2.1. International Migration and Remittances

Migration provides a safety net against adverse shocks or market failures in transnational households (^{de Haas, 2010}). For example, The New Theory of Migration groups the causes of migration processes into three hypotheses. First, families seek to improve their socio-economic conditions. Second, migra­tion may be the product of individuals’ desire to conduct income-generation activities (e.g., investments in education, agriculture, among others). Finally, migration can be caused by adverse shocks experienced by a household (e.g., social violence, natural disasters, among others).

The United Nations Department of Economic and Social Affairs (DAES) reports that by 2017, the number of migrants reached 258 million, compared to 173 million in 2000. One of the main economic results of these migratory flows has been the money transfers made by people to their families or friends in their places of origin. According to World Bank data, these flows reached a record high in 2018, with Latin America and the Caribbean, reaching a total of US$88 billion.

Colombia is no stranger to the impact of international migration and remit­tances. The Colombian Ministry of Foreign Affairs, estimates that by 2019 around 4.7 million Colombians were outside the country. The main destination countries include the United States, Spain, Chile, Ecuador, and Venezuela. This migratory flow has translated into a gradual increase in international remit­tances to Colombia (See Figure 1).

Nowadays, remittances represent one of the primary foreign exchange inflows for Colombia, even above some of the leading traditional export products, such as coffee, flowers, bananas, and textiles (^{Ramírez & Mendoza, 2013}). By 2012, Colombia’s central bank, Banco de la República, the institution in charge of monitoring international remittances, reports the United States (35 percent), Spain (29 percent), and the Bolivarian Republic of Venezuela (10 percent) as the leading countries in which remittance income originates. Meanwhile, the main destinations for remittances in Colombia are the regions of Valle del Cauca (28 percent), Antioquia (15 percent), and Cundinamarca (13 percent) (^{Asobancaria, 2013}).

According to the EECV for Santiago de Cali, the capital city of Valle del Cauca, remittances are mainly destined to recurring household expenses. Households report using this income primarily for household expenses (82 percent), which includes rent payments, utilities, food, health and clothing; home purchase (6 percent); education (5 percent); home improvement, vehicle purchase, or establishing a business (2.52 percent); other expenses (4.48 percent).

2.2. Child Labor

The Colombian Institute of Family Welfare, Instituto Colombiano de Bienestar Familiar (ICBF), is the state institution in charge of the protection of early childhood, adolescence, and the well-being of families in Colombia. This institution defines child labor as any activity, paid or unpaid, at the service of another person, by a child or adolescent under the age of 18, which, con­sequently, prevents their education and full development. According to data from the National Department of Statistics (DANE), by 2012, 10.2 percent of children/adolescents in Colombia under the age of 18 reported to be working.

It is essential to highlight that the Childhood and Adolescence Code states that children and adolescents must be protected against the worst forms of child labor and that the minimum age for admission to work is 15 years old. Children must be authorized by the Labor Inspection, or in its absence, the Local Territorial Entity (^{Law 1098 of 2006}). However, boys and girls under the age of 15 may receive authorization to perform remunerated artistic, cultural, recreational, and sports activities. The permission will establish the maximum number of hours and prescribe the conditions in which this activity must be carried out. In no case will the permit exceed fourteen (14) hours per week. Nevertheless, (^{Cuadros-Menaca & Gaduh, 2019}) report that these laws are not perfectly enforced, and a significant number of children start working in income-generating activities at the age of 12.

Figure 1. Remittance Flows to Colombia 2000-2018Notes: Remittance data obtained from Banco de la República de Colombia. 

To prevent and eradicate child labor, the local authorities in Santiago de Cali have created a strategy called “Cali, a territory free of child labor.” This pro­gram implements actions that aim not only to prevent child labor but also to commit society and the participation of private and public institutions to fighting this problem. According to the EECV, around 6.13 percent of children aged 12 to 18 reported to be working.

3.1. The Survey

We used the Encuesta de Empleo y Calidad de Vida (EECV-2012) applied in Santiago de Cali, and representative of the 22 districts of the city and its rural area. The purpose of the survey is to identify the labor situation of individuals in the city and design statistics regarding the population’s quality of life. The survey collects demographic characteristics for each household member and household characteristics, including housing assets and migration information.

3.2. The Sample

The total sample in the EECV consists of 30,458 individuals. Given that we are interested in child labor participation, we use the labor participation module that collects information for individuals over 12 years of age. We consider the definition of child labor and the General Law of Education described in the previous section. As such, we include children under 18 years of age, accord­ing to the ICBF definition, which, at the same time, captures grade repetition and late starters in the educational system. Our final sample includes 3,560 children/adolescents aged between 12 and 18. The total number of house­holds is 2,752.

3.3. Child Labor, School Attendance, and Remittances

We create two dummy variables to measure child labor and school attendance. The latter takes the value of one if the child reports to be participating in some market activity, which involves paid and unpaid work, domestic service, and family work. We construct our second dependent variable, attendance, with a dummy that takes the value of one if the child reports to be attending school.

As we described in Section 2, international remittances are a valuable source of income for households in the Valle del Cauca region. The EECV collects infor­mation about whether each member of the household receives remittances. We follow the unitary household model to build the main independent variable as a binary variable that takes the value of one if the child lives in a household where one of the members received remittances during the last twelve months.

The statistics from the EECV report that in terms of child labor, around 2 per­cent of the children in recipient households report to be working, while this statistic in non-recipient households is 6 percent. On the other hand, 86 percent of children in recipient households report to be attending school compared to the 84 percent in non-recipient households. These descriptive statistics suggest that remittances may improve children’s welfare in the remittances-receiving household, decreasing their exposure to child labor and increasing schooling.

3.4. Other Independent Variables

Other explanatory variables correspond to characteristics of the child, the head of household, and the household. In the child characteristics group, we include the age and gender of the child. For the set of characteristics relat­ing to the head of the household, we include the years of education and two dummy variables: gender and employment status. For the set of household characteristics, we include the number of adults (members over 18 years of age) and children (members under 12 years of age) in the household and an asset index. For the latter, we use assets and access to certain services to mea­sure the household standard of living.

We use the following variable to construct the index: the availability of elec­tricity, access to water services, sanitation installation, the type of material used for the floor, and the possession of certain items such as telephone and a television. Table 1 shows the list of variables used to construct the index. According to ^{McKenzie (2005)}, the main advantage of constructing an asset index instead of using income level to measure the standard of living lies in the quality of the measurement. This variable is likely to help reduce bias in the respondent’s memory by asking questions related to the assets in the household rather than about how much income is spent and on what precisely.

Table 1. Asset Index Variables 

^Note:Data is obtained from the total households in the EECV-2012.

Table 2 presents the descriptive statistics for our sample of children by type of household. Regarding individual characteristics, there are more girls in both recipient and non-recipient households. Besides, the average age is around 15. For the household-head characteristics, Table 2 indicates a higher proportion of women heads of household in recipient households. This may be related to the migration scenario of spouses. Moreover, the share of the employed household head is larger for non-recipient households. This may be explained by house­hold members in the country of origin using remittances to buy time away from the labor market. Lastly, we also observe that the household heads’ level of education in both types of households is around eight years, and slightly higher in recipient households.

Concerning household characteristics, we find that the number of younger members and the number of adults is similar in both types of households. At the bottom of the table, we report the share of children by the asset index. The results suggest that most of the children are in the second quartile, with a larger share of children in the lowest quartile in non-recipient households. Finally, around 6 percent of the children live in recipient households.

5.1. Remittances, Child Labor and Schooling

This section presents the baseline results for this paper. Table 4 reports the ATT of living in a remittance-receiving household on the probability of work­ing and attending school for children aged between 12 and 18. Panel A and B report the effect on child labor and school attendance, respectively. We fol­low ^{Acosta (2011)} in using the single nearest neighbor as our chosen algo­rithm. This matching method seems to perform better when the non-treated observations are more common than the treated.⁴ For the analysis, the stan­dard errors are calculated following ^{Abadie & Imbens (2011)}.

Before reporting the main results, we analyze whether the model is correctly balanced, given the independent variables used. In Figure 3, we present box plots of the propensity score over treatment levels for the raw data and the matched sample. The plot reveals that the matched sample box is the same over the treatment level, suggesting that the matched sample is balanced.

Table 4. ATT of Remittances on Child Labor and School Attendance (PSM) 

^Notes:ATT are obtained using data from EECV-2012. The last two columns present the results of the statistical test (^{Mantel & Haenszel, 1959}), suggested by ^{Becker & Caliendo (2007)}. indicates the value of under the assumption that we have overestimated the true effects. indicates the value of under the assumption that we have underestimated the true effects. We report robust standard errors in parenthesis following ^{Abadie & Imbens (2011)}. */**/*** denotes significance at the 10/5/1 percent level.

We also check whether the overlap assumption is satisfied. Violating the over­lap assumption implies that we cannot account for the unobserved outcomes for some individuals. We follow ^{Busso, DiNardo & McCrary (2014)} to check whether the estimated density has too much mass around 0 or 1. Figure 4 plots the density of the predicted probabilities that a child in a non-recipient household is, in fact, in a non-recipient household, and the estimated density that a child in a recipient household is in a non-recipient household. Neither graph indicates too much probability mass near zero and one suggesting that the overlap assumption is not violated.

Table 4 presents the baseline results of this paper. Under the PSM Nearest Neighbor Matching (NN) algorithm, each individual in the treatment group is paired with an individual in the closest control group, regardless of distance. ATT results indicate that living in a remittance-receiving household reduces the probability of working by 10.6 percentage points. This effect is significant at the 1 percent level. Unlike child labor, we find no evidence that living in a remit­tance-receiving household has any effect on the likelihood of attending school.

Figure 3. Balance PlotNotes: The figure is obtained using data from EECV-2012. 

Figure 4. Estimated Densities and Overlap AssumptionNotes: The figure is obtained using data from EECV-2012. 

We assess the magnitude of our estimates by comparing our results with two studies in Latin America, analyzing the recipient of remittances (at the exten­sive margin) and child labor. The study by ^{Acosta (2011)} uses propensity score matching to evaluate the effect of remittances on child labor in El Salvador. The author finds that with an incidence of child labor of around 10 percent in non-recipient households and 7 percent in recipient households, receiving remittances decreases the probability of working by 3.6 percentage points. In turn, ^{Alcaraz, Chiquiar & Salcedo (2012)}, using a difference-in-difference strategy and data from Mexico, find that with a child labor incidence of 14.6 percent, increased child labor as a response to a decrease in remittances is of 11.4 percentage points. With an incidence of child labor of 6 percent in our sample, our results are large in magnitude.

The last two columns of Table 4 report the results of the statistic proposedby ^{Mantel & Haenszel (1959)} to determine the critical levels of sensitivity ofthe bias generated by possible differences in unobservable characteristics. Inthe case of child labor, the columns report the maximum value that rejectsthe null hypothesis that the treatment effect is zero at a significance level of1 and 5 percent under the assumption that we have underestimated the trueeffect. We do this because, under the assumption of no hidden bias, the Q_MH statistic indicates a significant effect. Besides, the statistic becomes even moresignificant for increasing values of under the assumption that we have overestimated(more positive) the true effect (See Table A1 in the Appendix). Underthe assumption that we have underestimated the treatment effect, which inour case is more relevant given the negative coefficient in our estimate, theresult becomes insignificant at the 1 percent level with a value of = 1.6. Thisresult means that for an unobserved covariate, this variable should affect theodds of receiving remittances by a factor of 60 percent to alter the estimates.

In the case of school attendance, the p-values under the assumption that we have over(under)-estimated the true effect decrease as increases (See Table A1 in the Appendix). Moreover, even when the critical value of does not mean that there is no effect of the treatment in the outcome variable, the lack of significance under the assumption of no bias = 1 constitutes con­sistent evidence that receiving remittances does not affect school attendance. Based on these results, we report the first value of , where the effect becomes significant at the 1 and 5 percent significance level under both assumptions of the statistic. Under the assumption that we have over(under)-estimated the treatment effect, the results state that the confidence interval would not include zero at the significance level of 5 and 1 percent if an unobserved variable caused the odds ratio of treatment assignment to differ between those at recipient households and their comparison group by 2.4(1.2) and 2.9(1.5) respectively. Overall the estimates for child labor and school attendance are unlikely to suffer from hidden biases.

One of the negative impacts that child labor may have in Colombia is the potential adverse effects on schooling (^{Ministerio de Educación Nacional, 2012}). However, reducing child labor does not necessarily equal increased school attendance and vice versa. For instance, ^{Ravallion & Wodon (2000)} show that if schooling and leisure are perfect substitutes, increases in school attendance may come particularly by the reduction of leisure time and not a reduction in labor participation. Colombia is a good example of this. ^{Attanasio et al. (2010)} find that school and work time are not perfect substitutes, report­ing that even when the government program Familias in Acción increased school participation by reducing the cost of schooling, there was no reduction in income-generating work activities, suggesting that some of the increased time in school comes from reducing children’s leisure time. With the absence of children’s time allocation data in our case, our results can only be taken as new evidence from Colombia that child labor and schooling are far from being substitutes. Nevertheless, the reduction in child labor underlines the impor­tance of international remittances as a source of private transfers that may complement government programs such as Familias en Acción.

5.2. Robustness Checks

We now check the sensitivity of our results to different algorithmic methods. Estimates of the average treatment effect on the treated are reported in Table 5, with the baseline estimate in Column 1 for comparison. In Column 2, we report the results using the stratification method proposed by ^{Dehejia & Wahba (2002)}. Column 3 contains the results matching the treated units with a weighted average of all controls (Kernel Matching). This methodology gives greater weight to controls with smaller distances. In Column 4, we use a distance-weighted radius matching methodology introduced by ^{Lechner, Miquel & Wunsch (2011)}. One of the advantages of this methodology is that we do not need to decide the radius ad hoc, as it uses a partially data-driven choice of the radius size as a function of the distances in pair matching.

Panel A and B in Table 5 report the results on child labor and school atten­dance. Overall, we find that receiving remittances leads to a decrease in child labor. As in the case of nearest-neighbor matching, we find no evidence that remittances affect school attendance. In our baseline result, the coefficients are lower in magnitude but still within range of the estimates in the literature (see the previous section). In the stratification procedure, we find that with five blocks, the mean propensity score is no different from treated and control in each block. Moreover, we test the balancing property for each variable in each block and find that our covariates are balanced (See Table A2 in the Appendix). It is worth noting that the latter result is consistent with the covariates balancing test in our baseline specification (See Table 3). The results in the stratification procedure and the kernel matching indicate that living in a remittance-receiving household reduces the probability of work­ing by 4.2 percentage points. Although there may be some concerns about the reliability of the standard errors under these two specifications (^{Abadie & Imbens, 2008}), the results are still robust to our nearest neighbor matching. Finally, the results in Column 4 indicate that receiving remittances decreases child labor by 8.5 percentage points. Overall, the negative effect of remittances on child labor is robust to the different chosen algorithms.

Table 5. Robustness checks. The impact of remittances on Child Labor and School Attendance 

^Notes:TT are obtained using data from EECV-2012. The Propensity Score in Columns 2-4 is computed using the variables in Table 3. The Kernel matching algorithm uses a bwidth (0.06). We report standard errors in parenthesis. Standard errors in Column 1 are computed using ^{Abadie & Imbens (2011)}. Standard errors in Columns 2 and 3 are calculated using Bootstrapping. The radius matching estimator and standard error in Column 4 are computed based on ^{Huber, Lechner & Steinmayr (2012)}. */**/*** denotes significance at the 10/5/1 percent level.