Mexico-U.S. Migration: Do Spatial Networks Matter?

Mexico-U.S. Migration: Do Spatial Networks Matter? Leila Baghdadi December 7, 2005 Abstract Using individual data on Mexican migrants in the United States, this study analysis empirically the role of spatial networks in the migrant s destination choice. It evaluates two concepts of networks: local and spatial networks. The former concerns the attractiveness of locations related to the stock of migrants by area whereas the latter concept is related to the access by the migrants to networks in the other counties. I analyse several levels of networks taking into account migrant s community-of-origin, state-of-origin and nation-of-origin networks. Our results show that spatial networks are more important than local ones. Moreover, I show that a location surrounded by counties with stronger networks is much more "attractive". In addition, community-of-origin networks have a larger impact in migrants location choice than state and nation are more important than state and nation-of-origin networks, both at the local and spatial levels. However, I show that the influence of network access differs with migrant qualifications (skilled/unskilled) and status (legal/illegal): skilled and legal migrant are less attracted by locations with higher network potential. JEL classification: F22, J61 1 Introduction One of most intensive international migratory flow is the one taking place between Mexico and the United States of America. Chiquiar and Hanson (2005) show that Mexico is the largest source country for United States immigration, accounting for 31.3% of net arrivals during the 1990s. According to the US censuses, 4.3 million Mexican-Born individuals were living in the United States in 1990. In the year 2000 this number had grown to 9.2 million, which was equal to 9.4% of Mexico s total population. Because of the economic consequences this migration is a concern both the sending and receiving countries. Several studies have tried to explain the fundamental forces that drive Mexican migrants to the United States. Massey and Espinosa (1997) propose three main elements as motivating immigration from Mexico. The first one is market consolidation symbolized by the North American Free Trade Agreement and the two other causes -social capital formation TEAM, Université Paris 1 1

and human capital accumulation- are directly related to migrant s networks. The present paper focuses on this last driving force. A substantial amount of literature points out that the magnitude of networks externalities are determinant of the location choice of migrants. Bartel (1989) finds that post-1964 U.S immigrants tend to locate in cities with a higher concentration of immigrants of similar nationality. She further shows that the more educated migrants are less geographically concentrated. These findings are also supported by Jaeger (2000) who differentiates between immigrants of different admission status. Phillips and Massey (2000), Winters et alii (2001) and Munshi (2003), confirm these networks effects on the location decision of Mexican immigrants in the United States. These authors stress the fact that families (Winters et al (2001)) and communities-of-origin networks (Munshi (2003)) are even stronger in direct assistance and job process than nation-of-origin networks 1. According to Massey and Espinosa (1997), people who have already established in the U.S. are in a position to help friends and relatives travel northward, cross the border, and obtain work by providing information, contacts, and material assistance. Although number of studies have underscored the importance of networks in migrations, no one has yet taken into account spatial and geographical considerations. In this article, I seek to assess the importance of spatial networks for the immigrant s location choice and to provide some additional explanations to the phenomenon of the clustering of immigrants. I mean by spatial networks the access of migrants in a given US county to migrant s networks established in other American counties. Gravitational literature on demographics emphasizes the role of networks access on the spatial distribution of population. Stewart (1948) identified first the magnitude of this social process that he attributed to demographic gravitation: people need to live close enough together to form a city; each exerts on his immediate neighbors a demographic force of cohesion, which adds to that of demographic gravitation and is associated with additional economic value. and tried to build an abstract index of the intensity of possible contacts with networks. He labelled this measure "Population Potential". Following Stewart s considerations, I could say that migrants need to live close together in order to form not only networks but also spatial networks to get needed information on job opportunities 1 captured by the stock od people from the migrant s country or the number of people speaking the migrant s native language. 2

and to the migrants relevant labor market as proposed by Gross and Schmitt (2003). The latter support the idea that there is a labor market segment in which job characteristics (linked to the home country personal contacts, language etc.) are what matters. Demand for labor in this market segment comes from employers belonging to migrant s networks. In order to examine whether access to networks has a significant positive influence on migration choices, I used two spatial measures inspired from the Harris s (1954) term (a somewhat modified form of Stewart s term). The first tool that I called Networks Potential tries to identify the access by the migrants established in one area to all other county networks. The second one, Neighboring Network s Potential, measures the access by the migrants established in one area to the networks existing in the adjacent counties. In addition, I took into account some "spatial-of-origin" features (community-of-origin, state-of-origin and nation-of-origin) both for the local and spatial networks. I used a Mexican Migration Project data (micro data) over the period 1978-1998 to investigate the relevance of these spatial networks on migrant s location decisions. The paper proceeds as follows: Section (1) describes immigration patterns from Mexico and the data used. The second section discusses the econometric model and some estimation issues. I present in the third section our econometric results. I conclude in section (5). 2 Regional Patterns in Mexican Migration and Their Geographic Distribution in US Large scale migrations from Mexico to the United States first occurred in western Mexico when the first railway reached this region in 1885. At that time, the population on the Texas-Mexico border was small and scattered. Mexican workers were actively recruited, particularly in U.S. mining and agriculture in the early 1900s. This trend continued over the first half of the twentieth century, and especially during the application of the Bracero Accord (temporary work arrangement) from 1942 to 1964. From the 1920s to the 1960s, the states of western Mexico accounted for 44% to 56% of the Mexican migrants to the US (Durand, Massey and Charvet (2000)). Recently, Migration was a central topic when negotiating the establishment of a preferential trade agreement between both countries. As related by Martin (2002), NAFTA was expected to induce job creations in Mexico and thereby to contribute to stop migration into the United States. One of the richest data tracking 3

migration flows between these two countries is the one collected by the Mexican Migration Project (MMP). This survey-based data is fruit of a collaborative research project between the University of Pennsylvania and the University of Guadalajara. 2 Massey and Zenteno (1999) show that the Mexican Migration Project data s estimation is more accurate for western Mexico migration than the Mexico s National Survey of Population Dynamics. 3 They add that although the Mexican Migration Project data is not strictly representative of the states of western Mexico, MMP is nevertheless a good source of reasonably representative retrospective data on documented and undocumented migration to the United States. I use this data for the empirical evaluation of spatial and neighbouring networks as factors of migrations location. 2.1 Data The data collected within the Mexican Migration Project (MMP) is based on an ethno-survey approach, combining techniques of ethnographic fieldwork and representative survey sampling. Interviews were generally completed in December-January when U.S. migrants often return to Mexico. These are complemented with surveys of out-migrants from each community located in the United States. These surveys are performed during the summers following each winter survey. Each year since 1987, two to five communities in these states are surveyed. Each community is surveyed only once. In general, 200 households in each community are selected through random sampling. If the community is small, fewer households are chosen. Communities have been selected based on their diversity in size, ethnic composition and economic development, not because they were known to contain U.S. migrants. The data includes information on the socioeconomic characteristics of the migrant within household, such as age, education and marital status, their migration experience including information on year of migration, costs of border crossing, documentation and location in the United States. The sample used in this paper covers 1411 households head in 71 communities. The communities are all located in western Mexico with the exception of Baja California del Norte. The western Mexican states are Aguascalientes, Colima, Guanajuato, Guerrero, Jalisco, Michoacán, Nayarit, Oaxaca, Puebla, San Luis Potosí, Sinaloa and Zacatecas. These Mexican households head migrated to 61 different American Metropolitan Statistical 2 See Massey et al (1987), Massey, Goldring and Durand (1994), and Massey and Zenteno (1999) for descriptions of the data set. I use the MMP52 version of the data. The data is made available to users at www.pop.upenn.edu/mexmig/. 3 a large nationally representative survey fielded by Mexico s National Statistical Survey. 4

Areas (MSA). It is more appropriate to use MSA s than states as the location unit because in the economic model of location choice the key determinants of that choice are labor market conditions. Metropolitan Statistical Areas are generally viewed as close approximations to homogeneous labor market (Bartel (1989) and Jaeger (2000)). The MSA s vary in geographic unit, some are cities, some are parts of a county, and some are whole counties (See Table (11)) for a list of the locations). Figure 1: Distribution of Mexican Immigrants in the United States The communities differ in their overall concentration. The highest concentration is unregistered by immigrants from the state of Zacatecas (two communities), from Mexico State (one community), from the state of San Luis Potosi (one community) and Guanajuato (one community). Immigrants from the three first communities are highly concentrated in Los Angeles, California whereas immigrants from the State of San Luis Potosi and Guanajuato are concentrated respectively in Chicago and San Diego. Fig (1) presents the stocks of mexican immigrants that each location received and hence helps the understanding of the spatial patterns of networks variables. As showed in the map, Mexican migrants are highly concentrated in the USA states bordering Mexico (in the states of California, Texas, Arizona and New Mexico). In addition, they are concentrated in very close counties in this border zone. This agglomeration of mexican migrants in neighbourhooding counties illustrates the importance of the geographic proximity of migrants to their networks and hence the importance of spatial networks on their location decision. Los Angeles appears as the most attractive region 5

for the Mexican immigrants. In our sample, 60% of migrants are less then 6 years educated and only 5% of migrants are more than 15 years educated. We call this group "skilled migrants". Indeed, 43.03% of skilled migrants are concentrated in Los Angeles, California. In the other hand, 74.64% of Mexican immigrants are undocumented (illegal). 41.23% of legal immigrants live in Los Angeles. Further details concerning the sample can be found in the data appendix. Table (11) shows the regions that I choose in the sample and the number of Mexican migrants each one received. 2.2 Spatial Networks Measures I identified two types of networks: local networks and spatial networks. I mean by local networks the stocks of migrants established in a US location. I determined three levels of local networks according to migrants community-of-origin, state-of-origin and nation-of-origin networks. Using the event history file provided by the Mexican Migration, I calculated for each year t the stock of migrants from the Mexican community m in each U.S. MSA (Metropolitan Statistical Area) j, in t-1. This measure define the community-of-origin network CONet in U.S. MSA j at time t is: T K CONet jt = N imjt 1 (1) t=0 k=0 Where N imjt 1 is a dummy variable that takes the value 1 if an individual i in the Mexican community m migrates to the U.S. MSA j at time t-1. I applied the same procedure to compute state-of-origin networks. I calculated for each year t the stock of migrants from the Mexican state s in each U.S. MSA j, in t-1. The state-of-origin SONet in U.S. MSA j at time t is: T K SONet jt = N isjt 1 (2) t=0 k=0 Where N isjt 1 is a dummy variable that takes the value 1 if an individual i in the Mexican state s migrates to the U.S. MSA j at time t-1. I approximated nation-of-origin networks by the lagged stocks of Mexicans in each U.S. MSA. I used these local networks to construct our key variables : Spatial Networks. I defined two terms: Networks Potential and Neighbouring Network s Potential. The first one measures the access by migrants established in one area to all others county networks and the second measures 6

the access by the migrants established in one area to the networks established in adjacent counties. I used the Harris formulation (1954) to construct these measures: NP j = k K NO k f(d jk ) (3) Here NP j is the network potential for location j, NO k are networks of origin in location k, K is the set of locations. Networks of origin could be here community-of-origin networks (CON et) or state-of-origin (SONet) or nation-of-origin. f(d jk ) is a decreasing function of distance d jk between location j and location k that defining how distance hampers the exactness of information concerning a given network. 4 For Neighbouring Network s Potential (NNP j ), I follow the same idea: NNP j = k K border jk NO k f(d jk ) (4) where border jk = 1 if j and k are contiguous 5. Besides inter-regional distances (d ij ), one needs a proxy for internal distance since the potential term of equation (3) and (4) includes the local transport cost of migrants. The internal distance is proxied by d ii = 2 S i 3 π, where S i denotes the area of the region (Redding and Venables (2004)). This internal distance represent the proximity of migrants to their local networks. In order to monitor other factors that may affect the utility levels associated with a U.S. location; I included several variables to capture the economic and social characteristics of a location in the multivariate analysis. To control for job opportunities and the general level of economic activity, I included total population and wages. I also included the unemployment rate in a U.S. area in order to take account both job opportunities and potential wages. Studies often assume that the probability of choosing a particular location decreases with the unemployment rate in this location (see the discussion in Jaeger (2000)). In the other side, studies on gravitational literature on migration stresses the deterring effect of distance on migration. Distance serves as a proxy for transportation and psychic costs of movement, as well as the availability of information (Greenwood (1975) and Schwartz (1973)). 4 In this study, f(d jk ) is the inverse of the distance. 5 I suppose that border = 1 if j = k. 7

Migration costs have a direct effect on location choice. Most Mexican migrants have a very low income in their home country and the cost of migrating may be an important issue in determining the specific location to migrate. In order to monitor these costs I included a measure of direct (geodesic) distance, in which I assumed that migrants travel the minimum possible distance between two locations 6. The data sources and the expected signs of the explanatory variables are summarized in Table (1) and a detailed description of these variables is given in data appendix. Table 1: Independant variables: data sources and expected signs Variables Definitions Data Source Expected Signs NP Networks Potential lagged MMP and Encarta (?) (community, state and Census Bureau and Encarta nation-of-origin networks NNP Neighboring Network s Potential MMP and Encarta (?) lagged(community, state Census Bureau and Encarta and nation-of-origin networks) LN Local Networks lagged MMP and Census Bureau (+) (community, state and nation-of-origin networks) pop Population Census Bureau (+) ur Unemployment rate Bureau of Labor Statistics ( ) dis Distance Encarta CDROM ( ) wage Wages Bureau of Economic Analysis (+) The independent variables discussed are specific U.S. locations, as dictated by the conditional logit formulation I will discuss in the following section. In addition, I took into account several individual specific variables and examined how these individual dimensions interact with our network variables. In particular, I looked at the interaction of the specific location variables with those concerning skill level and legal status. Migrants with forty or less years of schooling are assumed to be unskilled; those with more than fifty years are considered to be skilled 7. According to Schwartz (1973), education increases a person s capability of obtaining and analyzing published information. Then, more skilled people are expected to be less dependant from networks. This 6 See Data Appendix (A): Distance Calculation. 7 15 years of education corresponds to the 95 percentile of the variable education years of migrants. Several regressions describe a change in migratory behavior only after 15 years of education. 8

hypothesis was confirmed by Bartel (1989). She find that skilled immigrants are less concentrated than unskilled ones. Hence, skilled people are expected to prefer high-wage and most-populated locations. Migrants report themselves whether they migrated legally (documented) or illegally (undocumented). I expect the migrant s use of networks or inclination to follow them will vary depending on these individual characteristics. Because of the difficulties to access labour markets and their lower level of education, illegal migrants may depend more on spatial networks, compared to legal (and more educated) migrants (Bauer et alii (2002)). 3 Multivariate Analysis In this section, I specify the econometric framework that I used to explore and to implement the determinants of Mexican Migration discussed in the precedent subsection (2.2). I describe two models: Part the conditional logit model (3.1) and the Random coefficient model (3.2). 3.1 The Conditional Logit Model To estimate the importance of spatial and neighbouring networks as attracting factors for migration I employ a conditional logit model (McFadden (1978)). Each Mexican migrant i face a choice between J alternatives U.S. counties. For the ith migrant, suppose that the utility of choice j is U ij = Z ij β + ɛ ij (5) Where Z ij is a set of location characteristics, β is the parameter to be estimated, and ɛ ij is the error term. The location choice results from a comparison of the perceived quality of life in the various locations approximated by the utility function. For empirical convenience, I assume that the location choice is designed to maximize utility function. Hence, the probability that a given individual i chooses location j is given by P rob(u ij > U ik ), k j (6) 9

In the conditional logit case, ɛ s are all assumed to be independently and identically distributed Weibull. Then, the probability that individual i choose the U.S. location j can then be written as J P (y i = j) = exp (Z ij β)/ exp (Z ij β) (7) The marginal effects of a change in the location characteristics of a U.S. location j (Z j ) on the probability that a Mexican migrant i will choose that location are given by the derivative of equation (7) with respect to those characteristics (Jaeger (2000)), i.e. j=1 P (y i = j)/ Z j = [( 1 J )(1 1 J )] β (8) Where J = 61. Hence, to obtain average marginal effects, the coefficients reported in our tables have to be multiplied by 0.0164. The assumed distribution of the error term in the construction of the logit model ensures a IIA property (Independence from Irrelevant Alternatives). This assumption implies that the ratio of probabilities of choosing two locations is independent of the characteristics of any third location. However, the test of independence from irrelevant alternatives often fails. In this case, an alternative to the conditional logit model will be needed. One way to relax this homoscedasticity assumption is the Random Coefficient Model 8. 3.2 The Random Coefficient Model The random coefficient model is the most general specification of discrete choice (Greene (2003)). The basis for the random coefficient model is the following utility function based on the econometric model given by Peter Haan (2003): U ij = Z ij β + ɛ ij (9) Again, the vector of observable variables that vary over the alternatives is denoted by Z ij and the error term ɛ ij follows an extreme independant and identic value distribution. The difference between the conditional logit model and the random coefficient model is captured in the vector 8 Another way to relax the IIA assumption is to use a Nested Logit Model as Head and Mayer (2004). Unfortunately, no geographic structure seems to be relevant in this work. 10

of coefficient β i. This difference become obvious when decomposing β i into a fixed and a random part: β i = β + µ i, µ i (0, W ) (10) The random part µ i captures non-observable individual effects, such as taste, which is distributed with mean zero and variance-covariance matrix W. If the variance of µ i turns out to be zero, the random coefficient specification becomes standard logit. In other words, rather than being fixed the coefficients vary over the individuals in the population with density f(β i β, W ), which is described by its mean β and variance W. Because β i cannot be observed and nor estimated but the distribution f(β i β, W ) is known, the parameters to be estimated in the random coefficient model are mean β, which is the fixed part of β, and the variance-covariance W, which describes the distribution of the random part µ i. In the random coefficient specification, the probability to choose alternative j is the integral over all possible values of β i : P (y i = j) = exp (Z ij β i ) [ J j=1 exp (Z ijβ i ) ] f(β i) dβ i, k J (11) This probability is a weighted average of the logit formula evaluated at different values of β i, with weights given by the density f(β i ). Therefore, the random coefficient model is often referred to as mixed logit model (Train (2003)). The distribution of the coefficients f(β i ) follows a normal distribution β i N(β, W ). To better understand the random coefficient model, assume that vector β i consists of a single random variable described by a known density f(β i β, W ). The probability that individual i chooses alternative j is calculated foe every possible value β i. The density f(β i β, W ) determines the weight of each calculated probability in the overall likelihood function. 4 Location Choice Results I began with an assessment of the contribution of our spatial networks in the conventional specification used in the literature. I therefore implemented the conditional logit estimation of the location choice of Mexican migrants in USA. 11

4.1 Conditional Logit Results Table (2) presents the estimation results concerning local networks at the tree different levels: community-of-origin, state-of-origin and nation-of-origin. All specifications shows consistent coefficients for the standard set of variables related to the different location characteristics (unemployment rate, wage, population and distance between home and host location). Table 2: Location Choice of Mexican Migrants in US and Local Networks (conditional logit model) Period 1978-1998 1411 household head choosing between 61 locations Community-of-origin State-of-origin Nation-of-origin Variables (1) (2) (3) ln dis 0.041 0.899-2.589 a (0.756) (0.729) (0.600) ln pop 0.206 a 0.118 a 0.376 a (0.040) (0.040) (0.036) ln ur -0.454 b -0.851 a -0.523 a (0.191) (0.180) (0.147) ln wage 0.152-0.269 1.066 a (0.361) (0.312) (0.297) ln LN 1.392 a 1.083 a 0.837 a (0.028) (0.024) (0.022) N 86071 86071 86071 Pseudo R 2 0.62 0.57 0.39 Note: Standard errors in parentheses. a, b represent respectively statistical significance at the 1%, 5% levels. The probability that migrants choose a particular U.S. location increases with the total population in that location. In the first column of table (2), it appears that the average marginal effect of an increase of population in a given location j by one percent increase the probability of migrant i to choose this location by 0.378% (= 0.206 0.0164). This positive impact reflects that migrants prefer to move to regions with a relatively large labour market, which increases the probability to find a job and to receive relatively high wage. Proximity to the home community are not significatif in the first two columns of table (2) but seems to play also a large role in determining locations and appears to be an important determinant in migrant s location choice in column (3) of table (2). The negative coefficient of the unemployment rate over our three 12

specifications confirms the effects expected. In Table (2), most of the specifications confirm the conventional wisdom that predicts a positive effect for wages and negative effect for unemployment rate. Immigrants are attracted to areas with higher wages and lower unemployment rate. The overall fit of the estimations is consistent with that found in comparable papers using conditional techniques on location choice. The coefficients on these classical variables reported in table (2) are confirmed by the others tables. The results in Table (2) reveal the importance of community-of-origin networks: The average marginal effect of an increase of community-of-origin networks in a given location j by one percent increase the probability of migrant i to choose this location by 2.28% (= 1.392 0.0164). These networks are stronger than state and nation-of-origin ones. This result confirms the findings of Winters et alii (2001) and Munshi (2003) that classify community networks as strong ties and more global (state and nation-of-origin) networks as weak ties. It appears also that stateof-origin networks are greater than nation-of-origin ones. This result confirm Bauer et alii s proposition (2002) : community-of-origin link are more important for the location choice of migrant than ethnic goods captured by nation-of-origin networks. In Table (3), I present the indicators concerning the simple and spatial networks at the community-of-origin level. It appears clearly that spatial networks are greater than simple one. Spatial networks coefficients are significantly positive and their magnitude are higher then simple networks, reflecting the attractiveness of "central" locations to networks in USA. The overall fit of the model is however reduced compared with specification (1). Moreover, table (3) reveals that neighbouring network s potential is greater than networks potential. This result stresses the attractiveness of locations surrounded by areas with higher community-of-origin networks. How substantial is the impact of spatial networks on location choice? Based on β = 2.898 from specification (2) in table (3), the average marginal effect of an increase of Neighbouring network s potential by one percent in a given location j by one percent increase the probability of migrant i to choose this location by 4.75% (= 2.898 0.0164). Based on β = 2.438 from specification (3) in table (3), the average marginal effect of an increase of Network Potential in a given location j by one percent increase the probability of migrant i to choose this location by 3.99%. 9 9 The pseudo R 2 are weak however they are somewhat comparable to others papers focusing on networks (Bauer et alii (2002)). In the other hand, we remark that in all tables when we introduce communities spatial networks, the signs of other variables are improved and correspond much more to our expectations. 13

Table 3: Spatial versus Local Networks at the Community of origin level (conditional logit model) Period 1978-1998 1411 household head choosing between 61 locations Variables (1) (2) (3) ln dis 0.041-3.908 a -3.884 a (0.756) (0.480) (0.480) ln pop 0.206 a 0.939 a 0.940 a (0.040) (0.032) (0.032) ln ur -0.454 b -0.796 a -0.795 a (0.191) (0.120) (0.120) ln wage 0.152 3.840 a 3.848 a (0.361) (0.274) (0.274) ln LN ln NNP ln NP 1.392 a (0.028) 2.898 a (0.484) 2.438 a (0.455) N 86071 86071 86071 Pseudo R 2 0.62 0.21 0.21 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. Table (4) shows us that spatial networks are not so important at the macro level (state and nation-of-origin) for migrant s location choice as for community-of-origin networks. This finding reflects once again the magnitude of community ties. Taking into account the primacy of micro level networks, I will concentrate only on community-of-origin networks in the following tables. I find the same results if I expand the sample to all individuals within households (see Table (9) in appendix). A number of personal characteristics are likely to exert important influences on the individual s location choice. Among these characteristics are his level of education and his legal status. Employment information and job opportunities are both expected to increase with increased education. Then, education may also reduce the importance of network s ties in location decision of skilled migrant because she allows migrants to access to others sophisticated modes of information than the one provided by networks (Schwartz (1973)). We expect that illegal migrant will 14

Table 4: Spatial Networks at the State and Nation of origin level (conditional logit model) Period 1978-1998 1411 household head choosing between 61 locations State-of-Origin Nation-of-Origin Variables (1) (2) (3) (4) (5) (6) ln dis 0.899-3.504 a -3.451 a -2.589 a 1.430 b 2.366 a (0.729) (0.491) (0.490) (0.600) (0.649) (0.630) ln pop 0.118 a 0.895 a 0.912 a 0.376 a 0.703 a 0.780 a (0.040) (0.032) (0.032) (0.036) (0.031) (0.031) ln ur -0.851 a -0.849 a -0.852 a -0.523 a -1.017 a -0.586 a (0.180) (0.123) (0.123) (0.147) (0.150) (0.143) ln wage -0.269 3.866 a 3.922 a 1.066 a 2.044 a 3.104 a (0.312) (0.277) (0.277) (0.297) (0.304) (0.309) ln LN 1.083 a 0.837 a (0.024) (0.022) ln NNP 0.826 a 0.811 a (0.057) (0.023) ln NP 0.780 a 1.383 a (0.057) (0.036) N 86071 86071 86071 86071 86071 86071 Pseudo R 2 0.57 0.23 0.22 0.39 0.38 0.37 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. 15

be more dependent in their location choice to networks than legal one. Because of their status, illegal migrants will be mostly hired by self-employed immigrants from their networks (found job opportunities in the black market or another kind of job and material assistance provided by their communities s contacts). Table 5: Qualifications of Mexican Immigrants and Local and Spatial Networks at the Community of origin level (conditional logit model) Period 1978-1998 1411 household head choosing between 61 locations All Individuals Skilled Unskilled Variables (1) (2) (3) (4) (5) (6) ln dis 0.041-3.879 a 0.322-4.787 b -0.066-3.868 a (0.756) (0.481) (2.910) (2.172) (0.785) (0.494) ln pop 0.206 a 0.930 a 0.482 a 1.354 a 0.192 a 0.912 a (0.040) (0.032) (0.186) (0.160) (0.042) (0.030) ln ur -0.454 b -0.801 a 1.474 b 0.480-0.569 a -0.866 a (0.191) (0.121) (0.723) (0.538) (0.198) (0.124) ln wage 0.152 3.855 a 1.745 2.308 b 0.045 3.929 a (0.361) (0.275) (1.587) (1.122) (0.371) (0.282) ln LN 1.392 a 1.086 a 1.417 a (0.028) (0.091) (0.029) ln NNP 2.487 a 1.412 a 2.639 a (0.338) (0.712) (0.356) N 86071 86071 4941 4941 81130 81130 Pseudo R 2 0.62 0.22 0.60 0.30 0.63 0.21 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. The second two results (3) and (4) columns in table (5) concerns skilled Mexican immigrants whereas the two last columns (5) and (6) concerns unskilled immigrants. The first group appears a little bit more concerned by economic indicators as population and wages than the first group as well when we introduce local networks or spatial networks. In the other side, unskilled group is strongly influenced by the unemployment rate. However, unemployment rate is not significatif for the Skilled group in the second two columns or have an unexpected sign as showed in column (3). This group is largely attracted by higher wages whereas the most important driving force for the second group is the probability to be employed. Local Network s coefficients are 16

not substantially different for the two groups but are more pronunciated for the unskilled one. At the spatial level, table (5) display a great difference between skilled and unskilled immigrant behaviour. For the first group Local and spatial networks have the same impact on their location s decision. However, spatial networks have a great influence on the location choice of the second one. This result support the idea that unskilled migrants are more in need to information about the neighbourhooding labour markets than in the local one whereas skilled migrants are able to drawn information about job opportunities in the local and neighbourhooding markets by themselves. Finally, distance coefficients appears improved and consistent with expected results when I introduce spatial networks. Skilled group appears more negatively influenced by distance than the second one. This result can be explained by the fact that 45.16% of the first group is concentrated in Los Angeles (as we showed in section (2.1)). Table 6: Status of the Mexican Immigrants and Local and Spatial Networks at the Community of origin level (conditional logit model) Period 1978-1998 1411 household head choosing between 61 locations All Individuals Legal Illegal Variables (1) (2) (3) (4) (5) (6) ln dis 0.041-3.879 a -1.817-6.107 a 0.276-3.394 a (0.756) (0.481) (2.048) (1.100) (0.816) (0.535) ln pop 0.206 a 0.930 a 0.185 b 0.980 a 0.208 a 0.925 a (0.040) (0.032) (0.113) (0.083) (0.043) (0.035) ln ur -0.454 b -0.801 a 0.470-0.406-0.603 a -0.858 a (0.191) (0.121) (0.489) (0.300) (0.208) (0.132) ln wage 0.152 3.865 a 0.584 4.207 a 0.126 3.814 a (0.361) (0.275) (1.002) (0.700) (0.388) (0.299) ln LN 1.392 a 1.403 a 1.390 a (0.028) (0.072) (0.030) ln NNP 2.487 a 1.489 b 2.755 a (0.388) (0.650) (0.376) N 86071 86071 12871 12871 73200 73200 Pseudo R 2 0.62 0.22 0.68 0.24 0.61 0.21 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. The second two results (3) and (4) columns in table (6) concerns legal Mexican immigrants 17

whereas the two last columns (5) and (6) concerns illegal immigrants. The first group like the skilled group in table (6) appears more concerned by economic indicators as population and especially wage than the second group. In the other side, illegal group is strongly influenced by the unemployment rate. However, unemployment rate is not significatif for the legal group in the first three columns. This group is largely attracted by higher wages whereas the most important driving force for the second group is the probability to be employed. It appears clearly that spatial networks don t matter for the first group but have a great influence on the location choice of the second one. Local Network s coefficients are not substantially different for the two groups. However spatial networks appears two times more important for illegal group than for legal one. This result correspond to our expectation that illegal immigrants are much more in need to information on job opportunities provided by their networks in the neighbourhooding labor market than legal group. Finally, it appears that legal group is more negatively influenced by distance than the second subpopulation. Two explanations could be provided. First and as for skilled/unskilled immigrants, this result is explained by the fact that a large part of the first group is concentrated in Los Angeles (as we showed in descriptive statistics). Second, unskilled migrants are a large share of the legal group and hence he is greatly influenced by distance. 4.2 Random Coefficient Model Results In this section, we report results for random coefficient model of the location choice of 1411 Mexican immigrants among 61 USA counties over the period 1998-1998. I assume Local networks, neighbourhooding networks potential and networks potential to be random respectively in specification (2), (4)and (6) in Table (7) 10. Estimations in table (7) are restricted to the community-oforigin networks because of their primacy on location choice revealed on the precedent results. In specifications (1), (3) and (5), I present the conditional logit results in order to compare with random coefficient results. The estimations are performed using gllamm (Generalized Linear Latent and Mixed Models) that has been developed by Rabe-Hesketh et alii (2001) 11. The random estimations suggest that unobserved heterogeneity is present in the model: The Akaike Information Criterion (AIC) indicates that the random specification is superior to the 10 In general, in the most flexible specification all coefficients should be varying randomly. However, the more flexible specification have very high computational costs. That s why, we assume only our variable of interest to be random. 11 More information about gllamm can be obtained at www.gllamm.org. 18

specification od the conditional logit model 12. That implies that the variance of the errors terms are not constant, and thus the IIA assumption is violated. Table 7: Spatial versus Local Networks at the Community of origin level (Random Coefficient model) Period 1978-1998 1411 household head choosing between 61 locations Variables (1) (2) (3) (4) (5) (6) Model Conditional Random Conditional Random Conditional Random ln dis 0.041-0.017-3.879 a -3.901 a -3.859 a -3.873 a (-0.05) (-0.02) (8.07) (8.10) (8.02) (8.04) ln pop 0.206 a 0.204 a 0.93 a 0.935 a 0.932 a 0.937 a (5.11) (4.85) (28.86) (28.89) (28.90) (28.90) ln ur -0.454 b -0.454 b -0.801 a -0.801 a -0.8 a -0.801 a (2.38) (2.32) (6.64) (6.63) (6.64) (6.63) ln wage 0.152 0.023 3.855 a 3.839 a 3.862 a 3.851 a (-0.42) (-0.06) (14.03) (13.94) (14.06) (13.99) ln LN 1.392 a 1.478 a (49.99) (32.71) ln NNP 2.487 a 2.659 a (7.35) (6.53) ln NP 2.351 a 2.418 a (7.34) (6.61) AIC 4379.635 4374.377 9112.314 9109.508 9114.488 9112.858 N 86071 86071 86071 86071 86071 86071 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. The coefficients in random specification (2), (4) and (6) are improved relatively to those performed with conditional logit model but there is not changed a great difference. These results leads to the conclusion that the standard discrete choice model, attractive for its simple structure, provides an adequate model choice for the location choice (Haan (2003)). 12 The model with the smallest AIC is the prferred. 19

5 Conclusion I analyze the determinants of location choices by Mexican immigrants in United States. I link the optimal location choice of Mexican immigrants to simple and spatial networks. The results confirm the primacy of community-of-origin ties over state and nation-of-origin one. This finding confirms the importance of this type of networks underlined by the existing literature. Conditional logit estimations also provide evidence on the relationship between location choice and spatial networks. Immigrants are highly attracted by central to community-of-origin networks. Moreover, they prefer areas surrounded by locations where established village networks. These results suggest that the demographic gravitation emphasized in Stewart (1948) explain the clustering phenomena of Mexican Immigrants in the United States. Because of the IIA limitation of the conditional logit model, I use random coefficient model to investigate the importance of spatial networks. Results provided by this model support the standard logit model and confirm the great influence of Spatial networks on Location decision of Mexican Immigrants in the United States during 1978-1998. Finally, spatial networks appear much more determinant in the location choice of unskilled and illegal immigrants than the skilled and legal immigrants. This result is explained by the fact that skilled and legal immigrants are more capable to obtain and analyze published information on local and spatial labour markets and hence are more independant from the information provided by their spatial networks. 20

References Bartel, A. P. (1989). Where do the new u.s. immigrants live? Journal of Labor Economics 7 (4), 371 391. Bauer, T., G. S. Epstein, and I. R. Gang (2002). Herd effects or migration networks? the location choice of mexican immigrants in the u.s. Centre for Economic Policy Research Discussion Paper (3505). Chiquiar, D. and G. H. Hanson (2005). International migration, self-selection, and the distribution of wages: Evidence from mexico and the united states. Journal of Political Economy 113 (2), 239 281. Durand, J., D. Massey, and F. Charvet (2000). The changing geography of mexican immigration to the united states: 1910-1996. Social Science Quarterly LXXXI, 1 15. Greene, H. (2003). Econometric Analysis. New Jersey: Prentice Hall. Greenwood, M. (1975). Research on internal migration in the united states: A survey. Journal of Economic Literature 13 (2), 397 433. Gross, D. and N. Schmitt (2003). The role of cultural clustering in attracting new immigrants. Journal of Regional Science 43 (2), 295 318. Haan, P. (2003). Discrete choice labor supply: Conditional logit vs. random coefficient models. Harris, C. (1954). The market as a factor in the localization of industry in the united states. Annals of the Association of American Geographers (64), 315 348. Head, K. and T. Mayer (2004). Market potential and the location of japanese investment in europe. Review of Economics and Statistics 86 (4), 959 972. Jaeger, D. (2000). Local labor markets, admission categories, and immigrant location choice. http://fsweb.wm.edu/dajaeg/research/wp/immloc.pdf. Martin, P. L. (2002). Economic integration and migration: The mexico-us case. WIDER Conference. Massey, D., L. Goldring, and J. Durand (1994). Continuities in transnational migration: An analysis of nineteen mexican communities. American Journal of Sociology 99, 1492 1533. Massey, D. S., R. Alarcon, J. Durand, and H. Gonzales (1987). Return to azlan: The social process of international migration from western mexico. Berkley and Los Angeles: University of California Press. Massey, D. S. and K. Espinosa (1997). What s driving mexico-u.s. migration? a theoretical, empirical, and policy analysis. American Journal of Sociology 102, 939 999. Massey, D. S. and R. Zenteno (1999). A validation of the ethnosurvey: The case of mexico-u.s. migration. International Migration Review, 766 793. McFadden, D. (1978). Modelling the Choice of Residential Location (A. Karlquist et al. ed.)., pp. 75 96. Amsterdam North-Holland. Munshi, K. (2003). Identification of network effects: Mexican immigrants in the u.s. labor market. The Quarterly Journal of Economics, 549 599. Phillips, J. and D. S. Massey (2000). Engines of immigration: Stocks of human and social capital in mexico. Social Science Quarterly 81 (1), 33 48. Rabe-Hesketh, S., A. Pickels, and A. Skrondal (2001). Gllamm manual. Department of Biostatistics and Computing Institute of Psychiatry, King s College, London. 21

Redding, S. and A. Venables (2004). Economic geography and international inequality. Journal of International Economics 62 (1), 53 82. Schwartz, A. (1973). Interpreting the effect of distance on migration. Journal of Political Economics 81 (5), 1153 69. Stewart, J. (1948). Demographic gravitation: Evidence and applications. Sociometry 11, 31 58. Train, K. (2003). Discrete Choice Models using Simulation. Cambridge, Massachusets: Cambridge University Press. Winters, P., A. Janvry(de), and E. Sadoulet (2001). Family and community networks in mexicou.s. migration. The Journal of Human Resources 36 (1), 159 184. 22

A Data Appendix This appendix briefly describes the procedure analysis variables Total Population: This variable was obtained from the U.S. Census Bureau for the censual years 1970, 1980 and 1990. Wages: Data for selected years 1978-1998 were obtained from the Bureau of Economic Analysis 13. We utilize the average wage per job. The average wage per job is wage and salary disbursements divided by the number of wage and salary jobs (total wage and salary employment). Wage and salary disbursements consist of the monetary remuneration of employees, including the compensation of corporate officers; commissions, tips, and bonuses; and receipts in kind, or pay-in-kind, such as the meals furnished to the employees of restaurants. It reflects the amount of payments disbursed, but not necessarily earned during the year. Unemployment Rate: The most recent information on the number of unemployed and the size of the civilian labor force at the county level was obtained for the years 1974 and 1976-1998 from the Bureau of Labor Statistics, Local Area Unemployment Statistics Division. For the early 1970s, no information by county is available although information on unemployment for the censual years 1960 and 1970 is available. For the years 1971-1973, the assumption was made that unemployment rates in a metropolitan area follow the same trends as that of the state. An estimate of the unemployment rate for 1975 was obtained by averaging the unemployment rates for 1974 and 1976. Migration Costs: Distance calculation To calculate geodesic distance, I converted two sets of latitude and longitude points (the latitude and latitude points for which are taken from ENCARTA CDROM) into Cartesian coordinates and then calculated the minimum-length arc that connect 2 points, and I assume that the Earth is a perfect sphere with radius equal to the means of the polar and equatorial radii (the polar radius is 6.357 kilometers; the equatorial radius is 6.378 kilometers). The Great Circle Distance Formula using radians is: 3963.0 arccos[sin(latitude1) sin(latitude2) + cos(latitude1) cos(latitude2) cos(longitude2 longitude1)]. (12) Local and Spatial Networks at the Community/state/nation-of-origin level: These variables are calculated as indicated in the text from the event history file. To calculate these networks, I use migration information on all individuals within households (7438 individuals for the period 1978-1998) not only migration of the household head. I compute our immigrant stocks using information on migration flows until the year 1906. Source: Mexican Migration Project 52. Nation-of origin (local and spatial networks): Total Mexican population in each US MSA. 13 www.bea.doc.gov 23

B Descriptive statistics Table 8: Data description Variable Obs Mean Std. Dev. Min Max Household Head 86071 706 407.32 1 1411 ln ur 86071 1.869 0.30 0.53 3.11 ln pop 86071 10.86 1.59 6.70 15.13 ln wage 86071 9.75 0.279 8.82 10.7021 ln distance 86071 8.99 0.0766 8.68 9.15 Community-of-origin networks ln LN 86071 0.115 0.603 0 6.003 ln NNP 86071 0.011 0.034 0 2.306 ln NP 86071 0.029 0.040 0 2.306 State-of-origin networks ln LN 86071 0.4158 1.22 0 7.029 ln NNP 86071 0.038 0.373 0 13.37 ln NP 86071 0.1247 0.419 0 13.37 Nation-of-origin networks ln LN 86071 10.26 1.93 0 15 ln NNP 86071 8.63 3.69 0 23.007 ln NP 86071 10.47 2.75 7.71 23.007 24

C Additional estimation results Table 9: Local and Spatial Networks at the Community/state/nation-of-origin levels (conditional logit model) Period 1978-1998 7438 Mexican Immigrants choosing between 61 locations Community-of-origin State-of-origin Nation-of-origin Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) ln dis -1.524 a -4.233 a -4.198 a -1.057 a -3.945 a -3.862 a -3.996 a 0.511 a 1.256 a (0.352) (0.201) (0.201) (0.334) (0.207) (0.207) (0.271) (-0.291) (0.277) ln pop 0.113 a 0.776 a 0.776 a 0.086 a 0.746 a 0.761 a 0.304 a 0.596 a 0.665 a (0.018) (0.013) (0.013) (0.016) (0.013) (0.013) (0.014) (0.013) (0.012) ln wage 0.703 a 5.494 a 5.503 a -0.013 5.44 a 5.545 a 1.542 a 3.335 a 4.717 a (0.165) (0.111) (0.110) (-0.143) (0.112) (0.112) (0.131) (0.128) (0.128) ln ur -0.065-0.666 a -0.663 a -0.343 a -0.724 a -0.727 a -0.11-0.727 a -0.337 a (-0.08) (0.049) (0.049) (0.072) (0.051) (0.050) (-0.058) (0.060) (0.057) ln LN 1.36 a 1.073 a 0.911 a (0.012) (0.010) (0.010) ln NNP 2.697 a 0.934 a 0.871 a (0.165) (0.022) (0.010) ln NP 2.322 a 0.884 a 1.44 a (0.142) (0.022) (0.016) N 453718 453718 453718 453718 453718 453718 453718 453718 453718 Pseudo R 2 0.67 0.21 0.21 0.6 0.23 0.23 0.42 0.4 0.39 Note: Standard errors in parentheses. a and b represent respectively statistical significance at the 1% and 5% levels. 25