Does Health Insurance Reduce Child Labour and Education Gaps? Evidence from Rwanda

Abstract

A common practice of poor households to informally deal with risk is to allocate children’s time away from school towards income-generating activities or household production. Focussing on Rwanda, this study investigates whether the provision of formal health insurance helps to prevent this undesired risk coping strategy. We find that children of households enrolled in health insurance work significantly less compared to those of not enrolled families, and also have better educational achievements. The results suggest that policy interventions which reduce household risk exposure may have additional benefits in terms of lower child labour supply and higher schooling levels.

Acknowledgements

The author would like to thank Conny Wunsch and Stefan Felder for their valuable comments. Thanks also to the participants of the World Congress of the International Health Economics Association in Sydney, the Annual Conference of the German Society for Health Economics in Munich and the Swiss Health Economics Workshop in Lucerne. The data used in this survey is available from the National Institute of Statistics of Rwanda; the code is available from the author upon request.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. Fitzsimons (2007) distinguishes the effect of idiosyncratic household income risk and covariate village income risk. She finds that only village-level risk has an adverse effect on years of schooling, while household-level risk has no significant effect. The latter result may indicate that households are sufficiently self-insured by other informal strategies. It might, however, also be explained by a long-term ex ante strategy of households that use the education of children as a form of insurance (see Section 3.2 for an illustration of this mechanism in a theoretical model).

2. The three pilot districts were Kabutare, Byumba and Kabgayi. They do not correspond to current districts, as in the course of an administrative reorganisation in 2006, the former 106 Rwandan districts were replaced by the 30 districts that exist today. After the first year of the CBHI pilot, 88,303 individuals were enrolled (Schneider & Diop, 2001).

3. The gross enrolment ratio is the ratio of the number of students actually enrolled at a specific schooling level (for example primary or secondary), regardless of age, to the population of the age group that corresponds to this level of education (World Bank, 2013).

4. H denotes the Hessian matrix.

5. For this test, we run a first stage regression of exogenous variables and instruments on the endogenous variable. The estimated residuals of this regression are then added to the second-stage regression of the outcome of interest. If the predicted residuals have no predictive power, that is the estimated coefficient is not statistically significant, we prefer simple (count data) regression models over the instrumental variable approach.

6. To qualify as a valid instrument, a variable must satisfy two conditions. Firstly, it must be highly correlated with the endogenous variable (relevance condition). Secondly, it must not be correlated with the error term of the outcome equation, that is it should neither qualify as a regressor on its own nor be correlated with unobserved characteristics captured by the error term in the outcome equation (exogeneity condition).

7. The community enrolment rate is calculated with the following formula: Cluster enrolment rate = (Number of enrolled households in cluster – Dummy if household is enrolled)/(Total number of households in cluster – 1). This instrumental variable was used in two other impact evaluations of Rwanda mutual health insurance (Lu et al., 2012; Shimeles, 2010).

8. Wagstaff and Lindelow (2008) use a similar instrumental variable for health insurance in China.

9. Running the IV regressions shows that standard errors nearly quadruple. This suggests that IV does not provide precise estimations and count data regression is the more adequate approach.

10. In the empirical literature, dependent variables with many zero values are often estimated with Tobit models for censored data. However, the consistency of Tobit estimates is based on the assumptions of normality and homoscedasticity. Testing for these assumptions (see Cameron & Trivedi, 2010) shows that Tobit is not the appropriate method for our data.

11. In our sample, there is usually more than one child aged between 7 and 15 in a household, probably resulting in intra-household correlations and the need to estimate the standard errors clustering at household level. However, we choose clustering at community level, as clustering at higher levels (that is a smaller number of clusters) is more conservative (Cameron & Miller, 2015).

12. Besides the assumption of selection on observables, common support is the second important condition in the PSM approach. It states that for inferring causality, only non-treated units which are comparable to the treated units should be utilised.

13. In the radius matching approach, a maximal distance of propensity scores that is tolerated is predefined. All those non-treated individuals that lie within this threshold are assigned to a treated individual. We use the STATA command radiusmatch proposed by Huber, Lechner, and Steinmayr (2015).

14. The first test is a two-sided t-test for all regressors. After the matching process, none of the used covariates should be statistically different between treatment and control group. As second test, we look at the standardised bias (Rosenbaum & Rubin, 1985) which is an expression for the standardised mean difference of a covariate between the two groups remaining after matching. Usually, covariates are seen as well-balanced if the bias is below 5 per cent or 10 per cent (Caliendo & Kopeinig, 2008; DiPrete & Gangl, 2004). The results of the tests on each of the covariates can be obtained from the author. The last two indicators are based on a re-estimation of the PS with the matched sample. As proposed by Sianesi (2004), we compare the predictive power of all covariates with regard to the participation decision before and after matching, expressed by the pseudo-R²s. The pseudo-R² should be rather low after the matching, indicating that the covariates are similarly distributed in the treatment group and the matched control group (Caliendo & Kopeinig, 2008). Similarly, the likelihood ratio test on joint significance measures if the used regressors are able to explain insurance enrolment. The null hypothesis that all coefficients in the PSM regression are equal to zero should not be rejected after the matching.

15. Common support is the default option in radiusmatch.

16. Recently, discussion on the validity of bootstrapped standard errors in matching methods came up. Abadie and Imbens (2008) show that bootstrapping in nearest-neighbour matching with replacement and a fixed number of neighbours fails to provide consistent standard errors. A major reason for this is that the sample of matched controls does not increase with the sample size and therefore efficiency gains are not realised in greater samples (Blundell & Costa Dias, 2009; Gangl, 2010). However, this seems to be less a concern for Kernel and Radius matching where the number of used control observations expands with sample size (Abadie & Imbens, 2008; Blundell & Costa Dias, 2009; Gangl, 2010; Wagstaff & Yu, 2007).

17. As the matching process does not lead to a satisfactory matching quality for the gender subsamples, we do not provide estimates of the treatment effects separately for boys and girls.

18. Shimeles (2010) uses, as we have, the EICV2 data set for his evaluations.

19. We abstract from controlling for child’s health (or the interaction of health insurance and child’s health) in our regressions to isolate this effect, since the variable is likely to be endogenous due to reverse causality and omitted variables, and since valid instruments are lacking.