Ethnic Enclaves and the Economic Success of Immigrants Evidence from a Natural Experiment *

Transcription

1 Ethnc Enclaves and the Economc Success of Immgrants Evdence from a Natural Experment * by Per-Anders Edn, Peter Fredrksson, and Olof Åslund ** December 14, 2000 Abstract Recent mmgrants tend to locate n ethnc enclaves wthn metropoltan areas. The economc consequence of lvng n such enclaves s stll an unresolved ssue. We use an mmgrant polcy ntatve n Sweden, when government authortes dstrbuted refugee mmgrants across locales n a way that may be consdered exogenous. Ths polcy ntatve provdes a unque natural experment, whch allows us to estmate the causal effect on labor market outcomes of lvng n enclaves. We fnd substantve evdence of sortng across locatons. When sortng s taken nto account, lvng n enclaves mproves labor market outcomes; for nstance, the earnngs gan assocated wth a standard devaton ncrease n ethnc concentraton s n the order of four to fve percent. Keywords: Immgraton, Enclaves, Labor market outcomes JEL classfcaton: J15, J18, R23 * We have benefted from dscussons wth Kenneth Carlng and from the useful comments of George Borjas, Jan Ekberg, Ed Glaeser, Clauda Goldn, Larry Katz, Magnus Löfström, Regna Rphahn, Dan-Olof Rooth, semnar partcpants at the Swedsh Insttute for Socal Research, Tnbergen Insttute, Uppsala Unversty, Unversty of Tlburg, the CEPR conference on Margnal Labour Markets n Metropoltan Areas, the 2000 meetng of EALE/SOLE, the NBER Summer Insttute 2000, and the CEPR European Summer Symposum n Labour Economcs We thank Lsa Fredrksson for expert data assstance. We also thank Sven Hjelmskog, Stg Kattlakosk, Chrstna Lndblom, Anders Nlsson, Krstna Sterne, and Lena Axelsson of the Immgraton Board, and Anna Gralberg of the Mnstry of Culture, who generously found tme to answer our questons. Ths research has been partly fnanced through a grant from the Swedsh Councl for Work Lfe Research (RALF). ** Department of Economcs, Uppsala Unversty, Box 513, S , Uppsala, Sweden. E-mal: per-anders.edn@nek.uu.se, peter.fredrksson@nek.uu.se, olof.aslund@nek.uu.se

2 1. Introducton In most countres, mmgrants tend to be spatally concentrated. In the US, for nstance, recent mmgrants often resde n ethnc enclaves, usually located n metropoltan areas; see LaLonde and Topel (1991). Another example s Sweden, where the share of the foregnborn populaton lvng n the three largest metropoltan areas outstrpped the share of the natve populaton by 18 percentage ponts n Ths paper deals wth the economc consequences of lvng n these enclaves. There are many potental explanatons for ths locaton pattern: 2 economc ncentves may dctate that new mmgrants use the establshed networks of prevous mmgrants; t may also reflect the mportance of ethnc tes per se, dscrmnaton n the housng market, or the ratonal response to mperfect nformaton. 3 A farly large, and predomnantly Amercan, lterature has examned the ndvdual consequences of lvng n enclaves (or ghettos / neghborhoods ). Cutler and Glaeser (1997), for nstance, fnd that blacks lvng n segregated areas have sgnfcantly worse outcomes than blacks n ntegrated areas. The earnngs effects are szable: a one standard devaton ncrease of segregaton reduces the earnngs of blacks relatve to whtes by 7 9 percent. The early studes concernng the effects of segregaton on ndvdual outcomes Kan (1968) s the semnal paper estmated the effects treatng the ethnc composton of an area wthn a cty as exogenous. The results from ths type of studes, however, are susceptble to Tebout bas, arsng because ndvduals choose n whch communty to resde. As Evans et al. (1992) llustrate, statstcally sgnfcant neghborhood ( peer group ) effects may dsappear f proper account s taken to the fact that there s scope for choosng the neghborhood. 4 Later studes, e.g. Cutler and Glaeser (1997), have utlzed the varaton across 1 The 25 largest metropoltan areas of the US hosted 75% of the mmgrant populaton and 40% of the natve populaton. 53% of the mmgrant populaton and 35% of the natve populaton lved n the three largest metropoltan areas of Sweden n Bartel (1989) has shown that mmgrants n the US choose to resde n regons where there are other mmgrants. Moreover, ther locaton decsons are less senstve to wage varatons n comparson to the natve populaton. 3 In the absence of other nformaton on job market opportuntes, new mmgrants may use the locaton of prevous mmgrants as an ndcator of labor market prospects. 4 Mansk (1993) has even questoned whether t s possble to dentfy the peer group effect. 2

3 metropoltan areas, argung that Tebout sortng s less problematc n ths case. 5 Yet another approach to ths ssue s to use parental choces of neghborhoods, where the assumpton s that ths choce s exogenous wth respect to the outcome of the offsprng; Borjas (1995) s an example. Although we agree that these later approaches may have reduced smultanety problems, we stll thnk that the valdty of the mplct exogenety assumptons s an open queston. In the US, for nstance, moves for non-housng reasons consttute almost 40 percent of total moblty; Greenwood (1997). The assumpton of no sortng across local labor markets means that these moves are not the response of, say, ndvduals whose labor market prospects are threatened by an nflux of mmgrants. Indeed the US evdence suggests that ncreased mmgraton mostly hurt prevous mmgrants rather than natves; see LaLonde and Topel (1997). Therefore, we beleve that the ndvdual consequences of lvng n enclaves are stll an unresolved ssue. We take a dfferent approach. Our exogenous source of varaton comes from a Swedsh government polcy concernng the ntal locaton of refugee mmgrants. Ths polcy was vable between 1985 and Government authortes placed refugees n localtes that were deemed sutable accordng to certan crtera. Intally, these crtera were supposed to be related to factors lke educatonal and labor market opportuntes. In practce, however, the avalablty of housng seems to have been the all-mportant factor. Our mantaned hypothess s that the polcy change mpled that the ntal locaton of mmgrants was ndependent of unobservable ndvdual characterstcs. Hence, ths natural experment enables us to reexamne the queston of the economc consequences of lvng n enclaves. 6 The government settlement polcy had real consequences for mmgrant locaton. Ths s llustrated n Fgure 1, whch plots the share of the mmgrant nflow and stock that resdes n Stockholm and the north of Sweden respectvely. Pror to 1985, refugees were allowed to settle n a neghborhood of ther own lkng. In 1985, the mmgrant shares n Stockholm and the north of Sweden stood at 36 and 5 percent respectvely. By 1991, the share lvng n Stockholm had been reduced by more than 3 percentage ponts, whle the 5 Bertrand et al. (2000), Dustmann and Preston (1998), and Gabrel and Rosenthal (1999) make analogous arguments. 6 Katz et al. (2000) uses a smlar approach. 3

4 share resdng n the north ncreased by 2 percentages. Thus, the polcy ntatve clearly ncreased the dsperson of mmgrants across Sweden. Fgure 1: Share of non-oecd mmgrant nflow (sold) and stock (dashed) located n Stockholm and n the North of Sweden respectvely, Stockholm North Stockholm North Notes: Stockholm refers to the county of Stockholm, North to the sx northernmost countes of Sweden. Own calculatons usng the LINDA mmgrant sample. Our results can brefly be summarzed as follows. We fnd pervasve evdence of sortng across local labor markets. The coeffcents on muncpalty characterstcs, whch s the level at whch our measures of, e.g., enclaves pertan, dffer rather drastcally between estmates that account for sortng and those that do not. For example, estmates that suffer from sortng bas assocate an (nsgnfcant) earnngs loss of 1.0 percent wth a standard devaton ncrease n ethnc concentraton. Our baselne estmates that do not suffer from ths problem suggest a sgnfcant earnngs ncrease of 4.2 percent. The remander of the paper s outlned as follows. By way of background, secton 2 compares the Swedsh and US experence wth respect to mmgraton and ethnc concentraton. Secton 3 gves a descrpton of the nsttutonal settng and dscusses whether we can treat the polcy shft n 1985 as a natural experment. In secton 4, we outlne a smple framework that we use as a gude to specfcaton and nterpretaton. Sectons 5 and 6 turn to the emprcal analyses. We focus on two outcome measures: earnngs and dleness. We use longtudnal mcro data derved from the database LINDA; 4

5 see Edn and Fredrksson (2000). Secton 7 concludes. 2. Immgraton and clusterng: Sweden and the US The purpose of ths secton s to put mmgraton to Sweden nto perspectve. To accomplsh ths objectve we compare the Swedsh mmgraton experence durng the past thrty years wth that of the US. The US s a natural pont of reference snce t s the most extensvely documented country n the lterature. The Swedsh fgures are calculated from LINDA. The US numbers are mostly taken from Borjas (1999). Relatve to the sze of each country, the mmgrant stock n Sweden s greater than that of the US a country that s sometmes referred to as a naton of mmgrants. In 1997, 11 percent of the Swedsh populaton was foregn-born. By comparson, 10 percent of the US populaton was foregn-born n The growth of the mmgrant populaton has, however, been somewhat lower n Sweden durng the past thrty years: whle the mmgrant share of the US more than doubled between 1970 and 1998, t grew by around 60 percent n Sweden between 1970 and The past thrty years has seen a radcal shft n the ethnc composton of mmgraton to Sweden. In 1970, mmgrants of Nordc descent consttuted 60 percent of the foregnborn populaton; by 1997, the share of Nordc mmgrants had been halved. Snce the md- 1980s mmgraton s predomnantly for poltcal reasons. Sweden shares the experence of a shft n the ethnc composton of mmgrants wth many other ndustralzed countres. Over two thrds of legal mmgraton to the US was from Europe or Canada durng the 1950s. By the 1990s less than 17 percent were of European or Canadan orgn. Assocated wth the shft n the ethnc composton of the mmgrant stock s a declne n the relatve skll of the foregn-born populaton. Male mmgrant earnngs declned from 95 percent relatve to male natve earnngs n 1970 to 88 percent n Durng the same tme perod, relatve earnngs of a male mmgrant of Nordc descent ncreased from 92 to 97 percent. In the US, there has been a smlar declne n the relatve skll of 7 The growth rate of the mmgrant to populaton rato was dramatcally hgher durng the 1960s, when t grew by 65% n a sngle decade. 8 Concomtantly, relatve rates of non-partcpaton among male mmgrants ncreased from 1.5 tmes to 2.3 tmes the rate of natve non-partcpaton. When calculatng relatve mmgrant earnngs we appled a lower earnngs lmt correspondng to the mnmum amount of earnngs that qualfes to the earnngs related part of the publc penson system. In 1997, ths amount was 36,300 SEK. 5

6 mmgrants: n 1960 the average mmgrant male earned 4 percent more than the correspondng natve; by 1998, ths had been turned nto an earnngs defct of 23 percent. As noted n the ntroducton, mmgrants are concentrated to metropoltan areas to a larger extent than natves. In 1997, 53 percent of mmgrants lved n the three largest local labor markets n Sweden (Stockholm, Göteborg, and Malmö), whch host only 35 percent of the natve populaton. 9 By comparson, the 25 largest metropoltan areas of the US hosted 75 percent of the mmgrant populaton and 40 percent of the natve populaton; see LaLonde and Topel (1991). In 1998, almost three-quarters of mmgrants lved n only sx US states (Borjas, 2000). It s well known that mmgrants to the US tend to lve n ethnc enclaves. Borjas (1998) has calculated a smple measure of the probablty of resdng n an ethnc neghborhood. An ethnc neghborhood s defned as a neghborhood where the share of the ethnc group n the resdent populaton s at least twce as large as the share of the ethnc group n the US populaton. Accordng to ths measure, 48 percent of an average member of an ethnc group resded n an enclave n 1979; see Borjas (1999). Ethnc concentraton seems to be partcularly hgh among mmgrants from non-ndustralzed countres: e.g. Mexcans, Puerto Rcans, and Cubans. 10 We have calculated ths measure for our sample of frst generaton mmgrants. It turns out that 42 percent of the average frst generaton mmgrant lves n an ethnc neghborhood n Among the top ten source countres, the probablty of lvng n an enclave s partcularly hgh among mmgrants from Turkey, Iraq, and Iran. 11 Thus, ethnc concentraton s a feature of the Swedsh mmgraton experence. Further, mmgrants from developng countres are more lkely to resde n enclaves The placement polcy The objectve of ths secton s to gve the reader a practcal sense about the workngs of the placement polcy. For our purposes we want the actual placement to be ndependent of 9 The defnton of a local labor market s roughly comparable to a US Standard Metropoltan Statstcal Area (SMSA). 10 An ethnc group s defned n terms ethnc ancestry and, hence, many generatons of mmgrants. Neghborhoods are defned n terms of area zp codes. The calculatons are based on the NLSY. 11 The top ten source countres n 1997 (ordered by ther sze) were Fnland, the former Yugoslava, Iran, Norway, Poland, Denmark, (East and West) Germany, Iraq, Turkey, and the former Sovet Unon. 12 For the purpose of ths calculaton, neghborhoods were defned n terms of parshes. The average sze of a parsh approxmately corresponds to the average sze of a US Census tract. 6

7 any unobservable characterstcs n our outcome equatons. To what extent s ths true? Were some ndvduals more lkely placed than others? Unfortunately, there s very lttle documentaton about the practcal mplementaton of the placement polcy. Therefore, we have partly approached the ssues by ntervewng placement offcers and other offcals of the Immgraton Board. We begn by gvng a bref account of the nsttutonal settng. We then proceed to descrbe the handlng of a typcal asylum seeker from the border to the fnal placement. 3.1 The nsttutonal settng 13 In 1985, handlng refugee ssues n Sweden became the formal responsblty of the Swedsh Immgraton Board, and the natonal government took a more mportant role n the handlng of refugee mmgrants. 14 The Immgraton Board assgned mmgrants to a muncpalty of resdence. Muncpal authortes, n turn, assgned mmgrants to an apartment. Recepton n the muncpaltes was regulated n agreements between the Board and the muncpalty n queston. After recevng a resdence permt, the refugee was to stay n the muncpalty for an ntroductory perod of about 18 months. 15 Ths ntal phase, among other thngs, nvolved ntroductory courses n Swedsh. Two aspects of the assgnment strategy should be noted. Frst, the strategy only pertaned to the ntal locaton. There were no restrctons aganst relocatng f ndvduals could fnd a place on ther own. However, moblty mpled the loss of elgblty for some of the specal ntroducton actvtes granted n the assgned muncpalty. In partcular, the mmgrant had to awat a new place n a language course. Second, not all poltcal mmgrants became enrolled n the Immgraton Board s asylum recepton. Durng , a ffth of the nflow of poltcal mmgrants was famly members who traveled drectly to a muncpalty, where the remander of the famly resded. The reform was a reacton to the concentraton of mmgrants to large ctes that had taken place. The dea was to dstrbute asylum seekers over a larger number of 13 Ths secton draws prmarly on The Commttee on Immgraton Polcy (1996) and The Immgraton Board (1997). 14 In practce, the Immgraton Board started handlng refugee ssues durng a tral perod n the autumn of The length of the ntroducton perod appears to have vared across muncpaltes and years; n many cases t was consderably longer. 7

8 muncpaltes that had sutable characterstcs for recepton, such as educatonal and labor market opportuntes. At frst, the ntenton was to sgn contracts wth about 60 muncpaltes, but due to the ncreasng number of asylum seekers n the late 1980s, a larger number became nvolved; n 1989, 277 out of Sweden s 284 muncpaltes partcpated. It was consdered a vrtue f every Swedsh muncpalty took ts share of mmgrants. The advantages of smaller communtes n terms of closeness between people were emphaszed, but the factors that ntally were supposed to govern the choce of locaton were more or less abandoned. Instead, the avalablty of housng became the decdng factor. Formally, the polcy of assgnng refugees to muncpaltes was n place from 1985 to In 1994, a new law was passed that gave mmgrants the rght to choose the ntal place of resdence provded that they could fnd an apartment on ther own. 16 However, the strctness of the placement polcy gradually eroded durng , when there was an mmgraton peak caused by the war n Bosna-Herzegovna. For our purposes, the post perod s less attractve, snce t contaned larger degrees of freedom for the ndvdual mmgrant to choose the ntal place of resdence. The strctest applcaton of the assgnment polcy was between 1987 and In 1988, a new law was passed whch requred extraordnary reasons for all others than famly members to get the rght to stay n a muncpalty nstead of a refugee center whle watng for a resdence permt. 17 In effect, t seems that the law formalzed a strcter practce, whch had been ntroduced n Durng , the placement rate,.e., the fracton of refugee mmgrants assgned an ntal muncpalty of resdence by the Immgraton Board, was close to 90 percent. 3.2 A typcal case of asylum the placement polcy n practce An asylum seeker was placed n a refugee center whle watng for a decson from the mmgraton authortes. Refugee centers were dstrbuted all over Sweden and there was no correlaton between the port of entry and the locaton of the center. However, 16 From then on more than 50 percent of the mmgrants have used ths opportunty. The Immgraton Board has placed the remander of the mmgrants. 17 Ths was a tghtenng of regulatons n the followng sense. Pror to the change, refugees could stay n a muncpalty of ther own choce whle watng for a resdence permt and, n general, the chance of beng assgned the muncpalty of resdence was greater than beng assgned another muncpalty. 8

9 mmgrants were sorted by natve language when placed n centers. There was a long wat for a resdence permt. The mean duraton between entry nto Sweden and the recept of a permt (condtonal on recept) vared between 3 and 12 months durng ; see Rooth (1999). Notce that whether ndvduals were subjected to the placement polcy or not depended solely on when they receved ther resdence permts. So an ndvdual enterng n 1986, but recevng the permt n 1987, was placed accordng to the practce n There was a much shorter wat for a muncpal placement after recevng the permt, partly because placement offcers had explct goals n terms of the duraton of ths spell. When t came to the muncpal placement, weght was gven to mmgrant preferences. Most mmgrants, of course, appled for resdence n the tradtonal mmgrant ctes of Stockholm, Göteborg and Malmö. There were very few apartment vacances n these locatons, however, n partcular durng the second half of the 1980s when the housng market was boomng. When the number of applcants exceeded the number of avalable slots, muncpal offcers may have selected the best mmgrants. There was no nteracton between muncpal offcers and refugees, so the selecton was purely n terms of observable characterstcs; language, formal qualfcatons, and famly sze seem to have been the governng crtera. When the muncpaltes could cream skm, they selected hghly educated ndvduals and ndvduals that spoke the same language as some members of the resdent mmgrant stock. Sngle ndvduals were partcularly dffcult to place, snce small apartments were extremely scarce. After havng been assgned to an apartment, mmgrants man source of ncome was welfare (.e. socal assstance). They could lve on welfare whle partcpatng n ntroductory Swedsh courses. Recept of welfare was not condtonal on resdng n the assgned muncpalty and the central government rembursed the local governments for ther welfare expendtures. So there was lttle ncentve for the mmgrants to stay on n an assgned muncpalty, f they could realze ther preferred choce. The man ndvdual cost, apart from movng costs, conssted of delayed enrollment n ntroductory language courses. 9

10 3.3 The placement polcy as a natural experment On the bass of the above descrpton, we thnk that t s realstc to treat the muncpal assgnment as exogenous wth respect to the random components of the outcomes of nterest, condtonal on observed characterstcs. For hghly qualfed ndvduals ths assumpton s potentally more problematc. Cream skmmng on the part of muncpal offcers suggests that the hgh-sklled may have been able to realze ther preferred opton. 18 The strctest applcaton of the assgnment polcy was between 1987 and We have chosen to base our emprcal work on placements durng the perod. The last year s an obvous choce, snce the probablty of beng exogenously placed s ncreasng n the total nflow of resdence permts and the tghtness of the housng market. There was a hke n the number of new resdence permts n 1989, and the housng market peaked n that year. Gven the choce of 1989 we can follow ndvduals for a maxmum of eght years. To ncrease the sze of the sample we added two addtonal years. We chose 1987 and 1988 snce we wanted to follow ndvduals over tme for as long as possble A stylzed framework The purpose of ths secton s to present a smple model, whch we thnk s a useful framework of thought. Although we examne dfferent outcomes n our emprcal analyss we focus here on the smultaneous determnaton of locaton and earnngs. We begn by examnng the bas of the OLS estmator. We then present the condtons that render dfferent estmaton approaches consstent. Throughout we keep the condtonng on observed characterstcs (apart from locaton) mplct. To llustrate our man ponts we adapt the schoolng model of Card (1999) to our settng. 4.1 The bas of ordnary least squares Consder an mmgrant who derves utlty from lvng wth other mmgrants and the consumpton of goods. The mmgrant maxmzes utlty by makng a locaton choce, 18 Below, we provde some evdence on ths ssue. On the whole, rates of post-placement moblty do not suggest that the hghly qualfed were more lkely to exercse ther preferred opton when beng assgned to a muncpalty. 19 As a gude to the selecton of years we calculated the rato of the nflow of resdence permts and the stock of vacant publc rental apartments. Ths rato stood at 10 n 1989; n 1988 and 1990, t equaled 4, and n 1987 and 1991, 2. 10

11 where each locaton s characterzed by some measure of mmgrant densty (m). For smplcty we assume that there s a contnuum of locaton choce where m spans the real lne between zero and unty: m [01, ]. Utlty s gven by U = u ( m ) + ln y (1) The objectve of the ndvdual s to maxmze (1) subject to a market opportunty locus: ln y = α + β m (2) where α reflects general apttudes n the labor market and β s the margnal return of lvng n an enclave or the apttude for enclavng. The frst-order condton s u + β = 0. To proceed, assume that the margnal utlty of lvng n an enclave s lnear n m,.e., u = µ km, where k s a postve constant. Wth ths assumpton we have that the optmal choce of m satsfes 20 m * = µ + β k (3) Now, let us revert to the earnngs equaton. Rewrte ths relatonshp as follows: ln y = α + βm + ( α α) + ( β β) m = α + βm + η (4) Our nterest concerns the parameter β the average return to lvng n an enclave. (Throughout we choose the conventon that non-ndexed varables are populaton averages.) Consder the OLS estmate of equaton (4). The probablty lmt of the OLS estmate, b OLS, s plmb OLS = β + λ + λ 1 m (5) 0 The parameters λ j are theoretcal regresson coeffcents: λ 0 = cov( α, m ) var( m ) and λ 1 = cov( β, m ) var( m ), where m s gven by (3). Assumng that the underlyng random 20 In order to avod dgressng nto detals we do not dscuss the possblty of corner solutons. 11

12 varables, α,, have some jont and symmetrc dstrbuton, the explct β, and µ expressons for λ j are 21 λ λ 0 1 = k σ = k σ 2 µ 2 µ σ + σ αµ 2 σβ 2 β + σ + σ + σ 2 β αβ + 2σ βµ + 2σ βµ βµ (6) The coeffcent λ 0 sgnfes ablty bas, whle λ 1 s related to bas because of self-selecton. In general, we cannot say much about the sgn of the bas, but t s nstructve to walk through some specal cases. We wll consder four such cases: (a) homogenety n the return to enclavng β = ex post returns dffer across ndvduals,.e. β; (b) locaton decsons are based on expectatons of ncome but β = β n (3); (c) preference homogenety µ = µ ; and (d) preferences are ndependent of apttudes σ = 0. (a) β = β. The bas of OLS then equals βµ = σ αµ σαµ bas = k (7a) σ 2 µ The sgn of the bas depends on the correlaton between general apttudes and preferences. Absent any such correlaton OLS s unbased; random varaton of preferences then effectvely trace out the structural earnngs equaton (2). In general, of course, sgn{ bas } = sgn{ σαµ }, but t s dffcult to have a concrete pror about σ αµ. (b) Ex ante homogenety but ex post heterogenety n β. Ths case s smlar to (a) n the sense that the bas of OLS s a functon of the statstcal dependence between preferences and the returns n the labor market: k bas = [σ m] 2 αµ + σβµ (7b) σ µ 21 Symmetry s mposed snce then we do not have to worry about moments of the thrd order. 12

13 (c) µ = µ. The bas of OLS equals σ α bas = k m + ρ (7c) σβ where ρ s the correlaton between ablty and the monetary pay-off of lvng n an enclave. If there s no such correlaton OLS s based upwards; those who gan most from lvng n enclaves choose to do so; the bas s thus purely due to self-selecton. However, we beleve that ρ < 0 s a realstc assumpton,.e., those who do poorly on the labor market for unobservable reasons are those who wth the hghest return to enclavng. One argument for ths belef would be that network effects are more mportant for ndvduals of low general apttude. If ρ < 0, OLS may plausbly be based downwards. (d) σ = 0. In ths case the bas expresson s very smlar to (7c): βµ = σ αµ σ β bas = k 2 2 σ + σ β 2 µ m + σ ρ σ α β (7d).e., the bas s proportonal to, but lower than n, case (c). Suppose that we can obtan a consstent estmate of β. Can we say anythng useful 2 β σ βµ about the sgn of the unknown covarances? If σ > ( ), the bas due to self-selecton ( λ 1) s postve. Snce ths seems lke an nnocuous assumpton, we wll mpose t. As t turns out, there s one nformatve and one unnformatve case. The nformatve case s when OLS s downward based, snce then sgn{ λ 0 } sgn{ σ αµ + σ } < 0. If σ = 0, ths = αβ mples that those of less general apttude derve greater utlty from enclavng. If σ = 0, low-ablty ndvduals have a greater return to lvng n an enclave. αβ αµ 4.2 On the estmaton of the outcome equatons What knd of assumptons do we need n order to estmate the average return to lvng n an enclave consstently? In practce, we have two estmaton alternatves an IV or a control functon approach. For our nstrument to be of any use t s clear that t must be ndependent of the random coeffcents n (4). Let us denote our nstrument, the ntal placement, by 0 m. Then we assume that 13

14 0 0 [( α) m ] = E[ ( β m ] = 0 E α β) (8) We also mpose an excluson restrcton n the sense that only varables assocated wth the current locaton have an effect on earnngs. Gven the above assumptons, two questons arse: What does (lnear) nstrumental varables estmate? Under what condtons does IV estmate the average return to lvng n an enclave consstently? These questons have gven rse to a scholarly debate; the essence of the debate can be found n Angrst et al. (1996) and Heckman (1997). An applcaton of IV to a settng that s smlar to ours s gven n Angrst (1990). He used the Vetnam draft lottery to estmate the effect of mltary servce on earnngs. An IV estmate of the effect of mltary servce on earnngs then s a weghted average of ndvdual treatment effects for those whose mltary servce status was changed because of the value of nstrument.e. for those who were complers. 22 Notce that: () ndvduals who would have enrolled n the mltary rrespectve of ther lottery number do not contrbute to dentfcaton; () the nterpretaton of the IV estmand does not hnge on lnearty of the relatonshps of nterest; and () the group of complers need not be representatve of the populaton. The fact that complers need not be representatve of the populaton s the bass for Heckman s crtcsm of (lnear) IV estmaton of models wth varable treatment effects. He argues that beng a compler nvolves a choce, whch partly may be based on the unobservable earnngs gans of gong nto the mltary. If ths s the case, the average causal response among complers s not representatve of the average treatment effect (ATE) n the populaton. A control functon approach, on the other hand, provdes consstent estmates of ATE even f there s selecton on unobservable earnngs gans, subject to mposng some addtonal structure. Let us provde more substance to our nformal dscusson. Gven (8), the consstency of IV requres 0 E [( β β) m m ] = κ, where κ s constant and ndependent of Wooldrdge (1997). Let m 0 * ( 1 s ) m 0 m ; see = s m +, where s = 1 f the ndvdual stayed on n 22 For ths weghted average to be well-defned a monotoncty assumpton s requred. Monotoncty means that gettng a lottery number that mpled elgblty for draft should not decrease the probablty of servng n the mltary and vce versa. If monotoncty holds, the weghts sum to unty. 14

15 the assgned locaton, s = 0 for the complementary event, and requres that * m s gven by (3). Then IV 0 E[( β β) m m ] = E[( β β) m m, s = 1]Pr( s + E[( β 0 0 β) m m, s = 0]Pr( s 0 = 1m ) 0 = 0m ) = κ (9) We frst note that the probablty of stayng wll generally depend on 0 m. 23 Hence, IV needs that the condtonal means of (9) are ether equal to zero, or that the two terms n (9) are equal n magntude but of opposte sgn. The condtonal means wll equal zero f there s no sortng on comparatve advantage ( β ) or nformaton correlated wth t. In terms of the specal cases consdered n secton 4.1, IV estmates the average treatment effect f the return to enclavng s homogeneous n the populaton (case a), and when there s ex ante homogenety but ex post heterogenety n β (case b), provded that σ = 0. However, as soon as there s self-selecton by comparatve advantage (cases c) and d) IV estmaton s problematc and requres specal assumptons n order to be consstent. So where does ths leave us? Should we use an IV or a control functon approach? Our strategy s to use both methods. The advantage of lnear IV s ts robustness; t gves a weghted average of treatment effects for complers. Wthout further restrctons ths s all the data can be nformatve about. If we are wllng to mpose some addtonal structure (such as lnearty n the outcome equaton and a specfc error structure) we can apply selecton correcton methods to estmate ATE. Subject to ths addtonal structure, the mplct behavoral (or nformatonal) assumpton of IV s, n prncple, testable. 24 βµ 5. Consequences of the reform: moblty and concentraton In ths secton we gve a quanttatve pcture of how the polcy reform affected subsequent moblty and locaton patterns. Subsequent moblty s a natural ndcator of how mmgrants perceved ther ntal placement. We also want to make sure that the polcy 23 0 * 0 The ndcator varable s equals s = I{ U ( m )( m m ) c }, where I s the ndcator functon and c * 0 2 denotes the cost of moblty. Usng (3) ths equals s = I{ k( m m ) c }. There are two cases where the probablty of stayng does not depend on the assgned muncpalty. One s when there are no moblty costs. In ths nstance, however, the rank condton fals, gven our mantaned assumpton that the ntal placement s exogenous to unobserved ndvdual characterstcs. The second example s f moblty costs * 0 2 are gven by c = δ ( m m ) and so s = I{ k δ }. 24 Heckman and Robb (1985) suggest a Hausmann test. 15

16 ntatve actually affected resdental locaton n the longer run. If not, the ntal placement wll not have much predctve value for current regon of resdence. 5.1 Data and sample selecton The emprcal analyss s based on data from the LINDA database. Among other thngs, LINDA contans a panel of around 20 percent of the foregn-born populaton. Moreover, the data are cross-sectonally representatve. Data are avalable from 1960 and onwards, and are based on a combnaton of ncome tax regsters, populaton censuses and other sources; for more detals, see Edn and Fredrksson (2000). We cannot dentfy refugee mmgrants drectly from our data. Instead we dentfy them by country of orgn. As a general rule we nclude mmgrants from countres outsde Western Europe that were not members of the OECD as of The only excepton from ths rule s Turkey, whch s ncluded snce t was the orgn of a substantal nflow of refugee mmgrants durng the perod. Furthermore, persons belongng to a household wth ether a Swedsh-born grown-up or a prevous mmgrant were excluded, snce these ndvduals were lkely to have mmgrated as famly members and, consequently, are not program partcpants. We also apply an age restrcton and base our analyss on ndvduals aged at the tme of entry nto Sweden. Lastly, we focus on the mmgraton waves durng for reasons outlned above. Another feature of the data that s relevant for our analyss s that we observe ndvduals regon of resdence at the end of the year. Thus, the observed ntal locaton may dffer from the actual ntal placement when ndvduals move durng ther frst year. Ths ntroduces a measurement error n ntal placement, an ssue we wll return to when assessng the stablty of our estmates. 5.2 Consequences of the reform In order to gve a quanttatve vew of how the placement polcy affected the locaton pattern of recent mmgrants t s necessary to construct a counterfactual. For ths purpose, we choose ndvduals who are dentfed as refugee mmgrants (accordng to the above crtera) durng the years We use these two samples of mmgrants, and , to llustrate dfferences n ntal and subsequent locaton patterns and 16

17 secondary moblty. Snce we want to use the 1981/83 cohort as an approxmaton of the counterfactual for the 1987/89 cohort, t s vtal that the cohorts are smlar n terms of observable and unobservable characterstcs. Wth respect to observable characterstcs, there were no mportant dfferences n terms of age and educaton. The representatve ndvdual of the 1987/89 cohort was 0.6 years older and had 0.2 years more of mputed years of educaton. The dfference between the two cohorts n terms of ethncty s a greater source of concern. It s well known that ethncty s an mportant determnant of success n the recevng country; ethncty s mportant as t nfluences language sklls and the level of formal educaton vares by orgn country (Borjas, 1994). The chef dscrepancy between the two cohorts s that the 1981/83 cohort has more of the mass among mmgrants from Eastern Europe. The later cohort, by contrast, has the greatest fracton of mmgrants orgnatng from the Mddle East. 25 To us the dfferences n terms of regon of orgn seem substantal. To generate the counterfactual locaton dstrbuton for the 1987/89 cohort, we rewegh observatons n the 1981/83 cohort such that the dstrbuton over regon of orgn conforms to the 1987/89 cohort. Whenever we talk about dfferences across cohorts n the sequel, we refer to the dfferences between the 1987/89 cohort and the weghted 1981/83 cohort. One ndcator of how mmgrants perceved the reform s post-mmgraton moblty. If a consequence of the government polcy was that mmgrants were placed n regons that they deemed nferor, we should observe greater moblty n the 1987/89 cohort n comparson to the earler cohort. The predcton that moblty should be greater n the program cohort s clearly contngent on mmgrants beng able to choose/dentfy ther most preferred regon upon arrval. There are plausble reasons why ths mght not be the case. For one thng, there s probably genune uncertanty about the regonal varaton n the pay-off to labor market sklls and, hence, the answer to ths queston s not obvous. We start by comparng moblty across the two cohorts; see Table 1. We note that, moblty s substantal n both cohorts. Around 50 percent chose to move from ther ntal 25 The share of refugee mmgrants from Eastern Europe declned from 37 percent (1981/83) to 18 percent (1987/89), whle the share from the Mddle East ncreased from 23 to 46 percent. The ncrease n refugee mmgraton from the Mddle East s manly due to the war between Iran and Iraq. The large share of Eastern Europeans n the earler cohort s due to a substantal nflow of mmgrants from Poland n 1982, followng the Soldarty upheavals. 17

18 locaton. Moreover, there are some dfferences between the cohorts. 26 The probablty of remanng n the ntal locaton s lower among those who were assgned a muncpalty by government authortes; the propenstes to emgrate are roughly equal, but there s more nternal moblty n the 1987/89 cohort. Table 1: Indvduals who stayed, emgrated, and relocated, percent. Immgrant cohort 1981/ /89 t and t+8 t and t+8 Stayed Emgrated Relocated Notes: Refugee mmgrants aged at mmgraton. Probablty of emgraton equals probablty of not beng n sample (.e. the fgures nclude deceased). t denotes year of mmgraton. Observatons n the 1981/83 cohort weghted to correspond to the (perod t) regon-of-orgn dstrbuton n the 1987/89 cohort. Table 2: Locaton patterns by populaton densty, percent. Immgrant cohort 1981/ /89 t t+8 t t+8 Regon 1 (Stockholm) Regon 2 (Göteborg & Malmö) Regon Regon Regon Regon 6 (Sparsely populated) Notes: Refugee mmgrants aged at mmgraton. Regon 1 most densely populated; Regon 6 least densely populated. t denotes year of mmgraton. Observatons n the 1981/83 cohort weghted to correspond to the regon-of-orgn dstrbuton n the 1987/89 cohort. Thus, post-mmgraton moblty seems to be hgh; ths s true for both cohorts. To what knds of regons dd the mmgrants move? We nvestgate ths queston n Table 2. As we have noted, the polcy reform was a reacton to the concentraton of the foregn-born to metropoltan areas, prmarly Stockholm, Göteborg, and Malmö. As a consequence of the reform, we should expect a shft n the ntal locaton pattern n favor of sparsely populated areas, often located n the northern part of Sweden. Table 2, whch tabulates regon of resdence accordng to populaton densty, shows that the dstrbuton of ntal locaton across the two cohorts s radcally dfferent. There s concentraton over tme n both cohorts, although much more pronounced n the 1987/89 26 To examne whether these numbers are drven by an overall ncrease n the probablty of relocaton, we have also calculated dfference-n-dfference estmates relatve to a sample of natves. The dfference between the cohorts s moderated slghtly. Gven the smlarty between the before-and-after and dfference-n-dfference estmates, we report the former n Tables 1 and 2. 18

19 cohort. Nevertheless, there s far from total convergence of the two dstrbutons. Thus, t seems that the reform dd have lastng effects on locaton. A tabulaton of muncpaltes by geographcal locaton conveys a smlar message, wth one provso: moblty s not just from a desolated North to the populous South, but also to the regonal centers n the north of Sweden. So, n lne wth our expectatons, there s more moblty and concentraton among mmgrants who were assgned a muncpalty by government authortes. 27 Nonetheless, two addtonal facts are strkng to us. Frst, there s also a lot of moblty n the cohort that was supposedly free to choose, suggestng that nformatonal problems may be of some mportance. Second, eght years after entry to Sweden, the post-reform dstrbuton of mmgrants has far from converged to the pre-reform dstrbuton of mmgrants. An mportant queston s whether some groups were less lkely to be placed than other groups. Our account of the workngs of the assgnment polcy suggests that less qualfed ndvduals may have been able to realze ther preferred choce to a lesser extent than the hghly sklled. An analogous argument holds for sngles. One natural avenue to examne these hypotheses s to look at mgraton propenstes by educaton and famly status. If the above hypotheses are correct we should expect greater moblty among the less educated than among the hghly educated as a consequence of the reform, and lkewse for sngles n comparson to couples. Wth respect to educaton we fnd no evdence n favor of the hypothess on the contrary: dfference-n-dfference estmates suggest that the reform ncreased the relatve relocaton rates of ndvduals wth a unversty degree. Moreover, we found no dfferences accordng to famly or martal status. 6. The effects of lvng n enclaves In ths secton we estmate the effect that lvng n ethnc enclaves has on economc outcomes. We begn by offerng a bref account of the lterature on why enclaves could affect outcomes. Then we present a set of baselne estmates that delver the gst of the results. Fnally, we subject the baselne estmates to a comprehensve set of specfcaton 27 In related work we have examned subsequent moblty n closer detal. It turns out that much of the raw dfference n moblty between the two cohorts dsappears when t s standardzed wth respect to ndvdual characterstcs; see Åslund (2000). 19

20 checks. We wsh to emphasze from the start that the concluson from these checks s that, f anythng, the baselne estmates understate the value of enclaves. 6.1 Why does lvng n enclaves affect outcomes? The purpose of ths secton s to gve a bref revew of the lterature on why enclaves nfluence the outcomes of ndvduals lvng there. We consder four types of explanatons: () slower rate of acquston of host country sklls; () network effects ; () spatal msmatch ; and (v) human captal externaltes. 28 Although we present each explanaton separately, they are not mutually exclusve. The hypothess that the enclave decreases the rate of host country skll acquston seems to have been among the prme motves for the reform that we are utlzng. Accordng to ths vew, the ethnc enclave provdes less nteracton wth natves and reduces the ncentves for acqurng, e.g., the language sklls that are necessary to succeed n the natonal labor market. Thus, the enclave hnders the move to better jobs and reduces earnngs n the longer run. More of a postve vew s contaned n stores that emphasze network effects. The enclave represents a network that ncreases the opportuntes for ganful trade n the labor market; e.g. Portes (1987) and Lazear (1999). Further, the network dssemnates valuable nformaton on, e.g., job opportuntes, and consttutes an envronment where the mmgrant s less exposed to the dscrmnaton encountered elsewhere on the labor market. The enclave would thus mprove labor market outcomes, n partcular for recent mmgrants and for ndvduals who have dffculty ntegratng nto the labor market. Of course, the enclave may also provde nformaton on matters that are not conducve to success n the labor market, such as welfare elgblty; e.g., Bertrand et al. (2000). The spatal msmatch hypothess emphaszes dscrmnaton n the housng market; see Ihlanfeldt and Sjoqust (1998). Snce mmgrants face restrctons n the housng market they are forced to segregate n an enclave. The enclave, n turn, may be dstant from areas that provde employment opportuntes. Therefore, ndvduals lvng n the enclave wll 28 To ths lst one could potentally add relatve factor supples and compensatng dfferentals. Ths story would go as follows. If the typcal mmgrant has preferences for lvng wth members of hs own ethnc group, then he s wllng to pay a prce for lvng n that area. The prce corresponds to the movement along the labor demand curve as labor supply ncreases. The equlbrum sortng of ndvduals wll feature a negatve correlaton between wages and ethnc concentraton. The correlaton s smply due to preferences and does not have a causal nterpretaton. For ths reason, ths story s not ncluded above. 20

21 fare worse than otherwse smlar mmgrants who have managed to escape housng market dscrmnaton. In ths vew, t s not the enclave as such that hampers success n the labor market, but rather that the enclave s located away from employment opportuntes. The stores based on human captal externaltes are also based on resdental segregaton. In ths nstance, however, segregaton s not necessarly bad t all depends on the qualty of the enclave, e.g. the stock of human captal; see Cutler and Glaeser (1997) and Borjas (1998). If resdental segregaton mples that sklled members of an ethnc group lve n the enclave, and ndvduals prmarly nteract wth members of ther own ethnc group, then dsadvantaged members such as recent mmgrants gan from lvng n the enclave. The concluson from ths bref revew s that the causal effect of lvng n an enclave s ambguous n sgn. The net effect on outcomes s thus an emprcal queston. To determne the net effect, we estmate what must be nterpreted as reduced form relatonshps between measures of labor market outcomes and, among other thngs, the sze of the enclave. We wll thus not be able to test any of the above hypotheses. However, we argue that our estmates have a causal nterpretaton. 6.2 Baselne estmates In ths secton we provde estmates of the effects of segregaton across muncpaltes n a cohort of recent mmgrants. We nvestgate to what extent the share of mmgrants (foregn ctzens) n a muncpalty, and the ethncty of these mmgrants, matter for the economc outcome of recent mmgrants. Snce mmgrants can choose muncpalty of resdence, muncpal varables cannot be assumed exogenous. Therefore, we wll make use of the settlement polcy ntroduced n 1985 to obtan nstruments for muncpal varables. In effect, we use varables pertanng to the ntal (assgned) muncpalty as nstruments for muncpal varables eght years later. Our mantaned presumpton throughout s that the placement polcy s ndependent of unobserved ndvdual characterstcs. Moreover, we assume that locaton does not have permanent effects on outcomes. We employ the followng baselne specfcaton 21

22 outcome k k k k j( t+ 8) = X (t+ 8 ) + βe ln e j( t+ 8) + βm ln m j( t+ 8 ) + βn ln n j( t+ 8) + δu j( t+ 8) + εj( t+ 8) a' (10) We focus on two outcomes log earnngs and dleness and standardze for a set of ndvdual characterstcs X, contanng gender, age, age squared, martal status, educaton, ethncty, and year of mmgraton. The outcome for ndvdual of ethnc group k s related to four muncpal varables (muncpaltes are ndexed by j): k k e j t 8) ( + denotes the sze of the ethnc group; m j ( t+ 8) = e j( t+ 8 ) the number of mmgrants of other l k l ethnctes than k; n j( t+8) the sze of the natve (.e. Swedsh-born) populaton; and u j( t+8 ) the muncpal unemployment rate (percent of the populaton aged 16 64). We ntroduce the levels of the populaton varables n logs, snce ths provdes a more flexble specfcaton than the perhaps more standard approach of usng populaton shares. 29 We are prmarly nterested n the effect of changes n the composton of the muncpal populaton,.e. changes n k e j t 8) ( + and k m j t 8) ( + holdng populaton ( p j e j + m j + n j ) constant. For nstance, the elastcty of earnngs (y) wth respect to ethnc and mmgrant concentraton, respectvely, equals k k η η m e d ln y d ln e d ln y d ln m dln p= 0 d ln p= 0 = βˆ e = βˆ m βˆ n βˆ n e n m n (11a) (11b) and analogously for the probablty of beng dle. 30 For all practcal purposes, the adjustment n (11a) s mmateral, snce ( e n ) s such a small number (0.004 on average). The adjustment n (11b) s materal, however, because m n equals on average. 31 Motvated by the dscusson n secton 4, we estmate (10) wth a varety of estmaton methods. As we have explaned, these methods dffer wth respect to the underlyng assumptons about the sortng process. To conserve space, we just report estmates of local unemployment and the elastctes defned n (11a) and (11b) here. The full set of estmates for OLS and IV are avalable n the appendx; see Table A2. Hence, we wll not 29 In our emprcal analyss, the specfcaton n terms of levels ft the data better than the specfcaton n shares. Also, the qualtatve results are smlar n both specfcatons. 30 If we want to get at the latter elastcty we need to dvde by the correspondng average probablty. 31 Summary statstcs for the muncpal varables and the ndvdual characterstcs of the sample are reported n Table A1. 22

23 comment on the coeffcents on ndvdual characterstcs. Let us just note that they are very smlar across specfcatons. Earnngs Table 3 reports the results of the basc specfcaton for earnngs. The outcome of nterest s defned condtonal on havng postve earnngs. As we go along n Table 3, we add more sophstcaton but also more structure. Column (1) reports OLS estmates where we treat the four local varables as exogenous. Column (2) gves the results of the IV (2SLS) procedure, whch uses the local varables for each mmgrant s ntal placement (n year t) as nstruments for current (year t+8) local condtons. 32 In column (3), we present the results from a control functon approach due to Garen (1984). The value of ths approach relatve to IV s that t allows E[( β β) m m ] to depend on 0 0 m. 33 Column (4), fnally, presents the result of the Heckman (1979) two-step estmator; we choose the two-step estmator, rather than the FIML estmator, snce the former s more robust to departures from bvarate normalty. Accordng to the OLS estmates we are led to beleve that ethnc and mmgrant concentraton do not matter for earnngs. Unemployment s the only local varable that s of mportance (n the statstcal sense) for earnngs. A standard devaton ncrease n ethnc concentraton s assocated wth an (nsgnfcant) earnngs gan of 1.4 percent. 34 The IV estmates, however, mply that these conclusons are premature, snce the OLS estmates on ethnc and mmgrant concentraton are downward based by a factor of three to four. 35 Interestngly, the precson of ethnc concentraton s not much affected by the IV-procedure. In ths nstance, a standard devaton rse n ethnc concentraton produces a sgnfcant earnngs ncrease of 4.2 percent. 32 Estmates where muncpal house prces were used as an addtonal nstrument are almost dentcal to those reported here. 33 It adds the assumpton that the condtonal expectatons of the ndvdual specfc error terms ( α α and β β ) can be wrtten as lnear functons of the local varables n t + 8 and t; see Card (1999). 34 The standard devaton s calculated wthn ethnc groups. 35 A Hausman test of the exogenety of the local varables decsvely rejects exogenety. The test statstc s F(4, 249) = 7.47 (degrees of freedom wthn parentheses), wth a p-value of