Mexican Migrants to the US: What Do Unrealized Migration Intentions Tell Us About Gender Inequalities?

World Development Vol. 59, pp. 535–552, 2014 Ó 2014 Elsevier Ltd. All rights reserved. 0305-750X/$ - see front matter www.elsevier.com/locate/worlddev

http://dx.doi.org/10.1016/j.worlddev.2014.01.036

Mexican Migrants to the US: What Do Unrealized Migration Intentions Tell Us About Gender Inequalities? ISABELLE CHORT * PSL, Universite´ Paris-Dauphine, LEDa, Paris, France Summary. — This paper exploits unrealized intentions to migrate to highlight mobility constraints. I analyze the discrepancies between Mexicans’ intention to migrate and their subsequent migration behavior using the two waves of the Mexican Family Life Survey panel (2002 and 2005–06). I first provide evidence that intentions contain behavioral information. Controlling for various shocks likely to affect the migration decision, I find that women’s probability to carry out their migration plans is systematically lower than men’s. Different interpretations are investigated, but empirical evidence suggests that women’s unrealized migration plans are due to femalespecific costs and constraints. Ó 2014 Elsevier Ltd. All rights reserved. Key words — migration, gender inequalities, intentions, shocks, Mexico

1. INTRODUCTION

affect the allocation of household resources and the nature of investments, even though Parrado and Flippen (2005) show in the Mexican context that migration does not mechanically result in women empowerment. Second, unfulfilled migration intentions may negatively affect mental health: in an experimental framework, comparing successful to unsuccessful Tongan applicants to a random ballot offering migration opportunities to New-Zealand, Stillman, McKenzie, and Gibson (2009) indeed show that migration improves mental health, especially for women. Third, constraints to women’s emigration may dampen the impact of migration, through transfers of norms, on education and fertility (Beine, Docquier, & Schiff, 2009) or female political empowerment (Lodigiani & Salomone, 2012). This paper relies on intention data which have seldom been used by economists since Manski (1990). Applied here to migration, the analysis of unrealized intentions provides a unique opportunity to identify constraints to individuals’ mobility. Consistent with the best-case hypothesis that individuals’ intentions are based on rational expectations (Manski, 1990), the non-realization of their migration plans could be explained by two types of factors. First, unexpected shocks may affect individuals between the moment when they state their intention and the moment when their behavior is determined, leading them to reconsider their decision given their new environment. 1 This paper investigates the impact of different types

The third Millennium Development Goal claims the importance of gender equality and women empowerment for development and poverty reduction. Particular emphasis is being placed on eliminating gender disparities in access to education, which are found to have a detrimental impact on growth (Abu-Ghaida & Klasen, 2004), and on the labor market. Although less investigated, the question of an equal access of men and women to migration opportunities, is particularly relevant. Indeed, although the feminization of migration flows has been acknowledged by scholars and institutions in the last decades, women’s mobility is still more constrained than men’s in many developing countries. In the context of Mexico, empirical evidence tends to show that women’s access to international migration is still partly subordinate on men’s (Donato, 1993; Kanaiaupuni, 2000; Curran & Rivero-Fuentes, 2003). Existing studies, however, are based on observed migration flows. This paper pushes the analysis of gender-specific constraints on women’s migration one step further by exploiting data on both intentions to migrate and actual migrations. The comparison between intentions and actual moves shows that women have both lower intention to migrate and a lower propensity to realize their migration plans than men. This paper thus uncovers gender-specific constraints affecting the realization of women’s migration plans that would have remained invisible without data on migration intentions. Constraints on women’s mobility may have numerous implications. Indeed, although this issue remains largely under-explored, gender inequalities in access to migration, either internal of international, may affect development through different channels, including in particular access to labor market, health, education and fertility decisions, or political participation. Among these many channels, the following three can be singled out. First, women’s constraints on geographical mobility are coupled with a limited access to job opportunities (as shown for example by Assaad & Arntz (2005) in the context of Egypt). By improving the economic prospects of women, migration is likely to affect the balance of power within the household through an increase in the share of household resources controlled by women. Women’s migration may then

* I am grateful to Sylvie Lambert for numerous discussions on earlier versions of this paper, and to Francßois Bourguignon, Eve Caroli, Philippe De Vreyer, Christelle Dumas, Alice Mesnard, Elisabeth Sadoulet, Akiko Suwa-Eisenmann, as well as to two anonymous referees for their helpful comments and suggestions. I also wish to thank the participants at the 2012 ESPE Conference (Bern), the 2012 4th Development Conference of the GRETHA/GRES (Bordeaux), the 2011 AFSE Conference (Nanterre), the 2011 NORFACE and CReAM interdisciplinary conference on migration (UCL, London), the 2010 Third AFD-World Bank International Conference on “Migration and Development” (Paris School of Economics) and at seminars in Grenoble (GAEL), Le Mans (GAINS), Nanterre (Economix), Nice, Paris (Paris School of Economics) and Pau for their comments. Final revision accepted: January 19, 2014. 535

536

WORLD DEVELOPMENT

of shocks that may affect several dimensions of households’ capital. Some shock variables used in the paper, such as death or illness of a household member, are measured at the household level while others reflect changes in the environment such as the evolution of violence and crime, and climatic and natural shocks (rainfall, hurricanes, and earthquakes). The emphasis put on climatic shocks, and in particular rainfall, is justified by their expected impact on households’ economic outcomes, directly through agricultural yields in rural areas, and through prices in urban environments. The impact of rainfall on migration has been documented in the case of Mexico in particular by Munshi (2003), Pugatch and Yang (2011) or Feng, Krueger, and Oppenheimer (2010). Halliday (2006), on the other hand, studies the emigration of Salvadorans in response to earthquakes. Note in addition that catastrophic natural events such as hurricanes and earthquakes are not uncommon in Mexico, as recently illustrated by the massive flood caused by Hurricane Ingrid in September 2013. More generally a vast literature has developed on the issue of climate-driven migrations, in particular in anticipation of climate change (Barrios, Bertinelli, & Strobl, 2006; McLeman & Smit, 2006; Tacoli, 2009). Second, some characteristics of individuals and their household may act as constraints on their mobility or induce higher migration costs. Although most of these characteristics are known when intentions are stated (for example gender or age), they may prevent, foster or delay the realization of migration plans. The rationale behind the distinction between those two types of factors is to separately examine the impact of exogenous unexpected factors (shocks), and (mostly) endogenous characteristics, known at the intention stage, on the realization of migration plans. A parallel can be made with the livelihoods literature (Chambers & Conway, 1992; Rakodi, 2002) that makes a difference between household capabilities and assets (captured here by individual and household characteristics), and stresses and shocks that households have to cope with. I focus in this article on unrealized intentions to migrate of Mexican adults. I exploit the panel structure of the Mexican Family Life Survey (MxFLS) to compare intentions to migrate to subsequent moves. I use data on migration intentions collected in the first wave of the survey (2002) and information on migration behaviors between the two survey waves, available in the second wave (2005–06). I first provide empirical evidence supporting the hypothesis that migration intentions contain behavioral information. I then investigate the role of exogenous shocks and individual and household characteristics in explaining observed discrepancies between stated intentions and actual migration. I estimate bivariate probit models for migration intentions and mobility behaviors. I construct different sets of shock variables, using in particular long time series of rainfall data. Climatic shocks are found to affect migration but they appear to be only part of the story: the largest shift in selection between intentions and migration is related to gender. Conditional on initial intentions, women are found to be much less likely to migrate than men, in particular when they have children or no family network abroad. This finding is consistent with the existence of female-specific constraints to migration, but alternative interpretations are plausible. I discuss and test the two main competing hypotheses of different time preferences of men and women, and joint migration decisions at the household level. Empirical findings do not support these last two interpretations and rather suggest that social and family constraints explain why women migrate less than men conditional on their intention.

This paper contributes to the literature in several ways. First, thanks to the use of intention data, I am able to highlight an undocumented aspect of gender inequalities in access to international migration. I add to the empirical literature on migration and gender in the context of Mexico (Donato, 1993; Kanaiaupuni, 2000; Curran & Rivero-Fuentes, 2003) by studying the role of gender in explaining the non-realization of migration plans. Among the different interpretations of the gender effect discussed in this paper, the most plausible points to the existence of specific constraints to women’s international migration. This paper brings empirical evidence that women’s migration, and in particular international migration, is more constrained than men’s, and emphasizes the role of family constraints and networks in explaining such a gender selectivity, consistent with earlier findings in the empirical literature (and in particular in Donato (1993), Kanaiaupuni (2000) and Curran & Rivero-Fuentes (2003), while an extensive review is provided by Chant & Craske (2003)). While this feature has implications on the skill composition of the female emigrant population (Docquier, Marfouk, Salomone, & Sekkat, 2012), it also raises the crucial issue of gender inequality. In one of the first papers specifically addressing the issue of Mexican women’s migration, Donato (1993) tests the assumption that migration, as a family strategy, results in sending male members to work in the United States while women, assigned to their traditional role of spouses and mothers, would stay behind. She indeed finds evidence of family migration, women’s migration being conditioned by the prior successful migration of either their husband or father. In a more recent study, Kanaiaupuni (2000) finds that women’s lower probability to migrate may be due to the traditional roles devoted to women in the family organization. The predominant family reunification motive for female migration tends to make women “associational” migrants (Kanaiaupuni (2000), quoting a term coined by Balan (1981)). Building on this empirical literature, Curran and Rivero-Fuentes (2003) investigate the dynamics of the gender dimension of Mexican migration flows by exploring the potentially different impact of migrant networks depending on their gender composition on male and female migration. Their findings confirm the existence of higher barriers to female than to male international migration. While the gender selectivity of actual international migration flows is not necessarily problematic, as truly noted by Chant and Craske (2003), it may be more so once it is established that part of it is due to women being prevented from realizing their migration plans. In this regard, by focusing on the comparison between migration intentions and actual moves, this paper unveils a novel dimension of gender inequalities related to migration. Second, this paper is one of the first to exploit divergences between intentions and behaviors after they had been analyzed by Manski (1990). Manski (1990) rightly argues that even in the “best-case” hypothesis of rational expectations, discrepancies between intentions and observed actions must be interpreted cautiously. Indeed, intentions are stated by individuals depending on their known characteristics but also on unknown realizations of future events. Using panel data containing retrospective information merged to additional data on climatic events and crime statistics, I am able to account for a large range of shocks that are likely to affect the realization of migration intentions. This set includes health shocks, death of a household member, economic shocks, rainfall, natural disasters (earthquakes, hurricanes), and crime. This paper explores the determinants of unrealized migration intentions, once various sets of potential shocks are accounted for. More specifically, this paper is one of the first economic

MEXICAN MIGRANTS TO THE US

contributions to the small literature on discrepancies between migration intentions and mobility behaviors. 2 Incidentally, this paper proves that using intentions as proxies for actual migration (Burda, Hardle, Muller, & Werwatz, 1998; Liebig & Sousa-Poza, 2004) may lead to biased conclusions since, in the context under study, it would lead to overestimate female migration. The following section presents the empirical strategy. The data are described in Section 3. In Section 4, I present the results from bivariate probit estimations of the impact of shocks and other individual and household characteristics on the realization of migration plans, with a particular focus on gender issue. I present robustness checks in Section 5. The main conclusions and implications are summarized in Section 6. 2. EMPIRICAL STRATEGY The first part of the empirical analysis aims at providing elements supporting a rational interpretation of intentions to migrate by exploring the correlation between intentions and migration and analyzing the determinants of migration intention. Indeed, under the “best case” hypothesis that intentions are rational, they should be correlated with individuals characteristics which are found to drive migration, and in particular education. I thus estimate the following model by a standard probit: Intentioni ¼ a2 þ X i b2 þ Z i;h;l c2 þ u1;i

ð2:1Þ

Intentioni

where denotes the latent migration intention of individual i and is only observed as: Intentioni ¼ 1fIntentioni >0g

ð2:2Þ

X i represents the usual set of individual characteristics that are likely to explain migration. It includes age and age squared, gender, and education dummies. Z i;h;l refers to individual, household, and community variables that are likely to affect migration costs. Z i;h;l in particular contains gender, marital status, household size, a dummy for the presence of children or elderly persons in the household, household per capita expenditures as a proxy for household income, household and locality network variables, and regional dummies. 3 All the variables used in the empirical section are listed and described in the Appendix A. Second, I intend to explain unrealized migration intentions, accounting for shocks and potential constraints. In order to identify the determinants of migration conditional on intentions, I simultaneously estimate the following two equations by a bivariate probit:  Intentioni ¼ a2 þ X i b2 þ Z i;h;l c2 þ m1;i ð2:3Þ Migrationi ¼ a3 þ X i b3 þ Z i;h;l c3 þ S h;l;s d þ m2;i with Intentioni defined as above, and where Migrationi denotes the latent migration status of individual i and is only observed as: Migrationi ¼ 1fMigrationi >0g

ð2:4Þ

Since individual unobserved characteristics are likely to affect both intentions and actual moves, the bivariate probit specification is the most appropriate because it allows the errors terms m1;i and m2;i to be correlated. The above described X i and Z i;h;l sets of variables enter both equations, while the S h;l;s vector is added to the migration equation only. It contains variables representing potential shocks that may have affected individuals between the two waves of the survey. These

537

shocks include rainfall shocks and hurricanes measured at the state level, natural disasters recorded at the community level, as well as shocks directly reported by the household such as natural disasters, and death or illness of a household member. The h; l; s, subscripts refer to the household, locality, and state levels. In addition, I take potential non-linearities into account in some specifications by allowing shocks to non-additively enter the migration equation (in particular, by interacting shock variables with gender and education dummies). The marginal effects presented in the following tables are the marginal effects of the independent variables entering the model on the probability to migrate conditional on migration intention: ProbðMigrationi ¼ 1jIntentioni ¼ 1Þ. In all regressions observations are weighted using survey weights and standard errors are clustered by household, except in Table 6 where standards errors are clustered by locality. 3. DATA The data used in this article come from different sources. Intention and migration data, as well as all other individual, household and locality level data, come from the MxFLS panel. 4 In order to control for exogenous unexpected shocks, I use three additional data sources for rainfall, hurricanes, and crime. All data sources are briefly described below. Summary statistics are presented in the Appendix A (Tables 10 and 11). The Mexican Family Life Survey (MxFLS) is a nationally representative household survey, with a longitudinal structure. 5 Two waves of data collection have been conducted up to now, in 2002 and 2005. 6 During the first wave, 8,440 households (19,177 individuals aged 18–64) were surveyed, in 150 communities distributed across 16 states, representing all regions of Mexico. Since this paper intends to compare intentions to subsequent decisions, the sample of interest is restricted to the 15,917 individuals (83% of the initial sample) who were present in the household at the time of the survey so that intention data could be collected. The tracking of households and individuals, in 2005, led to a high re-contact rate (92%), but the trail of 1,237 adults has been lost. Most of the results shown in the following section are thus based on a subsample of around 13,000 adults (aged 18–64), present in the household in 2002, with non-missing migration intention data, non-deceased and tracked in 2005. Attrition is a very important issue which is fully addressed in the Appendix A: indeed, attrition is most likely due to migration, either internal or international. (a) Intentions and migration Information on intentions to migrate is collected in both survey waves. Since I focus on unrealized intentions, I mainly use information on migration intentions collected in the first wave, but I also exploit migration intentions in 2005–06 to discriminate between alternative interpretations of my results (Section 4-c-iii). Individuals are asked a first general question about their migration intentions. 7 Individuals stating an intention to migrate are then questioned about their planned destination. I use information contained in both questions to define intentions to migrate abroad and intentions to migrate in Mexico. Note that the intention question does not specify any time-limit for the realization of migration plans. The absence of time-limit in the intention question does not challenge the relevance of intention data collected in MxFLS for the identification of mobility constraints. However, we need to

538

WORLD DEVELOPMENT

be cautious about the definition of unrealized intentions. In the following, unrealized intentions refer to migration intentions stated in 2002 that have not resulted (yet) in migration three years later, although, of course, individuals may have migrated after the second wave. For a better understanding of the structure of the intention and migration data used in the paper, a figure is provided in the Appendix A (Figure 1). I focus in this paper on international migration. The definition of migrants includes both individuals who either migrated abroad and returned between 2002 and 2005, and those who were currently abroad when their household was re-interviewed in 2005. I also exploit data on internal migration to test the robustness of my results in Section 5. Similarly, internal migrants are individuals who moved to another Mexican locality than the one in which they were first surveyed between 2002 and 2005, and either stayed in their new place of residence or returned. 8 (b) Shocks Rainfall data come from the global gridded datasets produced by the University of Delaware’s Center for Climatic Research. 9 Using monthly precipitations series available from 1949, I apply the same strategy as Pugatch and Yang (2011) and create state-level 10 yearly normalized rainfall variables (rainfall z-scores). 11 I also use in some specifications a dummy variable for high rainfall that equals one if at least one of the two years of interest (2003 and 2004) has been especially rainy (characterized by a z-score greater than two). The symmetrical dummy variable for low rainfall could not be constructed for lack of inter-state variability, 2003 and 2004 being especially rainy years all over Mexico (see also Section 4(b)). Using state-level data on hurricanes from the Coastal Services Center database of the National Oceanic and Atmospheric Administration, 12 I construct two state-level variables for natural disasters: first, a hurricane dummy that equals one if the state has been hit by a hurricane during 2002–04, and second, a variable for the number of storms (including hurricanes, but also tropical storms of lower intensity). Two additional climatic shock variables (hurricane and earthquake dummies) are constructed using information collected in the MxFLS community questionnaire. The rationale behind the use of different level data for the same phenomenon is that they are complementary. State level shocks may be interpreted as proxies for actual shocks (a hurricane recorded at the state level increases the probability for a given individual or household to have been directly hit), or proxies for changes in local labor market conditions: they would thus capture an indirect effect of adverse economic shocks. Note that, reassuringly enough, the correlation is positive and reasonably large between hurricane variables at the state and community levels (0.29). 13 Together with climatic data, in order to control for potential exogenous shocks that may have affected individuals’ plans between 2002 and 2005, I use data on the evolution of violence and crime. Violence is indeed a national issue in Mexico, with regional peculiarities and an increasing drug-related criminality. The crime data that I use are taken from the “Justice in Mexico” 14 and “Seguridad Pu´blica en Me´xico” 15 projects. The first variable included in the set of shock variables represents the evolution, in percentage, of the number of registered crimes (all included) at the state level. The variation between the 16 states represented in the survey is large (see summary statistics in Appendix A), whereas the country average (for all 32 entities) for the same period slightly decreased (5%). Note that this variable is a rather crude proxy for actual crime since crime statistics are known to reflect both police activity

and criminal activity. This is why this measure of crime is completed with two other variables measured at the community level, aimed at capturing the subjective perception of violence. These variables are based on the community informant’s subjective assessment of the evolution of violence in the community in the last 12 months (information contained in the MxFLS community questionnaire in 2005). Shocks directly affecting the household and recorded in the MxFLS household questionnaire are also controlled for in some specifications. They include a dummy for the death or illness of a member of the household, and a dummy that equals one for a range of economic shocks (unemployment) and natural disasters. 16 (c) Costs and constraints Besides unexpected shocks, another explanation of unrealized migration intentions relies on migration costs and constraints. Individuals incurring high migration costs or faced with different types of constraints may indeed not be able to migrate, especially over the three- year span of the survey. Based on the migration literature, proxies for different types of migration costs and constraints are thus included in the model. Migration costs and constraints are in particular captured by gender, marital status, household income proxied by expenditures, demographic characteristics of the household (size, presence of children, or elderly members), household and locality migrant networks, and regional dummies. The regional categorization of Mexican states is taken from Durand, Massey, and Zenteno (2001), and the label “historic region” refers to the Western states of Central Mexico which first experienced massive outmigration flows to the United States at the end of the 19th century and are today characterized by strong migration networks. All the variables used in the subsequent regressions are listed and described in the Appendix A (Table 9). Basic summary statistics are also provided in the Appendix A (Tables 10 and 11). I mainly focus in the remainder of my paper on individuals intending to move in 2002, and the possible inconsistencies between their intention and subsequent behavior. 4. RESULTS The section is organized as follows: I first investigate the behavioral content of intentions, since the information contained in intentions conditions the validity of the subsequent results. Second, I focus on individuals intending to migrate in 2002 and examine two sets of possible explanations to the fact that some of them had not (yet) materialized their plans in 2005: unexpected shocks, and costs or constraints. Both sets of reasons are found to matter, however, the effect of any shock on migration conditional on intention to migrate is found to be less important than the impact of gender. Third, I consider different interpretations of this gender effect, and present suggestive evidence of costs or constraints that seem to be specific to women, related in particular to their social role as mothers. (a) Intentions and selection at the intention stage The first concern that might be raised by the MxFLS data on intentions to migrate, is that given how general the intention question is, answers might be interpreted as reflecting interviewees’ dreams rather the result of a rational calculation. In the former case, intentions would certainly be overstated.

MEXICAN MIGRANTS TO THE US

However such an interpretation of intentions is not supported by the data: 14.7% of individuals state an intention to migrate in 2002, whatever the destination, and 2.9% only specifically intend to migrate abroad (to the US for the overwhelming majority). 17 More generally, as mentioned above, intentions tend to be mistrusted, in particular by economists, because they do not imply any commitment. Overstating the case, it could be claimed that intentions would be of no more use than if they were random. Again, the data do not support this interpretation: first, as shown in Table 1, the correlation between intentions to migrate and migration is positive and significant for both international and internal migration, which would not be the case if intentions were random. Second, intentions to migrate abroad are found to be correlated with human capital and network variables, consistent with self-selection models based on returns to skills differentials, as can be seen in Table 2. A probit regression for intentions to migrate abroad is run on the whole sample, and then separately for men and women (column 2 and 3). The fourth column provides the significance level of the difference between the coefficients for men and women. Men have a higher probability to intend to migrate abroad, which justifies the need for running separate regressions for men and women. Male and female intentions to migrate abroad mainly differ with regard to education. Whereas all three coefficients on education dummies are positive for women (though they are not significant on the sample restricted to non-attritors 18), they are negative for men. Men with tertiary or vocational education are found to be less likely to have the intention to migrate abroad. This result of opposite selection for men and women relative to education is consistent with the previous findings by Kanaiaupuni (2000) based on actual migration data. Moreover, as appears in Table 2, intentions are positively correlated with destination-specific network variables for both men and women, which supports the assumption that networks affect migration intentions by supplying individuals with information on the foreign labor market and helping them to form their anticipations. (b) Who migrates and who does not: are shocks the only reason? Since intentions are found to have behavioral content, the next step consists in understanding the heterogeneity in individual migration behaviors among individuals with the

539

intention to migrate in 2002. Two major sets of reasons have been mentioned above. First, unexpected and unpredictable shocks could have occurred before migration plans were carried out. They may have modified individual or household calculations, leading them to cancel or postpone their decision to migrate. Second, even in the absence of shocks, different kinds of constraints (including social, family, or liquidity constraints) or costs could have delayed or canceled the realization of migration intentions. In order to assess the respective role of shocks and other factors on the realization of migration intentions, I estimate a bivariate probit model for intentions to migrate abroad and international migration. Results shown in Table 3 are the marginal effects of the right-hand side variables on migration, conditional on the initial intention to migrate abroad. The set of regressors in both equations includes gender, human capital variables (measured in 2002), and geographical controls. Various sets of shock variables are added to the migration equation, in columns 3 to 12 of each table. Results in Tables 3–5 read in percentage points. In order to control for potential liquidity constraints, I include a variable for household per capita consumption, proxying for household wealth, geographic controls that proxy for the accessibility to credit facilities, and network variables, that could capture, among other aspects, the possibility for would-be emigrants to get informal loans to fund their travel, as assumed by Borger (2011). The models estimated in Table 3, columns 3 to 6, include the state-level shock variables described in the previous section. Regression models whose results are shown in columns 7 and 8 include climate shocks measured at the community level. Results for crime and violence at the state and community levels are presented in columns 9 to 11, while column 12 shows results on household shocks. First, as regards weather shocks measured at the state level, I find a negative correlation between normalized precipitations in 2003 and 2004 and international migration, conditional on intentions. These results are consistent with those of Pugatch and Yang (2011) who indeed find a negative correlation between rainfall in Mexico and Mexican immigration (measured from US data sources as the share of male Mexicans in the US labor force). The interpretation that they propose is twofold: this negative correlation could be explained by larger emigration flows from drought-afflicted Mexican states, or by lower return flows of Mexicans settled in the US to these areas. My results suggest a third interpretation. Measuring here migration at the source, using origin country data, I am not

Table 1. Probit regressions of migration during 2002–05 on intention to move in 2002

Intention to migrate abroad (2002) (d)

(1) Migration abroad (2005)

(2) Migration abroad (2005)

0.871*** (0.104)

0.708*** (0.107)

Observations Pseudo R2

(4) Migration within Mexico (2005)

0.519*** (0.070) 1.872*** (0.031)

Yes 0.809*** (0.298)

1.775*** (0.031)

0.453*** (0.073) Yes 0.836*** (0.251)

13038 0.026

13036 0.105

13038 0.024

13036 0.055

Intention to migrate within Mexico (2002) (d) Controls Constant

(3) Migration within Mexico (2005)

Robust standard errors in parentheses (clustered by household). Controls included are gender, age, education, regional and geographic dummies. (d) dummy variables. *** p < 0.01.

540

WORLD DEVELOPMENT Table 2. Probit regressions of intention to migrate abroad in 2002, comparison between men and women Dependent variable, column (1) to (3): Intention to migrate abroad

Male (d) Age Age squared Education: secondary (d) Education: preparatoria or tertiary (d) Education: professional (d) Married (d) Household size Children <15 (d) Elderlies in the household (d) Log per capita total expenditures Locality migration network abroad Relative (s) in the US (d) Shocks (1998–2001) Geographic controls Constant Observations Pseudo R2

(1) All

(2) Men

(3) Women

0.251*** (0.064) 0.027 (0.021) 0.000 (0.000) 0.016 (0.087) 0.050 (0.101) 0.075 (0.116) 0.257*** (0.088) 0.003 (0.016) 0.134 (0.103) 0.059 (0.096) 0.200*** (0.061) 0.037*** (0.013) 0.572*** (0.079) Yes Yes 3.345*** (0.642)

0.067** (0.029) 0.001 (0.000) 0.059 (0.130) 0.297* (0.154) 0.323* (0.190) 0.267 (0.165) 0.002 (0.022) 0.218 (0.171) 0.254* (0.142) 0.207*** (0.068) 0.040** (0.020) 0.504*** (0.104) Yes Yes 2.295*** (0.812)

0.023 (0.030) 0.001 (0.000) 0.081 (0.108) 0.191 (0.126) 0.179 (0.152) 0.264** (0.106) 0.005 (0.022) 0.077 (0.123) 0.076 (0.126) 0.198*** (0.070) 0.038** (0.015) 0.624*** (0.097) Yes Yes 4.314*** (0.737)

12,879 0.133

5261 0.136

7618 0.147

Diff (2) (3)

**

**

**

*

*

**

12,879

Robust standard errors in parentheses (clustered by household). (d) dummy variables. Geographic controls include regional dummies and dummies for different urban and rural strata Shocks include rainfall z-score in 2001, the number of storms (1998–2001), and crime evolution at the state level (1998–2001). The reference category for education variables is no education or primary education. Column (3) provides a test of equality of coefficients between Eqns. (1) and (2) using seemingly unrelated estimation tools. * p < 0.10. ** p < 0.05. *** p < 0.01.

concerned with return migration issues and directly observe a negative correlation between net migration outflows and rainfall. Nonetheless, the drought interpretation is challenged by the fact that I am presenting marginal effects, conditional on previous intentions to move. My results suggest that rainfall is correlated with migration for those who already had the intention to move. 19 Low rainfall, and thus drought, could indeed result in increased vulnerability and be an incentive to carry out migration plans earlier. But heavy rainfall, by improving local economic prospects, could make migration plans less or no more profitable by reversing the sign of the expected income differential between the US and Mexico. The fact that 2003 and 2004 happen to have been two especially rainy years supports this latter interpretation. Indeed, z-scores are negative in one state only in 2003 (Sonora) and two states in 2004 (Sinaloa and Morelos). For lack of inter-state variability, however, this interpretation cannot be convincingly tested by splitting the z-scores variables into negative and positive

ones. On the other hand, the fact that the number of storms at the state level has a positive and significant impact supports the interpretation in terms of increased vulnerability. At the community level results are different: adverse shocks (earthquake or hurricanes) are associated with a lower propensity to migrate. A possible explanation for observing opposite effects of similar shocks (storms and hurricanes 20) measured at different levels is as follows: at the state level, shock variables capture the indirect effect of weather shocks, and may thus proxy for unexpected adverse changes on state level labor markets or commodity prices. Thus, having less job opportunities locally, individuals (men especially) would be led to migrate earlier. At the community level, on the other hand, survey-based variables capture the direct effects of natural disasters: individuals living in communities that were actually affected by a hurricane most likely incurred financial losses. Such negative shock is unsurprisingly negatively correlated with their ability to migrate.

Table 3. Marginal effects on the probability to migrate abroad during 2002–2005 conditional on intention to migrate abroad in 2002 Marginal effects on Prob (migration abroad= 1j intention to migrate abroad= 1) after bivariate probit (1) Female (d)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

0.071*** 0.066*** 0.068*** 0.066*** 0. 066*** 0.067*** 0. 065*** 0.065*** 0. 066*** 0.066*** 0.066*** 0. 064*** (0.015) (0.013) (0.013) (0.013) (0.013) (0.014) (0.013) (0.013) (0. 014) (0.014) (0.013) (0.013)

State level weather shocks Rainfall z-score 2003

0.012** (0.006) 0.020*** (0.006)

Rainfall z-score 2004

0.004 (0.018)

High rainfall 2003–04 (d) Hurricane 2002–04 (d)

0.028 (0.021)

Community level shocks Earthquake (d)

MEXICAN MIGRANTS TO THE US

0.023*** (0.009)

Number of storms 2002–04

0.042* (0.023) 0.035* (0.019)

Hurricane (d) Crime and violence Crime evolution (state level) 2002–05

0.000*** (0.000) 0.003 (0.016)

Crime increased (community) 2004–05 (d)

0.019 (0.015)

Violence to women increased (community) 2004–05 (d) Household shocks Death/illness 2003–05 (d)

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

0.026** (0.013) 0.009 (0.018) Yes

Estimated probability Prob (migrant 2005= 1j intention 2005 =1)

0.148

0.107

0.106

0.107

0.108

0.108

0.108

0.106

0.107

0.107

0.106

0.079

Observations

13,038

13,036

13,036

13,036

13,036

13,036

13,036

13,036

13,036

13,036

13,036

12,842

Adverse economic shock 2003–05 (d) Controls

Marginal effects; Standard errors in parentheses. (d) for discrete change of dummy variable from 0 to 1. Controls included are age, education, regional and geographic dummies. * p < 0.10. ** p < 0.05 *** p < 0.01. 541

542

WORLD DEVELOPMENT Table 4. Marginal effects on the probability to migrate abroad during 2002–05 conditional on intention to migrate abroad in 2002 Marginal effects on Prob (migration abroad= 1j intention to migrate abroad= 1) after bivariate probit (1)

Female (d) State level weather shocks Rainfall z-score 2003 Female * rainfall z-score 2003 Rainfall z-score 2004 Female * rainfall z-score 2004

0.037 (0.023)

(2)

(3) ***

0.070 (0.018)

(4) ***

0.058 (0.013)

(5) ***

0. 068 (0.018)

(6) ***

0.072 (0.014)

(7) ***

0.067 (0.014)

0. 068*** (0.015)

0.013* (0.008) 0.000 (0.008) 0.013* (0.007) 0.018** (0.008) 0.006 (0.021) 0.004 (0.020)

High rainfall 2003–04 (d) Female * high rainfall 2003–04 (d)

0.061* (0.033) 0.054** (0.021)

Hurricane 2002–04 (d) Female * hurricane 2002–04 (d)

0.023* (0.012) 0.001 (0.014)

Number of storms 2002–04 Female * number of storms 2002–04 Community level shocks Earthquake (d)

0.066** (0.023) 0.075 (0.081)

Female * earthquake (d)

0.043** (0.022) 0.021 (0.036)

Hurricane (d) Female * hurricane (d) Household shocks Death/illness 2003–2005 (d)

Yes

Yes

Yes

Yes

Yes

Yes

0.041*** (0.015) 0.049 (0.034) 0.014 (0.024) 0.012 (0.024) Yes

Estimated probability Prob (migrant 2005= 1j intention 2005= 1)

0.106

0.107

0.108

0.108

0.108

0.106

0.079

Observations

13,036

13,036

13,036

13,036

13,036

13,036

12,842

Female * death/illness 2003–2005 (d) Adverse economic shock 2003–2005 (d) Female * economic shock 2003–2005 (d) Controls

Marginal effects; Standard errors in parentheses; (d) for discrete change of dummy variable from 0 to 1. Controls included are age, education, regional and geographic dummies. * p < 0.10. ** p < 0.05. *** p < 0.01.

As for crime variables, they are not found to affect migrations plans, either at the state-level (objective figures), or at the community level (subjective assessments). The significance of the coefficient in column 9 is exclusively driven by the outburst of crime in the Sonora state, 21 with a 146% increase during 2002–05. 22 Then, among household shocks, the death or illness of a household member is found to negatively affect the probability to migrate. 23

In conclusion, shocks are unsurprisingly found to affect the realization of migration plans abroad. However, the most striking result in Table 3 is the negative and significant coefficient on the female dummy: being a woman reduces by almost one half (around 7 percentage points) the probability to migrate abroad during 2002–05, conditional on intending to migrate in 2002. The sign, size, and significance of the coefficient on the female dummy are robust to the inclusion of any set of shocks in the model. Moreover, being a woman cuts by two

MEXICAN MIGRANTS TO THE US

543

Table 5. Marginal effects on the probability to migrate abroad during 2002–05 conditional on intention to migrate abroad in 2002, for men and women. Marginal effects on Prob (migration abroad= 1j intention to migrate abroad= 1) after bivariate probit (1) All mfx Female (d) Age Age squared Education: secondary (d) Education: preparatoria or tertiary (d) Education: professional (d) Married (d) Household size Children <15 (d) Elderlies in the household (d) Locality migration network abroad Relative (s) in the US (d) Log per capita total expenditures Geographic controls Shocks during 2002–05

***

0.060 0.002 0.000 0.005 0.027 0.017 0.009 0.009*** 0.042*** 0.020** 0.005*** 0.034*** 0.005

(2) Men

(3) Women

se

mfx

se

mfx

se

(0.013) (0.002) (0.000) (0.010) (0.017) (0.015) (0.013) (.003) (0.016) (0.010) (0.002) (0.010) (0.005)

0.000 0.000 0.011 0.049 0.033 0.001 0.013*** 0.028 0.063*** 0.007** 0.018 0.005

(0.004) (0.000) (0.021) (0.035) (0.031) (0.029) (0.005) (0.028) (0.021) (0.003) (0.019) (0.011)

0.002 0.000 0.002 0.011 0.005 0.004 0.005** 0.029** 0.001 0.002* 0.034*** 0.003

(0.002) 0.000 (0.007) (0.011) (0.009) (0.007) (0.002) (0.014) (0.006) (0.001) (0.012) (0.003)

Yes Yes

Yes Yes

Yes Yes

Estimated probability Pr (intention 2002= 1j non-migrant 2005= 1)

0.078

0.124

0.037

Observations

12,689

5185

7504

Robust standard errors in parentheses (clustered by household). (d) dummy variables. The reference category is no education or primary education for education variables. Shocks: Rainfall z-scores 2003 and 2004, number of storms, earthquake (community), death or illness (household). Geographic controls: historic region dummy and rural/urban strata dummies. * p < 0.10 ** p < 0.05. *** p < 0.01.

thirds the probability to move abroad, which is more than the largest marginal effect of any shock variable included in the model (40% for the earthquake dummy at the community level). (c) Explaining women’s unrealized intentions (i) Shocks and gender Based on the previous results, a first extension to the initial model consists in adding interactions between shock variables and the female dummy variable: results are presented in Table 4. The models including different sets of crime shocks are not shown here since none of these variables is found to significantly affect migration. With the exception of the first model, all results shown in Table 4 suggest that shocks do not affect women more than men. In models 2 and 4, the interaction terms are not significant, shocks are thus found to have similar effects on men and women. In models 3 and 5 to 7, the interaction terms and main effects have opposite signs and comparable magnitude, suggesting that shocks exclusively affect male migration (conditional on intention), although the interaction term is not always significant, maybe because of the relatively small number of observations. In all specifications, the magnitude and significance of the marginal effect of the female dummy remain constant. The results of the first specification requires further comments: once an interaction term between the female dummy and the rainfall z-score in 2004 has been added to the model, the female main coefficient is not significant any more, whereas the interaction term suggests that rainfall affects women twice more than men. Nonetheless, this specification is problematic because of the high and positive correlation (0.53) between

precipitations in 2004 and the center region dummy. 24 Actually, the cumulative negative effect of rainfall on female migration is likely to be driven by the exceptionally heavy rainfall that affected the two states of Mexico and Morelos, as well as the Federal District. 25 A plausible interpretation is thus that women living in the center region (and in particular in Mexico City) who plan to migrate may be less likely to do so because of a larger set of local job opportunities than in other regions. Indeed, the five federal states part of the center region (Mexico, Morelos, Oaxaca, Puebla, and the Federal District) form the largest employment area in Mexico: in 2003, they host one third of the national economically active population (INEGI 26). Note that the inclusion of interaction terms reveals the positive marginal effect of the hurricane variable at the state level on the conditional probability to migrate for men, whereas it does not affect women’s migration. (ii) Female-specific constraints I then test the hypothesis of different types of costs and constraints, and in particular social and family constraints, that would be specific to women and affect the realization of their migration plans. In Table 5, I present results from the same model estimated by a bivariate probit as in Table 3, with additional controls for demographic characteristics of the household in 2002. Marginal effects read in percentage points as in the previous tables. A set of shock variables is still controlled for but marginal effects of these additional controls are not shown, and separate regressions are run for men and women. 27 First, the variable proxying for household wealth has no effect on the probability to migrate conditional on intentions either for men or for women which suggests that the financial

544

WORLD DEVELOPMENT

cost of migration or potential liquidity constraints are accounted for at the intention stage, consistent with a rational interpretation of intentions. Second, one of the main findings is that having children under 15 is negatively associated with women’s migration. The marginal effect of the children dummy is negative and of similar magnitude for men, but it is not significant. This result supports the hypotheses made by Kanaiaupuni (2000) relative to the social and financial constraints to the emigration of mothers. It is also consistent with the earlier results of Donato (1993) who finds an inverse U-shape relationship between age and the probability for women to migrate. A similar pattern for age is also found here (not shown), but the age effect is entirely absorbed by the children dummy. 28 On the other hand, having relatives in the US is positively and significantly correlated with women’s migration only. These results are consistent with the interpretation that women’s ability to realize their plans and migrate abroad is limited by specific constraints, in particular caused by their having children. Moreover, the fact that women seem to rely more than men on family networks to migrate abroad is consistent with a family reunification motive, in which women would be “associational” migrants (Kanaiaupuni (2000), quoting a term coined by Balan (1981)). Note that because of the relatively small size of the sample I cannot test the effect of networks’ gender composition and compare my results to those of Curran and Rivero-Fuentes (2003). Alternative interpretations may however account for these findings: they are tested and discussed in the rest of this section. (iii) Female-specific time preferences First, women may have different preferences regarding the time of their migration. Since no time limit is explicitly mentioned in the intention question, it could be assumed that women systematically intend to migrate in a further future than men (for example, as regards mothers, when their children are grown up). Nonetheless, if different preferences were the whole story, since we are observing migration flows, and under the assumption that preferences are relatively stable over time, we should not observe any difference in the size of male and female migrant cohorts during 2002–05 (conditional on intention). The only difference would be that female migrants would have waited longer before they could migrate. However, this interpretation cannot definitely be rejected, if women’s migration preferences have changed over time and if for example younger women have a higher propensity to intend to migrate than their elders. Another element is provided by the comparison of the age distribution of men and women stating an intention to migrate, compared to the one of male and female migrants (see Figure 2 in Appendix A). Indeed, the share of individuals intending to migrate abroad aged between 25 and 35 years is larger for women than for men. The highest density is around 20 years old for men, and it then steeply declines. This finding is consistent with either an interpretation based on constraints delaying or preventing women’s migration, or different time preferences: women appear to intend to migrate as young as men, but they migrate later, thus swelling the density for ages up to 35 years. However, the same density plotted for migration abroad seems to rather support the first interpretation: if preferences were to explain the difference between male and female age distributions at the intention stage, we would expect the mode of the age distribution for women at the migration stage to be shifted to the right, as compared to men’s. And yet, the modes for male and female age distribu-

tions coincide at the migration stage, at around 23 years old. In addition, I find that age distributions for intentions are not much affected when including attritors in the sample (Figure not shown, available upon request). Another element supports the interpretation in terms of miscalculated costs and constraints: if women had longer term prospects and stated in 2002 their intention to migrate in a more remote future than men, they would then be expected to keep on intending to move longer than men, and in particular three years later, when they were surveyed again. Results shown in Table 6 are marginal effects on the probability to have the intention to migrate in 2005 conditional on not having migrated between 2002 and 2005, obtained after a bivariate probit for migration between 2002 and 2005 and intentions to migrate in 2005 on the subsample of those who had the intention to migrate in 2002. Because of the small number of observations, the destination (internal vs international migration) is not taken into account (as in Tables 7 and 8). The female dummy is not significant: women do not significantly persist longer than men in their migration plans, which does not support the assumption that preferences regarding the time pattern of migration is different for men and women. (iv) Household migration model It could also be argued that a household model would be better suited to the analysis of Mexican migration and could explain the lower realization of women’s migration plans. Women’s intentions to migrate could then be viewed as depending on a decision taken at the household level. Note that a household-level decision to participate in migration does not necessarily imply that all household members migrate: the household could move as a whole unit, or split, with one or more members migrating while the others stay. The former case is first investigated by exploring the correlation between migration intentions of male and female members from the same household. As appears in Table 7, intentions to migrate are correlated within the household. 29 Note however that this correlation exists for both men and women: the interaction term is not significant, which means Table 6. Marginal effects on the probability to have the intention to migrate in 2005 conditional on not having migrated during 2002–2005: subsample of individuals with the intention to migrate in 2002 Marginal effects on Prob (intention 2005= 1j non-migrant 2005= 1) obtained from a bivariate probit

Female (d) Age Age squared Education 2ry and higher (d) Married (d) Household size Children <15 (d) High rainfall 2003–2004 (d) Human shock 2003–2005 (d) Economic shock 2003–2005 (d) Geographic controls Observations

Marginal effects

Standard errors

0.026 0.000 0.000 0.075** 0.032 0.011 0.048 0.077** 0.017 0.145** Yes

(0.031) (0.010) (0.000) (0.034) (0.047) (0.009) (0.043) (0.033) (0.048) (0.063)

1670

Marginal effects; Standard errors in parentheses. (d) for discrete change of dummy variable from 0 to 1. Geographic controls: historic region dummy and rural/urban strata dummies. ** p < 0.05.

MEXICAN MIGRANTS TO THE US Table 7. Correlation between intention to move of individuals of different sex within the same household Probit regression, individual level Dependent variable

Intention to migrate, 2002 0.681***

Intention to move of another household member, different sex

(0.073) 0.044 (0.050) 0.002

Female (d) Intention to move of another household member, different sex* Female Controls Constant

(0.047) Yes 0.916*** (0.172)

Observations

13185

Robust standard errors in parentheses (clustered by household). (d) dummy variables. Controls included are age, education, regional and geographic dummies. * p < 0.10. *** p < 0.01.

545

only reflect their spouse’s migration plans. In order to test this interpretation, I investigate whether women’s intentions to migrate in 2002 are associated with a higher probability to observe male migration from the same household between the two survey dates. As presented in Table 8, this seems to be the case. However, the lower part of the table presents similar results for the reverse case: men’s intention to move in 2002 is positively correlated with observed female migration within the same household between the two survey waves. Note that here again, the sample is restricted to individuals living with at least one person of opposite sex whose intentions as regards migration are also recorded in 2002. 30 The latter two tables unsurprisingly emphasize the fact that members of the same household coordinate their migration decision but do not provide evidence of different interpretations of the intention question by men and women. The gender effect on the realization of migration plans emphasized above thus seems to be best explained by specific costs and constraints incurred by women. The following section provides several robustness checks.

5. ROBUSTNESS CHECKS Table 8. Correlation between male migration during 2002–05 and female intentions to migrate in 2002 at the household level

(a) Impact of shocks

Probit regression, household level

(i) Shocks and migration of individuals with no intention to migrate In order to check the consistency of my findings for shocks, I computed the marginal effects of all shocks included in the different models estimated in the previous section, on the probability to migrate conditional on not intending to migrate in 2002 (either within Mexico or abroad). Results for international migration (not shown, available upon request) prove consistent with the above findings: shocks are found to affect those who had no intention to move in a similar way (same signs on all coefficients) as those who had the intention to migrate, but the magnitude of the coefficient is unsurprisingly much lower.

Dependent variable At least one woman with intention to move in 2002 Controls Constant

At least one male migrant member during 2002–05 0.215**

0.167*

(0.093) No 1.241*** (0.049)

(0.090) Yes 1.864*** (0.275)

Dependent variable

At least one female migrant member during 2002–05

At least one man with intention to move in 2002

0.256***

0.198**

(0.080) No 0.979*** (0.045)

(0.081) Yes 2.757*** (0.289)

5081

5021

Controls Constant Observations

(b) Shocks and gender in the case of internal migration

Robust standard errors in parentheses (clustered by locality). (d) dummy variables. Controls included are the age of the head, highest level of education of any adult household member, weather shocks at the state level, household shocks, regional and geographic dummies. * p < 0.10 ** p < 0.05. *** p < 0.01.

that women’s intentions to migrate are not found to depend more than men’s on a spouse or parent’s own migration plans. Even in the case of migration as a household strategy consisting in selecting only one migrant member, it could be argued that the lower propensity for women to migrate, conditional on their intention, may reflect a different interpretation of the intention question by men and women. Women could voice more often than men a household decision to participate in migration, even though they may not migrate themselves: in that case, their stated intention to migrate would

Since internal migration is much less costly than international migration, the impact of variables proxying for costs and constraints on the realization of migration plans is expected to be lower in the case of internal migration. First, no gender-specific patterns appear as regards intentions to move within Mexico, as opposed to intentions to migrate abroad. 31 Intentions are indeed positively and very significantly correlated with education for both men and women. 32 Second, Table 12 in Appendix A, shows that shocks, in general, do not alter internal migration plans. This result is not surprising since internal migration is characterized by lower costs and benefits and may be driven by non economic motives (for example marriage or separation, 33 family, or health care). Note that as regards internal migration, unlike international migration, the consequence of being directly hit by a hurricane is to increase the probability to move conditional on intention. This finding is consistent with an interpretation in terms of costs and liquidity constraints, since migration costs are undoubtedly much higher in the international migration case. The negative impact of gender on the realization of intentions to migrate within Mexico is smaller than in the international migration case explored in the previous section. Being a woman cuts by one fifth (2 percentage points) the probability

546

WORLD DEVELOPMENT

to migrate in Mexico conditional on intention to migrate within Mexico. Women thus seem to be less constrained on the internal migration market, consistent with previous findings (Chant & Craske, 2003; Curran & Rivero-Fuentes, 2003). Note however that consistent with an interpretation in terms of costs the marginal effect of the female dummy is smaller in the internal migration case, but it is still negative and robust to the introduction of shocks. (c) Shocks in the US Shocks in the US may also affect the realization of migration intentions. For example, an adverse economic shock in the US between the two survey waves may have caused individuals who stated an intention to migrate in 2002 to change their plans and stay in Mexico. This hypothesis cannot be directly tested, since it would require attributing a specific US destination to all Mexicans who did not migrate, in order to create individual variation in the US shock variable. However aggregate figures at the country level do not provide any evidence of an economic shock in the US in the period of interest. Figure 3 plots the unemployment rate of Mexicans in the US during 1995–10. The 2002–06 period is characterized by the stability of the unemployment rate of Mexicans in the US. Although the hypothesis of local adverse economic shocks in the US cannot be ruled out, this figure does not support the hypothesis of a macroeconomic shock in the US, either positive or negative, which could explain the observed discrepancies between intentions and migration. The robustness checks conducted in this section do not challenge the interpretation supported by the empirical evidence presented above: the observed lower propensity for women to realize their intention to migrate is best explained by female-specific costs and in particular family constraints. 6. CONCLUSION This paper uncovers female-specific mobility constraints by exploring the determinants of unrealized migration intentions.

The analysis is based on the Mexican Family Life Survey panel collected in 2002 and 2005–06. I first show that the correlation between intentions to migrate and actual migration is strong and robust, which supports a rational interpretation of migration intentions. Second, I find that the realization of migration plans of Mexicans is affected by different shocks that occurred between the two survey waves, and in particular weather shocks at either state or community level. I find in particular that above-average rainfall lowers the probability that individuals with the intention to migrate abroad in 2002 actually migrated during 2002–06. However, shocks are only part of the explanation of unrealized migration intentions. Indeed, gender is the most important determinant of migration conditional on intentions. Women are found to have a much lower propensity to migrate abroad within the three-year span of the survey conditional on migration intentions stated in 2002. The analysis of unrealized migration intentions of Mexican adults thus reveals gender inequalities which affect women’s direct access to the benefits from migration: most women who want to move abroad are either constrained to stay or to delay the realization of their migration plans. Alternative interpretations, such as different time preferences of men and women or the voicing by women of a household migration plan are not supported by the data. Incidentally, this article proves that even though migration intentions contain valuable behavioral information, they may not be relevant proxies for actual migration. In the context under study, using intention data as proxies for actual migration would lead to a large overestimation of female migration. The above findings suggest new avenues of research on migration based on the unique information contained in intention data. Further research could in particular address the issue of the welfare impact of women’s unfulfilled intentions to migrate. Indeed while the gender selectivity of observed international migration flows may not necessarily be problematic if they result from different preferences, it may be more so once it is established that such gender selectivity is mostly due to constraints preventing women with the intention to migrate to realize their plans.

NOTES 1. Moreover, a large literature has documented the importance of migration as a strategy available to households to cope with shocks. 2. Among the very few papers that empirically analyze the link between intentions and mobility, at the crossroad of psychology and geography, Lu (2005) investigates the explanatory power of mobility intentions on subsequent actions and interprets the observed discrepancies between intentions and actions as behavioral inconsistencies, ignoring the criticism developed by Manski (1990). In the same line, Van Dalen and Henkens (2012) explain discrepancies between intentions to migrate and international migration in the case of the Netherlands by individuals’ personality traits and dissatisfaction with their home country, but do not account for shocks. 3. Other tested specifications (not shown) included a dummy for a having a permanent job in the formal sector, as a proxy for opportunity costs.

6. The second wave was actually completed in 2006. 7. The word “intentions” is used throughout the paper for the sake of simplicity. Note however that the exact question asked to interviewees is somewhat less precise: “Ha pensado usted en irse a vivir en un futuro, fuera de la localidad/colonia en la que vive actualmente?” which could be translated as “Have you ever thought of moving one day out of the locality which you are now living in?”. Wording is known to be very important for subjective data (Bertrand & Mullainathan, 2001). For that reason, the very general way the question is formulated may be an advantage since it is likely to reduce the propensity for interviewees to censor themselves in front of surveyors or other family members attending the interview. On the other hand, it may lead us to over-interpret mere dreams as actual migration plans. However, as shown in the following section the generality of the intention question does not deprive intentions from informational content. More importantly, nothing suggests that men and women would have different interpretations of the question. This issue is however further investigated in Section 4-c.

4. http://www.ennvih-mxfls.org/. 5. The survey design is documented in Rubalcava and Teruel (2006) and Rubalcava and Teruel (2008).

8. This definition of internal migrants excludes those individuals who moved within the same locality or within the most urbanized strata of the capital city of Mexico.

MEXICAN MIGRANTS TO THE US

547

9. Full documentation is available at: http://climate.geog.udel.edu/ c~limate/html_pages/Global2_Ts_2009/README.global_p_ts_ 2009.html.

at the 5% level for women, whereas only the coefficient on the vocational education dummy is significant at the 10% level for men, refer to Appendix A for a complete analysis of attrition.

10. For reasons of confidentiality, data permitting the geographical identification of localities in MxFLS (latitude, longitude, codes that could be compared to INEGI codes) are not made public. This is why I use statelevel rainfall data, even though global gridded datasets offer much more precision. State-level rainfall data are thus imperfect proxies for actual precipitations shocks: as such, they are likely to decrease the precision of estimated effects, but unlikely to bias the results.

19. The impact of shocks on migration for those who had no intention to migrate in 2002 is explored as a robustness check in the following section.

11. To construct these rainfall z-scores, I first assign grid points to states based on latitude and longitude coordinates, then sum up monthly data to obtain yearly rainfall variables and compute state-level averages for each year, state-level long term averages (1949–2005), and state-level standard deviations. The normalized variable is the state-level rainfall value minus the state-level long-run mean, divided by the state-level standard deviation over 1949–2005. For example, a positive value for year 2003 in state S means that 2003 has been an especially rainy year in state S. Conversely, a negative value means that precipitations have been lower than (long-term) average in that state in 2003. 12. http://csc.noaa.gov/. 13. In 2005, 17% of MxFLS community respondents declared that their community had been affected by a hurricane in the last three years. In 62% of cases, no hurricane was recorded at the state level (based on the exhaustive list provided by the National Oceanic and Atmospheric Administration): either hurricanes declared at the community level were only storms (of lower intensity), or recall errors caused hurricanes prior to 2002 to be declared. The state-level hurricane variable excludes hurricanes that occurred in 2005 since, for lack of the exact survey date, I cannot verify that they are prior to the second wave of the survey. Even without errors in hurricane statements at the community level, this would explain the unperfect correlation between state-level and community-level hurricane dummies. 14. http://justiceinmexico.org/. 15. www.seguridadpublicaenmexico.org.mx. 16. A larger set of household and individual shocks would ideally include positive income shocks (lottery gains, inheritance), or demographic shocks (marriage, births). Alternative specifications were tested, but their results (not shown) could not be exploited because of the very low occurrence of such events in the sample, in particular as regards international migration. For example, even though the percentage of individuals getting married between the two survey waves is not negligible (8.2%), cell sizes become too small for results to be robust when this variable is interacted with intentions to migrate abroad and gender. Specifications including demographic changes, and in particular marriage, between the two survey waves were nonetheless tested but the coefficients on these variables were not found significant in the case of international migration. The impact of demographic changes on the realization of internal migration plans is discussed in Section 5(b). 17. Pooling internal and international destinations, in 2002, 1,616 individuals aged over 18 state their intention to migrate. Among them, 294 actually migrated between the two survey waves. The migration rate is twice higher among those who had the intention to migrate (15.4%) than those who had no intention to migrate in 2002 (6.6%). 732 individuals fall into this category. 18. When attritors whose intentions are non-missing are included in the sample, signs of the coefficients on all education dummies are not modified, but the coefficient on the tertiary education dummy is significant

20. The coefficient on the hurricane dummy is indeed positive and significant for men, whereas hurricanes at the state level do not affect the realization of women’s migration plans, as is shown below, in Table 4. 21. Indeed, when the same regression is run on a subsample of observations excluding individuals surveyed in the Sonora region, the coefficient on the state-level crime variable is no longer significant. 22. According to the reported crime statistics used to construct the statelevel crime variable, see Section 3(b). Note that this evolution is related to illegal activities linked to the border with the US (including drug trafficking and illegal migration networks). 23. Unfortunately, the data do not allow us to know whether this correlation could be explained by the fact that illness of the individual who initially intended to migrate had forced him to change his plans since the affected household member cannot be precisely identified. A solution would consist in considering only death shocks, which could only affect another household member, but they concern such a low percentage of individuals (5.3%) that they cannot be used separately. 24. For that reason, the set of controls in this specification (in all tables) does not include the center region dummy. 25. Results are indeed similar, except that the marginal effect of the female dummy is significant, when the center region dummy instead of the 2004 rainfall variable is interacted with the female dummy. 26. http://www.inegi.org.mx/prod_serv/contenidos/espanol/bvinegi/productos/integracion/pais/historicas10/Tema5_Empleo.pdf. 27. Alternative models were tested: in particular, simple probit models for migration estimated on the subsample of individuals intending to migrate in 2002 provided similar results. 28. Donato (1993) controls for the presence of children in the household but not for own children. 29. Note that destination could not be taken into account either in this regression, because of the reduced sample size. The regression sample indeed includes only households of size at least equal to 2, including two persons of opposite sex. 30. 1,662 individuals being the only adult of their household, 5,458 women living in households with at least one other adult but no adult male member and 178 men living in households with two or more adults but no adult women are excluded from the regression sample. 31. Results are not shown but are available upon request. 32. The same result is obtained when attritors are included in the regression sample. 33. Indeed, in an alternative specification (not shown, available upon request) including demographic changes in the household between the two survey waves, both marriages and separations are associated with a higher probability to move within Mexico conditional on intentions.

548

WORLD DEVELOPMENT

REFERENCES Abu-Ghaida, D., & Klasen, S. (2004). The costs of missing the millennium development goal on gender equity. World Development, 32(7), 1075–1107. Assaad, R., & Arntz, M. (2005). Constrained geographical mobility and gendered labor market outcomes under structural adjustment: Evidence from egypt. World Development, 33(3), 431–454. Balan, J. (Ed.) (1981). Why people move: Comparative perspectives on the dynamics of internal migration. Paris: UNESCO Press. Barrios, S., Bertinelli, L., & Strobl, E. (2006). Climatic change and rural– urban migration: The case of sub-saharan africa. Journal of Urban Economics, 60(3), 357–371. Beine, M., Docquier, F., & Schiff, M. (2009). International migration, transfers of norms and home country fertility. Policy Research Working Paper Series 4925, The World Bank. Bertrand, M., & Mullainathan, S. (2001). Do people mean what they say? Implications for subjective survey data. The American Economic Review, 91(2), 67–72. Borger, S. C. (2011). Self-selection and liquidity constraints in different migration cost regimes, Unpublished manuscript. Burda, M. C., Hardle, W., Muller, M., & Werwatz, A. (1998). Semiparametric analysis of german east–west migration intentions: Facts and theory. Journal of Applied Econometrics, 13(5), 525–541. Chambers, R., & Conway, G. (1992). Sustainable rural livelihoods: Practical concepts for the 21st century. UK: Institute of Development Studies. Chant, S., & Craske, N. (2003). Gender in Latin America. Latin American Bureau. Curran, S. R., & Rivero-Fuentes, E. (2003). Engendering migrant networks: The case of mexican migration. Demography, 40(2), 289–307. Docquier, F., Marfouk, A., Salomone, S., & Sekkat, K. (2012). Are skilled women more migratory than skilled men?. World Development, 40(2), 251–265. Donato, K. M. (1993). Current trends and patterns of female migration: Evidence from mexico. International Migration Review, 27(4), 748–771. Durand, J., Massey, D. S., & Zenteno, R. M. (2001). Mexican immigration to the united states: Continuities and changes. Latin American Research Review, 36(1), 107–127. Feng, S., Krueger, A. B., & Oppenheimer, M. (2010). Linkages among climate change, crop yields and mexico-us cross-border migration. Proceedings of the National Academy of Sciences. Halliday, T. (2006). Migration, risk, and liquidity constraints in el salvador. Economic Development and Cultural Change, 54(4), 893–925. Kanaiaupuni, S. M. (2000). Reframing the migration question: An analysis of men, women, and gender in mexico. Social Forces, 78(4), 1311–1347. Liebig, T., & Sousa-Poza, A. (2004). Migration, self-selection and income inequality: An international analysis. Kyklos, 57(1), 125–146. Lodigiani, E., & Salomone, S. (2012). Migration-induced transfers of norms. The case of female political empowerment. IRES Discussion Papers 2012-1. Lu, M. (2005). Do people move when they say they will? Inconsistencies in individual migration behavior. Population and Environment, 20(5), 467–488. Manski, C. F. (1990). The use of intentions data to predict behavior: A best-case analysis. Journal of the American Statistical Association, 85(412), 934–940. McLeman, R., & Smit, B. (2006). Migration as an adaptation to climate change. Climatic Change, 76(1-2), 31–53. Munshi, K. (2003). Networks in the modern economy: Mexican migrants in the U.S. labor market. The Quarterly Journal of Economics, 118(2), 549–599. Parrado, E. A., & Flippen, C. A. (2005). Migration and gender among mexican women. American Sociological Review, 70(4), 606–632. Pugatch, T., & Yang, D. (2011). The impact of mexican immigration on U.S. labor markets: Evidence from migrant flows driven by rainfall shocks. Technical report. Rakodi, C. (2002). A livelihoods approach—Conceptual issues and definitions. In C. Rakodi, & T. Lloyd-Jones (Eds.), Urban livelihoods: A people-centered approach to reducing poverty (pp. 3–22). London: Earthscan. Rubalcava, L., & Teruel, G. (2006). User’s guide for the mexican family life survey first wave. Technical report.

Rubalcava, L., & Teruel, G. (2008). User’s guide for the mexican family life survey first wave. Technical report. Stillman, S., McKenzie, D., & Gibson, J. (2009). Migration and mental health: Evidence from a natural experiment. Journal of Health Economics, 28(3), 677–687. Tacoli, C. (2009). Crisis or adaptation? migration and climate change in a context of high mobility. Environment and Urbanization, 21(2), 513–525. Van Dalen, H., & Henkens, K. (2012). Explaining emigration intentions and behavior in the netherlands, 2005–10. Population Studies: A Journal of Demography.

APPENDIX A A.1 Intention and migration data Figure 1 illustrates the structure of the data and the way they are exploited in this paper. For simplification purposes, intention to migrate is represented as a continuous process with a beginning, and that ends either when the individual migrates or when she gives up her plan. It is assumed that migration is always preceded by the intention to migrate, although this period may be brief, since forced migrations are excluded in the Mexican context. The two survey waves are materialized on the graph by the vertical lines. Selected different cases of interest are represented in Figure 1: situation 1 represents individuals who are already migrants in the first wave of the survey. Observations 2 to 5 correspond to individuals who state their intention to migrate in the first wave. While individual 3 actually migrates, individual 2 gives up her plan between the two waves, and individuals 4 and 5 had not yet migrated at the time of the second wave. Individual 6 had no intention to migrate in 2002 but had the intention to migrate in 2005, and finally gave up. Situation 7 illustrates the case of individuals that migrate between the two waves although they had not yet the intention to move in 2002. Situations 1 and 6 are out of the scope of this article. This paper focuses on situations 2 to 5 and addresses the following questions: among individuals who had the intention to migrate in 2002, who are those who were able to carry out their plans between the two survey waves, when others could or would not? Which shocks, costs, or constraints matter ?. Figure 1 also shows that migration intentions stated in 2002 and not yet realized in 2005 are not symmetrical to observed migration in 2005 of individuals with no intention to migrate in 2002. This latter case is illustrated by observation 7. Indeed, although we could be tempted to interpret these migrations as

Figure 1. Migration intentions and time.

MEXICAN MIGRANTS TO THE US

549

Age distribution, without attritors With intention to migrate in Mexico

kdensity age 0 .01 .02 .03 .04 .05

kdensity age 0 .01 .02 .03 .04 .05

With intention to migrate abroad

20

30

40 x

50

Male

60

20

30

Female

40 x

50

Male

Female

kdensity age 0 .01 .02 .03 .04 .05

Internal migrants

kdensity age 0 .01 .02 .03 .04 .05

Migrants abroad

60

20

30

40 x Male

50

60

20

30

Female

40 x Male

50

60 Female

Figure 2. Intention and migration.

Figure 3. Unemployment rate of Mexicans in the US.

“surprise” migrations or inconsistent behaviors, they merely result from observational issues due to the timing of the survey. The fact that migration intentions “beginning” between the two waves are obviously not recorded in the first wave does not mean that they have never existed. We thus expect the same determinants to similarly affect migration, conditional on migration intention measured in 2002 and conditional on the absence of migration intention in 2002. The only difference is that shock variables are noisy in the latter case, since shocks may have occurred before the relevant period corresponding to the “beginning” of migration intentions. A.2 Analysis of attrition Attrition is a major issue for empirical research using panel data to address migration related questions since migration is very likely the most important cause of attrition. Another concern in this paper is the loss of observations due to missing intentions data at the baseline survey. Since these two sources of data loss are driven by very different reasons, they are analyzed separately.

The initial sample, in the first survey wave, is made of 19,177 adults aged 18–64 years. Intention data were collected for 15,917 adults only (83%). For the remaining 3,260 individuals, the section on intentions to migrate could not be filled in because they were not present in the household at the time of the survey. The problems that may arise from this reduction of my sample of interest are potentially biased estimation results, if these individuals were more likely to both have the intention to migrate and actually migrate. In that case, I would overestimate discrepancies between intentions and subsequent actions, by considering only those who were at home in 2002 and could be interviewed. However this may not be a major issue, since the main objective of this article is to understand the observed discrepancies and not assess their frequency or probability. In this respect I can consider that my sample of interest is made of the 15,917 individuals with recorded answers to the intention question. As regards attrition, the re-contact rate of these 15,917 individuals with non-missing intentions data is around 92%. 1,237 individuals are lost. My 15,917 individuals of interest form 7,474 households in 2002. 9.58% of them (673) are affected by attrition. In 668 cases, the whole household is lost, but they are mostly small households (84% are made of one or two individuals). The attrition rate is unsurprisingly higher in two of the three states at least partly included in the greater Mexico city (Federal district: 16.23%, Puebla: 13.55%). Similarly, the attrition rate is higher in urban areas. There is no gender bias since we lose 8.27% of women and 8.13% of men. As expected, as regards intentions to migrate, attritors have higher initial intentions to move (24.35% against 16.48% of non-attritors). Then, a third source of data loss is due to the temporary absence of some individuals when the second wave was collected in 2005. In 2005, these individuals are either declared to be currently residing household members by other members of their household, or considered “movers” and not migrants according to the definition I chose. As stated above, I do not define as migrants those individuals who moved within the same locality. In both cases, no information was collected on their past migration status (during 2002–05) and I cannot

550

WORLD DEVELOPMENT Table 9. List of variables

Variable Education: primary Education: secondary Education: Preparatoria or tertiary (d) Education: vocational (d) Married (d) Household size Children <15 (d) Elderlies in the household (d) Log per capita total expenditures Relative (s) in the US (d) Locality migration network in Mexico Locality migration network abroad Regional dummies Border region (d) Center region (d) Periphery region (d) Historic region (d) Urban strata1 (d) Urban strata2 (d) Urban strata3 (d) Rural strata (d) Adverse economic shock (d) 2002–05 Death or illness shock (d) 2002–05 Rainfall z-score 2003 Rainfall z-score 2004 High rainfall 2003–04(d) Hurricane 2002–04 (d) Number of storms 2002–04 Crime evolution 2002–05 Crime increased past 12 months (d) Violence to women increased, past 12 months (d)

Definition Equal to 1 if the interviewee’s highest education is primary schooling; 0 otherwise Equal to 1 if the interviewee’s highest education is secondary schooling; 0 otherwise Equal to 1 if the interviewee’s highest education is tertiary schooling; 0 otherwise Equal to 1 if the interviewee’s highest education is vocational; 0 otherwise Equal to 1 if the interviewee is married in 2002; 0 otherwise Number of persons living in the household in 2002 Equal to 1 if the interviewee has children aged less than 15 years in 2002; 0 otherwise Equal to 1 if at least one household member is aged 65 and over in 2002; 0 otherwise Annual amount declared in 2002 for the household Equal to 1 if the interviewee has relatives in the US in 2002; 0 otherwise Percentage of internal migrants among all adults from other households in the same locality Percentage of migrants abroad among all adults from other households in the same locality Equal to 1 if the household lives in 2002 in one of the following states; 0 otherwise: Baja California Sur, Coahuila, Nuevo Leo´n, Sinaloa and Sonora Me´xico, Morelos, Oaxaca, Puebla and Distrito Federal Veracruz and Yucata´n Durango, Guanajuato, Jalisco and Michoaca´n Equal to 1 if the household lives in 2002 in a locality with more than 100,000 inhabitants; 0 otherwise Equal to 1 if the household lives in 2002 in a locality with 15,000 to 100,000 inhabitants; 0 otherwise Equal to 1 if the household lives in 2002 in a locality with 2,500 to 15,000 inhabitants; 0 otherwise Equal to 1 if the household lives in 2002 in a locality with less than 2,500 inhabitants; 0 otherwise Equal to 1 if the household suffered unemployment, natural shock, loss of animals or crops during 2002–05; 0 otherwise Equal to 1 if the household suffered illness or death of any of its members during 2002–05; 0 otherwise State level normalized rainfall in 2003 State level normalized rainfall in 2004 Equal to 1 if min (rainfall z-score 2003, rainfall z-score 2004) >2; 0 otherwise Equal to 1 if at least one hurricane hit part of state area during 2002–04; 0 otherwise Number of hurricanes and tropical storms hitting state area during 2002–04; 0 otherwise State level percentage change in the number of registered crimes during 2002–05 Community level, equal to 1 if crime is estimated to have increased in 2005; 0 otherwise Community level, equal to 1 if violence to women is estimated to have increased in 2005, 0 otherwise

discriminate between return migrants and “stayers”. I thus chose to drop them from my sample. The percentage of these additionally lost individuals with the intention to move is higher than average (22.62%). There is a strong gender bias among these additionally lost observations: men are more often lost (14.31% of them are not in the final sample) than women (6.58%). In the remainder I include them in the category of attritors. In the results previously shown, I chose to drop all individuals whose migration status during 2002–05 could not be identified (including individuals who died between the two dates). The major problem is that this obviously leads me to underestimate migration, and may bias my results. To test their robustness I apply three different treatments to attritors and consider that: 1. all attritors migrated within Mexico, which amounts to overestimating internal migration and underestimating international migration 2. all attritors migrated abroad, with symmetrical implications: underestimation of internal migration and (large) overestimation of international migration 3. all attritors who had the intention to migrate abroad actually did it, and that the remainder migrated within Mexico.

All regressions whose results are shown above are then rerun on the sample augmented with attritors. Note that the few specifications requiring household variables measured in 2005 could not be tested (household shocks, Tables 3 and 4, last column). The reason why I apply the third treatment is the following: since I am interested in understanding the discrepancies between intentions to move and actual moves, and particularly understanding why individuals with the intention to move may not have migrated (yet) three years later, attrition would bias my results the most if all attritors who had the intention to move actually did so. The third treatment very likely overstates the consistency of individual behaviors with their previous intentions. Nonetheless, results (not shown, available upon request) are not challenged by this highly hypothetical assumption which represents the highest bound of the correlation between intentions and actual behaviors. Only minor changes are observed, and in particular, including attritors in the estimation sample of the male intention to move abroad equation makes negative coefficient on the dummies for the highest levels of education significant. The results presented in this article are thus robust to different treatments of attrition.

MEXICAN MIGRANTS TO THE US

551

Table 10. Summary statistics: individual and household variables Mean/pct Male (d) Age No education (d) Education: primary (d) Education: secondary (d) Education: preparatoria or tertiary (d) Education: vocational (d) Married (d) Household size Children <15 (d) Elderlies in the household (d) Log per capita total expenditures Relative (s) in the US (d) Locality migration network in Mexico Locality migration network abroad Border region (d) Historic region (d) Center region (d) Periphery region (d) Urban strata1 (d) Urban strata2 (d) Urban strata3 (d) Rural strata (d) Human shock 2002–05 (d) Adverse economic shock 2002–05 (d)

40.9 36.676 9.1 45.8 27.8 11.8 5.5 70.4 5.018 53.8 14.2 8.917 36.1 12.689 2.354 33.1 27.3 25.2 14.4 33.1 9.4 11.4 46.2 17.3 11.1

Observations

13,038

Standard error 13.029

2.187

0.904 7.008 2.667

(d) dummy variables.

Table 11. Summary statistics: state and community level shock variables

Rainfall z-score 2003 Rainfall z-score 2004 High rainfall 2003–04 (d) Hurricane 2002–04(d) Number of storms 2002–04 Crime evolution 2002–05 Observations Earthquake (d) Hurricane (d) Crime increased, past 12 months (d) Violence to women increased, past 12 months (d) Observations (d) dummy variables.

Mean/pct

Standard error

2.214 1.342 62.5 18.8 0.875 3.631 16 4.7 18.0 37.3 48.7

1.454 1.491

150

0.957 41.910

552

Table 12. Marginal effects on the probability to migrate in Mexico during 2002–05 conditional on intention to migrate in Mexico in 2002 Marginal effects on Prob (migration in Mexico= 1j intention to migrate in Mexico= 1) (1) Female (d)

(2) **

0.022 (0.010)

(3) **

0.018 (0.009)

State level weather shocks Rainfall z-score 2003

(4) **

0.019 (0.009)

(5) **

0. 018 (0.009)

(6) **

0.018 (0.009)

(7) **

0.019 (0.009)

(8) **

0. 018 (0.009)

(9) **

0.019 (0.009)

(10) **

0.018 (0. 009)

(11) **

0.018 s (0.009)

(12) **

0. 018 (0.009)

0.006 (0.005) 0.009* (0.005)

Rainfall z-score 2004

0.007 (0.015)

High rainfall 2003–04 (d) Hurricane 2002–04 (d)

0.004 (0.013)

Number of storms 2002–04

0.009 (0.005)

Community level shocks Earthquake (d)

0.003 (0.018)

Crime and violence Crime evolution (state level) 2002–05

0.001*** (0.000)

Crime increased (community) past 12 months (d) Violence to women increased (community) past 12 months (d) Household shocks Death/illness 2003–2005 (d)

0.010 (0.013) 0.004 (0.012)

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

0.021 (0.014) 0.029 (0.020) Yes

0.103 13,036

0.089 13,036

0.089 13,036

0.089 13,036

0.089 13,036

0.090 13,036

0.089 13,036

0.089 13,036

0.089 13,036

0. 089 13,036

0.089 13,036

0.078 12,842

Adverse economic shock 2003–2005 (d)

Marginal effects; Standard errors in parentheses. (d) for discrete change of dummy variable from 0 to 1. Controls included are age, education, regional, and geographic dummies. * p < 0.10. ** p < 0.05. *** p < 0.01.

Available online at www.sciencedirect.com

ScienceDirect

WORLD DEVELOPMENT

0.031** (0.014)

Hurricane (d)

Controls Estimated probability Prob (migrant 2005= 1j intention 2005= 1) Observations

0.023*** (0.009)