WPS4702 Policy ReseaRch WoRking PaPeR 4702 imPact evaluation seRies no. 24 Poverty Alleviation and Child Labor Eric V. Edmonds Norbert Schady The World Bank Development Research Group Human Development and Public Services Team August 2008 Policy ReseaRch WoRking PaPeR 4702 Abstract Does child labor decrease as household income rises? most likely to leave school and start work. Additional This question has important implications for the design income is associated with a decline in paid work that of policy on child labor. This paper focuses on a program takes place away from the child's home. Declines in work of unconditional cash transfers in Ecuador. It argues that for pay are associated with increases in school enrollment, the effect of a small increase in household income on especially for girls. Increases in schooling are matched child labor should be concentrated among children most by an increase in education expenditures that appears vulnerable to transitioning from schooling to work. The to absorb most of the cash transfer. However, total paper finds support for this hypothesis. Cash transfers household expenditures do not increase with the transfer have small effects on child time allocation at peak school and appear to fall in households most impacted by the attendance ages and among children already out of school transfer because of the decline in child labor. at baseline, but have large impacts at ages and in groups This paper--a product of the Human Development and Public Services Team, Development Research Group--is part of a larger effort in the department to understand the determinants of child labor. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors can be contacted at eric.v.edmonds@Dartmouth.EDU and nschady@worldbank.org. The Impact Evaluation Series has been established in recognition of the importance of impact evaluation studies for World Bank operations and for development in general. The series serves as a vehicle for the dissemination of findings of those studies. Papers in this series are part of the Bank's Policy Research Working Paper Series. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Poverty Alleviation and Child Labor Eric V. Edmonds Norbert Schady Dartmouth College, IZA, and NBER The World Bank We thank Caridad Araujo and Ryan Booth for helping preparing the data for analysis, in particular with construction of the expenditure aggregate. Correspondence to Edmonds at eedmonds@dartmouth.edu and Schady at nschady@worldbank.org. The views expressed in this paper are those of the authors, and they should not be attributed to the World Bank, its Executive Directors, or the countries they represent. 1. Introduction More than one in five children in the world are working. Most of these working children reside in poor countries. This paper is concerned with the relationship in poor countries between current family economic status and whether the child works. There are two distinct strands of research. The first considers whether working while young influences current economic status through the economic contribution of children to the household (Manacorda 2006) and child labor's impact on local labor markets (Basu and Van 1998). The second examines whether and why current economic status influences the decision to send children to work. Understanding the influences of economic factors on child time allocation is important for the design of child labor related policy and for understanding the political economy of existing child labor regulation (Doepke and Zilibotti 2005). This second strand of research is the focus of this study, which examines child time allocation responses to experimental variation in family income from a cash transfer program in Ecuador. In the recent literature on child labor responses to variation in economic status, there is a debate on the extent to which child labor responds to income among poor households. Basu, Das, and Dutta (2007) provide a recent discussion of the state of this literature. The theoretical literature has emphasized parental preferences (Basu and Van 1998) and liquidity constraints (Baland and Robinson 2000) as two reasons why there might be a strong causal relationship between poverty and child labor. Empirical evidence faces the problem of establishing that causation runs from variation in economic status to time allocation decisions. Many correlates of family income influence the economic structure of the household, and a large literature documents the impact of employment opportunities open to children on child time allocation (e.g. Fafchamps and Wahba 2006; Kruger 2007; Manacorda and Rosati 2007; Rosenzweig and Evenson 1977; Schady 2004). This study considers how child time allocation in Ecuador responds to receipt of the Bono de Desarrollo Humano (BDH) cash transfer. The evaluation of the Bono de Desarrollo Humano (BDH) program randomly assigned cash transfers to some poor households and not to others. The BDH transfer is $15 per month, slightly less than 1/10 the monthly income of recipient households, and does not come with any additional conditions attached. We use the random assignment from the evaluation of the BDH as our source of variation in economic status 1 in this study. Our sample consists entirely of BDH-eligible households who are in the poorest two quintiles of Ecuador. We find that random assignment of the BDH income is associated with less work for pay, less involvement in the family farm or business, increased schooling, and fewer children working in some way without attending school. Market work declines as schooling increases, but domestic work appears positively associated with the increase in income, although this increase is small and insignificant. These findings are consistent with a growing literature that has documented that cash and in-kind transfers can reduce child work and increase school enrollment (e.g. Attanasio et al. 2006; Edmonds 2006; Filmer and Schady 2008; Ravallion and Wodon 1998; Schultz 2004). Our paper adds to this literature in a number of important ways. First, we provide a theoretical framework which predicts that there should be substantial heterogeneity in program effects. Using an adaptation of the Basu and Van (1998) model, we show in section 3 that the effect on child labor of the BDH transfer should be concentrated among children most vulnerable to transitioning between school and work. Specifically, our hypothesis is that for young children, for whom schooling is less expensive and the opportunity cost of time in school low, additional income is likely to have little effect on child time allocation. For older children, already withdrawn from school and working, the cash transfer is likely too small to affect their time allocation. However, for children in school and not working but facing the dramatic increase in schooling costs that comes with completing primary school in Ecuador and the rise in opportunity costs that comes with age, their time allocation is potentially substantially altered by small changes in non-child labor income. The findings for work for pay are strongly consistent with this theory. The decline in work for pay associated with BDH receipt on average is concentrated in children most vulnerable to transitioning from work to school. Older children in poor households, especially girls and children in rural areas who are at the end of primary school or higher in the baseline survey are most likely to transition to work for pay and stop schooling in the control sample follow-up survey, taken 1.5 years later. It is this population of older children finishing primary or higher that experience the largest declines in work for pay and increases in schooling with the BDH transfer. These findings for work for pay and schooling are consistent with the results from the PROGRESA program in Mexico, which appears to have an impact on school enrollment 2 primarily among children making the transition from primary to lower secondary school (e.g. Schultz 2004; de Janvry and Sadoulet 2006). Second, we show that the decreases in child labor occurred despite the fact that the transfer was substantially smaller than the sum of the direct and opportunity costs of attending school. While the transfer is $15 per family per month, it is associated with an additional $9 per month spent on education for children most vulnerable to transitioning from schooling to work for pay outside of their family. Average monthly wages for a child working for pay are $84 per month. Thus, the foregone child labor income is greater than the additional education spending. Taken together, we do not see any significant increases in family expenditures associated with additional BDH income. In fact, total family expenditures appear to decline for those whose work for pay is most impacted by the BDH transfer, those most vulnerable to transitioning from school to work for pay. This empirical result is also a prediction of the Basu and Van (1998) model. Families send their children to work for pay when they cannot afford their desired alternatives such as schooling without the child's economic contribution. Our findings suggest that the BDH transfer of $15 per month allows families to continue the schooling of many of the children most vulnerable to leaving schooling for work for pay. The BDH program is described in the next section and we consider its effect on child labor in detail in section 3. The main findings are presented in section 4. We document that the randomized increase in income is associated with increased schooling and decreased work for pay in those most vulnerable to transitioning from school to work. The changes in work and school are large enough that the net effect of the transfer for this population is to decrease total expenditures. Section 5 concludes. Our findings highlight the importance of schooling costs in the decision to send the child to work and illustrate considerable scope for small, targeted changes in family income to have large effects on the child labor situation. 2. Background on the BDH program and its evaluation Ecuador has had a cash transfer program, the Bono Solidario, in place since 1998. Recipient households received $15 per month per family. While the intent of Bono Solidario was to assist poor families during an economic crisis, the program continued well past the economic crisis and the program was poorly targeted. Bono Solidario was replaced by Bono de Desarrollo Humano (BDH) beginning in mid 2003. A key difference between BDH and Bono 3 Solidario is that BDH is explicitly means-tested. Starting in 2001, the government invested into developing a family means test, called the Selben Index. Only families in the poorest two quintiles of the Selben index are eligible to receive BDH's transfer of $15 per month. Another difference between the Bono Solidario and the BDH is that the launch of the BDH was accompanied with a social marketing campaign that encouraged households to invest in the human capital of their children. However, unlike other transfer programs in Latin America, BDH transfers have never been made explicitly conditional on specific investments in child human capital (for example, school enrollment). The rollout of BDH explicitly contained a randomized component in 4 of Ecuador's 24 provinces. Within provinces selected for the evaluation, parishes (counties) were randomly drawn. Within these parishes, BDH eligible households were randomly sorted into BDH recipient households (lottery winners) and non-recipients (lottery losers). Households formerly receiving Bono Solidario transfers were excluded from the evaluation. Lottery losers were taken off the roster of households that could be activated to receive BDH transfers. An important feature of this experiment is that, unlike the PROGRESA evaluation, randomization is at the household level, rather than the community level. That is, within a community, we observe both lottery winners and lottery losers. The randomization appears to have been successful in attaining balanced treatment and control populations. Table 1 summarizes background characteristics of children and their families in treatment and control populations. These data are from the baseline survey data collected between June and August 2003 before households were assigned to treatment and control populations. Most of the background characteristics in table 1 appear similar. The control population looks a little more likely to be female and urban than does the treatment population, but these differences are not significant at 5 percent. We also consider our results on samples restricted to just rural and just female below. There appears to be considerable leakage of BDH into the control population. By design of the experiment, the control population was not supposed to receive the BDH. In reality 38 percent receive it. The treatment group is 73 percent more likely to receive BDH the control population, but the considerable leakage means that our empirical work later will need to be an intent-to-treat type analysis. 4 Several other studies have considered the impact of BDH transfers: Paxson and Schady (2008) show that transfers improved the health and development of preschool-aged children, and Schady and Rosero (2008) show that a higher fraction of transfer income is used on food than is the case with other sources of income. Most directly related to this paper, Schady and Araujo (2008) show that the program had large effects on child schooling. Though the transfers are small - 8.9 percent of expenditures in median household ­ the impact they have on children seems to be large. Time allocation is available in the BDH evaluation data for children 6 to 17. Table 2 summarizes their time allocation in the baseline data by the child's treatment status. There are no statistically significant differences in activities at baseline between the treatment and control samples. Three-fourths of children attend school, and three fourths work in domestic work such as cooking, cleaning, and other household chores. Less than half participate in market work. Most that do are engaged in unpaid family work. Less than one in ten children participate in market work for pay. 3. The effect of the BDH on child labor supply 3.1 Set-up In this section, we examine the response of child labor supply to the BDH transfer in a simple version of the model of child labor supply developed in Basu and Van (1998, hereafter BV). We consider the case of a two person household, one adult and one child. The BV model is built on two explicit assumptions. First, child and adult labor are perfect substitutes subject to a productivity shifter, a. One child worker is equivalent to adult workers, <1. This is known as the substitution axiom. wc is the child's labor income, wAis adult labor income (adult labor supply is inelastic). Equilibrium between the child and adult labor markets and the substitution axiom, together, imply wc =wA. Second, child labor occurs only if family income is very low. Denote si as the perceived subsistence level of family i. ci is the family's consumption. This second assumption, the luxury axiom, is written by BV as: (ci,0) f (ci +,1) if ci si (ci +,1) f (ci,0) if ci < si eq. (1) 5 for all > 0 . We think this characterization of preferences is most appropriately applied to the decision to send children to work outside of the family farm or business.1 Hence, we use work for pay outside of the family farm or business as our definition of child labor in the empirical portion of this study. Of course, we examine the transfer's effect on other forms of work as well in our empirical analysis. The discontinuity in child labor supply implied by the luxury axiom is central to BV's proof of the possibility of multiple equilibria for child labor supply in an economy. 2 Hence, the luxury axiom has been the focus of a large body of research. Tests of the luxury axiom typically look at whether child labor is income elastic. A negative income elasticity of child labor follows out of any Becker (1965) style time allocation model where leisure or an alternative use of time outside of work is a normal good (see for example, Cigno and Rosati 2005). In fact, the luxury axiom as codified in equation (1) does not imply that child labor is income elastic for all levels of income. Child labor does not respond to increases in income when such increases leave the household unable to cover subsistence costs, for example. Edmonds (2005) argues that, if we accept that households will vary in their perceptions of subsistence costs, the luxury axiom implies that the relationship between child labor and family living standards should be flat, and then decline rapidly in the range of perceived subsistence levels. He documents that the changes in child labor in Vietnam in the 1990s are consistent with what would be implied by the luxury axiom. The baseline data in the BDH evaluation are consistent with the basic pattern implied by equation (1) of rapid declines in child labor after a certain standard of living is attained. Figure 1 contains the plot of participation rates in work for pay outside of the household for children 6-16 against the log of annual per capita expenditures in the baseline, pre-BDH, data. The curve in 1Across countries, work for pay outside of the child's family is less prevalent and associated with lower school attendance rates than other forms of work. The lower school attendance rates among children who work for pay outside of the household across countries can be explained by differences in hours worked (Edmonds 2007). Work for pay outside of the household may be associated with less flexibility in hours worked and greater intensity for all of the reasons typically offered for the lumpiness of formal wage employment in the adult labor market. Families may view work for pay outside of the home differently than other forms of work because of its greater intensity directly, the implied incompatibility with schooling, or discomfort with having children working outside of the home. 2Swinnerton and Rogers (1999) emphasize that the multiple equilibrium result implicitly depends on the assumption that the distribution of capital income is sufficiently unequal. 6 Figure 1 is the result of a non-parametric (local-linear) regression.3 The curve does not represent a causal relationship and should not be taken as a prediction to what will happen to work outside of the child's household as they grow wealthier or receive the BDH transfer (as we discuss below). Nonetheless, the shape of the curve in figure 1 is striking. From the bottom of the distribution until monthly per capita expenditures are approximately $20 per person per month (when the log of annual per capita expenditures is 5.5), participation rates in work for pay outside of the household are roughly flat. After $20 per person per month, participation rates decline very rapidly. In the wealthiest households in the evaluation sample (drawn from the poorest two quintiles in Ecuador), participation rates in market work outside of the household are one-third of what they are for the poorest families. Heterogeneity in perceived subsistence levels is necessary to reconcile preferences such as equation (1) with the data and seems realistic. We model the family's perception of basic subsistence needs as having three components: the costs of maintaining a child h, the costs of educating the child k, and an idiosyncratic perception that is normally distributed with a mean 0 and variance 1. Both the maintenance and education costs vary with the child's age. Denote the age of the child in family i as ai . Hence, the family's perception of its subsistence needs can be written as: si = h(ai ) + k (ai ) -i eq. (2) with i N (0,1). We assume that both h and k are everywhere differentiable and increasing in child age. The assumption that maintenance costs increase with age follows from the increasing nutritional and energy requirements associated with puberty, physical growth, and increased physical intensity of work with age (over the ages relevant for our discussion). Schooling costs include school fees, schooling inputs, transport costs, etc. Conceptually, schooling costs may also include the family's perceived opportunity cost of child time in schooling if this opportunity cost is not determined by adult wages. Opportunity costs would increase with age with the child's ability to work and contribute productively. Direct schooling costs also increase with age. Figure 2 plots schooling costs per child by age for children that attend school in our data. Primary school is technically free in Ecuador, but families still face costs for transport, uniforms, 3We follow Fan and Gijbels (1995) nearest neighbor adaptive bandwidth selection rule for 60 nearest neighbors (approximately 2 percent of sample). We use a Gaussian kernel. 7 learning materials, and some other fees. Secondary school is not free and, at secondary school ages, direct schooling expenditures per child appear to be double their primary school age level (although part of this difference may be a result of differences in which families send older children to school). The family chooses consumption ci and whether the child works in the formal labor market. ei is an indicator that is 1 if the child works in the formal labor market and 0 otherwise. The household then maximizes utility (1) subject to the budget constraint: ci eiwc + wA + ti eq. (3) where ti is the household's non-labor income. Child labor supply and consumption then depend on whether adult labor income and non-labor income are enough to cover subsistence expenses. Liquidity constraints, as in Baland and Robinson (2000), are implicit within this model. Assuming non-satiation in consumption and the substitution axiom: ci = (wA++ti) if wA + ti si ((1 )w ) eq. (4) A +ti if wA + ti < si ei =01 if wA +ti si eq. (5). if wA +ti < si An interesting attribute of this model is that implies that increases in income can result in declines in consumption. Assume non-satiation in consumption so that the budget constraint (3) holds with equality: ci = (1+ei)wA +ti . Suppose wA < si so that without the transfer, the child works and expenditures are: ci = (1+)wA . Suppose that the transfer is sufficient to switch the family from having the child work to not: wA +ti si . Expenditures are then ci = wA + ti . If si - wA ti wA, then household expenditures fall with the transfer relative to what expenditures would be absent the transfer. The family is still better off. This result that the transfer can decrease family income follows directly out of the assumption on preferences in equation (1). 3.2 The effects of the cash transfer The probability a child works for pay is the probability that the family's income absent child labor is below subsistence. That is: 8 Pr(e =1) = Pr(wA +ti < si ) = Pr wA +ti < h(ai)+k(ai)-i ( ) = Pr i < h(ai)+k(ai)-wA -ti ( ) Given the assumption i N (0,1), we have Pr(ei =1) = F h ai + k ai - wA -ti ( ( ) ( ) ) eq. (6) where F is the cumulative normal. Three factors that influence whether the child works are emphasized in equation (6): the child's age, adult wages, and the transfer. Totally differentiating, we have: d Pr(ei =1) = f h ai + k ai - wA -ti ( ( ) ( ) ) h(aai) k(aai) + da -dwA -dti eq. (7) where f() is the standard normal density and is assumed to be everywhere positive. Higher maintenance and education costs increase the probability that we observe a child working. Additional adult income or non-labor transfers reduce the incidence of child labor. The change in schooling costs (associated with age) in equation (7) is: k ai a da . ( ( ) ) Changes in schooling costs have the same effect on the incidence in child labor as does an equivalently valued increase in transfers or adult income. Put another way, the implication of our set-up is that families consider whether they can afford to not send their children to work. They come to the same conclusion whether they are sufficiently rich or alternative uses of the child's time are sufficiently inexpensive. This point is important in interpreting the empirical results later in this study. Whether the BDH transfer affects child labor by increasing incomes absent child labor or lowering the perceived costs of schooling (because of misperceptions in some households that receipt of the transfer requires schooling, Schady and Araujo 2008), the BDH transfer causes the decision to forego work to become relatively more affordable. There is substantial heterogeneity between children in the effect of the transfer. Assume that the transfer does not influence equilibrium wages in the local labor market. Holding age constant, the change in the incidence of child labor in a family that starts receiving the transfer is then: d Pr(ei =1) = - f h ai + k ai - wA ( ( ) ( ) ) eq. (8) 9 The magnitude of the reduction in child labor with the BDH transfer depends on subsistence costs and adult income. Child labor supply will be more elastic to the transfer to households closer to the margin in the sense of wA si . That is, for families where adult income is well above subsistence, the transfer will not influence child labor supply, because subsistence needs is not a motive for child labor. For very poor households, with adult income well below the subsistence level, it is unlikely that a small transfer such as the BDH will be sufficient to eliminate the need to have a child work. The effect of the transfer should therefore be largest for families who are close to subsistence absent child labor. The effects of the transfer will vary with the child's age. Differentiating equation (8) with respect to age, we find: d Pr(ei =1) dti ( ( ) ( ) ) + . eq. (9) a = - f h ai +k ai -wA h(ai) k(ai) a a f' is the derivative of the density function at its argument (the difference between average subsistence costs and adult wages at age a). Both maintenance and schooling costs increase in age. Whether the impact of the transfer increases or decreases with age then depends on whether or not the child works (that is, whether adult income is above or below subsistence costs). Suppose the child works for pay at baseline, h(ai )+ k (ai ) > wA. This implies that f > 0 and thereby d Pr(ei =1) dti ( ) a <0. Among children who are working, the impact of the transfer is smaller as they age. Alternatively, suppose children are not working at baseline, h(ai ) + k (ai ) < wA. The effect of the transfer then gets larger as children age: d Pr(ei =1) dti a >0. Given that schooling costs appear to increase dramatically starting at ( ) age 12 and with the primary to secondary school transitions, this discussion suggests that the impact of the BDH should be largest for children age 10 or larger who do not work at baseline. Moreover, as the value of the normal density f is greatest when its argument is closest to zero, the impact of the transfer on children at the transition ages should be largest for poorer children whose adult income absent child labor is closer to subsistence, wA si . An examination of the baseline data is useful to identify what ages are most likely to be affected by the BDH transfer. Figure 3 pools baseline data for both the treatment and control 10 populations and plots participation rates at baseline by type of activity and age.4 Paid market work does not appear in the dataset until age 12 and does not exceed 20 percent of children until age 15. Schooling is nearly universal ages 6 through 11 and begins to decline rapidly thereafter, with the largest declines occurring between ages 12 and 13. The age patterns in schooling and paid market work are more vivid than those for work categories that take place inside the child's own home. Overall, unpaid market work occurs in the family farm or business, and by age 8, nearly 40 percent of children participate in unpaid market work. The prevalence of domestic work grows rapidly between age 6 and 9, and appears to be relatively stable thereafter. Figure 3 has two implications for our empirical work. First, given the age patterns in child time allocation it seems that there is little scope for a transfer to affect child involvement in paid work at ages 11 and under. Hence, we focus our analysis on children age 10 and above at baseline. Second, it is striking how the timing of the decline in schooling matches the growth in work for pay. This reflects the fact that when children transition from school to work, it is often paid work that they are transitioning to, especially in poorer households. Given that much of the variation in child time allocation with age in this population is from increases in work for pay and decline in schooling, we anticipate that the largest impact of the transfer will be on work for pay, rather than participation in unpaid market work in the child's home. 4. Main findings 4.1 Empirical methods Our empirical strategy approach follows from equation (6). Adult wages are determined by local labor markets. We treat the parish as the labor market and include parish fixed effects p . Maintenance and education costs vary by child age and possibly gender. We include a full set of age dummies a and a gender effect g. Non-labor transfers are affected by winning the BDH lottery. In our reduced form approach, we include an indicator for whether the family won the BDH lottery l as our measure of ti : eip = + p + a + gi + rli +ip eq. (10) 4Figure 3 ends at age 16 in the baseline data, because the evaluation did not collect time allocation information for children above 17 in the treatment period. 11 where ip is an error term that is 0 in expectation conditional on the other controls listed in equation (10). Standard errors are clustered at the parish level. In our preferred specification, we estimate equation (10), replacing li with an indicator for whether the family receives the BDH transfer as our measure of ti (the transfer does not vary among recipients). 38 percent of control households report receiving the BDH transfer, despite their exclusion from it in the evaluation design. Hence, take-up of the BDH transfer is non- random but the lottery is random by design. The lottery nearly doubles the probability that a household reports receiving the BDH transfer. Hence, we use the lottery indicator li as an instrumental variable for ti . The assumption in the instrumental variable specification is that the fact of winning the lottery does not itself influence child labor decisions beyond its effect on BDH take-up. We examine the impact of winning the BDH lottery on child time allocation for each of the different indicators of child time allocation described in table 2. For each outcome, e, we report reduced forms of winning the lottery and instrumental variables estimates of the impact of receiving the transfer. Our discussion of age trends above suggests that we would be unlikely to find an effect of the cash transfer at ages below 12. On average, there is 1.5 years between pre and post baseline data. Conservatively, we then expect the treatment to only be evident in children as young as 10 in baseline. We bifurcate the sample into above and below age 10 in estimating (10) Our theoretical discussion suggests that the effect of the transfer should be largest on the child on the margin between schooling and work. We identify marginal children in three ways. First, children working at baseline are revealed to have subsistence needs above adult income absent the transfer. With age, subsistence needs increase. Given the small value of the transfer, we expect its effect on work to be largest among children who are not working at baseline but are vulnerable to transitioning to work. Our first additional restriction is to look at children age 10 and above at baseline who attend school without participating in market work. Second, children are most likely to transition from school to work at the end of primary as direct schooling costs increase dramatically and perhaps there is a labor market return to primary. Our second additional restriction (in addition to greater than age 10 and older at baseline, in school, and not working) is to focus on children near or beyond the end of primary. We limit the sample to 12 children who do not work and are in grade five or higher. (There are six grades in primary school in Ecuador.) Third, we block children by the probability that they transition from this no work / schooling group to working. Specifically, for children 10 and older, we restrict the sample to the control population. We regress an indicator that a child participates in market work in the post round on age effects, gender effects, urbanity, parish fixed effects, baseline time allocation, and a second order polynomial in the natural log of per capita expenditures that we allow to vary with age*gender. We use the predicted probabilities from this regression to divide the entire evaluation sample into three equally sized groups: those with low, middle, and high probabilities of transitioning to work for pay. 4.2 At school to work transition ages, market work and work for pay decline with additional income Children age 10 and older who receive the additional BDH transfer income are less likely to work in market work, work for pay, unpaid market work. They are more likely to work in domestic work. The increase in domestic work is smaller than the decline in market work, and is similar in magnitude to the decline in work for pay outside of the child's family. The probability that a child works without enrolling in school decreases. The probability that a child enrolls in school increases. These findings are in table 3 for the full sample age 10+, separately for boys and girls, and separately for rural and urban. Each cell in table 3 contains the result from a different regression. For each population grouping (e.g. full sample, male, urban, etc.), the first line contains the reduced form coefficient on the lottery winner indicator from estimating equation (10) with the dependent variable indicated by the column header. The second line contains the coefficient from a separate regression of the dependent variable indicated by the column on an indicator for whether the family receives the BDH, with random assignment used as an instrumental variable. Thus, winning the BDH lottery is associated with an 8 percentage point decline in the probability a children age 10+ at baseline works in market work in the post round, a 15 percent decline. Slightly less than a third of those that win the BDH lottery do not take-up the transfer, and more than one-third who should not receive the transfer do so. The impact of actually receiving the BDH, correcting for the endogeneity in this decision, suggests that receipt of the BDH reduces 13 market work by 24 percentage points, or just under 50 percent. The BDH income increases school enrollment by 18 percentage points or 26 percent. Changes in market work, especially work for pay, associated with BDH income increase substantially when we focus our analysis on children that are most vulnerable to a school to work transition. In table 4 and remaining tables, we only present instrumental variables results of the effect of BDH receipt for economy. We limit the population used to estimate table 4 to children 10 and older who attend school and do not work in baseline. As discussed above, these children are most likely to be affected by the transfer. In fact, we observe a 44 percentage point decline in market work and a statistically significant 20 percentage point decline in the probability that a child participates in work for pay. The magnitude of the increase in domestic work (although not statistically significant) also increases. This implies a degree of substitutability between work for pay and domestic work that is not typically found in the data. School enrollment is also more elastic to the BDH transfer in this vulnerable group relative to the general age 10+ population. Magnitudes are still greater when we focus on children who are most likely to complete primary school or higher during our evaluation period. Table 5 presents our main findings of the impact of the BDH transfer for children 10 and older, in school in grade 4 or higher, and not working at baseline. Market work declines by nearly 50 percentage points and work for pay by nearly 30 percentage points. School enrollment is increased in this group by nearly 40 percentage points relative to the comparable population that does not receive the transfer. Effects on work for pay and schooling are largest for girls and in rural areas. These results are consistent with the overall idea that the effects of the additional income are largest for children that are most close to the margin of subsistence as both rural areas and girls have faster school to work transitions than urban areas and boys. We see much smaller effects of the BDH transfer in populations that we expect to be further from the margin of subsistence. These findings are in table 6. Children under 10 at baseline experience much smaller changes in time allocation compared to children 10 and over. This is to be expected as perceptions of subsistence will generally be lower because of lower maintenance costs and direct and indirect schooling costs. Hence, younger children are more likely to be in school regardless of the transfer. Children that are working at baseline also are less impacted by the transfer. Children 10 and older that are working and not attending school at baseline already have revealed that their subsistence level is above adult income absent the 14 child's labor. This difference will only grow with the child's age, and it seems unlikely that a small transfer will radically affect the family's perception of the need for the child's work. In fact, we see little change in child time allocation for this population (the second row of table 6). Children that are away from the end of primary school are also less likely to be impacted by the transfer. They face lower direct schooling costs (and perhaps indirect costs if there is a return to primary completion). Hence, they are less likely to transition out of school to work in the 1.5 years between the pre and post periods. Changes in time use associated with BDH receipt for this population are in the bottom row of table 6. While they do experience declines in unpaid market work (work in the family farm or business) that is greater in magnitude than the population completing primary school or higher during the evaluation period (table 5), this school group does not experience the changes in work for pay or school enrollment observed in the population that completes primary school during this evaluation period. That is, for this relatively less schooling-advanced population, the transfer seems to forestall participation in the family farm and business, but this group is not especially vulnerable to transitioning to work for pay and no schooling. Schooling is most likely to be orthogonal with paid rather than unpaid market work for a variety of reasons, including flexibility in the time and the hours worked. Hence, the effect of the transfer on types of work that is most incompatible with schooling is largest amongst the population transitioning from primary school or higher grades. The analysis associated with tables 4 through 6 is based on inferring who is most likely to be transitioning from schooling to work for pay, based on baseline age and school / work status. An alternative is to estimate the child's transition probability directly using the control sample and taking into account the child's family's baseline living standards. These results are in table 7 where we trifurcate children into groups based on the probability that they transition from schooling to work for pay during the evaluation period. While the declines in market work are similar in all groups, this is driven by declines in work in the family farm or business in most of the population. It is only the group that is most vulnerable to transition where the decline in work is concentrated in work for pay. Similarly, it is only this group with the highest probability of transitioning from school to work where we see schooling effects of the transfer as well. Thus, the effect of the transfer on child labor and schooling can be very large, but its effects seem fairly concentrated in one segment of the population. 15 4.3 School expenditures increase but per capita expenditures do not increase significantly with additional income The BDH transfer is $15 per family per month. Our estimates imply that families spend an additional $42 per child per year on schooling as a result of the transfer. This is $4.7 per month of the school year (the school year is 9 months). With 1.8 children per recipient household, more than half of the transfer is spent on the direct schooling costs of children. These findings are in table 8. Table 8 mimics table 3 in its construction. Each cell in the table comes from a different regression. Instrumental variables are presented for the full sample, male/female, and urban/rural. The first column contains the simple difference in annual school expenses per child in the post sample. The second column looks at changes in annual schooling expenditure per child between the pre and post periods. Given random assignment, it is expected that the first difference results look qualitatively similar to the double difference results in column 2, and we see this. Hence, we focus on the first difference results in column 1 for discussion. The increase in schooling expenditures is largest for girls and in rural areas. This is similar to the patterns in school enrollment observed in table 3. Total expenditures do not appear to increase substantively with the cash transfer despite greater school expenditures.5 The dependent variable in columns 3 and 4 is total annual household expenditures. In the full sample, an additional $15 per month is associated with a $213 reduction in annual expenditures, or $18 less per month. The decline in total expenditures is largest for households with girls; girls also experience larger increase in school expenses (column 1) and the largest decline in work for pay (tables 3-5). In fact, the decline in expenditures appears largest in families of children that were most vulnerable to transitioning from schooling to work for pay. This finding is explicit in table 9 which mimics the trifurcation of the data from table 7. Schooling expenditures per child increase by $77 per child per year with BDH receipt ($8.5 per month, $15.4 per household per month) among these children whose probability of attending school increases by 43 percentage points. Annual household expenditures decline by $430, $36 per month. Thus, while schooling expenses increase considerably relative to the control population with these children whose 5Schooling expenditures are included in total expenditures. 16 schooling status is protected with the transfer, overall their family seems to forego considerable consumption in order to send children to school rather than work for pay. These expenditure results are consistent with the results from table 7. In households most likely to transition to work, the BDH is associated with an additional $77 in schooling expenses per year, or $8.5 per month per child. There are 1.8 children, so total expenditures on education rise by $15. However, this increase in education experiences comes with a decline in income. The average monthly wage for a child in the control population working for wages in the post period is $84 per month. The probability a child works for pay declines by 37 percentage points for children most likely to transition to work. On average, then, a 37 percentage point decline in the probability that a child works earning $84 per month costs a household $31 in foregone income per child or $56 for its 1.8 children. The additional transfer income of $15 per month implies that total household income declines by $41 per month assuming no other behavioral changes. This is $5 above the $36 decline directly estimated in the data in table 9, one hundredth of the standard error on the estimated $36 decline. The model of section 3 posits a simple explanation for this surprising observation. Absent the transfer, families do not perceive themselves as able to cover their subsistence needs without child labor. The BDH transfer makes it more affordable for the family to continue the child's schooling, especially among families most likely to transition children from school to work. 5. Conclusion Work for pay among children in Ecuador is concentrated in the poorest households, and children appear to transition from school to work for pay starting at age 12. We find that a randomly assigned cash transfer maintains school enrollment and leads to a decline in work for pay among children vulnerable to transitioning from school to work. The declines in work for pay and increases in schooling, relative to the control population, are largest for girls and in rural areas. The additional income appears to have little influence on child time allocation or schooling related expenditures for children below school transition ages or already working and out of school in our baseline data. Among children vulnerable to transitioning from school to work, we observe a substantial increase in school related expenditures. Most of the cash transfer appears to be spent on schooling in this population. Despite increased school expenditures, the 17 decline in work for pay is large enough that total expenditures decline in families with children whose schooling is prolonged by the transfer. The decline in total expenditures is very close in magnitude to the foregone income implied by our estimates of the decline in work for pay. The Basu and Van (1998) set-up offers a simple theoretical interpretation of these results. It posits that children work when families feel they cannot afford alternatives to that work. For families vulnerable to transitioning from school to work, the transfer improves their ability to afford schooling. Hence, they forego the child labor income, total expenditures decline, and families are better off as a result.6 It is striking that for children vulnerable to transitioning from school to work, families appear to use all of the transfer to support child schooling and defer the transition to work for pay. To the extent that this educational investment is productive and multiplier effects from education are substantive, our finding suggests potentially very high aggregate returns to small, well-targeted transfers. But why would families invest so heavily to sustaining the education of these children at transition ages? The Basu-Van set-up frames the answer in terms of preferences, but why might preferences be such? The answer might depend on whether the transfer is perceived as transitory or permanent. If it is transitory, our findings are consistent with education as the highest return savings vehicle available to these poor families. If permanent, our findings may reflect nothing more than parental preferences (education or the absence of child labor brings utility directly) or something about liquidity constraints (education is the best long-term investment available and liquidity constraints were constraining investment). Of course, surrounding the transfer program is a social marketing campaign promoting investments in child human capital. It is possible that the results herein reflect a behavioral response to a combination of the cash transfer and the social marketing campaign. We do not have a comparable experiment to suggest how families would treat a similarly sized lottery award without the social marketing context. Understanding why families appear to prioritize schooling and the absence of child labor to the extent observed here is an interesting avenue for future research. 6This is a result of the "luxury axiom" in the model. The Basu and Van setting frames the child labor decision in the language of preferences. However, it is trivial to recharacterize the model as being one where child labor is driven by liquidity constraints, as in the Baland and Robinson (2000) model. 18 References Attanasio, O., E. Fitzsimons, A. Gómez, D. López, C. Meghir, and A. Mesnard (2006), "Child education and work choices in the presence of a conditional cash transfer programme in rural Colombia." IFS Working Paper, W06/01. London: Institute for Fiscal Studies. Baland, J. and J. A. Robinson (2000), "Is child labor inefficient?", Journal of Political Economy 108: 663-679. Basu, K. and P. Van (1998), "The economics of child labor", American Economic Review, 88: 412-427. Basu, K., S. Das, and B. Dutta (2007), "Child labor and household wealth: Theory and empirical evidence of an inverted-U," Working Paper no. 139 (Bureau for Research and Economic Analysis of Development Working Paper, Cambridge MA). Becker, G. (1965), "A theory of the allocation of time", Economic Journal 75: 493-517. Cigno, A. and F. Rosati (2005), The economics of child labour, (Oxford University Press, Cambridge). de Janvry, A., and E. Sadoulet (2006), "Making conditional cash transfer programs more efficient: designing for maximum effect of the conditionality", World Bank Economic Review 20(1): 1-29. Doepke, M. and F. Zilibotti (2005), "The macroeconomics of child labor regulation", American Economic Review 95: 1492-1524. Edmonds, E. (2005), "Does child labor decline with improving economic status?", Journal of Human Resources 40: 77-99. Edmonds, E. (2006), "Child labor and schooling responses to anticipated income in South Africa," Journal of Development Economics 81(2): 386-414. Edmonds, E. (2007), "Child Labor," in T.P. Schultz and J. Strauss, eds., Handbook of Development Economics, (Elsevier Science, Amsterdam, North-Holland) 3607-3710. Fafchamps, M. and J. Wahba (2006): "Child labor, urban proximity, and household composition", Journal of Development Economics 79: 374-397. Fan, J. and I. Gijbels. 1995. Local Polynomial Modelling and Its Applications. New York: Chapman & Hall. Field, E., (2007), "Entitled to work: Urban tenure security and labor supply in Peru", Quarterly Journal of Economics 122(4): 1561-1602. Filmer, D., and N. Schady (2008), "Getting girls into school: evidence from a scholarship program in Cambodia", Economic Development and Cultural Change 56(3): 581-617. Kruger, D. (2007), "Coffee production effects on child labor and schooling in rural Brazil", Journal of Development Economics 82: 448-463. Manacorda , M. (2006), "Child labor and the labor supply of other household members: Evidence from 1920 America", American Economic Review 96: 1788-1800. 19 Manacorda, M. and F. Rosati (2007), "Local labor demand and child labor", Working Paper (Understanding Children's Work Project, Rome). Mueller, E. (1984), "The value and allocation of time in rural Botswana", Journal of Development Economics 15: 329-360. Paxson, C., and N. Schady (2008), "Does money matter? The effects of cash transfers on child development in rural Ecuador", unpublished manuscript, Princeton University and the World Bank. Ranjan, P. (2001), "Credit constraints and the phenomenon of child labor", Journal of Development Economics 64: 81-102. Ravallion, M., and Q. Wodon (2000), "Does child labour displace schooling? Evidence on behavioural responses to an enrollment subsidy", The Economic Journal 110(462):158- 175 Rosenzweig, M. and R. Evenson, (1977), "Fertility, schooling, and the economic contribution of children in the rural India: An econometric analysis", Econometrica 45: 1065 ­ 1079. Schady, N. (2004), "Do macroeconomic crisis always slow human capital accumulation?", World Bank Economic Review 18: 131-154. Schady, N. and M. C. Araujo (2008), "Cash transfers, conditions, and school enrollment , and child work: Evidence from a randomized experiment in Ecuador", Economía 8(2): 43- 70. Schady, N., and J. Rosero (2008), "Are cash transfers made to women spent like other sources of income?", forthcoming, Economics Letters. Schultz, T. P (2004), "School subsidies for the poor: evaluating the Mexican Progresa poverty program", Journal of Development Economics 74(1): 199-250. Swinnerton, K. and C. Rogers (1999) "The economics of child labor: Comment", American Economic Review 89: 1382-1385. 20 12 nelies 11 Ba 10 atetaR 9 iontapic 8 rtiaPyaP 7 for 6 rko W 5 4 4.7 4.9 5.1 5.3 5.5 5.7 5.9 6.1 6.3 6.5 6.7 6.9 7.1 7.3 7.5 Ln Per Capita Expenditures at Baseline Figure 1: Work for Pay outside of the Household and Per Capita Expenditures at Baseline Local Linear Regression with 28 (1% of sample) Nearest Neighbor Adaptive Bandwidth Selection 21 200 175 student per 150 125 expenditure school 100 Annual 75 50 6 8 10 12 14 16 Age Figure 2: Annual School Expenditure for Enrolled Students by Age at Baseline 22 100 80 Rate 60 40 Participation 20 0 6 8 10 12 14 16 Age Enrolled in School Paid Market Work Unpaid market work Domestic Work Figure 3: School, Paid and Unpaid Market Work, and Domestic Work by Age at Baseline 23 Table 1: Child and Family Background Characteristics Variable Treatment Control Sample Size 1567 1437 Age 13.78 13.82 Male 0.46 0.52 Ever Married 0.00 0.00 Disabled 0.01 0.01 Indigenous 0.10 0.09 Father's Education 4.58 4.45 Mother's Education 3.63 3.51 Household Size 6.08 6.08 Ln Per Cap Expend 5.99 6.00 Rural 0.52 0.50 Receives BDH 0.68 0.38 Aside from the indicator that the child's family receives BDH, no treatmentcontrol differences are significant at 5% or less. Rural difference is significant at 10 percent. Table 2: Child Time Allocation at Baseline Full Sample Age 10 and older Variable Treatment Control Treatment Control Enrolled in School 0.767 0.770 0.696 0.704 Highest Grade Completed 4.620 4.589 5.693 5.650 Any Market Work 0.475 0.463 0.521 0.514 Paid Market Work 0.092 0.094 0.121 0.123 Unpaid market work 0.404 0.389 0.428 0.417 Domestic Work 0.798 0.787 0.823 0.827 Works without school 0.214 0.213 0.284 0.280 No treatmentcontrol differences are significant at 30 percent or lower. Table 3: Changes in Time Allocation with the BDH Children 10 and older at baseline (1) (2) (3) (4) (5) (6) Unpaid Market Work for Market Domestic Work w/o Enrolled in Work Pay Work Work School School Full Sample Randomization (Reduced 0.0778** 0.0228 0.0667** 0.0157 0.0634** 0.0581** Form) (0.0231) (0.0166) (0.0203) (0.0196) (0.0196) (0.0222) Receives BDH (2SLS) 0.244** 0.0716 0.209** 0.0492 0.199** 0.182** (0.0719) (0.0537) (0.0608) (0.0634) (0.0652) (0.0764) Male Randomization (Reduced 0.0941** 0.0140 0.0778** 0.0161 0.0338 0.0226 Form) (0.0333) (0.0304) (0.0246) (0.0279) (0.0348) (0.0377) Receives BDH (2SLS) 0.294** 0.0439 0.243** 0.0505 0.106 0.0707 (0.105) (0.0962) (0.0705) (0.0900) (0.108) (0.118) Female Randomization (Reduced 0.0604** 0.0288 0.0504* 0.0180 0.0906** 0.0875** Form) (0.0255) (0.0213) (0.0268) (0.0221) (0.0272) (0.0314) Receives BDH (2SLS) 0.191** 0.0910 0.159* 0.0569 0.286** 0.276** (0.0798) (0.0657) (0.0865) (0.0728) (0.0845) (0.105) Rural Randomization (Reduced 0.0686** 0.0406 0.0443 0.0165 0.0583** 0.0587* Form) (0.0331) (0.0252) (0.0280) (0.0305) (0.0264) (0.0312) Receives BDH (2SLS) 0.220* 0.130 0.142 0.0528 0.187* 0.188 (0.112) (0.0851) (0.0882) (0.102) (0.0979) (0.117) Urban Randomization (Reduced 0.0936** 0.00651 0.0925** 0.0175 0.0698** 0.0592* Form) (0.0319) (0.0220) (0.0271) (0.0246) (0.0320) (0.0320) Receives BDH (2SLS) 0.296** 0.0205 0.292** 0.0553 0.221** 0.187* (0.0951) (0.0699) (0.0837) (0.0781) (0.0948) (0.103) Notes: 1 Each cell contains the result from a different regression. The column indicates the dependent variable. The coefficient labeled "randomization" is the reduced form coefficient on an indicator that the household won the BDH lottery (equation 10 in the text). Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity when feasible. 3 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 4 * p<0.10, ** p<0.05 Table 4: Changes in Time Allocation with the BDH for children not working and in school at baseline Children 10 and older, in school, and not engaged in market work at baseline (1) (2) (3) (4) (5) (6) Unpaid Market Work for Market Domestic Work w/o Enrolled in Work Pay Work Work School School Full Sample Receives BDH (2SLS) 0.440** 0.199** 0.247** 0.157 0.277** 0.299** (0.136) (0.0946) (0.103) (0.115) (0.0934) (0.114) Male Receives BDH (2SLS) 0.505** 0.175 0.233 0.331 0.243 0.250 (0.233) (0.180) (0.169) (0.226) (0.187) (0.203) Female Receives BDH (2SLS) 0.417** 0.239** 0.241 0.131 0.361** 0.394** (0.145) (0.113) (0.149) (0.117) (0.155) (0.180) Rural Receives BDH (2SLS) 0.677** 0.379* 0.233 0.207 0.368* 0.377* (0.329) (0.224) (0.240) (0.343) (0.191) (0.207) Urban Receives BDH (2SLS) 0.416** 0.137 0.315** 0.156 0.246** 0.272* (0.163) (0.0999) (0.142) (0.0946) (0.118) (0.146) Notes: 1 Each cell contains the result from a different regression. The column indicates the dependent variable. Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity when feasible. 3 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 4 * p<0.10, ** p<0.05 Table 5: Changes in Time Allocation with the BDH for children that do not work and are close to the end of primary school Children 10 and older, in school in grade 4 or higher, and not in market work at baseline (1) (2) (3) (4) (5) (6) Unpaid Market Work for Market Domestic Work w/o Enrolled in Work Pay Work Work School School Full Sample Receives BDH (2SLS) 0.493** 0.290** 0.166 0.117 0.376** 0.378** (0.148) (0.122) (0.126) (0.145) (0.109) (0.123) Male Receives BDH (2SLS) 0.600* 0.298 0.105 0.230 0.346* 0.323 (0.329) (0.273) (0.258) (0.310) (0.204) (0.241) Female Receives BDH (2SLS) 0.559** 0.326** 0.275 0.171 0.383** 0.400* (0.185) (0.155) (0.169) (0.151) (0.188) (0.207) Rural Receives BDH (2SLS) 0.663* 0.529* 0.00269 0.0946 0.510** 0.512** (0.355) (0.268) (0.265) (0.370) (0.168) (0.206) Urban Receives BDH (2SLS) 0.507** 0.211 0.304* 0.147 0.339** 0.340** (0.189) (0.140) (0.159) (0.147) (0.152) (0.165) Notes: 1 Each cell contains the result from a different regression. The column indicates the dependent variable. Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity when feasible. 3 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 4 * p<0.10, ** p<0.05 Table 6: Changes in Time Allocation with the BDH, Counterfactuals Various subsamples (1) (2) (3) (4) (5) (6) Unpaid Market Work for Market Domestic Work w/o Enrolled in Work Pay Work Work School School Children under 10 at baseline Receives BDH (2SLS) 0.117 0.0205 0.143 0.0827 0.0853** 0.0401 (0.125) (0.0275) (0.123) (0.170) (0.0423) (0.0520) Children 10 and older, in market work, and not in school at baseline Receives BDH (2SLS) 0.0385 0.118 0.102 0.120 0.0129 0.000574 (0.102) (0.118) (0.129) (0.117) (0.134) (0.123) Children 10 and older, in school in grade 3 or lower, and not in market work at baseline Receives BDH (2SLS) 0.454 0.00548 0.494* 0.380 0.00775 0.0906 (0.318) (0.130) (0.295) (0.380) (0.254) (0.286) Notes: 1 Each cell contains the result from a different regression. The column indicates the dependent variable. Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity. 3 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 4 * p<0.10, ** p<0.05 Table 7: Effect of BDH on time allocation by estimated transition probabilities Children 10 and older at baseline Transition Probability: Lowest Middle Highest Participation Indicator For: Market Work Receives BDH (2SLS) 0.218** 0.316** 0.245* (0.101) (0.0907) (0.134) Work for Pay Receives BDH (2SLS) 0.0674 0.00543 0.365** (0.0493) (0.0707) (0.130) Unpaid Market Work Receives BDH (2SLS) 0.262** 0.336** 0.00634 (0.106) (0.0862) (0.104) Domestic Work Receives BDH (2SLS) 0.0101 0.0447 0.121 (0.100) (0.0936) (0.115) Work w/o School Receives BDH (2SLS) 0.140 0.127 0.447** (0.120) (0.0831) (0.140) Enrolled in School Receives BDH (2SLS) 0.0804 0.125 0.426** (0.133) (0.0958) (0.162) Notes: 1 Each cell contains the result from a different regression. The row heading (column 1) indicates the dependent variable. Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity. 3 The full sample of children age 10 and older at baseline is trifurcated based on the predicted probability the child starts working in paid market work outside of the family's home during the post survey. The predicted probability is computed by using the control sample to regress an indicator that a child works for pay in the post round on parish fixed effects, urbanity, indicators for the child's time allocation at baseline and whether the child attended grade 5 or higher at baseline, gender effects, age effects, and a second order polynomial in baseline per capita expenditures that is allowed to vary for each agegender group. The entire post population is then divided into three categories based on the resulting predicted probabilities from this regression. "Lowest" indicates the bottom 33 percent of predicted values, "middle" refers to the middle third of predicted probabilities, and "highest" refers to the top third of the population most likely to start working for pay between the baseline and post rounds. 4 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 5 * p<0.10, ** p<0.05 Table 8: Changes in Household Expenditures with the BDH Children 10 and older at baseline (1) (2) (3) (4) Per Child School Expenditures Total Household Expenditures Difference in Difference in First Difference Differences First Difference Differences Full Sample Receives BDH (2SLS) 41.64* 43.40** 213.3 168.5 (23.62) (21.04) (338.2) (340.0) Male Receives BDH (2SLS) 11.63 26.28 25.42 13.89 (29.91) (26.39) (321.6) (333.0) Female Receives BDH (2SLS) 58.57* 48.85 353.5 277.0 (32.55) (31.39) (480.5) (471.4) Rural Receives BDH (2SLS) 65.12** 47.61 118.4 106.4 (29.79) (28.36) (426.8) (409.3) Urban Receives BDH (2SLS) 18.50 39.94 305.5 243.4 (32.31) (29.15) (473.1) (488.7) Notes: 1 Each cell contains the result from a different regression. The column indicates the dependent variable. The coefficient labeled "Receives BDH" reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. Columns labeled first difference use only expenditure data from the followup survey. Columns labeled difference in differences report changes in expenditures from baseline to followup. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity when feasible. 3 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 4 * p<0.10, ** p<0.05 Table 9: Changes in Household Expenditures with the BDH by estimated transition probabilities Children 10 and older at baseline Transition Probability: Lowest Middle Highest Per Child School Expenditures Receives BDH (2SLS) 34.62 40.60 77.19** (49.16) (37.83) (37.59) Total Household Expenditures Receives BDH (2SLS) 158.9 232.5 430.0 (430.2) (495.8) (470.3) Notes: 1 Each cell contains the result from a different regression. The row heading (column 1) indicates the dependent variable. Receives BDH reports the coefficient on an indicator that the child's family receives the BDH, using the random assignment as an instrumental variable. 2 All regressions include parish fixed effects, a vector of age dummies, and controls for gender and urbanity. 3 The full sample of children age 10 and older at baseline is trifurcated based on the predicted probability the child transitions to working for pay described in the text and in table 7. 4 Standard errors in parentheses. Standard errors corrected for Parish level clustering. 5 * p<0.10, ** p<0.05