Our goal is to identify the extent to which variation in these child human capital–related outcomes can be attributed to the death of a parent in the tsunami. A natural starting place is to estimate the relationship between each of the shorter-term and longer-term outcomes, Y
, for child i at time t (where t spans the period before and after the tsunami), and parental death, D
, controlling time-varying and time-invariant child and family characteristics, X
Parental death, D, is vector-valued, distinguishing children who lost their mother, those who lost their father, and those who lost both parents in the tsunami. An important advantage of our research is that parental death does not reflect prior health-related behaviors but is the consequence of a large and unexpected natural disaster. Estimates from Eq. (1) can be interpreted as causal if parental death in the tsunami is exogenous in the model: that is, if unobserved heterogeneity is not correlated with covariates in the model including parental death. This strong assumption underlies much of the existing literature. In our context, it is reasonable to suppose that parents who survived the tsunami are stronger or better swimmers than other parents, or that they lived in more robust houses. If those parents also invested more in the human capital of their children prior to the tsunami, then the assumption that is unrelated to parental death will be violated.
If such differences exist and reflect traits that do not change during the study period, they can be taken into account in Eq. (1) by including a child-specific fixed effect. Specifically, separating unobserved heterogeneity into two components—a fixed effect that is time-invariant for each child () and a component that varies over time (), we rewrite Eq. (1) as
The fixed effect absorbs all characteristics of the parent and child that do not change over time and affect the outcome, Y
, in a linear and additive way. These include observed characteristics, X
, in model (1) along with unobserved characteristics that are included in in that model. The latter might include, for example, parents’ tastes for investments in their children; characteristics of the child, such as ability and ambition; and characteristics of the family and community in which they were living at the time of the tsunami.
Estimates of Eq. (2) require repeated observations of the same child before and after the tsunami. We examine indicators of schooling and time allocation that were measured for the same child before the tsunami and again after the tsunami.
One indicator was collected only after the tsunami. In the first resurvey, we asked about participation in programs implemented after the tsunami to assist families, including whether the child received a scholarship from such a program. Because this scholarship program did not exist at baseline, we assume that no child received one of these scholarships at that time.
Before presenting our empirical results, we assess whether parental death can be treated as exogenous in model (2). Indicators measured in the pre-tsunami baseline for children whose parents subsequently survived the tsunami are compared with indicators for children who lost one or both parents in the tsunami.
The first row of Table 2 shows that, on average, children whose parents survived the tsunami were 12.9 years old at baseline (column 1), whereas children who lost any parent were age 13.5 (column 2); this 0.6-year difference is significant (column 3). The differences for children who lost their mother, their father, or both parents relative to those whose parents survived are displayed in columns 4, 5, and 6, respectively; none of these differences are significant.
The second row of the table indicates that males constitute a significantly higher fraction of survivors among children who lost parents relative to children whose parents survived. The difference is largest for children who lost both parents; in this group, 19.9 % more young males survived than young females, and this difference is also significant.
Children whose parents died in the tsunami were also significantly better educated and significantly more likely to be enrolled in school prior to the tsunami. They were less likely to be working or engaged in housekeeping in the week before the pre-tsunami survey relative to those whose parents survived, although these differences are not significant. The rest of Table 2 compares characteristics of parents and households of those children whose parents survived the tsunami relative to those whose parents did not survive. None of these differences are statistically significant. We also estimate models that include a community fixed effect, which compares children within each community. In these models, we find no statistically significant differences in any of the indicators in the table between children who lost one or more parents and those who did not. Thus, part of the differences between orphans and non-orphans can be attributed to differences across study sites, including the likelihood of death of parents. For some outcomes, differences between children who were orphaned and those who were not are similar in magnitude in models with and without community fixed effects, suggesting the possibility that even within communities, preexisting differences may exist between children who were orphaned by the tsunami and children who were not orphaned.
Results in Table 2 establish that children who lost parents in the tsunami had higher levels of human capital before the tsunami than children whose parents survived the tsunami. To the extent that these pre-tsunami differences are not absorbed by observed characteristics in the model (in Eq. (1)), unobserved heterogeneity in the model will be correlated with parental death and estimates will be biased. If, however, these pre-tsunami differences reflect influences that are fixed for a child over the study period, they will be absorbed in the child fixed effect, and estimates in Eq. (2) can be given a causal interpretation. The results in Table 2 underscore the critical importance of having a pre-tsunami baseline in order to identify the causal effect of parental death on child outcomes.
We have established that males are more likely to have survived the tsunami than females. The male survival advantage also holds for adults, which has been attributed to the fact that males are stronger and, in Islamic Aceh, much more likely to know how to swim than females (Frankenberg et al. 2011). We estimate separate models for males and females.
We also explore whether other attributes are associated with children’s survival status (Table 5 in the Appendix). The only significant difference is that children who helped with housekeeping were also more likely to survive. This difference, however, is small in magnitude and is both smaller in magnitude and not significant when comparisons are drawn within communities. The evidence indicates that net of age and sex, children’s deaths are not significantly related to pre-tsunami own human capital, parental human capital, or household resources.
The final three columns of Table 5 compare the same indicators for respondents who were interviewed in the first follow-up with those who were not. None of the differences are significant; and, taken together, the indicators explain only 1.2 % of the variation in the probability that an individual was not interviewed in the follow-up survey (F statistic for the significance of all the covariates in Table 5 = 1.2; p value = .31). In short, we find no evidence that attrition is selected on observed characteristics measured at baseline.