Revisiting the Trade-Migration Nexus: Evidence from New OECD data

Transcription

1 Revisiting the Trade-Migration Nexus: Evidence from New OECD data Gabriel J. Felbermayr and Farid Toubal November, 2007 Abstract International migrants contribute to bilateral trade creation if their presence reduces information costs or entails additional demand for goods from their source countries. Using new data on stocks of foreign-born individuals by skill class, we try to separately quantify those two channels. We assume that improved information affects host countries imports and exports symmetrically, while the preference channel matters for imports only. On average, for differentiated goods, both channels contribute evenly towards the total trade-creating effect of migration. In line with expectations, the relative importance of the trade cost channel is largest with homogeneous goods and for high-skilled migrants. JEL Codes: F12, F22, Z13 Keywords: Migration, International Trade, Gravity Equation. The authors would like to thank Hartmut Egger, Josef Falkinger, Thierry Mayer, and Wilhelm Kohler for comments. Part of this research was undertaken while Felbermayr was visiting Zürich University. Benny Jung has provided excellent research assistance. The usual disclaimer applies. Felbermayr gratefully acknowledges financial support under a grant from Thyssen Foundation entitled Migration and Labor Market Integration. Department of Economics, University of Tübingen, Nauklerstrae 47, Tübingen, Germany. Tel.: gabriel.felbermayr@uni-tuebingen.de. Corresponding author: University of Paris I, Panthéon-Sorbonne, CNRS. Centre d Économie de la Sorbonne, Boulevard de l Hôpital, Paris. Tel.: toubal@univparis1.fr. 1

2 1 Introduction A growing body of economic literature discusses the trade-creating role of international migration. The presence of migrants can promote trade between their source and their host countries in at least two ways. First, they might help overcome informal barriers to international trade related to language, culture, or institutions, they may facilitate the creation of business relationships, and they may make valuable information on foreign sales and sourcing opportunities more readily available(dunlevy, 2006; Combes et al., 2005; Herander and Saavedra, 2005; Rauch and Trindade 2002). Second, migrants boost trade if they attach higher utility from goods produced in their host countries (Gould, 1994; Head and Ries, 1998; Girma and Yu, 2000; Wagner et al., 2002). We refer to the first channel as the trade-cost and to the second channel as the preference channel. Quantifying the relative importance of these mechanisms is important, since trade creation due to the alleviation of informational frictions constitutes a source of welfare gains for the host and source country. One cannot draw that conclusion if trade is higher due to specific features of preferences. Empirically, we distinguish between the information and preference channels by making the following identification assumption: Improved information affects host countries imports and exports symmetrically, while the preference channel matters for imports only. This is clearly a strong assumption; however, it receives support on conceptual grounds as well as through a number of robustness checks. The literature on the trade-migration nexus has made increasing use of the gravity model of bilateral international trade. We follow the recent contribution by Combes et al., (2005) and introduce a bilateral affinity parameter into the usual Dixit-Stiglitz utility function of the representative household. That parameter may depend on bilateral ethnic ties, thereby capturing the preference channel described above. We also allow bilateral trade costs to depend on migration; this is meant to account for the information channel described above. We use recently available data on the stocks of individuals born in some OECD country and residing in another. Compared to other cross-country data sets, that data has the virtue that it involves (almost) all OECD countries, so that the effects of immigrants and emigrants on bilateral trade can be studied simultaneously. Moreover, we have information on stocks rather than flows. This may be an advantage rather than a shortcoming since it is stocks that matter for the information and preference channels, not flows. The 2

3 data provides the number of foreign-born individuals rather than of persons with foreign nationality: Migration stocks are directly comparable across countries since the definition of a migrant does not depend on idiosyncratic characteristics of national naturalization practices (ius solis versus ius sanguis). Finally, and rather uniquely, the data distinguishes between different skill classes. This allows to check the widely held belief that the importance of the information channel relative to the preference channel varies systematically across skill groups. This rich structure of the data therefore provides ways to externally validate our identification assumptions. Our estimation results suggest that migrants have a positive, statistically and economically significant impact on imports. With aggregate trade, and using information on the total stocks of migrants, we find that the preference effect of migration on bilateral trade amounts to up to 63 percent of the total effect. This finding qualifies the intuition provided in earlier papers. 1 Running separate regressions for Rauch (1999) subaggregates of goods, we find that the trade cost effect is about three times larger with differentiated than with homogeneous goods. This confirms the hypothesis of Rauch and Trindade (2002) that migrants convey trade-relevant information on differentiated goods that are not already captured by the price system. Moreover, we find that the information channel dominates the total effect of high-skilled migrants on imports, leaving no statistically significant role for the preference channel. The finding is in line with the argument that high-skilled migrants are more likely to possess information that is relevant for international transactions, while their attachment to source country varieties is not strong. Overall, this paper makes the following contributions. First, it makes a systematic attempt to disentangle the channels through which migrants affect trade. Several authors have discussed both channels, but the literature does not hold any quantification of their relative importance. We propose a simple novel identification strategy that allows to disentangle the preference and the network effect of migration on trade. Second, we use new data that has not been explored in the present context yet. More precisely, we exploit information on the skill structure of migrants, and make separate inference for homogeneous and differentiated goods. The paper is structured as follows. In Section 2, we derive the theoretical framework. In section 3, we present the econometric specification, the methodology, and the data. In 1 See the discussion by Combes et al. 2005, p.11. 3

4 section 4, we present the econometric results. We conclude in Section 5. 2 Theoretical Framework In this section, we briefly sketch the theoretical foundations of a partial equilibrium gravity equation that allows for the preference and the trade cost channels of migration. We follow Combes et al. (2005) and assume that the representative agent in country i has a Dixit-Stiglitz utility function defined over domestic and imported varieties U i = n N j (a ij x ijh ) σ 1 σ, σ > 1, (1) j=1 h=1 where x ijh denotes consumption of a generic variety h produced in country j. N is the number of countries, n j is the mass of varieties produced in country j, and σ denotes the elasticity of substitution. The only modification relative to the standard specification is the inclusion of a bilateral affinity term a ij which describes the preference of the representative consumer in country i for country j s products. Maximizing (1) subject to an appropriate budget constraint yields an expression for country i s value of import demand for goods from country j c ij = a σ 1 ij T 1 σ ij n j p 1 σ j E i P σ 1 i, (2) where p j is the mill price of a variety produced in country j and is assumed identical over varieties. T ij > 1 is an ad valorem iceberg-type trade costs between country i and country j. Thus, c.i.f. prices are given by p ij = T ij p j. The aggregate price level in country i is ( ) 1 given by P i = j aσ 1 ij n j p 1 σ 1 σ ij while aggregate expenditure is given by E i. 2 There are two channels through which migrants might increase imports. First, the foreign-born population from country j in country i as well as the foreign-born population from country i in country j may provide information on trading opportunities between the two countries. Immigrants are familiar with the foreign country s language and culture, and have knowledge on both countries commercial, legal and political institutions. In that 2 One can easily interpret (1) as the subutility index belonging to some specific sector, and nest subutility indices, into, say, a Cobb-Douglas aggregator. Up to a constant multiplicative factor representing sectoral expenditure shares, bilateral trade flow equations for sub-aggregates (e.g., for groups of goods with different degrees of substitutability) will be formally similar to those derived from (1). 4

5 way, migrants lower trade costs and enhance bilateral trade (most likely both exports and imports). Second, the foreign-born population from country j in country i may have a special preference for varieties from their origin country, which ceteris paribus also creates trade. We assume that ad valorem trade costs T ij depends on traditional factors such as transportation and transaction costs, variables describing the stance of trade policy. The gravity literature discusses the different ways to measure the former variables, usually using geographical distance, a dummy for a common border (adjacency), a dummy for the use of a common language, a dummy for joint membership in a free trade agreement (FTA) or in the World Trade Organization (WTO). Following Combes et al. (2005), we posit that T ij also depends on the stock of information available in countries i and j about business conditions in countries j and i. To be more precise, we write information costs related to transactions between countries i and j as I (m ij, m ji ). We postulate that those costs depend on the share of individuals born in country j in the total population of country i, m ij, and on the share of individuals born in country i in the total population of country j. Leaving the functional form of I (m ij, m ji ) open for the time being, it is reasonable to assume that I (.,.) decreases in the shares of immigrants m ij and emigrants m ji. This specification has the plausible implication that information costs do not depend on the size of the two economies that form a trade relationship. In other words, the information-related tariff equivalent is invariant to a proportional increase in countries total and foreign-born populations. 3 Our specification is in line with the idea that the pro-trade effects of migrants networks are larger the higher the probability to meet a migrant coming from a partner country. Similarly, we assume that the bilateral affinity parameter in the utility function depends on the share of immigrants rather than the absolute sizes of immigrants and emigrants populations. Since we work with a representative agent framework, a higher share of foreign-born individuals in the population means that preferences are more strongly tilted towards the host country of those individuals: a ij = eᾱm ij, ᾱ > 0. (3) 3 Combes et al. specify the function I(.,.) in levels rather than in shares. This implies that for a given composition of the work force in the importer or exporter country, larger countries (who tend to receive and send more immigrants) have smaller iceberg trade costs. 5

6 The formulation implies that there is no systematic bias for imports from any country unless there is a strictly positive stock of foreign-born individuals from that country residing in country i. This captures the home market bias that immigrants may have; it is also consistent with the idea that the presence of immigrants in some country may by its own tilt the preferences of natives towards goods typically consumed by those immigrants. Similar to Combes et al., this formulation disallows for a special preference for varieties produced in countries with a stock of expatriates. 4 3 Econometric Specification and Data Estimation of (2) requires to make assumptions about the functional form of T ij. Moreover, one has to properly account for determinants of bilateral trade, such as the overall price index dual to (1), P i, or the number of varieties produced in every country, n j. We deal with both issues in a way that is well established in the received empirical literature. First, we assume that different types of trade costs all take the iceberg form and enter multiplicatively into the bilateral component of ad valorem trade costs T ij. We write T ij = T ij t χ 1 i t χ 2 j with T ij = I (m ij, m ji ) (1 + D ij ) δ e λ(1 LANG ij ) e γ(1 ADJ ij) e π(1 F T A ij) (4) where D ij denotes geographical distance, LANG ij is the common language dummy, ADJ ij is an adjacency dummy, and F T A ij is a dummy for joint membership in an FTA. The variables t i and t j capture multilateral components of trade costs. For example, if a country possesses a major international sea- or airport, it has lower trade costs with all of its trading partners. Similarly, if a large fraction of the population speaks the lingua franca of international commerce English or is at ease with the use of international currencies such as the US dollar, this should lower trade costs with all partners. Second, the price index P j in the market capacity term complicates the estimation since it depends in a complicated non-linear fashion on T ij and hence on the unknown parameters that govern that function. Moreover, the number, n i, of varieties produced in region i and the mill price, p i in the expression of the supply capacity are not observable. 4 Realistically, immigrants may attach a particular weight to varieties produced in their host countries. In contrast, it is difficult to find a convincing reason why source country consumers should specially value goods produced in the country where emigrants reside. 6

7 As Redding and Venables (2004), we refer to φ i c i Pi σ as the importer s market capacity of country i and to ϕ j n j p 1 σ j the exporter s supply capacity. We follow the established practice and apply a comprehensive set of exporter and importer fixed effects to capture unobservable origin and destination effects (Combes et al., 2005; Feenstra, 2004; Redding and Venables, 2004). The drawback of this method is that estimates of trade cost related coefficients are cross-sectional averages. In theory, those elasticities depend on country characteristics, as the components of iceberg trade costs enter the price index P i (Anderson and van Wincoop, 2003; Feenstra, 2004). 3.1 The standard (STD) model Substituting (4) and (3) into the demand equation (2) and taking natural logarithms we obtain a specification that we refer to as the standard (STD) model (since versions of it are widely used in the literature): ln c ij = (α + µ 1 ) m ij +µ 2 m ji +γadj ij +λlang ij +πf T A ij +δ ln D ij +φ i +ϕ j +ε ij (5) where φ i ln E i + (σ 1) (ln P i + χ 1 ln t i ), ϕ j ln n j + (1 σ) (ln p j χ 2 ln t j ), α (σ 1) ᾱ, γ (σ 1) γ, λ (σ 1) λ, π (σ 1) π and δ (σ 1) δ. Moreover, µ 1 = (σ 1) µ 1 and µ 2 = (σ 1) µ 2, where µ 1 and µ 2 are obtained from positing I (m ij, m ji ) = exp ( µ 1 m ij + µ 2 m ji ). We assume that the error term ε ij has the usual properties. Equation (5) raises two important identification issues. First, identification of the parameters of the trade cost function or of the bilateral affinity formula require external information on σ. Estimates of σ abound in the literature and can be used to transform gravity equation coefficients into ad valorem tax equivalents. 5 Moreover, we cannot separately identify the trade cost channel (µ 1, µ 2 ) from the preference channel α. The next section discusses an indentifying assumption that helps disentangle the two channels. 3.2 The symmetric trade cost (STC) model Identification of α requires to make an assumption about the nature of bilateral trade costs. In particular, we assume that T ij T ji = exp (ζ ij ), ζ ij N (0, σ ζ ), (6) 5 See the recent survey by Anderson and van Wincoop (2004) for a discussion of the available evidence. 7

8 where ζ ij is a normally distributed random variable with mean zero and variance σ ζ. In words, we claim that, on average, lower trade costs encourage more exports and more imports in a symmetric way. Hence, there is no systematic directional bias in bilateral trade flows. finds Using (6) in equation (2), dividing exports by imports, and taking logarithms, one c ij c ji = ( aij a ji ) σ 1 exp (ζ ij) 1 σ E i t 1 σ i P σ 1 i n j (t j p j ) 1 σ n i p 1 σ i E j P σ 1. j Substituting for a ij and using the property of ζ ij we have ln (c ij /c ji ) = α (m ij m ji ) + υ i + ν j + ɛ ij, (7) ) ln ( ni p 1 σ i ), νi ln ( E i t 1 σ i ) ( ) ln ni p 1 σ, ɛi where υ i ln ( E i t 1 σ i P σ 1 i P σ 1 i i (1 σ) ζ ij. Equation (7) relates the excess of country i s imports from j over its exports to the difference in the share of immigrants relative to the share of emigrants and to importer and exporter fixed effects. It allows the identification of the preference channel α = (σ 1) ᾱ. Note that the STC model makes clear that we need bilateral migration rates, i.e., immigration and emigration rates, for a sample of countries large enough to make meaningful inference. Our data has the advantage. Existing studies could not make use of our identifying strategy because they have information either on immigration rates or on emigration rates but not both (Dunlevy, 2006; Gould, 1994; Rauch and Trindade, 1999). We argue that (6) is a sensible assumption. As any identifying assumption, we cannot formally test it. However, there are good reasons to belief that it is not systematically violated. Much of the theoretical and, indeed, parts of the empirical literature make this assumption (e.g., Anderson and van Wincoop, 2003). It implies that the increased availability of information due to migration does not boost exports more than it boosts imports. There are many metaphors that may be used to illustrate this situation: for example, the tunnel under the Channel makes trade costs between Britain and France lower; trains run in both directions at the same speed. Hence, exports and imports benefit similarly. Similar allegories can be found for informational costs: both, firms engaged in exporting and those engaged in importing should make use of improved information and boost transactions. Clearly, there are important limitations to (6). First, σ ζ may be large. This does not per se invalidate our identification strategy, but it makes it harder to obtain statistically significant results. Second, and more importantly, Tij / T ji may deviate systematically 8

9 from unity if countries specialize on different types of goods. For example, trade in goods like printing paper requires less information than trade in car parts. Hence some country i that exports paper and imports car parts from country j may feature T ij / T ji < 1. We deal with this issue by focusing on merchandize trade only (thereby excluding services) and by dividing the spectrum of goods into three subclasses that differ with respect to the degree of differentiation (following Rauch, 1999). Moreover, the use of importer and exporter fixed effects effectively controls for the implicatios of different specialization patterns on trade costs (since t i and t j are absorbed into the fixed effects). Note that equation (7) is different from Combes et al. (2005) odds specification which also accounts for supply side considerations. Our specification is close to Davis and Weinstein (2002) who discuss bilateral trade balances in a gravity framework, without grounding their analysis in the monopolistic competition trade model. They control for fundamental drivers of bilateral trade deficits directly using GDP data. 6 In almost 50% of all cases, their empirical model predicts the wrong sign of the bilateral trade balances. It also displays what they call the mystery of the excess trade balances: their model predicts too little variance in bilateral positions compared to actually observed ones. Thanks to the use of importer and exporter fixed effects, our model is more general. It also has superior fit. It predicts the sign of bilateral trade balances correctly in about 75 percent of all cases. We are therefore confident that the STC model (7) provides a useful shell to investigate the trade effects of migration. To the extent that the taste parameter a ij in (1) is asymmetric, i.e. a ij a ji, bilateral trade positions need not be balanced, even if countries aggregate trade balances are zero. As described in Davis and Weinstein (2002), such a situation involves triangular trade. 3.3 Data and descriptive statistics Comparable data on stocks of foreign-born individuals by country of origin are rare. Immigrants are often defined as individuals with foreign citizenship, so that reported numbers depend on countries naturalization policies. The OECD has recently completed a project which draws on national census data to produce comparable information on the stocks of foreign-born persons by place of birth. The data decomposes stocks of 6 Moreover, they compute predicted trade deficits from fitted bilateral imports derived from a usual gravity equation. 9

10 immigrants according to their level of education and age. We focus on immigrants aged 15 years or older and report results for the total stock of immigrants and the subgroup of highly educated (high-skilled) individuals. 7 The data covers all OECD countries (with the exception of Iceland) and refers to the year However, the data is not balanced: only 652 out of potentially = 812 data points are actually available. The sample size shrinks to 536, since we require that data on emigration and immigration rates is available simultaneously. 8 Our empirical measure of m ij is the percentage of foreign born individuals from source country j in country i s population. Data on the US dollar value of total bilateral imports for the year 2000 are from the IMF s Direction of Trade Statistics. Following Rauch (1999), we aggregate goods into differentiated goods, goods traded on organized exchanges, and goods that display some reference price. We produce those aggregates using SITC 4 digits bilateral trade value data from the COMTRADE statistics cleaned by Feenstra et al. (2005). The other covariates used in the regressions are taken from the website of the Centre d Etudes Prospectives et D Informations Internationales in Pairs (CEPII). 9 The lower panel of Table 1 provides the summary statistics for the information used in the STD and STC model. Since the STC model draws on bilateral trade positions rather than on imports (or exports), it uses exactly half the number of observations than the STD model. The Appendix provides summary statistics on the disaggregate trade data (differentiated goods, reference-priced goods, goods traded on exchanges). Table 1 shows that on average bilateral imports where 0.4% higher than exports. That measure is not zero, because we are working with an unbalanced panel and triangular trade with the non-oecd world is not accounted for. For the same reason, the average net migration balance is not exactly zero neither. Note, however, that all net 7 Highly educated individuals have been enrolled or graduated from tertiary education. 8 The host countries included are (number of non-missing source countries in brackets): AUS(24), AUT(25), BEL(22), CAN(27), CHE(25), CZE(22), DNK(25), ESP(27), FIN(25), FRA(28), GBR (28), GER(18), GRC(25), HUN(25), IRL(24), ITA(27), JPN(9), KOR(5), LUX(20), MEX(20), NLD(10), NOR(24), NZL(24), POL(26), PRT(24), SVK(19), SWE(25), TUR(21), USA(28). 9 The full dataset will be made available on internet. See Dumont and Lemaître (2005) on the construction of the migration data. 10

11 Table 1: Summary statistics, Mean Std. Dev. Min Max STD Model: N = 536 Ln imports; Ln exports Share of emigrants/immigrants (percent) Share of high-skilled emigrants/immigrants (percent) Ln distance Contiguity Ln product of GDPs Common language, dummy Both countries in EU, dummy Both countries in NAFTA, dummy Accession treaties, dummy STC Model: N = 268 Ln (imports/exports) (Share of immigrants - Share of emigrants) (Share of high-skilled immigrants) Share of high-skilled emigrants) Since Imports (c ij ) are constructed from exports of from j, exports and imports are both in f.o.b and by construction have identical summary statistics. The same logic applies for immigration and emigration shares. bilateral positions are on average close to zero, while they exhibit relatively high standard deviations. Hence, there seems sufficient variance both in the dependent and the independent variables used in our STC model to identify the effects of migration. 4 Results We denote by ˆβ ST D EM (σ 1) µ 2 and by ˆβ ST D IM (σ 1)(ᾱ + µ 1) the estimated compound effects of emigration and immigration, respectively, as yielded by the STD model. ˆβST C (σ 1)ᾱ refers to the estimated preference channel, as computed under our identifying assumption in the STC model. In order to retrieve the tariff equivalent of a change in the population of migrants ( µ 1 ) and ( µ 2 ), we use external information on the elasticity of substitution. Following the discussion of available evidence in Anderson and van Wincoop 11

12 (2004), we assume that the elasticity is equal to 6 for aggregate trade, 4 for trade in differentiated goods, and 25 for goods traded on organized exchanges (i.e., homogeneous goods). 10 We organize the discussion of our results as follows. First, to estimate the total effect of migration on bilateral trade, we run the STD gravity model. We use aggregate bilateral trade or the Rauch (1999) sub-aggregates as dependent variables. Moreover, we also check whether the skill structure of immigration matters. In particular, we use shares of immigrants and emigrants computed using all or only high-skilled migrants, respectively. Second, we use the same trade and migration variables in our bilateral trade deficit model. Under our identifying assumption, this model estimates the size of the preference channel. Combining the results of both models, we retrieve our estimate of the trade cost channel. 4.1 The STD Model Table 2 and Table 3 report our results of the STD model, where Table 2 uses total bilateral imports as the dependent variable. In contrast, Table 3 studies the effect of migration on the sub-aggregates proposed by Rauch (1999) and distinguishes between imports of goods traded on organized exchanges (Org.), goods that possess a reference price quoted in trade publications (Ref.), and differentiated goods (Dif.). 11 We report findings for Rauch s conservative aggregation scheme (which minimizes the number of goods that are classified as either traded on an organized exchange or reference priced). Results using the liberal scheme (that minimizes the size of the group of differentiated goods) are similar, in particular for the STD specification, and are available upon request. Specification (S1) in Table 2 reports estimates of the simplest possible model, where 10 Writing iceberg trade costs as ln(t ij ) = ln(1 + τ ij ) τ ij, the coefficients ˆβ ST D EM ST D and ˆβ IM measure the derivative of the tariff equivalent τ ij with respect to the shares of migrants, m ij or m ji, respectively, multiplied by 1 σ. We use this procedure to compute tariff equivalents. 11 We refer to Ref. as homogeneous goods. 12

13 the shares of immigrants and emigrants are the only covariates (besides the comprehensive set of exporter and importer fixed effects and the constant; all not shown). That regression has an R 2 of about 79%. The coefficients for the shares of emigrants and immigrants both come out with positive, statistically and economically significant estimates. An increase of the immigration share by one percentage point, leads to a boost imports from the migrants host country of 56 percent. To gauge the overall importance of the migration variables towards explaining the total variance in the dependent variable, we compute standardized beta coefficients. 12 It turns out that the beta coefficient for the share of immigrants is about 0.16 ( /1.802), while that for the share of emigrants is of comparable size and approximately equal to 0.13 ( /1.802). Columns (S2) and (S3) include the additional covariates vindicated by equation (5). (S2) uses only geographical distance and a contiguity dummy, while (S3) also includes the product of the importer and the exporter countries GDPs, the common language dummy capturing cultural proximity, and a host of variables that are meant to capture the broad orientation of trade policy. It turns out that adding distance improves the F-statistic and R 2 of the regression, while the other covariates have only minor effects. Comfortingly, the elasticity of distance is about 1 in size, which is in line with earlier estimates. The adjacency dummy does not turn out different from zero, which is also a standard finding in samples of rich industrialized countries (like ours). Comparing (S1) to (S2) or (S3), it is striking to see that the coefficients of shares of migrants are approximately halved. This finding suggests as others have found in the literature that the effect of migrants partly substitutes for geographical or cultural proximity. In fact, in our data, distance and common language are strongly correlated to bilateral stocks of foreign-born individuals. However, in terms of beta coefficients, migrants remain important for bilateral trade volumes: 0.08 and 0.05, respectively. Not surprisingly, compared to, the effect of distance, 12 Beta coefficients are defined as the estimated coefficient times the standard deviation of its corresponding independent variable divided by the standard deviation of the dependent variable, which transform the estimated coefficients into units of sample standard deviation. 13

14 which has a beta coefficient of ( /1.802), the role of migration is relatively small. Yet, it does have substantial explanatory power in our bilateral trade flow model. Specification (S4) uses the share of high-skilled migrants in the respective populations, rather than the total stock of migrants in the respective populations. Using the same list of covariates as in specification (S3), the estimated effect of migration comes out more than double that found in (S3). Hence, an one-point increase in the share of high-skilled migrants has twice as strong a trade creating effect that an equivalent increase in the total share of migrants (with, supposedly, an even stronger difference when compared to unskilled migrants.) However, the beta coefficients are 0.06 for immigrants and 0.04 for emigrants ( /1.802 and /1.802, respectively), which is similar in size to the effect found for total migration. This confirms the earlier finding of Herander and Saveedra (2005). Different to us, those authors use data on trade relations of US-states. 13 In specifications (S1) to (S4), equality of the immigration and the emigration effects cannot be rejected. 14 Since ˆβ ST D IM specifications, we impose the restriction β ST D IM ST D and ˆβ EM are indistinguishable statistically in the STD = βst D EM and rerun specifications (S3) and (S4) and (S4). In the restricted models, we find a coefficient of for the share of migrants (S3 ) and of for the share of high-skilled migrants (S4 ). These coefficients compare roughly to those found in the unrestricted model. With an elasticity of substitution of σ = 6 for total bilateral trade, the ad valorem tariff equivalent of increasing the share of migrants by one percentage point is 3.7% (0.185/(6 1)). Table 6 collect the estimated tariff equivalents associated to the tradecost and the preference effects for our different regressions. The coefficient of the share of 13 Interestingly, the contiguity dummy turns out different from zero in (S4). This is not the case for the other specifications and signals that the shares of high-skilled migrants do not covary as strongly with adjacency than shares of the total stock of migrants. 14 The p-values of the F-test of equality of the two coefficients are 0.63, 0.27, 0.16 and 0.21 for specification (S1) to (S4), respectively. 14

15 Table 2: The total impact of migration on bilateral imports Unrestricted Restricted Model Model (S1) (S2) (S3) (S4) (S3 ) (S4 ) Share of immigrants a a a a (0.15) (0.064) (0.062) (0.049) Share of emigrants a b b a (0.14) (0.075) (0.075) (0.049) Share of high-skilled immigrants a a (0.14) (0.17) Share of high-skilled emigrants c a (0.21) (0.17) Ln geographical distance a a a a a (0.055) (0.060) (0.062) (0.048) (0.048) Contiguity, dummy b (0.12) (0.12) (0.12) (0.14) (0.13) Ln product of GDPs a a a a (0.062) (0.067) (0.0076) (0.0076) Common language, a a dummy (0.12) (0.12) (0.12) (0.13) Both countries in EU, c c a a dummy (0.14) (0.14) (0.11) (0.11) Both countries in a a a a NAFTA, dummy (0.40) (0.45) (0.31) (0.31) Accession treaties, dummy (0.18) (0.18) (0.14) (0.14) (0.72) (0.59) (3.12) (3.35) Observations R-squared Robust standard errors in parentheses. a, b, c denotes statistical significance at the one, five, ten percent levels of significance, respectively. All regressions include exporter and importer fixed effects as well as a constant (not shown). high-skilled migrants is equal to This corresponds to a tariff equivalent of about 12.2% (0.612/(6 1)). The strong trade creating effect of high-skilled migration confirms 15

16 within our broad cross-country OECD sample earlier results by Herander and Saavedra (2005). The trade-promoting effects of migrant networks is larger the better the ability of that group to receive and process information on trading opportunities. The estimates of ˆβ ST D EM ST D and ˆβ IM suffer from the fact that they compound the tradecost (information) and the preference channels of migration. Therefore, to the extent that the total effect of migration also captures the preference channel, the estimates discussed above represent upper bounds to the trade cost channel. In our STC specification, we will attempt to resolve that identification problem. Before doing so, we use the Rauch (1999) sub-aggregates rather than total bilateral trade as the dependent variable in our STD regressions. Table 3 reports the findings. In line with expectations, we find several differences between the specifications that explain imports of goods traded on organized exchanges (Org.), goods that have a reference price (Ref.), and goods that are classified as differentiated (Dif.). First, the gravity equation has the best fit for trade in differentiated goods. Second, we find that geographical distance restricts trade in homogeneous goods (Org., Ref.) more strongly than trade in differentiated goods. However, we cannot infer from this result that geographical distance has a lower impact on trade costs for differentiated goods. From the theoretical trade equation (5), the coefficient of distance is (σ 1) δ. Note that homogeneous products have a higher elasticity of substitution than differentiated ones so that the tariff equivalent of distance is difficult to compare across the Rauch (1999) sub-aggregate (as long as δ 0). Turning to the estimates of the share of migrants, table 3 reports positive and statistically significant effects for each Rauch (1999) sub-aggregate. As with the distance coefficient, the size of the estimates is not as such informative on the effect that migrants have on trade costs (or preferences). Within the group of homogeneous goods (Org., Ref.), the elasticity of substitution can be thought to be larger by an order of magnitude as compared to differentiated goods (Anderson and van Wincoop, 2004). It follows that 16

17 Table 3: Homogeneous versus differentiated goods in the STD model Total number of migrants High-skilled migrants Org. Ref. Dif. Org. Ref. Dif. Share of immigrants a a a (0.10) (0.063) (0.064) Share of emigrants b b b (0.11) (0.094) (0.072) Share of high-skilled immigrants b a a (0.34) (0.14) (0.16) Share of high-skilled emigrants c b b (0.31) (0.27) (0.21) Ln geographical distance a a a a a a (0.11) (0.068) (0.070) (0.11) (0.069) (0.072) Contiguity, dummy a b a a (0.21) (0.14) (0.14) (0.20) (0.14) (0.13) Ln product of GDPs a a a a a a (0.10) (0.079) (0.057) (0.11) (0.083) (0.062) Common language, dummy c (0.21) (0.14) (0.11) (0.21) (0.14) (0.11) Both countries in EU, dummy a c a c (0.28) (0.16) (0.13) (0.27) (0.16) (0.13) Both countries in NAFTA, dummy a a (0.66) (0.32) (0.45) (0.73) (0.38) (0.51) Accession treaties, dummy (0.37) (0.22) (0.15) (0.37) (0.23) (0.16) (5.24) (3.98) (2.82) (5.46) (4.20) (3.08) Observations R-squared Robust standard errors in parentheses. a, b, c denotes statistical significance at the one, five, ten percent levels of significance, respectively. All regressions include exporter and importer fixed effects as well as a constant (not shown). the effect of migrants on trade costs is larger for differentiated goods than for the other sub-aggregates. Hence, our results are in line with Rauch and Trindade (2002). 15 When turning to high-skilled migration (the left three columns in Table 3), we find 15 Rauch and Trindade (2002) do not base their model on the monopolistic competition model of bilateral trade. However, their estimates of ethnic links are always larger for differentiated than for homogeneous goods. Hence, allowing for differences in the underlying values of σ would make their results even stronger. 17

18 a picture that is qualitatively similar to the one for total migration. Quantitatively, the estimates of high-skilled migration are about twice as large, which roughly coincides with our results for total trade. Compared to homogeneous goods, the coefficient for differentiated goods is estimated with a higher degree of precision. As above, Table 6 shows tariff equivalents for differentiated and homogeneous goods (traded on organized exchanges, Org.). Under the statistically valid restriction of equal coefficients for shares of immigrants and emigrants, we find a tariff equivalent of 8.8 percentage points for differentiated goods and of 1.15 percentage points for homogeneous goods. 16 For high-skilled migrants, the tariff equivalents are 24.1 percentage points and 2.7 percentage points for differentiated and homogeneous goods, respectively. Again, all these estimates are upper bounds of the trade-cost effect, since they are confounded by the preference effect. 4.2 Results of the STC model The STC model relates bilateral trade deficits to the bilateral difference in migration shares and therefore draws on exactly half the number of observations than the STD model. The ratio of migrants shares have been centered around their means. The results of the STC model for total bilateral trade are shown in Table (4). Specification (S1) reports the results of a regression of bilateral net imports on difference between the immigration shares of the two involved countries. Specification (S2) reports the results using shares of high-skilled migrants only. All regressions include exporter and importer fixed effects. 17 In terms of R 2, the STC specification underperforms the STD model, but this may be 16 Note that the ad valorem tariff equivalent of the information channel of Chinese networks found by Rauch and Trindade (2002) and reported by Anderson and Van Wincoop (2004) is 6%. 17 Obviously, all symmetric determinants of trade, such as geographical distance, common language, or common RTA membership, drop out. 18

19 due to the lower number of observations. The use of exporter and importer fixed effects effectively controls for both countries fiscal and monetary policies, as well as for other (potentially unobservable) supply and demand side determinants. 18 Table 4: Results for the symmetric trade costs model (STC) (S1) (S2) (Share of immigrants c -share of emigrants) (0.060) (Share of high-skilled immigrants share of high-skilled emigrants) (0.18) Observations R squared Robust standard errors in parentheses. a, b, c denotes statistical significance at the one, five, ten percent levels of significance, respectively. All regressions include exporter and importer fixed effects as well as a constant (not shown). In Specification (S1), we find evidence for a statistically significant preference effect for the share of migrants, but fail to find a preference effect for high-skilled migrants. Without making skill-based distinctions, the difference between the immigration and emigration shares has the expected sign and is statistically significant, if only at the 10 percent level. We take the point estimate of at face value and interpret it as the size of the preference channel. Compared to the total effect of found in the restricted model of Table 2, the preference channel amounts to 63% of the total effect. Hence, as shown in Table 6, the tariff equivalent of the trade-cost effect is about 1.4 percentage points (again, assuming an elasticity of substition of 6). Interestingly, a different picture emerges for high-skilled migration. Here, the results of the STC model suggest that there is no preference channel associated to high-skilled migration. Hence, the tariff equivalent of a one percentage point increase in the share of 18 Therefore, we need not include a measure of the real exchange rate into the model. 19

20 migrants amounts to about 12.2 percentage points. High-skilled migrants do not bias the preferences of the representative host country individual towards goods imported from the source country in a discernible fashion. However, high-skilled migrants ease international transactions quite dramatically. In Table 5 we report estimates of the STC models for Rauch (1999) sub-aggregates. Again, we differentiate between an exercise where we do not differentiate across skills, and one where we only look at high-skilled migrants. Regardless of the skill class, there is no preference effect of migration for trade in homogeneous (Org., Ref.) goods. Without accounting for migrants education, there is a positive and statistically significant (at the 10 percent level) preference channel of about for trade in differentiating goods. Comparing this number to the total average effect computed in Table 3 (0.244) 19, we conclude that the preference channel amounts to about 54 percent. Hence, for differentiated goods, the tariff equivalent of a reduction in trade costs induced by an increase in migration by one percentage points leads to a reduction in ad valorem trade costs of about 3.8 percentage points. Our results suggest that the skill class of migrants seems to make an important difference. For high-skilled migrants, there is no evidence for a preference effect, regardless of whether we look at differentiated or homogeneous goods. Hence, the tariff equivalent of information conveyed by migrants is about 21.7 percent for differentiated goods Caveats, Qualifications, and Robustness checks Endogeneity concerns. The received literature makes important efforts in establishing the direction of causality to run from migration to trade, and not vice versa. Basically, 19 Restricted model. 20 The restricted estimate of the effect of migration on trade in differentiated goods is for highskilled migrants. 20

21 Table 5: Homogeneous versus differentiated goods in the STC model Total migration High-skilled migration Org. Ref. Dif. Org. Ref. Dif. (Share of immigrants c share of emigrants) (0.15) (0.076) (0.075) (Share of skilled immigrants share of skilled emigrants) (0.55) (0.26) (0.19) Observations R squared Robust standard errors in parentheses. a, b, c denotes statistical significance at the one, five, ten percent levels of significance, respectively. All regressions include exporter and importer fixed effects as well as a constant (not shown). Table 6: The pro-trade effects of migration: Decomposition of effects Ad valorem tariff equivalents (in percent) of a one-percentage point increase in the bilateral stock of migrants. Agg. Dif. Org. (σ = 6) (σ = 4) (σ = 25) Trade-cost effect Total migration Preference effect none Compound effect Trade-cost effect High-skilled migration Preference effect none none none Compound effect the main argument is that incentives of individuals to migrate to foreign countries depend on cross-country wage differentials. Those might be affected by aggregate (multilateral) trade. However, it is more difficult to see how bilateral trade flows should shape the motivation to migrate. The cross-sectional nature of our data prevents a direct test of the hypothesis that migrants influence trade. Where a time-dimension is available, studies provide evidence 21

22 supporting the asserted direction of causality from migration to trade; see Gould (1994, p.310, footnote 17), Dunlevy and Hutchinson (1999) and Dunlevy (2006) for data on US trade flows and migrants, or Combes (2005) for migration and trade across French regions. Identification. The symmetric trade costs (STC) model, which we use to identify the preference effect of migration on trade, demands more from the data than the standard gravity model (STD), since it draws inference on net bilateral migration and trade positions. As a result, the preference effects are estimated with some imprecision. The scenarios total migration/aggregate trade and total migration/trade in differentiated goods yield preference effects that are different from zero at the 10% level of significance. Therefore, our numerical results should be taken with a substantial pinch of salt. However, even when the estimated preference effects are reduced by one standard error, they still command substantial fractions of the total pro-trade effect of migration. Hence, in some cases, an important fraction of the overall effect of migration on trade actually comes from the preference rather than from the trade-cost (information) effect. Similarly, since we use external estimates of the elasticity of substitution in the calculations of ad valorem tariff equivalents, those number should be understood as suggestive rather than as definitive numbers. Robustness. We have carried out a number of robustness checks. First, qualitatively, our results continue to hold when we use stocks of migrants rather than their population shares as regressors. Second, while we present results only for the conservative aggregation scheme proposed by Rauch (1999), our main conclusions both qualitatively and quantitatively remain unchanged for the STD model. However, the STC model works less well. Finally, we use migrants aged above 15; using the entire populations of migrants essentially does not change the picture. Similarly, using low-skilled migration (i.e., that of individuals with at most primary education) yields results very similar to those obtained for total migration. This is not surprising, because the stock of high-skilled migrants is only a small subset of the stock of total migrants. 22

23 6 Conclusion We examine the impact of ethnic networks on bilateral trade between OECD countries. To that end, we use a new and comprehensive dataset on the stocks of foreign-born individuals by country of birth, recently made available by the OECD. Migrants networks are measured by the stock of immigrants from a partner country relative to the host country population. Assuming that trade costs are symmetric, we are able to separately identify the preference and the information channels of migration. This is important, since only the latter channel can be related to welfare-enhancing trade creation. We find that for total bilateral trade an increase of the migration share by one percentage point amounts to an equivalent reduction of ad valorem tariffs of about 3.7 percent. While that effect is comparable to the findings of Rauch and Trindade (2002), the estimate is the sum of a preference and a trade-cost (information) effect. Exploiting our identifying assumption of symmetric trade costs, we run a bilateral trade deficit model that isolates the preference effect. For total bilateral trade and total migration shares, the preference effect is strong, amounting to an ad valorem tariff equivalent of about 2.3 percent, which leaves 3.7 percent to the trade cost effect. However, the tariff equivalent of high-skilled migration is considerably higher: we find that high-skilled migrants reduce information costs amounting to a tariff equivalent of about 21.7 percent for differentiated goods. We do not find any preference effect for homogeneous goods, which may be interpreted as a validation of our identification strategy. Finally, assuming that the elasticity of substitution amongst varieties in the class of homogeneous goods is of an order of magnitude larger than in the class of differentiated goods, the information channel of migrants is of an order of magnitude smaller for homogeneous goods as compared to differentiated goods. Our analysis provides evidence on an economic mechanism through which the migration of high-skilled individuals might improve world welfare. The reason is that highskilled migrants embody valuable information on their host countries, thereby helping 23