Upload
cheresada-calderon-diaz
View
13
Download
1
Embed Size (px)
DESCRIPTION
Migración 2
Citation preview
The Impact of Economic Migration on Child Development:
Evidence from the Mexican Family Life Survey
By
Elizabeth T. Powers
Draft Date: November 30, 2009
Abstract: Data from the Mexican Family Life Survey are used to estimate the net impact of migration of a household member to the U.S. on the cognitive development of children remaining in Mexico. Single-equation estimates of a ‘value-added’ specification of early child development (ECD) suggest there may be adverse effects of migration on cognitive development when the parent migrates, but no effect when the migrant is a sibling. Effects tend to be larger in absolute magnitude for children who are younger and later-born. An instrumental variables strategy based on historical migration patterns is used to correct for potential biases due to omitted variables and endogeneity of migration with ECD. The IV coefficients of the migration effects are insignificant once the estimated standard errors are adjusted for clustering at the household level. The findings are robust with respect to a wide variety of specification changes.
Acknowledgments: This research was supported by a generous grant from the Inter-American Development Bank. The author benefitted greatly from the comments and suggestions of attendees at the first IDB discussion seminar on Improving Early Childhood Development in Latin American and the Caribbean. Seth Gitter very generously provided his instrumental variables data. I am grateful to Emilie Bagby for help with Spanish-English translation.
1
INTRODUCTION
Migration is a critical option for enhancing income for many families in Latin America
and the Caribbean (LAC). The World Bank reports that the LAC is the top remittance-receiving
region in the world, with remittances topping $48.3 billion in 2005 (Fajnzylber & Lόpez, 2007).
In 2004, over 50 percent of Haiti’s GDP was remittances, while remittances to Jamaica,
Honduras, El Salvador, Guatemala, Nicaragua, and the Dominican Republic all surpassed 10
percent of GDP (Fajnzylber & Lόpez, 2007).
Of all countries in the LAC, Mexico is the absolute leader in remittance volume, with a
total of $21.8 billion received (Fajnzylber & Lόpez, 2007).1 Migration to the U.S. for
employment is an important income source for the Mexican economy, and migration is
pervasive. According to Hanson & Woodruff (2003), the U.S. Mexican immigrant population in
the United States equaled nearly 8 percent of the total population of Mexico in 2000. Thus,
many Mexican households are directly and indirectly affected by the opportunity to migrate to
the U.S.
Economic migration may affect early child development (ECD) through several
mechanisms. By increasing household income, remittances are hypothesized to benefit ECD
through enhanced consumption, including of costly education services. Time investment
patterns and allocations among household members must change to accommodate migration.
This imposes additional constraints on choices influencing ECD, which may have a detrimental
effect on child outcomes. However, increased income from migration opportunities also allows
families to specialize in activities, possibly enhancing ECD. Finally, when an adult leaves the
1 Remittances from the U.S. are believed to have fallen precipitously with the recent global economic crisis.
2
household, it is possible that the identity of the de facto household decision maker changes in
ways that potentially benefit children.
Despite the widespread phenomenon of cycle migration in the LAC region and the
growing body of literature on ECD in the region, little is known about the effects of cycle
migration on ECD in the LAC region. Information enabling policies to be better tailored to the
fact of temporary economic migration is of immediate value. Understanding the extent to which
families and children are resilient in the face of major life changes such as migration is important
for better understanding child development and for crafting migration and family support
policies.
This project examines the impact of migration for economic opportunity to the United
States on the early cognitive development of Mexican children. Data from the Mexican Family
Life Survey (MXFLS) are employed to estimate the effect of sending a household member to the
U.S. on the cognitive development of children aged 5-12 who remain behind in Mexico.2
An immediate challenge to any analysis of this issue is that migration and child
investment decisions are jointly made. Endogeneity or simultaneity of the migration and
investment decisions, including the fact that migrant households and members self-select into
this status on the basis of both observed and unobserved characteristics, makes identification of
the causal effect of migration on child development difficult. The empirical strategy in this
paper exploits the unusually rich set of variables provided by the MXFLS and adds a distance-
based instrumental variables strategy. In theory the inclusion of parent cognitive scores goes a
long way towards addressing the problem of selection on unobservables, while an instrumental
2 The issues discussed in this paper also pertain to other areas of the globe. For example, Bryant (2002) estimates that 4.5 to 7.5 million children in the Philippines, Indonesia, and Thailand, and 10 to 20 percent of all children in the Philippines, have a parent working overseas.
3
variables strategy addresses both unobserved heterogeneity and endogeneity/ simultaneity
problems.
The empirical approach is as follows. I first estimate the basic value-added model of
child development (Todd and Wolpin, 2003), modeling a child’s current (wave 2) cognitive
score as a function of her prior (wave 1) score, other ‘family background’ variables (including
parent cognitive scores), and interim shocks and changes hypothesized to affect development.
The latter include the incidence of household members’ migration to the U.S. Inclusion of child
and parent cognitive scores in the specification controls for spurious correlations of the child’s
subsequent cognitive ability with household migration status due to selective migration. In
addition, because historical migration patterns are hypothesized to effect migration but not other
decisions, an instrumental variables strategy based on historical migration patterns further
identifies the causal effect of migration on cognitive ability. Ideally the IV strategy eliminates
bias from residual influences of unobservables (i.e., ‘residual’ once child and parent cognitive
scores are controlled) and identifies variation in migration that is exogenous with respect to
cognitive development.
To preview the findings, single-equation estimates indicate that migration often has a
significant effect on child development, with the level of significance and magnitude of the
effect depending upon the identity of the migrant, and the age, sex, and birth order of the child.
These preliminary findings indicate that children whose parents migrate to the U.S. between
waves 1 and 2 of the MXFLS fall significantly behind their peers in cognitive attainment for age.
However, when the migrant is a sibling, young brothers and sisters develop at the same average
rate as children who do not experience any migration from their household. These migration
effects are usually estimated to be insignificant in the entire sample of children, but are
4
significant and large for subsamples of children who are younger and later-born. To correct for
biases induced by unobserved heterogeneity and endogenous migration, instrumental variables
coefficients are also estimated. While the IV estimates are qualitatively similar to the single-
equation findings, the key coefficient estimates are never significant when errors are clustered at
the level of the household.
The paper proceeds as follows. The next section describes the present state of knowledge
about migration and ECD. A theoretical discussion follows that outlines a simple model in
which families make decisions about investments of adult time and purchased goods in children
(resulting in specific developmental attainments of children) and the economic migration of
adults. The key insights for empirical work are discussed. Next, a discussion of the data source
and a preliminary descriptive analysis are presented. After discussing the methodological
approach and empirical implementation, the main findings are presented. The final section
discusses the findings and draws conclusions.
Throughout the paper, the focus is on estimating the net impact of migration activity on
cognitive development, as measured by age-adjusted Raven scores. The particular channels by
which migration may affect ECD (e.g., increased remittances, reduced time spent with children)
are not identified.3
PRIOR LITERATURE ON MIGRATION AND ECD
Little research has been done on the topic of migration and ECD. Fajnzylber and Lόpez
(2007) review the literature and present their own original work on several topics (none of the
3 The analysis presented below can be readily extended to provide evidence on the impact of migration on other potential ECD measures and potential inputs to ECD such as household income, ‘quality’ time spent with children, and child care.
5
studies discussed treat migration as endogenous, with a single exception noted below). They
find that remittances positively affect anthropomorphic measures for Nicaraguan children and
provide some cross-country LAC evidence of increased school enrollment of 10-15-year olds
receiving remittances. Yang and Martinez (2006) find that greater remittances in the Philippines
increase school attendance and reduce child labor. However other research, also on the
Philippines, fails to find a beneficial effect of remittances on schooling (Bryant, 2002). Some
research for Ghana (Guzman, et al., 2007) suggests that migration affects child consumption
patterns in ways that may be beneficial for ECD.
Bryant (2002) draws conclusions from a survey of evidence from the Philippines,
Indonesia and Thailand.
[I]t appears that (i) migration of parents improves the material conditions of the
children left behind, which probably flows through to children’s health and
schooling, and (ii) the social costs are strongly mitigated by the involvement of
the extended family. In the Philippines, but less so in Indonesia and Thailand,
governmental and non-governmental organizations already provide a range of
services for children and migrants.
A handful of studies also provide evidence that children of migrants have better physical
abilities, no worse or better mental health, and are no more likely to engage in risky behavior as
older teenagers (Bryant, 2002).
Evidence on the school attainment of migrants’ children is mixed; while children of
migrants are more likely to attend private school, they are equally likely to be out of school and
there is little effect of migration on achievement as measured by grades (Bryant, 2002). Hanson
& Woodruff (2005), using an instrumental variables approach to address the potential
6
endogeneity of migration, estimate the effect of having any household-member migrant to the
U.S. on years of completed schooling of Mexican children ages 10-15. Children (particularly
girls) in migrant households where parents have low education levels complete significantly
more years of schooling. They argue that “the results are consistent with emigration helping
relax household credit constraints on the financing of education.”
There is some evidence that the extended family steps up its caregiving in response to
migration, as the hypothesis that migration is endogenous with child investment suggests.
Changes in caregiving and household arrangements have been documented in a number of
studies (see the citations in Bryant, 2002). Bryant (2002) also finds that “children of migrants
are more likely to have relatives from outside the nuclear family (i.e., cousin, aunt, uncle, or
grandparent) living in the same household, especially if both parents are overseas.”
Finally, there is little direct research on how migrants’ family roles moderate the effects
of migration on children. Surveys conducted in the Philippines indicate that children believe that
the migration of mothers would have a more detrimental impact on them (Bryant, 2002).
A THEORETICAL MODEL AND HYPOTHESES
The centerpiece of the basic model of human capital development is a human capital
production function that specifies the relationships of inputs to outputs (e.g., Behrman, Pollack,
Taubman, 1982). Families optimize with respect to consumption and human capital investments
in their child or children. The issue of migration can be analyzed in a straightforward way by
permitting adults to work outside the home. For simplicity’s sake, adult earnings are assumed to
be the sole income source. Note that the migration decision is not modeled discretely to avoid
complicating the analysis, nor are the financial and psychic costs of migration to the adults made
7
explicit. The latter, however, are readily handled by defining the wage to include earnings
opportunities abroad, net of migration costs.
Behrman (1998) lays out the human capital investment problem in the case of multiple
children with heterogeneous characteristics. Without loss of generality, suppose there are k=1,2
children and j=1,2 adults in the family. The family maximizes a welfare function whose
arguments are consumption (c) and the human capital attainment of the children (Hk), W(c, H1,
H2).4 For simplicity, there is no utility from leisure. A human capital production function for
each child is specified Hk=h(t1k, t2
k, ck;ak). The number of units of time investment of adult j in
child k is denoted tjk and ck is consumption of child k. The parameter vector ak summarizes key
characteristics of the child that affect its human capital production, including observed (e.g., age,
sex) and unobserved (‘teach-ability’) characteristics, as well as important characteristics of the
two adults (again, both observed and unobserved) that moderate the transformation of time and
consumption inputs into the realized human capital of child k. The time constraint for each adult
is tj1+ tj
2 + lj<=16, j=1,2, and the household budget constraint is c1+c2<=l1w1+l2w2.
For simplicity, assume the family utility function is separable in consumption. The
problem is to maximize V(H1, H2) subject to the technological, time, and budget constraints. The
first-order conditions from this optimization problem can be manipulated to reveal the
relationship, for each child k, between the substitutability of adult time in the production function
and relative wages, or
.
4 The possibility that migration itself affects the allocation of consumption to the child (e.g., through intrafamily bargaining) is not permitted.
8
For each child k, the optimal time investment contributions of heterogeneous adults are governed
by the substitutability of adult 1 and 2’s time investment in the human capital production
function (valued in utility terms), balanced against their relative wages. In the special case
where time contributions from adults 1 and 2 are perfect substitutes in the human capital
production function, the optimum requires specialization; the higher-wage adult makes no time
investment while the other adult provides all the time investment. In the special case where the
time investments of two adults are perfect complements, both adults contribute time investment,
regardless of their relative earnings power. Therefore, adult 1 is more likely to migrate to the
extent that he or she earns a sufficiently high wage abroad and/or to the extent that the time
investment of adult 2 is sufficiently substitutable for adult 1’s time investment.
For each parent, the distribution of his or her total time investment across the two
children is governed by
.
If the children are homogenous and parental preferences are “equal concern,” then each parent
divides his investment contribution equally among the children. In general, this will not be the
case. When children are heterogeneous, parents allocate their total contribution so as to equalize
the marginal benefit of an additional hour spent with the child. Heterogeneous child and adult
characteristics imply that an adult’s time investment is more productive for some children than
others. Thus, there could well be differential investment across the family’s children by the
same adult. Similarly, the relative value of time versus consumption investments in children
may vary, so that the consumption allocation is unequal across children.5
5 The relevant first-order condition is .
9
The model indicates that the ECD impacts of migration depend heavily on the family role
of the migrant. In turn, optimizing families take child and adult characteristics into account
when choosing who will migrate for economic opportunity. Migration may have little effect on
ECD if families can undertake compensatory adjustments, either by sending someone whose role
is not critical for ECD (e.g., an extended family member who does not normally reside with the
children) or if other household members are good substitutes for the migrant in the human capital
production functions. Families lacking a rich household roster of adults are predicted to be less
likely to choose economic migration, ceteris paribus. However, there are plausible
circumstances when families with sparse household rosters may also find it optimal to send an
economic migrant and reduce time investment. For instance, in very poor families with
extremely low consumption, the marginal value of an additional unit of consumption investment
may outweigh the developmental loss from reduced time investment in the child.
An obvious and important extension of this model is to multiple periods, as in Cunha and
Heckman (2007). In a dynamic model, human capital investment may be complementary over
time and ‘self-productive’ in the sense that increases in one period make investment more
productive in the next, giving rise to what Cunha and Heckman (2007) term “critical” and
“sensitive” periods of child development. An obvious implication of this extension is that the
timing and length of migration relative to the child’s developmental stage could greatly influence
ECD. In addition, in a dynamic model the relative substitutability of adults with respect to a
child could vary over time. However, the basic notion that having ‘substitute’ adults on the
household roster mitigates detrimental effects of economic migration on ECD, although made
more complex by consideration of the temporal dimension, continues to be a key hypothesis.
10
DATA AND DESCRIPTIVE ANALSYIS
Data from the first two waves of the Mexican Family Life Surveys (MXFLS) are used to
evaluate the impact of migration on early child development (ECD). The MXFLS is an ongoing,
longitudinal, nationally representative and comprehensive survey of Mexican households. Two
waves are currently available. The first wave consists of 8,440 households in 150 communities
surveyed in 2002. Follow-up interviews were conducted in 2006 and a third wave of surveying
is planned for 2010. The MXFLS contains detailed data on individuals and households,
including measures of cognitive development, remittances, and temporary migration experiences.
The MXFLS also provides information on the communities where the respondents live.
The MXFLS is an excellent resource for this study for several reasons. Flows of
temporary migration between Mexico and the U.S. are very large, so a substantial number of
households with workers abroad appear in the survey. The MXFLS has an excellent measure of
cognitive development, the Raven figure test. The identities of wave 1 household members who
reside in the U.S. in wave 2 are provided, as is complete information on family relationships that
can be used to infer the household role of the migrant.6
The major insight of the simple theoretical model is that the household roster is a major
determinant of migration decisions in the presence of children, as well as of migration’s potential
impact on ECD. Ceteris paribus, the presence of close-substitute caregivers on the household
roster increases both specialization in human capital investment among the adult household
members and the likelihood of migration of adult roster members with the best overseas earning
potentials. Thus, it is important to characterize the household roster of potential caregivers and
6 The analysis presented below can be broadened to exploit information contained in the MXFLS on remittance amounts, program participation, community characteristics, other child outcomes, and time investments in children.
11
workers in empirical work. Ideally this characterization extends beyond the current household
membership to include those whose time investments and earnings contributions are potentially
available to the household, rather than just observed, but this is impractical given typical data
limitations. Plausible initial assumptions are that men and women are less substitutable in the
child development function than adults of the same sex, that the human capital development
function may vary with the child’s age and sex, and that men in the middle of the domestic
earnings distribution have the greatest relative earning opportunities in the US.
In this section, the construction of the sample is explained and descriptive information is
presented on cognitive scores and sample characteristics according to migration activity.
Identifying Children with Migrating Relatives
The sample consists of children interviewed in wave 1 of the MXFLS who remain at
home in wave 2, but who have a wave 1 household member residing in the US in wave 2. A
wave 2 follow-up module tracks wave 1 household members who are in the U.S. in wave 2.
While this is the best way to identify all household members’ movements to the U.S., a
disadvantage is that families may benefit from non-household members’ migration. This sample
selection likely leads to understated estimates of the overall benefits of migration to ECD. On
the other hand, household member migration is arguably the type of migration likely to have the
greatest impact on the household’s children via time investment, so this type of migration is of
greater policy interest. A drawback is that remittance information is not provided.7
Using this module to identify those affected by migration to the U.S., there are 2,018
individuals in the wave 1 sample (of 34,674 total individuals) who are ‘left behind’ in Mexico by
a wave 1 household member who has migrated to the U.S. by wave 2. Of children with reported 7 There are at least two other sources of information on temporary migration in the MXFLS. It may be possible to combine information from several sources to gain a more detailed picture of migration.
12
wave 1 cognitive (Raven) scores ‘left behind’ by a U.S.-migrating family member, in 35 percent
of cases the migrating family member is their parent, their sibling in 67 percent of cases, and
another household member in 7 percent of cases. 11 percent of these children experience
migration by more than one type of household member.
Raven Scores
The indicator of cognitive progress used in this study is the Raven colored progressive
matrices instrument, designed to measure visual reasoning ability. test consists of a series of 18
figures that measure visual reasoning ability. A child is shown a series of two related two-figure
panels. They are then asked to select a figure so as to complete the third panel by composing the
pattern that is most consistent with the first two panels. For children age 5-12, this process is
repeated 18 times (I construct the score as the percentage of correct answers). A chief advantage
of the instrument is that it is designed to be “culture free;” it does not require knowledge of a
particular language or formal schooling on the part of the respondent. It is possible to study
some very young children (ages 5-8) in each wave, and it is also possible to track progress over
time for the same children between waves. An adult version of the test consists of 12 questions
and is administered to persons older than 12.
In the first wave, over 6,300 children are tested, along with over 19,800 adults. In the
second wave, over 5,500 children are tested, along with over 14,800 adults. Thus, it is also
possible to control for parents’ cognitive abilities when predicting children’s ECD. There are
474 sample children ages 5-12 in the ‘left behind’ group with reported wave 1 Raven scores.
246 of these children are tested with the ‘child’ test version in wave 2, while an additional 205
are re-tested with the adult version.
13
Figure 1 shows that Raven scores rise steadily with age, from a low of 49 percent correct
answers at the age-5 sample mean to 71 percent correct at age 12 (see the y-axis scale on the
right). The figure also shows (see the y-axis scale on the left) that the inter-wave correlation of
scores rises with age in wave 1, consistent with cognitive ability plateauing and coalescing at
older ages.8 At young ages, the inter-wave correlation is quite low, but it roughly triples by age
9. Differences in Raven scores by sex are modest, with boys slightly dominating girls
throughout the distribution, with the exception of the extreme left tail. There are no obvious sex
differences in the age pattern of inter-wave correlations (not shown).
Figures 2 and 3 present kernel density function plots of wave 2 Raven scores. In an
attempt to remove the strong trend of scores with age, each individual’s score is benchmarked
against the sample average score for the corresponding year of age. In particular, the two waves’
worth of Raven score observations are pooled and the mean determined for each calendar year of
age in the sample. Each observed Raven score is transformed by dividing through by the age-
specific sample mean for the appropriate in-sample calendar year of age. The transformed
‘relative’ scores are interpreted as “cognitive achievement for age.” Scores in excess of one
indicate above-average achievement.
Figure 2 presents age-adjusted Raven score distributions according to migration status.
The top panel shows the distribution for children ages 5-12 in wave 2, while the bottom panel
provides comparable estimates for the sample of adults from the same households. The
distribution of scores in the children’s samples appears to be shifted to the left for migrant-
sending households, indicating lower overall cognitive ability of children in households with a
migrant. The pattern for adults (Figure 2) is similar. Since one doesn’t expect adult’s cognitive 8 This analysis uses only observations of children who have taken the child version of the Raven figure test in wave 2.
14
development to be much affected by short-term migration, this suggests that the pattern for
children may simply be a product of adult self-selection for migration and heritability.
Figure 3 presents similar kernel density estimates, using further detail on the identity of
the migrant from the wave 1 household. The top panel contrasts the distributions of children in
migrant households whose sibling or parent migrates. Except at the extremes of the distribution,
the distribution of Raven scores for children appears less favorable when a parent migrates. In
the case of adults (bottom panel), the ‘parent’ group exhibits less density in the lower range of
scores. In contrast to Figure 2, the densities for children are not very similar to the adults.’ It is
not obvious that the pattern for children is simply due to selective migration.
Sample Characteristics by Migration Activity
Table 1 presents characteristics of migrants and non-migrants from wave 1 households
with children. Migrants tend to be much younger (nearly a decade on average), more often male,
and most often the child of the wave 1 household head, rather than the head themselves. They
are very unlikely to be the spouse of the head. Migrants are also less likely to be married and
tend to be better educated, having more often attended or graduated high school. Relatively
more migrants report working in the past 12 months, but migrants are no more likely to have
worked for pay than other sample members. As measured by the relative Raven score, migrants
have lower average cognitive ability than non-migrants.
Table 2 presents select characteristics of the households in which the sample children
reside, according to the household’s migration status. The categories examined are “household
has no migrant”, “household has any migrant,” and the subcategories of the latter group,
“household sends parent,” and “household sends sibling.” The latter three groups are not
mutually exclusive, as households may have multiple migrants.
15
Households that send a migrant to the U.S. are large and growing rapidly. Households
with a migrant are larger by around 1.4 members in wave 1 and gain around 0.40 additional
members (relative to other households) between waves 1 and 2. The largest migrant households
are those sending a sibling of a sample child, but households sending parents are also quite large.
Migrant households also tend to be well supplied with adults.9 Over 70 percent of migrant-
sending households have more than two adults in the household in wave 1. Despite the loss of a
migrant, households experiencing migration gain an entire adult member, on average. Thus, it
does not appear that migration leads to a ‘shortage’ of adult household members. Households
with migrants are also relatively ‘rich’ in male family members, which is expected given the
greater propensity of males to migrate. The ratio of male to female adults in wave 1 is 0.86 for
households with a migrant versus 0.64 for those without. In households sending a sibling, the
initial male-female ratio is nearly one. Finally, migrant households live in states with historically
high migration rates. The state-average 1950s migration rate for households sending a migrant is
just over 2.0, in contrast with a rate of just 1.7 for households without migrants.
METHODOLOGICAL APPROACH
The empirical strategy is to implement a “value-added” specification of child
development, augmenting this approach with parent Raven scores and instrumental variables in
order to correct for estimation biases due to unobserved selection on migration and the
simultaneous determination of migration and ECD investment choices.
9 Throughout the paper, the term ‘adult’ refers to individuals aged 15 or more.
16
Todd and Wolpin (2003) argue the value-added model is a relatively reasonable approach
to estimating ECD when choosing among imperfect alternatives. The basic specification is
, , , ,
Child i’s wave 2 relative Raven score (RKit) is modeled as a function of the wave 1 score
(RKi,t-τ) and other observed factors (reduced forms for investment and changes/shocks/events)
that influence the child’s development in the intervening period between two ability measures,
denoted Xi,t-τ. Migration of a household member in the intervening period is interpreted here as
one of these factors. The specification also indicates that intervening unobserved influences at
the child (a) and household (h) level may influence development.
An advantage of the ‘value added’ specification is that the impact of any systematic
unobserved differences between children in migrant and non-migrant households that occur up
through period ‘t-τ+j’ (including any time-invariant, ‘permanent’ household or child
heterogeneity) are reflected in RKt-τ. Remaining concerns about unobservables are thus limited
to non-permanent ‘shocks’ occurring in the intervening period between waves, a roughly three-
to-four year period.
The value-added specification directly addresses concerns about estimation bias due to
selection of migrant households with regard to child cognitive ability. However, other biases are
legitimately concerning, and these problems and potential solutions are now discussed in turn.
A practical concern of implementation of the value-added model is that children’s Raven
scores may be quite noisy. In particular, the earliest observed score, which plays the crucial role
in controlling for selection, can be a quite noisy indicator of the true score for very young
children. Therefore, the initial distribution of child cognitive ability with respect to migration
status may not be well characterized, and this problem may worsen when examining younger
17
subsamples. Maternal Raven scores have been shown to be better predictors of children’s later
cognitive attainment than the child’s own early scores, so a straightforward remedy is to include
parent Raven scores as ‘state’ variables in addition to RKt-τ.10
Many variables that should plausibly be included in ‘X’, such as family structure
changes, are endogenously determined with migration, leading to potentially inconsistent
estimates of all the parameters. In addition, other standard ‘inputs’ to cognitive development,
such as schooling, are also likely endogenous. Using instruments for migration that are
uncorrelated with the Xs, do not directly influence ECD, but that directly influence migration,
yields a consistent coefficient estimate for migration. An appropriate IV strategy also addresses
the problem of ‘transitory’ unobserved influences on ECD.
Inlcuding migration as an ‘X’ variable is an ad hoc extension of the value added model
because, as the theoretical discussion indicates, migration, child investments, and ECD are all
jointly determined. The theoretical model indicates that factors that directly influence ECD may
also influence migration. It may therefore be difficult to identify the migration coefficient when
these factors are also included along with migration in an ECD specification. An IV strategy
addresses the potential identification problem that arises from extending the value-added model
to encompass migration. An exogenous shifter of migration aids in the identification of its
coefficient.
10According to Cunha, Lochner, and Masterov (2005), measures like the Raven score, which they would characterize as “pure cognitive ability” do not predict adult IQ well. They argue that prior to age 5, maternal IQ is a better predictor of age-15 IQ than any available test score and that after age 10, “IQ becomes stable within the constraints of psychometric measurement error.” The scores of the younger children taking the Raven test in this sample could be subject to this problem.
18
Finally, a potentially important problem remains. Migration may be correlated over time,
so that families experiencing migration in the past are more likely to have migrants in the
present. If so, causality may run from the ‘current’ migration variable to the initial (‘baseline’)
child cognitive score. Such a relationship makes it difficult to identify the separate influences of
initial cognitive ability and intervening household migration status on current cognitive
development. It is not evident that the particular IV strategy taken in this paper can successfully
address this problem. Because the instruments are based on historical migration patterns, the
instrumented migration variable may well be correlated with baseline ECD, since baseline ECD
has presumably been influenced by past migration. Under the assumption that migration during
the child’s lifetime does not affect adults’ cognitive development, an alternative approach is to
replace the child’s initial Raven score with those of his parents.
IV Strategy
The IV strategy is developed in Woodruff and Zenteno (2001) and also implemented in
Gitter, Gitter, and Southgate (2008). Historical migration rates from the 1950s, by edo (state)
serve as instruments for current migration. Following Gitter, Gitter, and Southgate (2008), the
migration rates are also interacted with the region of Mexico. (These region dummies are always
included in the main ECD equation). The intuition behind the IV is that early migration patterns
were established by geographic barriers, transportation advances established and located prior to
the phenomenon of widespread migration (railroads), and the Federal Bracero program that
brought large numbers of Mexican migrants to the U.S. for the first time. The underlying
hypothesis is that migration networks (knowledge about and practical help with migration)
sprang up in communities that were initially advantaged in migration, and that this explains why
contemporary patterns of migration still strongly mimic earlier historical patterns.
19
FINDINGS
Single-Equation Estimates
Single-equation estimates of the value-added specification for children’s Raven scores
are presented in Tables 3, 4, and 5. Tables 3 and 4 present findings on how the key coefficients
evolve as sets of explanators are sequentially added to the specification.
Table 3 presents estimates for the entire sample of children (i.e., 5-12-year-olds in wave
1). All specifications include two dummy variables indicating migration status—whether there
is any migrant from the household and whether a sibling is a migrant from the household.
Column 1 presents the findings from the value-added specification without any additional
explanatory variables aside from the child’s wave 1 Raven score. The effect of migration is
negative at a confidence level exceeding 95 percent, but the estimated effect of sibling migration
is insignificant. As additional explanators are added to the model, the effect of having a migrant
from the household diminishes and becomes insignificant. The effect of a sibling migrant is
insignificant at standard confidence levels for every specification. The child’s Wave 1 Raven
score is always highly significant (at confidence levels exceeding 99 percent). Its effect tends to
decline as additional explanators are introduced to the model. The estimated magnitude of its
coefficient drops by almost one-third when moving from the least to most ‘saturated’
specifications. The (unadjusted) R-squared doubles as explanators are added. The only groups
of explanators that are not found to be jointly insignificantly different from zero at standard
levels of confidence are detailed family structure variables and interim shock and change
variables (i.e., variables in ‘X’).
The blocks of explanatory variables are as follows. “Dwelling characteristics” originally
included indicators of an indoor toilet, tap water access, and whether the home was paid for. The
20
other variables turn out to be redundant with the presence of an indoor toilet, so only the latter is
included as a dwelling characteristic. The effect of an indoor toilet on ECD is positive.11
The set of child characteristics includes whether an indigenous language is spoken at
home, the grade of school currently attended, the birth order of the child (expressed as dummy
variables indicating first born, second born, etc… up to fifth or higher-born), twin status, only-
child status, dummy variables indicating the presence of one, two, or three or more younger
siblings, whether the mother is reported to be the ‘caregiver’, sex, and a full set of dummies for
child age.12 Cognitive development is significantly slower for children who speak an indigenous
language, the later-born, and those with a greater number of younger siblings. The findings
suggest a declining effect of age on later achievement, although several individual age dummies
have large standard errors.13
Maternal and paternal characteristics consist of age, education, and work status variables,
and parent Raven scores. Parental factors that significantly increase ECD are parent Raven
scores and higher maternal and paternal educational attainment. Parental factors diminishing
ECD are maternal age and whether the father worked in the past year.
Family structure variables consist of the ratio of males to females, the number of
individuals in the household, the number of adults in the household, and a dummy variable
indicating the presence of three or more adults. None of these variables has a significant effect
and the variables are insignificant as a group. Geographic information includes the region of the
11 Throughout this discussion findings are reported as significant from the ‘saturated’ specification in column (7), which excludes potentially endogenous variables.
12 In practice, the variable reporting ‘mother is caregiver’ appears to capture whether the father is involved with the family.
13 Since the Raven scores are age-normed, this finding suggests that the likelihood of improving cognition for age is declining with age, consistent with cognitive ability coalescing in later childhood.
21
country and the size of the municipality. Central location has a negative effect on ECD, but
these variables are also insignificant as a group.
Table 4 presents the same specifications, but restricts the sample to young children (those
ages 5-8 in wave 1). In contrast to the first set of findings, the effect of sending a migrant is
estimated to be negative at standard confidence levels across the first seven specifications, and
the coefficient of the sibling migration variable is typically significantly positive. The findings
consistently indicate that non-sibling migration has a negative effect on ECD, while sibling
migration has no effect. The effect of the prior Raven score on the current Raven score is
smaller for the younger sample, as one might expect (i.e., the scores of younger children have
less predictive value), and the coefficient of the Raven score declines by about 40 percent from
the first to the seventh specifications. The effects of other variables are as described above.
To place the magnitude of these changes in perspective, the mean inter-wave change in
relative Raven scores for the young child sample is 0.106. According to the preferred
specification (7), sending a parent to migrate reduces the wave 2 Raven score by almost 0.058,
implying that more than half of expected average cognitive gains are lost.
Table 5 presents single-equation findings for other subsamples and specifications,
including interacting the migration variables with the sex of the child. Columns (1) and (2)
repeat the findings in tables 3 and 4 (column 7) for reference purposes. In column 3, the
specification is extended to include interactions of migration status with child sex for the sample
of young children. The F-tests reported below the coefficient estimates indicate that there are no
significant migration effects for boys or girls, but this is because all the migration coefficients are
now estimated very imprecisely.
22
Another characteristic of children that may plausibly influence the impact of migration is
birth order. There is evidence from the U.S. that first-born children receive more parental time
investment (e.g., see Price, 2008). Whatever mechanism produces this outcome may also be
protective against the adverse consequences of migration for first-borns. Column (4) applies the
preferred specification to the subsample of those children who are not first-born. Without
additional age restrictions, the effects of migration have large estimated standard errors and are
insignificant. However, when the sample is restricted to young children who are later-born
(column 5), migration by family members other than the sibling is estimated to have a negative
effect due to estimated coefficients for the age-restricted subsample. Finally, sex is interacted
with migration variables for the subsample of younger, later-born children. In this case, F-tests
indicate that there is a negative effect of migration on later-born boys unless the migrant is a
sibling. While migration effects are estimated to be insignificantly different from zero for late-
born girls, the coefficient values are comparable to the young child sample, but the standard
errors are large.
Instrumental Variables Estimates
The appendix table presents first-stage regression coefficient estimates for the
instrumental variables, along with F-tests for the joint exclusion of the instruments from the first-
stage specification. All the other model variables are also included in these specification (i.e.,
dwelling characteristics, child and parent characteristics, detailed family structure, and
geographic information), but their coefficients are not reported. For the entire sample of
children, variables that increase the chance that a household sends any migrant to the U.S. are a
toilet in the house, higher birth order, maternal age, the ratio of adult males to females, and living
in a non-rural area. Factors that reduce the probability of sending any migrant are mother is the
23
caregiver, mother worked in the past year, and father’s age. Neither cognitive scores for children
nor parents are significant predictors of migration from the household, nor is parents’ education.
The findings reported in the appendix table indicate that the strength of the instrument
set varies widely across subsamples. The instruments are strongest (as indicated by an F-test for
their joint significant in the first stage) for the entire sample and the sample of later-born
children. The instruments’ predictive power is substantially weaker when the sample is
restricted by age.
Table 6 presents the IV estimates. The scale has changed because the instrumental
variables model is estimated as a linear probability model in the first stage. Therefore, the right-
hand-side migration variables explaining ECD are now migration probabilities, not binary
outcomes. In all cases, standard errors of the IV coefficients are quite large when errors are
clustered at the household level, and none of the effects is estimated to be significantly different
from zero. (The migration coefficients are significant at standard confidence levels in the
absence of clustering).
Robustness of the Findings
As mentioned in the methodology discussion, child Raven scores may be subject to error,
particularly at young ages. Therefore, the robustness of the main findings is explored with
respect to dropping the child Raven score and letting parent Raven scores proxy for the child’s
‘baseline’ cognitive ability. Single-equation estimates yield significant migration coefficients
(with a negative effect of migration and no effect if the sibling is the migrant, as usual) for the
young child subsample and the subsample of later-born younger children. IV estimates continue
to be insignificant when errors are clustered.
24
An additional consideration is the interim ‘X’ variables, which have been excluded from
the main analysis to this point. While these variables may obviously be endogenous with
migration and their inclusion problematic, their exclusion may also introduce problematic biases.
Arguably, the instrumental variables strategy helps to identify the migration coefficients,
although other coefficients may be inconsistently estimated. Inclusion of these variables in the
IV estimation results in coefficients for the migration variables that are larger in absolute
magnitude but still insignificant at standard confidence levels.
Finally, the inclusion of interim ‘shock’ or ‘change’ variables may be problematic, since
these are doubtless endogenously determined with migration. However, the IV strategy may
permit unbiased estimation of the migration coefficients. Column 7 of Tables 3 and 4 indicate
that this group of variables (which consists of indicators for death of a household member,
serious illness of a household member, unemployment of a household member, experience of a
natural disaster or crop failure, an increase in the number of younger siblings, the change in total
household size, the change in the number of household adults, and a decrease in the number of
older siblings in residence) is jointly insignificant in single-equation specifications. Their
inclusion in the model has little effect on the IV estimates.
CONCLUSIONS
Data from the MXFLS are used to estimate the net impact of migration of household
member(s) to the US on the cognitive development of children remaining in Mexico. A simple
theoretical model indicates that migration and ECD are jointly determined. Optimizing families
with children are generally predicted to send household members to the U.S. who have good
potential relative earnings gains and who are either not influential for child development or who
25
are good substitutes for remaining family members in the household roster in the ECD
production function. Descriptive statistics and first-stage estimation of migration indicate that
households sending migrants are large, relatively more endowed in males and adults, and that
children higher in the birth order are more likely to experience migration from their families.
Single-equation estimates of children’s cognitive development indicate that the identity
of the migrant matters for ECD, as do the characteristics of the child. Generally, where
significant effects of migration are found, children experiencing migration from the household
have lower cognitive gains over the interwave period. However, when it is a sibling who
migrates, there is no difference in developmental gains according to migration status. Adverse
effects of non-sibling migration are largest for younger and later-born children. Further analysis
does not provide strong evidence for differences in the effect of migration according to the sex of
the child.
The value-added specification controls for the distribution of ECD in a way that arguably
addresses the issue of selective migration based on ‘permanent’ child and household
characteristics that also influence ECD. Inclusion of parent Raven scores also controls for the
unreliability of very young children’s scores and further controls for selective migration, since
parent cognitive ability is presumably determined prior to migration from the household.
However, potentially important problems of selection on intervening characteristics, reverse
causality from ECD to migration, endogeneity with intervening observed variables with
migration, and potential endogeneity of the initial measure of ECD with migration remain.
To that end, an instrumental variables strategy that uses historical migration patterns
interacted with region is applied in a two-stage-least squares framework. The IVs are good
predictors of migration from the household so long as the sample is not restricted to the youngest
26
children. The IV findings are qualitatively similar to the single-equation findings, in that the
estimated coefficient of ‘any migrant’ has a negative sign, and the effect of a sibling migrant is
of opposite sign and of roughly the same absolute magnitude. However, the significance levels
of the IV estimates are not robust with respect to clustering of errors at the household level.
The findings in this paper suggest several directions for additional research. The analysis
can be extended to consider other child development indicators, such as school progress, or
physical health. Further investigation of the exact mechanisms by which migration may effect
ECD (e.g., remittances, time use, consumption allocated to children) appears merited based on
the evidence presented here.
27
REFERENCES
Behrman, Jere R., "Intrahousehold Distribution and the Family," 1998. Chapter 4 in Volume 1A, Handbook of Population and Family Economics, Rosenzweig and Stark, editors.
Behrman, Jere R., Robert A. Pollack, and Paul Taubman, 1982. "Parental Preferences and Provision for Progeny," Journal of Political Economy 90(1, Feb.), 52-73.
Bryant, John, 2002. Children of International Migrants in Indonesia, Thailand and the Philippines: A Review of Evidence and Policies. Cunha, Flavio and Heckman, James, 2007. “The Technology of Skill Formation.” American Economic Review 97(2,May), pp. 31‐47. Cunha. F.. J. Heckman. L. Lochner, D. Masterov. 2005. "Interpreting the Evidence on Life Cycle Skill Formation." NBER Working Paper 11331. Cambridge. Mass Fajnzylber, Pablo, and J. Humberto Lόpez, 2007. “Close to Home: The Development Impact of Remittances in Latin America.” The International Bank for Reconstruction and Development/ The World Bank. Washington, DC. 89 pp. Gitter, Seth R., Gitter, Robert J., and Southgate, Douglas, 2008. “The Impact of Return Migration to Mexico.” Estudios Economicos 23(1, enero-junio), pp. 3-23. Guzmán, Juan Carlos, Andrew R. Morrison and Mirja Sjöblom, 2007. “The impact of remittances and gender on household expenditure patterns : evidence from Ghana ” in The international migration of women, Andrew R. Morrison, Maurice Schiff and Mirja Sjöblom, editors. Washington, DC : World Bank. Hanson, G. H. and C. Woodruff (2003): “Emigration and Educational Attainment in Mexico” mimeo, University of California, San Diego, California. Price, Joseph, 2008. “Parent-Child Quality Time: Does Birth Order Matter?” Journal of Human Resources 43(1): 240-265.
Todd, Petra E., and Kenneth I. Wolpin. 2003. “On the Specification and Estimation of the Production Function for Cognitive Achievement.” Economic Journal (113, February): pp. F3-33.
Woodruff, C. and R. Zenteno. 2001. Remittances and Microenterprises in Mexico, University of California San Diego (mimeo). Yang, Dean, and Claudia A. Martínez. 2006. “Remittances and Poverty in Migrants’ Home Areas: Evidence from the Philippines.”In International Migration, Remittances, and the Brain Drain, ed. Çag˘lar Özden and Maurice Schiff, 81–121. New York: Palgrave Macmillan.
28
Figure 1: Average Raven scores and variation by age
Source: Author’s computations from the MXFLS data.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
5 7 9 11
Age
Corr(W1,W2) scores
mean W1 Score
29
Figure 2a: W2 Raven scores of children, by household ‘sending’ status
Figure 2b: W2 Raven scores of adults, by household ‘sending’ status
0.5
11.
5D
ensi
ty
0 .5 1 1.5 2Wave 2 Raven score
Non-migrant householdMigrant household
Author's computations from 4,911 and 451 observations of children in the MXFLS.
Kernel density estimate0
.2.4
.6.8
1D
ensi
ty
0 1 2 3Wave 2 Raven score
Nonmigrant householdMigrant household
Author's computations from 12,211 and 900 adults in the MXFLS.
Kernel density estimate
30
Figures 3a: W2 Raven scores of children, by detailed household ‘sending’ status
Figure 3b: W2 Raven scores of adults, by detailed household ‘sending’ status
0.5
11.
5D
ensi
ty
0 .5 1 1.5 2rRaven_W2
Sibling migrant householdParent migrant household
Author's computations from 304 and 149 observations of children in the MXFLS.
Kernel density estimate0
.2.4
.6.8
1D
ensi
ty
0 .5 1 1.5 2rRaven_W2
Sibling migrant householdParent migrant household
Author's computations from 411 and 161 observations of adults in MXFLS.
Kernel density estimate
31
Table 1: Characteristics of adult migrants and non-migrants Non-migrant adults Migrant adults Age 28544
32.23 (20.20)
747 23.11
(12.20) Male 28593
0.474 (0.499)
748 0.607
(0.489) Child of head 28593
0.404 (0.491)
748 0.644
(0.479) Household head 28593
0.290 (0.454)
748 0.184
(0.388) Spouse of head 28593
0.216 (0.412)
748 0.064
(0.245) Married 28593
0.491 (0.500)
748 0.330
(0.471) No education 28593
0.108 (0.310)
748 0.040
(0.196) Elementary education 28593
0.349 (0.477)
748 0.362
(0.481) Secondary education 28593
0.251 (0.434)
748 0.3
(0.484) High school education 28593
0.109 (0.311)
748 0.131
(0.338) Worked in past 12 months 28593
0.442 (0.497)
748 0.508
(0.500) Any income in past 12 months 28593
0.360 (0.480)
748 0.382
(0.486) Raven score, W1 19242
0.956 (0.475)
580 0.878
(0.444) Notes: Each entry is number of observations with mean and standard deviation (in parentheses).
32
Table 2: Characteristics of ‘sending’ households
(1) (2) (3) (4) Non-sending
households
Sending householdsHousehold
sends sibling Household
sends parent Total household members, W1
3465 5.28
(1.79)
246 6.70 (2.11)
158 6.91
( 2.18)
90 6.52
(2.05)
Change in total household members, W1 to W2
3465 -0.046 (1.70)
246 0.434
(0.931)
158 0.361
(0.966)
90 0.322
(0.934)
Total number of adults, W1
3465 2.77
(1.28)
246 3.57
(1.50)
158 3.68
(1.54)
90 3.39
(1.50)
More than 2 adults in HH, W1
3465 0.440
(0.497)
246 0.720
(0.450)
158 0.760
(0.423)
90 0.633
(0.484)
Change in number adults, W1 to W2
3465 0.454
( 1.270)
246 1.08
(0.920)
158 1.18
(0.873)
90 0.900
(0.900)
Ratio of male to female adult members, W1
3413 0.638
(0.375)
246 0.839
(0.495)
158 0.900
(0.522)
87 0.810
(0.446)
Historical state migration rate
3463 1.51
(1.62)
246 2.12
(1.97)
158 2.21
(2.02)
90 1.95
(1.87) Notes: Sample in column (1) consists of households that report a children’s relative Raven score for some member in wave 1. Columns (2)-(4) are further restricted as indicated.
33
Table 3: OLS Estimates of children’s Raven scores
(1) (2) (3) (4) (5) (6) (7) (8) Sending household
-.067** (0.028)
-.064** (0.027)
-.064** (0.028)
-0.060** (0.028)
-0.054*** (0.028)
-.050*** (0.028)
-0.041 (0.028)
-0.039 (0.030)
Sibling is the migrant
0.027 (0.033)
0.031 (0.033)
0.054 (0.034)
0.062*** (0.034)
0.056*** (0.034)
0.054 (0.034)
0.047 (0.034)
0.040 (0.035)
Wave 1 Raven score
0.275* (0.015)
.254* (0.015)
0.235* (0.015)
0.201* (0.015)
0.193* (0.015)
0.193* (0.015)
0.189* (0.016)
0.187* (0.016)
Dwelling characteristics
YES YES YES YES YES YES YES
Child Characteristics
YES YES YES YES YES YES
Maternal characteristics
YES YES YES YES YES
Paternal characteristics
YES YES YES YES
Detailed family structure
YES* YES* YES*
Geographic information
YES* YES*
Shocks & change variables
YES*
R-squared 0.0739 0.0848 0.1024 0.1203 0.1276 0.1294 0.1331 0.1352
Observations 5362 5362 5347 5347 5347 5294 5294 5189
Cases with migration
451 451 450 450 450 448 448 435
Notes: Coefficient estimates are reported with standard errors in parentheses beneath. (*,**,**) indicates significance at the (99th, 95th, 90th) confidence level, respectively. Errors are clustered at the household level. See text for description of variables. All specifications include a constant. Samples consist of children with a child Raven score in wave 1 and a Raven score (child or adult test version) in wave 2. *The group of variables is insignificantly different from zero at standard confidence levels based on F-Test statistic.
34
Table 4: OLS Estimates of young children’s Raven scores (1) (2) (3) (4) (5) (6) (7) (8) Sending household
-0.80* (0.029)
-0.079* (0.028)
-0.080* (0.028)
-0.073** (0.029)
-0.067** (0.028)
-0.066** (0.029)
-0.058** (0.029)
-0.050*** (0.030)
Sibling is migrant
0.061 (0.038)
0.066*** (0.038)
0.089** (0.038)
0.099** (0.038)
0.090** (0.038)
0.089** (0.038)
0.086** (0.038)
0.071*** (0.039)
Initial Raven score
0.177* (0.015)
0.160* (0.015)
0.147* (0.015)
0.117* (0.015)
0.111* (0.015)
0.108* (0.015)
0.104* (0.015)
0.105* (0.015)
Dwelling characteristics
YES YES YES YES YES YES YES
Other child characteristics
YES YES YES YES YES YES
Maternal characteristics
YES YES YES YES YES
Paternal characteristics
YES YES YES YES
Family structure
YES* YES* YES*
Geographic controls
YES* YES*
Shocks and interim changes
YES*
R-squared 0.0529 0.0677 0.1006 0.1275 0.1397 0.1420 0.1456 0. 1465 Observations 3130 3130 3126 3126 3126 3105 3105 3063 Cases with migration
246 246 246 246 246 245 245 235
Notes: Coefficient estimates are reported with standard errors in parentheses beneath. (*,**,**) indicates significance at the (99th, 95th, 90th) confidence level, respectively. Errors are clustered at the household level. See text for description of variables. All specifications include a constant. Samples consist of children with a child Raven score in wave 1 and a child Raven score in wave 2. *The group of variables is insignificantly different from zero at standard confidence levels based on F-Test statistic.
35
Table 5: Single-equation Estimates of children’s Raven scores, subsamples and interactions
(1)
(2)
(3)
(4) (5) (6)
All children
Young children
Young children
Later born
children
Later born young
children
Later born young
children Sending household -0.043
(0.029) -0.058** (0.029)
-0.060 (0.388)
-0.063*** (0.037)
-0.076** (0.036)
-0.054 (0.049)
Sending household x male
0.050 (0.054)
-0.045 (0.062)
Sibling is migrant 0.048
(0.034) 0.086** (0.038)
0.086*** (0.049)
0.066 (0.042)
0.109** (0.044)
0.087 (0.057)
Sibling migrant x male
-0.001 (0.068)
0.046 (0.075)
W1 Raven score 0.190*
(0.016) 0.104* (0.015)
0.104* (0.015)
0.175* (0.018)
0.100* (0.018)
0.100 (0.018)
H0: Any migrant + male x any migrant = 0
F(1, 2346) = 1.92
Prob > F = 0.1655
F(1, 1765) = 4.77
Prob > F = 0.0290
H0: Sibling migrant + Sibling migrant x male =0
F(1, 2346) = 2.54 Prob > F
= 0.1111
F(1, 1765) = 5.08
Prob > F = 0.0244
Mean interwave change in relative Raven score
0.078 (0.395)
0.106 (0.397)
0.106 (0.397)
0.082 (0.4023)
0.110 (0.400)
0.110 (0.400)
R-squared 0.1323 0.1450 0.1450 0.1294 0.1508 0.1510 Observations 5294 3105 3105 3739 2254 2254 Notes: Coefficient estimates are reported with standard errors in parentheses beneath. (*,**,**) indicates significance at the (99th, 95th, 90th) confidence level, respectively. Errors are clustered at the household level. All specifications include a constant, child’s initial Raven score, dwelling characteristics, other child characteristics, maternal and paternal characteristics, family structure variables, and geographic controls.
36
Table 6: Instrumental Variables Estimates of children’s Raven scores, subsamples and interactions
(1) (2) (3) (4) All
children Young
children Later born children
Later born young children
Sending household
-1.71 (1.37)
-1.72 (1.06)
-2.59 (2.18)
-1.48 (1.08)
Sibling is migrant 1.61 (1.75)
1.88 (1.41)
2.66 (2.62)
1.59 (1.34)
W1 Raven score 0.195* (0.022)
0.111* (0.024)
0.190* (0.034)
0.106* (0.026)
Observations 5294 3105 3738 2254 Notes: Coefficient estimates are reported with standard errors in parentheses beneath. (*,**,**) indicates significance at the (99th, 95th, 90th) confidence level, respectively. Errors are clustered at the household level. All specifications include a constant, child’s initial Raven score, dwelling characteristics, other child characteristics, maternal and paternal characteristics, family structure variables, and geographic controls.
37
APPENDIX TABLE: First-stage estimates of household ‘sending’ status
(1)
(2)
(3)
(4)
(5) (6)
(7) (8)
All children Young children Later born children Later born young children
Any migrant
Sibling migrant
Any migrant
Sibling migrant
Any migrant
Sibling migrant
Any migrant
Sibling migrant
1950 migration rate
0.030* (0.004)
0.024* (0.003)
0.097 (0.076)
0.005 (0.059)
0.038* (0.005)
0.032* (0.004)
0.035* (0.006)
0.028* (0.005)
1950 migration rate x Border state
-0.022* (0.008)
-0.025* (0.007)
0.104 (0.076)
-0.009 (0.060)
-0.035* (0.011)
-0.036* (0.010)
-0.031** (0.014)
-0.035* (0.012)
1950 migration rate x North state
omitted omitted 0.125*** (0.076)
omitted omitted omitted omitted omitted
1950 migration rate x Center state
-0.010** (0.005)
-0.008*** (0.004)
0.116 (0.758)
0.016 (0.059)
-0.015** (0.007)
-0.014** (0.006)
-0.013 (0.008)
-0.012*** (0.007)
1950 migration rate x Capital
-0.016* (0.062)
-0.080 (0.051)
omitted omitted -0.199** (0.089)
-0.110 (0.078)
-0.160 (0.107)
-0.008 (0.089)
F-statistic for IVs
F(4, 5244)
= 24.76
F(4, 5244) = 21.67
F(4, 3055) = 13.09
F(4, 3055) = 11.82
F(4, 3691) = 22.35
F(4, 3691) = 19.40
F(4, 2206) = 12.20
F(4, 2206) = 10.94
P-value for F-statistic (Prob > F = )
0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Adjusted R-squared
0.1106 0.1328 0.1063 0.1486 0.1151 0.1287 0.1120 0.1490
Observations 5294 5294 3105 3105 3739 3739 2254 2254 Notes: Coefficient estimates are reported with standard errors in parentheses beneath. (*,**,**) indicates significance at the (99th, 95th, 90th) confidence level, respectively. Errors are clustered at the household level. All specifications include a constant, child’s initial Raven score, dwelling characteristics, other child characteristics, maternal and paternal characteristics, family structure variables, and geographic controls.