- Email: [email protected]
Not so fast! Cash transfers can increase child labor: Evidence for Bolivia
Economics Letters 179 (2019) 57–61
Contents lists available at ScienceDirect
Economics Letters journal homepage: www.elsevier.com/locate/ecolet
Not so fast! Cash transfers can increase child labor: Evidence for Bolivia✩ ∗
Alberto Chong a,b , , Monica Yáñez-Pagans c a
Georgia State University, United States Universidad del Pacífico, Peru c World Bank, United States b
highlights • • • •
We study how unconditional cash transfers can causally affect child labor. We exploit pension changes discontinuity at the eligibility age. We use eligibility as an instrumental variable for treatment on the treated. Probability that boys in rural areas will engage in child labor increase.
article
info
Article history: Received 12 December 2017 Received in revised form 13 March 2019 Accepted 20 March 2019 Available online 25 March 2019 JEL classification: O10 D13 I22 H55
a b s t r a c t Using data for Bolivia we study how a national-level unconditional cash transfer programs can causally affect child labor. We estimate intent-to-treat effects under a fuzzy regression discontinuity approach by taking advantage of the fact that the probability of receiving the pension changes discontinuously at the eligibility age. We also estimate average treatment effects on the treated by using eligibility as an instrumental variable for receipt. We find substantial increases in the probability that boys in rural areas engage in child labor. © 2019 Published by Elsevier B.V.
Keywords: Unconditional cash transfers Child labor Old-age pension Bolivia
1. Introduction According to the conventional wisdom cash transfers can help reduce child labor. This is further highlighted in a recent review of the academic literature that argues that: ‘‘there is no evidence that cash transfer interventions increase child labor in practice, but instead they lower child labor’s extensive and intensive margin’’ (De Hoop and Rosati, 2014). A straightforward policy implication from this conclusion is that cash transfers are a safe policy to improve child welfare. ✩ We are grateful to Gustavo Canavire-Bacarreza, Fernando Rios-Avila and seminar participants at Georgia State University, Universidad EAFIT and the World Bank for comments and suggestions. We are thankful to Angelo Cozzubo for excellent research assistance. The standard disclaimer applies. ∗ Corresponding author at: Georgia State University, United States. E-mail addresses: [email protected] (A. Chong), [email protected] (M. Yáñez-Pagans). https://doi.org/10.1016/j.econlet.2019.03.021 0165-1765/© 2019 Published by Elsevier B.V.
However, in this letter we provide strong evidence that cash transfers can increase child labor when families are unconstrained in the use of the transfers received. We exploit the introduction of a national old-age pension program in Bolivia called Bolivida, which does not target specific behavioral changes. It is a flat unconditional cash transfer program paid to all Bolivians aged 65 and older independently of income, contributions to the system, or living arrangements so that the each eligible elder in the household gets an independent payment.1 The cash transfer amounts around US 250 dollars per year and represents between 50 and 85 percent of the total annual income of the poor and extremely poor households, respectively (Von Gersdorff, 1997). 1 Since its inception, in 1997 the program has gone through name changes as well as reformulations in both the size of the benefit and the timing of the payments. We refer to it as Bolivida, as this was the name used for the program during our time of analysis.
58
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
age trends we include a third order polynomial expansion in the age of the oldest person in the household.2 In order to minimize endogenous pension take-up we focus on age eligibility rather than on actual take-up. Thus we estimate average intention-totreat (ITT) effects on time allocation decisions for child i, living in household h, as follows: yih = α + ϑ (Eligibleh ∗ Ruralh ) + θ (Eligibleh ∗ Urbanh )
+ β Ruralh + f (Xh ) + ξih
(1)
where yih is an indicator for whether the child is involved in paid or unpaid market or agricultural work; Eligibleh reflects the presence of an eligible member in the household; Ruralh is rural household; Urbanh is urban household; f (Xh ) is a linear function which includes a third order polynomial expansion in the age of the oldest person in the household; and ξih is an error term.3 All estimates are weighted to be nationally representative and corrected for clustered and stratified sampling design. We also exploit the instrumental variables connection to the fuzzy regression discontinuity design and parametrically estimate average treatment effects on beneficiaries using eligibility as an instrument for actual program receipt. This allows us to estimate the treatment-on-the treated (TOT) as non-compliance in this case is just one-sided. That is, eligible people may not have benefited from the program while chances are that non-eligible people did not benefit from it, as rules to access the benefit were stringent and not easy to manipulate.4 Therefore, the fuzzy design is actually identifying the treatment effect on the subpopulation of compliers or TOT, which is then estimated using a two-step approach. First, we estimate a non-linear model reduced-form for the probability of receiving the Bolivida as a function of eligibility. Second, we use the predicted values of this probability of receiving the Bolivida as instruments for program receipt as follows: Fig. 1. Children ages considered are 7 to 17 years of age. We do not find evidence of discontinuity neither in the case of boys in urban areas nor in the case of girls in both urban and rural areas.
In order to evaluate the impact of this national program on child labor we estimate intent-to-treat effects under a fuzzy regression discontinuity approach by exploiting the discontinuity at the eligibility age. We also measure average treatment effects on the treated by using eligibility as an instrumental variable for receipt. The next section describes the data and methodology. Section 3 presents findings. Section 4 presents robustness tests and Section 5 concludes. 2. Data and methodology The data come from a national household survey conducted by the Bolivian National Institute of Statistics in 2001. We focus on school-age children (7 – 17 years old) who live with an elder between the ages of 60 and 69. We exclude all households that do not have at least one member in the agricultural or labor markets and those whose total reported income is missing. Non-relatives and domestic workers living in the household are also excluded. Our final sample comprises 1172 school-age children and 816 elders distributed across 756 households. Table 1 presents the summary statistics. We employ a fuzzy regression discontinuity design and identify the impact of the old-age pension on child labor by comparing children living in households with an eligible elder to those living in households with a near-eligible one. In order to control for
(
)
(
ˆ ˆ yih = π + φ Boli v idah ∗ Ruralh + η Boli v idah ∗ Urbanh
+ δ Ruralh + g(Xh ) + νih
) (2)
ˆ where Boli v ida is the predicted probability of receiving the program; g(Xh ) is a linear function which includes a third order polynomial of the age of the oldest person in the household; and νih is an error term and data are clustered at the age of the oldest person in the household. As above, all estimates are weighted to be nationally representative and corrected for clustered and stratified sampling design. 3. Findings In Fig. 1 we provide simple graphical evidence that shows discontinuity at the cutoff point when testing the full sample for boys aged 7 to 17 as well in the case of boys in rural areas. We find no graphical discontinuity neither in the case of boys in urban areas nor in the case of girls in either urban or rural areas. This graphical evidence is further supported by more 2 We employ a third-order polynomial because with more fine data cuts higher order terms drop out. Second order polynomial yield similar results. 3 We tested specifications that allow for f(X) as different functions of age on either side of the cutoff point and we find analogous results. In general, when relaxing f(x), as the impact of eligibility vary with rural/urban status, we also find similar results. 4 In a fuzzy regression discontinuity design with two-sided non-compliance, the instrumental variables estimator gives the LATE. With one-sided non-compliance, the instrumental variables estimator gives TOT instead.
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
59
Table 1 Summary statistics.
Children characteristics Age (in years) Share of boys Share working Number of hours worked per week Observations
Eligible
Non eligible
Beneficiary
Non beneficiary
(1)
(2)
(3)
(4)
12.34 (0.17) 0.5 (0.03) 0.38 (0.03) 24.47 (1.95) 575
12.78 (0.17) 0.56 (0.03) 0.47 (0.04) 28.2 (2.01) 597
12.47 (0.26) 0.53 (0.04) 0.4 (0.05) 26.62 (2.84) 309
12.59 (0.14) 0.53 (0.02) 0.43 (0.03) 26.54 (1.65) 863
6.24 (0.21) 2.47 (0.12) 38.31 (1.21) 1995 (178.35) 39.5 (3.41)
6.27 (0.30) 2.58 (0.16) 39.94 (1.16) 1785 (162.05) 0.00* (0.00)
6.26 (0.27) 2.22 (0.13) 36.3 (1.29) 2100 (259.01) 74.03 (3.13)
6.28 (0.24) 2.63* (0.12) 40.15* (1.03) 1812 (135.58) 0.00* (0.00)
331
326
187
470
Household characteristics Household size Number of school age children in HH Share of school age children in HH HH income excl. Bolivida (monthly, Bs) Bolivida HH income (monthly, Bs) Observations
All estimates weighted to be nationally representative and corrected for clustered and stratified sampling design. Linearized standard errors in parenthesis. Asterisks in column (2) correspond to p < 0.05 of the difference between children in households with an eligible elderly person and those in households with a near-eligible elderly person. Asterisks in column (4) correspond to a p < 0.05 of the difference between beneficiary and non-beneficiary children/households. The average number of people per household receiving the pension is 1.36. Per capita household income adjusted for equivalence scales is defined by the World Bank: Adult Equivalence Scales (AES) = 1 + 0.7(adults − 1) + 0.5 children.
Table 2 Intention-to-treat. Dependent variable: child labor. Old-Age pension*rural HH Uncorrected p-value Stepdown p-value Old-Age pension *urban HH Uncorrected p-value Stepdown p-value Rural HH Uncorrected p-value Stepdown p-value
Boys
Girls
0.0767 (0.089) [0.094] −0.0048 (0.898) [0.922] 0.5268 (0.010) [0.020]
0.0325 (0.305) [0.319] −0.0389 (0.623) [0.761] 0.4579 (0.010) [0.020]
Stepdown p-values in brackets are based on Romano and Wolf (2005, 2016). Regressions based on empirical specification (1) in text. Numbers in bold correspond to treatment that is statistically significant at conventional levels. Step-down calculations include all outcomes for boys and girls that is, six hypotheses in total.
Table 3 Treatment-on-the-treated. Dependent variable: child labor. Old-Age pension*rural HH Uncorrected p-value Stepdown p-value Old-Age pension *urban HH Uncorrected p-value Stepdown p-value Rural HH Uncorrected p-value Stepdown p-value
Boys
Girls
0.1249 (0.035) [0.046] 0.0380 (0.192) [0.211] 0.5403 (0.010) [0.023]
0.0481 (0.132) [0.158] −0.0478 (0.341) [0.378] 0.4695 (0.011) [0.024]
Stepdown p-values in brackets are based on Romano and Wolf (2005, 2016). Regressions based on empirical specification (2) in text. Numbers in bold correspond to treatment that is statistically significant at conventional levels. Step-down calculations include all outcomes for boys and girls that is, six hypotheses in total.
formal estimates.5 In particular, we provide results using stepdown adjusted p-values to correct for multiple hypothesis testing (Romano and Wolf, 2005, 2016). The effect of a treatment may be heterogeneous in that it may vary across subgroups defined by observed characteristics and it would be desirable to determine for which of these subgroups a treatment may have an effect. In our case we have an outcome variables of interest, child labor and two sub-groups for the same treatment, boys and girls. Table 2 shows our findings for the intention-to-treat case. We find that unconditional cash transfer to elders is associated with a large and significant increase in the probability that boys in rural areas are involved in child labor. Boys in rural households who live with an eligible elder are 7.8% points more likely to be involved in working activities as compared to those who live with a near-eligible person. We do not find any statistically significant result in the case of boys living in urban areas or in the case of girls living in either rural or urban areas. In Table 3 we repeat the same exercise as above, but we present our treatment-on-the-treated results as described in Eq. (2). We find analogous results to the intention-to-treat case. Boys in rural households who live with an eligible elder are 12.5% points more likely to be involved in working activities as compared to those who live with a near-eligible person. As in the case of intention-to-treat, our findings are not statistically significant for boys living in urban areas or for girls living in either rural or urban areas.6
5 We also test whether other exogenous characteristics change at the discontinuity using both visual and statistical tests. We find that household size (t = 0.24), average children age (t = 0.29), number of children in the household (t = 0.16) and average household income (t = 0.19) are not statistically significant. We find analogous results when focusing on boys or girls as well as in rural or urban areas. 6 First stages are shown in the Appendix.
60
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61 Table 4 Eligibility, receipt, and household composition.
Panel A: boys Bolivida rural HH Bolivida urban HH Observations Panel B: Girls Bolivida rural HH Bolivida urban HH Observations
Pr. HH member 60 69 living elsewhere in last 5 years
Log HH size
ITT (1)
TOT (2)
ITT (3)
TOT (4)
0.0130 (0.025) −0.0210 (0.058) 269
0.0098 (0.043) −0.0108 (0.006) 269
−0.0435
−0.0226 (0.042) −0.0223 (0.022) 326
0.0081 (0.026) −0.0156 (0.019) 295
0.0061 (0.059) −0.0218 (0.029) 295
(0.050) −0.0299 (0.060) 326
−0.0019 (0.023) −0.0340 (0.025) 282
−0.0319 (0.020) −0.0785 (0.050) 282
All coefficients are OLS estimates. Standard errors in parenthesis; data are clustered at the age of the oldest person in the household. All estimates weighted to be nationally representative and corrected for clustered and stratified sampling design.
Table 5 ITT effects: children’s time allocation and old-age pension.
to the outcome variables of interest, results that further supports our findings.8
Child labor Boys (1) Eligible HH
−0.0140 (0.009)
Eligible rural HH Urban HH Rural HH Observations
Girls (2)
0.4062*** (0.069) 402
(3)
(4)
−0.0334 (0.089) 0.0187 (0.150) −0.0596 (0.075) 0.4268*** (0.056) 402
0.4529*** (0.027) 371
−0.033 (0.094) −0.0964 (0.090) 0.5036*** (0.034) 371
Baseline sample: 1999 and 2000. Intention-to-treat effects estimated from Eq. (1) using OLS. Standard errors in parenthesis; data are clustered at the age of the oldest person in the household. All estimates weighted to be nationally representative. All regressions also include a third order polynomial in the age of the oldest person in the household, and year fixed-effects. ***Statistically significant at one percent.
4. Robustness Households may incur in endogenous consumption by anticipating the income and adjusting member composition for instance, by welcoming elder relatives. Table 4 shows that there is no statistically significant link between elder’s migration in the last five years, household size, and old age pension status. The assumption that observed changes in household composition across Bolivida status are random seems like a reasonable assumption.7 Also, there may be other confounding factors at play. One way to test for this is to apply a falsification test to estimate the effect of the presence of an eligible person in the household in a baseline sample on our outcomes of interest. Table 5 reports our out-of sample validation for the intent-to-treat coefficients using two household surveys collected in 1999 and 2000, which is prior to our analysis and when the program was suspended and was not paying any benefits. We find no statistically significant link
7 During the time of analysis Bolivida had just resumed after several years on hold. It is unlikely that families welcome elders as a result of anticipating income.
5. Conclusions
We find that Bolivida, a national-level unconditional cash transfer program, leads to increases in the probability that boys engage in labor in rural areas. This is consistent with previous research (Martinez, 2004) that shows that rural households tend to invest the proceeds of unconditional transfers, which then translates in increases in output. In this context, Bolivida may trigger demand for labor in rural settings where labor markets are missing, where hiring labor cannot be easily substituted for family labor due to moral hazard and where returns to experience from child labor are high. The unconditional transfer may create incentives for child labor participation in rural settings by increasing the boy’s relative productivity in the labor market and therefore by increasing the opportunity cost of leisure and schooling for boys in rural areas. This is consistent with some recent evidence that shows a slightly upward trend in child labor in rural areas in the country during our period of study. Furthermore, given the adherence to traditional gender roles in Bolivia where males are the key family providers, boys in rural areas are more likely to work outside the home while girls in rural areas are more likely to be engaged in domestic work. Our definition of child labor does not include domestic work though, which is consistent with the fact that we find no impact of the unconditional transfer on girls.
Appendix
First Stages. See Table A.6.
8 We find analogous results in the case of treatment-on-the-treated, but for the sake of economy we do not include those additional findings here. They are available upon request.
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
61
Table A.6 Child labor Girls
Linear prediction
(2) Transfer* Rural
(3) Transfer
(4) Transfer* Rural
0.044*** (0.011)
−0.033*** (0.007) 0.106*** (0.013) 18.720*** (7.007) 0.311 0.308 8.0123 2392
0.046*** (0.014) −0.003 (0.017) 13.860 (12.039) 0.336 0.333 19.2613 2392
0.039*** (0.011)
Linear prediction*Rural Constant R-squared Adjusted R-squared F-statistic Observations
Boys
(1) Transfer
13.890 (12.046) 0.336 0.333 21.0268 2392
23.766** (11.733) 0.322 0.319 18.4606 2212
Clustered standard errors are shown in parenthesis. Regressions include year fixed effects and the following controls: age of household members, number of members in the household, education, experience, and gender. *Statistically significant at ten percent. **Statistically significant at five percent. ***Statistically significant at one percent.
References De Hoop, J., Rosati, F., 2014. World Bank Res. Obs. 29 (2), 202–223. Martinez, S., 2004. Pensions, Poverty and Household Investments in Bolivia Manuscript. University of California, Berkeley. Romano, J.P., Wolf, M., 2005. Stepwise multiple testing as formalized data snooping. Econometrica 73 (4), 1237–1282.
Romano, J.P., Wolf, M., 2016. Efficient computation of adjusted p-values for resampling-based stepdown multiple testing. Statist. Probab. Lett. 113 (C), 38–40. Von Gersdorff, H., 1997. The Bolivian pension reform: Innovative solutions to common problems, In: World Bank Working Paper, vol. 1832, Washington, DC.
Contents lists available at ScienceDirect
Economics Letters journal homepage: www.elsevier.com/locate/ecolet
Not so fast! Cash transfers can increase child labor: Evidence for Bolivia✩ ∗
Alberto Chong a,b , , Monica Yáñez-Pagans c a
Georgia State University, United States Universidad del Pacífico, Peru c World Bank, United States b
highlights • • • •
We study how unconditional cash transfers can causally affect child labor. We exploit pension changes discontinuity at the eligibility age. We use eligibility as an instrumental variable for treatment on the treated. Probability that boys in rural areas will engage in child labor increase.
article
info
Article history: Received 12 December 2017 Received in revised form 13 March 2019 Accepted 20 March 2019 Available online 25 March 2019 JEL classification: O10 D13 I22 H55
a b s t r a c t Using data for Bolivia we study how a national-level unconditional cash transfer programs can causally affect child labor. We estimate intent-to-treat effects under a fuzzy regression discontinuity approach by taking advantage of the fact that the probability of receiving the pension changes discontinuously at the eligibility age. We also estimate average treatment effects on the treated by using eligibility as an instrumental variable for receipt. We find substantial increases in the probability that boys in rural areas engage in child labor. © 2019 Published by Elsevier B.V.
Keywords: Unconditional cash transfers Child labor Old-age pension Bolivia
1. Introduction According to the conventional wisdom cash transfers can help reduce child labor. This is further highlighted in a recent review of the academic literature that argues that: ‘‘there is no evidence that cash transfer interventions increase child labor in practice, but instead they lower child labor’s extensive and intensive margin’’ (De Hoop and Rosati, 2014). A straightforward policy implication from this conclusion is that cash transfers are a safe policy to improve child welfare. ✩ We are grateful to Gustavo Canavire-Bacarreza, Fernando Rios-Avila and seminar participants at Georgia State University, Universidad EAFIT and the World Bank for comments and suggestions. We are thankful to Angelo Cozzubo for excellent research assistance. The standard disclaimer applies. ∗ Corresponding author at: Georgia State University, United States. E-mail addresses: [email protected] (A. Chong), [email protected] (M. Yáñez-Pagans). https://doi.org/10.1016/j.econlet.2019.03.021 0165-1765/© 2019 Published by Elsevier B.V.
However, in this letter we provide strong evidence that cash transfers can increase child labor when families are unconstrained in the use of the transfers received. We exploit the introduction of a national old-age pension program in Bolivia called Bolivida, which does not target specific behavioral changes. It is a flat unconditional cash transfer program paid to all Bolivians aged 65 and older independently of income, contributions to the system, or living arrangements so that the each eligible elder in the household gets an independent payment.1 The cash transfer amounts around US 250 dollars per year and represents between 50 and 85 percent of the total annual income of the poor and extremely poor households, respectively (Von Gersdorff, 1997). 1 Since its inception, in 1997 the program has gone through name changes as well as reformulations in both the size of the benefit and the timing of the payments. We refer to it as Bolivida, as this was the name used for the program during our time of analysis.
58
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
age trends we include a third order polynomial expansion in the age of the oldest person in the household.2 In order to minimize endogenous pension take-up we focus on age eligibility rather than on actual take-up. Thus we estimate average intention-totreat (ITT) effects on time allocation decisions for child i, living in household h, as follows: yih = α + ϑ (Eligibleh ∗ Ruralh ) + θ (Eligibleh ∗ Urbanh )
+ β Ruralh + f (Xh ) + ξih
(1)
where yih is an indicator for whether the child is involved in paid or unpaid market or agricultural work; Eligibleh reflects the presence of an eligible member in the household; Ruralh is rural household; Urbanh is urban household; f (Xh ) is a linear function which includes a third order polynomial expansion in the age of the oldest person in the household; and ξih is an error term.3 All estimates are weighted to be nationally representative and corrected for clustered and stratified sampling design. We also exploit the instrumental variables connection to the fuzzy regression discontinuity design and parametrically estimate average treatment effects on beneficiaries using eligibility as an instrument for actual program receipt. This allows us to estimate the treatment-on-the treated (TOT) as non-compliance in this case is just one-sided. That is, eligible people may not have benefited from the program while chances are that non-eligible people did not benefit from it, as rules to access the benefit were stringent and not easy to manipulate.4 Therefore, the fuzzy design is actually identifying the treatment effect on the subpopulation of compliers or TOT, which is then estimated using a two-step approach. First, we estimate a non-linear model reduced-form for the probability of receiving the Bolivida as a function of eligibility. Second, we use the predicted values of this probability of receiving the Bolivida as instruments for program receipt as follows: Fig. 1. Children ages considered are 7 to 17 years of age. We do not find evidence of discontinuity neither in the case of boys in urban areas nor in the case of girls in both urban and rural areas.
In order to evaluate the impact of this national program on child labor we estimate intent-to-treat effects under a fuzzy regression discontinuity approach by exploiting the discontinuity at the eligibility age. We also measure average treatment effects on the treated by using eligibility as an instrumental variable for receipt. The next section describes the data and methodology. Section 3 presents findings. Section 4 presents robustness tests and Section 5 concludes. 2. Data and methodology The data come from a national household survey conducted by the Bolivian National Institute of Statistics in 2001. We focus on school-age children (7 – 17 years old) who live with an elder between the ages of 60 and 69. We exclude all households that do not have at least one member in the agricultural or labor markets and those whose total reported income is missing. Non-relatives and domestic workers living in the household are also excluded. Our final sample comprises 1172 school-age children and 816 elders distributed across 756 households. Table 1 presents the summary statistics. We employ a fuzzy regression discontinuity design and identify the impact of the old-age pension on child labor by comparing children living in households with an eligible elder to those living in households with a near-eligible one. In order to control for
(
)
(
ˆ ˆ yih = π + φ Boli v idah ∗ Ruralh + η Boli v idah ∗ Urbanh
+ δ Ruralh + g(Xh ) + νih
) (2)
ˆ where Boli v ida is the predicted probability of receiving the program; g(Xh ) is a linear function which includes a third order polynomial of the age of the oldest person in the household; and νih is an error term and data are clustered at the age of the oldest person in the household. As above, all estimates are weighted to be nationally representative and corrected for clustered and stratified sampling design. 3. Findings In Fig. 1 we provide simple graphical evidence that shows discontinuity at the cutoff point when testing the full sample for boys aged 7 to 17 as well in the case of boys in rural areas. We find no graphical discontinuity neither in the case of boys in urban areas nor in the case of girls in either urban or rural areas. This graphical evidence is further supported by more 2 We employ a third-order polynomial because with more fine data cuts higher order terms drop out. Second order polynomial yield similar results. 3 We tested specifications that allow for f(X) as different functions of age on either side of the cutoff point and we find analogous results. In general, when relaxing f(x), as the impact of eligibility vary with rural/urban status, we also find similar results. 4 In a fuzzy regression discontinuity design with two-sided non-compliance, the instrumental variables estimator gives the LATE. With one-sided non-compliance, the instrumental variables estimator gives TOT instead.
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
59
Table 1 Summary statistics.
Children characteristics Age (in years) Share of boys Share working Number of hours worked per week Observations
Eligible
Non eligible
Beneficiary
Non beneficiary
(1)
(2)
(3)
(4)
12.34 (0.17) 0.5 (0.03) 0.38 (0.03) 24.47 (1.95) 575
12.78 (0.17) 0.56 (0.03) 0.47 (0.04) 28.2 (2.01) 597
12.47 (0.26) 0.53 (0.04) 0.4 (0.05) 26.62 (2.84) 309
12.59 (0.14) 0.53 (0.02) 0.43 (0.03) 26.54 (1.65) 863
6.24 (0.21) 2.47 (0.12) 38.31 (1.21) 1995 (178.35) 39.5 (3.41)
6.27 (0.30) 2.58 (0.16) 39.94 (1.16) 1785 (162.05) 0.00* (0.00)
6.26 (0.27) 2.22 (0.13) 36.3 (1.29) 2100 (259.01) 74.03 (3.13)
6.28 (0.24) 2.63* (0.12) 40.15* (1.03) 1812 (135.58) 0.00* (0.00)
331
326
187
470
Household characteristics Household size Number of school age children in HH Share of school age children in HH HH income excl. Bolivida (monthly, Bs) Bolivida HH income (monthly, Bs) Observations
All estimates weighted to be nationally representative and corrected for clustered and stratified sampling design. Linearized standard errors in parenthesis. Asterisks in column (2) correspond to p < 0.05 of the difference between children in households with an eligible elderly person and those in households with a near-eligible elderly person. Asterisks in column (4) correspond to a p < 0.05 of the difference between beneficiary and non-beneficiary children/households. The average number of people per household receiving the pension is 1.36. Per capita household income adjusted for equivalence scales is defined by the World Bank: Adult Equivalence Scales (AES) = 1 + 0.7(adults − 1) + 0.5 children.
Table 2 Intention-to-treat. Dependent variable: child labor. Old-Age pension*rural HH Uncorrected p-value Stepdown p-value Old-Age pension *urban HH Uncorrected p-value Stepdown p-value Rural HH Uncorrected p-value Stepdown p-value
Boys
Girls
0.0767 (0.089) [0.094] −0.0048 (0.898) [0.922] 0.5268 (0.010) [0.020]
0.0325 (0.305) [0.319] −0.0389 (0.623) [0.761] 0.4579 (0.010) [0.020]
Stepdown p-values in brackets are based on Romano and Wolf (2005, 2016). Regressions based on empirical specification (1) in text. Numbers in bold correspond to treatment that is statistically significant at conventional levels. Step-down calculations include all outcomes for boys and girls that is, six hypotheses in total.
Table 3 Treatment-on-the-treated. Dependent variable: child labor. Old-Age pension*rural HH Uncorrected p-value Stepdown p-value Old-Age pension *urban HH Uncorrected p-value Stepdown p-value Rural HH Uncorrected p-value Stepdown p-value
Boys
Girls
0.1249 (0.035) [0.046] 0.0380 (0.192) [0.211] 0.5403 (0.010) [0.023]
0.0481 (0.132) [0.158] −0.0478 (0.341) [0.378] 0.4695 (0.011) [0.024]
Stepdown p-values in brackets are based on Romano and Wolf (2005, 2016). Regressions based on empirical specification (2) in text. Numbers in bold correspond to treatment that is statistically significant at conventional levels. Step-down calculations include all outcomes for boys and girls that is, six hypotheses in total.
formal estimates.5 In particular, we provide results using stepdown adjusted p-values to correct for multiple hypothesis testing (Romano and Wolf, 2005, 2016). The effect of a treatment may be heterogeneous in that it may vary across subgroups defined by observed characteristics and it would be desirable to determine for which of these subgroups a treatment may have an effect. In our case we have an outcome variables of interest, child labor and two sub-groups for the same treatment, boys and girls. Table 2 shows our findings for the intention-to-treat case. We find that unconditional cash transfer to elders is associated with a large and significant increase in the probability that boys in rural areas are involved in child labor. Boys in rural households who live with an eligible elder are 7.8% points more likely to be involved in working activities as compared to those who live with a near-eligible person. We do not find any statistically significant result in the case of boys living in urban areas or in the case of girls living in either rural or urban areas. In Table 3 we repeat the same exercise as above, but we present our treatment-on-the-treated results as described in Eq. (2). We find analogous results to the intention-to-treat case. Boys in rural households who live with an eligible elder are 12.5% points more likely to be involved in working activities as compared to those who live with a near-eligible person. As in the case of intention-to-treat, our findings are not statistically significant for boys living in urban areas or for girls living in either rural or urban areas.6
5 We also test whether other exogenous characteristics change at the discontinuity using both visual and statistical tests. We find that household size (t = 0.24), average children age (t = 0.29), number of children in the household (t = 0.16) and average household income (t = 0.19) are not statistically significant. We find analogous results when focusing on boys or girls as well as in rural or urban areas. 6 First stages are shown in the Appendix.
60
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61 Table 4 Eligibility, receipt, and household composition.
Panel A: boys Bolivida rural HH Bolivida urban HH Observations Panel B: Girls Bolivida rural HH Bolivida urban HH Observations
Pr. HH member 60 69 living elsewhere in last 5 years
Log HH size
ITT (1)
TOT (2)
ITT (3)
TOT (4)
0.0130 (0.025) −0.0210 (0.058) 269
0.0098 (0.043) −0.0108 (0.006) 269
−0.0435
−0.0226 (0.042) −0.0223 (0.022) 326
0.0081 (0.026) −0.0156 (0.019) 295
0.0061 (0.059) −0.0218 (0.029) 295
(0.050) −0.0299 (0.060) 326
−0.0019 (0.023) −0.0340 (0.025) 282
−0.0319 (0.020) −0.0785 (0.050) 282
All coefficients are OLS estimates. Standard errors in parenthesis; data are clustered at the age of the oldest person in the household. All estimates weighted to be nationally representative and corrected for clustered and stratified sampling design.
Table 5 ITT effects: children’s time allocation and old-age pension.
to the outcome variables of interest, results that further supports our findings.8
Child labor Boys (1) Eligible HH
−0.0140 (0.009)
Eligible rural HH Urban HH Rural HH Observations
Girls (2)
0.4062*** (0.069) 402
(3)
(4)
−0.0334 (0.089) 0.0187 (0.150) −0.0596 (0.075) 0.4268*** (0.056) 402
0.4529*** (0.027) 371
−0.033 (0.094) −0.0964 (0.090) 0.5036*** (0.034) 371
Baseline sample: 1999 and 2000. Intention-to-treat effects estimated from Eq. (1) using OLS. Standard errors in parenthesis; data are clustered at the age of the oldest person in the household. All estimates weighted to be nationally representative. All regressions also include a third order polynomial in the age of the oldest person in the household, and year fixed-effects. ***Statistically significant at one percent.
4. Robustness Households may incur in endogenous consumption by anticipating the income and adjusting member composition for instance, by welcoming elder relatives. Table 4 shows that there is no statistically significant link between elder’s migration in the last five years, household size, and old age pension status. The assumption that observed changes in household composition across Bolivida status are random seems like a reasonable assumption.7 Also, there may be other confounding factors at play. One way to test for this is to apply a falsification test to estimate the effect of the presence of an eligible person in the household in a baseline sample on our outcomes of interest. Table 5 reports our out-of sample validation for the intent-to-treat coefficients using two household surveys collected in 1999 and 2000, which is prior to our analysis and when the program was suspended and was not paying any benefits. We find no statistically significant link
7 During the time of analysis Bolivida had just resumed after several years on hold. It is unlikely that families welcome elders as a result of anticipating income.
5. Conclusions
We find that Bolivida, a national-level unconditional cash transfer program, leads to increases in the probability that boys engage in labor in rural areas. This is consistent with previous research (Martinez, 2004) that shows that rural households tend to invest the proceeds of unconditional transfers, which then translates in increases in output. In this context, Bolivida may trigger demand for labor in rural settings where labor markets are missing, where hiring labor cannot be easily substituted for family labor due to moral hazard and where returns to experience from child labor are high. The unconditional transfer may create incentives for child labor participation in rural settings by increasing the boy’s relative productivity in the labor market and therefore by increasing the opportunity cost of leisure and schooling for boys in rural areas. This is consistent with some recent evidence that shows a slightly upward trend in child labor in rural areas in the country during our period of study. Furthermore, given the adherence to traditional gender roles in Bolivia where males are the key family providers, boys in rural areas are more likely to work outside the home while girls in rural areas are more likely to be engaged in domestic work. Our definition of child labor does not include domestic work though, which is consistent with the fact that we find no impact of the unconditional transfer on girls.
Appendix
First Stages. See Table A.6.
8 We find analogous results in the case of treatment-on-the-treated, but for the sake of economy we do not include those additional findings here. They are available upon request.
A. Chong and M. Yáñez-Pagans / Economics Letters 179 (2019) 57–61
61
Table A.6 Child labor Girls
Linear prediction
(2) Transfer* Rural
(3) Transfer
(4) Transfer* Rural
0.044*** (0.011)
−0.033*** (0.007) 0.106*** (0.013) 18.720*** (7.007) 0.311 0.308 8.0123 2392
0.046*** (0.014) −0.003 (0.017) 13.860 (12.039) 0.336 0.333 19.2613 2392
0.039*** (0.011)
Linear prediction*Rural Constant R-squared Adjusted R-squared F-statistic Observations
Boys
(1) Transfer
13.890 (12.046) 0.336 0.333 21.0268 2392
23.766** (11.733) 0.322 0.319 18.4606 2212
Clustered standard errors are shown in parenthesis. Regressions include year fixed effects and the following controls: age of household members, number of members in the household, education, experience, and gender. *Statistically significant at ten percent. **Statistically significant at five percent. ***Statistically significant at one percent.
References De Hoop, J., Rosati, F., 2014. World Bank Res. Obs. 29 (2), 202–223. Martinez, S., 2004. Pensions, Poverty and Household Investments in Bolivia Manuscript. University of California, Berkeley. Romano, J.P., Wolf, M., 2005. Stepwise multiple testing as formalized data snooping. Econometrica 73 (4), 1237–1282.
Romano, J.P., Wolf, M., 2016. Efficient computation of adjusted p-values for resampling-based stepdown multiple testing. Statist. Probab. Lett. 113 (C), 38–40. Von Gersdorff, H., 1997. The Bolivian pension reform: Innovative solutions to common problems, In: World Bank Working Paper, vol. 1832, Washington, DC.