SELF-SELECTION AND INTERNATIONAL MIGRATION:
NEW EVIDENCE FROM MEXICO
Robert Kaestner and Ofer Malamud*
Abstract—This paper examines the selection of Mexican migrants to the
United States using novel data with rich premigration characteristics that
include permanent migrants, return migrants, and migrating households.
Results indicate that Mexican migrants are more likely to be young, male,
and from rural areas compared to nonmigrants, but they are similar to
nonmigrants in cognitive ability and health. Migrants are selected from
the middle of the education distribution. Male Mexican migrants are
negatively selected on earnings, and this result is largely explained by differential returns to labor market skill between the United States and Mexico rather than proxies for differential costs of migration.
I.
Introduction
M
EXICAN immigrants account for one-third of all foreign-born residents in the United States and are the
fastest-growing immigrant group. It is estimated that in 2009,
11.4 million Mexican immigrants lived in the United States,
and this figure represents approximately 4% of the U.S. population and 11% of Mexico’s population.1 The size of the
Mexican immigrant population in the United States, and the
potential consequences associated with an increase in population of this magnitude, makes it important to identify characteristics of these immigrants that are likely to affect socioeconomic outcomes in both the United States and Mexico. For
example, if migrants are relatively less educated, this will
tend to reduce the relative scarcity of highly educated labor
and reduce earnings disparities by education in Mexico. Positive selection of migrants with respect to education would
have opposite effects. The education of Mexican immigrants
in the United States also has implications for U.S. labor market and immigration policy. Negative selection of Mexican
immigrants will tend to reduce the wages of less educated,
native-born persons and exacerbate earnings inequality,
although the evidence in support of this hypothesis remains
controversial (Borjas, Freeman, & Katz, 1996; Lalonde &
Topel, 1997; Borjas, 2003; Card, 2005). Similarly, the selection of migrants with respect to other characteristics such as
age, health, cognitive ability, and earnings has social and economic implications for both Mexico and the United States.
Received for publication November 29, 2010. Revision accepted for
publication October 11, 2012.
* Kaestner: University of Illinois, Chicago, and NBER; Malamud: University of Chicago and NBER.
We thank the editor, Alberto Abadie, and three anonymous referees for
their comments. Kerwin Charles, Jesus Fernández-Huertas Moraga, Jeff
Grogger, Bruce Meyer, Stephen Trejo, and Abigail Wozniak, as well as
seminar participants at the Georgetown University, University of Chicago, the University of Notre Dame, the University of Virginia, and the
Universitat Pompeu Fabra, provided very useful suggestions. All errors
and opinions are our own.
A supplemental appendix is available online at http://www.mitpress
journals.org/doi/suppl/10.1162/REST_a_00375.
1
Pew Hispanic Center, Statistical Profile: Hispanics of Mexican Origin
in the United States, 2009, http://www.pewhispanic.org/2011/05/26
/hispanics-of-mexican-origin-in-the-united-states-2009/.
This paper uses data from the Mexican Family Life Survey
(MxFLS) to accomplish two objectives. First, we compare
characteristics such as the age, education, health, and cognitive ability of Mexicans who migrate to the United States
with nonmigrants who remain in Mexico. This descriptive
analysis of Mexican migrant characteristics is relevant for
policy because regardless of the causes of the pattern of
selection, many consequences of migration for Mexico and
the United States depend on the characteristics of those who
choose to migrate. The MxFLS is ideally suited for this purpose because it contains information on individual characteristics prior to migration for a large, geographically diverse
sample of Mexicans. These data have not been previously
used to examine the selection of Mexican migrants, and they
enable us to address several of the data limitations that have
plagued earlier studies. The MxFLS avoids the potential
undercount of Mexican migrants and possible overstatement
of educational attainment in the U.S. Census. Moreover, in
contrast to most Mexican surveys, the MxFLS allows us to
identify all migrants, including those who have moved permanently to the United States or moved to the United States
with their entire households. Finally, the MxFLS provides
information on characteristics such as cognitive ability and
health that are generally unmeasured in other surveys.
Second, we use the MxFLS to describe the pattern of
selection of migrants with respect to earnings and examine
whether this pattern changes when we adjust for differences
in some of the costs and benefits of migration. Earlier studies have often focused on ‘‘skill’’ characteristics, such as
education and age (or experience), due to their strong associations with labor market earnings. Using the MxFLS, we
examine the pattern of selection with premigration earnings
directly. The MxFLS also contains detailed measures of
migration networks, credit availability, and household
assets, which serve as proxy variables for the costs of
migration. In addition, we construct measures of monetary
returns to migration using the Mexican and U.S. 2000 censuses. We use these variables to assess possible determinants of selection with respect to earnings by comparing
changes in the pattern of migrant selection with and without
adjusting for these measures of the costs and benefits of
migration.
Our descriptive analysis reveals a marked selection of
Mexican migrants with respect to age: both male and
female migrants are more likely to be younger, and there is
a near monotonic, negative relationship between age and
the probability of migrating. With respect to education, both
male and female migrants tend to be drawn from the middle
of the education distribution (four to nine years of schooling). There is little evidence of migrant selection with
respect to cognitive ability or health. Finally, rural residents,
The Review of Economics and Statistics, March 2014, 96(1): 78–91
Ó 2014 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
SELF-SELECTION AND INTERNATIONAL MIGRATION
unmarried persons, and those with a relative living in the
United States are much more likely to migrate than urban
residents, married persons, and those without a relative living in the United States, respectively.
With regard to earnings, which we analyze only for males
because of high rates of nonparticipation in the labor market
for women, Mexican migrants tend to come from the bottom
half of the (unconditional) earnings distributions. Interestingly, this pattern of selection of migrants with respect to
earnings is not due to the relationship between earnings and
three measures of the cost of migration: whether a person has
a relative in United States, has access to credit, or has significant assets. While some of these measures of costs of migration are significantly related to the probability of migration,
adding them to the regression has only a small impact on the
pattern of migrant selection with respect to earnings. In contrast, a significant part of the pattern of migrant selection
with respect to earnings can be explained by differences in
labor market returns to observed determinants of earnings
such as age and education. We find that Mexican males are
more likely to migrate to the United States when the labor
market return for people with their observed characteristics
is larger in the United States than in Mexico.
II.
Related Literature
Empirical analyses of the selectivity of migrants in the
economics literature are motivated by theories of migration
in which the decision to migrate depends on the costs and
benefits of migration (Sjaastad, 1962). Predictions about the
characteristics of migrants versus those of nonmigrants
depend on how these characteristics are related to the costs
and benefits of migration. For example, Borjas (1987) used a
Roy (1951) model to generate predictions about whether
migrants will be relatively low skilled or high skilled
depending on how skill affects the pecuniary benefits of
migration. In the absence of the differential time-equivalent
costs of migration by skill, migrants are positively selected
on skill when the returns to skill in Mexico are lower than in
the United States and negatively selected on skill when the
returns to skill are relatively higher in Mexico.
Several authors have drawn on the Borjas (1987) model
to motivate analyses of migrant selection with respect to
education and earnings. Chiquiar and Hanson (2005) used
Mexican and U.S. censuses to evaluate the selection of
Mexican immigrants in terms of observed skills and found
evidence of intermediate selection.2 Chiquiar and Hanson
(2005) first compared the educational attainment of nonmigrants in the Mexican Census to migrants in the 2000 U.S.
Census and showed that Mexican immigrants in the United
States are drawn from the middle of the Mexican education
distribution. They then compared the predicted wage distri-
bution for residents of Mexico using the 2000 Mexican
Census with the predicted wage distribution of Mexican
immigrants in the United States using the 2000 U.S. Census
had they been paid according to the skills prices in Mexico.
Predicted wages were used as a proxy for observed skill,
and wages were predicted using age, education, and marital
status. As with education, the authors found that migrants
are concentrated in the middle of Mexico’s observed skill
distribution. Given evidence that the returns to skill are
actually higher in Mexico than for migrants in the United
States, they conjectured that high-skilled individual may
have lower time-equivalent costs of migration (or that any
fixed costs of migration are relatively less important for
high-skill individuals).3
The use of the U.S. and Mexican censuses presents several empirical problems for Chiquiar and Hanson (2005)
besides the fact that there are no measures of past earnings
in Mexico for migrants: the likely undercount of immigrants in the U.S. Census, the possibility that Mexican
immigrants can obtain additional schooling after arriving in
the United States, return migration, and the likelihood that
unobserved characteristics are correlated with education.
Results remained relatively unchanged after attempts to
address these potential problems, but despite their comprehensive assessments, these concerns cannot be completely
alleviated.
Ibarraran and Lubotsky (2007) used the 2000 Mexican
Census to estimate differences in the educational attainment
between migrants and nonmigrants. They replicated Chiquiar
and Hanson’s (2005) finding that Mexican immigrants in the
2000 U.S. Census have more education (and are older) than
nonmigrants in the Mexican Census. However, they further
investigated potential overreporting of education by
migrants and undercounting of younger migrants in the
U.S. Census. They also exploited the fact that in the Mexican Census, heads of households were asked to list current
or past household members who had lived abroad during
the preceding five years. Unfortunately, the heads of household reported only a limited set of individual characteristics: age, gender, Mexican state of origin, and migration
patterns. Since no information about educational attainment
of migrants was available, the authors used measured characteristics to predict educational attainment. In contrast to
Chiquiar and Hanson (2005), Ibarraran and Lubotsky
(2007) found evidence of negative selection on education:
less educated Mexicans were more likely than highly educated Mexicans to migrate to the United States, and the
degree of negative selection was larger in regions with
higher returns to schooling. However, this approach is
based on the assumption that there are no differences
3
2
Feliciano (2005) also used the Mexican and U.S. censuses and found
evidence of positive selection on education of Mexican migrants in terms
of mean differences.
79
Caponi (2010) offers an alternative explantion for the U-shaped relationship between migration and education based on the imperfect transferability and intergenerational transmision of human capital. See also
Caponi (2011).
80
THE REVIEW OF ECONOMICS AND STATISTICS
between migrants and nonmigrants beyond those characteristics used to predict education.
Fernández-Huertas Moraga (2011) used data from the
Encuesta Nacional de Empleo Trimestral (ENET) survey to
address some of the limitations associated with using census data. The ENET survey follows households for five consecutive quarters and includes reports on whether any
household member has left for the United States. Consequently, it is possible to examine the selectivity of recent
migrants on their education attainment and actual earnings
prior to migration. Fernández-Huertas Moraga reported
finding negative selection on wages and ‘‘intermediate to
negative selection’’ on education for men aged 16 to 65
(although he reports positive selection for migration out of
rural Mexico).4 The ENET is not subject to the problems of
undercounting of migrants and overreporting of migrant
educational attainment that are present in U.S. Census.
However, these data still miss persons in households in
which all members have migrated to the United States.
Orrenius and Zavodny (2005) used data from the Mexican
Migration Project (MMP) to examine how various factors
influence the selectivity of undocumented migrants over
time.5 They found that migrants were drawn from the middle of the education distribution. They did not examine
other measures of skill. They also found that a greater cost
of migration, as measured by greater border enforcement, is
associated with relatively lower immigration of low-skilled
Mexicans. While the MMP includes both migrants and
nonmigrants living in Mexico, it does not include Mexicanborn household heads who have migrated to the United
States permanently. Furthermore, the data from the MMP
are collected retrospectively, which may result in recall
bias, and from a limited number of communities.
Finally, McKenzie and Rapoport (2010) used data from
the 1997 Encuesta Nacional de la Dinámica Demográfica
(ENADID) survey to examine the selection of recent Mexican migrants on education.6 This survey has information on
migrants in the United States who have a (coresident)
family member living in Mexico but not on migrants who
have permanently left Mexico and do not have family members who remained in Mexico. McKenzie and Rapoport
(2010) also incorporated the migration history of the community, as a measure of the cost of migration, into the analysis. They reported that migrants are negatively selected
on education in communities that have a high proportion of
persons who migrated, but that selection on education is
positive in communities that do not have high rates of
migration.
4
See Fernández-Huertas Moraga (2009) for further analysis of differences between urban and rural Mexico.
5
Munshi (2003) used the MMP to show that Mexican migrants are
more likely to be employed and hold a higher-paying job when their network in the United States is larger.
6
Durand, Massey, & Zenteno (2001) used the ENADID survey to show
that Mexican migrants have become more negatively selected over time
after controlling for cohort effects (though still predominantly drawn from
the middle of the education distribution).
This brief review highlights several gaps in the previous
literature examining the selectivity of Mexican migrants.
First, previous researchers have had to confront significant
data limitations. We avoid most of these issues by using
arguably the most appropriate data available to study the
selectivity of Mexican migrants.7 For example, studies that
used the U.S Census to identify Mexican migrants may miss
a significant portion of migrants because of undercounting,
and they may overstate the education of migrants if immigrants obtain additional schooling after arrival in the United
States. Other studies that have used Mexican surveys generally miss permanent migrants or migrants who move to the
United States with their entire households. The MxFLS
allows us to identify all migrants, even those who have
moved permanently to the United States or moved to the
United States with their entire households, and avoids the
potential undercount of Mexican migrants in the United
States. Census. Furthermore, the MxFLS includes return
migrants (those who will eventually return to Mexico) and
does not depend on recall to identify migrants and time of
migration.
Second, most previous research has focused on the selectivity of migrants with respect to education or predicted
earnings based on a limited set of observed characteristics
such as age and education. Actual earnings, however, differ
significantly from predicted earnings, and an important
question is where in the actual earnings, or skill, distribution Mexican migrants come from. Insofar as the MxFLS
has information on earnings of migrants before they leave
Mexico, we can examine the selection of migrants with
respect to actual earnings. We also consider the selectivity
of migrants with respect to two components of earnings: the
part of earnings determined by observed characteristics
such as age, education, and marital status and the remaining
(orthogonal) component of earnings. This is an interesting
comparison because the selectivity of migrants with respect
to these components of earnings may differ. Moreover, how
the pattern of selection of migrants for these two components of earnings changes after adjusting for measures of
the costs and benefits of migration provides additional
information about the causes of migrant selection with
respect to earnings and characteristics correlated with earnings such as education.
Third, the MxFLS enables us to examine the selectivity
of migrants with respect to other factors such as cognitive
ability and health, as well as potential proxies for the costs
of migration such as credit availability, family networks in
the United States, and household assets. Considering a
broader set of factors is important because of the important
social and economic consequences related to the selectivity
of migrants with respect to characteristics other than education and earnings. For example, the selectivity of migrants
7
Rubalcava et al. (2008) used these data to examine the health selectivity of migrants to the United States and found relatively little evidence of
differential selection on health.
SELF-SELECTION AND INTERNATIONAL MIGRATION
81
TABLE 1.—SUMMARY STATISTICS
Men
Migrated to United States
Rural
Age
Years of schooling
Worked during past year
Annual earnings
Cognitive ability
Married
In good health
Relative in United States
Household assets
Know person/places to borrow
Women
Mean
SD
Observations
Mean
SD
Observations
0.036
0.385
38.9
7.40
0.935
34,128
5.84
0.751
0.539
0.350
211,922
0.401
0.187
0.487
12.222
4.459
0.247
47,337
2.66
0.432
0.437
0.418
426,707
0.464
8,116
8,114
8,116
8,103
8,114
6,681
8,116
8,116
8,116
8,116
8,116
8,116
0.020
0.381
38.6
6.68
0.439
24,539
5.52
0.710
0.464
0.358
206,010
0.317
0.141
0.486
11.999
4.282
0.496
34,010
2.80
0.453
0.470
0.452
406,947
0.445
9,183
9,180
9,183
9,168
9,176
3,388
9,183
9,183
9,183
9,183
9,183
9,183
All summary statistics are calculated for Mexican males aged 21 to 65 and are unweighted. ‘‘Migrated to the United States’’ is an indicator variable for whether the individual moved to the United States between
2002 and 2006. ‘‘Rural’’ is an indicator variable for living in a community with fewer than 2,500 people. ‘‘Age’’ and ‘‘Years of schooling’’ are in years. Annual earnings’’ and ‘‘Household assets’’ are in 2002 pesos.
‘‘Cognitive ability’’ is the number of correct answers out of twelve on the Raven’s Progressive Matrices test. ‘‘Married’’ is an indicator variable for being married. ‘‘In good health’’ is an indicator variable for reporting one’s current health as ‘‘very good’’ or ‘‘good’’ rather than ‘‘regular,’’ ‘‘bad,’’ or ‘‘very bad.’’ ‘‘Relative in United States’’ is an indicator variable for having a relative in the United States. ‘‘Know person/places to
borrow’’ is an indicator for reporting that one knows person or places to borrow. See the online data appendix for more details on the construction of these variable. Source: Mexican Family Life Survey.
with respect to health has implications for local financing of
health care in areas where immigrants are concentrated.
Finally, few papers have directly assessed whether differences in earnings between Mexico and the United States are
associated with the probability of migration, a direct test of
economic models of migration. We examine whether differential labor market returns predict migration. While this
analysis is subject to certain caveats, we believe it is useful
in helping to identify the causes of selection with respect to
earnings. Specifically, it is reasonable to interpret the pattern of migrant selection with respect to earnings that
remains after adjusting for the dominant source of monetary
benefit of migration as due to the relationship between earnings and the costs of migration. Similarly, we examine
whether measures of costs of migration are significant determinants of migration and show how the pattern of migrant
selection with respect to earnings changes after adjusting for
the differences in the costs of migration.
III.
Data and Empirical Strategy
A. Data
The Mexican Family Life Survey is a longitudinal survey
of households from geographically diverse communities in
Mexico (Rubalcava & Teruel, 2006a, 2006b).8 The first
round of the survey was conducted from April to July 2002
and collected information from a sample of approximately
8,400 households or 35,000 individuals across 150 different
communities in 16 of Mexico’s 32 states. The baseline survey included both household information and individuallevel information for all members living in the household.
The survey covered many topics: educational attainment,
8
Sampling weights are available that can be used to produce representative estimates at the national, urban-rural, and regional level. We do not
use sampling weights in the main analyses because we stratify the sample
by age and gender. However, the appendix table presents robustness
checks that include sampling weights.
labor market participation and earnings, networks, assets,
credit availability, and cognitive ability. The second round
of the survey began in mid-2005 and completed in 2006,
with a 90% recontact rate. Those who had migrated to the
United States were contacted at a rate of over 91%.
Although information from the second-round interviews of
migrants to the United States has not been released, an indicator for whether the individual has moved to the United
States is available in the public-use data; note that all
migrants should be identified even if they could not be
found in the second round. In cases where entire households
were no longer present at the original address, the interviewers searched for the final destination of that particular
household by contacting close relatives, including those living in the United States. Indeed, approximately 13% of
migrants in the MxFLS migrate to the United States with
their entire households.
We restrict the sample to persons aged 21 to 65 in the
first round of the MxFLS.9 For analyses that focus on earnings, we restrict the sample to men because of the low labor
force participation rates of women and the absence of earnings information for women who do not work. Table 1 presents summary statistics for this sample. The fraction of
men who migrated to the United States over the three-year
period between the first and second round of the MxFLS is
3.6%, or 295 of 8,116 men. Among women, 2% migrated
during the same period. These rates of migration are consistent with other sources. Hanson and McIntosh (2010) used
consecutive Mexican censuses to estimate migration rates
and reported ten-year migration rates of approximately 12%
to 14% for young men in their 20s and 5% to 10% for men
in their 30s. In our sample, the three-year migration rates for
men in their 20s and 30s are 5.6% and 3.3%, respectively,
and are therefore consistent with those reported by Hanson
9
Given our focus on education and earnings, analyses that include persons under age 21 are complicated by the fact that education has not been
completed for a nontrivial fraction of young people and that many young
people do not work.
82
THE REVIEW OF ECONOMICS AND STATISTICS
and McIntosh (2010). Passel and Cohn (2009) used CPS data
to estimate that approximately 550,000 Mexican immigrants
arrived in the United States each year between 2003 and
2006. This figure represents about 1.1% of the nonelderly
adult population of Mexico. In the MxFLS, approximately
2.8% of the baseline adult (male and female) sample migrated between 2002 and 2005, which is roughly equivalent
to the annual rate reported by Passel and Cohn. Unfortunately, the actual number of migrants in the MxFLS is relatively small and prevents analysis of subsamples such as by
urban and rural status.
The remaining variables in table 1 are from the first
round of the MxFLS in 2002, and are reported by the individual, the head of household, or proxy through some other
knowledgeable person in the household. We combine information from these different sources to fill in missing values
and maximize the sample size.10 Notably, there is substantial agreement in information that can be derived from selfreports, proxy reports, and roster reports (those from the
head of household). For example, the correlation between
self-reported years of schooling, proxy-reported years of
schooling, and roster-reported years of schooling is 0.94.
Similar agreement is found with respect to whether a person
worked in the past year. There is also considerable agreement between self-reported, proxy-reported, and rosterreported earnings, though there is greater scope for
measurement error because individual reports of annual
earnings include earnings in the main job as well as any
earnings received from a secondary job, and several materially important sources of earnings such as profit sharing
and Christmas bonuses.11 We assess the sensitivity of our
findings with respect to earnings in two ways. First, we drop
all persons without self-reported earnings. Second, we predict earnings for those without self-reported or proxyreported earnings using estimates from a regression of selfreported or proxy-reported earnings on roster-reported earnings and other covariates. In both cases, results, which are
presented in the appendix table and discussed below, are
similar to those reported in the text.
The proportion of our sample living in rural areas is
approximately 38%, a little lower than the 41% of an analogous sample in the 2000 Mexican Census. The average age
of men and women in our sample is about 39 years, a little
higher than the 37.5 years of age in the 2000 Mexican Census. Furthermore, the average years of schooling of men
and women in our sample is almost identical to the 7.3 and
6.7 years of schooling reported for a similar sample of men
and women aged 21 to 65 in the 2000 Mexican Census. The
MxFLS also provides information on cognitive ability, a
10
This is also the procedure recommended by the principal investigators of the MxFLS (Rubalcava & Teruel, 2006a).
11
Federal law requires firms to participate in a profit-sharing program
in which employees receive 10% of the firm’s annual profits. Firms are
also required to pay a year-end Christmas bonus (Aguinaldo) to all
employees equivalent to at least two weeks of pay.
dimension of skill that is usually not measured in most data
sets. Cognitive ability is measured as the number of correct
responses to twelve multiple-choice questions of a Raven’s
(Raven, Court, & Raven, 1983) Progressive Matrices test,
with an average of about five to six correct answers. This
test is designed to assess general intelligence by measuring
the ability to form perceptual relations and reason by analogy independent of language and formal schooling.
Finally, there is information about marital status and a selfreported measure of health that we collapse to a dichotomous variable.
In addition, the MxFLS contains unique information on
plausible proxies for the costs of migration: presence of
relatives in the United States, household assets, and availability of credit. Approximately 35% of individuals have
relatives in the United States and about 30% and 40% of
men and women, respectively, know persons or places from
which to borrow. The presence of a relative in the United
States is a direct measure of the potential availability of a
network in the United States that will plausibly lower the
costs of migration (Massey, 1986; Curran & Rivero-Fuentes,
2003; McKenzie & Rapoport, 2007). This variable is similar
in spirit to the use of contemporaneous or historical migration rates in a person’s home community or state that has
been used by previous researchers as a measure of migration networks (McKenzie & Rapoport, 2010; Woodruff &
Zenteno, 2007). The other two variables are motivated by
concerns that liquidity constraints prevent migration. Having
access to credit and sufficient household assets, which we
measure, likely alleviates those constraints and makes it possible to migrate if desired (Stark & Taylor, 1991; Orrenius &
Zavodny, 2005; McKenzie & Rapoport, 2007). In the regression analysis, we collapse the measure of household assets
into tertiles: households with low, medium, or high levels of
assets.
B. Empirical Strategy
Our first objective is to describe differences between
migrants and nonmigrants with respect to several characteristics related to labor market skill: education, age (as a
proxy for experience), and cognitive ability. All of these
factors have an independent association with wages in both
Mexico and the United States. We highlight the characteristics of migrants most closely related to earnings, often
referred to as measures of skill, because of the large role
that earnings play in determining the benefits of migration
and thus the decision to migrate. In addition, we describe
differences between migrants and nonmigrants with respect
to several other characteristics, including health, marital
status, and whether the person lives in an urban or rural
area.
The availability of earnings prior to migration also allows
us to examine the selectivity of migrants with respect to
earnings directly instead of having to predict premigration
earnings of migrants using a limited number of covariates.
SELF-SELECTION AND INTERNATIONAL MIGRATION
Further, we can decompose earnings into two parts: the part
correlated with observed measures of skill (age, education,
and so on) and the part uncorrelated with these skill measures (residual earnings). While this decomposition is
somewhat arbitrary and depends on what observed factors
are used to partition earnings, we do it for two reasons. First,
it provides a link to past studies that did not have direct measures of earnings and therefore had to use a predicted measure of earnings. Second, the selectivity of migrants with
respect to the two components of earnings may differ, and
these differences can provide information as to the causes of
selection with respect to characteristics used to partition
earnings. We also explore how these results are affected
by using different sets of observed factors to partition earnings.
To decompose earnings into two parts, we estimate earnings regression models using two different specifications.
The first specification is a saturated model that includes a
complete set of interactions between indicator variables for
age (five categories), education (five categories), and marital
status (two categories). The second specification, which is
also a fully interacted specification, adds indicators for cognitive ability (five categories) and health (two categories).
The predicted values from these regressions represents the
part of earnings correlated with the measured variables; the
residual from these regressions is the remaining part of earnings. We divide each of these components of earnings, as
well as total earnings, into quintiles to assess the pattern of
migrant selection with respect to these earnings measures.
Our second objective is to examine how patterns of selection with respect to earnings change when explicit measures
of costs and benefits of migration are included in the regression model. This analysis is intended to shed light on
the causes of selection of migrants with respect to earnings. The logic underlying the analysis is straightforward.
Migrant selection with respect to any characteristic, for
example, earnings, depends on how that characteristic is
related to both costs and benefits of migration. Therefore,
including measures of one or more of the benefits and
costs of migration will affect the pattern of migrant selection with respect to earnings if earnings are correlated with
costs and benefits. We acknowledge that this analysis is
subject to important caveats because some measures of the
costs and benefits of migration may themselves depend on
migration. For example, a household’s future plans about
migration may influence asset accumulation, education,
and other determinants of the costs and benefits of migration. This is a general problem that characterizes most
research on migrant selection, and there are few good
solutions. Below, we describe how we attempt to address
the reverse causality problem when measuring the benefits
12
In the case of education, the potential bias is likely to reinforce the
pattern of negative selection because higher returns to schooling in Mexico would lead future migrants to accumulate less education than similar
nonmigrants (McKenzie & Rapoport, 2011).
83
of migration by differential labor market returns between
the United States and Mexico and when using the presence
of a relative (network) in the United States as a measure of
the costs of migration.12
We measure the costs of migration using three variables:
whether person has a relative in United States, whether a
person has access to credit, and whether a person has a significant amount of assets. Earnings are likely to be related
to these measures of costs, particularly whether a person
has access to credit and the amount of household assets.13
We assess these associations directly and discuss them later
in the paper, but we are mainly interested in the mediating
effect of these variables on the pattern of migrant selection
with respect to earnings. We measure the benefit to migration by calculating differential labor market returns between
the United States and Mexico to several characteristics
known to influence earnings: age, education, and marital
status. We allow the U.S.–Mexico difference in labor market returns to differ by Mexican state and urban or rural status because previous studies have shown that labor market
returns to age and education differ by these geographic divisions (Ibarraran & Lubotsky, 2007).
IV.
Results
A. Selection of Migrants with Respect to Age, Education,
Cognitive Ability, and Other Characteristics
We begin by describing the nature of selection of migrants
by education, age, cognitive ability, health, and other characteristics. Table 2 shows the proportion of migrants and nonmigrants by several characteristics. On average, migrants are
younger than nonmigrants, and there is a steep gradient in
the likelihood of migrating by age, particularly among men.
Compared to nonmigrants, migrants are significantly overrepresented among those aged 21 to 29 and significantly
underrepresented among those aged 48 to 65. Given that
earnings rise with age and that the labor market returns to
age are relatively higher in the United States than in Mexico
(authors’ calculations), a simple theory of migration would
predict that older persons would be more likely to migrate.
However, the longer time horizon and the likely lower costs
of migration due to community and family attachments may
explain much of why younger workers are so much more
likely to migrate. Indeed, age is a good example of how the
nature of selection on a characteristic depends critically on
the correlation between that characteristic and both the benefits and costs of migration.
13
McKenzie and Rapoport (2007) develop a model in which the association between household assets and migration is U-shaped for agricultural (rural) workers. The intuition from this model is that at low levels of
wealth (assets), increases in wealth facilitate migration because the family
has the capability to finance the move. At higher levels of wealth, increases
in wealth reduce migration because of the loss of household production of
the migrant.
84
THE REVIEW OF ECONOMICS AND STATISTICS
TABLE 2.—SELECTED CHARACTERISTICS BY MIGRATION STATUS
MEN
Age
21–29 years old
30–38 years old
39–47 years old
48–56 years old
57–65 years old
Education
0–3 years of schooling
4–6 years of schooling
7–9 years of schooling
10–12 years schooling
More than 12 years of schooling
Cognitive ability
5th (bottom) quintile
4th quintile
3rd quintile
2nd quintile
1st (top) quintile
Health status
Not good
Good
Urban or rural status
Rural
Urban
Marital status
Not married
Married
Family networks
No relative in United States
Relative in United States
Credit availability
No borrowing options
Have borrowing options
Household assets
3rd (bottom) tertile
2nd (middle) tertile
1st (top) tertile
WOMEN
Fraction of
Migrants
Differences from
Reference Category
Fraction of
Migrants
Differences from
Reference Category
0.0565
0.0335
0.0367
0.0184
0.0148
0.0230***
0.0199***
0.0381***
0.0417***
0.0294
0.0201
0.0159
0.0146
0.0129
0.0093**
0.0135***
0.0148***
0.0165***
0.0295
0.0413
0.0484
0.0346
0.0173
0.0118**
0.0189***
0.0050
0.0123**
0.0149
0.0223
0.0224
0.0260
0.0171
0.0074**
0.0075*
0.0111**
0.0022
0.0327
0.0302
0.0391
0.0396
0.0353
0.0025
0.0064
0.0069
0.0026
0.0163
0.0166
0.0215
0.0231
0.0235
0.0003
0.0052
0.0067
0.0072
0.0377
0.0293
0.0084*
0.0202
0.0187
0.0015
0.0560
0.0240
0.0320***
0.0260
0.0167
0.0093***
0.0450
0.0335
0.0116**
0.0252
0.0183
0.0070**
0.0228
0.0523
0.0295***
0.0075
0.0410
0.0335***
0.0358
0.0401
0.0043
0.0191
0.0237
0.0046
0.0313
0.0449
0.0328
0.0136**
0.0015
0.0140
0.0218
0.0251
0.0078**
0.0111***
***, ** and * indicate statistically significant differences between migrants and nonmigrants using a t-test at the 1%, 5%, and 10%, level respectively. The sample is Mexican males aged 21 to 65 and unweighted.
Source: Mexican Family Life Survey.
With respect to education, table 2 indicates that rates of
migration are significantly higher among those in the middle of the education distribution (four to nine years of education) than among those with the least amount of education. This is similar to the findings in Chiquiar and Hanson
(2005), Orrenius and Zavodny (2005), and Fernández-Huertas
Moraga (2011). It is also the case that most migrants come
from the middle of the education distribution: 63% of male
migrants and 58% of female migrants have between four and
nine years of schooling.
In contrast to age and education, migrants are relatively
similar to nonmigrants with respect to cognitive ability.
Rates of migration do not differ significantly by categories
of cognitive ability, although there is some evidence of
positive selection for women. The absence of any strong
evidence for the selection of migrants with respect to cognitive test scores is surprising given the selection on education and the relatively high positive correlation between
years of schooling and cognitive test score (r ¼ 0.49). This
finding implies that education and cognitive ability do not
have similar associations with the costs and benefits of
migration even though the two variables are correlated.14
There is also only limited support for selection with
respect to health. Among males, those who report better
health are less likely to migrate than those who report
poorer health, but the difference is only marginally significant. For women, self-reported health status is unrelated to
the probability of migration. Moreover, there were no statistically significant correlations between the probability of
migrating and several other measures of health, including
objective measures of health, such as obesity, high blood
pressure, and low hemoglobin. These findings were also
reported by Rubalcava et al. (2008), but they are notable
because of substantial evidence showing that at the time of
14
Notably, cognitive ability is significantly related to earnings among
males. In a simple bivariate regression, each correct answer on the
Raven’s test is associated with 7.6% higher earnings. Controlling for education, age, and state of residence, each correct answer is associated with
only 1.9% higher earnings, but this estimate remains statistically significant.
SELF-SELECTION AND INTERNATIONAL MIGRATION
85
TABLE 3.—ASSOCIATION BETWEEN MIGRATION AND EARNINGS
DEPENDENT VARIABLE: MIGRATED TO UNITED STATES
2nd quintile
3rd quintile
4th quintile
5th quintile
(1)
(2)
(3)
(4)
(5)
(6)
0.005
[0.007]
0.003
[0.007]
0.009
[0.007]
0.015**
[0.007]
0.006
[0.007]
0.005
[0.007]
0.013*
[0.007]
0.018**
[0.007]
0.005
[0.007]
0.001
[0.007]
0.005
[0.007]
0.005
[0.007]
0.025***
[0.007]
0.006
[0.007]
0.003
[0.007]
0.008
[0.007]
0.008
[0.007]
0.028***
[0.007]
0.005
[0.007]
0.003
[0.007]
0.009
[0.007]
0.015**
[0.007]
0.006
[0.007]
0.005
[0.007]
0.013*
[0.007]
0.018**
[0.007]
Differential returns
Differential wage levels
0.000
[0.000]
Relative in the United States
Credit availability
Household assets (medium)
Household assets (high)
Sample size
6,675
0.026***
[0.006]
0.006
[0.005]
0.008
[0.006]
0.000
[0.006]
6,675
6,675
0.026***
[0.006]
0.008*
[0.005]
0.009
[0.006]
0.007
[0.006]
6,675
6,675
0.000
[0.000]
0.026***
[0.006]
0.006
[0.005]
0.008
[0.006]
0.000
[0.006]
6,675
Robust standard errors clustered at the household level are in brackets. ***, **, and * indicate statistical significance at the 1%, 5%, and 10 % level, respectively. The dependent variable is equal to 1 if migrated
to the United States between 2002 and 2006, and 0 otherwise. The independent variables are quintiles of log annual earnings. Differential returns and wage levels constructed from Education. Age categories in the
2000 Mexican and U.S. censuses. All regressions are estimated using linear probability models for Mexican males aged 21 to 65, and are unweighted.
Source: Mexican Family Life Survey.
arrival, Mexican immigrants are healthier than U.S.-born
persons of similar age and education.
Figures in table 2 indicate significant selection of
migrants with respect to several other characteristics. Rural
residents have much higher rates of migration than urban
residents, particularly among men. Married persons have
significantly lower rates of migration than unmarried persons. Persons with a relative in the United States have significantly higher rates of migration than those without a
relative in the United States. For men, there is a nonmonotonic relationship between household wealth and migration,
with the highest rates of migration for men in the middle of
the wealth distribution. This is consistent with the predictions and results in McKenzie and Rapaport (2007). For
women, migration rates are highest from the top of the
wealth distribution. Finally, there is some weak evidence
for a positive association between migration and the availability of credit, but this is not significant.
To summarize, the estimates in table 2 show that Mexican
migrants differ from nonmigrants along several observable
dimensions: migrants tend to be younger and unmarried,
overrepresented in the middle of the education distribution,
and come from rural areas. Surprisingly, migrants and nonmigrants tend to have similar distributions of cognitive test
scores and similar health. We note that the finding of selection from the middle of the education distribution is similar
to some previous research and is somewhat of a puzzle given
that the returns to education in Mexico are generally higher
than in the United States. All else equal, we expect migrants
to come from the lowest-educated groups. However, table 2
also shows significant differences between migrants and
nonmigrants with respect to a number of other variables that
are correlated with education and the probability of migrat-
ing, such as age, health, and rural status. Thus, it may be difficult to infer the causes of the association between education and migration from the simple association between
education and migration reported in table 2.
B. Selection of Migrants with Respect to Earnings
One of the advantages of the MxFLS survey is the availability of earnings prior to migration for both migrants and
nonmigrants. This information allows us to examine the
pattern of migrant selection with respect to actual earnings
instead of earnings predicted from a few observed characteristics. Table 3 presents estimates of the selection of
migrants with respect to earnings for males. We do not use
the sample of females because of high rates of labor market
nonparticipation and the absence of earnings for those
who do not work. For males, approximately 90% of both
migrants and nonmigrants work, and there are no statistically or economically significant differences in the probability of working between male migrants and nonmigrants.
We divide persons into earnings quintiles and examine the
association between the probability of migration and these
quintiles of earnings. Results are not qualitatively different
if we use quartiles or tertiles of earnings instead of quintiles
(not shown but available on request).
Estimates in column 1 of table 3, which are unadjusted
for any other variables, indicate a pattern of negative selection with respect to earnings. The probability of migrating
is substantially lower among men in the top two quintiles of
the earnings distribution (42% and 25% respectively) relative to the bottom quintile, although only the estimate pertaining to the top quintile is statistically significant. This
result is similar to the only other study to examine earnings
86
THE REVIEW OF ECONOMICS AND STATISTICS
directly; Fernández-Huertas Moraga (2011) reported that
migrants were negatively selected on wages with low earners most likely to migrate. This pattern of selection of
migrants with respect to earnings suggests that earnings are
negatively associated with the benefits of migration or positively associated with the costs of migration.
In column 2, we present estimates of the pattern of
migrant selection with respect to earnings after adjusting
for differences in three measures related to the costs of
migration: whether a person has a relative in United States,
whether a person has access to credit, and whether a person
has significant assets. Estimates in column 2 suggest a
slightly stronger pattern of negative selection with respect
to earnings than estimates in column 1, with coefficients on
the top two quintiles more economically and statistically
significant. The change in the pattern of selection between
columns 1 and 2 is due to the positive association between
earnings and access to credit and household assets; those
with higher earnings are more likely to have access to credit
and have more household assets and are therefore more
likely to migrate. There is little systematic association
between earnings and the presence of a relative in the United States, although it is a strong predictor of migration.15
In addition, the slightly stronger pattern of negative selection evident in column 2 than in column 1 implies that the
benefits of migration are negatively related to earnings.
To assess whether the benefits of migration are a possible
explanation for the pattern of migrant selection with respect
to earnings, we construct a measure of the benefits of
migration based on differences in earnings in Mexico and
the United States. We do this by calculating the differential
in labor market returns between recent Mexican immigrants
in the 2000 U.S. Census and Mexican natives in the 2000
Mexican Census. We focus on labor market returns to
determinants that are observed in both data sets: age, education, and marital status. We also allow these labor market
returns to vary by urban and rural status and state of residence in Mexico. Specifically, we estimate Mincer-type
earnings regressions that include a full set of interactions
between age, education, and marital status and allow these
effects to differ by urban and rural status and state of residence in Mexico. We then merge the differences in estimated returns between the United States and Mexico by
age, education, and marital status.
Adjusting for differential labor market returns to skill
substantially alters the pattern of migrant selection on earnings. Estimates in column 3 of table 3 indicate little selection of migrants with respect to earnings. If anything, there
15
An important concern is that our proxies for migration costs, such as
having a relative in the United States, are correlated with unobserved
earnings ability abroad. Previous researchers (starting with Woodruff &
Zentano, 2007) have used historical migration rates from geographic
areas to proxy for the size and quality of migration networks. We followed this literature and used historical migration rates from 1955 to
1959 to instrument for whether a person has a relative in the United
States. These results, not shown, were very similar to those reported in
text and are available on request.
is some evidence that persons in the lowest quintile (reference category) are more likely to migrate than those in
other parts of the earnings distribution. The change in the
pattern of selection after this adjustment suggests that differential labor market returns are an important determinant
of migration and the primary explanation of the negative
selection of migrants with respect to earnings. Differential
labor market returns are also significantly correlated with
migration; a 1-unit increase (100 percentage point) in the
differential return increases the probability of migrating by
2.5 percentage points, or approximately 66% of the mean.
Estimates in column 4 of table 3 come from a model that
includes measures of the cost of migration and the differential return to labor market characteristics. The pattern of
selection revealed by estimates in this column is similar to
that shown by estimates in column 3 and indicate a slight
degree of negative selection of migrants with respect to earnings mainly due to a 15% higher rate of migration among
those in the bottom quintile of the earnings distribution
relative to those in other parts of the earnings distribution.
It is important to acknowledge the inherent selection bias
associated with these simple calculations of the labor market returns to migrants and nonmigrants from Mexico. After
all, the empirical evidence presented in this paper demonstrates that Mexican migrants to the United States are
selected on a variety of observed characteristics. Consequently, our estimates of the labor market return to skill for
Mexican migrants in the United States may not represent
the actual returns faced by a random person in Mexico.16
Addressing this selection problem in a compelling manner
is extremely challenging, but we do try to assess the severity of the problem with two alternative approaches. First,
we obtained selection-corrected estimates of the labor market return to skill in the United States using Heckman’s
(1979) two-step estimator.17 Second, we calculated the
returns to skill for alternative samples in the U.S. Census:
all Mexican immigrants in the United States (as opposed to
only recent immigrants who arrived in the previous ten
years), other Spanish-speaking immigrants from Central
and South America, and U.S. natives who report Mexican ethnicity. In all cases, results from these alternative approaches
(not shown but available on request) yielded findings very
similar to those presented below. While we cannot rule out
the possibility that the selection problem is confounding our
estimates, we believe it is unlikely that the conclusions of the
16
However, to the extent that individuals in Mexico are not aware of
the selection problem and make their migration decisions on the basis of
the observed returns of migrants to the United States, these biased estimates are the appropriate ones to include in the model.
17
In particular, we used a probit regression to predict migration from
Mexico based on age and education categories as well as marital status
and number of children (assumed to be excluded from the second-stage
model). We then used the parameters of this model to construct the
inverse Mills ratio associated with the probability of being a migrant for
the sample of Mexican migrants drawn from the U.S. Census and reestimated the earnings regressions, including the selection term.
SELF-SELECTION AND INTERNATIONAL MIGRATION
analysis would change qualitatively even if we could more
effectively address the selection bias.
It is also important to note that we have followed the
model of Borjas (1987) to measure the earnings benefit of
migration by using differences in the relative return to skill
(log differences in earnings). We consider an alternative
characterization of earnings benefits proposed by Grogger
and Hanson (2011), for which we use skill-specific difference in the level of earnings between the United States and
Mexico. In the last two columns of table 3, we present estimates of the pattern of migrant selection with respect to
earnings from regression models that use the level difference in labor market returns of observed characteristics
instead of the log difference. In this case, the differential
labor market returns are not significantly related to the
probability of migration and adding them to the regression
has virtually no impact on the pattern of selection.
Overall, estimates in table 3 indicate that Mexican migrants
are negatively selected with respect to earnings and suggest that a potential cause for this pattern of selection is
due to differences in the benefits of migration that are
related to earnings. High earners in Mexico have smaller
relative benefits than low earners, as measured by differential labor market returns, and this makes them less likely
to migrate. Costs of migration, as measured by the presence of family networks in the United States, availability
of credit, and household assets, do not have a very strong
influence on the pattern of selection with respect to earnings. However, there is some evidence that differential
costs of migration may slightly weaken the negative selection which suggests that earnings are negatively related to
these costs measures.
Finally, in the appendix table, we explore whether our
main results are robust to different specifications of earnings. Specifically, we assess the sensitivity of our findings
to the following: limiting the sample to those with selfreported earnings; imputing earnings for those without selfreported or proxy earnings by predicting earnings using
estimates from a regression of self-reported or proxy earnings on roster-reported earnings and other covariates (education, age, state of residence, marital status, and urban or
rural status); using sample weights to estimate weighted
least squares; and limiting the sample to young men (using
combined information on earnings from self reports and
household roster). Overall, estimates in the appendix table
are consistent with those in table 3 and indicate negative
selection of migrants with respect to earnings. There is
somewhat stronger evidence of negative selection among
younger men than among the full sample (or older men),
and there is somewhat weaker evidence of negative selection when only those with self-reported earnings are used
instead of the full sample with earnings information from
either self-reports or the household roster. The use of
imputed earnings or sample weights has little effect on the
estimated pattern of migrant selection with respect to earnings. Our results (not shown but available on request) are
87
also robust to excluding past migrants, return migrants, and
individuals who migrated with their entire households.
C. Selection of Migrants with Respect to Predicted
Earnings and Residual Earnings
As noted in our literature review, previous studies examining Mexican migrant selection that lacked information on
pre-migration earnings in Mexico have used predicted earnings to assess the selection of migrants with respect to skill,
or human capital. To link our study to this prior work, we
also construct predicted earnings using labor market returns
to several observed components of skill. We construct this
index of observed skill by regressing observed earnings on
various combinations of age, education, marital status, cognitive test scores, and health to form a predicted measure.
However, because we have information on actual earnings,
we can also examine migrant selection with respect to residual
earnings—the part of earnings uncorrelated with observed
components of skill. We assess whether the probability to
migrate differs by these measures of predicted earnings and
residual earnings. Again, we divide predicted and residual
earnings into quintiles, although results are similar when using
quartiles or tertiles.
Table 4 presents regression estimates of the association
between the probability of migrating and predicted and residual earnings. Panel A of table 4 presents estimates from a
model that uses age, education, and marital status to predict
earnings (and residual earnings). Panel B presents estimates’from a model that uses age, education, marital status,
cognitive ability, and health to predict earnings. For each
dependent variable, we estimate two models—one without
covariates and one that includes the three measures related
to the cost of migration: whether person has a relative in
United States, whether a person has access to credit, and
whether a person has a significant amount of assets. In this
analysis, we do not include differential labor market returns
as a measure of the benefits of migration because of the
high degree of collinearity between predicted earnings and
differential labor market returns, which are constructed
from the same set of observed characteristics.
The estimates in these specifications are all relatively
imprecise, in part because we adjust the standard errors to
account for the two-step procedure in which predicted and
residual earnings are estimated from a separate regression.18 The point estimates in column 1 of the panel A of
table 4 provide some suggestive evidence for selection from
the middle of the predicted earnings distribution, with individuals in the third quintile about 1 percentage point more
likely to migrate to the United States relative to the bottom
quintile. However, this estimate is not statistically significant. The probability of migrating is 1.3 percentage points
lower among those in the top quintile of the predicted earn18
Specifically, we bootstrap the standard errors by performing 500
replications of the two-step estimation procedure.
88
THE REVIEW OF ECONOMICS AND STATISTICS
TABLE 4.—ASSOCIATION BETWEEN MIGRATION AND PREDICTED AND RESIDUAL EARNINGS
DEPENDENT VARIABLE: MIGRATED TO THE UNITED STATES
Predicted Earnings
(1)
2nd quintile
3rd quintile
4th quintile
5th quintile
Relative in the United States
Credit availability
Household assets
Sample size
2nd quintile
3rd quintile
4th quintile
5th quintile
Relative in the United States
Credit availability
Household assets
Sample size
Residual Earnings
(2)
(3)
A. Partitioned by Age, Education, and Marriage
0.007
0.008
0.002
[0.011]
[0.011]
[0.008]
0.010
0.010
0.002
[0.011]
[0.010]
[0.008]
0.005
0.004
0.006
[0.009]
[0.009]
[0.008]
0.013*
0.014*
0.012*
[0.008]
[0.008]
[0.007]
X
X
X
6,678
6,678
6,678
B. Partitioned by Age, Education, Marriage, Cognitive Abiility, Health
0.003
0.003
0.001
[0.009]
[0.010]
[0.009]
0.009
0.005
0.007
[0.009]
[0.008]
[0.007]
0.013
0.007
0.008
[0.009]
[0.009]
[0.008]
0.004
0.013*
0.014**
[0.007]
[0.007]
[0.007]
X
X
X
6,678
6,678
6,678
(4)
0.004
[0.008]
0.000
[0.008]
0.007
[0.008]
0.016**
[0.007]
X
X
X
6,678
0.003
[0.008]
0.009
[0.008]
0.008
[0.008]
0.014**
[0.007]
X
X
X
6,678
Robust standard errors clustered at the household level are in brackets. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% level, respectively. The dependent variable is equal to 1 if migrated
to the United States between 2002 and 2006, and 0 otherwise. The independent variables are quintiles of predicted log earnings and wages, constructed by regressing annual earnings or hourly wages on (five edlucation. (five age, and marital status, fully interacted. All regressions are estimated using linear probability models for Mexican males aged 21 to 65 and are unweighted.
Source: Mexican Family Life Survey.
ings distribution, and this estimate is marginally significant.
including the three measures of migration costs in the model
has little effect on the pattern of selection with respect to
predicted earnings, as evidenced by the similarity of estimates in columns 1 and 2 of the top panel of table 4. This
result suggests that the selection from the middle of the
predicted earnings distribution is not caused by the correlation between predicted earnings and these proxies for
costs of migration.
The point estimates in columns 3 and 4 of the top panel
of table 4 do not indicate selection from the middle. Instead,
we find evidence of negative selection of migrants with
respect to residual earnings, or the part of earnings uncorrelated with age, education, and marital status. Males in the
top quintile of residual earnings have a migration rate that
is 1.2 percentage points greater than males in the bottom
quintile, although the estimate is only marginally significant. The addition to the model of the three measures of the
costs of migration slightly increases the degree of negative
selection, and the coefficient on the top quintile becomes
statistically significant at conventional levels. Thus, migration costs, as measured, are not a particularly important
influence underlying the pattern of negative selection on
residual earnings.
In panel B of table 4, we constructed predicted earnings
using cognitive ability and health in addition to age, education, and marital status. According to this alternative specification, selection from the middle of the predicted earnings
distribution, is more apparent in the third and fourth quintiles, and there is no evidence of negative selection in the
absence of controls for the costs of migration. However,
adding these measures in column 2 in the panel B of table 4
leads to substantially less selection from the middle of the
predicted earnings distribution and indicates that those in
the top quintile of predicted earnings are the least likely to
migrate. The pattern of selection by residual earnings is
qualitatively similar between the top and bottom panels of
table 4, with those in the top quintile of the residual earnings distribution being 1.4 percentage points less likely to
migrate as compared to those in the bottom quintile.
As estimates in the panels indicate, the pattern of selection with respect to predicted (and residual) earnings
changes along with the variables used to construct these
measures. However, is an analysis of migrant selection with
respect to these measures useful? From a descriptive, policy
point of view, these measures are of limited use because predicted and residual earnings are arbitrary distinctions and
migrants’ human capital, or skill, is most accurately characterized by actual earnings. Moreover, it is difficult to link this
pattern of selection and the underlying cause of it to specific
characteristics because predicted earnings are a function of
several factors. Those in the middle of the predicted earnings
distribution have different ages, levels of education, and marital status. Therefore, it is not possible to link any specific
characteristic to the pattern of selection or the potential cause
of selection. The same issues apply to the interpretation of
SELF-SELECTION AND INTERNATIONAL MIGRATION
the pattern of selection with respect to residual earnings. For
this measure, estimates suggest that the benefits of migration
are largest for those in the lower end of the residual earnings
distribution. However, those in the lower part of the residual
earnings distribution are also a mix of persons with different
observed characteristics, and it is difficult to link these characteristics to potential explanations of the cause of migrant
selection.
V.
Conclusion
In this paper, we presented new evidence of the nature of
selection of Mexican migrants to the United States during
the period 2002 to 2005. Our analysis was based on a previously unused, high-quality data source, the Mexican
Family Life Survey (MxFLS), which is ideally suited to
study the selection of Mexican migrants. In particular, the
MxFLS has information on the education, cognitive ability,
health, and prior earnings of Mexican migrants, including
those in which the entire household migrated to the United
States. These data allow us to examine the nature of selection
of recent migrants on a more comprehensive set of characteristics than used in previous studies. Furthermore, the MxFLS
contains variables that are good proxies for the costs of
migration, and we use this information to help identify the
pattern of selection of migrants with respect to earnings.
Descriptive analyses found that Mexican migrants, both
males and females, were more likely to be young and
unmarried, to come from the middle of the education distribution (four to nine years of schooling), and to come from
rural areas. Interestingly, there was relatively little selection
of migrants with respect to cognitive ability and health,
although among women, there was some evidence that
migrants were more likely to be positively selected among
those with greater cognitive ability. With respect to earnings, which we examined only for males because of low
labor force participation rates among females, migrants
tend to come from the first through third quintiles of the
earnings distribution.
Analyses that investigated the causes of migrant selection
with respect to earnings revealed that the pattern of migrant
selection with respect to earnings may be primarily due to
the association between earnings and labor market benefits
of migration. Persons in the upper end of the earnings distribution had larger relative returns to labor market skills than
persons in the lower end of the earnings distribution. Controlling for these differences greatly diminished the negative pattern of selection with respect to earnings. In addition, differential labor market returns was a large and
significant predictor of migration. In contrast, three measures of the costs of migration were not strongly related to
earnings, and adjusting for differences in these variables
had relatively little effect on the pattern of migrant selection
with respect to earnings. Estimates indicated that earnings
were negatively related to these costs of migration. Overall,
our results suggest that there is negative migrant selection
89
with respect to earnings, or skill, and that this selection is
largely driven by differential benefits associated with earnings. The finding of negative selection with respect to earnings that is largely the result of higher benefits of migration
for those with relatively low earnings is consistent with the
most basic economic model of migrant selection such as
that presented by Borjas (1987).
We also presented analyses of the pattern of selection
with respect to predicted and residual earnings, which is
similar to analyses conducted in some previous papers on
the selection of Mexican migrants. The primary explanation
for the use of predicted earnings is the absence of earnings
information in Mexico for both migrants and nonmigrants.
These analyses indicated that Mexican migrants are selected
from the middle of the predicted earnings distribution and
the lower end of the residual earnings distribution. No previous study has analyzed the selection with respect to residual earnings and the difference in the pattern of migrant
selection with respect to predicted and residual earnings, is
interesting. Notably, and not surprisingly, these findings were
somewhat sensitive to the manner in which earnings were
partitioned into predicted and residual earnings. The arbitrary nature of the division of earnings into predicted and
residual earnings and the difference in the pattern of
migrant selection between predicted and actual (residual)
earnings raises questions about the efficacy of using predicted earnings as a measure of skill to assess migrant
selection with respect to skill. Similar concerns arise with
respect to using individual components of skill such as
education to assess migrant selection with respect to
skill.
In sum, we have used arguably the most appropriate data
to examine the pattern of Mexican migrant selection with
respect to several characteristics of interest to policy and
theory. However, our study, like most previous studies, has
some limitations. We use a sample of recent migrants, from
2002 to 2006, and this sample may not accurately characterize the full population of migrants, in the United States
(although the pattern of selection among recent migrants
may be most relevant for informing current and future policy). The sample of migrants is also too small to assess
whether patterns of migrant selection differ by urban or
rural status, network availability, and other potential causes
of heterogeneous associations. In addition, the measures of
the costs and benefits of migration we used in our study
have certain shortcomings. For example, whether a person
has a relative in the United States may be related to unobserved factors that affect the migration of both the person
and the relative. While we cannot rule out this possibility,
we note that our results (not shown) are similar when we
instrumented for this variable using historical migration
rates. Finally, we were able to make only limited headway
toward identifying the cause of migrant selection with
respect to earnings. This is a complicated problem that has
been underappreciated in the literature. Migrant selection
depends on the costs and benefits of migration, and identi-
90
THE REVIEW OF ECONOMICS AND STATISTICS
fying the pattern of selection with respect to any characteristic requires a full specification of the causal links
between the variable of interest and costs and benefits, and
between the variable of interest and other variables correlated with it and with costs and benefits of migration.
Almost no research has addressed this more fundamental
problem, which should be the focus of future research in
this area.
REFERENCES
Borjas, George J., ‘‘Self-Selection and the Earnings of Immigrants,’’
American Economic Review 77:4 (1987), 531–553.
Borjas, George J., ‘‘The Labor Demand Curve Is Downward Sloping:
Reexamining the Impact of Immigration on the Labor Market,’’
Quarterly Journal of Economics 118:4 (2003), 1335–1374.
Borjas, George J., Richard B. Freeman, and Lawrence F. Katz,
‘‘Searching for the Effect of Immigration on the Labor Market,’’
American Economic Review 86:2 (1996), 246–251.
Caponi, Vincenzo, ‘‘Heterogeneous Human Capital and Migration: Who
Migrates from Mexico to the US?’’ Annals of Economics and Statistics 97/98 (2010), 207–234.
——— ‘‘Intergenerational Transmission of Abilities and Self-Selection
of Mexican Immigrants,’’ International Economic Review 58:2
(2011), 523–547.
Card, David, ‘‘Is the New Immigration Really So Bad?’’ Economic Journal 115:207 (2005), 300–323.
Chiquiar, Daniel, and Gordon H. Hanson, ‘‘International Migration, SelfSelection, and the Distribution of Wages: Evidence from Mexico
and the United States,’’ Journal of Political Economy 113:2
(2005), 239–281.
Curran, Sara R., and Estela Rivero-Fuentes, ‘‘Engendering Migrant Networks: The Case of Mexican Migration,’’ Demography 40:2
(2003), 289–307.
Durand, Jorge, Douglas S. Massey, and Rene M. Zenteno, ‘‘Mexican
Immigration to the United States: Continuities and Changes,’’
Latin American Research Review 36:1 (2001), 107–127.
Feliciano, Cynthia, ‘‘Educational Selectivity in United States. Immigration: How do Immigrants Compare to Those Left Behind?’’ Demography 42:1 (2005), 131–152.
Fernández-Huertas Moraga, Jesus, ‘‘Wealth Constraints, Skill Prices or
Networks: What Determines Emigrant Selection?’’ mimeograph
(2009).
——— ‘‘New Evidence on Emigrant Selection,’’ this REVIEW 93:1 (2011),
72–96.
Grogger, Jeffrey, and Gordon H. Hanson, ‘‘Income Maximization and the
Selection and Sorting of International Migrants,’’ Journal of
Development Economics 95:1 (2011), 42–57.
Hanson, Gordon H., and Craig McIntosh ‘‘The Great Mexican Emigration,’’ this REVIEW 92:4 (2010), 798–810.
Heckman, James J., ‘‘Sample Selection Bias as a Specification Error,’’
Econometrica 47:1 (1979), 153–161.
Ibarraran, Pablo, and Darren Lubotsky, ‘‘Mexican Immigration and SelfSelection: New Evidence from the 2000 Mexican Census,’’ in G. J.
Boras (ed.), Mexican Immigration to the United States (Chicago:
University of Chicago Press, 2007).
LaLonde, Robert J., and Robert H. Topel, ‘‘Economic Impact of International Migration and the Economic Performance of Migrants,’’ in
M. R. Rosenzweig and O. Stark (eds.), Handbook of Population and
Family Economics, vol. 1B (Amsterdam: Elsevier Science, 1997).
Massey, Douglas S., ‘‘The Settlement Process among Mexican Migrants
to the United States,’’ American Sociological Review 51:5 (1986),
670–685.
McKenzie, David, and Hillel Rapoport, ‘‘Network Effects and the
Dynamics of Migration and Inequality: Theory and Evidence from
Mexico,’’ Journal of Development Economics 84:1 (2007), 1–24.
——— ‘‘Self-Selection Patterns in Mexico-U.S. migration: The role of
migration Networks,’’ this REVIEW 92:4 (2010), 811–821.
——— ‘‘Can Migration Reduce Educational Attainment? Evidence from
Mexico,’’ Journal of Population Economics 24:4 (2011), 1331–
1358.
Munshi, Kaivan, ‘‘Networks in the Modern Economy: Mexican Migrants
in the US Labor Market,’’ Quarterly Journal of Economics 118:2
(2003), 549–599.
Orrenius, Pia M., and Madeline Zavodny, ‘‘Self-Selection among Undocumented Immigrants from Mexico,’’ Journal of Development Economic 78:1 (2005), 215–240.
Passel Jeffrey, and D’Vera Cohn, ‘‘Mexican Immigrants: How Many Come?
How Many Leave?’’ Pew Hispanic Center report (July 22, 2009).
Raven, John C., John H. Court, and John E. Raven, Manual for Raven’s
Progressive Matrices and Vocabulary Scales: Standard Progressive Matrices (London: Lewis, 1983).
Roy, A. D., ‘‘Some Thoughts on the Distribution of Earings,’’ Oxford
Economics Papers, n.5. 3 (1951), 135–146.
Rubalcava, Luis N., and Graciela M. Teruel, ‘‘User’s Guide for the Mexican Family Life Survey First Wave,’’ (2006a), www.mxfls.uia.mx.
——— ‘‘Mexican Family Life Survey, Second Round,’’ working paper
(2006b), www.ennvih-mxfls.org.
Rubalcava, Luis N., Graciela M. Teruel, Duncan Thomas, and Noreen
Goldman, ‘‘The Healthy Migrant Effect: New Findings From the
Mexican Family Life Survey,’’ American Journal of Public Health
98:1 (2008), 78–84.
Sjaastad, Larry A., ‘‘The Costs and Returns of Human Migration,’’ Journal of Political Economy 70:5, pt. 2 (1962), 80–93.
Stark, Oded, and J. Edward Taylor, ‘‘Migration Incentives, Migration
Types: The Role of Relative Deprivation,’’ Economic Journal
101:408 (1991), 1163–1178.
Woodruff, Christopher M., and Rene M. Zenteno, ‘‘Remittances and
Micro-Enterprises in Mexico,’’ Journal of Development Economics 82:2 (2007), 509–528.
SELF-SELECTION AND INTERNATIONAL MIGRATION
91
TABLE APPENDIX—ROBUSTNESS CHECKS FOR THE ASSOCIATION BETWEEN MIGRATION AND EARNINGS
DEPENDENT VARIABLE: MIGRATED TO THE UNITED STATES
Imputed Earnings
(1)
2nd quintile
3rd quintile
4th quintile
5th quintile
Relative in the United States
Credit availability
Household assets
Differential returns
Sample size
0.001
[0.013]
0.001
[0.010]
0.006
[0.009]
0.017**
[0.008]
6,539
(1)
2nd quintile
3rd quintile
4th quintile
5th quintile
Relative in the United States
Credit availability
Household assets
Differential returns
Sample size
0.004
[0.009]
0.009
[0.009]
0.001
[0.009]
0.012
[0.007]
6,675
(2)
Self-Reported Earnings Only
(3)
A. Alternative Earnings Specifications
0.001
0.001
[0.012]
[0.012]
0.001
0.002
[0.010]
[0.010]
0.008
0.003
[0.009]
[0.010]
0.021**
0.013
[0.010]
[0.009]
X
X
X
X
X
X
X
6,539
6,539
B. Alternative Sample Specifications
Weighted Sample
(4)
(5)
0.010
[0.009]
0.007
[0.009]
0.002
[0.008]
0.009
[0.007]
3,665
0.010
[0.009]
0.008
[0.009]
0.002
[0.009]
0.009
[0.008]
X
X
X
3,665
(6)
0.013
[0.009]
0.010
[0.009]
0.002
[0.007]
0.000
[0.008]
X
X
X
X
3,665
Young Men (under 40 Years) Sample
(2)
(3)
(4)
(5)
(6)
0.003
[0.009]
0.008
[0.009]
0.003
[0.009]
0.015**
[0.007]
X
X
X
0.003
[0.009]
0.010
[0.009]
0.001
[0.009]
0.007
[0.007]
X
X
X
X
6,675
0.007
[0.010]
0.014
[0.010]
0.016
[0.010]
0.018*
[0.010]
0.008
[0.010]
0.015
[0.010]
0.019*
[0.010]
0.023**
[0.010]
X
X
X
3,829
3,829
0.008
[0.010]
0.013
[0.010]
0.012
[0.010]
0.012
[0.010]
X
X
X
X
3,829
6,675
Robust standard errors clustered at the household level are in brackets. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% level, respectively. The dependent variable is equal to 1 if migrated
to the United States between 2002 and 2006, and 0 otherwise. The independent variables are quintiles of predicted and residual log earnings and wages. Columns 2, 4, and 6 include controls for marital status, relative
in the United.State, visited the United States, having U.S. documentation, and thought about moving to the United States. All regressions are for Mexican males aged 21 to 65, and are unweighted.
Source: Mexican Family Life Survey.