Conditional cash transfers tools to combat child labor: Evidence from a randomized controlled trial in Costa Rica

The Government of Costa Rica collaborated with a research team to conduct a randomized controlled trial of their Working Children and Adolescents program. The program provided working youth with a monthly cash transfer with the conditions that they attend school regularly and complete their grade. This study examines the effect of the cash transfer on (1) child labor and hazardous child labor participation as well as hours worked; (2) school enrollment, attendance, and completion; and (3) self-reported health. The main findings provide evidence of a statistically significant reduction of more than 4 hours worked per week by children. The findings also suggest null effects on labor participation and school outcomes. Cost-effectiveness analysis shows that the program demonstrates a transfer effectiveness and cost-effectiveness comparable to similar interventions in Latin American countries. The subsidy alone does not seem enough to improve schooling outcomes, justifying the necessity of additional education policies to complement the cash transfer program.


| INTRODUCTION
Child labor seriously threatens well-being, human development, productivity, and equality. In 2020, approximately 1 out of 10 children worldwide was engaged in work, accounting for 160 million child workers (International Labour Office and United Nations Children's Fund, 2021). What is more troubling is that 79 million of these children are engaged in what is considered to be hazardous child labor that endangers their safety. The determinants of child labor are contextually diverse. They often include poverty, cultural views toward work, insufficient legal frameworks, and weak enforcement, among others. By extension, the policies to combat child labor are also diverse. They include poverty alleviation schemes, changes in legal and cultural norms, educational interventions, rights awareness campaigns, and child protection policies.
The Government of Costa Rica (GCR) created the NNAT (Niños, niñas y adolescentes trabajadores) program to address child labor and hazardous child labor. The NNAT program's objective was to reduce child labor by incentivizing school retention and reinserting youth into public secondary schools. The program was designed to compensate the family for the opportunity cost of income lost by a child attending school instead of working. The subsidy is expected to reduce child labor, serving as an income effect. At the same time, a school enrollment conditionality is expected to reduce child labor through a substitution effect from allocating time to schooling. The main objective of the evaluation is to estimate the effects of the NNAT program on youth beneficiaries' labor market, education, and health outcomes, specifically, child labor and hazardous child labor participation, hours worked, school enrollment, attendance, and grade completion. To accomplish this objective, four research questions were developed: (1) What are the effects of the NNAT program on labor outcomes for program participants?
(2) What are the effects of the NNAT program on school outcomes for program participants?
(3) How has the health of program participants changed after receiving the subsidy? (4) What is the program's cost-effectiveness, and how does it compare with similar interventions in other countries?
The literature on conditional cash transfers (CCT) focused on addressing child labor has been under constant development during the last two decades. De Hoop and Rosati (2013) argued that CCTs enable household consumption smoothing and prevent youth from pursuing work to support their families during economic shocks. De Hoop and Rosati (2014) also conducted a systematic review of 30 cash transfer programs across 12 countries. They found that cash transfers lowered participation in child labor and hours worked and discouraged parents from using child labor to meet their household economic needs. Additionally, they found that transfers have greater economic impacts on boys and larger decreases in household chores for girls. Regarding schooling outcomes, the authors argued that there is little evidence of the effectiveness of including schooling as a condition for the transfer in reducing child labor. Additionally, they found that CCTs lower the risk of students dropping out due to economic shocks but did not reduce child labor as a risk-coping strategy for families. De Janvry et al. (2006) use data from Mexico's Progresa CCT program to study how adverse economic shocks (unemployment or illness of the household head or natural disasters) can be offset in families protected by a CCT program, particularly for school and child labor outcomes. Their findings show that student beneficiaries were able to remain in school during negative economic shocks; however, they were more likely to increase their labor participation. Reynolds (2015) finds that an expansion of the Bolsa Familia CCT program in Brazil to cover previously ineligible adolescents ages 16 and 17 years does not impact their time allocation for household chores and work. The author also reports an increase of between 6 and 7 points in school attendance but does not find an associated decrease in their hours worked, suggesting that the trade-off between schooling and work is weak. Instead, students continue to work and attend school to receive the additional income. Cepaluni et al. (2022) and Peruffo and Ferreira (2017) also use simulated data from Brazil to investigate the long-term effects of the program on school and labor outcomes. Their main long-term findings include an increase from 52% to 90% in school completion, a reduction of 20% in hours worked, and a decrease of 22.5% in the share of working children. Cepaluni et al. (2022) use a multiple-births instrument to analyze whether or not Brazil's Bolsa Familia is able to reduce child labor participation. The authors hypothesize that an income effect from the transfer will offset wage earnings. In contrast, a substitution effect would displace potential wage earnings by reducing work time available due to school attendance. Their study finds that receiving the transfer does not reduce child labor participation and does not improve school outcomes, implying a failure of the income effect. Regarding the substitution effect, the authors do not find effects on child laborers' literacy or school attendance.
Not all studies concur in their findings. Dammert et al. (2018) found that CCTs can also increase the risk of child labor if the policies decrease the households' income-generating activities and productivity, particularly when the child is essential in the family's production of goods for sale. In another recent study, Parker and Vogl (2018) describe how CCT programs designed to reduce child labor and improve educational outcomes are currently being implemented in more than 80 countries. The authors found that, in the short-term, CCTs are more effective than unconditional subsidies in discouraging child labor and improving educational outcomes for youth. When looking at the long-term effects of CCTs for youth, the authors found positive long-term effects on labor market outcomes, including lower unemployment, transitions from informal to formal labor, and higher earnings for program participants, with the greatest effects being for women. De Hoop et al. (2019) evaluated the Pantawid CCT program in the Philippines through a randomized controlled trial. The authors find that children that were not studying or working were able to increase their school participation because of the program. However, they also increased their work for pay by 5%. The author's explanation for this puzzling finding is that the cash transfer is insufficient to cover the school costs incurred after enrolling in school. An important implication of this research is that the design and contextualized transfer amount is crucial for the program to meet its goals.
After reviewing the international evidence, this literature review focuses on CCT programs in Costa Rica. Costa Rica has implemented two CCTs to increase education completion: (1) Superémonos, which started in 2000 and was suspended in 2002, and (2) Avancemos, which was established in 2006 and expanded nationally in 2007, and is still ongoing. Duryea and Morrison (2004) evaluated the role Superémonos played in slowing the rise in child labor and increasing school attendance among program participants. Their study found that program beneficiaries were more likely to attend school than their non-beneficiary counterparts. However, the program did not lead to measurable reductions in child labor. Hern andez and Mata (2015) and Meza-Cordero et al. (2015) found that Avancemos beneficiaries completed more years of schooling than students who were not participating in the program, but neither study focused on child labor.
The proliferation of CCTs led researchers and policymakers to question which interventions achieve better results, given the elevated costs incurred. García and Saavedra (2017) conducted a meta-analysis of 94 education-oriented CCT studies in low-and middle-income countries. Their paper provided a cost-effectiveness estimation of nine CCT programs, for which they have data on both transfer-effectiveness ratios (impact units per dollar of transfer) and cost-transfer ratios (CTR) (administrative costs per dollar of transfers). The authors described three types of costs associated with CCT programs: (1) administrative costs encompassing the costs of operations, personnel, and equipment, (2) transfer costs related to the amount of the transfer and any materials given to beneficiaries, and (3) private costs accounting for the costs that beneficiaries must incur to receive the transfer. In their results, the authors reported that the average monthly transfer for a secondary school intervention is $23.20. They also found an average of 21 cents of administrative costs per year for each dollar of transfer. The lowest ratio was for programs in Ecuador and Bangladesh, with 4 cents of administrative costs to deliver one dollar of transfers to beneficiaries. In comparison, the cost to deliver 1 dollar of transfers to beneficiaries in Nicaragua was 63 cents. The programs in Ecuador and Bangladesh delivered 1.3 and 13 percentage points of enrollment impact per dollar of administrative cost per year, respectively. In contrast, programs in Mexico and Brazil delivered 0.4 and 0.5 percentage points of enrollment impact per dollar of administrative cost per year. Mexico's Progresa delivered a 0.21 percentage point increase in secondary enrollment per dollar of administrative cost per year. Colombia's Familias en Acci on delivered a 0.12 percentage point increase in secondary enrollment per dollar of administrative cost per year. Colombia's Familias en Acci on delivered a 0.10 percentage point increase in secondary graduation per dollar of administrative cost per year.
The evaluation of the NNAT program provides evidence of the effectiveness of a CCT targeted at working youth in a middle-income country. This is the first experimental evaluation conducted on a program designed to reduce child labor and improve educational outcomes in Costa Rica. The evaluation of NNAT also provides academics, researchers, and the GCR with data-driven findings that measure the effects on education and labor outcomes. These results will allow policymakers to better understand the current child labor problem, assess NNAT's impacts, and make timely corrections on the basis of empirical evidence. The main findings are that the program was able to significantly reduce hours worked per week by more than 4 hours. The estimations do not show any effects from the program on labor participation decisions and on school or health outcomes. Finally, NNAT's cost-effectiveness analysis documents transfer effectiveness and cost-effectiveness comparable to similar interventions in Latin American countries.
This paper contributes to the literature on cash transfer programs in middle-income countries by conducting a randomized controlled trial of the Government of Costa Rica's leading public programs for combating child labor. The paper also contributes to policy crafting and correcting by providing evidence of the impacts of a unique CCT design tailored to a vulnerable group that is of the highest priority to most country governments. The remaining part of the paper is structured in the following way: Section 2 describes the background of child labor in Costa Rica and the design of the NNAT Program. Section 3 describes the evaluation design and random assignment, Section 4 describes the primary data collected and the data on program costs, Section 5 presents the empirical strategy, Section 6 details the results, and Section 7 lists the limitations and robustness checks performed. Section 8 concludes.

| BACKGROUND ON CHILD LABOR IN COSTA RICA AND ON THE NNAT PROGRAM
Although Costa Rican children have widespread access to school, some are still involved in child labor and hazardous child labor. Child labor in this middle-income country has proven hard to eradicate, as it is a multi-causal problem whose roots include poverty and economic inequality, lack of interest in formal education, youth pregnancy, embedded cultural work values, and family disintegration. In recent years, child labor rates in Costa Rica have been significantly reduced. However, the remaining cases are concentrated primarily in the agriculture and fishing sectors. Construction and manufacturing, domestic work, and street vending also employ children on a lower scale. The Costa Rican National Household Survey (Encuesta Nacional de Hogares; ENAHO) collects data on working children every 5 years. 1 This official data from the 2011 and 2016 surveys is presented in Table 1 and shows that child labor rates reduced from 4.36% to 3.09%, respectively.
The Government of Costa Rica acknowledges child labor and hazardous child labor as social problems that adversely affect the physical, intellectual, moral, emotional, social, and educational development of working children and adolescents. As part of its efforts to eliminate child labor and hazardous child labor, GCR ratified all major international conventions on child labor and also established corresponding national laws. Conventions 138 and 182 of the International Labour Organization (ILO) were ratified without exception in 1976 and 2001, respectively. Costa Rica's National Childhood and Adolescence Code (C odigo de la Niñez y la Adolescencia; CNA), made law in 1998, goes beyond the ILO's conventions by prohibiting minors under the age of 15 years from participating in any form of labor, including light work and unpaid household work.
In Costa Rica, labor is illegal for all persons under 15 years of age, and hazardous labor is unlawful for anyone under 18 years off age. Based on these standards, the statistical measure of illegal child labor includes all individuals aged 5 to 15 years engaged in any work. 2 Youth between the ages of 15 and 17 years working under conditions prohibited by the CNA are considered to be engaged in hazardous child labor. Hazardous child labor is defined on the basis of working hours and schedules, work activities, and working conditions. Youth between the ages of 15 and 17 years may not work more than 6 hours a day or 36 hours a week. They are also forbidden to work between 7 p.m. and 7 a.m., with a few exceptions allowing for work until 10 p.m. The CNA includes a list of prohibited hazardous occupations and economic activities as well as a list of hazardous conditions and environments prohibited because they may be harmful to the health, security, or morality of minors. Examples of occupations considered hazardous on the basis of the economic activity include construction and mining. Examples of occupations considered hazardous on the basis of the environment include bars, casinos, and other establishments that serve alcohol and are typically open at late hours. This study's operational definitions of child labor and hazardous child labor adhere to the international standards determined by the ILO's International Conference of Labor Statisticians (ICLS) and the national legislation of the GCR. The 18th ICLS Resolution concerning statistics on child labor (2008), with an amendment approved at the 20th ICLS (2018), is the current international standard for child labor statistics. The ICLS has also determined that each national government is responsible for defining hazardous child labor. Figure 1 summarizes the definitions of permissible and non-permissible work used in this research.
The Government of Costa Rica designed the NNAT program. Government technicians investigated the reported earnings of potential beneficiaries and found that their income was, on average, between 120 to 240 USD per month. By covering this opportunity cost through a subsidy, the program enabled the family to cover the costs of living and schooling, such as transportation and supplies. The monthly amount of the subsidy was set at 90,000 Costa Rican Colones ($158.60 at 2017 current rates). For reference, according to the Central Bank of Costa Rica (2017), the GDP per capita in 2017 was 6,708,295.20 Colones ($11,820.50 at current rates). 3 Therefore, a family with a beneficiary student would receive 1,080,000 Colones annually ($1,903.10 at current prices). Following the annual income per capita, the transfer accounts for approximately one-sixth of the average annual income. It is essential to point out that these transfers are directed to households in the lowest socioeconomic group, particularly those below the poverty line. Therefore, the transfer represents a much higher percentage of their income in comparison with the total population.
The NNAT program was implemented by two GCR institutions: (1) the Ministry of Labor of Costa Rica (Ministerio de Trabajo y Seguridad Social, MTSS) and (2) the program administrator, the Joint Institute of Social Aid (Instituto Mixto de Ayuda Social; IMAS). The Ministry of Labor provided the technical expertise to identify and assess youth participating in child labor who were eligible for financial support through public funds. The program administrator was responsible for the adding beneficiaries into its service delivery database and distributing the subsidies to participating families on a monthly basis.
The program targeted working youth between ages 12 and 17 years. The NNAT eligibility requirements are detailed as follows. First, youth must either be between 12 and 15 years of age and working in any occupation (for pay or not for pay) or between 15 and 17 years of age and working in a hazardous occupation, which the Ministry of Labor verifies. 4  must be in the IMAS objective population system and qualify under their poverty or extreme poverty criteria. Third, a youth's parents must provide IMAS with proof of school enrollment and bank account information. The GCR's administrative process for enrolling beneficiaries has two stages: verification and processing. First, each youth's child labor status must be verified through a technical report prepared by a social worker from the Ministry of Labor. This technical report documents the child labor status and the child's age. After verifying the child labor status and age, the Ministry of Labor sends an official note to the program administrator, IMAS, requesting the case be processed as a NNAT beneficiary. The IMAS administrative office checks that the child meets the age and poverty condition requirements and requests proof of school enrollment and the bank account information of the mother (or legal guardian). Once a working youth is approved, the subsidy is given to the mother or legal guardian every month. The monthly subsidy continues through the school year if the child continues attending school. The NNAT program is similar to other CCT interventions in Latin America. However, the main differences include that it is a program targeted at a particularly vulnerable group (as opposed to more extensive nationwide interventions targeting the population in poverty) and that, although smaller in the target population, it provides a much larger subsidy amount. It follows that a rigorous study of this intervention will allow for the documentation of the effects of transfers on child labor outcomes and comparisons with the effects of similar interventions, targeted or non-targeted.

| EVALUATION DESIGN AND RANDOM ASSIGNMENT
This evaluation used an experimental design to identify the effects generated by the NNAT program. The effects are identified by comparing the outcomes of program beneficiaries with nonparticipants (the counterfactual). A close collaboration between researchers and the GGR took place in 2016 to identify a population of cases of child labor. Because NNAT had a limited budget, an agreement was made with the GCR to study 551 child laborers verified as eligible to participate. Over 600 cases of child labor were reported to the Ministry of Labor's Child Labor Attention Office by school principals and counselors. The potential child laborers were visited and interviewed by social workers trained by the Ministry of Labor. 5 The social workers prepared the required technical report for 554 corroborated child labor cases and conducted the baseline evaluation surveys on all of them. 6 The final study sample included 551 youths who either both worked and attended school or left school to work. This entire population was surveyed at baseline before being randomized into a treatment group or a control group. In late 2016, random assignment of students took place: 276 were allocated into the treatment group to receive the cash subsidy during the 2017 school year, and 275 were assigned into the control group. The application of this randomized controlled trial (RCT) method allowed for the determination of the unbiased effects of the NNAT program on beneficiaries after a full year of the intervention. The estimators from this design measure the average treatment effect on those treated during 2017.
A lottery was implemented in January 2017 to randomly assign the youth to the treatment and control groups. The lottery randomly assigned the 551 participants, 276 to the treatment group and 275 to the control group. After random assignment, the list of treatment group members was shared with the GCR so that their age and child labor condition could be verified and they could be processed as program participants. The criteria for NNAT eligibility and administrative process are described in Section 2.
Ninety-one cases from the treatment group failed the verification and processing stages; 41 turned 18 years old during those months, and another 52 failed to meet at least one of the verification and processing requirements. The enrollment process resulted in 183 youth fully becoming NNAT participants. These beneficiaries were notified of the approval and started receiving the intervention in May 2017. The program administrator, IMAS, distributed the monthly cash subsidy to the bank account of the mother (or legal guardian) from May through December 2017. To continue receiving the subsidy, the participants were required to be enrolled in school and maintain regular attendance. They were also required to complete their current grade to continue receiving the subsidy in the next academic year. Beneficiaries who turned 18 years of age after verification continued to receive the subsidy as long as they satisfied the attendance and completion requirements.

| DATA & DESCRIPTIVE STATISTICS
This section describes the collection of primary data at baseline and endline. It first describes the design, cognitive testing, and piloting of the survey instrument. The next paragraphs describe the baseline and endline data collection. Primary data was collected through a projectspecific survey instrument designed to gather information about sociodemographic characteristics and the outcomes of interest. The instrument was developed to be administered directly to the potential program participants (individuals reported as engaged in child labor) through a personal interview. The survey was composed of modules that gathered personal and sociodemographic information, educational information, work information, workplace conditions, self-reported health status, and information about household chores. The survey also included modules tailored to identify child labor and hazardous child labor according to the operational definitions from section 1.
Cognitive testing was conducted to ensure the target population understood the questions, had adequate response categories, found transitions between topics to be smooth, and clearly understood the interviewer's instructions. Collecting data from working children was complex, as it required (1) enumerators who had training on child protection laws, (2) approaching extremely vulnerable populations, and (3) safety protocols. The strategy was to have the Ministry of Labor train enumerators on child labor legislation and guidelines for interviewing vulnerable children. As a second step, a data collection partner trained the enumerators on the proper use of the survey instrument and how to enter the data into tablets. After training, a survey pilot was conducted among youth by a nonprofit organization recommended by the Ministry of Labor that serves vulnerable youth between the ages of 12 and 17 years. Baseline data collection started immediately after the pilot.
Baseline data were collected between October and November 2016 by social workers jointly trained by the Ministry of Labor and the data collection partner. Specialized training was required because of the age and vulnerable status of the target population. A total of 551 potential beneficiaries were surveyed at baseline, 276 of them became part of the treatment group and 275 became part of the control group. Baseline data served to document pre-program characteristics from all eligible participants. A baseline equivalence analysis was performed through a t-test after the randomization to verify that the treatment and control are identical in observable characteristics. Table 2 presents the baseline sociodemographic statistics of the sampled youth, divided into three columns by the randomly allocated treatment and control groups, plus a column with the difference between the groups.
The average age of the treatment-group youth was 15.5 years, with male participants comprising approximately 54% of the group. The average age of the control group was 15.6 years, and the group was approximately 58% male. Youth in the treatment group came from households with approximately 4.5 members, whereas youth in the control group came from households with approximately 4.7 members. More than 55% of sample youth in both the treatment and control groups lived in households headed by their mother. In the treatment group, 66% of the youth reported that the head of their household had primary education, compared with 63% in the control group. Approximately 15% of youth in the treatment group reported that their head of household had a secondary education, compared with 12% of youth in the control group. Youth in the treatment group reported earning approximately 23,170 Costa Rican Colones (about 37 USD) in the previous week. In comparison, the control group reported earning approximately 19,826 Costa Rican Colones (about 32 USD) during the same period. None of the differences across groups was statistically significant. Table 3 presents the baseline child labor summary statistics. The table shows that approximately 94% of the youth in the treatment and control groups were engaged in some type of work. Treatment-group youth spent 22 hours working in the previous week, whereas controlgroup youth spent 24 hours working in the previous week. Youth in both the treatment and Note: Summary statistics of sociodemographic characteristics present the mean for each variable of interest, and underneath each, the standard error is indicated in parentheses. The third column reflects the difference in the means across groups, as well as the range of values for a 95% confidence interval. The t-tests for group comparison reflect no significant difference at the 95% level across groups at baseline. HH = household.
control groups reported beginning work when they were 15 years old, and almost a quarter of sampled youth reported that work interfered with their plans to study. A total of 58% of youth in the treatment group and 65% of youth in the control group reported working for a wage, salary, commission, or in-kind payment. Additionally, 23% of youth in the treatment group and 75% of youth in the control group reported participating in unpaid household labor. There also were no significant differences in the responses of treatment and control youth for any of these questions. Endline data was collected between March and May 2018. The enumerators were instructed to track and visit each of the 183 students who received the NNAT subsidies as well as the 275 students in the control group. In the treatment group, 11 participants changed their contact information and left their schools, so they were unreachable. In the control group, Note: Summary statistics of employment characteristics present the mean for each variable of interest, and underneath each, the standard error is indicated in parenthesis. The third column reflects the difference in the means across groups, as well as the range of values for a 95% confidence interval. The t-tests for group comparison reflect no significant difference at the 95% level across groups at baseline. 37 participants were unreachable, and 19 refused to answer the survey. The final number of completed surveys was 391, 172 from the treatment group and 219 from the control group. A potential source of bias that emerged in this evaluation has to do with differential attrition due to the delays in the verification of the treatment group and the provision of the transfer to the mother or legal guardian's bank account. Participants who turned 18 years of age during the verification process or who were missing any paperwork throughout this process could not receive cash subsidies. This raised a concern of selection bias from the assigned treatment cases that failed to be completely verified and processed. The concern was addressed by conducting an ex-post group equivalence analysis using baseline values. The analysis utilized t-tests to test if the data from both groups interviewed at endline were equivalent at baseline. The analysis showed no indication of systematic bias in any of the key variables and is presented in Table 4.
Data on costs comes from a document review of official sources. According to the Ministry of Labor of Costa Rica, a cooperation agreement with IMAS was ratified in 2016 to determine the program's administration. The Ministry of Labor would provide IMAS with the approved child labor cases plus the funding for the transfers. IMAS would be responsible for verifying the eligibility requirements and delivering the transfers every month. The transfer costs encompassed $1,903.05 per student per year, which accounts for $525,241.85 for the 276 beneficiaries. To calculate the administration costs of NNAT, the entire expense structure of IMAS was investigated to calculate the percentage designated for personnel costs related to the transfers to beneficiaries. The Contraloría General de la República (2019) reported that 73% of the institution's budget was destined to transfers, 7% to the provision of services, 10% to salaries, and the remaining 10% to equipment and other costs. Using these ratios, and assuming that the provision for this shared effort between the Ministry of Labor and IMAS of the NNAT transfers requires similar administrative costs compared with other IMAS interventions, the costs of the personnel responsible for administrating NNAT are estimated to be 10% of the amount of the transfer, or $52,524.19 in 2017.

| EMPIRICAL STRATEGY
The identification strategy relies on the random assignment of students to treatment and control groups. A single-difference analysis ensures that any differences in outcomes between the two groups are causally attributed to the program. The program impacts are measured through a multivariable regression model that leverages information collected through an endline survey. This section presents the models that are used as part of the empirical strategy and that estimate the program impacts. The benchmark regression model used for this estimation is presented in Equation (1).
Where: Y i determines the outcome of interest for individual i (such as hours worked or grade completion); T i determines the treatment indicator, which equals 1 if the individual i was assigned to the treatment condition and 0 otherwise; and X i is a set of individual observable characteristics of individual i to be included as covariates. Covariates include age, age squared, gender, zone, size of the household, having a mother as the head of the household, and education of the head of the household, and u i is an independently and identically distributed error term with a pooled mean of zero and variance of σ 2 .
The parameter of interest in this model is β, which is the regression-adjusted average effect of the intervention.
To disaggregate the results by age group, gender, and rural/urban zone, this study included three additional specifications with interaction terms. Equations (2), (3), and (4) allow for the capture of the differentiated effects generated by the interaction between the treatment indicator and gender, age group, and urban zone dummies, respectively.
In these equations, G is a dummy equal to 1 for boys, A is a dummy equal to 1 for the age cohort between 12 and 14 years, and Z is a dummy equal to 1 for participants living in urban zones. λ represents the differentiated effect of being a boy. In contrast, μ represents the differential effect of being in the 12 to 14 years age group, and η represents the differential effect of living in an urban zone. Note: Ex-post baseline analysis utilizes baseline data from those individuals interviewed at endline. Summary statistics of the sociodemographic characteristics present the mean for each variable of interest, and underneath each, the standard error is indicated in parentheses. The third column reflects the difference in the means across groups, as well as the range of values for a 95% confidence interval. The t-tests for group comparison reflect no significant difference at the 95% level across groups at baseline. HH = household.
Equation (5) is an encompassing model that combines all three interactions, used to assess how the estimators may change.
Each outcome variable was estimated using all five models. Binary outcomes, such as labor participation, school enrollment, and self-reported good health were estimated using logistic regressions. Their estimators represent changes in the likelihood of the outcome happening, for example, the change in the probability of continuing to work. Continuous outcomes, such as hours worked or income, were estimated using linear regression models. Their estimators represent the numeric change in the outcome, for example, the difference in weekly hours worked. The intervention's transfer-effectiveness (TE) and cost-effectiveness (CE) analysis were conducted. These analyses enable the quantification of how relatively expensive the intervention is per student and permit cross-country comparisons. The TE analysis presents the transfer amount associated with each intended educational and labor outcome of the program. In contrast, the CE analysis determines the cost associated with each intended educational and labor outcome of the program. Both analyses quantify the monetary cost associated with a percentage increase in the outcome of interest.
The program costs, which include all administrative costs plus the total amount of transfers distributed, are analyzed as average transfer and costs across all beneficiaries. TE is defined as the ratio of the program impact on the educational outcome per beneficiary (school attendance, years completed, work status, and hours worked) to the transfer program's costs per beneficiary. The CE ratio is similarly obtained, as the numerator continues to be the impact on each outcome per beneficiary. At the same time, the denominator consists of the implementation costs to run the program (Dhaliwal et al., 2013). Another commonly used measure in the costefficiency of CCT programs is the cost-transfer ratio (CTR). The CTR is the ratio of the administrative cost to the transfer cost, that is, the costs for every dollar delivered in cash to beneficiaries, and is also detailed in per-beneficiary terms. The NNAT impacts chosen for each outcome are the average treatment effect on the treated (ATT) obtained from Equation (1). The mean transfer amount is obtained by dividing the total expenditure on transfers by the number of beneficiaries. The mean administration costs are obtained by dividing total administration costs by the number of beneficiaries. Owing to the nature of the NNAT program (the provision of a transfer to households with the expectation to influence numerous outcomes), the cost-effectiveness results are presented as the ratio of total indivisible expenditures (transfers and administrative costs) to each outcome of interest independently.

| RESULTS
This section presents the estimation of program impacts on the outcomes of interest. Tables 5 through 10 present the average treatment effect on the treated (ATT) estimator, β. The standard error (SE) is reported in parenthesis underneath. Significance levels are measured at the 5%, 1%, and 0.1% levels. Each table also describes the additional interaction terms in each model, as well as the inclusion of covariates, the mean of the outcome variable in the control group, the R-squared, and the number of observations used.
The estimation of each model is presented in five columns. The first column represents the estimates using the basic model, which includes only the treatment indicator and the set of covariates serving as control variables. The second model presents the estimates from the treatment indicator and adds a gender interaction term to reflect the incremental effect of being a boy. The third model uses an age group interaction term to reflect the additional effect of being in the 12 to 14 years of age cohort. The fourth includes a zone interaction to show the differential effect of living in an urban area. Finally, the fifth column presents the model that gathers all three interactions, allowing for the comparison of specifications. The tables for child labor and hazardous child labor outcomes are presented first, followed by school and health outcomes.
Table 5 presents the findings on child labor participation for all children ages 12 to 17 years who reported working, for pay or not, during the previous week. The estimates range between À0.44 and 0.191 and show no definitive evidence of the NNAT program reducing the likelihood of labor participation. While these findings appear to show large effects for some of the specifications, these effects are not statistically significant at the 5% level. The results also show that boys and younger students had a lower likelihood of engaging in child labor, although these findings are also not statistically significant. Table 6 presents the program's effects on hazardous child labor participation among the older youth. These estimates are provided separately from those of the younger cohort because there are differing triggers for illegal hazardous child labor depending on age. This estimate includes all working children ages 15 to 17 years who reported working in a hazardous occupation during the previous month, in hazardous conditions during the previous 6 months, or for a hazardous number of hours during the previous week. The estimates are inconclusive, showing positive and negative effects on the likelihood that older beneficiaries would participate in hazardous child labor. These effects range from À0.212 to 0.099. However, none of them are statistically significant.
The NNAT program's effects on the number of hours worked are presented in Table 7. This table includes all study participants ages 12 to 17 years who reported working, for pay or not, during the previous week. The estimates show that the NNAT program reduced the number of hours worked among beneficiaries, between 1.544 and 6.36 hours per week, depending on the model specification. The basic model shows a reduction of 4.95 hours. This estimate is statistically significant at the 99% level. The model with the gender interaction term shows an incremental reduction of 6.1 hours worked among boys. This finding is statistically significant at the 95% level. The model with the age group interaction term shows that the program had a larger (but not statistically significant) effect on the hours worked for the younger cohort over the older cohort. Finally, the model with the zone interaction shows an insignificant incremental positive effect of 2.9 for those who live in urban areas.
This section presents the findings about the NNAT program's impacts on schooling outcomes, including enrollment, attendance, and grade completion. Table 8 shows that impacts on school enrollment range from 1.42% to 5.58%. Although the treatment estimates show a positive likelihood of enrollment, they are not statistically significant. This result was expected, as most youth in the treatment and control groups were already enrolled in school at baseline.
The program impact on self-reported school attendance (of over 90% attendance) shows effects ranging from À0.548% to 0.382%. All five models do not show a significant impact. The full list of school attendance estimates is presented in Table 9. Table 10 details that program effects on grade completion are positive, ranging from À0.051 to 0.191 on average. However, none of these effects are statistically significant at the 5% level. The final impacts analyzed are self-reported health outcomes. All five models show no significant effects on health. The estimates range from 0.002 to 0.496. These estimates are summarized in Table 11.
The administrative costs of NNAT cover the processing and monitoring of beneficiaries, and transfer costs comprise the amounts given to beneficiaries through direct deposit to their bank accounts. Private (family) costs for this program are negligible. The program's total expenditures include the amount of the transfers plus an administrative cost of approximately 10% of the transfers, as described in Section 4. This ratio of administrative costs for transferring the amounts distributed can also be seen as a CTR of 10%, which is very similar to the CTR of similar programs in the region, such as Progresa in Mexico, with a CTR of 10.6%, and Familias en Acci on in Colombia, with a CTR of 11.7% (García & Saavedra, 2017). Counterfactual mean at endline = 0.81 Note: Child labor participation is a binary variable (equal to 1 for "yes" and equal to 0 for "no"); the model is estimated using a logit regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.
T A B L E 6 Program impacts on hazardous child labor participation, youth ages 15 to 17 years

Note:
Hazardous child labor participation is a binary variable (equal to 1 for "yes" and equal to 0 for "no"); the model is estimated using a logit regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.

Model
(1) Note: Hours worked is a numeric variable; the model is estimated using a linear regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.
The total program expenditures for 2017 in current USD were calculated to be $525,241.85 in transfers plus $52,524.19 in administrative costs, totaling $577,766.04. A total of 276 students received the transfer in this year, resulting in an annual average cost per student of $1,903.05. The impacts, costs, TE and CE ratios, and CRT for each outcome considered in the paper are presented in Table 12.
The overall transfer effectiveness for child labor participation and hazardous child labor participation are À0.0092% and À0.011% points per dollar, respectively, for a representative student. This same representative student had a transfer effectiveness of À0.26 hours worked per dollar. The school enrollment and attendance estimates are 0.03% and 0.02% points, respectively, while the grade completion is 0.0087% points per dollar spent on the transfer. Self-reported health has a transfer effectiveness of 0.002% points per dollar. As administrative costs were found to represent 10% of the transfer amounts, it follows that, by construction, the cost:transfer ratio is fixed at 10% and the cost-effectiveness mean is exactly 10% higher than the associated transfer-effectiveness estimate for each outcome.

| LIMITATIONS AND ROBUSTNESS CHECK
This study had the limitation that school administrative records were not available. The evaluation design originally planned to use administrative records to complement the self-reported primary data. Early on, it was noticed that the Ministry of Education would not participate in Counterfactual mean at endline = 0.94 Note: School enrollment is a binary variable (equal to 1 for "yes" and equal to 0 for "no"); the model is estimated using a logit regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Under 15 years of age interacting with treatment predicts success perfectly, so it has been omitted. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.
verifying NNAT school attendance or performance and that the schools' own administrative records on attendance and performance were not comparable across project schools. The lack of administrative records forced the evaluation to rely solely on student self-reported school outcomes measures, such as self-reported attendance. Another challenge was accessing the program's administrative data on verification and approval of NNAT participants. Participants were verified by the Ministry of Labor, which transfers their data to the program administrator (IMAS) office for processing. Approved student cases are then sent to the administrator's regional offices, which verify school enrollment and distribute the cash subsidy. This lengthy process led to many child laborers who should have been able to participate in the NNAT program aging out and becoming ineligible, so they had to be dropped from the study.
A final limitation of this evaluation was that the timeframe of the data only allowed for the analysis of the short-term effects of the intervention. The literature shows how longer-term evaluations can study the effects of CCTs on labor market outcomes such as employability and earnings, which this study could not measure by design and will have to be explored in future research.
A difference in difference (DID) strategy was implemented as a robustness check to verify the validity of the estimates. The effects of the transfer provision were identified by comparing the outcomes of the treated students to a counterfactual. The DID estimator provides the unbiased effects of the intervention by quantifying and subtracting the time trends of the control students (to serve as the counterfactual trend for treated students). The notation for the DID model is the same as that of Equation (1). The period will be defined by Ρ. P = 0 is defined as T A B L E 9 Program impacts on participant-reported school attendance Counterfactual mean at endline = 0.93 Note: School attendance is a binary variable (equal to 1 for "yes" and equal to 0 for "no"); the model is estimated using a logit regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Male interacting with treatment predicts success perfectly, so it has been omitted. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.
the baseline period, and P = 1 is the endline period. Equation (6) represents the equation used to obtain the DID estimator: In this model, δ identifies the effects of the program and μ i represents any unobservable fixed effects for individual I, and the effects of the baseline control variables, γ t , are allowed to change over time. By subtracting the baseline period from the post-intervention period at the individual level, any time-invariant fixed effects are eliminated, resulting in the following Equation (7): where γ is equal to Δγ and η i is equal to Δε i . Since T i is the dummy variable indicating treatment, γ indicates the average treatment effect on the treated. Continuous variables were estimated through ordinary least squares, while binary variables were estimated with logistic regression. Standard errors are clustered at the zone level to account for contextual educational differences between rural and urban areas. The model is estimated through a balanced panel consisting of 784 observations. The attendance and grade completion variables are not included, as they were only asked at endline. Note: Grade completed is a numeric variable; the model is estimated using a linear regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.
not fully consistent with the analogous model in Equation (1), as not all the effects are identical or of the same sign. However, there is consistency in all insignificant effects. The only significant effect, hours worked, does show a robust statistically significant reduction of nearly minus 5 hours in the RCT specification and nearly minus 3 hours in the DID specification, although the latter is not statistically significant.
T A B L E 1 1 Program impacts on self-reported health of children Counterfactual mean at endline = 0.82 Note: Self-reported good health is a binary variable (equal to 1 for "yes" and equal to 0 for "no"); the model is estimated using a logit regression. Standard errors are clustered at the zone level and presented in the parenthesis below each estimate. Column 1 reflects the estimation of the model without treatment interaction terms. Columns 2, 3, and 4 include a unique treatment interaction with being a male child, ages 12-14 years, and living in an urban area, respectively. Column 5 includes all treatment interaction terms simultaneously. Every model includes a set of covariates reflecting age, age squared, gender, urban, household size, having a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01. Child labor participation, hazardous child labor participation, enrollment, and self-reported health are binary variables (equal to 1 for "yes" and equal to 0 for "no") estimated using a logit regression. Hours worked is a numeric variable estimated using a linear regression. Standard errors are clustered at the zone level and presented in the parentheses below each estimate.

T A B L E 1 2 Transfer effectiveness and cost-effectiveness analysis
The estimation of the DID model does not include any treatment interaction terms and includes a set of covariates reflecting age, age squared, gender, urban, household size, have a mother as the head of the household, and the education level of the head of the household. Significance levels for the estimates are determined as: * p < .1, ** p < .05, *** p < .01.

| CONCLUSIONS
The underlying hypothesis of this research is that a conditional cash transfer would offset the opportunity cost for working children (an income effect) while at the same time elevating their schooling outcomes and reducing the time available for work (a substitution effect from the conditionality). The estimates from this impact evaluation show that the NNAT program reduced the likelihood of labor participation for the full sample of children between ages 12 and 17 years. However, these results are not statistically significant at the 5% level. A main finding is that program participants reduced the number of weekly hours worked at statistically significant levels, which is consistent with the existing literature. The benchmark model provides evidence that NNAT beneficiaries reduce their weekly hours worked by 4.95. This is a very significant change considering that the control group average during the same period was 16 hours worked per week. It is particularly interesting that, when adding a gender interaction component, the estimates show an even larger reduction of hours worked by boys, with 6.1 fewer hours worked per week. Age group and zone of residence do not show heterogeneous effects. These child labor findings suggest that the income effect generated from the subsidy was not enough to offset child labor participation fully. A potential explanation is that a weak deterrence mechanism (lack of information about child labor laws and harms and inadequate protective services enforcement) led children to continue working while enrolled in school, only limiting their hours available to work because of the time allocated to school attendance. The estimates show positive effects on school enrollment, attendance, and grade completion, although these results are not statistically significant at the 5% level. These findings are not surprising, as in this context, children continue to be enrolled in school while working. 7 The same argument for enrollment can be made for school attendance. Surprisingly, the NNAT program did not lead to significantly higher grade completion. It was expected that the reduction in hours worked would lead to more time for homework and better rest and that the grade completion requirement would lead to better grades and higher rates of grade completion. Finally, positive health outcomes were detected but not statistically significant. A potential explanation is that health did not encounter much variation because students did not fully stop working. These findings suggest that a substitution effect due to the reallocation of time does seem to be present. Still, it is not large enough to improve school and health outcomes in a statistically significant way. When considering heterogeneous effects through interaction terms, no differential effects were found for younger students or based on their residence zone.
The analysis of program impacts and costs suggests that the program is considerably efficient, utilizing 10% of total transfer amounts to cover administration costs (a CTR of 0.1). When analyzing transfer efficiency, this study shows that a dollar increases attendance by 0.02%, which is very similar to the multi-country average of 0.026% documented by García and Saavedra (2017). The estimate for the cost-effectiveness of a 0.2% increase in attendance per dollar is also consistent with the literature for other CCT programs in the region. 8 It is noteworthy that, although the program seems to be achieving a higher impact than other countries in terms of attendance when costs are considered, these gains per cost appear to converge with those observed in other countries. The cost-effectiveness analysis detailed in this paper additionally fills gaps in the existing literature by providing benchmark estimates for measures not commonly documented, particularly health, labor participation, and hours worked.
This study provides empirical evidence that a monetary subsidy can assist in reducing child labor. However, the findings did not show a large effect on other outcomes of interest. The results show that child labor in Costa Rica is primarily driven by financial needs. Still, they suggest that a monetary subsidy alone is not enough to have a strong effect on other expected behaviors, such as school outcomes. Finally, the implementation processing challenges and subsidy disbursement delays also suggest that the results from this evaluation would be underestimated in comparison to a more-effectively run intervention.
The evaluation of the NNAT program suggests numerous recommendations and policy implications for implementers and other policymakers. The findings suggest that most youth reduced the number of hours worked but did not completely stop working. Policymakers should consider that CCT programs provide an incentive to increase the time in school and reduce the financial need from working. However, they do not seem to have enough capacity to fully prevent children from working. The NNAT program did not lead to significant improvements in schooling outcomes. Programs to improve schooling outcomes should include an additional educational component to address students' motivation and school completion targets. Public policy to mitigate child labor should continue crafting interventions that can be rigorously evaluated to document program impacts and make timely corrections that will result in a better use of resources. As learned in this evaluation, service delivery, starting with the identification of beneficiaries and ending with the disbursement of cash subsidies, is crucial for meeting the program's goals. Initially, a slow administrative process will lead eligible participants to become ineligible, and delays in the provision of the transfers could lead to opposite effects, as children may drop out of school and return to labor activities. Child labor policy studies should also focus on the quality of the program's administration, particularly regarding the verification of enrollment requirements and timing of the provision of the transfers.
Further research on the NNAT program should include an assessment of how schools and government offices are able to verify eligible participants through official channels in an expedited fashion instead of requiring parents to present the paperwork. An additional important topic for future research on conditional cash transfers in education has to do with the timing of the school subsidy. Certain school expenditures are occasional (such as uniforms and materials at the beginning of the school year), and others are constant (such as transportation and food); hence, it would be important to evaluate the best timing and transfers amounts to minimize the likelihood of drop-out and child labor outcomes. 9 In addition, public policy programs such as NNAT are expected to have the largest impacts when families that are receiving other government subsidies periodically do not have to withdraw from them and instead receive NNAT as an incremental transfer after approval. 10 The NNAT program proved that a CCT transfer can reduce student beneficiaries' hours worked but is not able to fully eliminate their child labor practices. The continuation and holistic expansion of services provided by NNAT could be conducive to continuing the reduction of the remaining child labor in Costa Rica. A final technical lesson learned from this study is that when designing a randomized controlled trial for evaluating a social program, it is crucial to collect baseline data for conducting group equivalence tests, but also for implementing difference in difference methods as robustness checks, particularly useful in cases when the groups may suffer from differential attrition or any other potential source of bias.
Saurabh Singhal, Arya Gaduh, and other participants at the Conference in Honor of John Strauss for their valuable comments.

DATA AVAILABILITY STATEMENT
The data supporting this study's findings are openly available in the US Department of Labor, International Bureau of Labor Affairs repository at https://www.dol.gov/agencies/ilab/ourwork/impact-evaluations.
2 For this study, the operational definition of child labor participation also includes adolescents ages 15 to 17 years legally engaged in non-hazardous occupations and conditions. 3 According to the OECD Income Distribution Database, in 2015 the mean disposable income of Costa Rica was 5,162,885 Colones ($9,657.47 at current rates), while the median disposable income was of 3,386,500 Colones ($6,334.64 at current rates). 4 Owing to the existence of remote locations in the country and of illegal child labor activities that may lead individuals to not truthfully declare work status, the pool of eligible program participants does not necessarily reflect the problematic of child labor in Costa Rica to its full extent. 5 The social workers write a technical report for cases that they identify as child labor and do not complete a technical report for cases that they determine are not child labor. 6 Three cases age out (turned 18 years old) between the period of the interview and the randomization, so they were excluded. 7 As part of the fieldwork activities, we learned that most of the children beneficiaries worked and studied simultaneously; while their parents expressed an interest in keeping their children enrolled in school, they also needed them to continue to contribute financially to the household. 8 Progresa in Mexico has a CE of 0.143% and Familias en Acci on in Colombia has a CE of 0.038% (García & Saavedra, 2017). 9 In a child labor context, cash transfers should be designed as year-round, as a household's basic needs and opportunity costs from child labor are constant. In fact, the subsidies would potentially have the greatest effect if they were disbursed before the beginning of the school year, when schooling costs are the highest. The first payment of the school year should even be somewhat larger than the others to help families cover the cost of uniforms and materials. 10 For example, if the family is already receiving the secondary school subsidy for families living under poverty, Avancemos, they should not be forced to withdraw from it and have to wait for the processing and approval of NNAT, but instead they should continue receiving the original subsidy continuously and, when approved for NNAT, start receiving the complementary amount in addition to account for the full NNAT transfer.