The Trade Effects of Skilled versus Unskilled Migration

Transcription

1 The Trade Effects of Skilled versus Unskilled Migration Peter H. Egger ETH Zurich Maximilian von Ehrlich University of Bern September 29, 2014 Douglas R. Nelson Tulane University Abstract In this paper, we assess the role of skilled versus unskilled migration for bilateral trade. Using a large data-set on bilateral skill-specific migration and a flexible novel identification strategy, the impact of different levels of skilled and unskilled immigration on the volume and structure of bilateral imports is identified in a quasi-experimental design. We find evidence of a polarized impact of skill-specific immigration on imports: highly concentrated skilled or unskilled immigrants induce higher import volumes than a balanced composition of the immigrant base. This effect turns out particularly important when institutions are weak. Regarding the structure of imports, we observe that skilled immigrants specifically add to imports in differentiated goods. Both bits of evidence are consistent with a segregation of skill-specific immigrant networks and corresponding trade patterns. Keywords: Skilled vs. unskilled immigration; Migrant networks; Bilateral trade; Quasi-randomized experiments; Generalized propensity score estimation JEL classification: C14; C21; F14; F22 We thank the participants of the CEPR Workshop on the Economics and Politics of Immigration in Turino as well as the participants of the 2nd TEMPO Conference on International Migration in Vienna for numerous helpful comments. Contact information: Egger: egger@kof.ethz.ch. Egger is also co-affiliated with CEPR, CESifo, Ifo, and WIFO; von Ehrlich: Maximilian.vonEhrlich@vwi.unibe.ch; Nelson: dnelson@tulane.edu; Nelson and von Ehrlich are co-affiliated with CESifo

2 1 Introduction Empirical work on the role of bilateral migration for bilateral trade typically relies on three paradigms: (i) that the skill structure of migration is irrelevant for the impact on trade; 1 (ii) that the functional form of the impact of migration on trade is linear or log-linear; 2 and (iii) that migration is exogenous conditional on a linear function of other determinants of bilateral trade. 3 See Gaston and Nelson (2011) for a survey of the literature on the nexus between migration and trade. A violation of (i) would result in a heterogeneous impact of migration on trade, depending on the skill-specific composition of migration, and this impact heterogeneity is masked by the approach usually adopted. A violation of (ii) would mean that the estimates of the trade-response to migration might be meaningless to the extent that marginal changes in migration might have largely different effects on trade, depending on the level of migration in the outset. A violation of (iii) would entail that bilateral migration is not exogenous conditional on a linear function of control variables, leading to inconsistent estimates of the relationship between migration and trade. Inter alia but not only, the latter results from a non-log-linear impact of the determinants of trade on trade in general (see Anderson and van Wincoop, 2003; Santos Silva and Tenreyro, 2006) and a non-log-linear impact of migration on trade in particular. This paper aims at relaxing the three aforementioned paradigms by employing a flexible econometric approach which permits skill-specific migration to be an endogenous determinant of trade while allowing for a flexible, nonparametric functional form of the relationship. This is done on the basis of a so-called generalized propensity score framework for continuous endogenous variables (treatments; skill-specific bilateral immigration in our case) and continuous outcome (bilateral imports). Exploiting this flexibility and using two large cross-sections of bilateral stocks of skilled and unskilled immigrants and bilateral import flows for 1990 and 2000, respectively, we are able to provide novel evidence on the effect of skill-specific immigration on bilateral imports. In particular, the findings suggest that a concentration on skilled or unskilled immigrants leads to a bigger response of bilateral imports 1 A number of the papers on trade and migration have considered different levels of skill and found that skilled immigrants are strongly associated with trade creation, though intermediate and low levels of skill seem to have no such relationship (see Felbermayr and Jung, 2009; Hatzigeorgiou, 2010; Javorcik et al., 2011; Felbermayr and Toubal, 2012). This work tended to assume a (log- )linear relationship between trade and skill-specific immigrantion and a random assignment (i.e., exogeneity) of immigration. 2 A handful of papers considered a non-linear functional form. See Head and Ries (1998), Wagner et al. (2002); Bryant et al. (2004) for parametric evidence and Egger et al. (2012a) for non-parametric evidence, focusing on total rather than skill-specific immigration. 3 Egger et al. (2012a) permit total immigration to be endogenous in a nonparametric framework. 2

3 than a mixed composition of the same level of total immigration. Moreover, an improvement of the institutional quality in a migrant s origin country displays the biggest positive effect on bilateral imports if, at the same time, the mix of immigrants is skill-wise relatively balanced. We rationalize these findings in a number of ways, but they are consistent, in particular, with the presence of (at least partially) segmented skilled and unskilled migration networks and their effectiveness for trade in different domains of goods. Finally, the nonlinear impact of the two types of immigration on import volume and specific categories thereof suggests that the insights gained in this paper could not easily be derived in the framework of log-linear gravity equations as often used to analyze bilateral migration stocks or flows. The remainder of the paper is organized as follows. Section 2 introduces the concept of generalized propensity score estimation with multiple (or multivariate) treatments for inference of the impact of skilled versus unskilled immigration on bilateral imports. Section 3 introduces the data used for inferences and associated descriptive statistics. Section 4 summarizes the results regarding the causal impact of bilateral skilled versus unskilled immigration on bilateral international imports, including an analysis on the structure of imports and Section 5 provides a rationalization of those findings against the background of economic theory and earlier research on the nexus between migration and trade. The last section concludes with a summary of the most important findings. 2 Econometric approach 2.1 General outline For our purposes, we need a methodology that is able to handle two determinants of bilateral trade (i.e. skilled and unskilled immigration) which are (i) continuous and endogenous, and (ii) exhibit an impact on outcome (bilateral imports) whose functional form is unknown. In many earlier empirical studies, bilateral migration may be an endogenous regressor in the model of bilateral trade for two main reasons. First, the set of determinants of bilateral trade flows is specified in a relatively parsimonious way so that an exclusion of joint determinants of bilateral migration and trade flows is likely, whereby bilateral migration is correlated with the residual. This entails a classical omitted variables bias. Second, the functional form of the relationship between bilateral migration and trade is not necessarily log-linear as is typically assumed. With a log-nonlinear functional form, log bilateral migration is again correlated with the residual in the trade model which generates endogeneity. Econometric theory offers a range of remedies to the problem of endogeneity. Of those, instrumental variable and control function estimation rely on the idea 3

4 of splitting endogenous explanatory variables into the endogenous part and the exogenous part, the former being correlated and the latter being uncorrelated with unobservable determinants of outcome (the residual; see Wooldridge, 2010). In linear models, instrumental variable and control function regressions are two sides of the same coin: while the former replace endogenous variables with their fitted values from a first-stage regression in the outcome model so that the residual term there includes the unobservable determinants of outcome as well as the endogenous part of the explanatory variable(s), the latter adds the first-stage residual(s) as explanatory variable(s) in the outcome regression in addition to the original endogenous variables. In the present context, suppose that a researcher wanted to run the regression M = W β + u, where M and W are vectors of log bilateral imports and log bilateral immigration, respectively, and u is a vector of disturbances that contains all the other determinants of bilateral imports. Obviously, β is the unknown parameter of interest. With instrumental variables and control function estimation, one assumes that immigration can be linearly decomposed in two components: W = [W u, W c ] where W u and W c are the components which are uncorrelated and correlated with u, respectively. The task is to estimate W u and W c and to either estimate the regression by plug-in instrumental variable estimation, M = Ŵuβ IV + u IV, or by control function estimation, M = W β CF + Ŵcγ CF + u CF. Notice that Ŵu is the prediction of the first-stage model for W, while Ŵc is the residual from the same regression. The term Ŵcγ CF is a control function that absorbs the asymptotic bias of the parameter of interest such that β CF is consistent. The same idea prevails for nonlinear models. An important difference between instrumental variables and control function estimation pertains to the set of instruments. Instrumental variable estimation requires at least as many instruments as there are endogenous and exogenous regressors in the outcome equation. These instruments need be exogenous (uncorrelated with all possible unobservable determinants of treatment and outcome) and they need to be relevant (be correlated with treatment) while not directly affecting outcome. With suitable instruments, the parameters of interest in the outcome equation can be estimated consistently. Control function estimation requires to condition on all relevant factors (instruments) that determine outcome and treatment simultaneously so that outcome is independent of treatment status after conditioning on the instruments. With nonlinear models, the plug-in instrumental variable estimator is not generally feasible while the control function approach is, since nonlinearities in the model do not permit simply replacing endogenous variables by their fitted values from first-stage regressions (see Terza et al., 2008). Notice that also nonparametric instrumental variables estimation would build on a control function approach (see Newey et al., 1999). The approach employed here only differs from nonparametric 4

5 instrumental variable estimation with regard to the underlying assumptions about the instruments. 2.2 Notation and formal model outline Let us denote the cross-sectional units of observation (here, country-pairs) by i = 1,..., N and economic outcome (in the present context, the value of bilateral goods imports) by M i. The goal of this paper is to determine the causal impact of skilled and unskilled immigration and their interaction on M i. For econometric identification, let us think of the levels of skilled and unskilled immigration as potentially endogenous treatments. Yet, unlike in most of the evaluation literature in econometrics, these treatments are not binary but continuous. 4 Let us denote specific potential treatment levels of skilled and unskilled migrants by s, u. Moreover, denote the lower and upper bounds of these potential treatment levels as {s 0, u 0 } and {s 1, u 1 }, respectively. The potential treatment levels are then associated with sets of potential treatment levels of S [s 0, s 1 ], U [u 0, u 1 ], respectively. 5 While potential treatment levels of migrants are denoted by lower-case letters, corresponding realized treatment levels of country-pair i will be denoted by upper-case letters, S i, U i. For each country-pair i, we may now define the set of potential outcomes in terms of a unit-level dose-response function for l = s, u as M i (l) M i (s, u) for s S, u U and the corresponding average dose-response function as µ(l) µ(s, u) = E[M i (s, u)]. We adopt a reduced-form specification to model immigration of any kind as a function of a vector of covariates, Z i, determining S i and U i. Suppose S i and U i were determined in a structural model by the respective vectors of covariates X Si and X Ui such as S i = f(x Si, δ S ) + ε Si ; U i = f(x Ui, δ U ) + ε Ui, 4 See Lee (2005) or Wooldridge (2010) for an extensive discussion of models with binary endogenous treatments. Also, see Lechner (2001) for econometric models with multiple binary endogenous treatments. Finally, see Hirano and Imbens (2005) or Imai and van Dyk (2004) for a generalization of propensity score estimation with univariate continuous treatments. The novelty of the approach adopted in this paper is the generalization of the concept of generalized propensity scores for multiple continuous treatments. 5 Obviously, the proposed generalized propensity score approach is in no way restricted to two treatments only. Egger and von Ehrlich (2013) provide a proof and a Monte Carlo study to illustrate identification issues with one, two, and more endogenous continuous treatments. The question at stake calls naturally for a bivariate approach, which is also better amenable to graphical illustration than higher-dimensional models. 5

6 where f( ) are flexible functions, δ S and δ U are unknown parameters, and ε Si and ε Ui are disturbances. Then, we could think of Z i as the joint set of exogenous regressors or independent columns in [X Si, X Ui ]. The reduced-form representation of the model may then be written as S i = f(z i, γ S ) + ν Si ; U i = f(z i, γ U ) + ν Ui (1) with γ S and γ U being unknown vectors of reduced-form parameters and ν Si and ν Ui the corresponding disturbances. We observe Z i, the continuous treatments S i and U i, and outcome M i (S i, U i ) and assume that they are suitably measurable. Denote any possible vector of exogenous covariates determining treatment by z and define the bivariate conditional joint density of s, u given z as g(s, u, z) = f Si,U i Z i (s, u z). Then, the generalized propensity score (GPS) is defined as with the property that G i = g(s i, U i, Z i ) Z i 1{S i = s, U i = u} g(s, u, Z i ). The latter states that the probability of the observed treatments being equal to some potential treatment combination {s, u} is independent of the covariates in Z i once we condition on the GPS. Accordingly, the treatment status is independent of the outcomes conditional on the GPS under the assumptions stated in Subsection 2.3. For our identification strategy, this implies that we need to condition only on one scalar, namely G i for unit i, in order to remove the selection bias in the unconditional impact of skilled and unskilled immigration on imports instead of all covariates in the vector Z i. This allows a maximum of flexibility regarding the functional form of the import response to skilled and unskilled immigration. 2.3 Assumptions For identification, we have to assume unconfoundedness (or conditional mean independence) as stated in Rosenbaum and Rubin (1983) for the binary propensity score and, more specifically, weak unconfoundedness as in Hirano and Imbens (2005) and Imai and van Dijk (2004) for the generalized propensity score with a univariate, continuous treatment. The latter means that conditional independence holds at each value of treatment without requiring joint independence of all potential outcomes. 6

7 Assumption 1. Weak unconfoundedness M i (s, u) S i, U i Z i s S, u U. In our setup, the untestable assumption of weak unconfoundedness means that, conditional on the vector of covariates Z i, the potential outcome M i is independent of the treatment status in the two treatment dimensions S i, U i. Assumption 2. Balancing of the covariates The balancing of covariates is a testable assumption, and it means that the functional form of the density of the residual treatment levels is appropriately chosen, in the sense that there is sufficient overlap between observations with different levels of treatment S i, U i but similar levels of both the covariates in Z i and the GPS, G i. Hence, for all strata of the GPS the probability of the treatments (i.e. levels of skilled and unskilled immigration) does not depend on the values of Z i. We will illustrate in Section 4.1 that the data comply with this assumption. 2.4 Implementation and parametrization The implementation of the approach proceeds in five steps. In Step 1, we estimate reduced-form parametric first-stage models determining S i and U i. The parametrization of these models is a polynomial form which is linear in parameters such that f(z i, γ S ) = Z i γ S and f(z i, γ U ) = Z i γ U in (1). Accordingly, (1) is estimated by ordinary least squares. The polynomial order of the terms included in Z i is chosen according to an information criterion (the Akaike information criterion). In Step 2, we utilize the estimated residuals from Step 1, (ν Si, ν Ui ), assume a functional form for the density function, namely bivariate normality, and compute the GPS as Ĝ i = 1 { 2π σ exp 1 [ νsi + ν Ui 2 ν ]} Si ν Ui ρ, (2) SS σ UU 1 ρ 2 2 (1 ρ 2 ) σ SS σ UU σ SS σ UU where σ SS and σ UU are the variances of the disturbances in the equation for S i and U i, respectively and ρ is the estimated covariance between those disturbances (see, e.g., Greene, 2011, for a general treatment of bivariate normals). With the treatments S i and U i measured in logarithmic terms, the normality assumption is approximately met. In Step 3, we ensure common support, which means comparing only observations with similar levels of predicted skill-specific immigration but different realizations of the treatments. The GPS represents an objective criterion to select comparable units. Hence, we restrict our analysis to the common support of Ĝi. In the binary 7

8 treatment case this is straightforward as there are only two types of observations treated and untreated while with continuous treatments we have to discretize the sample into treatment groups. As a benchmark, we split the data into four groups in S i -U i space where we denote the group an observation i belongs to by Q i {1, 2, 3, 4}. Following Hirano and Imbens (2005), Kluve et al. (2012), and, in particular, Flores et al. (2012) who all study problems with a single treatment variable, we evaluate the GPS at the median treatment level of each treatment group q, independently of whether the observation belongs to the group or not. This yields for each observation four scalar values Ĝq i, where q = 1,..., 4. Finally, we compare the distribution of Ĝq i for observations that belong to the respective group, i.e., with Q i = q and observations that do not belong to that group Q i q. An observation i satisfies the common support criterion for group q, if [ ] Ĝ q i max{min {j:qj =q}ĝq j, min {j:q q}ĝq j j }, min{max {j:q =q}ĝq j j, max {j:q q}ĝq j j }. (3) Only those observations that are comparable across all groups are kept for the analysis such that each observation has to fulfill the common support criterion (3) for all groups q simultaneously. Naturally, the common-support sample depends on the number of groups we choose for the discretization. We report the analysis with four and nine groups. 6 Step 4 is concerned with assessing the validity of the balancing assumption. As argued by Imai at al. (2008), individual tests of the balancing assumption may be misleading such that it is important to pursue different testing strategies. Hirano and Imbens (2005) suggest splitting the sample along two lines. First, use again the groups in S i -U i space as defined in Step 3. Second, split the sample of observations within each group in Ĝi-space into blocs. These blocs are determined by the GPS evaluated at the median treatment level of the respective group Ĝq i. In the benchmark specification with four groups, we use sixteen blocs such that observations in bloc 1 of Ĝq i are those with a very high propensity for receiving the median treatment intensity of group q. With an identical number of blocs per group, there is approximately the same number of observations per bloc and group. We then conduct t-tests about the equivalence of the averages in the data of each linear term of the covariates. For instance, we take observations with a similar level of GPS evaluated at the median treatment level of group q, Ĝq i, and compare the covariates between observations Q i = q and Q i q. This is done for all groups and blocs and provides information on whether the covariates differ significantly across groups once we condition on the GPS. Hence, in the conditional comparison, we estimate 6 In the working paper version of this paper we report also results for a 16 group common-support sample (see Egger et al., 2012b). 8

9 16 t-values per group and covariate and then calculate the average over blocs. Of course, the sample sizes differ slightly across groups and blocs so that test statistics should be adjusted. This is done by weighting the data properly by the number of observations used in each bloc. A sufficient balancing of the covariates is reached, if these t-tests turn out insignificant. Our second test of the balancing assumption follows Imai and van Dyk (2004) (this test is implemented in a similar fashion by Kluve et al., 2012). For this, we regress each covariate on the treatment variables with and without conditioning on the distribution of the GPS. In the conditional models we evaluate the GPS at the quartiles of S i and U i. The test essentially represents a more flexible version of the aforementioned approach by Hirano and Imbens (2005). Ideally, if the GPS absorbs all information relevant for treatment assignment, the covariates Z i are uncorrelated with S i and U i, once we condition on the distribution of the GPS. The last step pertains to estimating parametric and nonparametric regressions of the second-stage, outcome equation. The parametric model specifies E(M i S i, U i, G i ) = H i λ as a function which is linear in parameters. The vector H i depends on a polynomial function of the terms {S i, U i, G i } only (apart from a constant). As with the first-stage models in Step 1, the polynomial order is determined by the Akaike information criterion and the chosen regression approach is ordinary least squares. The parameter estimates ˆλ from this regression are neither interpretable nor of interest themselves. The reason is that E(M i S i, U i, G i ) represents only a so-called unit-level dose-response function, where the marginal effect of {S i, U i } depends on the level of Z i and G i. What one would like to know is an average dose-response function, where the marginal effect of treatment is independent of Z i. The latter can be obtained as follows. First, one considers the range of treatment levels of interest, [s 0, s 1 ] and [u 0, u 1 ]. Second, one discretizes these intervals into bins, calculates all possible values S i s and U i u and employs these instead of ˆν Si and ˆν Ui in the GPS in Step 2. 7 Suppose one distinguishes 40 bins of equal size in [s 0, s 1 ] and [u 0, u 1 ] each. This leads to 40 2 = 1, 600 combinations of {s, u} and, hence, as many different values of Ĝ i (s, u) as well as of E(M i S i s, U i u, Ĝi(s, u)). One then computes the average of E(M i S i s, U i u, Ĝi(s, u)) for each of the 1, 600 combinations of {s, u} across all observations N. Each one of the averages is now independent of the country-pair-specific level of Z i. The average dose-response function in this example is the relationship between all 1, 600 levels {s, u} and average associated (potential) outcome (log bilateral imports), E[M(s, u)]. Alternatively, one may model the unit-level dose-response function by a nonparametric estimator. A flexible approach to estimate such a model is to employ a 7 Note that one keeps ˆσ SS, ˆσ UU, and ˆρ constant at their originally estimated levels for this exercise. 9

10 multivariate local linear estimator based on an Gaussian product kernel with bandwidth h (see Fan and Gijbels, 1996). We obtain predictions for E[M(s, u)] from the estimate ˆα of the local linear estimator: N arg min [M i α λ 1 (S i s) λ 2 (U i u) λ 3 (Ĝi Ĝi(s, u))] 2 K h, (4) α,λ i=1 where E[M(u, s)] = ˆα and K h = K h (U i u)k h (S i s)k h (Ĝi Ĝi(s, u)). Obviously, the critical parameter for the nonparametric estimation of the unit-level dose-response function is the bandwidth. Conventional procedures for rule-of-thumb bandwidth selection are constrained to univariate local polynomial regression (see Fan and Gijbels, 1996). Therefore, we implement a leave-one-out cross-validation procedure of optimal bandwidth selection. 8 Both the parametric polynomial and even more so the nonparametric versions of the avg. dose-response function are relatively flexible and they can best be visualized graphically by way of three-dimensional plots. The confidence bounds of the parametric and nonparametric estimators can be estimated by bootstrapping all estimation steps together, including the common support restriction, the firststage estimation of Ĝ i (s, u), and the second-stage regressions (according to Efron and Tibshirani, 1993, 200 replications should be sufficient for this). Apart from the avg. dose-response function, the avg. treatment-effect function may be of interest. With two continuous, endogenous potential treatment levels s and u, there are two such avg. treatment-effect functions, and they correspond to de[m(s, u)]/ds and de[m(s, u)]/du. They can be estimated by taking the respective derivatives of the unit dose-response function and subsequent averaging over all observations. Again, the confidence intervals of the avg. treatment-effect functions have to be bootstrapped. 3 Data and descriptive statistics The variables entering our analysis encompass direct drivers of bilateral imports as well as direct drivers of skilled and unskilled bilateral immigration. All variables entering the analysis may be broadly grouped into dependent variables and independent variables. 8 Due to the immense computational intensity of a three dimensional cross-validation we apply the same optimal bandwidth in all three dimensions. This comes only at a small loss of precision since we standardize all three variables before implementing the cross-validation and estimating the local linear regressions by their respective standard deviation. 10

11 3.1 Dependent variables The dependent variables are bilateral import flows (which we also refer to as outcome) and bilateral stocks of skilled and unskilled immigrants (which we refer to as treatments). The goal will be to determine the impact of (endogenous) bilateral skilled and unskilled immigration treatments on bilateral import outcome. Bilateral skill-specific immigration data are available for the years 1990 and 2000 from Docquier et al. (2009) for OECD destination countries and from Artuc et al. (2014) for non-oecd destination countries. This dictates the sample coverage of the data in the empirical analysis. In the benchmark, we define skilled immigrants as those with at least a secondary level of education. Note that our results are invariant to an alternative aggregation with skilled immigrants characterized by at least tertiary education. As the outcome, we use data on bilateral imports from the United Nations Comtrade Database for the average year within the period and , respectively. Accordingly, we consider immigration and imports for two time periods. In order to determine the effect of immigration on the structure of imports we distinguish between homogeneous and differentiated goods according to the classification by Rauch (1999). We assign sectors to the respective classifications using the Standard International Trade Classification (SITC) three-digit data. This obtains an additional dependent variable measuring the ratio between differentiated and homogeneous goods imports denoted by M D i /M H i. Table 1 here Altogether, we cover 98 countries of origin and 29 (65) OECD (Non-OECD) countries of residence of migrants and trade between those countries in our analysis (see the Appendix for a list). Table 1 provides some descriptive features of the data on the dependent variables covered in our analysis. We measure immigration as well as imports in logarithmic terms such that our empirical analysis captures only country pairs characterized by positive values of both bilateral imports and immigration of skilled and unskilled. 3.2 Independent variables As described in the previous section, we have to employ independent variables determinants of bilateral import flows as well as skilled and unskilled immigration stocks as elements in the vector Z i in order to remove the selection bias in an assessment of the impact of the two types of immigration S i and U i on bilateral imports M i of country-pair i (see Felbermayr and Jung, 2009; Mayda, 2010; Felbermayr et al., 2010; Felbermayr and Toubal, 2012; for important determinants of bilateral migration). 11

12 Generally, the vector of observables Z i includes both continuous and multi-valued discrete variables for both exporters/countries of origin and importers/countries of residence in a flexible 3rd-order polynomial functional form and binary variables in a linear functional form. 9 Specifically, we include a parametric (polynomial) function of exporter/origin- and importer/residence-specific log GDP, log GDP per capita, and log population to account for effects of economic market size, per-capita income, and population size in a fairly flexible way. These variables are taken from the World Bank s World Development Indicators Moreover, we control for origin- and residence-country GINI coefficients, unemployment rates, life expectancies, fertility rates, literacy rates, and real exchange rates between residence and origin countries as measures of unemployment risk, inequality, and economic well-being beyond per-capita income. These variables come from the World Bank s World Development Indicators 2009, the CIA World Factbook, and United Nations Educational, Scientific, and Cultural Organization (UNESCO). Furthermore, we control for bilateral distance between residence and origin countries as a continuous pair-i-specific geographical determinant of immigration and imports as well as for common language, colonial relationship, common religion, goods-trade-agreement membership, services-trade-agreement membership, bilateral migration impediments, OECD membership, and membership in the Warsaw Pact as pair-specific binary geographical, cultural, political, and economic control variables. All geographical variables are based on information from the Centre d Études Prospectives et d Informations Internationales (CEPII) geographical database. Trade agreement indicator variables as well as data on migration impediments are based on information from the World Trade Organization (WTO). Also, we include a number of other control variables for both the country of origin and residence, capturing political and institutional factors that may be relevant for both migration and trade. First of all, we include measures of the originand residence-specific number of armed conflicts from the Peace Research Institute at Oslo, the corruption perception index from Transparency International, political freedom (the POLITY-IV index) from the Center for Systemic Peace (see Marshall et al., 2011), and a number of measures determining labor market features from the International Labor Organization (ILO): the bargaining power index, the working condition index, the worker discrimination index, and the child labor index. Second, we capture several dimensions of institutional quality from the World Bank s Governance Indicators. 10 All covariates in Z i are measured prior to the endogenous 9 Binary variables enter linearly in order to keep the degree of multicollinearity of the model reasonably low. In general, we will use acronyms of the variables for the sake of brevity in tables. These acronyms are defined in Table While we only include country-pair-specific variables, it should be noted that the results are 12

13 variables M i, S i, and U i in each of the two periods. 11 Table 2 here Table 2 provides information on moments of the data for all (first-order) independent variables akin to both blocs of Table 1. 4 Results 4.1 Multivariate GPS estimation, common support, and the balancing property We include all main effects of the covariates listed in Table 2 together with quadratic and cubic terms of all non-binary regressors in the regressions explaining S i and U i. 12 Altogether, there are 104 explanatory variables in the two equations. We report parameter estimates and standard errors for the 3rd-order polynomial reduced-form model specifications for S i and U i based on 104 regressors. The coefficient signs are impossible to interpret due to the nonlinear form of the first-stage regressions. However, it is generally the case with propensity score estimation that the specific parameters and marginal effects of covariates in the treatment equation(s) are of subordinate interest. For completeness, the parameters and standard errors of the regressors are summarized in Table 3. What matters is that the joint contribution of the covariates to the variances of S i and U i is decently large, and that they are balanced (for given estimated values of the GPS). Table 3 here robust to the inclusion of third-country effects with regard to all of the origin- and residencecountry-specific covariates on treatment (see the working paper version of this manuscript: Egger et al., 2012b). The inclusion of the latter would serve the purpose of accounting for interdependence of origin and residence countries in supplying and attracting migrants (see Anderson, 2011). 11 Trade agreement membership, corruption perception, and ILO s indices on labor market conditions are measured in 1990 and OECD membership, geographical, cultural, historical variables, and bilateral migration impediments are time-invariant. All other elements of Z i are measured as averages over and for the two periods covered, respectively. 12 The 3rd-order polynomial model specification is not arbitrary. We did a selection of the optimal order of the polynomial on grounds of the Akaike information criterion searching across models which involve 1st-order up to 5th-order polynomials. Based on this search, we selected the 3rdorder polynomial version as the preferred one, since no substantial decrease in the Akaike criterion could be achieved when choosing a higher-order polynomial. 13

14 Indeed, the regressions do feature a decent predictive power with adjusted R 2 s of about 0.72 and 0.80 for S i and U i, respectively. Hence, the covariates are jointly highly relevant. Models based on 1st-order or 2nd-order would have achieved lower tuples of adjusted R 2 s of (0.64, 0.72) and (0.70, 0.76), respectively. Based on the estimates in Table 3, one may compute the GPS explicitly by assuming bivariate normality of the disturbances as in (2). With the treatments measured in logarithmic terms, the normality assumption is approximately met. What remains to be enforced is common support of the GPS as in Step 3 in Section 2.4, and what ought to be checked is whether the common support assumption is tenable as outlined in Step 4 in Section 2.4. We do this for four and nine groups (called Q i above), respectively, in S i -U i -space and with 16 and 8 blocs, respectively. Enforcing common support with four and nine groups leads to samples of 2,525 and 1,352 observations, respectively. We illustrate the distribution of t-statistics about the equivalence of the averages of each linear term of the covariates in Table 3 across groups and illustrate them in Figures 1 and 2 for four and nine groups, respectively, where Panels A refer to the unconditional covariate comparison and Panels B to the comparison conditioning on the GPS. Figures 1 and 2 and Table 4 here Two insights can be gained from an inspection of the two figures. First, the histogram plots illustrate that a large mass (namely 60% and 47% for the four and nine group common support samples) of the t-values of unconditional comparisons lie outside of the [ 2.576, ] interval which (approximately) indicates significance levels of less than one percent. When taking t-statistics in absolute terms, the interquartile ranges amount to [1.68, 6.46] and [1.18, 4.35] with four and nine groups, respectively. By way of contrast, only a very small mass of the distribution of t- values of conditional comparisons (on blocs of the GPS) lie outside of that interval (0.5 percent of the t-values are bigger than in absolute terms). The interquartile range of absolute t-values for the conditional comparisons amount to [0.26, 0.80] and [0.22, 0.71]. This is evident from the much more narrow-waisted distribution of conditional-comparison t-statistics around zero relative to the unconditionalcomparison t-statistics in Panels A and B of Figures 1 and 2. Second, the mean and median values of the absolute t-statistics which we report at the bottom of the two figures are much smaller once we condition on the GPS. For instance, among the unconditional-comparison t-statistics, the average absolute values are 4.86 and 3.32 and the median values are 3.30 and Among the conditional-comparison t-statistics, the average absolute values are 0.60 and 0.54 and the median values are 0.48 and This illustrates that conditioning on the GPS is extremely powerful in the data, independent on how stringent a common support we require. Conditioning 14

15 on the GPS and enforcing common support substantially raises the comparability of country pairs in the dimensions of interest. Hence, we hypothesize that there is little chance that the observable variables included in the empirical model explaining S i and U i confound the impact of S i and U i on bilateral imports, M i. The alternative balancing test based on Imai and van Dyk (2004) and Kluve et al. (2012) supports this conclusion, as can be seen from Table 4. The unconditional models feature significant correlations between treatments and covariates where the mean t-statistic ranges from 3.09 to 5.10 for the two common support samples and treatments S and U. Hence, ex ante there is a severe selection issue. In contrast, after conditioning on the GPS, the reduction of the t-values is remarkable as the maximum t-statistic across all conditional models is 2.14 and the mean over all covariates ranges from 0.15 to 0.60 for the conditional models. These results confirm that the control function performs well in solving the selection issue with regard to observable information. Hence, we may proceed to estimate the dose-response function by means of parametric and nonparametric estimators as outlined in Step 5 of Section Parametric estimates of the multivariate dose-response and treatment-effect functions Utilizing the GPS as a compact (scalar-function) balancing score to reduce dramatically the endogeneity bias of S i and U i in determining M i by invoking the underlying assumptions, we propose the parametric second-stage regression E(M i S i, U i, G i ) = H i λ H i = [1, S i, U i, S 2 i, U 2 i, S i U i, Ĝi, Ĝ2 i, S i Ĝ i, U i Ĝ i, S 2 i Ĝ2 i, U 2 i Ĝ2 i ]. (5) This regression serves to predict the unit dose-response function. The parameters are estimated by ordinary least squares and the standard errors are estimated by a bloc-bootstrap procedure (with 200 replications) in order to respect two features: first, that each unit i is observed in two years (1990 and 2000) so that the variancecovariance matrix may have i-specific blocs and, second, that (5) involves estimates Ĝ i rather than true GPS scores G i. We ensure the common support criterion within each bootstrap replication such that the sample size may vary slightly across bloc bootstrap draws. The regression results corresponding to (5) are summarized in Table 5. Table 5 here 15

16 The point estimates of the second-stage regression are not immediately informative with regard to the causal effect of immigration on imports. However, the significance of the terms provides information on the severeness of the endogeneity issue and the performance of the propensity score in absorbing information that determines treatment and outcome. The main effects and interactive terms involving Ĝ i are jointly highly significant (the F-statistic on all terms involving Ĝi is 14.98), which is a strong indication of selectivity across different levels of S i and U i. Hence, the GPS is relevant and helps reducing the bias of the estimated response of (log) bilateral imports (M i ) to changes in (log) bilateral immigration of the skilled (S i ) and the unskilled (U i ). Controlling for observable information affecting selection into treatment, we still observe that skilled and unskilled immigration have a substantial impact on imports as is evident from the joint significance of the S and U terms. In order to judge the direction and the magnitude of the effects, we compute the avg. dose-response function for a grid in s u-space which is governed by the observed range [min(u i ); max(u i )], [min(s i ); max(s i )]. We utilize the parameters summarized in Table 5 to compute E[M(s, u)] = 1 N N (h iˆλ) (6) h i = [1, s, u, s 2, u 2, su, Ĝi(u, s), Ĝ2 i (u, s), sĝi(u, s), uĝi(u, s), s 2 Ĝ 2 i (u, s), u 2 Ĝ 2 i (u, s)]. i=1 In Figure 3 we illustrate the avg. dose-response function as computed from (6) based on four groups and note that the result for nine groups is very similar. The plot contains three areas which indicate significant positive (in blue), insignificant positive (light blue), and insignificant negative (yellow) responses. The significance statements in the figure are based on 90% confidence intervals. Figures 3 and 4 here A key insight from Figure 3 is that import volume is not maximized at the diagonal where skilled and unskilled immigration reach similar levels but at the edges where the immigration stock is either dominated by skilled or by unskilled individuals. Hence, our results suggest that bilateral import flows are stimulated mostly by homogeneous immigrant communities while a heterogeneous mix between skilled and unskilled immigrants yields ceteris paribus a lower import volume. According to the results, there is a statistically significant (at 5%) positive level of log imports for almost any form of bilateral immigration. However, Figure 3 suggests that 16

17 the import-maximizing immigration treatment corresponds to a polarization of immigrant types, irrespective of whether unskilled or skilled immigration dominates. Note that the range of observed skilled-unskilled immigration combinations does not support all of the cells in the figures. However, the polarization result is found also in the S i -U i -subspace that is supported by the data. The polarization result becomes even more evident from inspecting the so-called avg. treatment-effect functions plotted in Figure 4. Dark blue areas correspond to significant positive and red to significant negative predictions while light blue and yellow mark insignificant positive and negative predictions, respectively. The treatment-effect functions represent the partial derivatives of the dose-response function with respect to the two types of treatment. Panels A and B in Figure 4 correspond to de[m(s, u)]/du and de[m(s, u)]/ds, respectively. Consider the avg. treatment-effect function of unskilled immigrants: starting from approximately the diagonal in Panel A of the figure, a marginal increase in u yields a significantly positive increase in the treatment effect on import volume. In contrast, the marginal effect of an increase in unskilled immigration u is insignificant or even negative if the country pair s point of origin is skewed towards relatively more skilled than unskilled immigrants. The reverse holds true if we consider the treatment effect for skilled immigration de[m(s, u)]/ds in Panel B: a marginal skilled immigrant induces imports only if a country pair has a U i -S i combination with relatively more skilled than unskilled immigrants. Hence, we find evidence for a polarized impact of skill-specific immigration on imports at the diagonal of the skilled-unskilled immigration space. Moreover, we can reject the null hypothesis of a positive marginal effect of skilled immigration in areas where unskilled immigration dominates while we can reject it for unskilled immigration in areas with predominantly skilled immigration levels. 4.3 Nonparametric estimates of the multivariate dose-response and treatment-effect functions The above results may depend to some extend on the functional form assumptions in (5). We performed sensitivity checks with a number of higher order polynomials which confirm our polarization result. Yet, an ultimate test requires a fully nonparametric specification of the dose-response and treatment-effect functions as described in (4) and in the implementation Step 5 in Section 2.4. The nonparameric estimates of the avg. dose-response and the avg. treatment-effect functions are illustrated in Figures 5 and 6, again using the common support sample with four groups as a reference point and noting that results are robust to using nine groups. 17

18 Figures 5 and 6 here In contrast to the parametric results in Section 4.2 the estimated surfaces display a higher degree of nonlinearity but the general v-shape around the diagonal is found again and is clearly visible. Bilateral imports are maximized at either a high concentration of skilled or unskilled immigrants while a balanced mix of the two groups yields a lower level of bilateral imports. A noticeable difference to Figure 3 is the asymmetry of the nonparametric surfaces at the edges with full polarization. According to the nonparametric avg. dose-response functions, a concentration of high-skilled immigrants leads to relatively more imports than a concentration of unskilled immigrants, but we estimate for either of them a higher level of imports than at the diagonal. For the nonparametric avg. treatment-effect functions illustrated in Figure 6 we employ a local quadratic estimator as it is generally advisable to choose an odd difference between the polynomial order of the local regressions and the degree of the derivative (this is preferable in terms of bias reduction). It is evident that positive (blue) treatment effects with regard to unskilled immigration de[m(s, u)]/du are confined to the areas with relatively more unskilled than skilled immigrants. On the other side of the diagonal, we estimate mainly negative (yellow or red) values for de[m(s, u)]/du. A similar pattern evolves for the treatment effect with regard to skilled immigration de[m(s, u)]/ds: positive (blue) treatment effects are confined to s u combinations with a dominating stock of skilled immigrants. As a general feature, the nonparametric estimates come at a lower degree of precision in terms of their confidence intervals. However, the nonparametric avg. dose-response and treatment-effect functions confirm the parametric findings: an immigrant mix which is skewed towards either the skilled or the unskilled maximizes bilateral imports as compared to a balanced mix of skilled and unskilled immigrants. 4.4 Immigration effects on the structure of imports A question of interest to the matter is whether the composition of immigrants has consequences for the composition of bilateral imports. Are the effects of skilled and unskilled immigrants driven by specific product groups? Since an investigation at the very disaggregated product level is not feasible for reasons of presentation, we resort to an analysis at the level of aggregates of product classes. A widely accepted way of grouping products is the one proposed by Rauch (1999; the so-called Rauch classification), which distinguishes between differentiated products, homogeneous products, and an intermediate category. Hence, each single observation on aggregate bilateral import flows may be split into the corresponding three sub-aggregates. 18

19 Rauch (1999) offers two classification schemes, one dubbed liberal and one conservative. Since the results turn out virtually identical for the two schemes, let us focus on the liberal classification, here. Specifically, we consider the ratio between differentiated and homogeneous bilateral imports Mi D /Mi H as an alternative outcome to total import volume. Since the nonparametric estimation strategy described above is preferable in terms of flexibility, we focus the analysis in this section to nonparametric estimates based on the four-group common-support sample. 13 We illustrate the corresponding finding in Figure 7. Figure 7 here The figure suggests that the ratio between differentiated and homogeneous goods imports, M D i /M H i, is unambiguously increasing in the stock of skilled immigrants, while for unskilled immigrants the opposite is true. Hence, the two types of immigrants seem to stimulate different types of bilateral imports in terms of Rauch s categories. While the aggregate level of imports ceteris paribus rises with a high concentration of either immigrant type, a high level of bilateral imports in differentiated goods seems to be facilitated by a high concentration of skilled immigrants. Hence, the pattern identified in Figure 7 is consistent with the results in Rauch (1999), Rauch and Trindade (2002), Briant et al. (2009), Tai (2009), Felbermayr et al. (2010), Peri and Requena-Silvente (2010), Felbermayr and Toubal (2012), and Genc et al. (2012), who found that the share of high skilled migrants is strongly associated with imports of differentiated goods and goods traded on organized markets, but less so with goods associated with reference prices. 4.5 Effects of migration to OECD versus non-oecd resident countries on imports Our benchmark estimates focus on migration from OECD and non-oecd countries of origin to OECD destination countries. Recent research by Artuc et al. (2014) made bilateral information on migration among non-oecd countries as well as migration from OECD to non-oecd countries available. While South-South migration is as common as South-North migration, it is largely driven by other factors. As pointed out by Ratha and Shaw (2007) South-South migration tends to take place predominantly between countries with contiguous borders and with relatively small differences in income. In many instances it occurs not for economic but for political reasons, especially due to conflicts and wars. The latter shows in the fact that 13 Note that parametric estimates along the lines described in Section 4.2 yield very similar results. The findings are also robust to using the nine-group common-support sample. 19

20 most refugees and asylum seekers migrate among developing countries. Accordingly, the first-stage coefficients on the observable determinants of bilateral immigration in our estimation approach should be allowed to differ for South-South and other migration instances. Moreover, we expect the measurement error to be more pronounced with data reported from developing (southern) countries of residence, no matter of whether North-South or South-South migration is concerned. Therefore, we decided to estimate the effects separately for the sample of non-oecd residence countries. We again follow the approach outlined in Section 2 in estimating the avg. dose-response and treatment-effect functions. For the sake of brevity, we only report the avg. dose-response function, though, and provide it in Figure 8. Figure 8 here It is apparent from the figure that the predicted import response to combinations of skilled and unskilled immigrants for non-oecd residence countries is very similar to the benchmark sample based on OECD residence countries. Hence, the polarized effect of immigration by skill-type on imports appears to be generic and independent of the sample. 4.6 Effects of immigration as well as of emigration on imports A further extension concerns the role of bilateral emigration. Suppose that emigrants from one country to another one engage in business activity with their country of origin. This may give rise to an additional import response which could bias the insights gained in the benchmark analysis which focused on immigration. In order to address this point, we specify total emigration as third endogenous treatment beyond S i and U i as a function of the same determinants as in Table 3, estimate the GPS, G i, based on a trivariate normal, using the residuals of the three first-stage models, impose the common support condition as before, and estimate the avg. doseresponse function from a second stage model which is based on three treatments: S i, U i, and log bilateral emigration (apart from the GPS). Notice that this avg. dose-response function adjusts for potential differences of emigration across country pairs, while the one in the outset did not. However, since we are not primarily interested in the (in comparison to immigration relatively more indirect) effect of emigration on imports per se, integrate its effect out (i.e., we average over the level of emigration) and display the results for the avg. dose-response function in S i -U i - space as before. Clearly, while such an analysis is not informative about the impact of emigration as such, the avg. dose-response function for skill-specific immigration 20

21 may not be biased due to an omission of emigration. We display the corresponding avg. dose-response function in Figure 9. Figure 9 here Two observations stand out. First, the shape of the avg. dose-response function and, accordingly, the polarization result remains unaffected by the consideration of bilateral emigration. Second, computing the difference of the predicted import responses with and without controlling for emigration reveals that the effect increases slightly on average. This suggests that the positive correlation between bilateral immigration and emigration leads to some mis-attribution of import responses to immigration, where emigration is the cause. However, the magnitude of the potential mis-attribution does not vary with skilled relative to unskilled immigration so that the qualitative pattern of effects is the same as in the outset. 5 An interpretation of the polarization result of immigration on imports and some further evidence on the role of institutions This section s aim is to provide one interpretation of the paper s main finding of a by-skill-type polarized effect of immigration on imports based on the literature on the the structure and the functioning of social networks. The argument that immigration affects trade through network effects as such is not new (Rauch, 2001). Distinct aspects of those networks discussed in that literature are two: one is the role of networks in mediating the economic relationship between two dense networks (representing the populations of residence and source countries); and the other one is the internal structure of the networks that do the spanning. Following Burt (2005, 2009), network research often refers to these two dimensions as brokerage and closure. Brokerage involves building connections across groups to increase exposure to diverse opinion and practice. Brokerage is associated with growth and innovation. Closure involves strengthening connections within a group to focus the group on a limited set of opinions and practice. Closure is associated with trust and alignment, ultimately enhancing efficiency. Closure plays a particularly central role in dealing with institutional failures and asymmetric information problems. Work on closure starts from the problems of contracting in certain types of goods or environments. The idea is that, in the absence of relatively complete contracts and/or effective legal environments, the risk of loss due to opportunistic behavior is sufficiently high that 21

22 many mutually beneficial contracts would not be made in the absence of some alternative source of assurance. These factors have been emphasized in explaining the role of ethnic networks and diasporas in the organization of trade across political jurisdictions or, more generally, in the absence of effective protection of contractual/property rights (Polanyi, 1957, 1968; Geertz, 1963, 1978; Cohen, 1969, 1971; Bonacich, 1973; Curtin, 1984). Accordingly, disasporas have been found to play a particularly strong role in mediating trade links in the context of institutional problems and asymmetric information problems. Moreover, earlier work suggests that the internal structure (ethnic homogeneity or, more broadly, social proximity) of groups plays an important role in dealing with contracting problems (see Blalock, 1967; Bonacich, 1973; Landa, 1981, 1994; Greif, 1989, 1991, 2006; Iyer and Jon, 1999; Burt, 2005, 2009). We argue that the pronounced effect of polarized bilateral immigration on bilateral imports may be suggestive of a decisive role played by skill-specific networks and, in particular, the formation of stronger networks among immigrants with more homogeneous skills than heterogeneous ones. Immigrants are naturally brokers between their source country and the host country in which they settle. That is, they carry information about commodities available in the source country to consumers in the host country, and vice versa. This will tend to increase demand for products of the immigrant home market in their new host country (see Gould, 1994, for an early argument along those lines; the results in Head and Ries, 1998; Wagner et al., 2002; Bryant et al., 2004; and Egger et al., 2012a; are consistent with this view). In particular, existing evidence on a diminishing import-fostering effect of rising immigration and the larger role of immigrants for differentiated (arguably more information-intensive) goods trade is consistent with the information link effect of immigrant networks. Our results suggest that, by comparison to migrant communities with a wide range of skills, migrations characterized by a strong concentration of a given skill group will form more effective networks, generate better bridges, and thus produce a stronger link between migration and trade. Networks play an essential role in the location choices of emigrants. Because networks reduce the costs of migration, it is well-known that immigrants drawn from well-defined sending regions tend to go to equally well-defined locations in the receiving country (see, e.g., Taylor, 1986; Massey et al., 1987; Winters et al., 2001; McKenzie and Rapoport, 2007). Thus, when we consider the pre-existing social bonds between any group of migrants, the claim that similarity of education creates closer bonds gains considerable plausibility. The finding that the effect of immigration on imports is stronger for immigration flows made up of people with relatively homogeneous skills is consistent with the hypothesis that the degree of closure within an immigrant community drives the 22

23 network effects of trade. The role of closure i.e. reputation-building and punishment via the exclusion from group benefits should be particularly pronounced for countries characterized by poor enforcement of contract and property rights. A high degree of closure within an immigrant community permits information exchange and bonding that thereby substitutes to some extend for the low quality of institutions. The homogeneity of observable traits such as education represents a proxy for the degree of closure. Accordingly, we should expect that homogeneous migrant groups exert a stronger effect on imports from countries with low quality of institutions compared to countries where the enforcement of contracts is guaranteed by an efficient legal system. In order to assess this hypothesis, we analyze whether the polarization result varies with the quality of institutions as measured by indicators on the control of corruption, government effectiveness, political stability, rule of law, regulatory quality, and absence of violence which are contained in the World Bank s World Development Indicators database. Specifically, we proceed as follows to shed light on the role of institutions in conjunction with the avg. dose-response function in S i -U i - space. We raise each one of the just-mentioned six measures of institutional quality simultaneously. Using the estimated first-stage-regression parameters, we change the residuals accordingly and give the change in institutional quality the interpretation of an exogenous, random shock. This results in an alternative level of the GPS. The latter and also all potential skilled and unskilled immigration levels, {s, u}, together with these random shocks from the first stage are used to compute predictions based on the original second-stage regression parameters but using counterfactual values associated with raised institutional quality. Based on these, we compute the predictions for an alternative counterfactual avg. dose-response function, associated with the improved institutional quality. We display the difference between this counterfactual avg. dose-response function and the original one in Figure 10. Figure 10 here The figure is capable of informing us about where i.e., for which combination of potential skilled and unskilled immigration the impact of higher institutional quality on the role of skill-specific immigration for imports is biggest and where it is smallest. In the figure, blue color indicates s u combinations where the predicted import response is increased whereas red color indicates s u combinations where the predicted import response is reduced due to the higher level of institutional quality. The figure suggests that institutional quality has the biggest beneficial impact for country pairs with a relatively balanced composition of skilled and unskilled immigrants, while it tends to have a negligible and sometimes negative impact where the composition is polarized. This suggests that homogeneous groups of immigrants 23

24 stimulate imports particularly in cases where institutions are weak in their origin countries. This is consistent with the aforementioned closure-effect hypothesis of immigrant networks. 6 Conclusions This paper assesses the role of skilled versus unskilled immigration for bilateral imports in a large data-set of country pairs. Flexible parametric and nonparametric reduced-form models are postulated, where the stocks of skilled and unskilled migrants at the country-pair level are determined as endogenous continuous treatments. By invoking conditional mean independence and weak unconfoundedness, the impact of different levels of skilled and unskilled immigration on the volume and structure of bilateral imports is assessed in a quasi-experimental design. This is accomplished through a generalized estimation procedure for an assessment of causal effects of univariate continuous treatments on outcome. Two sets of results from this analysis stand out. First, we find evidence of a polarized impact of skill-specific immigration on imports: highly concentrated skilled or unskilled migrants induce higher import volumes than a balanced, more homogeneous composition of the immigrant base. Second, while a polarization of skilled migrants seems to foster primarily differentiated goods trade, a polarization of unskilled migrants mainly stimulates homogeneous goods trade. Either piece of evidence is consistent with a segregation of skill-specific immigrant networks and corresponding trade patterns. XXX THIS PART READS A BIT STRANGE That a polarization of migrants in the skill dimension stimulates trade is in line with this argument per se. Moreover, our results are consistent with the hypothesis that social ties generated in homogenous immigrant communities facilitate contracting and thereby can substitute for weak institutions in the migrants countries of origin. XXX That polarizations towards skilled versus unskilled migrants lead to a polarization of trade towards differentiated versus homogeneous goods suggests that the nature of networks and of information in different migrant groups are skill-specific. XXX SHALL WE INCLUDE A SENTENCE ON skill specific migration policy and trade liberalization???...xxx References Anderson, J.E., The gravity model. Annual Review of Economics 3, Anderson, J.E., van Wincoop, E., Gravity with gravitas: A Solution to the border puzzle. American Economic Review 93,

25 Artuc, E., Docquier, F., Ozden, C., Parsons, C A global asessment of human capital mobility: The role of non-oecd destinations. World Development, forthcoming. Blalock, H.M., Toward a theory of minority-group relations. Wiley, New York. Bonacich, E., Theory of middleman minorities. American Sociological Review 38, Briant, A., Combes, P.-P., Lafourcade, M., Product complexity, quality of institutions and the pro-trade effect of immigrants. Paris School of Economics Working Paper # Bryant, J., Genç, M., Law, D., Trade and migration to new zealand. New Zealand Treasury, Working Paper Series #04/18. Burt, R.S., Brokerage and closure: An introduction to social capital. Oxford University Press, Oxford. Burt, R.S., Network duality of social capital, in: Bartkus, V.O., Davis, J.H. (Eds.), Social capital: Reaching out, reaching in. Edward Elgar, Cheltenham, UK, Cohen, A., Custom and politics in urban africa: A study of Hausa migrants in Yoruba towns, new ed. Routledge & Kegan Paul, London. Cohen, A., Cultural strategies in the organization of trading diasporas, in: Meillassoux, C. (Ed.), The development of indigenous trade and markets in West Africa. Oxford University Press, Oxford, Curtin, P.D., Cross-cultural trade in world history. Cambridge University Press, Cambridge Cambridgeshire; New York. Docquier, F., Lowell, B.L., Marfouk, A A gendered assessment of highly skilled emigration. Population and Development Review 35, Efron, B., Tibshirani, R.J., An introduction to the bootstrap. Chapman & Hall/CRC Monographs on Statistics & Applied Probability; London. Egger, P.H., Ehrlich, M. v., Generalized propensity scores for multiple continuous treatment variables. Economics Letters 119, Egger, P.H., Ehrlich, M. v., Nelson, D.R., 2012a. Migration and trade. The World Economy 35, Egger, P.H., Ehrlich, M. v., Nelson, D.R., 2012b. The trade effects of skilled versus unskilled Migration. CEPR Discussion Paper Fan, J., Gijbels, I., 1996., Local polynomial modelling and its applications. Chapman and Hall; London. 25

26 Felbermayr, G.J., Jung, B., The pro-trade effect of the brain drain: Sorting out confounding factors. Economics Letters 104, Felbermayr, G., Jung, B., Toubal, F., Ethnic networks, information, and international trade: Revisiting the evidence. Annales d Economie et de Statistique, Felbermayr, G.J., Toubal, F., Revisiting the trade-migration nexus: Evidence from new OECD data. World Development 40, Flores, C..A, Gonzalez, A., and Neumann, T.C., Estimating the effects of length of exposure to instruction in a training program: The case of job corps. Review of Economics and Statistics 94, Gaston, N., Nelson, D.R., International migration, in: Bernhofen, D.M., Falvey, R., Greenaway, D., Kreickemeier, U. (Eds.), Handbook of international trade. Palgrave Macmillan, Basingstoke, Geertz, C., Peddlers and princes; social change and economic modernization in two indonesian towns. University of Chicago Press, Chicago,. Geertz, C., Bazaar economy: Information and search in peasant marketing. American Economic Review 68, Genc, M., Gheasi, M., Nijkamp, P., Poot, J., The impact of immigration on international trade: A meta-analysis, in: Nijkamp, P., Poot, J., Sahin, M. (Eds.), Migration impact assessment. Edward Elgar, Cheltenham, Gould, D.M., Immigrant links to the home country: Empirical implications for united states bilateral trade flows. Review of Economics and Statistics 76, Green, W.H., Econometric analysis. Prentice Hall, Upper Saddle River; NJ. Greif, A., Reputation and coalitions in medieval trade: Evidence on the Maghribi traders. Journal of Economic History 49, Greif, A., The organization of long-distance trade: Reputation and coalitions in the Geniza documents and Genoa during the 11th-century and 12th-century. Journal of Economic History 51, Greif, A., Institutions and the path to the modern economy: Lessons from medieval trade. Cambridge University Press, Cambridge. Hatzigeorgiou, A., Migration as trade facilitation: Assessing the links between international trade and migration. The B.E. Journal of Economic Analysis & Policy 10, Article 24. Head, K., Ries, J., Immigration and trade creation: Econometric evidence from canada. Canadian Journal of Economics 31,

27 Hirano, K., Imbens, G.W., The propensity score with continuous treatments, Applied Bayesian modeling and causal inference from incomplete-data perspectives. John Wiley & Sons, Ltd, Imai, K., King, G., Stuart, E.A., Misunderstandings between experimentalists and observationalists about causal inference. of the Royal Statistical Society: Series A (Statistics in Society) 171, Imai, K., van Dyk, D.A., Causal inference with general treatment regimes: Generalizing the propensity score. Journal of the American Statistical Association 99, Iyer, G.R., Jon, M.S., Ethnic entrepreneurial and marketing systems: Implications for the global economy. Journal of International Marketing 7, Javorcik, B.S., Ozden, C., Spatareanu, M., Neagu, C., Migrant networks and foreign direct investment. Journal of Development Economics 94, Kluve, J., Schneider, H., Uhlendorff, A. Zhao, Z., Evaluating continuous training programmes by using the generalized propensity score. Journal of the Royal Statistical Society: Series A (Statistics in Society) 175, Landa, J.T., A theory of the ethnically homogeneous middleman group: An institutional alternative to contract law. Journal of Legal Studies 10, Landa, J.T., Trust, ethnicity, and identity: Beyond the new institutional economics of ethnic trading networks, contract law, and gift-exchange. University of Michigan Press, Ann Arbor. Lechner, M., Identification and estimation of causal effects of multiple treatments under the conditional independence assumption, in: Lechner, M., Pfeiffer, F. (Eds.), Econometric evaluation of labour market policies, vol. 13. Physica-Verlag HD, Marshall, M. G., Jaggers, K., Gurr, T.R, Polity IV Project: Dataset Users Manual. Center for Systemic Peace: Polity IV Project. Massey, D.S., Alarcón, R., Durand, J., Gonzalez, H., Return to Aztlan: The social process of international migration from western Mexico. University of California Press, Berkeley. Mayda, A.M., International migration: A panel data analysis of the determinants of bilateral flows. Journal of Population Economics 23, McKenzie, D., Rapoport, H., Network effects and the dynamics of migration and inequality: Theory and evidence from Mexico. Journal of Development Economics 84, Newey, W.K., Powell, J.L., Vella, F., Nonparametric estimation of triangular simultaneous equations models. Econometrica 67, Peri, G., Requena-Silvente, F., The trade creation effect of immigrants: Evidence from the remarkable case of Spain. Canadian Journal of Economics 43,

28 Polanyi, K., Trade and market in the early empires; economies in history and theory. Free Press, Glencoe, Ill. Polanyi, K., Primitive, archaic, and modern economies; essays of Karl Polanyi, [1st ed. Anchor Books, Garden City, N.Y.,. Ratha, D, Shaw, W., South-South Migration and Remittance. World Bank Working Paper, No Rauch, J.E., Networks versus markets in international trade. Journal of International Economics 48, Rauch, J.E., Business and social networks in international trade. Journal of Economic Literature 39, Rauch, J.E., Trindade, V., Ethnic Chinese networks in international trade. Review of Economics and Statistics 84, Rosenbaum, P.R., Rubin, D.B., The central role of the propensity score in observational studies for causal effects. Biometrika 70, Santos Silva, J, Tenreyro, S., The log of gravity. Review of Economics and Statistics 88, Tai, S.H.T., Market structure and the link between migration and trade. Review of World Economics 145, Taylor, J.E., Differential migration, networks, information, and risk, in: Stark, O. (Ed.), Migration, human capital and development, research in human capital and development, vol. 4. JAI Press, New York, Terza, J.V., Basu, A., Rathouz, P.J., Two-stage residual inclusion estimation: Addressing endogeneity in health econometric modeling. Journal of Health Economics 27, Wagner, D., Head, K., Ries, J., Immigration and the trade of provinces. Scottish Journal of Political Economy 49, Winters, P., de Janvry, A., Sadoulet, E., Family and community networks in Mexico- US migration. Journal of Human Resources 36, Wooldridge, J.M., Econometric analysis of cross section and panel data. MIT Press, Cambridge, Mass.

29 Tables and Figures Table 1: Descriptive Statistics Dependent Variables Mean Std. dev. Min Max Obs. OECD residence countries ln(u i ) ,178 ln(s i ) ,178 ln(m i ) ,178 Mi D/M i H ,669 Non-OECD residence countries ln(u i ) ,939 ln(s i ) ,939 ln(m i ) ,939 Notes: We have observations for 98 countries of origin and 29 OECD countries of residence of migrants as well as 65 Non-OECD countries of residence. Countries of residence are importers of goods while countries of origin are the exporters. U i and S i refer to skilled and unskilled immigrants where we define skilled immigrants as those with at least secondary level of education. M i denotes total bilateral imports while M D i and M H i refer to differentiated and homogeneous goods imports. The classification of goods follows Rauch (1999). For most pairs with non-oecd residence countries we lack data on imports by goods classification such that we refrain from distinguishing between differentiated and homogeneous goods trade for these pairs.

30 Table 2: Descriptive Statistics Independent Variables Acronyms Description Mean Std. dev. Min Max GDP R i log GDP residence country GDP O i log GDP origin country GDP P CR i log GDP per capita residence country GDP P CO i log GDP per capita origin country P OP R i log population residence country P OP O i log population origin country GINIR i GINI coefficient residence country GINIO i GINI coefficient origin country UNEMP R i unemployment rate residence country UNEMP O i unemployment rate origin country REALEXCH i real exchange rate btw, residence and origin country CP IR i corruption perception index residence country CP IO i corruption perception index origin country ILOBARGAINR i ILO bargaining power index residence country ILOBARGAINO i ILO bargaining power index origin country ILOLABORR i ILO working condition index residence country ILOLABORO i ILO working condition index origin country ILODISCRR i ILO worker discrimination index residence country ILODISCRO i ILO worker discrimination index origin country ILOCHILDR i ILO child labor index residence country ILOCHILDO i ILO child labor index origin country P OLIT Y 2R i Polity IV index residence country P OLIT Y 2O i Polity IV index origin country LIF EEXP R i life expectancy residence country LIF EEXP O i life expectancy origin country F ERT ILR i fertility residence country F ERT ILO i fertility origin country DIST i log distance btw. residence and origin country LIT R i literacy rate residence country LIT O i literacy rate origin country W ARSAW i member of Warsaw Pact CONF LICT R i number of armed conflicts residence country CONF LICT O i number of armed conflicts origin country COMLANG i common language in residence and origin country COLONY i colonial relationship btw. residence and origin country GT A i goods trade agreement ST A i service trade agreement OECDO i OECD member RELIGION i common religion in residence and origin country MIGIMP ED1 i migration impediments: re-admission agreements MIGIMP ED2 i migration impediments: social security system MIGIMP ED3 i migration impediments: exchange of information MIGIMP ED4 i migration impediments: labor market regulation MIGIMP ED5 i migration impediments: respect for human rights GOV CCO i institutions origin country: control of corruption GOV GEO i institutions origin country: government effectiveness GOV P SO i institutions origin country: political stability GOV RLO i institutions origin country: rule of law GOV RQO i institutions origin country: regulatory quality GOV AV O i institutions origin country: absence of violence Observations 5,117 Notes: We summarize information for those observations that have non-missing, positive levels of M i, U i, and S i as well as for all covariates. When estimating the effects for subgroups of bilateral imports the respective dependent variable determines the number of observations (see Table 1).

31 Table 3: First-Stage Estimation of GPS S i U i Coef. Std.err. Coef. Std.err. Const ( ) ( ) GDP R i (45.854) (37.962) GDP R i (1.728) (1.428) GDP R i (0.022) (0.018) GDP O i (5.610) (4.815) GDP O i (0.223) (0.191) GDP O i (0.003) (0.003) GDP P CR i ( ) ( ) GDP P CR i (40.189) (32.193) GDP P CR i (1.399) (1.120) GDP P CO i (5.621) (4.755) GDP P CO i (0.689) (0.582) GDP P CO i (0.028) (0.023) P OP R i (45.652) (39.917) P OP R i (2.742) (2.400) P OP R i (0.055) (0.048) P OP O i (5.276) (4.159) P OP O i (0.309) (0.244) P OP O i (0.006) (0.005) GINIR i (2.221) (1.908) GINIR i (0.066) (0.056) GINIR i (0.001) (0.001) GINIO i (0.136) (0.115) GINIO i (0.003) (0.003) GINIO i (0.000) (0.000) UNEMP R i (0.097) (0.079) UNEMP R i (0.012) (0.010) UNEMP R i (0.000) (0.000) UNEMP O i (0.048) (0.038) UNEMP O i (0.004) (0.003) UNEMP O i (0.000) (0.000) REALEXCH i (0.014) (0.012) REALEXCH i (0.002) (0.002) REALEXCH i (0.000) (0.000) CP IR i (2.375) (2.102) CP IR i (0.374) (0.330) CP IR i (0.019) (0.017) CP IO i (0.321) (0.257) CP IO i (0.061) (0.049) CP IO i (0.004) (0.003) ILOBARGAINR i (0.046) (0.038) ILOBARGAINR i (0.002) (0.002) ILOBARGAINR i (0.000) (0.000) ILOBARGAINO i (0.018) (0.016) ILOBARGAINO i (0.001) (0.001) ILOBARGAINO i (0.000) (0.000) ILOLABORR i (0.075) (0.062) ILOLABORR i (0.002) (0.002) ILOLABORR i (0.000) (0.000) ILOLABORO i (0.019) (0.017) ILOLABORO i (0.001) (0.001) ILOLABORO i (0.000) (0.000) ILODISCRR i (0.097) (0.081) ILODISCRR i (0.004) (0.003) Continued on next page

32 Table 3: First-Stage Estimation of GPS S i U i Coef. Std.err. Coef. Std.err. ILODISCRR i (0.000) (0.000) ILODISCRO i (0.021) (0.017) ILODISCRO i (0.001) (0.001) ILODISCRO i (0.000) (0.000) ILOCHILDR i (0.153) (0.132) ILOCHILDR i (0.035) (0.031) ILOCHILDR i (0.002) (0.002) ILOCHILDO i (0.068) (0.053) ILOCHILDO i (0.016) (0.012) ILOCHILDO i (0.001) (0.001) P OLIT Y 2R i (0.816) (0.708) P OLIT Y 2R i (0.170) (0.146) P OLIT Y 2R i (0.009) (0.008) P OLIT Y 2O i (0.017) (0.014) P OLIT Y 2O i (0.002) (0.001) P OLIT Y 2O i (0.000) (0.000) LIF EEXP R i (34.237) (29.310) LIF EEXP R i (0.451) (0.386) LIF EEXP R i (0.002) (0.002) LIF EEXP O i (0.367) (0.290) LIF EEXP O i (0.006) (0.005) LIF EEXP O i (0.000) (0.000) F ERT ILR i (9.869) (8.319) F ERT ILR i (5.385) (4.531) F ERT ILR i (0.980) (0.825) F ERT ILO i (0.389) (0.308) F ERT ILO i (0.102) (0.080) F ERT ILO i (0.008) (0.006) DIST i (3.148) (3.148) DIST i (0.415) (0.407) DIST i (0.018) (0.017) LIT R i (5.974) (4.997) LIT R i (0.071) (0.060) LIT R i (0.000) (0.000) LIT O i (0.047) (0.036) LIT O i (0.001) (0.001) LIT O i (0.000) (0.000) W ARSAW i (0.400) (0.321) CONF LICT R i (0.369) (0.309) CONF LICT O i (0.096) (0.077) COMLANG i (0.108) (0.084) COLONY i (0.166) (0.134) GT A i (0.130) (0.102) ST A i (0.153) (0.121) OECDO i (0.165) (0.128) RELIGION i (0.066) (0.054) MIGIMP ED1 i (0.389) (0.277) MIGIMP ED2 i (0.498) (0.382) MIGIMP ED3 i (0.245) (0.200) MIGIMP ED4 i (0.479) (0.363) MIGIMP ED5 i (0.221) (0.184) adj. R AIC Obs. 3,178 3,178 Notes:,, denote significance levels at 1, 5, and 10%, respectively.

33 Table 4: Balancing Test - T-statistics 4 Groups 9 Groups Unconditional Conditional Unconditional Conditional U i S i U i S i U i S i U i S i GDP R i GDP O i GDP P CR i GDP P CO i P OP R i P OP O i GINIR i GINIO i UNEMP R i UNEMP O i REALEXCH i CP IR i CP IO i ILOBARGAINR i ILOBARGAINO i ILOLABORR i ILOLABORO i ILODISCRR i ILODISCRO i ILOCHILDR i ILOCHILDO i P OLIT Y 2R i P OLIT Y 2O i LIF EEXP R i LIF EEXP O i F ERT ILR i F ERT ILO i DIST i LIT R i LIT O i W ARSAW i CONF LICT R i CONF LICT O i COMLANG i COLONY i GT A i ST A i OECDO i RELIGION i MIGIMP ED1 i MIGIMP ED2 i MIGIMP ED3 i MIGIMP ED4 i MIGIMP ED5 i Mean Obs. (common support) 2,525 1,352 Notes: We regress each covariate on the treatments U i and S i whereby the conditional model controls for the GPS evaluated at the quartiles of the treatments (see also Imai and van Dyk, 2004 for an analogous test of the balancing property). For each of these regressions we report the t-statistics on whether the effect of U i and S i on the respective covariates is significantly different from zero. The unconditional and conditional models are estimated for the common support samples based alternatively on 4 and 9 groups. Note that restricting the sample to the common support of the GPS already improves the balancing compared to the corresponding unconditional estimates based on the total sample.

34 Table 5: Second-Stage Estimation of the Unit Dose-Response Function 4 Groups Coef. Std. err. U i (0.228) S i (0.323) Ui (0.036) Si (0.054) U i Ĝi (0.693) S i Ĝi (0.814) U i S i (0.084) Ĝ i (4.434) Ĝ 2 i (12.606) Constant (0.688) Observations 2,525 adj. R F-statistic Ĝi terms F-statistic U i terms F-statistic S i terms Notes:,, denote significance levels at 1, 5, and 10%, respectively. Standard errors in parentheses. U i and S i refer to the logarithm of the stock of unskilled and skilled immigrants, respectively, who reside in the importer country and originate from the exporter country. Ĝ i refers to the generalized propensity score calculated according to equation (2) using the coefficients from the first-stage regressions in Table 3. We estimate the standard errors of the avg. dose-response function by bootstrapping with 200 draws that take into account that the second-stage estimates involve imprecision from first-stage estimates.

35 Figure 1: Balancing of Covariates - 4 Groups Mean: 4.86; Median: 3.30; 60% sign. at 1% level A. Unconditional Mean: 0.60; Median: 0.48; 0.5% sign. at 1% level B. Conditional on Bivariate GPS Note: The histograms contain the t-statistics for all 44 linear terms of the covariates. We test for the equivalence of the covariates between 4 groups which yields 176 t-statistics. In the conditional comparison we split the distribution of Ĝ q i into 16 blocs and conduct the t-tests for subsamples belonging to the same bloc. A weighted average over these blocs is computed for each treatment group and each covariate.

36 Figure 2: Balancing of Covariates - 9 Groups Mean: 3.32; Median: 2.45; 47% sign. at 1% level A. Unconditional Mean: 0.54; Median: 0.44; 0.5% sign. at 1% level B. Conditional on Bivariate GPS Note: The histograms contain the t-statistics for all 44 linear terms of the covariates. We test for the equivalence of the covariates between 9 groups which yields 396 t-statistics. In the conditional comparison we split the distribution of Ĝ q i into 8 blocs and conduct the t-tests for subsamples belonging to the same bloc. A weighted average over these blocs is computed for each treatment group and each covariate.

37 Figure 3: Avg. Dose-Response Import Volume Note: Blue corresponds to significant and positive and light blue to insignificant and positive. We obtain standard errors from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation.

38 Figure 4: Avg. Treatment Effect Import Volume A. Unskilled Migrants (U) B. Skilled Migrants (S) Note: Blue corresponds to significant and positive, light blue to insignificant and positive, yellow to insignificant and negative, and red to significant and negative predictions. We obtain standard errors from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation.

39 Figure 5: Nonparametric Avg. Dose-Response Import Volume Note: Blue corresponds to significant and positive, light blue to insignificant and positive, and yellow to insignificant and negative. The surface is predicted from a multivariate local linear regression. An optimal bandwidth is obtained by cross validation. Standard errors stem from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation.

40 Figure 6: Nonparametric Avg. Treatment Effect Import Volume A. Unskilled Migrants (U) B. Skilled Migrants (S) Note: Blue corresponds to significant and positive, light blue to insignificant and positive, yellow to insignificant and negative, and red to significant and negative predictions. The surface is predicted from a multivariate local quadratic regression. An optimal bandwidth is obtained by cross validation. Standard errors stem from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation.

41 Dose-Response Ratio of Differenti- Figure 7: Nonparametric Avg. ated and Homogeneous Goods Note: Blue corresponds to significant and positive, light blue to insignificant and positive, and yellow to insignificant and negative. The outcome refers to the ratio between imports in differentiated and homogeneous goods M D i /M H i. The surface is predicted from a multivariate local linear regression. An optimal bandwidth is obtained by cross validation. Standard errors stem from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation.

42 Figure 8: Avg. Dose-Response Non-OECD residence countries Note: In this Figure we limit the sample to pairs with non-oecd country of residence and compute the avg. dose-response function on the basis of the parametric version of the secondstage. We obtain standard errors from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation. Dark (light) blue color reflects positive and significant (insignificant) conditional expectations on imports.

43 Figure 9: Avg. Dose-Response Controlling for emigration Note: In this Figure we control for emigration as a third endogenous treatment and compute the avg. dose-response functions on the basis of the parametric version of the second-stage. We obtain standard errors from a bloc bootstrapping routine with 200 replications. This routine includes the common support restriction, the first-stage estimation of the GPS as well the second-stage estimation. Blue color reflects positive and significant conditional expectations on imports.

44 Figure 10: The Role of Institutions Note: This figure displays the difference between a counterfactual avg. dose-response E [M(U, S)] function where we have raised the quality of institutions by one standard deviation and the actual avg. dose-response function E[M(U, S)]. See Section 5 for details. Both avg. dose-response functions are computed on the basis of the parametric version of the secondstage. We use the estimates on the control of corruption, government effectiveness, political stability and absence of violence/terrorism, rule of law, regulatory quality, and voice and accountability from the World Bank as proxies for institutional quality. We raise each of these dimensions at the same time. Hence, level of the difference E [M(U, S)] E[M(U, S)] reflects the gain in bilateral trade volume due to a one-standard deviation increase of institutional quality at different combinations of skilled and unskilled migration.