Department of Economics. Ethnic Concentration and Language Fluency of Immigrants in Germany

Transcription

1 Department o Economics Ethnic Concentration and Language Fluency o Immigrants in Germany Aleander M. Danzer Royal Holloway University o London Firat Yaman 1 City University London Department o Economics Discussion Paper Series No. 11/09 1 Department o Economics, City University London, Social Sciences Bldg, Northampton Square, London EC1V 0HB, UK. irat.yaman.1@city.ac.uk. F: +44 (

2 Ethnic Concentration and Language Fluency o Immigrants in Germany 1 Aleander M. Danzer a,b Firat Yaman c a Royal Holloway, University o London, Department o Economics, Egham, Surrey TW20 0EX, UK, Tel/Fa : +44 ( /439534, a.m.danzer@rhul.ac.uk b IZA Bonn, Schaumburg-Lippe-Strasse 5-9, Bonn, Germany c City University London, Department o Economics, Whiskin St, London EC1R 0JD, UK, e- mail: Firat.Yaman.1@city.ac.uk Abstract Studies that investigate the eect o the regional ethnic composition on immigrant outcomes have been complicated by the sel-selection o ethnic minorities into speciic neighbourhoods. We analyse the impact o own-ethnic concentration on the language proiciency o immigrants by eploiting the act that the initial placement o guest-workers ater WWII was determined by labour demanding irms and the ederal labour administration and hence eogenous to immigrant workers. Combining several data sets, we ind a small but robust and signiicant negative eect o ethnic concentration on immigrants language ability. Simulation results o a choice model in which location and learning decisions are taken simultaneously conirm the presence o the eect. Immigrants with high learning costs are inclined to move to ethnic enclaves, so that the share o German-speakers would increase only modestly even under the counteractual scenario o a regionally equal distribution o immigrants across Germany. Keywords: Enclave, ethnic concentration, language proiciency, immigrants, instrumental variable, random utility model JEL Codes: J61, R23, F22 1 The authors would like to thank Herbert Brücker, Jason Abrevaya, Dan Hamermesh, Victor Lavy, Stephen Treo, Jonathan Wadsworth, Natalia Weisshaar, Kata Wol as well as seminar participants at Austin, Royal Holloway, IAB Nürnberg and Essen. Special thank goes to Jan Goebel rom the SOEP department at DIW Berlin. All remaining errors are our own. 1

3 1. Introduction Immigration and the social and economic perormance o immigrants have been controversial policy issues or decades, both in North America and Europe. From the immigrants perspective, leaving behind a amiliar social contet and adapting to a new environment can be challenging; however, the eperience can be eacerbated i immigrants do not succeed in integrating into the host country s society, a state oten associated with the ailure to learn the maority language. The eistence o segregated parallel societies which are said to be characterised by poverty risk and unemployment, has ueled the debate on the integration o immigrants in Germany. The political concern is that certain immigrant groups might orm selsuicient enclaves and challenge the lie-style as well as ormal or inormal institutions o German society. While those ears may be eaggerated and largely related to the until recently maintained oicial denial o Germany being a country o immigration 2, their eistence and impact on the immigration debate is a potential cause or mutual resentment. The scope o this paper is to analyse the eect o regional ethnic concentration on language proiciency o irst-generation immigrants in Germany. By eploiting the act that the post WWII guest-worker immigration was a quasi-natural eperiment through the eogenous placement o immigrants in irms across West Germany, we estimate the causal eect o ownethnicity concentration on a basic type o human capital, namely German speaking and writing proiciency. By merging several representative data sets and addressing potential endogeneity bias with an IV approach we provide robust evidence o a small negative eect rom ethnic concentration on language luency. Although similar questions have been addressed or more traditional immigration countries like the USA, Australia and Canada, no research eists on this question or Germany so ar. 3 The paper ocuses on the language skills o immigrants or the ollowing reasons: Language skills are a crucial part o the human capital endowment o an immigrant and the earnings implications are well documented (see below. By looking at language as an 2 Germany is not an immigration country. was the leading principle or immigration and oreigner -policies in the coalition contract between conservatives and liberal democrats in 1982 (Herbert, 2001, pp See also the essay Integration ist machbar by Bade in the daily newspaper Die Welt ( Sociological research has dealt with the neighbourhood quality o ethnic clusters in Germany (Drever,

4 endogenous variable we dissect one o the proimate determinants o labour market outcomes. I the costs o learning German are o a non-monetary nature (e.g., eort while beneits are largely relected in wages, the assumption o a monotonic and continuous dependency o wages on ethnic concentrations might give rise to misspeciications, as will be shown in the net section. The beneits o having a good command o German etend to many areas outside the labour market (e.g., participation in the civil society or use o health care and have been used to measure successul integration o immigrants. The economic consequences o language command have been studied intensively and or many countries (see or eample Chiswick and Miller (2002 and (2005 and Bleakley and Chin (2004 or the USA, Dustmann and Fabbri (2003 or the UK, Dustmann (1994 and Dustmann and van Soest (2001 and (2002 or Germany, and Chiswick and Miller (1995 or Australia. Comparing the luency penalty across the cited articles is complicated by dierences in immigration histories (e.g., rates o return migration, and by dierences in methodologies including the survey instrument to investigate sel-assessed language proiciency (e.g., the U.S. and Australian censuses distinguish between our levels o English proiciency, whereas the German GSOEP data have ive levels; however, the entire literature conirms that immigrants with good speaking and writing abilities perorm better in the labour market in terms o earnings and employment compared to immigrants who speak and write poorly. Another strand o the immigration literature has ocused on the inluence o ethnic enclaves on economic perormance and/or language luency. Theoretical arguments or the inluence o ethnic capital and its transmission through neighbourhoods on immigrants perormance have been made by Boras (1995 and (1998. Most studies that we are aware o ind a negative association between ethnic concentration and language proiciency (Cutler et al. (2008, Chiswick and Miller (2005, Lazear (1999 or the USA, Warman (2007 or Canada, and Dustmann and Fabbri (2003 or the UK, and Chiswick and Miller (1996 or Australia. Only the paper by Cutler et al. (2008 attempts to correct or the potential sel-selection o immigrants into speciic neighbourhoods (ghettos by using an occupational instrument matri. For Germany, no study analyses the link between ethnic concentration and language proiciency. The cited papers vary substantially in the size o the regions or which ethnic concentrations are deined, but the 3

5 negative eect is consistently stronger when the regions are deined on less aggregated levels. I immigrants who are less willing or able to learn a language cluster in local neighbourhoods and counties this inding is not surprising. Stronger identiication attempts have been made when looking at earnings o immigrants and how they are aected by ethnic concentration. Damm (2009 and Edin et al. (2003 use eogenous placement policies or immigrants in Denmark and Sweden. Initial placement o arriving immigrants is eploited to instrument or current eposure to their own ethnic group. Cutler and Glaeser (1997 use instruments pertaining to the administration and topography o the regions, such as the number o local governments and rivers in a metropolitan statistical unit. The results with respect to earnings are not as unambiguous as or language luency. Edin et al. ind that living in enclaves improves earnings o less skilled immigrants while no signiicant eect pertains or immigrants with more than 10 years o education. Damm inds that higher ethnic concentrations increase earnings irrespective o skill levels, Warman inds negative eects o enclaves on income growth and Cutler and Glaeser ind negative eects o segregation or Arican-Americans. This paper combines the latter two strands o literature by using an initially eogenous placement policy in order to instrument the eect o regional ethnic composition on language ability. Apart rom being the irst study o its kind or Germany, our paper adds to the literature a learning and location choice model which yields testable implications or the link between ethnic concentration and language proiciency, and which is able to eplain why studies on the eect o enclaves on earnings remain contradictory. Furthermore, the model allows simulating counteractual outcomes or changes in regional ethnic concentration or average immigrant characteristics. This eercise is inormative or gauging the impact o potential uture immigration when Germany ully opens her labour market or the Central and Eastern European countries o the EU in The remainder o the paper is as ollows: In Section 2, we set up our simple learning and location choice model. Section 3 gives a brie overview o the guest-worker programme in Germany and underlines speciic characteristics that resulted in eogenous placement o immigrants across German regions. Section 4 provides a detailed overview o the identiication strategy used throughout the paper, a description o data sources as well as a discussion o the choice o the regional aggregation level. Section 5 provides the results rom the econometric 4

6 analyses. Section 6 discusses potential eplanations or the dierence in OLS and IV estimates as well as the potential bias rom measurement error. Section 7 contains some brie policy simulations based on our structural model, while Section 8 concludes. 5

7 2. Theory In this section, we turn to the random utility model which derives location and learning choice probabilities through utility maimizing behavior. Suppose learning German is costly, and the cost o immigrant i can be described by some observable characteristics X i, a vector o parameters β, and an unobservable component ε, such as ability, which is assumed to be continuous and which we allow, but do not require, to dier across regions. Assuming the cost to be linear in variables we write: c i = X β + ε (1 i i We assume that an immigrant enoys some beneit rom the number or share o people he can interact with. An eample is the model by Lazear (1999, in which two people in a region are matched randomly and trade occurs with a ied payo i both can interact, that is, speak the same language. In this case the beneit would be the epected payo beore a match occurs and it would be linear in the share o people an agent can interact with. For the moment we ust use a generic unction ξ ( where location, so that takes the value equals the raction o people the immigrant can interact with in i the immigrant does not speak German (the subscript standing or oreign and + i he speaks German (n standing or native. The shares o n natives and all oreigner groups (denoted by an indeing set F have to sum to 1: + z F = 1, 0. (2 n z n z Dierent locations are then characterised by: 1. variables diering across locations but equal or all immigrants in that location, W, 6

8 2. variables characterizing the ethnic concentration,, which dier n across locations and across immigrant groups (but not across immigrants o the same country o origin, 3. an unobservable, continuously distributed component r i. Denoting by S i an indicator taking on the value o one, i the immigrant learns German, and zero otherwise, utility o choosing S and location or a given immigrant i is thus: U S ξ X β S ξ W, = *( ( + + (1 * ( + γ + r Sε (3 i ( S, i n i i i i i The irst term describes the deterministic part (rom the point o the econometrician o utility rom learning German. The immigrant can interact with both natives and immigrants o his own group, but incurs the cost other members o the immigrant group only. The term X i β. The second term is the utility enoyed by interacting with γ describes the utility speciic to W + r i the region or the immigrant, regardless o whether or not he learns German, and the last term is an unobservable part o the cost o learning German. The choice set consists o all unordered distinct pairs o ( S, and the chosen alternative is ( ( s, s {0,1}, J S*, * = arg ma{ U } (4 Equation (4 simply states that learning and location decisions are part o the same choice, a act that was acknowledged but not ormalised earlier by Lazear (1999 and Bauer et al. (2005. For notational simplicity, denote the observable part o utility by 7

9 V V i,(1, i,(0, = ξ ( n = ξ ( + + W γ X β + W γ i and the composite error term r i Sε by ω (,. Omitting the individual inde, the i i i, S probability o learning German and choosing location is given by P( S = 1, = P( V V ω ω =Φ (1, ( V (1,1 (1,, V (0,1 ( s, k,..., + V (1, ( s, k, V (0, (1,,..., V s {0,1}, k (1, J, V (0, J (5 with Φ 1, (V being the distribution unction o ( ω 1,1, ω(0,1,..., ω(1,, ω(0, or ( ( J J P ( S = 1, at V. The second equality is simply saying that the probability o choosing a particular ( S, is increasing in the associated utility and decreasing in the observed utility o any other alternative. In order to be able to make statements about the reaction o learning probabilities to changes in the immigrant share within a region we need to introduce an assumption concerning the payo unction ξ : Assumption 1. ξ ( is dierentiable and strictly increasing in its argument. It ollows: Proposition 1. Let the choice problem o the immigrant be described by equations (2 and (3. Under assumption 1, and holding constant the shares o all immigrant groups other than : P( S =

10 P( S = 1, 2. 0 P( k 3. 0, P( 0, k P( S = with strict inequalities i ω, has strictly positive density everywhere. ( s Proo. 1. The probability o learning German conditional on location is: P S = = P U ( > U = P( V V >ω ω ( 1 (1, (0, (1, (0, (0, (1, Taking the derivative with respect to gives: P( S = 1 = g * ( ξ ( + n ( 1+ 1 ξ ( 0 where g denotes the density o ω 0, ω(1, ( 2. Taking the derivative o P ( S = 1, with respect to gives: P( S = 1, Φ = V (1, (1, * ξ ( + n Φ *( ( V (1, (0, *( ξ ( 0 3. The probability o moving to k is: P ( k = P( S = 1, k + P( S+ 0, k =Φ (1, k + Φ (0, k 9

11 P( k Φ (1, k Φ (0, k Then = *( ξ ( + *( ξ ( 0 ( V ( V (0, (0, P( k P( P( Since + = 0,, it ollows that 0. k 4. The unconditional probability o speaking German is: = P( S = P ( S = 1 1 k P( k with the derivative k P( S = 1 P( S = 1 P( k = P( + P( S = 1 k k P( S = 1, P( k = + P( S = 1 k 0 k Q.E.D. All o the inequalities above ollow the same intuition: An increase o increases the observable part o utility o only one choice, which is moving to and not learning German. In particular, it leaves the utility o choice (S=1, unchanged, since the increased immigrant share ust replaces natives and does not change the interaction possibilities or a German-speaker. This eect is captured in the (-1+1 terms in the proo. Thus, all options including learning German are decreased in value relative to (S=0,. Furthermore, since all choice probabilities other than P(S, are decreasing in V ( S,, the probability o moving to any location k is also decreasing. I we assume ξ to be concave, the condition that a higher share o the own immigrant group replace the respective share o natives can be relaed. Furthermore, the results can be generalised to the case where and n stand or the absolute number o immigrants and natives in a region. At least in the latter case, concavity would not be an innocuous assumption i 10

12 there eist eternalities in beneits rom or some range, such as threshold values or the supply o certain goods and services. >> Figure 1 about here << An interesting implication o the model arises when the costs o learning German are unobserved (time and eort spent learning, but beneits are to some etent relected in higher earnings (see Figure 1. The earnings (solid line o immigrants will then be increasing in the ethnic concentration or immigrant i or concentrations above a certain threshold value with > i i being the concentration o own group members at which immigrants stop to learn German. For all values below i the immigrant learns German and her earnings are invariant to [ 0, i ]. I i is smaller or less educated immigrants (they do not learn German even at low ethnic concentrations, empirical studies might ind a positive concentration eect on earnings or less educated and no eect or better educated immigrants (which or eample is ound by Edin et al., or might ind inconclusive results. To compare the quantities in proposition 1, we need another assumption: 1 as Assumption 2. The probability o learning German conditional on location reaches approaches zero: lim P( S 0 = 1 = 1 Proposition 2. Let the choice problem o the immigrant be described by equations (2 and (3. Under assumptions 1 and 2: P( S = 1 P( S = 1 > 11

13 12 or small ( approaching 0. Proo. We need to show: = > = + = + = = = > = + = k k S P P S P k P k S P S P S P S P k P k S P P S P 1 ( lim ( 1 ( lim ( 1 ( lim 1 ( lim 1, ( lim 1 ( lim ( 1 ( lim ( 1 ( lim since 0 0, ( lim 0 = = S P is implied by assumption 2. With 1 1 ( lim 0 = = S P and = k k P P ( ( : = > = + = = k S P k S P k P S P S P 1 ( lim 1 1 ( ( ( lim 1 ( lim 1, ( lim which holds since = > = = S P S P S P 1 ( lim 1 ( lim 1, ( lim 0 0 0, 0 ( < k P and ( < = k S P. Since ξ is dierentiable, there is a neighbourhood around or which the inequality holds. Q.E.D. Note that the small condition is suicient, and less restrictive and/or alternative conditions can be ound. For eample, the inequality will hold i the marginal utility rom contact

14 with other people goes to ininity as the share o people one can interact with goes to 0, limξ ( =, or whenever 0 k P( k P( S = 1 k k P( k P( S = 1 The intuition o the proposition is that i the immigrant is limited to one location, he cannot escape the incentive to learn German by moving to another location. Lowering the immigrant share in a location where it was low initially is not going to change the learning decision o the immigrant, since he can choose rom a multitude o locations. The results are airly general and do not require any distributional assumptions other than continuity on the ω. In particular, no covariance structure is assumed. To illustrate the working o the model we provide a short eample. Let the ω be independently (across choices and individuals and identically distributed type I etreme value errors, resulting in the wellknown multinomial logit model with the choice probabilities given by ( S, e P( S, = (6 V(1, k V( 0, k ( e + e k V Let *( ξ ( (1 *( ξ ( V ( S, = S + + S, so that observable utility is given only n by the composition o the population. Finally, let ξ ( = ln(. It is easy to veriy that assumptions 1 and 2 hold under this speciication. We would have: 13

15 1 P( S = 1 = 1+ P( S = 1 1 = (1+ J P( S = 1 = J + P( S = 1 J = ( J + k k [(1/ 2,1] k 2 [(1/ 2,1] k [ 1, (1/ 4] 2 [ (1/ J, (1/ 4J ] It can be veriied that P( S = 1 lim 0 P( S = 1 < lim 0 The model is highly stylised to highlight the decision problem and the tradeos that immigrants ace, and naturally it has some shortcomings. First, we accept the payo unction ξ as a black-bo mechanism. Agents beneit rom increased communication prospects with other agents, but we do not link these beneits to any deep parameters or structures. 4 A more serious problem might be the omission o moving costs. Here, we are mainly led by data restrictions in our decision not to model moving costs. The bottleneck in the empirical part is the number o immigrants in the German Socio-Economic Panel, with roughly 2,000 observations in 1985 and 1,000 observations in Very ew o those move across regions, as we deine them, and we cannot know or what reasons they change their location. Our conecture is that moving costs would bring the marginal probabilities o learning German conditional and unconditional on location closer to each other, since escaping a region becomes more costly. While the working o the model as summarised above is instructive or understanding the choice situation and the trade-os each immigrant aces as well as or thinking about 4 While this could be done (Lazear (1999 being a possible starting point, it would be only o secondary interest in answering our research question. 14

16 counteractuals, the estimation strategy should depend on the hypothesis to be tested. In the empirical section we aim at estimating P( S = 1 or its intuitive interpretation (as a treatment eect. Identiication will rely on our assumption that initial placement and location choices or a certain time period ater arrival in Germany were eogenous to immigrants with respect to their willingness/ability to learn German. A more holistic estimation (allowing or simultaneous learning and location choices o immigrants o responses to dierent ethnic concentration counteractuals will be carried out in the eperiments section. Naturally, the latter will require more o the structure outlined above (and consequently will be more restrictive, but can be carried out without the use o instruments. Thus, it provides a robustness check o the direct estimation o the treatment eect. 3. The Guest-Worker Programme in West Germany The 1950s and 60s in Germany have become known as the time o the Wirtschatswunder (economic miracle, an episode o rapid post-war reconstruction and economic growth. The miracle has been acilitated by an inlow o reugees rom East Germany and territories ormerly belonging to the German Reich or inhabited by a German-speaking population. As this inlow (8.3 million until 1950 ebbed o, labour shortages became evident, and between 1959 and 1962 the number o vacancies overtook the number o people registered as unemployed. The guest-worker recruitment in Germany began with the German-Italian Recruitment Treaty signed in December 1955 to meet the hunger or labour o the German economy. 5 Subsequent treaties were signed with Greece and Spain in 1960, Turkey in 1961, Portugal in 1964, and Yugoslavia in >> Figure 2 about here << 5 The description o the recruitment history and its technicalities draws on Herbert (

17 Figure 2 shows the development o the share o the oreign population in Germany, where oreign is deined as not holding German citizenship. Until 1960 the presence o guestworkers was a marginal phenomenon, but we see that recruitment gained momentum in the early 60s and increased steadily until A dip in the share o oreign employees occurred in 1967 as the result o a brie recession, which however did not aect the urther inlow o the oreign population. Within 13 years, the share o oreign employees rose rom less than one to twelve percent. Recruitment was halted in 1973 as a consequence o a more severe economic recession; however, the upward trend o the oreign population continued modestly due to amily reuniication. The composition o the oreign population has been subect to substantial changes, as seen in Figure 3. While Italians constituted the most numerous group o oreigners in 1969, the Turkish population overtook all other groups in 1971 and has been widening the gap ever since. Notably, the numbers o Turks never decreased ater the recruitment stop, as it did or other guest-worker groups. >> Figure 3 about here << Technically, the recruitment was perormed by a recruitment commission in the sending country which was ointly set up by the Federal Employment Agency o Germany and the Labour administration o the sending country. German irms requested workers according to their needs and the commission assigned workers rom an application pool to speciic irms. Workers signed one-year contracts with their irst employers at decentralised labour oice branches beore arriving in Germany. Permits to live in Germany or the duration o one year were issued, but the permission was conditional on employment with the employer o the contract. Accommodation and travel costs were covered by the employer, so that monetary and administrative costs o the application and the move were essentially zero or the guest-worker. The recruitment was designed to attract workers with very low skill requirements. In Germany, most guest-workers became employed in manuacturing, notably in the construction, mining, metal and errous industries. As o 1966, 72% o the oreign workorce comprised unskilled workers. 4. Identiication 16

18 The basic question we attempt to answer in this research is whether the ethnic composition in their neighbourhood negatively impacts on the language luency o immigrants. We use the quasi-natural eperiment o the guest-worker immigration that took mainly place in the 1960s and 1970s in order to establish a causal link between area composition and individual ability to speak and write German. Guest-workers were little educated and generally without any knowledge o the German language upon arrival thus reducing the problem o selective migration. 6 As guest-workers were contracted in their home countries based on the (mostly manual labour demand o German irms and administered by outlets o the German Labour Oice, migrants had no control over their placement in Germany. 7 The idea is then to compare immigrants who were placed in areas with dierent ethnic compositions and thus with dierent incentives and costs to learn German. The natural counteractual or a person living in a cluster with a high concentration o own ethnic co-residents is a person o the same ethnicity in a lowconcentration area. Comparing persons o the same ethnicity levels out the potential bias rom linguistic distances between languages. The ideal set-up o our investigation would be to have a data source with obective measures o language speaking and writing luency or immigrants who were randomly distributed over Germany without ever changing their place o residence. In this case, we could simply estimate the basic OLS model y = α + β+ γ ( ysm + κ + µ + u (7 where y stands or language ability, stands or ethnic concentration, (ysm stands or eposure to the host country language (years since migration, κ are country o origin ied eects, µ are regional ied eects and u is a random error term. The estimated coeicient β would report the own-ethnic concentration eect which should carry a negative sign in case we 6 In a recent study on linguistic integration o immigrants in Germany, still more than 90 percent o Turkish immigrants responded that they had no usable German knowledge upon arrival (Rother, Given this procedure, the initial placement was eogenous to the guest-workers. From the perspective o amily members moving to Germany in the ramework o the amily reuniication, the location was also eogenous. 17

19 epect ethnic concentration to inhibit learning German, that is i assumption (1 in the theory section holds. Data In order to estimate the causal eect o ethnic concentration on language luency, this paper combines dierent data sources. As we are interested in language ability o individual immigrants, we make use o the guest-worker sample B o the German Socio-Economic Panel (GSOEP which was started in 1984 and which provides detailed inormation on individual and household characteristics. This sample initially comprised 1,393 households with either a Greek, Italian, Spanish, Turkish, or Yugoslavian household head. Due to the limited sample size o the GSOEP we have to use administrative data in order to generate regional concentration measures o guest-workers. Unortunately, the 1984 wave o the GSOEP does not allow suiciently detailed regional merging with other data sources: instead we use the 1985 wave comprising 2,346 immigrants with ull inormation rom the ive most important guest-worker countries. The main outcome o interest is language knowledge. As mentioned beore, we would be interested in using obective language measures, which are to date, however, unavailable in Germany. 8 As a consequence we are let with indicators o sel-assessed language luency and writing ability which are measured on a ive-category ordinal Likert-scale ranging rom not speaking at all (lowest category to speaking very well (highest category. For most o the analysis, we use a binary variable or speaking and writing ability which takes on the value one or the two highest scores on the Likert scale and zero otherwise. As can be seen rom Table 1, less than hal o the sample claimed to speak or write German at least at a good level in >> Table 1 about here << Demographic inormation comprises gender, marital status, country o origin, age at migration, years since migration, years o schooling, a dummy variable or education abroad and a dummy indicating the presence o children in the household. Table 1 urther reveals that the average immigrant entered Germany at relatively young age (23 years and had spent almost 15 8 The ederal oice or Migration and Integration initiated an integration panel which started in 2007 with a ocus on the eect o language course participation on language ability. Even in this study, no obective language evaluation was possible due to legal uncertainties and the absence o a coherent test scheme (Rother,

20 years in the country. Educational attainments are rather low (at nine years o schooling which is consistent with the act that the vast maority o educational degrees was attained in the home country. The gender mi as well as the common presence o children in immigrant households relect the migration or amily uniication, which became dominant ater the recruitment stop in Given the scope o the guest-worker programme it might be surprising that the German government never collected detailed inormation on where guest-workers moved and or how long they stayed, leaving us with general data sources. To generate ethnic concentration measures, we use the IAB 9 Beschätigtenstichprobe o 1975, a two percent sample o all persons with social security insurance in Germany. This employee-sample comprises 2% o the entire employee population plus recipients o certain social transers like unemployment beneits. Employers mandatorily register employees or the payment o social security taes, so that employees can be tracked through their social security numbers until dropping out o the labour orce. 9 The research institute ailiated with the Federal Employment Agency o Germany. 19

21 Instrumental variable approach Although the placement o immigrants in Germany was eogenous to them, they were in reality allowed to move ater one year o work (including being allowed to return. However, until the economic recession in mid 1970s, guest-workers would move only to ollow labour demand, and normally only short distances (i.e., within region. These moves must be understood as steps towards settling down in Germany, ater many guest-workers had spent the irst time in employer-provided accommodation. The act that immigrants moved across regions might imply that the propensity to move into ethnically homogeneous regions (enclaves is correlated with some unobservable characteristics o migrants. For instance, migrants who are less able or willing to learn German could sel-select into ethnic clusters in order to reduce the costs adherent to absent language skills. I this was the case we would epect naive OLS estimates o the enclave eect to be biased away rom zero. To overcome this bias, we use an instrumental variable approach where we estimate the ollowing system o equations: y= α + β+ γ ( ysm + κ+ µ + u = γ + λz+ e (8 0. The IV estimator where z is the instrument which satisies the assumptions Cov(z, 0 and Cov(z, u = ˆ β IV n i= 1 = n i= 1 ( z z( y i ( z z( i i i y can be interpreted as the ratio between the reduced orm relationship between y and z over the irst stage relationship o y and. The main instrument used in this application is the ethnic composition o regions in 1975, thus ten years beore our language ability measures were 20

22 taken and at the time when the guest-worker programme had ust come to an end. At this time, the placement o guest-workers was still predominantly eogenous to them, implying that z is uncorrelated with any unobservable actors that are accumulated in u. We identiy the eect o ethnic concentration on language acquisition through the use o the eogenous 1975 ethnic composition as an instrument or ethnic concentrations in 1985, the earliest year or which we have a sample o immigrants with German proiciency and residence county inormation. The identiying assumption is that until 1975 the guest-workers have not changed their locations according to characteristics that are correlated with the ability or willingness to learn German. We do not need to assume that guest-workers never moved, but that whatever inluenced their moving decision (i they moved was not correlated with unobservable characteristics inluencing the learning decision. Given the economic boom until 1973 and the pervasively low levels o education and skills among the guest-workers, we do not epect much sorting across regions until 1975; since transitions rom one employer to another were most likely to happen within rather than across regions, it helps that regions are deined at a airly aggregate level (discussed in the net section. Our instrument might contain measurement error, as we instrument the ethnic concentration eposure in 1985 with the regional concentration o 1975, although we cannot observe individual places o residence in This is potentially problematic i guest-workers have moved within this ten year period, and in consequence ethnic concentrations have become stronger or those who were not likely to learn German in the irst place. Using data rom the employee-sample o the IAB we investigate whether there are systematic dierences in the eposure to ethnic concentration beore and ater moving across regions. We base our calculations on all guest-workers who were present in the sample in 1975 and in 1984, and who moved ; those are migrants who were registered or work (or beneit receipt in dierent regions in 1975 and We base our analysis on the workplace location rather than residence, because o higher non-responses or the latter. 10 We observe that 17% o the guest-workers have moved across regions between 1975 and 1984, as compared to 14% o German nationals. 11 Some but not the entire dierential is due to the act that the immigrants who 10 Reporting residence was not mandatory. 11 When looking at inter-regional moves, the level o mobility seems low when compared to the USA. 21

23 moved were predominantly younger and male. Fity two percent o those who moved turned to neighbouring regions. We also construct a variable DIFF deined on guest-workers who moved between 1975 and 1984, which is the dierence in the regional ethnic concentration that a guest-worker eperienced between 1975 and For eample, i a Turk lived in Munich in 1975 and in Berlin in 1984, DIFF would be the concentration o Turks in Berlin in 1984 minus the concentration o Turks in Munich in Thus, the change in concentration ater moving cannot be attributed to dierential trends in the overall population o dierent immigrant groups. Figure 4 plots the density distribution o changes in ethnic concentration (DIFF o 2,523 guest-workers in the IAB sample who moved between 1975 and The distribution peaks around zero (the mean o DIFF being and has somewhat more mass to the right. For 54% o the movers the concentration o their own ethnic group changed by less than one percentage point and or 70% by less than 1.5 percentage points. >> Figure 4 about here << It could still be the case though, that guest-workers who moved to regions with lower concentrations dier systematically rom those who moved to higher concentration locations. We thus regress the variable DIFF on educational attainment, age, and nationality dummies (all as reported in I systematic sorting was present, we would epect educational attainment and age to correlate (albeit imperectly with ability or willingness to learn German. For this test we group the educational inormation in our categories, educ1 being education less than high-school (Gymnasium, qualiying or college without vocational training, educ2 high-school degree or vocational training (but not both, educ3 high-school degree and vocational training, and educ4 college degree. >> Table 2 about here << Table 2 reports these OLS results, with educ4 and the Greek dummy being the omitted categories. I anything, lower educational attainments show some weak correlation with a positive change in ethnic concentration, although none o the dummies is signiicant at conventional levels. The size o coeicients is very small, and the mean change in concentrations or guest-workers with the lowest education remains below two tenths o a percentage point when compared to workers with college degrees. Age seems to be irrelevant. Turks and Yugoslavs 22

24 were more likely to move to regions with higher concentrations, but here, too, the magnitudes o the coeicients are modest. When we include interactions between ethnic and educational category dummies (column 2, even the ethnic dummies lose their signiicance (none o the interactions comes close to signiicance. In general, the variable DIFF is eplained very poorly by the regression, with R² not even reaching We conclude that sorting o guest-workers along any observable characteristics has been absent or very modest between 1975 and To urther test the validity o our instrument we also employ a second variable, the regional election result o the Social Democratic Party o Germany (SPD in the national elections o This instrument is suiciently correlated with the 1985 ethnic composition o regions, as guest-workers were predominantly placed in regions with dominant mining and heavy industry sectors, which were traditional strongholds o the SPD. Beyond the link through ethnic composition, the instrument is not correlated with individual language ability, as guest-workers were not entitled to vote in the national election unless having adopted German citizenship. At that time, this was true only or a negligible raction o guest-workers and language knowledge was no criterion or the admission to German citizenship. Additionally, the political landscape in Germany largely ignored the act that guest-workers were starting to settle down and that the intended rotation principle o the migration lows (guest-workers should return ater a irst employment spell never came into eect. Consequently, none o the political parties broached the issue o integration or language policy at that time. Choice o regional level o aggregation Conditional on data availability, ethnic concentrations can be measured at several levels o aggregation. However, there is a qualitative trade-o between small units o aggregation that closely relect the idea o ethnic neighbourhoods (e.g., census tracts in the US ghettoisation literature with an average size o three to ive thousand inhabitants; Cutler and Glaeser, 1997; or municipalities in Sweden with a median population size o 16,000 inhabitants; Edin, Frederiksson and Aslund, 2003 and larger units, that circumvent the potential bias rom sel-selection into neighbourhoods (e.g., through the use o metropolitan level data (CMA; Warman, 2007; Cutler, Glaeser and Vigdor, The latter approach assumes that the problematic sel-selection o individuals into ethnic enclaves mainly takes place within cities rather than across. In our analysis we use so-called Anpassungsschichten, which are regional units comprising a larger city and the 23

25 economically linked hinterland. In West Germany including West Berlin, there were 111 Anpassungsschichten in 1985 with an average population size between 135 and 500 thousand inhabitants, respectively. By including Anpassungsschicht and ethnicity ied eects, we eploit only variation in ethnic concentrations that is not systematic across ethnicities or across regions. I the chosen level o aggregation eectively reduced the bias rom sorting, our OLS estimates should be very close to the true eect o own-ethnic concentration. 5. Results In the ollowing we provide empirical evidence o a German language penalty rom living among members o the same ethnicity which is robust when accounting or the endogeneity o immigrants post-initial-placement location choice. Figure 5 gives an initial idea o the correlation between ethnic concentration in the location o immigrants (here the log o the normalised requency and their average language luency in German (as a share o immigrants who speak German well or very well. The correlation between the two variables o interest is negative, with the variance across regions being substantial. It becomes evident that larger regions contain higher ethnic concentrations. >> Figure 5 about here << Main results Table 3 indicates that there is a signiicantly negative return to language luency rom living in an area with higher own-ethnic concentration. When including control variables, the coeicient becomes more pronounced and is Equivalently, i the ethnic concentration increases by one standard deviation, the probability that a person is luent in German decreases by 2.6 percent. Although the eect o own-ethnic concentration might seem small at irst, one has to consider the high level o aggregation it reers to. Other authors have ound similar eects at high levels o aggregation or the USA (Chiswick and Miller, 2005 or Canada (Warman, The table urther reports results rom speciication (2 which comprises an instrumental variable 12 The largest part o the eect stems rom variation across regions (50 percent. Thirty-nine percent o the eect is due to variation across ethnicities. 24

26 approach. The use o the instrument (in columns 3 and 4 returns a very similar coeicient, modestly urther rom zero than our OLS estimate. 13 >> Table 3 about here << Table 4 reports results rom the same estimations using writing luency as the dependent variable. Interestingly, the concentration eects are equally precisely measured when compared to Table 3, however, the eects are substantially closer to zero and signiicantly dierent thereo only in the IV estimation (columns 3 and 4. In Table 5, we repeat the analysis o speaking luency and test the robustness o our results by using alternative measures o ethnic concentration. Columns 1 and 4 use the absolute number o own-ethnic minority members, which simply relects a transormation o our initial results (see also theory section. The remaining columns use alternative measures o ethnic concentrations. The dissimilarity inde ranges between zero and one with the corner solution representing the state o perectly equal distribution across space and the state o perect concentration o all minority members in one region. 14 The isolation inde is a measure ranging between zero and one which relects the degree o isolation which an average member o an ethnicity aces on top o the equal distribution o this ethnicity across space. 15 As can be seen rom Table 5, our results are robust to the use o alternative measures o ethnic concentration or segregation; two-stage-least-squares estimators are consistently more negative than the OLS estimates. >> Table 4 about here << >> Table 5 about here << In Table 6 we add urther robustness concerning our dependent variable. So ar, we have used a binary indicator or speaking and writing luency. These variables are, however, generated rom ordinal rankings o ive answer categories. Columns 1 and 2 report basic results 13 We also use a variety o transormations o this instrument (e.g., ranks yielding qualitatively the same results. 14 The ormula or the Dissimilarity inde is N 1 # ethnicityi # non ethnicityi or ethnicity i in region. 2 # ethnicity non ethnicity i, total # i, total N # ethnicity 15 i # ethnicityi # ethnicityi, total The ormula or the Isolation inde is # ethnicity population population or ethnicity i in region. i, total # ethnicity min 1, population i, total smallest i # ethnicity population i, total i, total i, total 25

27 or OLS regressions that use the ull inormation o the language sel-assessment. Although the coeicients are hard to interpret, one can iner the robustness o our results rom them. 16 Columns 3 to 6 use a transormed binary concentration measure that takes the value one i the ethnic concentration o an ethnicity in a region lies above the 75th percentile o the entire ethnic concentration distribution, and zero otherwise. Due to the loss o inormation, the precision o the estimation in column 3 is lower compared to the one with continuous concentration measures. As column 4 shows the contact rate with natives might matter more or language acquisition than simply living in own-ethnicity enclaves. Although living with ewer Germans outside enclaves might be beneicial or language acquisition, 17 the absence o native speakers inside enclaves has a strongly negative impact on language knowledge. Columns 5 and 6 report dierential eects or older and younger migrants according to their age at migration. The comparison o both columns shows that older immigrants bear most o the negative impact rom enclaves while those who immigrated at younger age have no disadvantage rom living in an enclave; these results reconirm indings or other countries (Warman, The oint coeicient o a young immigrant in an enclave is signiicantly positive (s.e >> Table 6 about here << Table 7 shows urther instrumental variable estimation results. Given the relatively small sample size, we preer the use o only one instrument. However, we have a second instrument at hand with which to test or over identiication o the equation. Using only the second instrument the election outcome or the Social Democratic Party o Germany (SPD in 1976 the 2SLS estimator becomes even more negative. Employing both instruments at once we produce an over identiied model: the estimated coeicient moves very close to our initial result and the Hanson test statistics conirms that our instruments satisy the orthogonality condition. Even when introducing a number o interactions (column 4 we cannot clearly reect the null hypothesis that the instruments are invalid. Columns 5 and 6 produce the reduced orm results or both instruments. >> Table 7 about here << 16 The results are also robust to the use o ordered probit estimation. 17 Generally, immigrants tend to have more social contacts with other immigrants irrespective o ethnicity. As a result, German might be the language o communication among immigrants rom diverse ethnic backgrounds. 26

28 >> Table 8 about here << Table 8 adds evidence rom a non-parametric perspective. We perorm nearest neighbour propensity score matching to generate the closest counteractuals o our observations artiicially. 18 For two dierent binary treatment variables, the irst line reports the result without matching. The remaining rows are dierent versions o the matching estimator employing dierent numbers o nearest neighbours. As evidenced in the table, applying the matching estimator increases the language ability o the control group, i.e. in the unmatched sample we underestimate the language ability o those residing outside enclaves. Also, the average treatment eects are clearly signiicant, lending urther robustness to our earlier results. 6. Measurement Error The model estimated in (7 has several sources o potential measurement error which will be discussed in this section. More speciically, we wish to eplain the act that 2SLS estimates are more negative than standard OLS results. Models using language ability as an eplanatory variable (e.g., in wage regressions have discussed the measurement error inherent to sel-assessed language knowledge (Dustmann and van Soest, 2001; Bleakley and Chin, Survey respondents might generally misudge their language ability, and the deviation o sel-assessed rom obective luency might be correlated with level o education (i.e., better educated might have a better idea o their true language ability and level o language ability (i.e., those in the upper part o the luency distribution have less room or over-estimating their ability with the reverse being true or the other etreme o the luency distribution. 19 In our application, language luency is, however, the dependent variable and measurement error herein reduces precision while it does not introduce any bias into the estimates. This can be seen rom estimating the basic model y * = α + β+ u 18 The matching estimators are well-itted with ull support. 19 In our sample, there is a strong central tendency in the ive category Likert scale with only 15 percent o respondents claiming to have no (category 1 or very good (category 5 language ability. 27

29 where y* is the observed dependent variable which however relates to the true dependent variable in the ollowing way: y * = y + true e yields where e is a random error term. Estimating the variance o the coeicient o interest ( β 2 2 ~ σ u + σ e Var = NVar ( X implying a larger than true coeicient variance and standard error. More serious than in the dependent variable is measurement error in independent variables as it may bias the estimated coeicients. As such, this type o error might potentially drive OLS estimates closer to zero and eplain our inding o more negative 2SLS estimates. Our ethnic concentration measures are computed or ive ethnicities (Greek, Italian, Spanish, Turkish, and Yugoslav rom the IAB Employee sample 1985 which comprises two percent o all individuals with social security insurance in Germany. It seems reasonable to assume that these densities suer rom measurement error, especially in regions which comprise a generally low share o oreign population or ew individuals o one single ethnicity. In support o these measures, it should be noted that social security insurance was compulsory in Germany at that time (and still is and that unemployed individuals are also included in the sample. Further, due to the demand-driven nature o the guest-worker programme, ethnic minorities were more equally distributed across German regions than one would epect under more labour supply driven arrangements. As such, the etent o measurement error is probably not correlated with characteristics o the region other than size and thus should be o little concern in our estimation. Attenuation bias towards zero could be shown in our data when the instrument has better measurement properties than the original density measures. In our case, this seems rather 28