. The Effect of Remittances and Migration on Human Capital: Evidence from Mexico Alfredo Cuecuecha ITAM February 2008 Abstract.- In this paper we disentangle the effects of migration and remittances on human capital using a large data base which includes information on both the reception of remittances and the existence of migrants in the household. We are able to separate out the positive income effect of remittances from the allocation effect that migration has on education. The allocation effect can be positive or negative, depending on gender and time since the last return of the migrant. For migrants that left less than five years ago, the allocation effect is positive for males and negative for females. The allocation effect is negative and dominates the income effect for the children and teenagers from migrants that left more than five years ago. These results are based on using instrumental variables, that exploit differences between municipalities in the extent of migrant networks, as well as differences between municipalities in the importance of remittances as percentage of income. JEL Classification: F22, D91, J61,D84, O15, O24 Keywords: Remittances, Migration, Human Capital. Remmitances. Cuecuecha. February 2008 1 2 Introduction In recent years the amount of remittances of migrants to their native countries have increased substantially to represent very large amounts: in Mexico they reach a total of $20 billion during 2005, 3% of total GDP, while in El Salvador the importance is even bigger as remittances reach as much as 25% of GDP. In this paper we focus on determining how remittances and migration affect human capital accumulation. On the one hand, remittances imply faster human capital accumulation, as they increase the family’s income; that is the income effect. On the other hand, migration generates changes in the household that can have negative (positive) effects on education. First, parents migrating to other countries to remit money to their household leave their children behind at home, reducing thereby the time children devote to education (Hanson and Woodruff, 2003).1 Second, the household can reallocate household chores following market incentives. For example, if the labor market pays more for males, the household can increase education efforts on males and specialize females in household chores. Third, the household can increase labor market participation of the children if the migrant sent away does not send money back home. We refer to all these changes as the allocation effect. This is the first paper to disentangle allocation and income effects. Using a large representative data set that contains detailed information on the exposure of households to migration and remittances, we show that the allocation effect is negative for females and positive for males. We also find that the income effect is positive for all individuals. The combined effect of migration and remittances is positive for all individuals, which implies that for females the income effect dominates. We also show that the allocation effect overtakes the income effect when migrants left for more than five years. Studies so far have estimated only the combined effect, but have not disentangled the income and the supervision effect. Hanson and Woodruf (2003) proposed the 1 This is called in the literature the supervision effect. Remmitances. Cuecuecha. February 2008 3 concept of supervision effect for the fact that children of migrating parents that receive less attention from their parents devote less time to education, increase leisure time and time devoted to household chores and even labor force participation. Using data from the 2000 Mexican Census they find that for female head of household remittances increase the education of girls, but not of boys. As an instrument for remittances they use historical data on migration patterns from Mexico to the United States at the state level. On the contrary, Borraz (2005), using the same data and instruments, but focusing on communities with less than 2500 inhabitants, finds positive effects for both boys and girls. He also includes the geographical distance from the municipalities where individuals reside to the US as an instrument. The literature on the effects of remittances on human capital also contains studies that have used aggregate data or non-nationally representative data. Lopez-Cordoba (2006) uses aggregate data on municipalities and finds that among Mexican rural municipalities, remittances generate higher human development indicators. In particular, he finds that infant mortality, child illiteracy and some poverty measures tend to reduce. His analysis controls for potential endogeneity of the remittance variable using as an instrument rainfall patterns at the municipal level. Mckenzie and Rapoport (2005) use data from the Mexican Migration Project, which is primarily rural and it is not representative of the entire country. They find a relation between inequality in education and migration. They argue that remittances first increase inequality as only well off households have access to migration, but that as poorer households get access the situation reverses. However, they find a negative effect of migration and remittances on 16 to 18 year old individuals.2 The remainder of the paper is organized as follows: the second section presents 2 The paper is also related to a larger literature that has look for the effects of remittances in other indicators of development or reduction of poverty. Some authors have analyzed the relation between remittances and migration with entrepreneurial activities (Durand, Parrado and Massey, 1996; Lindstrom, 1997; Dustman and Kirchkamp, 2002). Esquivel (2004) analyzes the effect of remittances on poverty. Rodriguez and Cox (2005) analyze how remittances relate to labor force outcomes. Mckenzie and Rapoport (2004 ) analyze how remittances are related to wealth. Others have studied how remittances relate to spending patterns (Zarate-Hoyos,2004; Adams and Cuecuecha, 2007). Remmitances. Cuecuecha. February 2008 4 the data, the third section presents our empirical strategy, which consists of using as instruments the number of migrants in the municipality where the child lives, excluding the family of analysis, and the amount of remittances received in the municipality where the child lives, excluding the family of analysis. The empirical section also presents an analysis based on random matching estimators, and a non-parametric analysis that demonstrates the existence of selection in unobservable components. The fourth section concludes. 2 Data The primary source of data for this paper is the 9.1% sample of the 2000 Mexican census. Our sample consists of children and young adults between the ages 8 to 19 years old that belonged to households for which the information on remittances, migration, education of the children, and education of the head of household was not missing. We restrict attention to children from the head of household. After applying all our selection criteria, we have approximately 720 thousand individuals, 15 to 19 years old. We also have 1.3 million individuals 8 to 14 years old.3 The census asks the households about the amount of remittances from other countries that they receive. We generated the indicator function r = {1, 0} for the event the household receives remittances. The census also asks the households if anybody from the family has migrated in the last five years. We generated the indicator function m = {1, 0} for the event the household has or had a migrant during the last five years. Tables 1 and 2 show that the two indicator functions partition the data into four household types (S = {1, 2, 3, 4}): (1) households with migrants and remittances, (2) households with migrants and no remittances, (3) households with remittances and no migrants, (4) households with neither remittances nor migrants. 3 The total number of observations for 8 to14 years old individuals is 1.6 million, 50.65% are males. The total number of observations for 15 to19 years old individuals is 1 million, 49.4% males. If we only include children from the head, the numbers go down to 1.5 million individuals, 8 to 14 years old, and 800 thousand individuals 15 to 19 years old. Remmitances. Cuecuecha. February 2008 5 In particular, 86% of the children in our sample belong to households that neither receive remittances nor have migrants. Approximately 8% of the children belong to households with migrants that receive no remittances, while 5% of the households can be classified in the category no migrants and receive remittances. Finally, 1% of households is classified in the category migrants and remittances. An explanation for the existence of the category s = 3, is that the census authority deleted from the migrant files all individuals who either migrated more than five years ago or did not belong to the household at the moment of the migration. Consequently, the category is formed by children that belong to potentially different categories of households: a. in some households the migrants left more than five years ago and remit home money, and the family still considers the migrants part of them; b. in some other households the migrants where not part of the household at the moment of migration but they still send money to their relatives. Because the status of the migrants in this type of household is not clear, we decided to keep this type of households apart from those of type s = 1. Moreover, to simplify our reference to this type of household, from now on we will refer to this type of households as households that have long term migrants (LTM) and receive remittances.4 Given that we are looking at children of different ages, we decided to standardize the education years according to the age of the children. Define h as the standardized ¡ ¢ education years given by h = H(a) − H(a) /sd(H(a)) with H(a) equal to the edu- cation that a child with age a has, H(a) equal to the mean education for children with age a , and sd(H(a)) equal to the standard deviation for children with age a. Tables 1 and 2 compare the standardized education outcomes for the children depending on the type of household they live in. All teenagers (i.e. 15 to 19 years old) living in households with migrants and/or remittances have at most the education of the children from households type s = 4. The same applies for children 8 to 14 years old. These differences are all significant at the 1% level. This result suggest a negative cor4 The census misses a fifth type of household comprised by those households that left entirely the country and have no extended family members in Mexico that will provide information about them. Remmitances. Cuecuecha. February 2008 6 relation between the accumulation of human capital and the reception of remittances and/or migration to the US. However, they do not represent causal relations because they may reflect the correlation between observed and/or unobserved characteristics of the different household types and the education choices for their children.5 A potential concern with our classification is that if the individuals migrated more than five years ago and do not remit money, then we can not identify such households. However, if an individual has left for more than five years and has not send money to the household, it implies that such individual are no longer part of the household. The concern, however, comes from the fact that such type of households could differ systematically from households of type s = 4. One would expect such household to be poorer, since they lost some members and the members do not remit money. This would specially be true for households where the migrant was the main breadwinner and the separation took place more recently, giving less time to the remaining members of the household for reorganization. If the children from these households are poorer, then the average education of children for households with no remittances and no migrants would be lower down and the average education of children with migrants and remittances would be increased. Given that the first group of households is larger, the effect of this misclassification would most likely be to increase the average education of children with remittances and migrants. Following the above idea, it should be the case that children and teenagers from households with LT migrants and no remittances are more likely to be found as outliers in the distribution of children with no remittances and no migrants. So in an attempt to attenuate this problem we will perform an estimation eliminating children that have a standardized education above +-2.5. 5 The exogenous characteristics of the children, their households and communities are presented in the appendix. Remmitances. Cuecuecha. February 2008 3 7 Empirical Methodology 3.1 Lee method To disentangle the effects of remittances and migration we will use a method proposed by Lee (1983). This is a two stage method, where it is assumed that there is a different equation for the endogenous variable for each type of household. In the first stage, the household selects its type, while in the second stage a different equation for each household type is estimated. The households can choose from the following types: (1) Have migrants with less than five years away from their households and receive remittances; (2) Have migrants with less than five years away from their households and receive no remittances; (3) Have long term migrants and receive remittances; (4) Have no migrants (and remittances). Assume that the household chooses the type that maximizes household utility and chooses the education of their children conditional on their type. Assume that the decision is taken in two stages. In stage one, the household chooses the action that will maximize utility. In stage two, conditional on the action and the characteristics of the children, the household chooses the level of education for their children. Consequently unobserved characteristics of the children and the household that influence the schooling decision are correlated with the actions taken by the household. Moreover, if the household solves the problem by backward induction, in stage one, the evaluation of the value that the family can achieve by being in each different state includes a level of education for each children that can potentially be different depending on the actions taken by the household. Let those levels of education be equal to: hs = X 0 β s + es For each choice, there is a latent variable: (1) Remmitances. Cuecuecha. February 2008 8 Is = Z 0 αs + es (2) where Z includes all variables in X plus a set of instrumental variables. Utility maximization implies that: I = s if Is > Ij (j = 1, 2, 3, 4; j 6= s) (3) εs = MaxIj − η s (j = 1, 2, 3, 4; j 6= s) (4) Let: If η s follows a type I extreme value distribution, Domencich and McFadden (1975) show that εs has the following distribution function: Fs (ε) = Prob(εs < ε) = Define: exp(ε) P exp(ε) + j6=s exp(Z 0 αj ) ε∗s = Js (εs ) = Φ−1 (Fs (ε)) (5) (6) where Φ represents the standard normal cdf. Equation 5 represents the first stage of the Lee (1983) method. The second stage consists in estimating the conditional expectation of equation 1, which Lee (1983) showed to be: hs = X 0 β s + σ s ρs ϕ (Js (εs )) + vs Φs (εs ) (7) where σ 2s = var(es ), ϕ is the pdf of the standard normal and ρs is the correlation coefficient between es and ε∗s . Moreover, E(vs |X, Z) = 0. In the estimation of the above model there are three main concerns: identification, the unit of analysis and the appropriate standard errors. The identification of the Remmitances. Cuecuecha. February 2008 9 model is done based on both the nonlinearity of the model and the existence of instrumental variables. While the nonlinearity of the model could be enough to identify the estimator, we use variables that are correlated with the choices of family type but that are otherwise assumed uncorrelated with the education of the children. Our instruments are: the migration rate in the municipality where the individual lives, excluding the household of analysis; remittances as a fraction of income in the municipality where the individual lives, excluding the household of analysis; an interaction of the two mentioned instruments; and the capital to labor ratio in the municipality.6 The intuition for using four instruments comes from the identification condition that is obtained using a linear probability model. In a linear probability model, each choice becomes a linear equation and hence a system of four endogenous variables and a structural equation is formed. The intuition for using the migration rate at the municipality level is that different authors have found that migration networks help individuals to migrate (Durand, Parrado and Massey, 1996) and if the household lives in a place with more migrants it is more likely that the migrant can be exposed to migration. Hence variations across municipalities in the migration network will identify our treatments. This type of identification has been used in the literature by other authors who used historic state-level migration rates (see Hanson and Woodruff, 2003; Mckenzie and Rapoport, 2004; or Borraz, 2005). Our variable has variation between communities and inside the community since by not considering the household of study, we generate a variable that varies depending on whether the family is a migrant or not. The intuition for using remittances as a fraction of the income in the municipality, excluding family i, is that in places where remittances constitute a bigger fraction of the income of the population, it is more likely that there would be more business dedicated to the reception of money. 6 7 The appendix shows the mean and standard deviations for the instruments. Moreover, the relevance of these instruments is shown by their correlation with the choices of the individuals, which is shown to be significant later in the results section. 7 While postal offices are one way through which individuals receive money orders in Mexico and they are found, in principle, in each municipality, electronic transfers are by and large the main Remmitances. Cuecuecha. February 2008 10 Consequently, we exploit differences across municipalities in the fraction of income that represent the remittances to identify the effect of remittances. By excluding the household of study of this measure, we generate a variable that will also have variation within the municipalities. We obtained a third instrument by interacting our two instruments. Our fourth instrument is the capital to labor ratio in the community.8 These instrument are used conditioning in the family income, the gdp per capita in the municipality and the investment over output ratio. The intuition for the use of the fourth instrument is that conditional on output and investment in the community, a larger capital to labor ratio must be correlated with better economic conditions that will reduce migration or the need for remittances, while the correlation between the capital labor ratio with education is controlled for using the municipality and household characteristics included in our equation. This variable also provides with variations at the level of communities. A second issue in the estimation is about the unit of analysis. In principle, a household makes a decision and then the decision applies to all children of the household. However, note that the objective of the Lee method is to control for the correlation between unobserved factors that are correlated with the decision of the family and the unobserved factors that determine the schooling level of a given children. If children belonging to the same household have different unobserved children characteristics, assigning the same predicted value for the selection correction to those children will measure with error the true selection. Consequently, we treat each children as an independent observation. The final concern has to do with the standard errors to be used in the study. We have at least three concerns about it. The first is that a two stage method needs to take into account the estimation error that comes from the first stage. We bootstrap the standard errors to solve this problem. The second is that the instruments are mechanism used by individuals to remitt money. According to Banco de Mexico (2008), in 2006 94% of all transactions were done via electronic transfers, and 3% where done using money orders. 8 The appendix shows a linear probability model in which the validity of the instruments is formally tested. Remmitances. Cuecuecha. February 2008 11 measured at the level of the municipality which can generate correlation between the observations. We clustered the standard errors by municipality. The third is that we treat observations for all children as independent observations, but there are children coming from same households. We will perform robustness checks clustering the standard errors by household. 3.2 Average Treatment Effects on the Treated Once estimated the equations for each type of household, we can now define the effects that will allow us to disentangle the effects of migration and remittances. To do so we will apply techniques related to the analysis of two treatments in individuals, when the assignment to treatment is non-random. Following Lechner (2002) define the average treatment effect of treatment k compared to treatment l on the participants in treatment k as:? θk,l = E (hk |s = k, x) − E (hl |s = k, x) (8) Where E (hk |s = k, x) is the standardized education years that children from households that chose s = k have, conditional on the observed and unobserved characteristics that make the household select s = k. E (hl |s = k, x) represents the standardized education years that children from households that chose s = k would have, should the household have chosen s = l. Pairwise treatment effects are sufficient to identity specific questions when a situation involves more than one treatment (Lechner, 2002). The pairwise treatments will allow us to disentangle the income effect and the allocation effect. Specifically, estimate five pairwise comparisons: (a) θ2,4 = E[h2 |s = 2, x] − E[h4 |s = 2, x] finds the ”allocation of time” effect. We do so by obtaining the expected value of education for children from households that have migrants and no remittances, and obtaining a counterfactual level of education for those children, should they live in households where migration is equal to zero. Remmitances. Cuecuecha. February 2008 12 (b) θ1,2 = E[h1 |s = 1, x] − E[h2 |s = 1, x] finds the ”income effect”. We obtain the expected value of education for children from households with remittances and migrants, as well as the counterfactual level of education that those children would have should they live in households with migrants and no remittances. (c) θ1,4 = E[h1 |s = 1, x] − E[h4 |s = 1, x] finds the ”combined effect” of migration and remittances. We accomplish this by obtaining the expected value of the education for children from households with migrants and remittances and obtaining the couterfactual expected value of education for that type of children, should they live in households with no remittances and no migrants. (d) θ1,3 = E[h1 |s = 1, x] − E[h3 |s = 1, x] finds the effect of long term migration. We accomplish that by obtaining the expected value of education for children from households with remittances and migrants, and estimating the counterfactual expected level of education if the children would live in households with remittances and with long term migrants. (e) θ3,4 = E[h3 |s = 3, x] − E[h4 |s = 3, x] finds the combined effect of migration and remittances in households with long term migrants. We do so by obtaining the expected education for children from households with remittances and long term migrants, and obtaining the counterfactual expected level of education if the children would live in households with neither remittances nor migrants. 3.3 Results Tables 5 and 6 present the results for the first stage of the Lee method. Table 5 shows a multilogit model fitted on individual, household and community characteristics for individuals between 15 to 19 years old. Table 6 shows the corresponding results for children 8 to 14. Table 5 shows that the instruments work very well for teenagers 15 to 19, with all instruments being always significant at the 10% level. For children 8 to 14, not all the instruments are independently significant but they are all jointly significant. Remmitances. Cuecuecha. February 2008 13 Tables 7 and 8 show the results of the second stage. The most important factor for these tables is the σ s ∗ ρs , which shows the importance of the selection into each of the equations. The tables show that selection matters for both teenagers and children, eventhough the results indicate that selection is more important in the case of teenagers.9 Tables 9 and 10 show the ATT that are obtained using the coefficients from the Lee method for all the individuals, and for males and females.10 Let´s first discuss the combined effect of remittances and migration. Teenagers 15 to 19 years old that live in households with remittances and migrants increase their education .35% (.028 schooling years), compared to the education they would have without remittances and no migrants. The effect is found to be bigger for males (.74%) and non-significant for females. Children 8 to 14 years old from households with remittances and migrants increase their education 4.6% (.192 schooling years) compared to the education they would have without remittances and no migrants. The effect for males is 3.3% and 4.2% for females. In regard to the income effect, we find that teenagers that live in households with migrants and remittances increase their education .29% (.023 schooling years) compared to the education that they would have in households with migrants and no remittances. The effect is bigger for males (.46%) and it is non-significant for females. For children 8 to 14 the effect is considerably larger 3.2%, especially for females (4.2%). The allocation of time effect has mixed results. Teenagers that live in households 9 There are different potential explanations for this result. The first is that children 8 to 14 are in shooling years that are supposed to be mandatory in Mexico, while most of teenagers 15 to 19 are in schooling years that are not mandatory. A second explanation is that child work is allowed by law only at age 14. A third explanation is that migration of the young starts as early as 14 years old. A fourth potential explanation is that because primary school (6 to 12 years old) and junior high school (13-15) are mandatory there are more public schools offering such levels at the municipality level, than high schools and colleges offering public education at the municipality level. In all these potential explanations, either the direct cost or the opportunity cost of attending school increases at age 15. Further research into this issue is needed and is left for future work. 10 The multilogit models and second stage equations used for males and females are similar to those shown in tables 5 to 8, except that exclude the sex variable. Results available upon request. Remmitances. Cuecuecha. February 2008 14 with migrants and no remittances have a similar level of education to the one that they would have if their households would be with no migrants. However, we find that for males the effect is positive and almost a half of the total positive combined effect, while for females the effect is found to be negative (-.45%). For male children the effect is also found to be positive (.36%) while for females it is also found to be negative (-.55%). The effect of long term migration is also found to have mixed results. Teenagers that live in households where the migrants left less than five years ago and receive remittances have less education (-3.5%) compared to what education they would have if they would live in households with remittances and where the migrants left more than five years ago. This implies that over time the positive effects of migration and remittances accumulate. On the other hand, children 8 to 14 years old that live in households with migrants and remittances have more education (2.8%) compared to what would they have if they would live in households with remittances and migrants that left more than five years ago. These results apply for both boys and girls. Finally the combined effect of remittances and migration for families with long term migrants is found to be negative. Specifically, teenagers that live in households with remittances and long term migrants have less education (-.73%) compared to what would they have if they would live in households with no remittances and no migrants. The effect is negative for males and females and also for children 8 to 14 years old (-.77%). 3.4 Selection in observables We present this section for two main reasons: first as a robustness check on our previous estimation, and because if observable characteristics are the main reason for selection bias to exist, the propensity score matching (Rosenbaum and Rubin, 1983) would be our best option to estimate the ATT. This method involves the estimation of a propensity score that identifies the probability for an observation to receive the Remmitances. Cuecuecha. February 2008 15 treatment, and later on matching individuals that are very similar in their probability of receiving the treatment. Under the assumption that conditional on the propensity score, the treatment is uncorrelated with the characteristics of the individuals, we can identify the ATT. This characteristic is called the "Balancing Hypothesis" and implies that for a given propensity score, exposure to treatment is random (Becker and Ichino, 2002). Following (Heckman, et al,1996 ) we will limit testing of this hypothesis to individuals that belong to the common support.11 We followed the approach of calculating each pairwise ATT using individual models of the binary relations implied for each pair analyzed. 12 Each pairwise comparison involves the estimation of a different propensity score. We, however, fix the model to be estimated in each case using a set of exogenous characteristics and instruments that were also used in our instrumental variable estimation.13 Then, the ATT was estimated using three different methodologies: stratification, nearest neighbor, and kernel.14 Table 11 presents our results for males and females 15 to 19 years old, while Table 12 presents the results for children 8 to 14 years old. Unlike our estimation using the Lee method, the combined effect is found to be insignificant for males 15 to 19 years old, while it is found to be negative for females 15 to 19 years old. Similarly, the combined effect is found to be insignificant for females 15 to 19 years old. In contrast, the combined effect is found to be positive for males 8 to 14 years old, which is similar to our estimation with the Lee method. The income effect is found to be positive and significant for all our subsamples, 11 The estimation is done using Becker and Ichino (2002) STATA program. An alternative methodology that can be used in the context of multiple treatments was also implemented. Qualitatively similar results to those mentioned in the text were obtained. Those methodologies are based on Lechner (2002). Results available from the authors upon request. 13 We also follow the approach of letting each pairwise probability to be estimated according to models that will satisfy implications of the balancing hypothesis (Becker and Ichino: 2002). Specifically, we specify a model under which we could partition the observations in the common support into blocks that would satisfay the condition that treated and non-treated individuals have equal means for all exogenous variables used in the estimation of the score. Results are qualitatively similar to those presented using a common model for all pairwise probabilities. The appendix shows the variables used in the models fitted to obtain the propensity scores. 14 The estimations are done using the STATA programs provided by Becker and Ichino(2002). 12 Remmitances. Cuecuecha. February 2008 16 just as it was found with the Lee method. The results from the random matching estimators find a negative allocation effect for both males and females. This is similar to what is found for females using the Lee method and different in sign for males. The random matching estimations show that the effect of long term migration (LTM) is non-significant for males 15 to 19 years old and negative and significant for females 15 to 19 years old. These results agree with those founds with the Lee method for females. The effect of LTM is found to be positive and significant for males 8 to 14 years old and insignificant for females. These results also agree with the Lee method results for males. The combined effect for long term migrants is found to be insignificant for 15 to 19 years old individuals, which does not contradict the negative and significant effect found with the Lee method. The combined effect for long term migrants is found to be insignificant for females 8 to 14 years old, while it is found positive and significant for 8 to 14 boys, which contrasts with the negative and significant effects found with the Lee method. 3.5 The importance of selection Given the discrepancy of results between the Lee and the random matching estimator for the combined effect of migration and remittances, we decided to attempt to study the type of selection implied by the exposure of migration and remittances and its relation with the human capital of children and young adults. The type of selection in unobservable characteristics is determined using the non-parametric methodology of DiNardo. et al (1996) which has been applied to migration by Chiquiar and Hanson(2005). In order to do so, let us write the distribution of the standardized education h in a way that we can see its relation with the type of household that we are analyzing: Γi = Γ(h|s = i) = Z Φi (h|x) Ψ(x|s = i)dx (9) Remmitances. Cuecuecha. February 2008 17 where Φi (h|x) is the conditional distribution of education for children from households type s = i and exogenous characteristics x. Also Ψ(x|s = i) is the conditional distribution of characteristics for households from type s = i. Therefore, the distribution of education of children from type i households is conformed by how households optimally choose education for children with characteristics x, and the distribution of characteristics among children of such households. Having defined the conditional distribution of education, we can now define counterfactual distributions that will allow us to find how the exposure to migration and remittances can affect the education distribution. Define the counterfactual conditional distribution as: Γi,j = Z Φj (h|x) θi,j Ψ(x|s = j)dx (10) where θi,j is a reschaling factor given by: θi,j = Ψ(x|s = i) Ψ(x|s = j) (11) θi,j is obtained using Bayes law (DiNardo, Fortin and Lemiuex , 1996; Chiquiar and Hanson, 2005) and we write it as: θi,j = Ψ(x|s = i) Pr (s = i|x) Pr(s = j) = Ψ(x|s = j) Pr(s = j|x) Pr(s = i) (12) where Pr (s = i|x) is obtained using a parametric multilogit model, and Pr(s = i) is obtained using the observed sampling rates.15 Γi,j can be understood as showing how the distribution of education for children of households of type i would look like if the optimal choices of education in their households would follow the optimal decisions of households of type j. Define the difference in education distribution ∆i,j as: ∆i,j = Γi,j − Γj = 15 Z (θi,j − 1) Φj (h|x) Ψ(x|s = j)dx (13) The multilogit model used here is the same that is shown in the first stage of the Lee method. Remmitances. Cuecuecha. February 2008 18 Notice that children with a weight larger than one contribute positively to the difference. If θi,j > 1, we have that children with characteristics x have higher probability of living in households of type i than probability of living in households of type j. If we graph the above difference with the education of the children, we can learn how the ratio of probabilities correlates with education. If the difference ∆i,j shows positive mass above the education mean, it implies that children with larger probability of being in households of type i have education above the mean. If the difference shows positive mass below the mean, then we would say that there is negative correlation between the probability of being in households of type i and education. If the difference shows in the middle of the distribution we would speak of a sort of "selection from the middle" (Chiquiar and Hanson, 2005) in the sense that children that have higher probability of living in households of type i are most likely to be found in the middle of the education distribution of children from type j. We estimated 3 differences: ∆1,4 =Difference between the counterfactual distribution of education for children with the characteristics of households with migrants and remittances, and observed distribution of education for children with the characteristics of households with no migrants and no remittances, holding the conditional education distribution constant and equal to that of children with no remittances and no migrants. ∆2,4 =Difference between the counterfactual distribution of education for children with the characteristics of households with migrants and no remittances, and observed distribution of education for children with the characteristics of households with no migrants and no remittances, holding the conditional education distribution constant and equal to that of children with no remittances and no migrants. ∆3,4 =Difference between the counterfactual distribution of education for children with the characteristics of households with migrants that left more than five years ago and no remittances, and observed distribution of education for children with the characteristics of households with no migrants and no remittances, holding the Remmitances. Cuecuecha. February 2008 19 conditional education distribution constant and equal to that of children with no remittances and no migrants. Figures 1 and 2 show the results of the counterfactual experiments for teenagers 15 to 19 years old and children 8 to 14 years old. For teenagers, the figures show clearly that individuals with characteristics that will make them more likely to live in households exposed to migration or remittances are more likely to be found above the mean of the education distribution for children from households with no migrants and no remittances. The only exception is that of households types s = 2, which are also shown with some positive mass below the mean. For children, the type of selection depends in the types of households being compared. Children from households type s = 1, are seen to be selected from the middle of the education distribution of the children with no migrants and no remittances. Children from households type s = 3, seem to selected from the bottom of the education distribution of the children type s = 4. Finally, children from households type s = 2, seem to be selected from the top of the education distribution of children from households s = 4. 4 Conclusion Remittances and migration can have effects of opposite sign in the education of children. We refer to the effects generated by the additional income brought by remittances as the income effect. Every other effect we called it the allocation effect. Using the 2000 Mexico census we identify whether households have members in the US and whether they receive remittances. This information allow us to disentangle the income and the allocation effects, as well as the combined effect of remittances and migration on the education of children and teenagers in Mexico. We use instrumental variables to accomplish our purpose and we have four main results: First, we identify the income effect to be positive: on average children and teenagers increase their levels of education above what would they have in the absence Remmitances. Cuecuecha. February 2008 20 of remittances. The first experiment an increase of 3% in education years, while the latter increase .3%. Second, we identify the allocation effect to be non-negative for males and negative for females. For male teenagers the effect is non-significant while for male children the effect is an increase of .45% in education years, compared to the education they would have with no migrants in the household. For females, it represents a reduction between —.55% (children) and -.45% (teenagers). Third, the combined effect of migration and remittances is found to be positive for males and non-negative for females: it represents an increase of 4.6% for male children and an increase of .74% for male teenagers, compared to the levels of education that they would have with no migrants and no remittances in the household. For female children the increase is 3.3%, while no effect is found for female teenagers. Fourth, we also identify the combined effect of remittances and long term migration to be negative: teenagers have -.73% education years and children have -.77% education years compared to what would they have with no remittances and no long term migrants. Our analysis also found that selection in unobservable characteristics is important in the education equations. We found this importance using both parametric and semiparametric techniques. This implies that the use of instrumental variables to identify the effects of remittances and migration is the best strategy to follow. Our results have important implications for research and policy. First, they imply that while remittances and migration can be positive for human capital due to the additional resources that the household can acquire through them, factors that reduce the sending of remittances or difficult the contacts between families must have negative impacts on the education of children. Second, our finding that effects are different by gender, being either not so positive for females or even negative, imply that families can either be discriminating within the household, or that their decisions respond to market incentives by using resources selectively among household Remmitances. Cuecuecha. February 2008 21 members. Further research is needed to better understand the effects found. 5 5.1 Appendix 1. Characteristics of children and teenagers Tables A1 and A2 present the characteristics of the children and teenagers depending on the household type in which they live. Teenagers have in average 17 years in all households and half of them are males. Children are on average 11 years old and half of them are males. There are significant differences in terms of household income, education and age of the father, and the characteristics of the municipalities where they live in. Teenagers and children that receive no remittances live in households where the father is more educated and the family income is higher. They also live in municipalities that have higher GDP per capita. 5.2 2. Instrumental variables and regions Table A3 shows the mean and standard deviations for the instruments used in the paper. Migrants represent around 2% of the population, while remittances represent around 2% of municipality income. The capital labor ratio is calculated using gross fixed investment by municipality divided by population employed in the municipality. This data comes from INEGI (1999). The table also shows the fraction of children that lives in the five regions in which the entire country is divided. These regions are based on Chiquiar (2005). The table also shows the states included in each region. 5.3 3. Variables used in propensity score models Table A4 shows the variables used in each of the models fitted for the estimation of the propensity scores used to form the pairwise matches. Those variables are the same used in most of the regressions presented in the paper, with the exception of Remmitances. Cuecuecha. February 2008 22 regional dummies. The introduction of the regional dummies consistently generated few matches and we decided to drop them from the propensity score models. Sex was dropped from all regressions since the matches were always done conditional on gender. 5.4 4. Alternative methods Under the assumption of random assignment the average treatment effect on the treated can be estimated using appropriate OLS equations. To illustrate, let suppose we are interested in comparing children from households s = 1, 2, 3 with households with s = 4. In other words, our control group are households with no remittances and no migrants. An appropriate OLS equation would be: hj = Xj0 β + Tj0 δ 4 + ej (14) Where hj is the standardized years of education of children j; Xj represents a vector of exogenous characteristics; Tj is a vector of indicator functions for children j and ej is the error term. Vector Tj is formed by three indicator functions: T 1 which is an indicator for the event : household receives remittances and has migrants”; T 2 an indicator function for the event ”the household has migrants and receive no remittances”; and T 3 an indicator function for the event : ”the household receives remittances and has no migrants”. Each element δ k,4 from the vector δ 4 identifies the average treatment effect on the treated (ATT) as follows16 : E (hk |s = k, x) − E (h4 |s = k, x) = δ k,4 16 (15) Under the assumption that the treatments are exogenous, the ATT is identical to the average treatment effect in the population. At the mean of the entire sample, each element δ k,4 from the vector δ 4 identifies the following difference: E (hk |s = k, x, ) − E (h4 |s = 4, x) = δ k,4 Remmitances. Cuecuecha. February 2008 23 Note that unlike the Lee method, where constants and coefficients are allowed to vary depending on the type of household studied, or the random matching estimators where a semi-parametric method is used to obtain the ATT, equation 14 only allows the ATT to occur in one parameter. Consequently, differences between this OLS approach and the Lee method, as well as with the random matching estimator, come not only from the assumption of random assignment, but also from the use of a simpler model in the case of OLS. The assumption of random assignment can easily be rejected in the context of migration because the literature on migration has showed the importance of selection in observable and unobservable characteristics (Borjas, 1987), particularly for the Mexican case (Chiquiar and Hanson, 2005). Consequently, the coefficients δ k,4 from the OLS estimation are biased estimators of θk,4 . An alternative estimation is the use of an instrumental variable approach and a linear probability model. This method estimates an structural equation given by equation 14, together with three reduced form equations, one for each indicator variable Tj .The difference between this IV estimation and the Lee method comes not only from the assumption of a linear probability model, but also from the simpler model assumed in equation 14. A similar argument applies for the difference between the IV estimator and the random matching estimator. The instruments used are the same used in the Lee method and the estimation of the propensity scores for the random matching. Hausman tests show the need for instruments and all Sargan test performed show the validity of the instruments. Table A5 shows the results from the estimation. The combined effect is found to be positive and significant for all subsamples, as with the Lee method. The income effect is found to be positive and significant for all subsamples. The allocation effect is found to be negative and significant for all subsamples. The effect of long term migration is found to be negative and significant for all subsamples. The combined effect for LT migrants is found to be negative and significant. Remmitances. Cuecuecha. February 2008 24 In conclusion, the effects found with the IV method are straightforward: the combined effect is positive, the income effect is positive and the allocation effect is negative. The income effect is then obtained positive with any of the methods. The combined effect is found non-negative with any of the methods. The allocation effect is found non-positive for females in any of the three methods. On the other hand, the allocation effect for males is found with similar signs for the Lee method and the random matching method, but not for the linear IV method. Given that the linear IV imposes more restrictions in the model estimated our favorite estimations are those obtained with the Lee method. Remmitances. Cuecuecha. February 2008 25 [1] R. Adams and A. Cuecuecha. Remittances, Household Expenditure and Investment in Guatemala. World Bank, Working Paper, 2007. [2] S. Becker and A. Ichino. Estimation of Average Treatment Effects Based on Propensity Scores. Stata Journal, 2(4): 358-377. 2002 [3] F. Borraz. Assesing the Impact of Remittances on Schooling: The Mexican Experience. Global Economy Journal, 5(1):1-32. 2005 [4] D. Chiquiar and G. Hanson. International Migration, Self Selection, and the Distribution of Wages. Journal of Political Economy, 113(2):239-281. 2005. [5] D. Chiquiar. Why Mexico’s regional income convergence broke down. Journal of Development Economics, 77(1): 257-275. 2005. [6] J. DiNardo, N.M. Fortin, and T. Lemieux. Labor Market Institutions and the Distribution of Wages, 1963-1992: A Semiparametric Approach. Econometrica, 64(5): 1001-1044. 1996. [7] J. Durand, E. Parrado, and D. Massey. Migradollars and Development: A Reconsideration of the Mexican Case. International Migration Review, 30(2): 423444. 1996. [8] C. Dustmann, and O. Kirchkamp. The Optimal Migration Duration and Activity Choice after Remigration. Journal of Development Economics, 67(1):351-372. [9] G. Esquivel and A.H. Pineda. Remittances and Poverty in Mexico: A Propensity Score Matching Approach. El Colegio de México. Working Paper. 2006. [10] G. Hanson and C. Woodruff. Emigration and Educational Attainment in Mexico. NBER. Working Paper. 2003. [11] J. Heckman, H Ichimura, J Smith and P. Todd. Sources of Selection Bias in Evaluating Social Programs: An interpretation of conventional measures and evidence of effectiveness of matching as a program evaluation method. Proceedings from the National Academy of Sciences USA, 93(23):13416-13420. 1996. [12] M. Lechner. Some Practical Issues in the Evaluation of Heterogenous Labour Market Programmes by Matching Methods. Journal of the Royal Statistical Society, 165 (1):59-82. 2002. [13] Lee, Lung-Fei. 1983. Generalized Econometric Models with Selectivity. Econometrica 51, no. 2: 507-512. [14] D.P. Lindstrom. Economic Opportunity in Mexico and Return Migration from the United States. Demography, 33(3):357-374. 1996. [15] E. Lopez-Cordoba. Gloablization, Migration and Development: The Role of Mexican Migrant Remittances. Mimeo, Inter American development Bank. 2004. [16] D. Mckenzie and H. Rapoport. Network Effects and the Dynamics of Migration and Inequality: Theory and Evidence from Mexico. Stanford University and Bar-Ilan University, Working Paper. 2004. [17] D. Mckenzie and H. Rapoport. Migrant Network, Migration Incentives and Education Inequality in Rural Mexico. Stanford University and Bar-Ilan University, Working Paper. 2005. [18] E. Rodriguez-Oregia and A. Cox-Edwards. The Effect of Remittances on Labor Force Participation in Mexico. Universidad Iberoamericana, Working Paper. 2006. [19] G. Zarate-Hoyos. Consumption and Remittances in Migrant Households: Toward a Productive Use of Remittances. Contemporary Economic Policy, 22 (4):555565. 2004. Remmitances. Cuecuecha. February 2008 26 [20] Banco de México. Familiy Remittances Datasheet [online]. Consulted on January 2008. Available at: http://www.banxico.org.mx/polmoneinflacion/estadisticas/ balanzaPagos/balanzaPagos.html [21] INEGI. Censos Económicos 1999. Remmitances. Cuecuecha. February 2008 Table 1, 2000 Mexican Census, 15 to 19 years old Education years and standardized ed. yrs. of young adults by type of family Differences with respect to families with no rem. and no mig. in parenthesis All Rem.and Rem. and No rem. No rem. Mig. LT mig. and Mig. and no mig. All Ed. yrs. 7.95 7.61 7.99 7.52 7.99 Standardized ed yrs. -.01 -.11 .01 -.16 .01 Difference (-.12)** (-.0004) (-.17)** Males Ed. yrs. 7.81 7.57 7.86 7.47 7.84 Standardized ed yrs. -.06 -.12 -.18 -.04 -.04 Difference (-.07)* (.003) (-.13)** Females Ed. yrs 8.11 7.66 8.15 7.58 8.16 Standardized ed yrs. .05 -.11 -.13 .06 .07 Difference (-.18)** (-.004) (-.21)** N (males and females) 721,978 1,294 38,637 54,115 627,932 % (sample) 100 .18 5.35 7.49 86.97 % (weighted) 100 .21 4.32 8.89 86.58 Source: Calculations done by the author using the 9.1% public sample of the 2000 Mexico census. Sample includes only children of the head of household, for whom information on their education, the education of the head, and the migration information in the households is available. **1 % significance level. 27 Remmitances. Cuecuecha. February 2008 Table 2, 2000 Mexican Census, 8 to 14 years old Education years and standardized ed. yrs. of children by type of family Differences with respect to families with no rem. and no mig. in parenthesis All Rem. and Rem. & No rem. No rem. Mig. LT mig. and Mig. and no mig. All Ed. yrs 4.14 4.13 4.15 3.93 4.16 Standardized ed. yrs .001 .05 .002 -.097 .011 Difference (.04)* (-.01)* (-.10)** Males Ed. yrs 4.09 4.11 4.11 3.93 4.11 Standardized ed. yrs -.03 .05 -.03 -.11 -.03 Difference (.003) (.001) (-.17)** Females Ed. yrs 4.20 4.15 4.19 3.94 4.23 Standardized ed. yrs .04 .05 .04 -.08 .05 Difference (-.001) (-.01)? (-.14)** N (males and females) 1 ,2 9 8,3 6 2 2,405 69,555 102,880 1 ,1 2 3 ,5 2 2 % (sample) 100 .18 5.35 7.92 86.55 % (weighted) 100 .22 4.29 9.35 86.14 Source: Calculations done by the author using the 9.1% public sample of the 2000 Mexico census. Sample includes only children of the head of household, for whom information on their education, the education of the head, and the migration information in the households is available. **1 % significance level. *5% significance level. ?10% significance level. 28 Remmitances. Cuecuecha. February 2008 Table 3, Multilogit model for Probability that a household is of type S (First stage of Lee Method) Teenagers 15 to 19. Household Type Migration & Migration & LT Migrants Remittances No Remittances & Rem. Age -.022 -.011 -.001 (.028) (.007) (.005) Sex .012 -.0008 -.001 (.078) (.021) (.014) Family Income -5.60e-07* 1.94e-07 -1.80e-06*** (3.07e-07) (1.49e-07) (6.47e-07) PIB pc -.00001** -1.19e-06** -2.34e-06*** (6.67e-06) (6.00e-07) (4.59e-07) FBK in -6.04e-06** .467* -.389*** municipality ( 3.31e-06 ) (.273) (.198) Migration rate 37.52*** 32.07*** -4.097*** in municipalitya (1.145) (.156) (.423) Remittances as % 8.09*** -24.25*** 7.61*** of municipal incomea (1.213) (1.309) (.127) Mig rate in mun* 189.85*** 125.81*** 165.03*** Rem/income in mun (16.683) (10.36) (12.417) Capital/Labor -.0008* .001*** -.0003*** in municipality (.0004) (.00005) (.00006) Constant -10.80 -5.43*** -3.30*** (.258) (.031) ( .020) N=715743. Pseudo R2 : 51%. The regression includes the education and age of the head of household, as well as four region dummies. *** Significant at 1% level; **Significant at 5%. * Significant at 10 %. a Excludes members of household i. 29 Remmitances. Cuecuecha. February 2008 Table 4, Multilogit model for Probability that a household is of type S (First stage for Lee method) Children 8 to 14. Household Type Migration & Migration & LT Migrants Remittances No Remittances & Rem. Age -.022 .006 .0002 (.028) (.006) (.004) Sex .012 .012 .013 (.078) (.026) (.016) Family Income -5.60e-07* -1.84e-06 -1.26e-06 (3.07e-07) (1.51e-06) (7.85e-07) PIB pc -.00001** -1.38e-06 -3.24e-06 (6.67e-06) (1.38e-06) (9.75e-07) FBK in -6.04e-06** 4.288*** -.218 municipality ( 3.31e-06 ) (.442) (.339) Migration rate -234.45 -11.60 83.58 in municipalitya (756.424) (120.614) (99.651) Remittances as % 8.16*** -8.61*** 6.63*** of municipal incomea ( 1.002) (.203) ( .127) Mig rate in mun* 23328.22 5574.50 -4325.06 Rem/income in mun (58892.36) (12825.97) (10719.44) Capital/Labor -.001*** .00003 -.0005*** in municipality ( .0002) (.0003 ) ( .0001) Constant -10.80*** -2.45*** -2.69*** (.258) ( .043) ( .037) N=1287929. The regression includes the education and age of the head of household, as well as four region dummies. *** Significant at 1% level; **Significant at 5%. * Significant at 10 %. a Excludes members of household i. 30 Remmitances. Cuecuecha. February 2008 Table 5, Regressions for Standardized Education (Second stage of Lee method) Teenagers adults 15 to 19. Families with Migration & Migration & LT Migrants No Mig. Remittances No Remittances & Rem. No Rem. Age -.059** .007 -.002 -.006** (.024) ( .005) (.005) (.002) Sex .005 -.036 -.108*** -.118*** (.086) (.023) (.014) (.004) Family Income .00002** 1.96e-07* 4.09e-06*** 5.74e-07*** (9.85e-06) (1.00e-07) (1.36e-06) (1.35e-07) PIB pc -.00001** -6.44e-06*** -5.16e-07 -1.02e-06*** (6.67e-06) (2.29e-06) (4.52e-07) (3.64e-07) FBK in 1.096 -.988 -.453* -.371** municipality (.965) (1.147) (.235) (.148) Sigma i -.264 -.494*** .202*** -.093** (.316) (.169) (.061) (.043) Constant .275 -.089 -.690*** -.206*** (.849) (.139) (.126) (.043) N 1278 53087 37992 619682 R2 .11 .07 .08 .05 All reg. include the ed. and age of the head of hh. . Clustered s.e. by mun..**Significant at 1% level; *Significant at 5%. a Instruments: K/L ratio in municipality, mig. rate in mun.exc. hh i, and rem./gdp in mun. exc. hh i, interaction of mig. rate in mun. and rem/gdp in mun. both exc. hh i. 31 Remmitances. Cuecuecha. February 2008 32 Table 6, Regressions for Standardized Education (Second Stage Lee Method) Children 8 to 14. Families with Migration & Migration & LT Migrants No Mig. Remittances No Remittances & Rem. No Rem. Age -.039** -.037*** .004 .002 ( .016) (.005) (.003) (.001) Sex .036 -.024* -.085*** -.090*** ( .059) (.013) (.010) (.002) Family Income 2.86e-06* -9.34e-08 6.67e-07 3.38e-07*** (1.48e-06) (1.90e-07) (7.35e-07) (9.96e-08) PIB pc -5.39e-06 -4.16e-07 5.65e-07 4.02e-07 (9.31e-06) (3.16e-06) (6.06e-07) (3.16e-07) FBK in -.525 .160 -.015 -.1102845 municipality (2.122) (.630) (.156) (.115) Sigma i -.018 -.033** .008 -.009 (.013) (.013) (.018) ( .006) Constant .097 -.138* -.203*** -.210 (.161) ( .083) ( .042) ( .022) N 1494 98670 68457 1109547 R2 .04 .05 .03 .03 All reg. include the ed. and age of the head of hh. . Clustered s.e. by mun..**Significant at 1% level; *Significant at 5%. a Instruments: K/L ratio in municipality, mig. rate in mun.exc. hh i, and rem./gdp in mun. exc. hh i, interaction of mig. rate in mun. and rem/gdp in mun. both exc. hh i. Table 7, Predicted Standardized Education and Counterfactual Standardized Education Obtained for the Pairwise Average Treatment Effect on the Treated. (Lee Method) Teenagers adults 15 to 19. Pairwise E[zs |x,h=s] E[zk |x,h=s] ATT ATT ATT. comparison (1) (2) (Difference 1-2) Males Females Combined Effect -.059 -.087 .028*** .059*** -.004 s=1; k=4 (.010) (.007) (.007) (.010) (.012) Income Effect -.059 -.083 .023*** .037*** .008 s=1; k=2 (.010) (.008) (.007) (.011) (.013) Allocation Effect -.162 -.163 .001 .032*** -.033*** s=2; k=4 (.001) (.001) (.0007) (.0009) (.001) LT migration effect a -.059 .195 -.254*** -.246*** -.266*** s=1;k=3 (.010) (.008) (.007) (.011) (.012) b Combined for LT -.009 .044 -.053*** -.051*** -.058*** s=3;k=4 (.001) (.001) (.0005) (.0008) (.0007) *** Significant at 1% level; *Significant at 5%. a Compares children in families with remittances and migrants (s=1) with the education that children from those families would have if the migrants of the familiy would have more than five years without returning. b Compares children in families with remittances and migrants that left more than five years ago (s=3) with the education that children from those families would have if no remittances and migrants would exist in the family. Remmitances. Cuecuecha. February 2008 · 33 Table 8, Predicted Standardized Education and Counterfactual Standardized Education Obtained for the Pairwise Average Treatment Effect on the Treated. (Lee Method) Children 8 to 14. Pairwise E[zs |x,h=s] E[zk |x,h=s] ATT ATT ATT. comparison (1) (2) (Difference 1-2) Males Females Combined Effect .082 -.109 .192*** .138*** .173*** s=1; k=4 (.007) (.004) (.005) (.004) ( .006) Income Effect .082 -.053 .136*** .115*** .176*** s=1; k=2 (.007) (.006) (.005) (.004) (.007) Supervision Effect -.114 -.121 .007*** .015*** -.023*** s=2; k=4 ( .0008) (.0006) (.0003) (.0004) ( .0005) Supervision for LTa .083 -.034 .117*** .010*** .103*** s=1;k=3 (.007) (.005) (.004) (.004) (.005) Combined for LTb -.007 .024 -.032*** -.036*** -.031*** s=3;k=4 (.0007) ( .0006) (.0002) ( .0004) (.0003) a *** Significant at 1% level; *Significant at 5%. Compares children in families with remittances and migrants (s=1) with the education that children from those families would have if the migrants of the familiy would have more than five years without returning. b Compares children in families with remittances and migrants that left more than five years ago (s=3) with the education that children from those families would have if no remittances and migrants would exist in the family. Table 9, Average Treatment Effect on the Treated, Pairwise comparisons Random Matching Estimators, 15 to 19 years old Males Combined Income Allocation LT migration Combined Remittance Effect Effect Effect Effect effect for LTM Effect 1,4 1,2 2,4 1,3 3,4 Method θ θ θ θ θ θr Stratification -0.015 0.079 -0.181*** -0.025 -0.072 -0.028 (0.060) (0.060) (0.058) (0.058) (0.053) (0.057) Nearest -0.050 0.164** -0.169*** -0.102 -0.083 0.002 Neighbor (0.084) (0.079) (0.060) (0.080) (0.065) (0.078) Kernel -0.009 0.098* -0.192*** -0.017 0.002 -0.036 (0.058) (0.060) (0.066) (0.042) (0.056) (0.050) N 8,938 8,938 377,811 8,938 177,803 186,741 Females Stratification -0.096 0.093 -0.238*** -0.100 -0.037 -0.099 (0.068) (0.069) (0.063) (0.064) (0.062) (0.063) Nearest -0.177** 0.082 -0.243*** -0.191** -0.111 -0.212** Neighbor (0.089) (0.092) (0.065) (0.086) (0.080) (0.087) Kernel -0.081 0.145*** -0.237*** -0.093 0.001 -0.091 (0.060) (0.052) (0.068) (0.067) (0.060) (0.062) N 7,561 7,561 315,938 7,561 159,573 167,134 ***Significant at 1%. ** at 5%. *at 10%. Source: 9.1% public sample of the 2000 Mexico census. Sample includes only children for whom information on the edu., and age of the head of household is known. Remmitances. Cuecuecha. February 2008 34 Table 10, Average Treatment Effect on the Treated, Pairwise comparisons Random Matching Estimators. 8 to 14 years old Males Combined Income Allocation LT migration Combined Remittance Effect Effect Effect Effect effect for LTM Effect 1,4 1,2 2,4 1,3 3,4 Method θ θ θ θ θ θr Stratification 0.082** 0.154*** -0.059 0.040 0.051 0.082** (0.038) (0.048) (0.044) (0.042) (0.040) (0.043) Nearest 0.126** 0.159** -0.076* 0.131** 0.080* -0.011 Neighbor (0.065) (0.062) (0.045) (0.061) (0.049) (0.058) Kernel 0.094** 0.144*** -0.084** 0.031 0.051 0.079** (0.040) (0.044) (0.036) (0.044) (0.035) (0.041) N 16,203 16,203 668,716 16,203 310,073 326,276 Females Stratification -0.015 0.221*** -0.198*** -0.027 0.057 -0.015 method (0.044) (0.053) (0.045) (0.046) (0.046) (0.043) Nearest 0.026 0.176*** -0.179*** -0.070 -0.034 -0.093 Neighbor (0.061) (0.066) (0.047) (0.063) (0.053) (0.058) Kernel 0.001 0.194*** -0.206*** -0.016 0.047 0.021 (0.040) (0.050) (0.040) (0.043) (0.046) (0.039) N 15,423 15,423 650,017 15,423 295,241 310,664 ***Significant at 1%. ** at 5%. *at 10%. Source: 9.1% public sample of the 2000 Mexico census. Sample includes only children for whom information on the edu. and age of the head of household. Remmitances. Cuecuecha. February 2008 35 Table A1, 2000 Mexican Census, 15 to 19 years old Characteristics of young adults, their households and their municipalities (Differences with respect to families with no mig. and no rem. in parenthesis) All Rem. and Rem. & No remi. No rem. Mig. LT mig. and Mig. and no mig. Age 16.8 16.7 16.8 16.79 16.8 (-.01) (.00) (-.002) Males (%) 52.9 54.2 52.7 54.4 52.7 (.01) (-.001) (.01)** Family income 5704.63 6214.39 4669.14 7081.1 5613.6 (annual pesos pc) (600) (-944)** (1467)** Education of HH 7.45 5.92 4.74 6.76 7.66 Head (-1.74)** (-2.92)** (-.90)** Age of HH head 43.9 48.8 51.11 41.9 43.7 (5.13)** (7.37)** (-1.83)** % of children in hh 9.10 100 0 100 0 with mig. % of children in HH 4.53 100 100 0 0 with rem. GDP pc 10,376 6,835 8,455 6,431 10,779 in mun. (-3,943)** (-4,347.74)** (-2323.37)** Gross fixed investment .037 .045 .035 .042 .037 /gdp in mun. (.008)** (-.002)** (.004)** N (thousands) 721,978 1,294 38,637 54,115 627,932 Source: Calculations done by the author using the 9.1% public sample of the 2000 Mexico census. Sample includes only children of the head of household, for whom information on their education, the education of the head, and the migration information in the households is available. **1 % significance level. Remmitances. Cuecuecha. February 2008 36 Table A2, 2000 Mexican Census, 8 to 14 years old Characteristics of children, their households and their municipalities (Differences with respect to families with no mig. and no rem. in parenthesis) All Rem. and Rem. and No rem. No rem. Mig. LT mig. and Mig. and no mig. Age 10.94 10.87 10.94 10.88 10.95 (-.07)* (-.007) (-.06)** Males (%) 50.7 51.2 51.9 50.7 50.6 (.01) (.01)* (.001) Family income 4510 5303 3876 6055 4371 (annual pesos pc) (931) (-494)** (1683)** Education of HH 7.46 6.08 4.74 6.81 7.67 Head (-1.58)** (-2.91)** (-.85)** Age of HH head 43.8 47.9 51.11 41.7 43.7 (4.23)** (7.37)** (-1.99)** % of children in hh 9.58 100 0 100 0 with mig. to the US % of children in HH 4.51 100 100 0 0 with rem. GDP per capita in mun. 5,717 3,544 3,805 4,374 5,964 (-2,420)** (-2,158)** (-1,589)** Gross fixed investment/ .037 .047 .034 .042 .037 gdp in mun. (.009)** (-.002)** (.005)** N 1,298,362 1,294 38,637 54,115 627,932 Source: Calculations done by the author using the 9.1% public sample of the 2000 Mexico census. Sample includes only children of the head of household, for whom information on their education, the education of the head, and the migration information in the households is available. **1 % significance level. Table A3, 2000 Mexican Census, Instrumental variables and regions 8 to 14 yr old 15 to 19 yr old Mig rate in municipality exc. hh i. .02 .02 (.06) (.06) Rem /Income in municipality exc. hh i. .02 .02 (.03) (.04) Capital/labor ratio in municipality 140.53 141.45 (188.31) (190.54) Border 11.49 11.75 North 11.92 11.70 Center 36.95 36.98 Capital 15.08 16.92 South 24.53 22.62 Source: 2000 Mexico census. Border: all Mexican states in the US border. North: Baja Sur, Nayarit, Zacatecas, Aguascalientes, San Luis Potosí. Center: Jalisco, Colima, Guanajuato, Michoacan, Queretaro, Hidalgo. Puebla, Tlaxcala, Veracruz, Morelos. Capital: DF, Edomex. South: all others. Remmitances. Cuecuecha. February 2008 37 Table A4 Variables used in the estimation of the pairwise propensity scores Family income Age of children Education of head of household Age of head of household Gross fixed investment in municipality GDP pc in municipality K/L in municipality Mig rate in mun. excluding family i Rem/GDP in mun. excuding family i Mig rate in municipality*Rem/GDP in mun. (excluding family i) · Table A5, Average Treatment Effect on the Treated, Pairwise comparisons IV method-linear probability model Combined Income Allocation LT migration Combined Effect Effect Effect Effect effect for LTM θ1,4 θ1,2 θ2,4 θ1,3 θ3,4 Males 15-19 6.63** 6.92* -.81*** 7.45* -.28*** (3.15) (3.23) (.27) (3.16) (.10) 8-14 5.36* 3.78 -.61*** 4.13* -.18** (3.16) (2.31) (.22) (2.26) (.09) Females 15-19 9.22** 9.67** -.88*** 10.11*** -.47*** (3.90) (4.00) (.29) (3.90) (.13) 8-14 7.95* 5.86* -.74*** 6.15** -.37*** (4.23) (3.00) (.23) (2.93) (.12) ***Significant at 1%. ** at 5%. *at 10%. Source: 9.1% public sample of the 2000 Mexico census. Sample includes only children for whom information on the edu., and age of the head of household is known. Remmitances. Cuecuecha. February 2008 Figure 1: 38 Remmitances. Cuecuecha. February 2008 Figure 2: 39