average skill level of eachgross labour ow was worked out by calculating analternative skill measure based on ranking individualsin the predicted wage distribution, as has been proposedby BORJAS et al.(1992).10 The underlying idea is thatwages reect the marginal product of labour and maythus proxy for abilities and skills. The quality of the cor-responding measure, however, depends on the skillcharacteristics used in the wage regression. Recentstudies point towards the relevance of particular typesof skills that are used in certain occupations. In particu-lar, FLORIDAet al. (2012) and BACOLOD et al.(2009)show that it is primarily analytical skills that receivean urban wage premium. It is therefore important tond proxies for these different skills. More precisely,

individual is daily gross wage11

in region jand year tover the time period 19942004 is estimated with thefollowing xed-effects (FE) model for all individuals

in the sample:

log wijt= b0 + b1REGIONj+ b2YEARt

+ b3Xijt + ci+ uijt (10)

where cicaptures the individual-specic, time-constantunobserved effect; and uijtcorresponds to a remainingidiosyncratic error term. Wage is a function of avector of dummy variables indicating the workplacelocation (REGION), a vector of dummy variables indi-cating the year of the observation (YEAR), and a vectorof individual- and time-specic observable skill charac-teristics (X) described above. Not only are age, agesquared and educational attainment therefore includedto proxy for experience and formal education, but alsoan individuals occupation (12 categories), industry

sector (28 categories) and the establishment size includ-ing its squared term. The idea is that the skill mix appliedby an individual is well proxied by the occupationalstatus and industry afliation. In addition,establishmentsize might also affect the applied skill mix.12 In addition,part-time employment is controlled for because hourlywages are not observed but only daily wages that maydiffer between full- and part-time employees due todifferent working hours.

The wages for all workers in the sample are then pre-dicted based on the vector of observable skill character-istics (X) only. In this way the skill measure does notreect differences in predicted wages due to region-and year-specic factors. A region- and time-invariantskill distribution is thus constructed. The skill contentSkjtof each labourow is then measured by calculatingthe average predicted log wage for theNmovers of eachlabourow:

Skjt=1

N

Ni=1

log wikjt (11)

Note that equation (11) is the empirical operationaliza-tion of the average skill level of migrants following aparticular migration ow in equation (6). In order tocompare movers and stayers, the average predicted logwage for the stayers in the sending region is alsocalculated.

Note that a focus is made on observable skills only,although the unobserved skills of migrants could alsobe calculated by estimating the time-constant unob-served effect ci in equation (10), as proposed byBORJAS et al. (1992). However, one problem withthis approach is that unobservable skills and theirregion-specic returns are not really separable. Sincemotivation and the like are remunerated differently

across regions, cidiffers depending on the regions inwhich an individual is observed. Put differently, aregion-invariant distribution of unobservable skills



8/21

cannot really be constructed. For this reason, it wasdecided to stick to observable skills only.

Table 1 reports descriptives on the number ofmigrants and the average predicted daily gross wage asa proxy for the average skill level for the 7020 gross

labour

ows across the ten-year period. On average,506 migrants with a predicted daily wage of 79.40follow a particular migration path in any year and arethus a positive selection with regard to observableskills compared with immobile workers whose averagepredicted daily wage of 67.60 is shown in thebottom panel. Also note that the variation in themigrants average skill level across ows and time islarge. This suggests both a substantial spatial reallocationof human capital at a particular point in time as well as aresponsiveness of the skill content of a particularmigration path to changing (economic) conditions.Finally, note that the number of migrants is low for

some rather distant origindestination pairs. In orderto check whether ows with only few migrantsproduce outliers that dominate the estimates, sensitivitytests were run by excluding labour ows with fewerthan 50 migrants (7.96% of all ows).

Regional wage and employment distributions

In order to test the theoretical predictions presented inthe second section, the means and standard deviations(SDs) of the wage and employment distribution needto be estimated for each region and year.

For the construction of the regional wage distri-bution, the wages of the regional workforce that result

from separate region- and year-specic ordinary leastsquares (OLS) wage regressions analogous to equation(10) are predicted. By estimating this model separatelyacross years and regions, varying returns to observableskill characteristics across years and regions are allowed

for. The same covariates as in equation (10) are usedsince the aim is to measure the regional differences inthe returns to the characteristics that also reect theskill measure that is used for the labour ows. Themean and the SD of the predicted wage distributionare then calculated for each year and region.13

For the regional employment distribution, it is notimmediately clear which measure should be chosen.One might think about using the probability of receiv-ing a job in a particular region, i.e. the job-ndingchances. However, for the expected income, theexpected probability of being employed in the regionon any particular future workday given the region-

specic risk of losing a job, being long-term unem-ployed and nding employment again should be ofcrucial concern. Individuals are not expected to haverational expectations on this complex probability.Hence, the relevant employment conditions areproxied by the number of days that someone can cur-rently expect to be employed in a particular region.This is done based on a 2% random sample of theemployment register data since full-spell informationon periods of employment is needed. As with wages,the predicted employment distribution of the regionalworkforce is then constructed. However, since the

number of days employed during a year come withmass points at 0 and 365 employed days, one needs to

Table1. Summary statistics for gross labourows and employees staying in the sending region, 19952004

Variable Mean SD Minimum Maximum Number of observations

Gross labour ows between k = 1, . . . ,27 sending and j= 1, . . . ,26 receiving regionsAverage number of migrants

Overall 506 827 5 11955 K J T = 7020Between 805 15 8700 K J = 702Within 193 1710 5554 T = 10

Average predicted daily gross wagea

Overall 79.4 8.1 49.5 132.3 K J T = 7020Between 4.9 65.2 94.7 K J = 702Within 6.5 50.7 120.5 T = 10

Immobile employees in the sending regions k = 1, . . . ,27Average number of stayers (in 1000)

Overall 494 211 150 1112 K T = 270Between 213 170 1050 K = 27Within 24 368 634 T = 10

Average predicted daily gross wagea

Overall 67.6 4.9 56.1 82.1 K T = 270Between 3.5 62.4 75.5 K = 27Within 3.5 60.3 75.7 T = 10

Note: aAverage gross daily are calculated as 1

N

Ni=1

wikjt

.



9/21

take account of this unusual distribution by modellingthe different cases separately. For this, let Iijt= 0, 1, 2denote an individual-specic indicator function thatdepends on the number of days dijtthat an individual iis employed during a particular yeartin region j:

Iijt=

2 if dijt= 3651 if 0 , dijt, 3650 if dijt= 0

Individual is observed number of employed daysdepends on the probability of being employed all year(Iijt= 2), employed between 0 and 365 days (Iijt= 1),and being unemployed all year (Iijt= 0). According tothe law of total probability, the conditional number ofdays employed in regionjat timetcan be written as:

E[dijt|Xit] =P(Iijt= 1|Xit)E[dijt|Iijt= 1,Xit]+ P(Iijt= 2|Xit)365]

(12)

The conditional probabilities P(Iijt= 0|Xit),P(Iijt= 1|Xit) and P(Iijt= 2|Xit) are estimated for eachregion and year by predicting conditional probabilitieswithin a multinomial logit framework. The condition-ing set is the same as in equation (10) except for estab-lishment size, which is not available in the 2% randomsample of the dataset. The expected number of daysemployed conditional on being employed between 0and 365 days, E[dijt|Iijt= 1,Xit], is estimated running

separate region- and year-specic OLS regressions. Justas with wages, the mean and SD of the predictedemployment distribution are then calculated for eachregion and year. When comparing the ofcial unem-ployment rate across the ten-year period to the shareof days not employed that is implied by the employmentmeasure, very similar patterns were found, conrmingthat the measure captures a meaningful concept.14

Note that despite the censoring, there are enoughworkers employed between 0 and 365 days, so thatregions should differ in the SD of employed days, con-ditional on average employment.

Fig. 1shows the average parameters of the employ-ment and wage distribution across the ten-year period(absolute changes are shown in Fig. A1in AppendixA). For better interpretation the mean and SD of theexponentiated predicted log wage are used. Theexpected EastWest divide is mainly found, withaverage wages and average employment in WesternGermany clearly exceeding levels in Eastern Germany.However, some disparities are also found betweenSouthern and Northern Germany, with the latterbeing in a less favourable labour market situation. More-over, the absolute wage dispersion in Eastern Germanyis below the wage dispersion in Western Germany.

Relative to mean wages, though, wage inequality isquite comparable between both parts of the country,as has also been suggested by BURDA and HUN T

(2001) and GERNANDT and PFEIFFER (2009) for theperiod after 1995. In contrast, the employment dis-persion in Eastern Germany strongly exceeds theemployment dispersion in Western Germany in bothabsolute and relative terms. Thus, the risk of being

unemployed is not only higher on average in EasternGermany, but also it is distributed more unequallyamong the local workforce.

Note that the regional indicators are highly corre-lated, posing a challenge for the identication of themodel. However, since 27 regions are distinguished,there is still a lot of variation in the level of labourows. Moreover, the time variation in each of theows between the 27 regions is exploited. Thus,while the EastWest divide dominates the picture in across-section perspective, the later estimation approachexploits the variation in the skill content of eachlabour ow across the ten-year period and relates this

to changes in interregional disparities. In the subsequentanalysis, these refer to the difference between the receiv-ing and sending region in the standardized wage andemployment parameters and are denoted as Dmw, Dme,Dsw and Dse. For the standardization of the regionalparameters we subtract the mean and divide by theSD which are calculated across regions and years. Anincrease of 1 unit in Dmw, for instance, thus correspondsto a mean wage 1 SD higher in the receiving relative tothe sending region. Note that a simultaneity bias is miti-gated by measuring ows between 30 June of two con-secutive years, while the wage distribution relates to 30

June prior to observing the ows. Since information isneeded for an entire year for the employment distri-bution, the corresponding parameters are estimated forthe year prior to observing the destination state in thenext year so that there is only a limited overlapbetween the timing ofows and the estimation of theregional employment distribution.

DESCRIPTIVES

If the predictions of the theoretical framework hold,interregional differences in the mean and dispersion ofwages and employment would be expected to increasethe relative return to migration for high-skilled relativeto low-skilled workers. As a consequence, these returnswould be expected to increase when ranking labourows according to their skill content. In order toprovide some descriptive evidence on this predictedrelationship, the observable skill distribution Skjtis rststandardized and the average standardized skill level iscalculated for each gross labour ow. All gross labourows between and within Eastern and WesternGermany are then ranked according to their averagestandardized observable skill level and create quintiles

of this distribution. The lowest quintile among theEastWest ows, for instance, corresponds to the 20%of all EastWest ows with the lowest skill level,



10/21

while the fth quintile captures the 20% of all ows withthe highest skill level. The paper distinguishes betweenows within and between Eastern and WesternGermany to examine whether the relationship isdriven by the huge EastWest disparities only or

whether it holds within each part of the country as well.Table 2 shows the interregional returns to skills

measures for the period 19952004 by these quintiles

of the labour ows. First, note that the averagemover even of the least skilled ows in Table 2 hasan above-average skill level, conrming once againthat movers are a positive selection with regard toobservable skills. More importantly, it is found that

within all ow directions most of the parametersrespect the expected ranking across the quintiles.Regions offering higher wage and employment

Fig.1. Parameters of the regional wage and employment distribution at the level of 27 aggregated planning regions



11/21

returns to high-skilled workers attract labour owswith a higher average skill level. The interregional dis-parities in employment dispersion, however, aredecreasing in skills when the ranking of the meanemployment is increasing and vice versa. One expla-nation may be that regions with high average employ-ment rates tend to have a lower employmentdispersion so that Dmp and Dsp are strongly negativelycorrelated with r= 0.51. Multivariate analyses willthus have to be run in order to disentangle theeffects. Still, the descriptive evidence tends toconrm that high-skilled migrants move to regionswith relatively higher skill premiums in terms of bothwages and employment.

Also note that differences are found in the averageskill levels across quintiles of different ow directionswith ows within Eastern Germany being least skilled.Relating such differences to the interregional differencesinFig. 1, however, may be misleading since the averageskill level should also be affected by other factors thanregional differentials in wages and employment suchas, for example, amenity differentials. For this reason,

note that the aim is not to explain fully the observedskill composition, but to test whether changes in theskill composition ofows are related to changes in the

interregional differences in employment and wages astheoretically predicted.

Hence, a better descriptive test is to look at howchanges in employment and wage differentials acrosstime are related to changes in the skill composition forany particular labour ow. Of course, such an analysisis not feasible for the 702 available ows. As anexample, a focus is therefore made on the owbetween Eastern and Western Germany, which is ofparticular interest given the strong interregional differ-ences that still persist after reunication. Fig. 2 showsthe corresponding (labour) income differential calcu-lated based on equation (4) for an individual withaverage (y= 0), above average (y= 1) and belowaverage (y= 1) skills. In 1995, an average individualearns 10000 per year more in Western than inEastern Germany. This income differential increases to16000 in 2001 reecting deteriorating averageemployment chances in Eastern relative to WesternGermany while wage differentials remain rather con-stant. In 2001, the increase in the income differentialcame to a halt before stagnating from then on. In light

of this development, standard migration models on netmigration suggest an increasing net loss of migrants inEastern Germany until 2001.

Table2. Average interregional disparities by skill quintile of the labourows within and between Eastern and Western Germany,19952004

Observations Interregional standardized values

(1) (2) (3) (4) (5) (6) (7)Quintile of

observable skillsdistribution

Number ofows

Number of

movers (in1000)

Average

standardized skilllevel

Averagewage Dmw

Average

employmentDme

Wage

dispersionDsw

Employmentdispersion Dse

EastWest Flows

1 251 81 0.11 1.97 2.07 0.19 1.672 252 101 0.25 1.97 1.92 0.16 1.633 252 92 0.32 2.13 1.92 0.02 1.714 252 94 0.38 2.23 1.96 0.22 1.825 253 60 0.49 2.37 2.01 0.40 1.91

WestEast Flows

1 251 55 0.06 2.12 2.13 0.23 1.722 252 69 0.20 2.09 2.12 0.03 1.833 252 69 0.27 2.07 2.04 0.07 1.844 252 60 0.35 2.18 1.87 0.09 1.705 253 45 0.48 2.20 1.72 0.01 1.66

WestWest Flows

1 839 615 0.11 0.07 0.08 0.17 0.102 840 607 0.24 0.07 0.07 0.12 0.083 840 566 0.30 0.02 0.01 0.02 0.014 840 451 0.37 0.05 0.04 0.10 0.055 841 257 0.47 0.11 0.10 0.20 0.11

EastEast Flows

1 59 49 0.10 0.05 0.01 0.12 0.012 60 65 0.01 0.03 0.03 0.05 0.033 60 84 0.07 0.03 0.05 0.03 0.044 60 78 0.13 0.10 0.04 0.21 0.015 61 54 0.21 0.14 0.06 0.24 0.01

Total 7020 3551



12/21

The spread between the upper and the lower line inFig. 2reects the skill premium for high-skilled (y= 1)relative to low-skilled individuals (y= 1) whichaccording to the theoretical framework should drivethe skill composition between both parts of thecountry (equation 6). The increasing gap between theincome differentials of high- and low-skilled workersuntil 2001 reects the increasing skill premium inWestern relative to Eastern Germany that is mainlyrelated to the deteriorating average employment

chances in Eastern relative to Western Germany and,in addition, to an increasing average wage and wage dis-persion in Western as compared with Eastern Germany(see alsoFig. A1in Appendix A). Note that the increas-ing employment dispersion in Eastern compared withWestern Germany obviously did not close the spreadbetween high- and low-skilled workers (cf. Fig. 2).Hence, according to theory, an increasing net transferof observable skills from Eastern to Western Germanysimilar to net migration should be observed.

Indeed,Fig. 3shows that the net loss of migrants aswell as the net loss of skills in Eastern Germany increaseduntil 2001 and thus conrms the above derived predic-tions. In particular, in 2001 about 30 000 more migrantsmoved from Eastern to Western Germany than viceversa. Moreover, the average EastWest migrantbecame more skilled relative to the average WestEastmigrant, thus aggravating the already existing braindrain with regard to observable skills. While this devel-opment is consistent with the development of incomedifferentials until 2001, the somewhat decreasing netloss in migrants and skills after 2001 cannot be fullyexplained by the income differentials after 2001.

The descriptive ndings are in line with the fewexisting studies by BURDA and HUN T (2001), Hunt

(2006), and BRCKER and TRBSWETTER (2007)that consider the selectivity of EastWest migrationows. In particular, BURDA and HUN T (2001) and

HUN T (2006) study the years between reunicationand the millennium and nd that EastWest migrantstend to be young and better educated compared withstayers. The authors, however, do not explicitly studythe forces driving the skill composition of EastWest migration ows. In contrast, BRCKER andTRBSWETTER (2007) show for the period 19941997 that EastWest migrants also constitute a positiveselection based on unobservable characteristics.

EMPIRICAL ANALYSIS

In order to identify the determinants of the average skilllevel of gross migration ows, the variation in theaverage observable skill level across 7020 gross labourows that were observed during the period between1995 and 2004 was exploited. The panel is balanced,i.e. for all 702 region pairs there are ten-year obser-vations. Since unobserved effects in the error termsuch as amenity and price differentials are likely to becorrelated with regional wages and employment, asimple OLS regression should be biased. Therefore,the following labour-ow xed-effects (FE) panel

model is estimated:

Skjt= b0 + b1Dmwt+ b2Dmet+ b3Dswt

+ b4Dset+ ckj+ ukjt (13)

whereSkjtis to the average observable skill of a migrantmoving from region k to j in a particular yeart= 1, . . . , 10 with k= j, as shown in equation (11).The right-hand side of equation (13) contains the inter-regional differences in the returns to skills, namely thedifferences in wages and employment between the des-

tination regionjand the region of originkas dened inthe third section. The composite error consists of theow xed effect ckj as well as an idiosyncratic error

Fig.2. (Labour) income differentialDp(y) = pwest(y) peast(y) for an average-skilled

(y= 0), high-skilled (y= 1) and low-skilled (y= 1)individual, 19952004

Fig.3. Net migration and net skill transfer in EasternGermany, 19952004



13/21


14/21

(2) and (3) with the observed net skill loss in EasternGermany (Fig. 3). The predictions of the income-based selection model perform much better inexplaining the observed net skill loss in EasternGermany compared with the purely wage-basedmodel. The test underlines the importance of extendingthe framework on skill selective migration to allow forregional employment disparities.

Of course, looking at the impact of regional wageand employment differentials on the skill compositiondoes not indicate much about which migrants drive

the results. As an example, one could achieve thesame theoretically proposed effects with high-skilledindividuals being attracted to regions with a highreturn in wages or employment, by low-skilled individ-uals being distracted from these regions, or a combi-nation of both. From the predictions made in thesecond section, individuals would be expected to be

increasingly attracted to regions with higher meanwages and mean employment the higher their skilllevel. In addition, individuals with below-averageskills would be expected to be distracted from regionswith a high wage and employment dispersion, while

individuals with above-average skills should be attractedto these regions.In order to test these predictions, the quintile of the

observable skill distribution for each worker in thesample is determined and respective migration ratesfor each quintile of the 702 ows across the ten-yearperiod are computed, i.e. the absolute number ofmigrants fromk to jby quintile of the skill distributionare calculated and these numbers are divided by thesize of the labour force in k of the correspondingskill quintile.16 The rst quintile therefore containsthe rate by which individuals ranking in the lowestquintile of the skill distribution follow a particular

migration path.The results, shown inTable 4, can be interpreted as

follows: an increase in the mean employment differen-tial by 1 SD increases the migration rate of the rst quin-tile (20% less skilled migrants) by 18.4%. As expected,the effects of mean wage and employment differentialstend to be positive for all quintiles, although an increas-ing and signicant pattern across quintiles can be foundfor the mean employment differential only. Thus, therelated positive selection in Table 3 seems to be theresult of migratory responses along the entire skill distri-bution. In contrast, mean wage differentials signicantly

affect migration choices for the best skilled individualsonly. Since this group is more likely to earn a wageindependent of any centrally bargained tariff, wagerigidities may be less pronounced for this group. Inthat case, mean wage differentials are likely to reectregional differences in income opportunities for thegroup of high-skilled individuals only, and so it

Fig.4. Prediction of EastWest migration selectivity based onthe wage-based selection model (3) versus the extended income-

based selection model (2) inTable 3, 19952004

Table4. Log migration rates by quintile of the skill distribution, labour-owxed-effects estimation, 19952004. Dependentvariable: log migration rate by quintile of the skill distribution

1. Quintile 2. Quintile 3. Quintile 4. Quintile 5. Quintile

Mean wage (Dmw) 0.130 0.055 0.079 0.100*

0.210***

(1.50) (0.78) (1.23) (1.66) (3.30)Mean employment (Dme) 0.165*** 0.216*** 0.272*** 0.312*** 0.321***

(9.85) (14.15) (18.68) (22.01) (20.59)Wage dispersion (Dsw) 0.021 0.052* 0.024 0.068*** 0.033

(0.64) (1.91) (0.97) (2.99) (1.39)Employment dispersion (Dse) 0.096*** 0.132*** 0.158*** 0.157*** 0.205***

(2.84) (5.07) (6.33) (7.29) (9.87)Constant 8.360*** 7.814*** 7.544*** 7.312*** 7.420***

(553.65) (679.28) (726.19) (774.30) (726.33)

N 7020 7020 7020 7020 7020F 51 72 118 155 136R2 0.105 0.151 0.238 0.321 0.345AdjustedR2 0.104 0.149 0.237 0.320 0.344

Root mean square error (RMSE) 0.360 0.284 0.250 0.215 0.221Note:Robust t-statistics are given in parentheses; signicance levels: *10%, **5% and ***1%; all models include dummies for the origin anddestination region.



15/21

appears plausible that signicant effects are found for thefth quintile only.

Regarding the effect of the wage dispersion, no con-sistent pattern across all ve quintiles is found, butinstead a negative signicant effect for the second and

a positive signi

cant impact for the fourth quintile isfound that are both in line with the theoretical predic-tions. In contrast, the employment dispersion has a sig-nicantly positive and increasing impact on migrationalong the entire skill distribution. The related positiveselection effect in Table 3 thus results from theseincreasing patterns rather than individuals with below-average skills avoiding regions with a higher employ-ment inequality. While this may seem at odds withthe theoretical predictions, this observation is theresult of two empirical facts.

First, the empirical distribution of the number ofemployed days turned out to be bimodal with the

majority of individuals being almost always employed.Thus, although the risk of unemployment is spreadunequally across individuals, only a rather small shareof individuals appears to be affected by a positiveunemployment risk. If these individuals were theleast skilled as postulated in the theoretical framework,however, a negative sign would be expected for thelowest quintile only as well as insignicant effects forthe remaining quintiles. As an alternative explanation,unobservable skills play a major role. Due to the rel-evance of unobservable skills in determining an indi-viduals employment chances, even the individual

with the highest observable skill level has a non-zerochance of unemployment, while even the individualwith the lowest observable skill level may be continu-ously employed. Therefore, an increasing employmentinequality may be benecial for the majority of indi-viduals with favourable unobservable characteristicswithin each skill quintile, yet the share of those forwhom an increasing employment inequality increasesthe risk of unemployment decreases across the quintilesof the observable skill distribution. As a result, a con-tinuously increasing pattern of positive coefcientsacross quintiles is obtained in Table 4. This patternthus reveals that the theoretical framework needs tobe further extended to selective migration withregard to unobservable skills in order to capture itsfull complexity.

ROBUSTNESS

The main problem with the previous labourow xed-effect model is that it assumes the independent variablesto be uncorrelated with past and possibly current realiz-ations of the error term. However, migration could beboth the cause and the consequence of regional wages

and employment. In fact, several studies explicitlyexamine the effect of selective migration on the devel-opment of regional wages and employment (BURDA

and WYP LO SZ , 1992; FRATESI and RIGGI, 2007).Although the timing of the dependent and independentvariables rules out a direct simultaneity between interre-gional differences and the skill composition of labourows, the described dynamics may still bias the previous

labour

ow

xed-effects estimations.In order to deal with endogenous regressorsandreversecausality, the average skill level of the gross labourows isestimated with the rst-difference GMM estimator, asproposed by ARELLANOand BON D(1991). The estima-tor is designed for panel data, where the number of timeperiods,T, is small and the number of observations,N, islarge. The underlying idea is to instrument the endogen-ous variables in the differenced equation using thelagged versions of the endogenous variables. As Arellanoand Bond note, lagged variables dated t2 and earliercan potentially be orthogonal to the error and thereforeact as valid instruments.17

In addition, not only is this difference GMM testedfor the average skill level of the gross ows, but alsothis test is run for the average skill level relative to theskill level of the sending region. Such a test would beunnecessary if, as assumed by the theoretical framework,there was a skill distribution such that the position of anindividual within this distribution is region-invariant.However, this need not be the case, especiallybetween Eastern and Western Germany where edu-cational systems were separate until 1990. In fact, differ-ences are found in the average skill level of the regionalpopulation, especially between Eastern and Western

Germany. While some of these differences may be theresult of previous skill-selective migration, such differ-ences may also indicate that the skill distribution is notregion-invariant.

Table 5reports the difference GMM results for boththe average skill level and the average relative skill levelof a migrant. Thus, in the latter case migrants may nowbe high-skilled relative to the source populationdespite being low-skilled compared with other ows.Still, the sorting of skills across space would beexpected to follow the same predicted pattern asbefore. In all models, all available lags of the dependentvariable and of the four returns to skills variables datedt 6 and earlier as instruments for the transformedequation are used.18 In addition, models (2) and (4)add lags of the independent variables in order toallow for the possibility that there is a time lagbetween regional differentials and migratory responses.According toTable 5, the orthogonality restrictions ofthe instruments and the estimated residuals areaccepted in all models by the SarganHansen test. Asa test for autocorrelation, the ArellanoBond test isconducted on the residuals in differences. The AR(2)test rejects the hypothesis of autocorrelation ofsecond order, which argues against a dynamic model

with lags of the dependent variable. In fact, includinglags of the dependent variable turned out to be insig-nicant in all models.



16/21

Model (1) shows that the results remain quite robustwhen taking into account the potential endogeneity of

the regressors, therefore con

rming the results inTable3. The size of the coefcient for the mean employmentdifferential is, however, slightly higher compared withthe basic ow xed-effects model inTable 3. The samepattern holds when using the relative skill measure incolumn (3). When adding lags of the independent vari-ables, the signicant effect of mean employment disap-pears, while the contemporaneous employmentdispersion continues to attract better skilled migrants onaverage. Note, however, that all lags are insignicant,thus suggesting that lagged responses to regionalincome differentials are not relevant such that models(1) and (3) remain the preferred specications.

Overall, the robustness checks conrm the previousndings. Interestingly, it seems to be more importantto take into account time-constant interregional differ-ences that seem to bias cross-sectional estimationsstrongly than taking care of the potential endogeneityof the regressors due to reversed causality.

CONCLUSION

This paper has examined the factors driving the skillselectivity of internal migration by proposing a frame-

work for skill selective migration that takes account ofthe returns to skills in terms of both wages and employ-ment. In a European context with strong and persistent

employment disparities and with employment ratherthan wages responding to shocks, regional employment

rather than wage disparities might be the driver of skill-selective migration. The income-based model of skill-selective migration predicts the average skill level of amigrant to be a positive function not only of regionaldifferentials in wage inequality as suggested by thewage-based framework proposed by BORJAS et al.(1992), but also of differentials in mean wages andmean employment rates as well as in employmentinequality, i.e. differences in how employmentchances are spread across the workforce.

Being able to exploit the variation in the skill compo-sition of 702 gross labourows across a ten-year periodin Germany, these predictions are tested based on a

labour-ow xed-effects model that takes account oftime-constant, ow-specic unobservables such asamenity differentials. This way, the model identiesthe effects of interregional differences in the wage andemployment distribution on migration selectivitybeyond other relevant time-constant regional dispar-ities. Comparing the outcomes to pooled OLS indicatesthat controlling for time-constant unobservables on theow level is important in order to prevent biased esti-mates. This puts doubt on the reliability of previouscross-sectional estimates.

The ndings suggest that regional employment

differentials turn out to be important. In particular, aregion attracts an increasingly skilled inow of migrantsthe higher its average employment. The same is true for

Table5. Generalized method of moments (GMM) estimation of the average and relative average skill level of gross migrationows.Dependent variable: average observable skill of a migrant (Skjt)

(1) (2) (3) (4)Average skills Average skills Relative average skills Relative average skills

Mean wage (Dmw) 0.186 0.115 0.195 0.179

(1.52) (0.68) (1.60) (1.05)Mean employment (Dme) 0.094** 0.066 0.090** 0.035

(2.16) (1.05) (2.11) (0.56)Wage dispersion (Dsw) 0.034 0.044 0.030 0.050

(1.24) (1.39) (1.10) (1.59)Employment dispersion (Dse) 0.031** 0.030** 0.022* 0.025*

(2.45) (2.05) (1.72) (1.71)L. Mean wage (Dmw) 0.012 0.061

(0.09) (0.44)L. Mean employment (Dme) 0.001 0.027

(0.04) (0.88)L. Wage dispersion (Dsw) 0.023 0.015

(0.91) (0.60)L. Employment dispersion (Dse) 0.021 0.017

(1.43) (1.19)

N 6318 5616 6318 5616Sargan (p-value) 0.160 0.102 0.264 0.189Hansen (p-value) 0.291 0.260 0.437 0.436AR1 (p-value) 0.000 0.000 0.000 0.000AR2 (p-value) 0.355 0.464 0.226 0.477Number of instruments 39 38 39 38

Note:Robustt-statistics are given in parentheses; signicance levels: *10%, **5% and ***1%; GMM estimations are one-step estimates.



17/21

an increasing employment inequality. The moreunequal employment is spread across the regional work-force, the more a region attracts an increasingly skilledinow of migrants. In contrast, regional differentials inthe wage distribution exert no signicant effect on the

skill composition of labour

ows. In a context whereemployment rather than wages tend to respond toregional shocks and resulting employment disparitiesare quite persistent, the allocation of human capitalacross space is thus driven by employment rather thanby wage disparities. This paper thus establishes amissing link in understanding the self-reinforcingnature of interregional employment disparities. Usingthe ow between Eastern and Western Germany as anexample, the paper demonstrates that employment dis-parities have a much better predictive power for theobserved skill composition between both parts of thecountry than wage disparities. However, wage dispar-

ities in response to regional shocks might gain in impor-tance since institutions that give rise to regional wagerigidities in Germany such as collective wage agreementshave started to erode (OCHEL,2005).

From a policy perspective, the results indicate thatattempts to control migration ows in order toprevent an extensive brain drain should not focus onwage policies alone. In fact, attempts to speed upwage convergence articially, as has been the case inEastern Germany in the years following reunication,are likely to increase unemployment rates, therebyagain fostering an increased brain drain. The average

EastWest migrant, for example, became more skilledrelative to the average WestEast migrant, thus aggra-vating the already existing brain drain with regard toobservable skills.

The implications also apply to European regionsthat are particularly characterized by increasingemployment disparities. As noted by PUG A (2002),regional employment inequalities in Europe areincreasingly driven by disparities within rather thanbetween countries. During the recent recession,cross-country disparities have also been increasing dra-matically. As a result, intra-European labourows havealready been diverted towards prospering high employ-ment countries such as Germany (BERTOLI et al.,2013). Given the insights of this paper, such owsare likely skill biased with a brain drain from Southernto Northern European regions, thus potentially aggra-vating the current NorthSouth divide. Therefore,Europe needs to aim at reducing regional employmentdisparities in order to mitigate related polarization ten-dencies among its member states.

Finally, note that the predictions regarding skillselective migration might become more complexthan suggested by the framework. In particular,routine and codiable jobs typically ranking in themiddle of the wage distribution have recently been

found to experience lower wage and employmentgrowth than the tails of the wage distribution (fora recent overview, see AUTOR,2013). Moreover,AUTOR and DOR N (2013) show that a correspondingwage and employment polarization mainly takes placein local labour markets that initially specialized inroutine tasks. As a consequence, interregional utilitydifferentials may increase for both tails of the wagedistribution at the same time suggesting that aregion may be increasingly attractive for both low-and high-skilled workers, a prediction that is incom-patible with a Borjas-type framework of skillselectivity. However, this prediction only holds if

polarization patterns are really observed. ForGermany, SENFTLEBEN and WIELANDT (2012)show that wage and employment gains are mainlylimited to the upper tail of the wage distribution,which would predict increasing high-skilled migrationcompatible with the theoretical model. Hence, therelevance of recent polarization tendencies for skillselective migration is far from resolved and a promis-ing route for future research.

AcknowledgementsThe authors thank the scholars at:the Annual Congress of the German Economic Association

(VfS), the 51st Annual Congress of the Regional ScienceAssociation International (RSAI), the 25th Annual Confer-ence of the European Society for Population Economics(ESPE), the 4th Summer Conference in Regional Scienceof the German-speaking section of the RSAI, the 3rdWorld Conference on Economic Geography, theNORFACE Migration Network Conference, the 3rdCentral European Conference in Regional Science (CERS),the 3rd Workshop on Geographical Localisation, IntersectoralReallocation of Labour and Unemployment Differentials(GRUNLAB), the 6th international workshop onFlexibilityin Heterogeneous Labour Marketsof the German ResearchFoundation (DFG), as well as seminar participants at ZEW

Mannheim, Heidelberg University, Regensburg Universityand especially Stephan Dlugosz for helpful comments. Allremaining errors are the authorssole responsibility.

Funding The authors gratefully acknowledge nancialsupport by the German Research Foundation (DFG)through research project Potentials for the Flexibization ofRegional Labour Markets. An Empirical Study on the Con-sequences of Regional Mobility.



18/21

APPENDIX A

Fig.A1. Absolute changes of the regional wage and employment distribution at the level of 27 aggregated planningregions between 1995 and 2004



19/21

NOTES

1. The Borjas hypothesis has also been tested for inter-national migration. In particular, BORJAS (1991,1987)provides empirical evidence in favour of the BorjasSelf-Selection Model, assuming migration to be mostly

driven by interregional wage differentials. The resultshave been challenged by CHISWICK (1999) andCHIQUIAR and HANSON (2005) who suggest thatmigration costs are likely to be inversely related to earn-ings, thus leading to different outcomes for immigrantselectivity. For a recent overview and discussion of the lit-erature, see Bodvarsson and van den BER G(2013).

2. Their framework is linked to the self-selection ofworkers, as described by ROY (1951), and its extensionto the self-selection of immigrants, as developed byBORJAS(1987).

3. While this assumption may be unproblematic withinWestern and Eastern Germany, it is less clear whether

the assumption can be applied to migration across theformer German border. Some sensitivity analyses willthus be run in the sixth section.

4. The present empirical approach controls for time-con-stant regional differentials and, thus, takes account ofmuch of these factors.

5. Although it might seem unlikely that individuals takeaccount of such abstract measures in their utility maximi-zation, it is argued that individuals gather such infor-mation when looking for employment and better jobsacross regions. Since individuals mainly move afternding a better job match, migration depends on thechances of getting attractive job offers from a particularregion, a probability that can be explicitly modelled in

the context of a search model (for an example, seeDAM M and ROSHOLM, 2010). Moreover, searchmodels allow for a simultaneous search across differentlabour markets and are thus able to explain moves toregions that within a utility framework are suboptimal.Applying a search model would thus be an interestingextension. It was still decided to stick to the simplerutility-maximizing framework because aggregate owsare modelled for which the utility-maximizing frame-work gives comparable predictions with regard towages and employment.

6. Note that, analogous to BORJAS et al. (1992), relativeprices of all skills are assumed to be region invariant so

that one does not have to operate with a multifactormodel of ability.7. In Germany, individual taxes are mainly invariant across

regions apart from a minor tax for church afliation.The framework thus abstracts from taxes.

8. Men attending military or civilian service are furtherexcluded since they are centrally registered so that theidentication of their exact location is not possible, andapprentices and all employment spells with minor employ-ment are neglected since its denition changed in 1999.

9. Details on the algorithm are available from the authorsupon request.

10. Ideally, individuals would be ranked in the income distri-bution. However, the income distribution for the full

BeH data could not be estimated because the data arereduced to a cross-section that lacks information on theprevious employment history. Extending the data to

include the full employment history is impossible dueto the resulting size of the data.

11. Unfortunately, around 15% of all wages are top-coded atthe contribution limit of the social security. Therefore,the censored wages are imputed with an estimation pro-cedure described by GARTNER (2005). This procedure

adds a randomly drawn error term to the predictedwage level and, thereby, avoids a strong correlationbetween the error term and the explanatory variables.

12. Ideally, the skills used by each individual would beobserved. On the other hand, including occupation andindustry dummies in addition to establishment sizeshould allow the skill content to be captured quite pre-cisely although the types of skills that are rewarded inthe labour market could not be explicitly pinpointed, ashas been done by BACOLOD et al. (2009) and FLORIDAet al.(2012).

13. The selection of high-skilled individuals into labourmarkets that best reward their skills as is predicted by

the theoretical framework may give rise to an upwardbias in the returns to skills, as has been shown by DAH L(2002). For this reason, Dahl used bias-corrected returnsto skills in an estimation of skill-selective migration.Despite the upward bias in the returns to skills,however, estimation results for the migration modelwith uncorrected and corrected returns to skills yieldedvery similar results, which presumably reects their highpositive correlation. No attempt was made to correctthe estimated returns, especially since transferring themethodology proposed by Dahl is not straightforwardin a context where polychotomous choices were repeatedacross a ten-year period.

14. For the denition of unemployment one needs to deal

with gaps in the employment record, whenever an indi-vidual is out of the labour force, self-employed, a civilservant or unemployed without any receipt of unem-ployment transfers. For this, following Fitzenberger andWILKE (2010), non-employment periods can becounted only as unemployment if there has been atleast one initial receipt of unemployment benets.

15. Equation (11) was also estimated using absolute migrationas a dependent variable in order to test whether the speci-cation replicates migration patterns that have been foundin the literature. As expected, it was found that increasingmean wages andmean employment chances in the receiv-ing relative to the sending region raises gross migration

levels. Moreover, consistent with similar studies oninternal migration, employment differentials have a stron-ger impact than wage differentials (PUHANI, 2001;EDERVEEN et al., 2007; MCCORMICK, 1997). Forstudies on the particular German case, see PARIKH andLEUVENSTEIJN(2003) and DECRESSIN (1994).

16. Since the log migration rate is used for the estimations, a 1is added to the numerator in order to avoid missing for

year-ow observations with zero migrants.17. The model is also estimated with the system GMM esti-

mator proposed by BLUNDELLand BOND(1998). Sincethe SarganHansen test was rejected in most of the systemGMM estimations, the results for the difference GMMestimator are only presented.

18. The SarganHansen test on the joint validity of theinstruments failed to pass the test for lags dated prior tot6.



20/21

REFERENCES

ABRAHAMF. (1996) Regional adjustment and wage exibility in the European Union,Regional Science and Urban Economics26(1),5175.

ARELLANOM. and BONDS. (1991) Some tests of specication for panel data: Monte Carlo evidence and an application to employ-ment equations,Review of Economic Studies 58(2), 277297.

ARNTZM. (2010) What attracts human capital? Understanding the skill composition of interregional job matches in Germany,Regional Studies44(4), 423441.AUTORD. (2013) The task approachto labor markets: an overview, Journal for Labour Market Research 46, 185199.AUTORD. H. and DORND. (2013) The growth of low-skill service jobs and the polarization of the US labor market, American

Economic Review103(5), 15531597.BACOLODM., BLUMB. S. and STRANGEW. C. (2009) Skills in the city, Journal of Urban Economics 65(2), 136153.BADDELEYM. C., MARTINR. L. and TYLERP. (2000) Regional wage rigidity: the European Union and United States compared,

Journal of Regional Science40(1), 115141.BERRYC. R. and GLAESER E. L. (2005) The divergence of human capital levels across cities, Papers in Regional Science 84(3),

407444.BERTOLIS., BRCKERH. and FERNNDEZ-HUERTASMORAGAJ. (2013)The European Crisis and Migration to Germany: Expectations

and the Diversion of Migration Flows. Discussion Paper No. 7170. Institute for the Study of Labor (IZA), Bonn.BLANCHARDO. J. and KATZL. F. (1992) Regional evolutions, Brookings Papers on Economic Activity 23(1), 176.BLUNDELLR. and BONDS. (1998) Initial conditions and moment restrictions in dynamic panel data models, Journal of Econometrics

87(1), 115143.BODVARSSON. B. and VAN DENBERGH. (2013) Who immigrates? Theory and evidence, in The Economics of Immigration: Theory

and Policy, pp. 79106. Springer, New York, NY.BORJASG. J. (1987) Self-selection and the earnings of immigrants, American Economic Review77(4), 53153.BORJASG. J. (1991) Immigration and self-selection, in ABOWD J. M. and FREEMAN R. B. (Eds)Immigration, Trade and the Labor

Market, pp. 2976. National Bureau of Economic Research (NBER), Cambridge, MA.BORJASG. J., BRONARSS. G. and TREJOS. J. (1992) Self-selection and internal migration in the United States, Journal of Urban

Economics 32(2), 159185.BRCKERH. and TRBSWETTERP. (2007) Do the best go west? An analysis of the self-selection of employed EastWest migrants in

Germany,Empirica 34(4), 371395.BURDAM. C. and HUNT J. (2001) From reunication to economic integration: productivity and the labor market in eastern

Germany,Brookings Papers on Economic Activity 32(2), 192.BURDAM. and WYPLOSZC. (1992) Human capital, investment and migration in an integrated Europe, European Economic Review

36(2

3), 677684.CHIQUIAR D. and HANSONG. H. (2005) International migration, self-selection, and the distribution of wages: evidence fromMexico and the United States, Journal of Political Economy 113(2), 239280.

CHISWICKB. R. (1999) Are Immigrants Favorably Self-Selected?Discussion Paper No. 131. Institute for the Study of Labor (IZA),Bonn.

DAHL G. B. (2002) Mobility and the return to education: testing a Roy model with multiple markets, Econometrica 70(6),23672420.

DAMMA. P. and ROSHOLMM. (2010) Employment effects of spatial dispersal of refugees, Review of Economics of the Household8(1),105146.

DECRESSINJ. W. (1994) Internal migration in West Germany and implications for EastWest salary convergence,Review of WorldEconomics 130(2), 231256.

DECRESSINJ. and FATSA. (1995) Regional labor market dynamics in Europe, European Economic Review39(9), 16271655.EDERVEENS. NAHUISR. and PARIKHA. (2007) Labour mobility and regional disparities: the role of female labour participation,

Journal of Population Economics 20(4), 895913.

ETZOI. (2011) The determinants of the recent interregional migration ows in Italy: a panel data analysis,Journal of Regional Science51(0), 948966.

EUROSTAT (2013)Eurostat Regional Yearbook 2012. Publications Ofce of the European Union, Luxembourg.FIELDSG. S. (1976) Labor force migration, unemployment and job turnover, Review of Economics and Statistics 58(4), 407415.FITZENBERGERB. and WILKER. A. (2010) Unemployment durations in West Germany before and after the reform of the unem-

ployment compensation system during the 1980s, German Economic Review11(3), 336366.FLORIDAR., MELLANDER C., STOLARICK K. and ROSS A. (2012) Cities, skills and wages, Journal of Economic Geography 12(2),

355377.FRATESIU. and RIGGIM. R. (2007) Does migration reduce regional disparities? The role of skill-selective ows,Review of Urban and

Regional Development Studies 19(1), 78102.GARTNERH. (2005)The Imputation of Wages Above the Contribution Limit with the German IAB Employment Sample. FDZ Metho-

denreport 2, Institut fr Arbeitsmarktund Berufsforschung (IAB), Nrnberg.GERNANDTJ. and PFEIFFERF. (2009) Wage convergence and inequality after unication: (East) Germany in transition,Labor Markets

and Economic Development73, 387404.GIESECKEJ. (2009) Socio-economic risks of atypical employment relationships: evidence from the German labour market,EuropeanSociological Review25(6), 629646.



21/21

HELPMANE., ITSKHOKIO. and REDDINGS. (2010) Unequal effects of trade on workers with different abilities,Journal of the EuropeanEconomic Association8(23), 421433.

HUNTJ. (2006) Staunching emigration from East Germany: age and the determinants of migration, Journal of the European EconomicAssociation 4(5), 10141037.

HUNTG. L. and MUELLERR. E. (2004) North American migration: returns to skill, border effects, and mobility costs, Review ofEconomics and Statistics86(4), 9881007.

JUHNC., MURPHYK. M., TOPELR. H., YELLENJ. L. and BAILYM. N. (1991) Why has the natural rate of unemployment increasedover time?, Brookings Papers on Economic Activity 1991(2), 75142.

KANBURR. and RAPOPORTH. (2005) Migration selectivity and the evolution of spatial inequality, Journal of Economic Geography5(1), 4357.

LEUVENSTEIJNM. V. and PARIKHA. (2002) How different are the determinants of population versus labour migration in Germany?,Applied Economics Letters 9(11), 699703.

LUCASR. J. (1988) On the mechanics of economic development, Journal of Monetary Economics 22(1), 342.MAUROP. and SPILIMBERGOA. (1999) How do the skilled and the unskilled respond to regional shocks?: the case of Spain, IMF

Staff Papers 46(1), 117.MCCORMICKB. (1997) Regional unemployment and labour mobility in the UK, European Economic Review41(35), 581589.MERTENSA. (2002) Regional and industrial wage dynamics in West Germany and the United States, Journal of Economics and Stat-

istics 222(5), 584608.MORETTIE. (2013) Real wage inequality, American Economic Journal: Applied Economics5(1), 65103.

MORRISON

P. S. (2005) Unemployment and urban labour markets, Urban Studies 42(12), 22612288.NICKELLS. and BELLB. (1995) The collapse in Demand for the unskilled and unemployment across the OECD, Oxford Review ofEconomic Policy11(1), 4062.

NIEBUHRA., GRANATON., HAASA. and HAMANNS. (2012) Does labour mobility reduce disparities between regional labourmarkets in Germany?, Regional Studies 46(7), 841858.

OCHELW. (2005) Decentralizing wage bargaining in Germany a way to increase employment?, Labour19(1), 91121.PARIKHA. and LEUVENSTEIJNM. V. (2003) Interregional labour mobility, inequality and wage convergence,Applied Economics

35(8), 931941.PISSARIDESC. A. and MCMASTERI. (1990) Regional migration, wages and unemployment: empirical evidence and implications for

policy,Oxford Economic Papers 42(4), 81231.PUGAD. (2002) European regional policies in light of recent location theories, Journal of Economic Geography 2(4), 373406.PUHANIP. A. (2001) Labour mobility: an adjustment mechanism in Euroland? Empirical evidence for Western Germany, France

and Italy,German Economic Review2(2), 127140.RIDDELLW. C. and SONGX. (2011) The impact of education on unemployment incidence and re-employment success: evidence

from the U.S. labour market, Labour Economics 18(4), 453463.ROMERP. M. (1986) Increasing returns and long-run growth, Journal of Political Economy 94(5), 100237.ROYA. D. (1951) Some thoughts on the distribution of earnings, Oxford Economic Papers 3(2), 135146.SCHBR. and WILDASIND. E. (2007) Economic integration and labor market institutions: worker mobility, earnings risk, and

contract structure,Regional Science and Urban Economics 37(2), 141164.SENFTLEBENC. and WIELANDTH. (2012)The Polarization of Employment in German Local Labor Markets. Discussion Paper No. 2012

013, Humboldt University, Berlin.TODARO M. P. (1969) A model for labor migration and urban unemployment in less developed countries, American Economic

Review59(1), 138148.WILKER. A. (2005) New estimates of the duration and risk of unemployment for West Germany,Journal of Applied Social Science

Studies 125(2), 207237.


Documents

Arntz (2014) Allocation Human Capital in Space