International market access and internal migration

Transcription

1 International market access and internal migration Laura Hering Rodrigo Paillacar October 14, 2009 Abstract This paper studies the role of international trade in internal migration patterns. Using regional differences in access to international markets, we analyze bilateral migration between 26 Brazilian states. Bilateral migration rates are obtained from household survey data for the years 1993 to Our findings highlight the importance of external factors, in this case the demand from other countries, in the shaping of domestic migration patterns. Results show that workers migrate to states with higher market access, where higher labor demand leads to better job opportunities. These findings are robust to the control for self-selected migration, the introduction of regional wages, and other control variables. A comparison of migration determinants between different educational levels indicates that international market access plays a more important role for less educated workers. JEL classification: R1, F1, J61. Keywords: Economic geography, migration, international trade, Brazil. We thank Fabian Gouret, Philippe Martin, Sandra Poncet, Jose de Souza, Cristina Terra, Vincent Vicard and participants of the ELSNIT conference (Florence), the Migration and Development Conference (Lille) and the CREST Macro seminar (Paris) for excellent suggestions and fruitful discussions. We also want to thank two anonymous referees for helpful comments. CREST and Centre d Economie de la Sorbonne, Université de Paris 1, CNRS Boulevard de l Hôpital Paris Cedex 13, France. laura.hering@gmail.com Centre d Economie de la Sorbonne, Université de Paris 1, CNRS Boulevard de l Hôpital Paris Cedex 13, France. rodrigo.paillacar@univ-paris1.fr 1

2 1 Introduction Migration should act as an adjustment mechanism for spatial disequilibria. To understand migration patterns and to design successful regional policies, the identification of the macro-economic determinants of the migrants location choice is essential. The aim of this paper is to emphasize the role of international trade on migration by looking at internal migration within Brazil. Even though the link between migration and trade has been at the center of many studies, these have mainly focused on the reverse causality, the impact of international migration on international trade (see for example Gould (1994), Head and Ries (1998), Rauch and Trindade (2002)) or the impact of intranational migration on intranational trade (Combes et al., 2005). Yet, in an open economy, trade should affect migration patterns. When a region increases its exports in response to a higher demand, its labor demand is likely to increase also. Higher labor demand attracts migrants. When imports increase, the effect is more ambiguous. On the one hand, imports may substitute to local production and the following decrease in labor demand may lead to the emigration of workers. On the other hand, for a good to be imported, it is likely that it is cheaper to buy than to produce it. Increasing imports then result in lower living costs which in turn attract migrants. A region s access to export markets and suppliers can thus shape migration patterns. With ongoing globalization, the importance of trade in shaping migration patterns will even increase. The two studies treating the impact of international trade on internal migration we are aware of are Aguayo-Tellez et al. (2008) and Kovak (2008). As in our paper, both studies apply to Brazil. Aguayo-Tellez et al. (2008) analyze one particular aspect of exports by looking at the role of multinational enterprises. They find that migrants are attracted by the growth of employment opportunities in states with high concentration of foreign owned establishments. Kovak (2008) highlights the negative effect of imports on labor demand. He shows that decreases in regional wage levels resulting from tariff cuts will trigger migration between the Brazilian states. This paper concentrates on the export channel of international trade by using regional differences in access to export markets to explain internal migration pat- 2

3 terns. In Brazil, access to international markets varies substantially across its 27 states. Internal migration is exceptionally high: in 1999, 40% of all Brazilians lived in a state different from the state they were born in (Fiess and Verner, 2003). 1 theoretical framework that is well suited to study the link between access to markets and migration is the New Economic Geography (NEG) theory. This theory explains the agglomeration of economic activity and how workers will choose their location following changes in trade. It underlines the importance of the region s proximity to consumers, which is modeled by the region s access to markets. Regions closer to consumer markets (i.e. with higher market access ) experience lower transport costs and therefore enjoy higher income levels. Also, the costs of living in high market access regions are lower because consumption goods incur lower transport costs (Krugman, 1991). The fact that previous work on migration determinants has seldom focused on the trade channel, is partly due to the fact that international trade studies concentrate on labor relocation rather than on migration per se. Using such an approach, researchers explore changes in the employment levels, implicitly assuming that factors are mobile. In the context of NEG models, the effect of market access on employment has been estimated by Hanson (1998) for Mexico and Head and Mayer (2006) for the European Union. The first study looking at the impact of market access on migration in a NEG framework is Crozet (2004). He derives a quasi-structural migration equation from Tabuchi and Thisse (2002) to test how cost advantages due to higher market access shape migration flows within the countries of the European Union. His approach has been adopted by Pons et al. (2007) to explain migration flows within Spain and by Kancs (2005) for migration flows between the Baltic states. 2 1 For a comparison: Dahl (2002) finds that in 1999 in the US, a country known for its high labor mobility, only 30% of all male workers no longer live in the state in which they were born. 2 There is also a long tradition of economic models treating the connection between industrialization processes and rural- urban migration in developing countries, backing at least at Lewis (1954) and Harris and Todaro (1970). Our point is that this literature is not addressing the specific features of a trade channel as NEG models do. In addition, these approaches focus on unilateral rural-urban migration and are thus not suited for analysis where we observe migration flows in different directions. 3 A

4 In our study, we obtain yearly bilateral migration rates between the Brazilian states from a micro data set for the period 1993 to In a first step, we test whether regional differences in access to foreign markets have an impact on bilateral migration rates within Brazil. We further check whether this effect is robust to the introduction of various control variables relating to alternative explanations of migration. In a second step, we investigate whether the sensitivity to market access varies across educational levels. According to Redding and Schott (2003), under certain conditions, highly skilled workers benefit more from a higher market access. If this model correctly describes the case of Brazil, highly educated migrants should react more strongly to market access. Our results indicate that people indeed migrate to states with high international market access. In line with the migration literature, which highlights the importance of spatial wage differentials, we add also regional wage levels as migration determinants. Findings of a significant impact of market access are robust to the inclusion of regional wages and a number of other controls. We find that the sensitivity to market access is lower the higher the educational level of workers. We thus find no evidence for the Redding and Schott (2003) prediction. Altogether, these results suggest that the geography of regional access to international markets can help explain migration patterns within a country. Our paper hence contributes to the literature in at least two ways. We are the first to highlight the link between international trade and internal migration in an economic geography framework. We complement the study from Crozet (2004), who concentrates on the role of domestic market access by focusing on the access to foreign markets. Since we are interested here in the role of trade, we measure access to markets using revealed trade flows, as do Redding and Venables (2004). The use of bilateral trade data between the Brazilian states and the foreign countries reveals both observed and unobserved trade costs and market characteristics, both for foreign countries and for the Brazilian states. These allow us to obtain a structural measure of market access. For the Brazilian states this is of a particular importance. If international market access is estimated as a simple function of the country s income 4

5 weighted by the bilateral distance as often done in the literature, it follows that the Northern regions have a higher market access than the Southern regions. This is because they are closer to the US and the European market. Yet, most of the country s export activity is located in the South. Using trade data thus reflects much better the restricted access to foreign markets of the isolated Northern region than the simple income over distance measure. The second contribution is that we are the first to draw on individual data to demonstrate the link between economic geography and migration. Using bilateral migration rates that are obtained from micro data has several advantages. First, it allows the disaggregation of migration flows and to compare different types of migrants. Previous studies looking at the impact of economic geography on migration have not done this before. Here, we focus on differences in migration behavior across three educational levels. Second, it allows us to take into account the migration literature, which highlights the aspect of a possible self-selection of migrants towards certain destinations. By estimating regional wage levels following Dahl (2002), we control for a possible self-selection bias in this variable. Even though we are looking at bilateral migration rates, thanks to the use of micro data, our empirical strategy is based on the individual migration decision. Our estimated migration equation is derived directly from the utility differential approach that is widely used in the migration literature. The remainder of this paper is structured as follows. Section 2 presents our theoretical framework to identify the link between economic geography and migration. In section 3 we derive the empirical specification. The fourth section describes data sources and the construction of our main variables. It further gives some descriptive statistics on the migration patterns and the states international market access. Section 5 reports results from the empirical analysis. The last section concludes. 5

6 2 Theoretical background: how does market access impacts migration? In this section, we identify the link between trade and internal migration based on NEG theory. This literature combines imperfect competition, increasing returns to scale and transport costs to explain regional disparities and how these effect workers and firms location choice (Krugman, 1991, and Fujita et al.,1999). 2.1 Market access One of the central tenets of NEG theory is the importance of proximity to consumers, as represented by the region s market access. 3 A region with high market access is located close to consumer markets and benefits from low transport costs. A high market access thus gives the region a more competitive position concerning the export of its product. Another consequence of lower transport costs is that firms located in high market access regions can afford to pay higher wages and still break even. This relation between nominal wages and market access is modeled in the NEG wage equation. It follows that wages should be higher at the economic center (where market access is high) and lower at the periphery (where market access is low). Furthermore, most NEG models predict that a region with high market access has a lower price index. This has two origins. First, low transport costs allow lower prices for imported goods. Second, low transport costs attract more firms. A higher number of firms increases competition and thus leads to a downward pressure on prices. The combination of high nominal wages and low price index results in higher real wages for workers in high market access regions. In empirical analysis, however, this last prediction can be contested because regions at the economic center often suffer from congestion costs, as for example high rents, that could increase the price index significantly. Redding and Schott (2003) go one step further by modeling how market access can impact wages for skilled and unskilled workers differently. They argue that if 3 Appendix A derives the formula for market access from the theoretical model. 6

7 skill-intensive sectors have higher trade costs, more pervasive input-output linkages or stronger increasing returns to scale, remoteness reduces the skill premium. As a consequence, these sectors should locate in the economic center, while sectors less demanding in skills should be found in the periphery. The incentives to increase their skills for individuals living in the periphery are thus low. As a result, skilled workers should be disproportionately found in regions characterized by high market access. 2.2 Market access and labor mobility What happens to the economy of location i when its market access increases (e.g. as a result of a decrease in bilateral trade costs)? An increase in market access leads to an increase in the demand for local production and can thus be seen as a positive demand shock. NEG theory proposes two different adjustment channels to this shock: the price and the quantity version (Head and Mayer, 2006). It is possible that the economy adjusts using only one mechanism, but it is more likely that both mechanisms operate simultaneously. The extent to which each mechanism is at work depends on the degree of labor mobility between locations. In the case of low labor mobility, the additional demand generated by higher market access can only partially be satisfied, because of a lack of additional labor force. As a consequence, prices and then wages go up. Increasing spatial price disparities between region i and j are the corollary. This adjustment is called the price version. In this paper, we are interested in the second adjustment mechanism, the quantity version, which occurs when labor mobility is high. Given that regions with better market access can afford to pay higher wages, workers from low market access regions are expected to migrate to regions with higher market access. Of course, the key determinant of migration are real and not nominal wages, but Fujita et al. (1999) argue that if in a given region expenditure for manufacturing goods is high, real wages might be high both because the nominal wage is high and because the price index is low. Workers coming to region i ensure the satisfaction of the additional demand and 7

8 offset the upward pressure on wages. In the case of perfect adjustment by this mechanism, prices and wages are driven down back to their initial level. In the long run, prices stay constant, only the spatial distribution of the labor force and the production changes. Whereas empirical studies on the adjustment by prices have been manifold by now, empirical work assessing the second channel is still rare. 4 For Brazil, Fally et al. (2008) have shown that market access has a positive impact on wages. Hence, demand shocks are adjusted by changes in prices. Nevertheless, the high internal migration in Brazil suggests that in this country the quantity effect could also play a role in the adjustment of local demand shocks. In this paper, we thus test the quantity version by estimating the impact of regional differences in demand on interregional migration. Given free labor mobility in Brazil, will people react to market access differentials and go to regions with better market access? Can the demand coming from the international market impact domestic migration patterns? By including the states nominal wage levels in our regression, we control partially for the price mechanism, the impact of market access on nominal wages. The impact of market access on migration that we then observe is the migration motivated by better job opportunities in regions with high market access. Further, we want to test the hypothesis of Redding and Schott (2003). If remoteness is indeed reducing the skill premium, highly educated individuals should react even stronger to differences in market access. 3 Empirical specification The derivation of the empirical specification of our migration equation presented in this section follows Grogger and Hanson (2008) and Sorensen et al. (2007). We consider a log utility model as in Grogger and Hanson (2008). The individual location choice M kij in a general utility differential approach is considered as: 4 For the price mechanism see: Redding and Venables (2004), Mion (2004), Lin (2005); Hering and Poncet (2009) among others. For the quantity mechanism see: Hanson (1998) and Head and Mayer (2006). 8

9 M kij = 1 if and only if V kij = max (V ki1,..., V kir ), = 0 otherwise. Every individual k coming from location i maximizes his indirect utility V kij across all possible destinations j. The utility V kij can be decomposed as follows. V kij = X ij β + ξ ij + e kij (1) The utility of migrating to region j for an individual k from origin i is determined by X ij, the characteristics of the location j. The product X ij β represents the utility the individual receives from these characteristics, where β is a vector of marginal utilities. The subscript i is included because some characteristics of location j can vary across original locations, as for example bilateral distance. The error term ξ ij represents unobserved location characteristics. X ij β and ξ ij assign the same utility level to all individuals coming from i and going to j. In order to still allow individuals from the same region to choose different locations, we include an idiosyncratic error term that varies across both individuals and locations, e kij. We assume that this error term follows an i.i.d. extreme value distribution. Given that individuals choose the location that maximizes their utility, equation (1) leads to: P r(v kij > V kim ) j m P r(e kij e kim > X im β X ij β + ξ im ξ ij ) j m (2) McFadden (1974) shows that thanks to the i.i.d. extreme value distribution of the individual error term, integrating out over the distribution of the logistic distribution yields the following migration probabilities: P r(m kij = 1) = exp(x ijβ + ξ ij ) Σ J j=1 exp(x ijβ + ξ ij ) (3) Using the methodology proposed by Berry (1994), this probability of migration from i to j can be interpreted as the share of individuals from i migrating to j. As in Sorensen et al. (2007), we then write the share of migrants from i to j, s ij, as 9

10 s ij = P r(m kij = 1) = exp(x ijβ + ξ ij ) Σ J j=1 exp(x ijβ + ξ ij ) and the share of stayers of region i, s ii, as s ii = P r(m kii = 1) = exp(x iiβ + ξ ii ) Σ J j=1 exp(x ijβ + ξ ij ) (4) (5) Dividing equation (4) by equation (5) and taking the log yields: ln( s ( ) ij exp(xij β + ξ ij ) ) = ln = X ij β X ii β + ξ ij ξ ii (6) s ii exp(x ii β + ξ ii ) Adding a time dimension t and replacing the vector X by our location variables of interest (expressed in logarithms), we obtain our migration equation : ln m ijt = ln s ijt s iit = β 1 + β 2 MA jt β 3 MA it + β 4 ŵ jt β 5 ŵ it + ϕ ij + η t + v ijt (7) where the dependent variable m ijt is defined as the ratio of the number of migrants in year t going from i to j over the number of stayers. MA and ŵ correspond to the state s international market access to the state s wage level. The construction of these two variables is described in Section 4. η represents yearly fixed effects and v ijt is a i.i.d. bilateral error term. With the bilateral fixed effects ϕ ij, we attempt to capture time invariant unobserved location characteristics of the destination and the fact that they might vary depending on the state of origin. Additional control variables are income per capita and population. Due to data limitations, we cannot add variables on the states amenities even though we suspect them to play a significant role in shaping migration patterns within Brazil. However, if they are constant over time, they are captured by the bilateral fixed effects. The same applies to information on unemployment rates. As a consequence of controlling for bilateral and time fixed effects, we exploit the panel dimension of our data. The coefficients of market access indicate how a change in the origin s or destination s market access affects the evolution of migration between a given pair of states. A last important feature of our model is the independence of irrelevant alternatives (IIA). This implies that the probability of choosing one state relative to the probability 10

11 of choosing another is independent of the characteristics of a third state. IIA arises from the assumption that the error terms are i.i.d. across alternative destinations. IIA may be violated if two or more destinations are perceived as close substitutes by potential migrants. Hausman and McFadden (1984) note that if IIA is satisfied, then the estimated regression coefficients should be stable across all destination sets. To test whether IIA is violated in our sample, we follow Grogger and Hanson (2008) by re-estimating our model 26 times, each time dropping one of the destinations. Coefficients of our market access and wage variables stay similar across samples, suggesting that the IIA property is not violated here. 4 Data In this section we present the data sources and explain how our dependent and independent variables are constructed. In the last subsection, we will give some descriptive statistics about migration and market access. 4.1 Calculation of migration rates Our migration data set regroups nine rounds of the Brazilian household survey (PNAD). The data for migration covers the years 1993 to 2003, excluding 1994 and 2000 where no survey was conducted. In line with migration literature, we restrict our sample used for the calculation of migration rates to the information of household heads only. We further limit our sample to workers declaring a positive salary, having at least 20 years at the time of the interview, but are not older than 65. For the regression analysis we drop migration rates where either the origin or the destination state is Tocantins, leaving us with 26 out of 27 Brazilian states. Tocantins has officially separated from Goiás in Trade data for this state is not available for all years of our sample and for the first half of the 90s, statistical information on this state is not yet reliable. To determine bilateral migration rates, we use information on the individual s current state of residence and his residence state five years ago. An individual registered 11

12 in a different state today than five years ago is considered as a migrant. Given our interest in bilateral migration within Brazil, individuals having lived abroad at that moment are excluded from our analysis. Our final data set used for the computation of bilateral migration rates then consists of 58,000 (1993) to 74,000 (2003) individuals for each round of the PNAD, summing up to the total of around 700,000 individuals. We will use the same panel for the estimation of regional wage variables, except that we will also include data for This allows us to introduce a wage variable that is lagged by one year without losing In Section 5.1, the dependent variable m ijt is defined as in equation (7) as the log of s ijt /s iit : the number of persons that migrate from state i to j (i j) over those who originate from i and stay in i in year t. Given that we have 26 states and run regressions on seven years in our sample (we lose 2000 and 1995 because we cannot introduce lagged wage variables for these years), the maximal number of migration rates is 26x25x7=4,550. In our final regression, we will end up with only 2,542 migration rates because we do not observe positive migration for all combinations of states in all years. In Section 5.2, we will explicitly treat the problem of zero value migration flows. In Section 5.3, we are interested in differences in the migration behavior across educational levels. Therefore, we classify individuals according to their years of schooling into three groups: primary, secondary and tertiary education. 5 We then calculate for each of the three groups migration rates based on the sample of individuals with the respective educational level, e, where e stands either for primary, secondary or tertiary education. The dependent variable m e ijt is thus defined as s e ijt/s e iit, the number of persons with educational level e that migrate from state i to j (i j) over the number of persons with educational level e who originate from i and stay in i. If we could observe migrants for all origin-destination combination and for each 5 Primary: one to eight years of schooling (corresponding to primary school); secondary: nine to eleven years of schooling (corresponding to high school); and tertiary: more than eleven years of schooling. 12

13 educational level, this would yield 26x25x3=1,950 possible rates for each year, leading to 13,650 possible flows for all seven years. However, for about 70% of the possible origin-destination pairs, we don t observe any migrant. In our regressions, we use only the available 5,115 positive migration rates. Given the high share of zero flows, we compute also immigration shares for each state and educational level. state for each year and for each educational category. This guarantees us to have an observation for each Immig jt is defined as the percentage of all migrants with educational level e in year t that have come to state j: s e jt/ N=26 N=26 i j s e ijt. Here, we have a balanced panel with 26x7=182 observations for each educational level. This measure will be used in Section 5.4 and also in the descriptive statistics in Section 4.4 given that it better summarizes information on migration patterns than bilateral migration rates. 4.2 Calculation of international market access The central variable in our study is the state s access to international markets as defined by NEG theory. The formal definition of market access for a location i is the distance-weighted sum of the market capacity of surrounding locations (Fujita et al., 1999). Market access of each exporting state i can then be written as: where T ij MA i = R j=1 T 1 σ ij mc j, (8) represents the bilateral transport costs, often proxied by the inverse of the bilateral distance between i and j. mc j stands for the market capacity of each surrounding region j and is defined as the region j s expenditure deflated by the price index prevailing in region j. The price index decreases with the number of firms active in j and the price at which they sell. σ corresponds to the elasticity of substitution across varieties. For the formal derivation of the market access definition from the model, see Appendix A. In the construction of our market access variable, we follow Redding and Venables 13

14 (2004). By estimating market capacity and transport costs through regressions on trade flows at the state level, this methodology allows us to obtain for each of the Brazilian states structural estimates of its access to international markets. For each year, we estimate the following gravity trade equation to get estimates for bilateral transport costs and market capacity of each trade partner. 6 Exports ij = n i p ij q ij = n i p 1 σ i }{{} F X i T 1 σ ij mc j }{{} F M j (9) Under the assumption of homogeneous firms, total exports from country i to country j can be written as a function of the quantity q ij sold at the price p ij on market j by the number of firms in i, n i. The first term on the right hand side of the equation, n i p 1 σ i, is considered as the supply capacity of region i. It is defined as the number of firms active in i times the f.o.b. price. Supply capacity and market capacity are country-specific for country i and country j and can be captured by exporter and importer fixed effects, F X i and F M j. Taking logs, our estimated specification of the trade equation yields ln Exports ij = F X i + F M j + δ ln d ij + λ 1 C ij + u ij. (10) Here, bilateral trade costs, T ij, are defined based on variables that deter or enhance trade such as bilateral distance, d ij, and contiguity, C ij. 7 u ij is a random bilateral error term. i and j can be either a Brazilian state or one of the 210 countries in our data set. Hence, a state s access to international markets can be described as 6 Since regressions are run separately for each year, all parameters vary over time. The index t for year t has been dropped for simplicity in this section. Detailed derivation of the trade gravity equation, the market access construction and additional estimations can be found in Paillacar (2006) and Fally et al. (2008), available at 7 Gravity equations often include also common language or colony links. Given that all Brazilian states have the same language and a common colonial link, these variables are not relevant for our purpose here, which is to identify market access differentials. 14

15 countries ( ) ) MA i = exp F M j dˆδ ij exp ( λ1 C ij j i (11) In NEG theory, a country s market access includes, on top of access to international markets, the access to the own country. For a Brazilian state, this measure would include also the access to the rest of Brazil and to the own state. Since our main interest here is the role of international trade, our analysis concentrates on international market access. In the empirical analysis, the use of only the international part of market access has also the advantage that it reduces the risk of endogeneity, which is highlighted by Head and Mayer (2004). In our case, endogeneity between market access and migration can come from the fact that high migration inflows increase local demand for goods and thus local market access. By focusing on the international component of market access, we mitigate the endogeneity concern that has often been addressed to studies using market access including domestic demand. Lastly, there is a very practical reason for why we do not calculate a complete indicator of market access: the unavailability of trade flows between Brazilian states. 8 In the regression analysis, we will proxy domestic demand by the state s income per capita and population. For the annual estimations of the gravity equation we use trade data between Brazilian states and foreign countries and trade flows among foreign countries. 9 Coefficients of distance and contiguity have the expected sign and magnitude and are significant and similar for all years of our sample. 10 Summary statistics of our market access indicator can be found in Table 1. We see that international market access varies a lot between states. In 1991, the lowest 8 Actually, data on trade flows between the Brazilian states exist for Fally et al. (2008) use this data to construct a market access indicator including domestic demand for this year. 9 The first data set comes from the Brazilian Ministry of Trade. The second data source is BACI (Base pour l Analyse du Commerce International) from the CEPII. Information on the country s contiguity are from the CEPII and latitudes and longitudes of the states (to calculate bilateral distances) are from the CEPII and the Brazilian Institute of Geography and statistics (IBGE). 10 The coefficient of distance ranges between -0.4 and The coefficient for contiguity between 0.24 and

16 market access (belonging to Roraima in the North) is only 10% of São Paulo s, where market access is the highest throughout the sample period. 4.3 Region-education specific wages Regional wage differentials are considered as one of the key determinants of migration patterns. Also, as described in Section 2.2, market access can simultaneously impact wages and migration. The simplest way to include a wage variable in our specification would be to take the average wage of each region. There are two reasons, not to do so. First, given that an individual s wage depend crucially on his education, it would be better to allow the wage variable to vary across the three educational levels. Second, as highlighted in the literature, individuals currently living in state j are not a random sample of the population, because similar individuals tend to move to the same locations. This is reflected by the fact that immigrants in a specific region often share common characteristics like gender, ethnic group or educational level. Since these variables are also the main wage determinants, the wage level in a region will correspond to the average characteristics of the people living there. As a consequence, the average wage level is subject to a potential selection bias and is thus not appropriate for our purpose. To avoid this bias, Falaris (1987) proposes to use predicted wages corrected for self-selected migration. Since we have access to the PNAD survey data for the years 1992 to 2003, we can run for each state and each year a wage equation and predict individual wages. In our estimation of the wage equations, we rely on Dahl (2002). He proposes to include the individual s migration probabilities into the state-specific wage equation. For this, we first estimate a multinomial logit model on the individual s location choice. This gives us the probabilities of an individual to migrate to each of the Brazilian states. These migration probabilities are then added as additional regressors in the wage equations and the obtained estimates are used to predict wages for each individual in each of the 26 states Traditional selection bias corrections (like the conditional logit model) are not very well suited 16

17 Our aim is to compare the wage an individual gains in its actual state of residence with the wage she could gain in any other of the 26 Brazilian states. In the computation of average wage rates by state, we follow De Vreyer et al. (2009). Since our migration equation is at the aggregated level, our wage variable for the state of origin i, w i, will be the average of all workers that actually lived in i five years ago (more precisely, the mean of the predicted wages). In contrast, the wage variable for a given destination j, w j, varies according to the state of origin of the migrants. For each destination j, we compute the wage variable w j according to the origin of the residents. The wage variable of Sao Paulo thus differs between individuals having as origin Amazonas and those coming from Acre. For the former we take the mean of the predicted wages for Sao Paulo of all the workers coming from Amazonas, and for the later we take the mean of the predicted wages for Sao Paulo for the individuals with Acre as origin. By doing so, we use the wages of the same individuals for the origin and the destination variable. In section 5.3, we distinguish migration rates and wages according to the educational level of migrants. Our wage variables, w e j and w e i, are obtained in the same way as on the aggregated bilateral level, but allow different values for skilled and unskilled workers. 12 In section 5.4, the destination s wage, wage e j is the mean of the predicted wages for the respective state of all workers of the educational level e. By introducing the state s wage level in our migration equation, we also control for an impact of market access on wages. The market access impact we observe in our regression is the effect on migration beyond the effect on wages. to cases where individual migration decisions imply numerous potential destinations. 12 Note that we are using nominal wages here, while the pertinent variables for migration should be real wages. Unfortunately, information on price indexes covers only main Brazilian cities in a limited number of states. By including dyadic fixed effects in our final regression, we can control at least for time invariant differences in price indices. 17

18 4.4 The pattern of migration and its correlation with market access This section gives some stylized facts and summary statistics about the migration patterns observed in our data and the relation between migration and the state s access to international markets. For a detailed description of internal migration in Brazil see Fiess and Verner (2003). In contrast to other studies which normally use the birth region as the origin, we define migration as a change in the state of residence that took place only within the last five years previous to the survey year. This short time span allows a much more precise analysis of the determinants of migration than migration rates calculated with the birth state. Table 2 reports total migration rates for 1993 and We see that in 1993, 3.7% of the individuals in our data set had changed the state of residence during the last five years. In 2003, migration has slightly declined but is still at 3.2%. Differences in the migration behavior across educational levels are displayed in Tables 2 to 4. First, the higher the educational level, the higher the percentage of workers that migrate. Second, migrants with different educational achievements seem to have different location preferences (Table 3). Migrants with less education tend to move relatively more often to states in the Northwest (21-29), whereas people with higher education prefer the South (41-43), which is known for its good climate and lower criminality levels. 13 The higher the education, the higher also the probability of moving to the Distrito Federal (53), where Brazil s capital Brasilia is located. 13 The Brazilian states are grouped in five macro-regions (see Table 1). This classification is based on the structural and economic development of the different states, regrouping states with similar characteristics. The North is sparsely populated, poor, and largely inaccessible. The Northeast is the poorest macro-region of Brazil with the lowest life expectancy and wages, little access to mineral deposits or navigable rivers, and the highest proportion of low educated persons. The Center-West combines a diverse set of characteristics, mixing poor rural areas, dense forests, and the federal capital city of Brasilia, where income and education levels are high. The Southeast and the South are the most economically developed regions of Brazil. Education levels, income and life expectancy are all high in these regions, and dense highway networks make it easy to get around. These regions offer high economic opportunities and have a high population density. 18

19 These differences in the migration patterns signal that the utility of migrating to a specific state might vary across educational levels. They are also an indicator for the presence of self-selection of migrants within Brazil. Before we come to the regression analysis, we report some simple statistics and graphics to show the link between immigration shares and international market access. Immigration shares are defined here as the described in Section 4.1. Figure 1 plots the share of migrants a state attracts to its level of market access in We can see a positive correlation between the two variables, indicating that the higher international market access, the higher the share of migrants. 14 The map in Figure 2 gives an impression of the evolution of market access during the ten years of our sample. It shows how a state s market access has changed relatively to the market access of the other states between 1993 and The different colors indicate changes in the rank hierarchy in terms of access to markets. States colored in dark grey experienced an important progress in their ranking (advancing at least two ranks). This is the case for most of the inward states in the North (five states) and Center-west (two states) and reflects a catch-up with respect to the traditionally more developed coastal region. Also Rio Grande do Sul and Paraná, states in the South, closer to Argentina and Uruguay, exhibit an important progress in their international market access over the period, which could result from the increase in trade following the establishment of MERCOSUR. Where there are winners, there must be losers: the seven states that lost at least two ranks are colored in light grey. The remaining states (notably São Paulo and Rio de Janeiro) experience no change or a change of one place (colored in intermediate grey). As explained before, we exclude Tocantins from our analysis, so it is left blank. The map in Figure 3 shows which states have experienced a net immigration over the period of our sample. Immigration states are colored dark grey. 15 With the 14 Graphs and correlations are very similar throughout all years in our sample. 15 We consider a state to be an immigration state when net immigration over all years is superior to 200 individuals. A state is considered as an emigration state when net emigration exceeds 200 individuals. 19

20 exception of Espírito Santo at the Coast and the Distrito Federal, all of the states with a high positive net immigration have also experienced an important progress in their international market access. From the category of states with high net emigration, colored in light grey, only two states (Pará and Piauí) are at the same time regions with important progress in market access. Hence, we can observe also a correlation between internal migration and international market access over time. 5 Results In this section, we present panel data estimations for the migration equation presented in Section 3, equation (7). First, we analyze the impact of international market access on bilateral migration at the state level and test for the robustness of our results. In a second subsection, we use bilateral migration rates calculated separately for each educational level to test whether the impact of market access varies between the different education levels. Finally, we concentrate on immigration shares to highlight the importance of international market access. 5.1 Bilateral migration - Panel data estimations In this subsection, our panel variable is defined as the origin-destination state couple. Accordingly, we include dummy variables for each combination of origin and destination state. The dummy for the couple ij (São Paulo - Amazonas) differs from the dummy for the couple ji (Amazonas - São Paulo). These bilateral fixed effects take into account time-invariant specificities concerning migration between two particular states. Next to differences in climate, price indices or institutions that are stable over our sample period, these dummies capture typically also the stock of migrants (the number of migrants originating from region i and living in region j), which is known to have a positive impact on future migrant flows from i to j (Card, 2001). A gravity model of migration, as we use, should normally also include bilateral distance as a proxy for moving costs between two states (Greenwood, 1997). Also these costs, if 20

21 constant over time, are controlled for by dyadic fixed effects. In order to explain bilateral migration, we include each independent variable once for the state of origin and once for the state of destination. Our main interest lies in the state s access to international markets, M A. We expect market access of the origin, MA i, to have a negative impact on migration: the higher MA i, the higher the demand for work. Thus, the individual can attain already a high utility in his home region and is not necessarily motivated to look for a job in another state. For market access of the destination (MA j ) we expect a positive coefficient: the higher this indicator, the more the region attracts people in search for a job in export-oriented industries. All our independent variables are lagged one year in order to reduce endogeneity. Market access is computable only from 1991 on and regional wages from 1992 on. Using a higher number of lags will thus reduce significantly our number of observations. However, we run also regressions with lags of two and three years, which confirm main results. In all estimations, standard errors are clustered at the origin-destinationcouple-level. We report bootstrapped standard errors in all our regressions, since the two-step procedure of the market access computation leads to biased standard errors. Before showing results with dyadic fixed effects, we run regressions with macroregion fixed effects instead of bilateral fixed effects to test for the impact of distance. The negative and significant parameter of the distance shows that migrants tend to stay close to their home. Both market access variables are significant, with the coefficient for origin of higher magnitude than the rest of specifications. In the following we will always introduce bilateral fixed effects, unless otherwise indicated. In the second column of Table 6, we regress bilateral migration rates only over market access. Significant results for both market access variables suggest that this indicator plays indeed a role in the migration choice. The positive impact of MA j indicates that states that experience an increase in this variable attract workers. The negative coefficient of MA i shows that outflows are lower with increasing market access. From the test on the comparison of the magnitudes of both market access 21

22 coefficients we can conclude that absolute magnitudes are not statistically different. The introduction of our wage variable for each state in Column (3) does not alter coefficients or significance of market access. The non-significant impact of the average wage level both for origin and destination are however not uncommon in the literature. In his survey, Greenwood (1997) underlines that empirical evidence has often failed to confirm the importance of wage differentials on migration. According to this author, the non-significant effect of wages could result from the omission of time variant amenities. While our fixed effects control for location-specific features that are constant over time, it is possible that some amenities evolve in the opposed direction of nominal wages. A further explanation for the non significant effect of wage differentials is that they are reflecting spot wage differentials. But migrants might value other job characteristics like stability or career opportunities more than the actual wage. Aguayo-Tellez et al. (2008) have shown at the individual level that the impact of spot wage differentials on the location decision of Brazilian migrants is not robust. Another reason could be due to the fact that the regional average of the predicted wages is too general and neglects a strong heterogeneity between educational levels. In the next subsection, where we distinguish between different educational levels of workers, we will thus use state-education specific wage levels. In the last two columns, we introduce two additional controls, population and GDP per capita. 16 Both variables can be seen as a proxy for local economic activity and are strongly correlated with our market access variable (see Table 5). It is thus possible, that the significant coefficient of our variable of interest captures, at least partially, the impact of the evolution of population or of the per capita income. Indeed, when adding the log of the state s population, the coefficient of market access of the state of origin decreases slightly but stays significant. States with a high increase in population are less likely to see people leave. This can be explained by the fact that a higher population is normally associated with a 16 Data on population and GDP for the Brazilian states are from the IPEA (Instituto de Pesquisa Econômica Aplicada). 22

23 bigger surface and a greater number of cities. It is thus more likely that an individual can find a (better) job within the same state and is less likely to be urged to search for a job further away (keeping in mind that distance has a strong negative impact on the location choice of migrants). An increase in GDP per capita has a less important impact than an increase in population and does not alter much the magnitude of market access. Also here, coefficients are according to our expectations. Richer regions are much less likely to send migrants. 5.2 Bilateral migration - GMM and ML estimations We see that also when including additional controls, the coefficient of the destination s market access remains the same. International market access could thus be considered as a good indicator for export-oriented economic opportunities. There are however still some concerns about the robustness of our results linked to possible omitted variables and the high number of missing trade flows that could affect our results. First, there is a concern of reverse causality. A region s market access is likely to increase when new migrants come to the region. Regressing only international market access on internal migration reduces this causality concern. There is however still a risk of some time-varying unobservables like changes in amenities that could be correlated with the error term. To tackle these issues of endogeneity, we perform a two-step GMM estimator applied to first differenced data for the regression reported in Table 6 Column (5). 17 GMM results are presented in Table 7. Conditions for the validity of GMM are met: the transformed error terms exhibit first-order correlations, and the test rejects second-order correlations. Instruments are not rejected by the Hansen and Sargan tests. In the interest of space, we report only the coefficients for the lagged dependent variable and market access. The three regressions presented are identical except for the number of lags chosen for the instruments. Following the recent literature (Roodman, 2009) we minimize the use of instruments to avoid over-fitting. For instance, 17 Given that wages are not significant we exclude them from our GMM regressions. 23

24 the first column shows the regression resulting from using only the third and fourth lags as instruments. Standard errors were estimated by using the correction proposed by Windmeijer (2005). The first lag for our migration variable is never significant, suggesting the absence of persistence in bilateral migration flows after having controlled for origindestination-pair fixed effects. Whilst more imprecisely estimated, coefficients for destination market access of the GMM estimations are supportive of the results found in our panel regressions above. Regarding origin market access, we fail to find significant coefficients in the GMM regressions. As a last test for the robustness of our market access indicator, we address the concern of zero-value flows. Both the survey data from which we obtain migration rates and the trade data we use to compute our market access variable are characterized by a high number of zero-value migration flows. In this context, OLS regressions on our gravity equations can be criticized because they lead to biased estimates in the presence of many zero value flows. A high number of zero value trade flows is likely to result in biased estimates of the parameters used to compute market access. To treat zero value flows in a gravity equation researchers often recur to non-linear estimations (e.g. Gamma or Poisson regressions) that allow to deal with some specific problems of trade data. A definitive method on how the problem of zero flows should be treated has however not yet emerged (Silva and Tenreyro (2006); Martinez-Zarzoso et al. (2007); Martin and Pham (2008); Helpman et al. (2008); Silva and Tenreyro (2009)). The question is even more difficult to answer when non-linear estimations are combined with location fixed effects (Buch et al., 2006). When the proportion of zeros becomes too important, iterative methods for Poisson or Gamma estimations often do not converge. In order to explore the sensibility of results to a nonlinear estimation including zeros, we reduce the dimensionality of the zeros, by dropping all countries that trade with less than 20 partners. The proportion of zeros then falls to only maximal 50% of missing trade flows. Given the limits of this method, we will go back to the market access indicator obtained by OLS estimations as our main variable in 24

25 the following sections. In Table 8, we report Poisson and Gamma regressions on all possible bilateral migration flows, including migration rates that take the value zero in our data set. Two kinds of location fixed effects were considered: macro region fixed effects and origin and destination state fixed effects. Regressions with origin-destination-pair fixed effects as used in the panel regressions reported in Table 6 do not converge and are thus not displayed here. The absence of dyadic fixed effects, allows the inclusion of the distance coefficient. In the first four columns, we use the same market access variable as in Table6, but the migration equation is estimated using Poisson or Gamma methods including all zero value migration flows. We see that the count data methods confirm a positive impact of the destination s market access. However, including the zeros leads to a non significant impact of the origin s market access. In the second half of the table (columns 5 to 8), we repeat the same regressions, but use a market access variable that is computed including zero trade flows. Hence, estimations in both stages are made using count data methods. We can conclude that regressions using non-linear estimations, as well as the use of a MA measure including zero trade flows confirm most of the results given by the previous OLS results reported in Table 6. Results here are also in line with those found in the GMM regressions, which suggest that the destination s market access is robust across specifications. 5.3 Bilateral migration by education In this section, we split our sample according to the three levels of education. We identify our panel variable as the origin-destination pair for a given educational level and introduce corresponding fixed effects (origin-destination-education dummies). In the first column of Table 9, we pool all three types of migration rates and repeat the estimation in Column (5) of Table 6. Results for market access are quite similar to those found at the aggregated level. The wage variable used in this subsection is the mean of the predicted wages for each 25

26 state by level of education. It varies between primary, secondary and tertiary education and corresponds therefore much better to the expectations of each individual. In fact, in Column (1) of Table 9, the coefficients have the expected sign, and wages for the state of destination are significant at the 5% level. In the following columns, we split our sample according to the three educational levels to see whether one category of migrants is more sensitive to changes in international market access than the others. Fally et al. (2008) show that for Brazil the impact of international market access varies between workers with high and low education by finding a stronger impact on wages of low educated workers. It is therefore likely that the impact of market access on migration varies also across educational levels. According to Redding and Schott (2003), who predict a higher wage premium of market access for skilled workers, we could expect that highly educated workers have a stronger incentive to go to regions with good access to foreign markets. In this case, the coefficient of market access of the destination should be higher for highly educated individuals. But the opposite effect is also possible: highly educated people might react less to differences in market access because, thanks to their skills, they are more likely to find a job in the region they like most. Highly educated individuals are known to be more sensitive to non-pecuniary location characteristics like pollution, crime rate, schools, parks etc. For example, Levy and Wadycki (1974) have shown for Venezuelan data that educated individuals tend to value amenities much more than low qualified individuals and Schwartz (1973) argues that the negative impact of distance on migration flows diminishes with educational attainment. The regression results for primary, secondary and tertiary education are reported in Columns (2) to (4) of Table 9. Economic opportunities associated to international trade are most important for the location choice of low educated individuals. We see that the impact of the destination s access to markets decreases with education and is even non significant for secondary and tertiary education. Results in Column (1) and in Table 6 seem to be mainly driven by the sensitivity to market access of 26

27 less educated workers. Note however that the numbers of observations for secondary and tertiary are lower. Coefficients are therefore likely to be of expected signs but estimated imprecisely. This can lead to non significant parameters. For all educational groups, population size is a significant determinant of bilateral migration rates. Even though this variable is highly correlated to market access (see Table 5), it does not impact the significance of our market access indicator. Estimating standardized coefficients show that market access is the second important push and pull factor after population. The splitting of the sample leads to positive and significant coefficients for the destination s wage level for workers with secondary and tertiary education the wage of the state of destination has a positive and significant impact. Only wages for workers with primary education are contrary to our expectations, with the wage of the origin being positive and significant and the destination s wage non significant. According to these results, high wages are important for educated workers, whereas less educated workers, who are more likely to be unemployed, value more the job opportunities. The differences in migration patterns between educational levels can thus be explained in part by their different sensitivity to market access and wages. The migration literature cited above highlights the fact that with a higher educational achievement non-pecuniary benefits are crucial for the location choice. Unfortunately, we do not have data on amenities so that we cannot directly control for their effects. Of course, it is also possible that international market access and amenities evolve simultaneously and into the same direction. In this case, we risk to overestimate the effect of market access. But given that market access is not significant for individuals with tertiary education, who are known to react the most sensitively with respect to amenities, the coefficients on market access are unlikely to be biased by omitted amenities. 27

28 5.4 Immigration shares The previous sections established that the direction of migration is following regional differences in international market access. To give a broader picture on the impact of market access, this section concentrates only on migrating workers. We are looking at the share of the whole pool of migrants that a state attracts and test whether a high international market access increases this share. Immigration shares obtained from our data show important shifts during the period under study. For example the share of migrants São Paulo has attracted has been reduced from 14.7% in 1993 to 10.9% in 2003; the share of migrants going to Goiás has increased from 5.4% to 8.5%. As discussed in Section 4.4, our data suggests a strong correlation between the state s international market access and its immigration share. The dependent variable used in Table 10 is the state s share in all migrants. Since we are interested in the share of migrants attracted by a given state, regressions include only the characteristics of the state of destination. Our panel variable is defined as the destination-education combination. In Column 1, we find that also when controlling for population, wage, gdp per capita and destination-education and time fixed effects, market access seems to be an important pull factor for all educational categories of migrants. Splitting our sample according to the three different educational levels confirms results from the previous section: the higher the educational attainment the lower the sensitivity to market access. The wage variable has the expected positive and significant impact on the attraction of migrants for tertiary education. Given that regressions on immigration shares cannot control for distances between states, the impact of certain variables can be biased. It is thus important to take into account the state s remoteness even when looking at immigration rates. Market access, which measures the state s relative location, is significant for all levels of education. 28

29 6 Concluding remarks In this paper, we analyze the role of international trade in internal migration patterns using an economic geography framework. Bilateral migration rates between the Brazilian states are explained by differences in the state s access to international markets. The use of individual data allows us to control for a self-selection of migrants and to disaggregate migration rates and look at specific groups of migrants. Here, we are focusing on differences in the migration behavior across three educational levels. We find that international market access, especially that of destinations, has indeed a significant impact on the migration pattern within Brazil. These findings highlight the importance of external factors, here the demand from other countries, in the shaping of domestic migration. When splitting our sample according to three educational levels (primary, secondary and tertiary education), we see that individuals with primary education react much stronger to changes in market access. This heterogeneous impact of market access across can thus contribute to the explanation of differences in migration patterns across educational levels. References Aguayo-Tellez, E., M.-A. Muendler,, and J. P. Poole (2008). Globalization and formal sector migration in Brazil. WIDER Research Paper 22, UNU-WIDER. Berry, S. T. (1994, Summer). Estimating discrete-choice models of product differentiation. RAND Journal of Economics 25 (2), Buch, C. M., J. Kleinert, and F. Toubal (2006). Where enterprises lead, people follow? Links between migration and FDI in Germany. European Economic Review 50 (8), Card, D. (2001, January). Immigrant inflows, native outflows, and the local labor market impacts of higher immigration. Journal of Labor Economics 19 (1),

30 Combes, P.-P., M. Lafourcade, and T. Mayer (2005). The trade-creating effects of business and social networks: evidence from france. Journal of International Economics 66 (1), Crozet, M. (2004). Do migrants follow market potentials? An estimation of a New Economic Geography model. Journal of Economic Geography 4 (4), Dahl, G. (2002). Mobility and the return to education: Testing a Roy model with multiple markets. Econometrica 70 (6), De Vreyer, P., F. Gubert, and F. Roubaud (2009). Migration, Self-selection and Returns to Education in the WAEMU. Journal of African Economies Forthcoming. Falaris, E. M. (1987). A nested logit migration model with selectivity. International Economic Review 28 (2), Fally, T., R. Paillacar, and C. Terra (2008). Economic geography and wages in Brazil: Evidence from micro-data. Technical report. Available at Fiess, N. and D. Verner (2003). Migration and human capital in Brazil during the 1990s. World Bank Policy Research Working Paper 3093, World Bank. Fujita, M., P. Krugman, and A. Venables (1999). The Spatial Economy: Cities, Regions, and International Trade. Cambridge: MIT Press. Gould, D. M. (1994). Immigrant links to the home country: Empirical implications for U.S. bilateral trade flows. The Review of Economics and Statistics 76 (2), Greenwood, M. (1997). Internal migration in developed countries. In M. Rosenzweig and O. Stark (Eds.), Handbook of Population and Family Economics, Volume 1, Chapter 12, pp Amsterdam: Elsevier. Grogger, J. and G. H. Hanson (2008). Income maximization and the selection and sorting of international migrants. NBER Working Paper

31 Hanson, G. (1998). Regional adjustment to trade liberalization. Regional Science and Urban Economics 28 (4), Harris, J. and M. Todaro (1970). Migration, unemployment and development: A two-sector analysis. American Economic Review 60 (1), Hausman, J. and D. McFadden (1984). Specification tests for the multinomial logit model. Econometrica 52 (5), Head, K. and T. Mayer (2004). The Empirics of Agglomeration and Trade. In V. Henderson and J. F. Thisse (Eds.), Handbook of Regional and Urban Economics, Volume 4, pp Amsterdam: Elsevier. Head, K. and T. Mayer (2006). Regional wage and employment responses to market potential in the EU. Regional Science and Urban Economics 36 (5), Head, K. and J. Ries (1998). Immigration and trade creation: Econometric evidence from Canada. Canadian Journal of Economics 31 (1), Helpman, E., M. Melitz, and Y. Rubinstein (2008). Estimating Trade Flows: Trading Partners and Trading Volumes. Quarterly Journal of Economics 123 (2), Hering, L. and S. Poncet (2009). Market access impact on individual wages: Evidence from China. Review of Economics and Statistics. Forthcoming. Kancs, D. (2005). Can we use NEG models to predict migration flows? An example of CEE accession countries. Migration Letters 2 (1), Kovak, B. (2008). Regional labor market effects of trade policy: Evidence from Brazilian liberalization. Mimeo. Krugman, P. (1991). Increasing returns and Economic Geography. Journal of Political Economy 99 (3), Levy, M. B. and W. J. Wadycki (1974). Education and the decision to migrate: An econometric analysis of migration in Venezuela. Econometrica 42 (2),

32 Lewis, W. (1954). Economic development with unlimited supplies of labour. The Manchester School of Economic and Social Studies 22, Lin, S. (2005). International trade, location and wage inequality in China. In R. Kanbur and A. Venables (Eds.), Spatial Inequality and Development. Oxford University Press. Martin, W. and C. S. Pham (2008). Estimating the Gravity Equation when Zero Trade Flows are Frequent. MPRA Paper 9453, University Library of Munich, Germany. Martinez-Zarzoso, I., F. Nowak-Lehmann, and S. Vollmer (2007). The Log of Gravity Revisited. Center for European, Governance and Economic Development Research (cege) Discussion Papers 64, University of Goettingen (Germany). McFadden, D. (1974). Frontiers of Econometrics, Chapter Conditional Logit Analysis of Qualitative Choice Behavior, pp Academic Press. Mion, G. (2004). Spatial externalities and empirical analysis: The case of Italy. Journal of Urban Economics 56 (1), Paillacar, R. (2006). Market potential and worker heterogeneity as determinants of Brazilian wages. Working paper, University of Paris 1. Available at Pons, J., E. Paluzie, J. Silvestre, and D. Tirado (2007). Testing the New Economic Geography: Migrations and industrial agglomerations in Spain. Journal of Regional Science 47 (2), Rauch, J. E. and V. Trindade (2002, February). Ethnic Chinese networks in international trade. The Review of Economics and Statistics 84 (1), Redding, S. and P. K. Schott (2003). Distance, skill deepening and development: Will peripheral countries ever get rich? Journal of Development Economics 72 (2),

33 Redding, S. and A. Venables (2004). Economic geography and international inequality. Journal of International Economics 62 (1), Roodman, D. (2009). How to Do xtabond2: An Introduction to Difference and System GMM in Stata. Stata Journal 9 (1), Schwartz, A. (1973). Interpreting the effect of distance on migration. Journal of Political Economy 81 (5), Silva, J. M. C. S. and S. Tenreyro (2006). The Log of Gravity. Review of Economics and Statistics 88 (4), Silva, J. M. C. S. and S. Tenreyro (2009). Comments on The Log of Gravity Revisited. Manuscript, London School of Economics. Sorensen, T., P. Fishback, S. Allen, and S. Kantor (2007). Migration creation, diversion, and retention: New Deal grants and migration ( ). IZA Discussion Papers 3060, Institute for the Study of Labor (IZA). Tabuchi, T. and J.-F. Thisse (2002). Taste heterogeneity, labour mobility and economic geography. Journal of Development Economics 69 (1), Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics 126 (1),

34 Appendix A We consider a monopolistic competition framework with product differentiation, including firm-level increasing returns to scale and trade costs. The agricultural sector produces a homogeneous agricultural good (A), under constant returns and perfect competition, which is freely tradeable. The manufacturing sector M produces a large variety of differentiated goods, under increasing returns and imperfect competition, where each firm produces a different variety. In both sectors, labor is the only factor of production. All consumers of region j share the same Cobb-Douglas preferences for the consumption of the goods A and M: U j = M µ j A1 µ j, 0 < µ < 1, (12) where µ denotes the expenditure share of manufactured goods. M j is defined by a constant-elasticity-of-substitution (CES) sub-utility function of n i varieties: M j = R ( ) σ/(σ 1) n i q (σ 1)/σ ij, σ > 1, (13) i=1 where q ij represents demand by consumers in region j for a variety produced in region i and σ is the elasticity of substitution. Given the expenditure of region j (E j ) and the c.i.f price of a variety produced in i and sold in j (p ij ), the standard two-stage budgeting procedure yields the following CES demand q ij : q ij = µ p σ ij G σ 1 j E j, (14) where G j is the CES price index for manufactured goods, defined over the c.i.f. prices: [ R ] 1/1 σ. (15) G j = i=1 n i p 1 σ ij In contrast to the agricultural good, transporting manufactured products from one region to another is costly. The iceberg transport technology assumes that p ij is proportional to the mill price p i and shipping costs T ij, so that for every unit of 34

35 good shipped abroad, only a fraction ( 1 T ij ) arrives. Thus, the demand for a variety produced in i and sold in j eq. (14) can be written as: q ij = µ (p i T ij ) σ G σ 1 j E j. (16) To determine the total sales, q i, of a representative firm in region i we sum sales across regions, given that total shipments to one region are T ij times quantities consumed: where q i = µ R j=1 (p i T ij ) σ G σ 1 j MA i = R j=1 E j T ij = µp σ i MA i, (17) T 1 σ ij G σ 1 j E j, (18) represents the market access of each exporting region i (Fujita et al., 1999). The market capacity of each region j, mc j, is defined as G σ 1 j E j. 35

36 7 Tables and Figures 36

37 Figure 1: Correlation between international market access and immigration Immigration share and market access in lxm_j_migrants ln(ma_j) Fitted values 37

38 Figure 2: Variation of international market access between 1993 and 2003 Figure 3: Net migration between 1993 and