Income Elasticity of Child Labor: Do Cash Transfers have an Impact on the Poorest Children?

According to the International Labour Organization (ILO, 2017), 152 million children aged 5–17 were involved in child labor in 2016: 10% of children in this group. Most of the working children live in low- and middle-income countries. Sub-Saharan Africa shows the highest incidence of child labor, with one in five children involved in it.

Living standards of the household, market imperfections, availability of schools, and relative returns to education and work play a crucial role on the allocation of children’s time.

For a review, see Edmonds (2005) and Cigno and Rosati (2005).

A recurring debate on the causes of child labor concerns the role of poverty as one of its main determinants. Understanding whether household income affects child labor is essential for policy design and in particular for cash transfers (CTs), one of the main social protection instruments advocated (and implemented) also to combat child labor. Both targeting and transfer size are essential elements of CTs (conditional or not) and, of course, considering the heterogeneity of response at different levels of household income is an important element for their efficient design.

Cross-sectional and cross-countries studies, however, do not indicate the presence of a substantial child labor income elasticity. As pointed out in the study by Edmonds (2005), this can be due to endogeneity problems and also the inherent nonlinearity in the relationship between child labor and income. In fact, assuming that the income elasticity of leisure is close to zero for poor households, the Basu and Van’s (1998) subsistence hypothesis implied that an increase in income affects child labor supply so far as it allows households to raise above subsistence. The empirical evidence on the nonlinearity of income elasticity is not very large. Edmonds (2005) showed that the effect of the increase in income in Vietnam, between 1993 and 1998, was present mainly for the households that moved above the poverty line, and hence it was limited for households who experienced an increase in income not sufficient to bring them above the poverty line. A few impact evaluation studies, mainly focused on conditional CTs (CCTs) in Latin America, study heterogeneity by income, but the issue remains unsettled. They found different results: Galiani and McEwan (2013) for Honduras and Sparrow (2007) for the Jaringam Pengaman Social in Indonesia found larger reduction of child labor and increase in school enrolment among the children belonging to the relatively poorer household. Glewwe and Olinto (2004) for the Honduran PRAF-II and Dammert (2009) for the Social Safety Net in Nicaragua did not find any heterogeneous impact, while Ranzani and Rosati (2004) found that the impact of Oportunitades in Mexico on child labor increases slightly with the level of household income.

In this study, we provide novel evidence on the nonlinear relationship between child labor and income and, to our knowledge, the first causal study on the effects on child labor of unconditional CT (UCT). Assessing the impact of a UCT provides solid evidence about the nonlinearity of the income effect as it generates a pure income effect, without additional changes in the budget sets (as, e.g., those induced by a CCT).

We show in a simple theoretical model, built in the Basu and Van’s (1998) spirit, that the income elasticity of child labor is null for extremely poor households, and it becomes negative (and decreasing with income) for relatively less poor households and finally becomes zero as households become more affluent. We will also discuss how CCTs are more likely with respect to UCTs to affect also the poorest households, thus identifying a possible reason for the differences between our results and some of those in the literature.

We make use of an experiment relative to an UCT in Lesotho and offer new evidence on the income elasticity of child labor and on the effectiveness of CT for the poorest households. Our results indicate that at least in poor rural communities, UCT programs lead to an impact that is consistent with the theoretical predictions outlined: extremely poor households do not change children’s time allocation, while relatively less poor households reduce child labor and increase school attendance.

We estimate the impact of Phase 1—Round 2 of the Child Grant Programme (CGP), randomly assigned to poor households in Lesotho from 2011 to 2013. Using survey data from an experimental evaluation, we find that it generates an increase in consumption expenditure for children (uniforms and shoes) in extreme poor households. In less poor households, children reduce their participation in economic activities by 17% and work on average 3 hours less per day and almost 1 day less per week as a consequence of the CGP.

Our results contribute to the limited existing literature by offering causal evidence relative to the nonlinearity of the relationship between child labor and income and by extending the evidence on the heterogeneous effects of CTs. In particular, we show that the effectiveness of CTs does not necessarily increase with the level of deprivation of the household. Moreover, to the best of our knowledge, this is the first attempt to analyze heterogeneous impacts by income on child labor of UCT in sub-Saharan Africa. As discussed, the literature on heterogeneous effects by income is mainly focused on conditional transfers in middle-income countries characterized by higher urbanization and higher child employment in paid activities outside the household. Instead, the CGP is an unconditional transfer and has been implemented in rural areas of Lesotho, where children are mainly involved in farming and livestock activities inside the household.

From a policy point of view, our results confirm that targeting is very important for insuring the effectiveness of CTs. However, to obtain the desired effects in terms of child labor reduction, the transfers should be large enough to modify household behavior. Some forms, even simplified, of means testing of the amount of benefits would be necessary to improve the efficiency of the transfer. Our findings also point to the need to pay more attention to the size of the transfer in assessing its impact, a somehow obvious point but often neglected in the literature.

See De Hoop and Rosati (2014).

We also test for the presence of spillover effects, by assessing the impact of the CGP on non-eligible children. The results indicate the absence of spillover effects, as neither child labor nor child education in non-eligible households were influenced by the program.

This article is organized as follows. The next section presents the theoretical framework. Section 3presents the program, the experimental design, and the implementation of the CGP. The data, the descriptive statistics, and the balance analysis are discussed in Section 4. Section 5 illustrates the estimation approach and Section 6 presents the results and robustness checks. Concluding remarks are presented in Section 7.

We develop a simple overlapping generation model in the spirit of Basu and Van (1998), but without assuming an exogenous level of subsistence consumption. In particular, we consider a two-period overlapping generation model in which adult household members value current household consumption and children’s future consumption. The latter is assumed to be a function of the investment in education in the first period. Adult labor supply is assumed to be inelastically fixed. Parents decide about children’s time allocation between work and education.

For simplicity of exposition, we do not consider that time can also be allocated to leisure. This assumption will not change our results in a substantial way and the implication of relaxing it will be discussed later.

To keep the exposition simple, we make several additional assumptions. Fertility is assumed to be exogenous. More importantly, we assume that households cannot save or borrow. To allow for savings will not alter the results, while in absence of credit constraints, decisions relative to consumption and investment in education would be separable

See, inter alia, Cigno and Rosati (2005).

As mentioned, the unitary household derives utility from current consumption, C₁, which includes parents’ and children’s consumption, and from children’s future consumption, C₂. Children have 1 unit of time that can be allocated either to work, H, remunerated with a wage w, or to schooling, S. Beside its opportunity cost, education also has a direct cost of e. Children’s future consumption is a concave function of the human capital accumulated through education, g(S). We assume that individuals have an innate amount of human capital, so that g(0) = k > 0. Parents inelastically supply 1 time unit of work in period 1. Labor income plus any additional nonlabor income constitute the resources available to the household in the first period, Y₁. The households’ maximization problem can hence be written as follows:

where U(.) is a concave utility function with U(.)' > 0 and U"(.) < 0 and τ is an UCT. Expressing child labor supply in terms of schooling and allowing for corner solutions, the Lagrangian function, the First Order Conditions (FOC) and the complementary slackness conditions are:

The maximization problem has three possible solutions: one interior solution and two corner solutions. Taking as given all the other parameters of the model, it is easy to see that the solution depends on the level of Y₁. There is a level of Y₁, Y*, such that for Y₁ < Y*, we have λ₂ > 0, S=0,g′U′C2<U′C1(e+w),and∂S∂τ=0. $S=0,{g}'{{{U}'}_{C2}}<{{{U}'}_{C1}}\left( e+w \right),\ \text{and}\frac{\partial S}{\partial \tau }=0.$ The level of resources of the household is very low, given the other parameters of the model, that the household allocates the time of their children only to work and a marginal change in income does not affect such allocation. For Y** > Y₁ > Y*, we have an interior solution and children’s time is allocated according to:

with H > 0, S > 0, λ₁ = 0, λ₂ = 0. The amount of time dedicated to each activity is determined as to equate the marginal rate of substitution between current and future consumption to the relative price of future consumption. In this case, it is easy to see ∂S∂τ>0and∂S∂τ∂τ<0⋅$\frac{\partial S}{\partial \tau }>0\text{and}\frac{\partial S}{\partial \tau \partial \tau }<0\cdot$

As Y₁ grows above Y**, we have the other corner solution with λ1>0,S=1,g’U’C2>U’C1${{\lambda }_{1}}>0,\ S=1,\,g{'}U{{'}_{C2}}>U{{'}_{C1}}$(e+w)and, obviously, ∂S∂τ=0⋅ $\left( e+w \right)\ \text{and, obviously, }\frac{\partial S}{\partial \tau }=0\cdot $ The households with relatively higher income send their children only to school to transfer as many resources to the future as the time constraint allows.

In conclusion, this simple model indicates that very poor households do not send their child to school at all; moreover, a marginal increase in current income does not change their behavior (unless the increase is such that Y₁ + dτ > Y*). For relatively less poor households an increase in income reduces child labor and increases schooling, but at a decreasing rate up to a point where children completely stop working.

In other words, a UCT causes heterogeneous effects on work and education according to the level of income of the household, with null effects for extremely poor households and negative (for work) but decreasing effects for relatively less poor households. Following the introduction of a UCT, we can therefore expect to observe children from very poor households (those with income below Y*) to continue working, unless the transfer is such to bring them above the income threshold in which case some of the children will begin to attend school without necessarily stop working. For children from less poor households (for those with Y* < Y₁ < Y**), we should observe an increase in school attendance and a reduction in work, with some children stopping work altogether if the transfer is such that Y₁ becomes greater than Y**.

If transfers were conditional on school attendance, the results simply discussed would be partially different. In the following text, we present a heuristic discussion of such differences. A CCT offers a transfer τ conditional on a minimum investment in education S*. For households with income Y₁ > Y*, a CCT will have qualitatively the same impact of a UCT, unless S* is greater than the optimal S the household would have chosen with a transfer τ. In this latter case, the household might not find optimal to accept the transfer. If the household has an income Y₁ < Y**, the effect will also depend on the amount of the transfer. If τ > w (1 – S*) + eS*, i.e., if the transfer covers both the opportunity and the direct cost of sending a child to school, the household will accept the offer, send the child to school, and reduce child labor, as U (Y₁ + w(1 – S*) + τ – eS*, g(S*)) > U (Y₁ + w), i.e., as the lifetime utility sending the children to school for the required time and accepting the transfer is higher with respect to the baseline one.

On the other hand, if τ < w(1 – S*) + eS*, the effect is ambiguous, as in this case U(Y₁ + w(1– S*) + τ – eS*, g(S*)) can be higher or lower than U(Y₁ + w).

Therefore, a CCT might reduce child labor and increase school attendance also for children belonging to households below subsistence if the transfer is large enough to cover direct and opportunity cost of education (but not necessarily large enough to move them above subsistence).

For a discussion of the impact of a partial CCT subsidy, see De Hoop et al. (2019).

Table 2 presents the characteristics of the data, including attrition rates

The number of households with completed interviews in both surveys is 2,151, of which 1,354 eligible and 797 non-eligible. The mismatch is due to both the loss of observations from baseline and the addition of new households at follow-up. The new observations are a consequence of changes in demographic structure of households. Some children moved out from the original households at baseline and the new households where they live at follow-up constitute new observations. The new households reported in Table 2 are not eligible for the program and, for sake of this study, not included in the final sample. Whereas, for split households still eligible for the program, the one with the higher probability of receiving the transfer is matched with the corresponding original household in the baseline (Pellerano et al., 2014). These households are grouped with all the other households not split.

We use sampling weights to make inference on the entire “study population,” i.e., the NISSA population. Moreover, to address the potential attrition bias, sampling weights are multiplied by the inverse of the probability to remain in the sample at follow-up. Following Pellerano et al. (2014), sampling weights adjusted for attrition bias are constructed as follows:wij=((Aimiaij)(Nijknijk)p), ${{w}_{ij}}=\left( \left( \frac{{{A}_{i}}}{{{m}_{i}}{{a}_{ij}}} \right)\left( \frac{{{N}_{ijk}}}{{{n}_{ijk}}} \right)p \right),$ where A_i is the total number of households in the sample frame of villages in the i-th ED, m_i is the number of villages sampled in the i-th ED, a_ij is the number of households interviewed in village j in the i-th ED, N_ijk is the total number of households of type k in village j in the i-th ED, n_ijk is the number of households interviewed of type k in village j in the i-th ED, and p is the inverse of the probability of households to remain in the sample at follow-up. Sampling weights adjusted for attrition bias are used throughout this study.

The main goal of this study is to estimate the impact of the CGP on children’s work and education. Our final sample, therefore, is constituted of all matched children aged 6–15 at baseline and 8–17 at follow-up. We end up with a sample constituted of 2,928 children and 1,603 households, of which 2,098 eligible children and 1,107 eligible households.

The outcome variables considered refer to children’s education and work. Work outcomes include both the intensive and the extensive margins for all economic activities (household business, farming/livestock, and paid activities outside the household) and for farming/ livestock activities only, the sector where most children are employed. The household questionnaire provides information on the participation into economic activities during the last 12 months prior the interview and on the number of hours and days worked during the last 7 days prior the interview. For education, we focus on school enrolment and on time devoted to study outside the school in a typical school day. Finally, we look at school expenditures for each child since the beginning of the academic year and to their disaggregation in school fees and expenditures for uniforms and shoes.

Table 3 presents the descriptive statistics and balance tests at baseline for the outcome variables (Panel A) and for the covariates (Panel B) used in the estimates. There are not systematic differences between treatment and control groups, as we do not reject the null hypothesis of the t-test on the equality of means at 1% significance level. Treated and control households significantly differ only in one demographic characteristic: treated households have a somehow larger household size than control households.

Overall, 30% of eligible children worked in 12 months preceding the survey. Almost the totality of working children was employed in farming and livestock activities within the household (98%).

Given the small sample size of children working in the household business and in paid activities outside the household, extensive and intensive margins of labor in these two groups are not reported in Table 3.

Given the satisfactory results from the balance analysis presented in Table 3, we exploit the randomized treatment assignment to estimate the impact of the CGP on children’s labor and education by comparing outcome variables of treatment and control children at follow-up. We estimate the intention-to-treat (ITT), by considering all eligible households in treated and control EDs in the sample. The compliance rate is high, with 94% of eligible households in treated EDs actually receiving the transfer and very little spillover.

A negligible number of eligible households in control EDs (1%) and non-eligible households in treated EDs (5%) received the treatment; no non-eligible households in control EDs have managed to receive the transfer.

In particular, we estimate with Ordinary Least Squares (OLS) the following equation:

where Y_iv indicates the outcome for child i in ED v at follow-up. We include a set of relevant observable characteristics at child and household level at baseline, X_iv, to increase precision of the estimates. The impact of the transfer is given by the coefficient β₂, relative to the dummy variable T_v equal to 1 for treatment EDs and equal to 0 for control EDs.

The set of control variables includes sex and age of the child; a dummy equal to 1 if child is orphan (0 otherwise); household size; a dummy equal to 1 if children aged 0–5 are present in the household (0 otherwise); gender and age of the household head; the highest level of education reached by any adult member of the household; a dummy equal to 1 if the household has been affected by a serious economic shock during the 12 months prior the interview (0 otherwise); fixed effects for the 10 CCs. Robust and clustered standard errors at ED level and sampling weights adjusted for attrition bias are used throughout the analysis.

To analyze the heterogeneity of the impact of the transfer according to the level of resources of the household (as well as to other characteristics), we estimate the following OLS regression:

where C_iv is the characteristic of interest. In particular, we interact the treatment variable with four dummy variables for each quartiles of monthly expenditure per capita (at baseline) and with demographic characteristics of children to assess heterogeneity by age and sex.

Estimation results of Eqs (4) and (5) are shown in Section 6.1.

As robustness check for heterogeneity by level of expenditure per capita at baseline, we estimate Eq. (5) interacting the treatment variable with dummy variables of alternative measures of deprivation, such as the quartiles of the asset index and the mean and the median of expenditures at baseline. The results of the robustness checks are discussed in Section 6.2 and presented in Appendix.

We also estimate the probability to remain in the sample for beneficiary and non-beneficiary households, including the quartiles of expenditure per capita as covariates, to reassure that the estimation results are not driven by sample attrition. As shown in Table A.1 in Appendix, the coefficients relative to the quartiles are not statistically significant, indicating that the attrition is not driven by the economic condition of the households.

Finally, we test for the existence of spillover effects comparing non-eligible children in treatment and control EDs. The estimation results are presented in Section 6.3.

The possible nonlinearity of the income elasticity of child labor is an essential element for understanding its causes and for designing effective policies, especially as one of the main interventions advocated and implemented to address it relies on various forms of income transfer.

Causal evidence of the (nonlinear) impact of income on child labor is scarce, far from unambiguous and (with the exceptions discussed in the introduction) mainly based on experimental evidence deriving from CCT implemented in middle-income countries.

There is no experimental evidence based on UCT programs that mimic better than CCTs, a pure income effect and, therefore, allow to test for the nonlinearity of the impact of income changes on child labor.

As we have seen, theory indicates that below a critical (“subsistence”) income level, changes in income do not affect child labor and school attendance. Only above this critical level, changes in income affect (at a decreasing rate) household decisions concerning child labor.

We have analyzed the possible nonlinear response to income changes by evaluating the impact of the CGP (Phase 1—Round 2) on two specific dimensions of child well-being: children’s work and education. The CGP is a UCT randomly assigned to poor households of Lesotho providing a regular money transfer every quarter.

Looking at the aggregate effects on all the beneficiaries, we find that the CGP generated an increase of the enrolment rates by about 4%, of the time spent on studying by 13% and of the expenditures on uniforms and shoes. No significant effect on child labor was identified.

However, we find substantial heterogeneous treatment effects by household income. Significant reduction in both extensive and intensive margins of children’s work and increase in enrolment rates and expenditure on school fees can be identified only for children belonging to relatively less poor households. The poorest households apparently used the transfer only to increase expenditures on school uniforms and shoes, without changing children’s time allocation.

These findings are consistent with the theoretical framework developed and appear to support the hypothesis that the effectiveness of CT can increase, at least within a given range, with the level of income of the beneficiary households, thus offering solid evidence toward the hypothesis of nonlinearity of the income effect on child labor.

Our analysis indicates that, at least in low-income countries, UCT might not affect the decisions of the extremely poor in terms of school attendance and child labor. From a policy point of view, our results stress the importance of insuring that the amount of the transfer is sufficient to bring the household above “subsistence” to affect the decisions of the extreme poor. In contrast to CCT, for UCT is not sufficient that the transfer covers the (direct and opportunity) cost of taking children out of work and into school,

Remember this is a sufficient, but not necessary condition.