79426 AUTHOR ACCEPTED MANUSCRIPT FINAL PUBLICATION INFORMATION Impact of Remittances on Household Income, Asset and Human Capital: Evidence from Sri Lanka The definitive version of the text was subsequently published in Migration and Development, 1(1), 2012-10-25 Published by Taylor and Francis THE FINAL PUBLISHED VERSION OF THIS ARTICLE IS AVAILABLE ON THE PUBLISHER’S PLATFORM This Author Accepted Manuscript is copyrighted by the World Bank and published by Taylor and Francis. It is posted here by agreement between them. Changes resulting from the publishing process—such as editing, corrections, structural formatting, and other quality control mechanisms—may not be reflected in this version of the text. You may download, copy, and distribute this Author Accepted Manuscript for noncommercial purposes. Your license is limited by the following restrictions: (1) You may use this Author Accepted Manuscript for noncommercial purposes only under a CC BY-NC-ND 3.0 Unported license http://creativecommons.org/licenses/by-nc-nd/3.0/. (2) The integrity of the work and identification of the author, copyright owner, and publisher must be preserved in any copy. (3) You must attribute this Author Accepted Manuscript in the following format: This is an Author Accepted Manuscript of an Article by De, Prabal K.; Ratha, Dilip Impact of Remittances on Household Income, Asset and Human Capital: Evidence from Sri Lanka © World Bank, published in the Migration and Development1(1) 2012-10-25 http://creativecommons.org/licenses/by-nc-nd/3.0/ © 2013 The World Bank Impact of Remittances on Household Income, Asset and Human Capital: Evidence from Sri Lanka Prabal K. De The City College of New York Department of Economics and Business 160 Convent Avenue, NAC 5106B New York, NY 10031 Email: pde@ccny.cuny.edu Phone: +1 212 650-6208 Fax: +1 212 650-6341 pde@ccny.cuny.edu Dilip Ratha The World Bank 1818 H Street, NW Washington DC, 20047 Phone: +1 202 458 0558 Fax: +1 202 522-2578 Dratha@worldbank.org 1 Abstract This paper explores the developmental impacts of international remittance income on the recipient households. The empirical analysis proceeds in two parts. In the first part, we show that remittance income largely accrues to the families belonging to the bottom quintiles of the income distribution helping the recipient families move up the income ladder. In the second part, we show that remittance income has positive and significant effect on children health and education, but not on conspicuous consumption or asset accumulation. We argue that remittance income is targeted better and not as fungible as other sources of transfer income, as the senders closely monitor it. We use bias-corrected matching estimators to control for self-selection issues. JEL Classification: F24, I2, O15, O53 Keywords: Remittances, Development, South Asia, Migration, Asset Formation 2 Acknowledgements We are grateful to Richard Adams, Vidhi Chhaochharia, Priya Deshingkar, William Easterly, and Jonathan Morduch for valuable comments and Ihsan Ajwad for kindly providing with the data. All remaining errors and omissions are our own. The views expressed are those of the authors and should not be attributed to the World Bank. 3 Impact of Remittances on Household Income, Asset and Human Capital: Evidence from Sri Lanka 4 I. Introduction Developing countries received more than 325 billion dollars in remittances in 2010. In the recent past, international remittances have outpaced most traditionally important international financial flows such as official development assistance and foreign direct investment in some developing countries (Ratha, 2003; Yang, 2011). A large proportion of these remittance transfers occur at the household level when migrant workers send money to their families and friends living in their home countries. Anecdotal evidence on migrant workers supporting families and themselves and eventually climbing up the social ladder abound.1 Remittance flow is different from the other international financial flows such as official development assistance, foreign aid and foreign direct investments (FDI). While most of the development assistance is essentially official (though some of it can be construed as inter-agency flows such as flows from foreign to domestic non-governmental organizations), and FDI is private institutional in nature, remittance is a purely private household-level flow. The amount and the potential use of remittance income are often decided upon jointly by the sender and the recipient. Unlike earned (and some forms of unearned) income, the recipients often do not have the full discretion to spend the remitted money in an unrestricted way. In many cases, remittance income is tied to specific uses. Unfortunately, surveys rarely collect data on the both sides of a remittance transaction making direct information on the motive unavailable. Moreover, due to fungibility of income, survey subjects do not report different sources for different categories of expenditure. This lack of survey-based information forces us to infer on the use of remittance income from the observational data. 1 For instance, (Deparle, 2007, June 24) reports that for many Filipino families, migrant workers sustain their families better. He also reports that remittances sustain economic activities and precipitate political change by migrant money in the West African country of Cape Verde (Deparle, 2007, April 22). 5 However, not all remittance payments may be tied to specific uses and the recipient may treat the money in the same way as any unrestricted transfer income such as unemployment benefits or pension. Moreover, even if the money is sent for some intended use, the sender may not be able to monitor the recipient perfectly. The second issue gives rise to a moral hazard problem – remittance flows may have unintended negative consequences. It can potentially lead to unproductive dependency on transfer income, laziness and conspicuous consumption owing to a moral hazard problem (Chami et al., 2008). At the macro level, remittance income may impact real exchange rate in a way that may be detrimental to the health of the recipient economy (Amuedo-Dorantes and Pozo, 2004). The focus of the present study is to examine this polemic at the household level using detail household survey data and to make inference on the impacts of remittance in the context discussed above. We want to examine to what extent remittance flow is “developmental” in nature – does it help ameliorate poverty and contribute to children’s human capital formation? We also test the other conjecture in the literature that remittance recipient families accumulate durable assets meant for luxury consumption such as motor vehicles, and landholding.2 As the first step, we test the hypothesis that income is fungible so that the effects of remittance income will be no different from the effects of other income. Our data rejects this hypothesis. Then we go on to test if outcomes for the remittance-recipient families are significantly different from the non-recipient families. While we find that remittance income contributes to an increase in human capital accumulation among children, we do not find any evidence that it significantly increases household asset accumulation. The challenge in identifying the effects of remittance, as in any evaluation of a treatment, is that we do not observe one potential outcome for each agent – outcome of a recipient family in 2 Arguably, the definition of “luxury” item varies from society to society. In the context of a developing country like Sri Lanka in 2000, a car for personal consumption, for example, would be deemed a luxury item. 6 case it did not receive remittance and outcome of a non-recipient family in case it did receive remittance. Matching estimators allow us to estimate the unobserved potential outcome for each observation in the sample (and consequently, identify the effects of treatment, under certain assumptions). The critical assumption here is that the treatment is random for individuals and households with similar values of the covariates, so that we could use the average outcome of some similar individuals or households who were not treated to estimate the untreated outcome. In other words, for each individual, matching estimators impute the missing outcome by finding other individuals in the data whose covariates are similar but who were not exposed to the treatment. This identification strategy comes with the caveat that even after controlling for many covariates such as location, religion, and household characteristics, there may be individual unobserved characteristics of migrants that are not orthogonal to the migration and remittance- sending decisions. However, since we study households that are in the home country, individual characteristics of the migrants are less likely to affect the decisions their households beyond the “treatment” that is remittance sending. In the absence of experimental evidence, the other option is to look for instruments. There are several difficulties of using instruments in the current setting. First, unlike Latin American countries like Mexico, there is no long legacy of migration study and data collection process for Sri Lanka. Therefore, the popular instruments such as migrant networks in the destination country cannot be used here. Second, cross section data rules out difference-indifference types of estimates. Finally, instrumental variable estimates, in the absence of strong identifying natural experiments, are also prone to bias. As Mckenzie et al. (2006) note, a bias-corrected matching estimator, similar to the one we have used, also works as a second-best. There is a small literature on the impacts of international remittances in Sri Lanka. Deshingkar (2006) shows how international remittance income acted as an insurance flow for the Sri Lankan economy. As far as other South Asian countries are concerned, in India, Rajan (2004) and Mullick (2011) discuss the impacts of migration and remittances on the recipient economy. 7 One of the earliest papers that showed that remittance income helped poor people build some forms of assets was based on a dataset from Pakistan (Adams, 1998). Remittance-development literature has flourished since then. In later studies in Guatemala and Ghana, (Adams, 2004; 2006) found evidence that though international remittance income helped reduce the level, depth and severity of poverty; they had a greater impact on reducing the severity as opposed to the level of poverty, where the severity of poverty is measured by squared poverty gap. In two later papers, (Adams and Cuecuecha, 2010a; Adams and Cuecuecha, 2010b) documented the developmental impacts of remittances in Indonesia and Guatemala respectively. In Mexico, Amuedo-Dorantes and Pozo. (2011) find that for many households, remittance income helps in income smoothing. Walker & Brown (1995) found for the Tongan and Western Samoan migrant households that remittances were not used exclusively for consumption purposes and played an important role in contributing to both savings and investment in the migrant-sending countries. They also found that remittances were not driven exclusively by altruistic sentiments and the need for family support, but also, among some migrant categories, by the motivation to invest. There appears to be substantial scope for policy intervention on the part of Pacific Island governments to increase the flows of remittances into their economies. More recently, a study on the Pacific islands of Fiji and Tonga generally found that remittance income has led to a fall in poverty and economic inequality (Brown, 2008). Estimating the effects of international remittances, Edwards and Ureta, (2003) found in El Salvador that remittance income not only lowered the propensity to dropout from school, but also was more effective in doing so compared to non-remittance income – result similar to what we have found. Mansuri (2006) also found that remittance income helps children’s education, particularly for girls. In India, Mueller and Shariff (2011) has recently found similar positive effects of remittances, though such remittances were internal, rather than international. 8 Many of these studies suffer from identification problems arising out of endogeneity of remittance income. Endogeneity problems may arise due to both simultaneity bias and omitted variables. The decision and amount to be remitted may depend on the various outcome variables such as children’s education, asset building and changes in consumption pattern. Moreover, omitted variables may affect both remittance decision and outcome variables. A remitter may be a driven, enthusiastic and caring person who monitors her child’s education directly. Literature has often bypassed this issue because without randomized control trials, it is difficult to establish causality. Research using observational data has taken two routes – using instrumental variables that affect remittance, but not the outcome variables directly; and using matching estimators that estimate the differences in outcome between the recipient families and the non-recipient families that are similar based on observable characteristics. For example, in order to estimate the impact of remittance income on the household welfare for the overseas Filipino workers, variations in exchange rates arising out of the East Asian currency crisis were used as an instrument for remittance income and showed that remittance income helps reduce poverty and acts as an insurance payment (Yang & Choi, 2007; Yang, 2008). 3 Another popular instrument has been constructed on the intuition that historical events such as railroad construction (Adams and Cuecuecha, 2010a), and destinations of migrant workers (Amuedo- Dorantes and Pozo, 2011). Examples of the matching estimator can be found in Acosta, (2006) and Esquivel and Huerta-Pineda (2007) who take the second route in estimating the effects of remittance on poverty and education and Mexico and El Salvador, respectively. In the absence of either experimental or panel data, we employed two strategies to identify the effects of remittance income separate from other sources of income. First we 3 There is a rich literature on the impacts of international migration aside from remittances. However, even though they are closely related, the analysis at the household levels warrants different treatments. Many households in our sample do not qualify as a “migrant household” as there is no immediate family member living abroad, but receive remittances from friends and extended families. 9 estimated an over-identified model where both total income (including remittance income) and remittance income are included. The intuition is that if income is fungible, then remittance income should not show any additional significance in explaining the dependent variable over and above the effects of total income. Second, we use matching estimators to control for any systemic difference between remittance-receiving families and other families and examine if remittance-receiving families’ behavior is different from the non-recipients. Matching estimators have been widely used in the program evaluation. They impute the missing potential outcome by using average outcomes for individuals with “similar” values for the covariates. Identification issues have been discussed in more detail in the empirical section of the paper. We make two contributions to the literature. First, to our knowledge, this is one of the first studies that examine the behavior and impacts of remittances in South Asia. Second, our identification strategies help to test the fungibility of income and directly assess the impacts of remittances vis-à-vis other sources of income. We also improve upon the propensity score matching method so far used in this literature. Even though the methodology is not free from caveats such as roles played by unobserved heterogeneity, we believe that in the field of international migration where randomized trials are costly and difficult, such nuanced results from observational studies involving cross-section data advance our understanding. The rest of the paper is organized in the following way. In the next section we will discuss the international migration and remittance profile for Sri Lanka, followed by a brief review of the related literature. Section III discusses the dataset used in this paper and provides an exploratory analysis of the data. The following two sections describe the empirical strategy and results respectively. We conclude with a summary and policy implications of our results. II. International Migration and Remittance in Sri Lanka 10 The trend in out-migration in Sri Lanka started in the late 1970s as an effect of slow growth in the domestic economy and large-scale oil production in the Gulf countries that demanded a large number of unskilled labors.4 This trend was supplemented by a relatively recent trend in hiring female housemaids in those gulf countries. Table 1 shows these changing trends in terms of occupational mix for the migrants over time at a disaggregated level. Please insert Table 1 here Keeping up with the increasing migration, international remittance inflow has increased steadily for Sri Lanka over the past few years. It has outpaced the other two important sources of external finance such as official development assistance (ODA) and foreign direct investment (FDI). Figure 1 illustrates this trend. While both ODA and FDI have remained flat over time, international remittance flow has increased, even in the face of global recession. Please insert Figure 1 here Though parts of this significant increase is due to better recording of remittances in recent times and increasing tendency to send money through legal channels owing to reduction of remittance fees and improvements in technology, anecdotal evidence suggests that a significant amount of remittances still flow through informal channels and go unrecorded. Therefore, the official remittance figures potentially underestimate the actual extent of money transfer. III. The Dataset and an exploratory analysis of data The primary source of data for this study is the Sri Lanka Integrated Survey 1999-2000. The survey was conducted across all nine provinces in the country between October 1999 and the third quarter of 2000. The data is based on interviews of 7,500 households and includes data on 35,181 individuals. The survey is comprehensive and dependable and contains information on a 4 For an excellent review of the migration pattern from Sri Lanka, see (Sriskandarajah, 2002). 11 large number of variables such as demographics, occupation, income, education, health and asset holding allowing us to construct a dataset with various observable control variables. There is a separate module for international migration, where current and past migration information such as destination country, year of departure and family migration history are recorded. Remittance income information is recorded in two modules. In the migration module, families were asked about the amount received in the last 12 months as international remittances. Remittance income is also included in the category “other transfer income.” There is no perfect 1:1 matching between migrant and remittance receiving families – some migrant families do not receive any remittances, while some families with no immediate family members abroad do have remittance income. Please insert Table 2 here Table 2 shows the destination country profile for Sri Lankan migrants. It confirms the macro- level finding in table 1 that Gulf countries account for a majority of demand for migrant workers from Sri Lanka. In Table 3 we compare various demographic, education and socio-economic characteristics across families that receive remittances and families that do not. The top panel, Panel A, reports characteristics of the household head. Not surprisingly, a lower proportion of households are headed by a male among the remittance recipient families as the male member is likely to be abroad. The average age of the household heads for non-recipient families is higher. While non-recipient family heads have lower inheritance in terms of land, they have a higher education level (measured in terms of the years of schooling). Interestingly, there are no significant differences between the remittance-recipient families and other families in terms of either income or the value of landholdings. Please insert Table 3 here 12 In terms of characteristics, there are no significant differences between the children of recipient vs. the children of non-recipient families in terms of enrollment and dropout (previously enrolled, but not currently enrolled). However, recipient family children access private tuition at a significantly higher rate. We control for most of these observed differences in our empirical analysis. IIIa. Income Mobility This section analyzes the income dynamics of Sri Lankan households who received remittance income from abroad. If remittance income is received by families that are already wealthy, it may contribute little to upward income mobility or reduction in inequality. On the other hand, remittance income for poor families can help them climb up the income ladder. Suppose we divide all the families into 10 income deciles according to their pre- remittance income such that the first decile contains the poorest 10% families and the tenth decile contains the richest 10%. Let us now include the remittance income for all the remittance- receiving families and re-draw the income distribution according to that income data. A straightforward measure of mobility will be the following – number of families in each decile from 1-9 that have moved up the income ladder according to the new distribution. Please insert Figure 2 here The basic matrix illustrating this information is presented in Figure 2. This matrix compares the income decile of a recipient family before receiving remittances with its income decile after receiving them. Each row and column in figure 1 represents one decile of income distribution. The rows represent the initial income decile the recipient families belong to as if they did not receive any remittance income. The entries along the diagonal show those recipient 13 households who remained in the same decile before and after receiving remittance income. Off- diagonal entries show those families that moved between deciles. For example, the entry at the top of the first column indicates that 23 (16.3%) recipient households who were in the lowest decile before receiving remittances remained in the lowest quintile even after receiving it. This means 83.6% of the recipient families from the lowest quintile moved up in the income ladder after receiving remittances. Similarly, the second entry along the diagonal (i.e., second row, second column) shows that 28% =(16/57)*100) of the families from the second decile remained in the same strata after receiving remittances and 72% of these families moved to a different decile. Figure 3 summarizes the data on upward mobility – percentages of families moved up in the income ladder after receiving remittance income. In this figure the bottom label indicates which decile the families came from, and the bars show what percentage of the recipient families ended up in a higher decile. Figure 3 shows that though remittance income contributes to income mobility for families in all income strata, it is more pronounced for the lower half of the distribution. This is not surprising, as all families got a boost income from the baseline of no remittance income. Please insert Figure 3 here The problem in this type of analysis is that the potential income in the absence of migration is not controlled for. Migrants sending money from abroad would most likely have earned income in their home country if they had not migrated. In other words, we cannot observe the counterfactual income in the event that a migrant had not migrated. There is no perfect methodology to create counterfactual income, particularly since we have a cross section of data and no history of wage or other earning for the migrant workers. There have been two excellent attempts at estimating the counterfactual income. Barham and Boucher (1998) and Adams and Cuecuecha (2010a) create model-based prediction of income 14 without migration. Since we do not have data to follow such methodology, we have adopted an ad hoc method of assigning counterfactual income to a migrant family – we assign to the family the median income of the income decile of the group to which they belong to according to the distribution of pre-remittance income. In particular, we follow the following two-step procedure to create a more refined measure of income mobility. In the first step, we create the income distribution of all families according to their pre-remittance income (for families not receiving any remittance income, this is same as the total household income). We then calculate the median income of each decile as the representative income of a particular income group. Next, for each income group, we add the median income to the pre-remittance income of the recipient families belonging to the respective income group. This income stream constitutes the counterfactual income (or potential outcome in the nomenclature of (Rosenbaum & Rubin, 1984)). In the second step, we perform the previous exercise summarized in figure 1 and 2, but on the stream of counterfactual income and actual income. We create ten deciles of counterfactual income, locate the remittance-recipient families in various income strata, recalculate the income distribution according to the post-remittance income, and calculate the percentage of families that have moved up the income ladder according to the latter distribution vis-à-vis the counterfactual distribution.5 Please insert Figure 4 here The results are summarized in figure 4. In this version of the mobility graph, we see that families in the lowest decile have the highest incidence of upward mobility. The overall message 5 Note that while calculating the income distribution, all families are included regardless of their remittance status. Also note that families can potentially go into a lower income decile while being positioned in the actual income distribution vis-à-vis the counterfactual one as median income may be higher than the actual remittance income. This cannot be ruled out on reality also – sometimes migrants find that income in their destination country is less than potential income foregone in the home country. 15 from this graph is that remittance income helped the poorer sections of the society leading to an amelioration of inequality. IV. Estimating the Impacts of Remittance Income on Education, Health and Asset Accumulation IVa. Specification for Overidentified Regressions In examining the effects of remittance income on children’s welfare in terms of education and health, we start with a linear specification where the dependent variable will be various individual welfare measures representing health and education. On the right hand side of the regression equation, we have remittance income, total income and a set of control variables. Therefore, our basic regression specification is (1) Yi = µ + β.Ri + Xi ´λ + εi where Yi is the dependent variable in question for individual i, Ri is remittance income, and Xi is a vector of other covariates. The coefficient of interest throughout is β, the effect of remittance income of individual welfare. Finally, εi is the random error term. Depending on the outcome variable, we use either ordinary least squares or Probit model. IVb. Outcome Variables We study two different classes of outcome variables- children’s human capital in terms of education and health, and family asset accumulation in terms of value of durable assets and land. Children’s human capital – health and education In the first class, we use the anthropometric measure of weight of a child less than 5 years of age. Anthropometric measures such as weight, height and body mass index (BMI) are becoming increasingly popular in the development literature as they give a direct signal about individual 16 health and welfare. Further, since we have only cross section data it makes less sense to work with the adults as marginal impact on their health indicators seems to be less significant for a year. Our second measure is whether or not a school going child receives private tuition. The rationale behind this measure lies in the success of school education system in Sri Lanka. Sri Lanka has almost complete literacy and enrollment in school as seen in Table 2 and 97% of the students attend government-run schools. Therefore, traditional measure of enrollment is not effective here. Having a private tutor in this environment signals superior access to education. Unfortunately, we do not have any test score data to see if having a tutor directly translates itself into good scores. Since the outcome variable takes only binary values, we estimate a probit model. Household Asset Accumulation and Land Holding As discussed in the introduction, it has been argued that remittance income, being transfer income, goes into conspicuous consumption as idle asset building. Also, if the rich landed class of the society receives remittances and remittance contributes to further accumulation of such assets, it cannot be deemed as a development flow. Therefore, we use the measures of the value of durable assets and land holding to see if the effects of remittance income are different than those of the children’s variables. IVc. Explanatory variables Independent variables vary depending on the outcome variables in the model concerned. We use a set of individual characteristics as well as household level characteristics. For every individual, we control for gender and age. We also control for family income, parental education, gender of the head of the household and remittance income. 17 To recall, according our identification strategy described above, if income is fungible and remittance income has no effects over and above the effects of total income including remittances, we should have β = 0. If β, however, is significantly different from zero, the aforementioned hypothesis cannot be rejected.6 IVd. Specification for Matching Estimators Same set of outcome variables have been used in computing the matching estimators. To recall the basic logic of the matching estimators, each individual or household head belongs to either of the two groups – (i) remittance-recipient or (ii) remittance non-recipient. Therefore, for each individual, there is an actual outcome and there is a potential outcome that we cannot observe. For example, potential outcome for a remittance-recipient child will be the one when she did not belong to a recipient family. Econometrically, the potential outcome for unit i is obtained by imputing the average of outcomes for its matches and then the difference between the treatment and control group is computed as an average treatment effect. Matching Variables We use the same set of variables that have been used as control variables in our regression analysis above. IVe. Identification issues An ideal framework for assessing the causal effects of remittances would be to conduct an experimental trial in which individuals would be randomly assigned into international migration and stipulate that they send remittances to their families back home. Then families of 6 An equivalent way of testing this: we can take income without remittances and remittance income on the right hand side and test the null that the corresponding coefficients are equal. We have taken this approach because it’s easier to test and interpret the null β = 0. 18 these migrants can constitute the “treated” group as remittance-receiving families, and other non- recipient families can serve as controls. Practical considerations preclude such experiments in most of the cases.7 Problems with non-experimental, observational approaches are that the effects of remittances are confounded by omitted variables that influence both outcome variables and other control variables. For instance, a father may seek and obtain a job abroad because he cares about his children and makes sure to both send money home and monitor his child’s progress regularly. There are reverse causality issues also. How much money someone sends may depend on the quality of children’s health and education. As an alternative to the experimental approach, several non-experimental methods have been proposed. This includes using an instrumental variable for migration and Heckman two-step procedure.8 Though these studies deal mostly with migration and not remittances, and as we have argued above, they are not identical. There are two exceptions Acosta (2006) and Esquivel & Huerta-Pineda (2007) find positive effects of remittances on reducing poverty among Mexican households and on education in El Salvador respectively using propensity score matching among other methods. As discussed below, we attempt to improve upon simple propensity matching methods by using more reliable bias-adjusted matching estimators. In the absence of true exogenous variation that credibly serves as an instrument, we adopt two alternative identification strategies. Our first identification strategy stems from the intuition that remittance income flow is a special flow. Migrants who send money tend to monitor it closely. Therefore, it is not the same as, say, other transfer income a family receives. This leads us to an overidentification test. If income is fungible and a dollar is a dollar (in this case, Sri Lankan Rupee), then remittance income should have no impact on the outcome variables once the 7 There are such experiments where potential migrants apply through visa lotteries. 8 See Mckenzie (2006) for various approaches. 19 total income including remittances is controlled for. This strategy is similar to Thomas (1990), where he compared the effects of female vs. male income. Our second identification strategy is to use the latest matching estimators. The major advantages of a matching procedure are that it does not require parametric functional form and exclusion restrictions. A matching estimator is based on a simple idea: for each recipient family, or a child belonging to a recipient family, the procedure finds a group of comparable families or individuals who have similar observable characteristics among the non-migrants. Within each set of matched individuals or families, one can then estimate the impact of remittances by the difference in the sample means. Therefore, matching estimators approximate the virtues of randomization mainly by balancing the distribution of the observed attributes across remittance- recipients and non-recipients. Dehejia & Wahba (1999; 2002) showed that matching provides a significantly closer estimate for the treatment effects than the standard parametric techniques. We use the latest matching estimators developed by (A. Abadie & Imbens, 2002; A. Abadie, Drukker, Herr, & Imbens, 2004) with and without correction for bias. To our knowledge, this is the first attempt at estimating the effects of remittances in this framework. V. RESULTS Va. Results from Overidentified Models Table 4 reports results from the estimation of the first set of overidentified models where the dependent variables are weight of a child less than 5 years of age and private tuition decision for the students currently enrolled respectively. Table 4 shows us that weights of children less than 5 years of age are positively correlated with age and male sex. Female-headed households seem to have a positive significant impact on the health of the child. This is consistent with the findings of the intra-household 20 decision making literature. Further, while parental education has positive and significant impact on child health as expected, father’s education has a stronger effect. Finally, total remittance income has a positive and significant impact on child health even after controlling for total income providing evidence that income may not be fungible between remittance and non- remittance income.9 Please insert Table 4 here If income is not fungible, remittance income is attached to specific type of expenditure and remittance income is used to accumulate expensive assets such as durable goods and land, then we would expect similar results for durable assets and landholding. However, as Table 5 shows, this is not the case. Table 5 shows the results from the specification where the dependent variables are values of a household head’s durable assets and landholding respectively. However, none of these models took into account that remittance recipient families may potentially be different from the non-recipient families. Matching estimators attempt to correct this and provide a more accurate estimate of the impacts of remittance income. Please insert Table 5 here Vb. Results from Matching Estimators In this section we show estimates for the ATT using different specifications. Table 6 shows results for children’s human capital variables when matching is done on a wide set of covariates. We have used two groups of covariates for two dependent variables respectively. For the private tuition decisions, we have used individual characteristics such as age, gender, and current grade and region dummies. Covariates used for matching children less than 5 years of age consist of similar variables. 9 We also tested an alternative specification where we used remittance income dummy (=1 if remittance receiving family, zero otherwise). 21 The simple matching estimator will be biased in finite samples when the matching is not exact. The bias-corrected matching estimator (nnmatch using the biasadj() option) adjusts the difference within the matches for the differences in their covariate values. Therefore, Table 6 also includes biased-corrected matching estimators along with the simple matching estimator. This also provides a robustness test. Finally, we have used 4 matched across all specification as is the general norm. The results are robust to alternative number of matches. For the specifications at hand in Table 6, we conclude that the treatment effect or the effect of belonging to a remittance-recipient family is significantly greater than zero at the 1% or 5% levels for both the dependent variables at hand. Please insert Table 6 here Similar analysis has been performed for the asset variables and as Table 7 shows, remittance income has no impacts on asset holding, implying that there is little evidence to support the view that remittance-recipient families accumulate assets using remittance income. Please insert Table 7 here Vc. Sensitivity Analysis The previous estimates do not control for the problem of covariate imbalance – when the treatment and the control groups are observational, rather than being carefully designed for an experiment, values for pre-treatment variables may differ widely making the groups incomparable. By design, matching estimators try to exploit the covariates to find treatment and control units that are similar or balanced. However, there are three problems with this. First, in large datasets, the covariate set can be so large that meaningfully matched sample may be small. 22 Second, the process first chooses the matching method, produces estimates and then checks the resultant covariate balance. Third, it is model-dependent. Iacus, King and Porro (2008) and King et al. (2010) propose a solution, Coarsened Exact Matching (henceforth CEM), which attempts to create ex ante covariate balance. The procedure temporarily coarsens each variable into several strata, matches on these coarsened data and then only retains the original (uncoarsened) values of the matched data. It offers several advantages. First, the process involves creating matched data before the estimation process and hence model- independent (Ho et al., 2007). A simple average treatment effect can be estimated with the matched data now that the empirical distributions of the covariates in the groups are more similar. Second, in this process, adjusting the imbalance on one variable has no effect on the maximum imbalance of any other. However, the procedure of CEM nonetheless suffers from the same problem that matching estimators suffer from – there is a trade-off between the size of matching bin and the size of matched sample. While larger bins ensure more matched units, it creates less precise matches. In view of this, we believe the use of CEM will provide an excellent robustness test for our previous results coming from more conventional and widely used matching estimators. Please insert Table 8 here The results are summarized in Table 8. As seen from the sign and significance of the coefficients, the results are qualitatively similar to our previous results. The Average Treatment Effect (ATT) of remittance receipt is positive and significant for recipient family children’s health and education variables. However, the such effect is not significant for the various classes of assets. VI. Conclusion 23 International remittances have become a globally important financial flow in recent years. However, evidence on the developmental consequences of this essential transfer income at the household level is mixed. The questions we ask in this paper are whether remittance income (i) helps recipient families in moving up in the income ladder, (ii) positively and significantly affects children’s health and education and (iii) is spent on buying luxurious durable assets and land. Using very detailed household level data from the Sri Lankan Integrated Survey, we find that remittance income does help in income mobility and children’s human capital accumulation. However, we find no evidence that households use remittance income in building assets. Our identification is based on using bias-corrected matching estimators, whereby children from non- recipient families that are “similar” to recipient families in terms of observable covariates are compared to their brethren belonging to the latter group. To our knowledge, this is one of the first papers to discuss the implications of international remittances in the context of a South Asian country. As shown in the paper, both international migration and remittance inflow have been steadily growing in Sri Lanka making it an important source of foreign currency. This is also true for other South Asian countries such as India and Pakistan. Since these countries in the subcontinent are similar in many cultural, religious and economic aspects, our results are likely to have some external validity. Our evidence on the positive effects of remittance income calls for policies that ease remittance flow – reduction in fees or tax breaks. Future research should focus on better data collection, particularly longitudinal data. Surveys linking both segments of the family living in Sri Lanka and abroad will also be very useful in answering some of the questions that we could not address – monitoring of remittance income, response of remittance flow to family income shocks and children’s human capital formation before and after migration and remittance receipt. 24 References Abadie, A., & Imbens, G. W. (2002). Simple and bias-corrected matching estimators for average treatment effects. NBER Technical Working Paper 28. Abadie, A., Drukker, D., Herr, J. L., & Imbens, G. W. (2004). Implementing matching estimators for average treatment effects in Stata. Stata Journal, 4(3), 290-311(22). Acosta, P. (2006). Labor supply, school attendance, and remittances from international migration: The Case of El Salvador. Policy Research Working Paper Series, Adams, J.,Richard, & Cuecuecha, A. (2010a). The economic impact of international remittances on poverty and household consumption and investment in Indonesia. SSRN eLibrary. Adams Jr., R., & Cuecuecha, A. (2010b). Remittances, household expenditure and investment in Guatemala. World Development, 38(11), 1626-1641. Adams, Jr.,R. (2006). Remittances and poverty in Ghana. SSRN eLibrary. Adams, Jr.,R. (2004). Remittances and poverty in Guatemala. SSRN eLibrary. Adams, R., Jr. (2005). Do international migration and remittances reduce poverty in developing countries? World Development, 33(10), 1645-1669. Adams, R., Jr. (2009). The determinants of international remittances in Developing countries. World Development, 37(1), 93-103. 25 Adams, R. (1998). Remittances, investment and rural asset accumulation in Pakistan. Economic Development and Cultural Change, 47(1), 155-173. Amuedo-Dorantes, C., & Pozo, S. (2004). Workers' remittances and the real exchange rate: A paradox of gifts. World Development, 32(8), 1407-1417. Amuedo-Dorantes, C., & Pozo, S. (2006). Remittances as insurance: Evidence from Mexican immigrants. Journal of Population Economics, 19(2), 227-254. Amuedo-Dorantes, C., & Pozo, S. (2010). Accounting for remittance and migration effects on children's schooling. World Development, 38(12), 1747-1759. Amuedo-Dorantes, C., & Pozo, S. (2011). New evidence on the role of remittances on healthcare expenditures by Mexican households. Review of Economics of the Household, 9(1), 69-98. Amuedo-Dorantes, C., & Pozo, S. (2011). Remittances and income smoothing. The American Economic Review, 101(3), 582-587. Barham, B., & Boucher, S. (1998). Migration, remittances, and inequality: Estimating the net effects of migration on income distribution. Journal of Development Economics, 55, 307- 331. Brown, R. P. C. (2008). Remittances and development in the pacific: Effects on human development in Fiji and Tonga. Paper presented at the Bangkok, ESCAP, EGM Meeting on Migration and Development, Chami, R., Barajas, A., Cosimano, T., Fullenkamp, C., Gapen, M., & Montiel, P. (2008). Macroeconomic consequences of remittances. IMF Occasional Paper 259, 26 Cox, A., & Ureta, M. (2003). International migration, remittances, and schooling: Evidence from El Salvador. Journal of Development Economics, 72, 429-461. Dehejia, R. H., & Wahba, S. (1999). Causal effects in non-experimental studies: Re-evaluating the evaluation of training programs. Journal of American Statistical Association, Volume 94, Number 448(448), 1053-1062. Dehejia, R. H., & Wahba, S. (2002). Propensity score-matching methods for nonexperimental causal studies. Review of Economics and Statistics, 84(1), 151-161. Deparle, J. (2007, April 22, A good provider is one who leaves. The New York Times, Retrieved from http://www.nytimes.com/2007/04/22/magazine/22Workers.t.html Deparle, J. (2007, June 24, Border crossings: In a world on the move, a tiny land strains to cope. The New York Times, Retrieved from http://www.nytimes.com/2007/06/24/world/africa/24verde.html Deshingkar, P. (2006). Remittances in crisis: Sri lanka after the tsunami. HPG Background Paper. Esquivel, G., & Huerta-Pineda, A. (2007). Remittances and poverty in Mexico: A propensity score matching approach. Integration and Trade, 27, 45-71. Hildebrandt, N., & Mckenzie, D. J. (2005). The effects of migration on child health in Mexico. Economia, 6(1), 257-89. Ho, D. E., Imai, K., King, G., & Stuart, E. A. (2007). Matching as nonparametric preprocessing for reducing model dependence in parametric causal inference. Political Analysis, 15(3), 199-236. 27 Iacus, S., King, G., & Porro, G. (2008). Matching for causal inference without balance checking. SSRN Archive King, G., Blackwell, M., Iacus, S., & Porro, G. (2010). Cem: Coarsened exact matching in stata. Stata Journal, 9(4), 524-546. Lokshin, M., Bontch‐Osmolovski, M., & Glinskaya, E. (2010). Work‐Related migration and poverty reduction in Nepal. Review of Development Economics, 14(2), 323-332. Mallick, H. (2011). Inflow of remittances and private investment in India. Singapore Economic Review. Mansuri, G.. (2006). Migration, school attainment, and child labor: Evidence from rural Pakistan. Policy Research Working Pape 3945, World Bank, Development Research Group. McKenzie, D., Gibson, J., & Stillman, S. (2008). How important is selection? Experimental vs. non-experimental measures of the income gains from migration. IZA Discussion Paper 2087. Bonn, Germany. Mueller, V., & Shariff, A. (2011). Preliminary evidence on internal migration, remittances, and teen schooling in India. Contemporary Economic Policy, 29(2), 207-217. Rajan, Irudaya. (2004). Dynamics of international migration from india: Its economic and social implications. The ICFAI Journal of Employment Law 2 (II): 25-46. 28 Ratha, D. (2003). Workers’ remittances: An important and stable source of external development finance. Global Development Finance, 157, 75. Rosenbaum, P. R., & Rubin, D. B. (1984). Reducing bias in observational studies using subclassification on the propensity score. Journal of the American Statistical Association, 79(387), 516-524. Siddiqui, T., & Chowdhury, A. (2003). Migrant worker remittances and microfinance in Bangladesh. Working Paper no. 38, Social Finance Programme International Labour Office, Dhaka, Bangladesh Sriskandarajah, D. (2002). The migration–development nexus: Sri Lanka case study. International Migration, 40(5), 283-307. Thomas, D. (1990). Intra-household resource allocation: An inferential approach. The Journal of Human Resources, 25(4), 635-664. Walker, A., & Brown, R. P. C. (1995). From consumption to savings? interpreting Tongan and western Samoan sample survey data on remittances. Asian and Pacific Migration Journal, 4(1), 89-116. Yang, D. (2008). International migration, human capital, and entrepreneurship: Evidence from Philippine migrants' exchange rate shocks. Economic Journal, 118, 591-630. Yang, D. (2011). Migrant remittances. Journal of Economic Perspectives, 25(3), 129-52. Yang, D., & Choi, H. (2007). Are remittances insurance? Evidence from rainfall shocks in Philippines. World Bank Economic Review, 21(2), 219-248. 29 Figure 1: Various Sources of External Finance for Sri Lanka Source: World Development Indicator Database 30 Pre- remittance Post-remittance Income Decile Income Decile 1 2 3 4 5 6 7 8 9 10 Total 1 23 23 14 20 16 17 12 8 7 1 141 2 1 16 11 9 4 8 2 3 2 1 57 3 8 8 16 8 10 6 5 1 62 4 6 7 6 6 4 3 32 5 7 13 14 12 5 2 53 6 1 9 16 16 6 2 50 7 14 27 21 4 66 8 1 25 32 8 66 9 1 25 28 54 581 Figure 2: Matrix of Initial income decile and post-remittance income decile 31 Figure 3: Upward Income Mobility among Remittance Receiving Families 32 Figure 4: Upward Income mobility with projected counterfactual income 33 Table 1: Migration Patterns in Sri Lanka Occupational 1980 % 1992 % 1995 % 2000 % Group High_level 2517 6 1245 1 887 <1 983 <1 Middle-Level 4116 10 6225 5 7070 4 10203 6 Skilled 11964 28 22409 18 26806 16 36028 20 Unskilled 17681 41 9960 8 23469 14 35087 19 Housemaids 6467 15 84655 68 114208 66 98636 54 Total 42745 100 124494 100 172440 100 180937 100 Source: Sriskandarajah (2002) Notes: For a complete list of occupational categories, please see Sri Lanka Bureau of Foreign Employment statistics http://www.slbfe.lk/article.php?article=68 34 Table 2: Destination of International Migrants from Sri Lanka Destination Freq Percent KUWAIT 150 30.4 SAUDI ARABIA 117 23.7 U.A.E (DUBAI, 44 8.92 LEBANON 41 8.32 EUROPE 40 8.11 OTHER MIDDLE 33 6.69 EASTERN BAHARAIN 13 2.64 AMERICA 11 2.23 JORDAN 10 2.03 MALDIVE ISLANDS 8 1.62 OMAN 6 1.22 SINGAPORE 4 0.81 OTHER ASIAN 4 0.81 COUNTRIES JAPAN 3 0.61 INDIA 2 0.41 SYRIA 2 0.41 HONG KONG 1 0.2 IRAN 1 0.2 IRAQ 1 0.2 AFRICA 1 0.2 AUSTRALIA 1 0.2 Total 493 100 Source: Authors’ calculations based on Sri Lanka HIS 1999-2000 35 Table 3: Descriptive Statistics Variable Non-Recipient Recipient t-stat p-value Panel A: Characteristics of Household Heads Male 0.84 0.76 5.20*** 0.00 (0.37) (0.43) - Age 49.20 50.63 2.56*** 0.01 (13.56) (13.81) Married 0.82 0.76 3.82*** 0.00 (0.39) (0.43) Father owned land 0.33 0.25 4.57*** 0.00 (0.47) (0.43) Education (years of schooling) 7.77 8.01 -1.71* 0.09 (3.31) (3.17) Value of Land Holding 209342.88 193175.00 0.27 0.79 (627740.59) (689829.07) Income in current LKR 104654.15 57377.97 0.84 0.40 (1376381.60) (105693.56) Panel B: Characteristics of Children of age 6-18 Age 12.31 12.56 -1.80** 0.07 (3.76) (3.72) Male 0.51 0.50 0.78 0.43 (0.50) (0.50) - Private Tuition 0.33 0.46 7.37*** 0.00 (0.47) (0.50) Current Enrollment 0.87 0.86 1.17 0.24 (0.34) (0.35) Never School 0.02 0.02 0.10 0.92 (0.13) (0.12) Standard Errors are in parenthesis *** p<0.01, ** p<0.05, * p<0.1 Notes: Land holding and income are in current Sri Lankan rupees. 36 Table 4: Effects of Remittances on Children’s Human Capital (1) (2) Private Tuition VARIABLES Dummy Weight of children less than 5 years age 0.007 2,549.473*** [0.019] [113.694] age squared -211.557*** [27.071] Male Dummy -0.090** 408.523*** [0.040] [94.947] Current Class 0.126*** [0.019] Father Education 0.069*** 38.339* [0.008] [20.264] Mother Education 0.059*** 27.125 [0.008] [20.518] Male Head of Family -0.352*** -20.747 [0.095] [138.709] Log of Total Income 0.126*** 160.656*** [0.023] [57.190] Log of Remittance Income 0.041*** 45.187** [0.008] [21.348] Observations 4694 1500 R-squared 0.644 Robust standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 37 Table 5: Effects of Remittances on Asset Accumulation (1) (2) Value of Value of Durable Land VARIABLES Goods Holding Male 915.160 6,266.117 [6,619.942] [43,337.527] Dummy for Father Owning Land 38,967.110*** 62,522.382 [10,889.917] [39,678.901] Highest Education 11,545.205*** 11,554.864** [1,516.521] [5,544.962] Log of total Income 88,062.045*** 103,687.170* [21,839.486] [53,145.946] Log of total remittances 2,989.627 -6,882.295 [1,974.347] [4,298.829] Observations 5193 1052 R-squared 0.105 0.045 Robust standard errors in brackets *** p<0.01, ** p<0.05, * p<0.1 38 Table 6: Matching Estimators for Children’s Human Capital Estimator M Private Tuition Child Weight ATT (S.E) ATT (S.E) Simple 4 0.150267*** (0.03) 493.1441** (246.52) Matching Bias-Adjusted 4 0.149297*** (0.03) 424.6501** (227.20) No. of 4835 1536 observations Standard Errors are in parenthesis *** p<0.01, ** p<0.05, * p<0.1 39 Table 7: Matching Estimators for Asset Accumulation Estimator M Durable Asset Value of Land Holding ATT (S.E) ATT (S.E) Simple Matching 4 16916.76 (12021.74) -5078.658 (4581.88) Bias-Adjusted 4 17346.96 (12021.74) -5382.176 (4581.88) No. of 5342 5411 observations Standard Errors are in parenthesis *** p<0.01, ** p<0.05, * p<0.1 40 Table 8: Results with Covariate Balancing Child weight Private durable asset land Tuition holding ATT 708.2235*** 0.144904*** 17961.05 0.0129504 Std. Err. 285.5364 0.022578 15013.26 0.0184446 Standard Errors are in parenthesis *** p<0.01, ** p<0.05, * p<0.1 Note: Average Treatment Effect (ATT) was estimated comparing remittance recipient families (treatment) with non-recipient (control) groups having created the covariate balance. 41 42 Endnotes 43