The Migration network effect on international trade

The Migration Network Effect on International Trade R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni September 13, 2014 Contents 1 Outline 2 Migration and Trade Review 3 Multilateral trade resistance terms Review 4 Spatial Econometrics and the weight matrix 5 Empirical strategy 6 Results 7 Conclusion R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 1/20 ...near things... Everything is related to everything else, but near things are more related to each other (Waldo Tobler, 1970) How to define near things? asked a student at the 2013 SEAI summer school. - ”Even God Doesnt know”. (Anil Bera, 2013) R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 2/20 Outline Outline ◮ To measure the relations between International Trade of different goods categories and Global Migration. ◮ Using spatial econometric gravity of trade (where near things are based on human migration) to highlight both the direct and the indirect (network) effect of migration on trade, using a pooled panel (1970,1980,1990,2000). ◮ Controlling for reverse causality and importers strength ◮ Obtaining, in a gravity model frameworks with results consistent with the existing literature, a new way to measure the effect of the ethnic groups all togheter. R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 3/20 Direct/indirect migration effect on trade ◮ Standard Heckscher-Ohlin model suggests that the movement of goods across borders can provide a substitute for the movement of migrants. The more recent works (Gould et al.1994, Nijkamp et al. 2011) stressed that the two actually complement each other. ◮ Two strands of research investigate 1) on the one hand the direct relation between trade and migration, 2) on the other hand the relation between migration network and trade (i.e. the effect of the neighbors) ◮ Rauch et al., in a series of papers (and Felbermayr et al. 2010) looks at the role of ethnic immigrants networks in facilitating trade. In so doing he not only looks at the direct effect of migration on trade between the same pair of countries, but at the indirect (or network) effect. Migration and Trade Review R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 4/20 Homogeneous versus differentiated goods ◮ Rauch’s 2002 classification distinguish between homogeneous and differentiated goods. The former have a reference price, whether it be a result of organized exchanges or simply of price quotations in a specialized journal, while the latter lack a reference price and can be thought of as branded commodities. ◮ The central result in Rauch and Trindade (2002) is that the positive effect of migration on trade is larger for differentiated goods (i.e. those items that are not homogeneous and that are not traded in organized exchanges), therefore rendering that knowledge about counterpart reputation particularly valuable. ◮ This is indeed the result found by Rauch and Trindade (2002) and one of the aspects we test in this paper, with a new approach. Migration and Trade Review R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 5/20 Multilateral resistance terms review Unconstrained Version: Fij = β0 si sj dij ◮ Anderson, Van Wincoop (2003) introduced and motivated the constrained version, that considers Multilateral Trade Resistance (MTR) (overall set of trade barriers that exporter and importer countries face). MTR equations are not linear in the parameters. ◮ Different version of Fixed effects (FE) and other approaches to proxy MTR. Feenstra Fixed Effects(2004), Baier Bergstrand Bonus Vetus OLS (2009), Patuelli et al. (forthcoming) Spatial Filtering, Behrens et al. (2012) spatial autoregressive models. ◮ Behrens et al. asserted that FE approach does not completely filters out the spatial (network) autocorrelation in the residuals. ◮ Summarizing ... Multilateral trade resistance terms Review R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 6/20 Spatial autoregressive models SDM y = ρWy + Xβ + WXγ + ǫ (1) which becomes the SEM model if included and excluded variables are not correlated (common factor tests can be performed (LeSage Pace 2008) SEM y = Xβ + λWǫ + ǫ. (2) and becomes the SAR model when λ = 0 SAR y = ρWy + Xβ + ǫ (3) y = ρWy + Xβ + λWǫ + ǫ (4) ...confused models SAC ...and the augmented version of the SDM model (also called Manski) that fully accounts for all possible spatial dependency: Manski y = ρWy + Xβ + WXγ + λWǫ + ǫ (5) where y is the dependent variable, X is the matrix of the explanatories and ǫ represents the residuals. W is the (spatial) weight matrix, β, γ, λ and ρ are the coefficients to be estimated. Spatial Econometrics and the weight matrix R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 7/20 The weight matrix ◮ To model the spatial autoregressive component one generally uses an n × n weight matrix (W ) that defines the set of neighbors: most frequently W is based on spatial contiguity, so that [wij ] = 1 if i and j share geographical borders, and 0 otherwise. ◮ The matrix can be both spatial or non-spatial. Accordingly, several proposals have been made in the literature, such as using the technological similarities or the transport costs (Case 1993, Behrens et al. 2012) instead of spatial metrics. LeSage and Pace (2011) discussed the possibility of jointly modeling spatial and non-spatial dependence through a double autoregressive component ◮ In general, spatial matrices are assumed to be exogenous, while non spatial are likely to be endogenous (respect to trade). ◮ The first attempt to estimate a SAR model when endogenous matrix is used is proposed by Kelejian and Piras (2014). Spatial Econometrics and the weight matrix R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 8/20 Empirical strategy Exemplifying representation of the indirect migration effect (co-etnicity) as in Felbermayr et al. (2010) Empirical strategy Exemplifying representation of the direct and indirect migration channels (origin-side) R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 9/20 Our W matrix ◮ To control for network structure we use the migration network matrix (year:2000), so that topological distance replaces the more usual spatial weight matrix. ◮ In order to identify the significant links (i.e. dyads where migrants are more than expected), we use a stochastic benchmark based on the hypergeometric distribution, as recently done in Riccaboni et al. 2013. ◮ In order to account for the effect described in the figure, we need to apply a Kroeneker transformation to our matrix: WKM = W M ⊗ I . ◮ In a panel framework one needs to account for the time index. In order to do that, the matrix has to be pre-multiplied by a diagonal matrix of dimension t: WKM,t = It ⊗ WKM . Empirical strategy R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 10/20 Data Sources ◮ ◮ ◮ ◮ Migrants data come from the World Bank’s Global Bilateral Migration dataset Ozden et al. (2011): it is composed of matrices of bilateral migrant stocks spanning from 1960 to 2000 taking into account a total of 232 countries. For trade, we use the NBER-UN dataset described by Feenstra et al.(2005). For each country it provides the value (in thousands of US dollars) exported to all other countries. In our analysis, we focus on the decades from 1970 to 2000. We retain only the countries present in both of them to enhance comparability, ending up with a set of 146 countries (nodes) that have active interactions in both trade and migration. All the other controls used in the regressions (e.g. GDP, population, contiguity, common language, etc.) have been retrieved from the CEPII dataset (Mayer and Zignago 2011) Empirical strategy R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 11/20 Model diagnostics ◮ Two approaches: 1. Fixed Effect (FE) model that properly accounts for MTR, we opted for importer and exporter time-varying FE as suggested by the most recent literature (Felbermayr et al. 2012, Head and Mayer 2013). 2. Spatial Durbin model (SDM) that should filters out the Multilateral Trade Resistances, following a similar approach to Berehens et al. (2012). ◮ The causal relationship between trade and migration can hold both ways, to disentangle the effect we adopt an instrumental variable strategy. Among others (Felbermayr et al. 2012, Briant and Lafourcade 2014), the method a la Antonji and Card (1991) is used by Peri, Requena - Silvente (2010) and Bratti et al. (2011). ◮ Since migrants bring knowledge and competences that are relevant for producing and exporting differentiated goods, migration from i to its neighbor may erode i’s ability to export specific goods to other markets making h a better competitor. We include the total import of the dest. country j net of imports from i to filters out this effect: im.strength.netj = Empirical strategy X Thj − Tij R.Metulini, P.Sgrignoli, S.Schiavo, h6=i M.Riccaboni (6) 12/20 The final model ◮ ❲t▼ T + X X β K T =ρ k k k=1 ◮ + K X k=1 ❲t▼ X γ k k +ǫ (7) where T is the dependent variable, ρ is the scalar coefficient of the lagged trade term to be estimated, β and γ are the k × 1 vectors of coefficients to be estimated for, respectively, the explanatories and the lagged explanatories Xk , where the regressors k are the following: distance, GDPpc sum, population sum, GDPpc sim, population sim, migration, importstregth, contig , comcur , comlang , colony and rta. WtM is the n2 ∗ t × n2 ∗ t network weight matrix relative to migration. Likelihood ratio tests was performed in order to end up with the Durbin model (SDM) specification. Empirical strategy R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 13/20 A discussant of myself ◮ Why do you used Durbin Model (SDM)? 1) We want to explicitly take into account for the indirect effects. 2) Elhorst (2010) argues that the SDM is the only model that provides unbiased parameter estimates and correct standard errors, even if the true data-generation process is any of the other spatial regression models mentioned above. ◮ Why you did not consider zero flows? Zero flows are present in the dataset (about 20 %), however, we have to choose between two curches: Estimation approaches accounting for zero flows in the framework of durbin model are not existing so far. I spoke with T. Scherngell (AIT) that in 2009 tries (with M.Fischer) to propose an estimation procedure for various spatial econometric models with zero flows, but the contra were more than the pro (i.e. the estimator was biased). Empirical strategy R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 14/20 Baseline model results Non instrumented base total trade diff. goods homog. goods ols ols fe ols fe ols fe migration .129***.128***.133***.140***.109***.113*** R 2 adj .639 .639 .752 .629 .820 .604 .716 obs 29784 24105 27217 20908 23467 22256 24813 total diff. homog. trade goods goods Correlation between Tradet and Migrationt 0.35 0.37 0.29 Correlation between Tradet and Migrationt−1 0.28 0.29 0.22 First stage test for the validity of the instrument >37.75>37.75>37.75 Durbin-Wu-Hausman for the endogeneity in the model 14.16 4.70 12.77 Instrumented migration R 2 adj obs Results total trade diff. goods homog. goods ols fe ols fe ols fe .088***.121***.109***.135***.070***.105*** .636 .746 .608 .806 .589 .707 17448 18551 15261 16124 16211 17039 R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 15/20 Impacts with Durbin model *I voluntary skipped the coefficients of the other variables Table: Controlling for Reverse Causality total trade different. goods homogen. goods directindirect totaldirectindirect totaldirectindirect total migration0.092 -0.0210.0710.115 -0.0090.1060.073 -0.0420.031 Table: Reverse causality and import strength total trade different. goods homogen. goods directindirect totaldirectindirect totaldirectindirect total migration0.092 -0.0230.0680.141 0.0020.1430.062 -0.0420.021 Results R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 16/20 Spatial residual autocorrelation ◮ Results, using Moran I, shaw that, confirming Beherens et al. (2012) work, the FE still shaw some (negative) autocorrelation in the residuals, while SDM significantly filters out all noise in the residuals. ◮ We interpret it saying that, controlling for migration network effect, we filter out for the overall set of trade barriers (i.e. migration network breaks down the multilateral barriers to trade). ◮ Apart from the economic interpretation, the Durbin model shaw i.i.d residuals, meaning that the durbin coefficients are unbiased. Table: Moran I test for autocorrelation on the residuals of the gravity model estimated by OLS (column i.), FE (column ii.) and SDM (column iii.) OLS FE SDM total 0.077 - 0.008 0.001 z-score (p-val)28.01 (0.000)-3.49 (0.000) 0.139 (0.444) differentiated 0.081 -0.009 -0.012 z-score (p-val)25.44 (0.000)-3.43 (0.000)-2.297 (0.011) homogeneous 0.081 -0.011 0.001 z-score (p-val)27.31 (0.000)-4.16 (0.000) 0.299 (0.381) Results R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 17/20 Results ◮ ◮ ◮ Conclusion The effect of migration for the differentiated goods (when accounting for import strength) is signicantly higher than for homogeneous goods (0.143 versus 0.021) using spatial econometric approach. The migration indirect effect is positive for differentiated goods, and negative for homogeneous. Maybe these coefficients are a sum of two effects: 1) competition (negative) and 2) positive network spillover. The negative one is stronger for homogeneous goods, while the positive is stronger for differentiated. Infact, differentiated are more difficult to substitute for, so that they suffer less competition from third countries. Using Durbin model, we filter out the residual autocorrelation. Anyway, the Durbin model shaw i.i.d residuals, meaning that the coefficients are unbiased. R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 18/20 Future developments Conclusion ◮ To account for endogeneity of migration in a better way, finding a good instrument for migration variable and migration matrix ◮ To accomodate for more networks, by mean more than one weight matrix togheter ◮ To switch to Virtual Water trade (VWT) and to accomodate spat.econometric and network techniques in the analysis of the relation between VWT and migration. R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 19/20 Thanks Thank you for your attention. Any comment or suggestion is welcome rodolfo.metulini@imtlucca.it Conclusion R.Metulini, P.Sgrignoli, S.Schiavo, M.Riccaboni 20/20