attributes which might influence both fostering and school enrollment. In the following child fixed
effects specification, I measure the impact of fostering on that child’s educational enrollment, con-
ditional on the child’s unobserved attributes:
Sijt = βo + ai + βι(EverFosteredij * AfterFosteringjt) + ηt + Ψijt (2)
where Sijt and EverFosteredij * AfterFosteringjt are as previously defined, ηt are time dummies
to capture any secular time effects in school enrollment, ai refers to the child fixed effect, and ψijt
is a random, idiosyncratic error term.11 The child fixed effects specification is identified by within
child variation over time and relies on the identification assumption that any unobservable factors
that influence fostering and school enrollment do not vary over time. All time-invariant factors,
such as a child’s ability or personality, will be captured by the fixed effects.12
While these two estimation strategies (household and child fixed effects) improve measurement
of the fostering impact on school enrollment, most panel datasets are only able to compare foster
children with their current host siblings and are still not able to fully measure the fostering impact.
Even if the foster child is treated poorly and is worse off after the fostering relative to his new
host siblings, the foster child still might be better off in terms of school enrollment relative to the
treatment he would have received if he had stayed with his biological family. It is impossible to
measure the “true” counterfactual that would compare the school enrollment change for the foster
child if he is sent to a host family with the school enrollment change for the same foster child in the
same time period if he had instead remained behind. However, with this dataset, it is possible to
11In equation 2, I do not include the term EverFosteredij because it will be absorbed by the fixed effects.
12If the source of the fostering’s endogeneity is time varying, the fixed effects estimation strategy will not be
able to address this problem. To deal with any time-varying unobservable factors, I tried an instrumental variables
estimation strategy using household level agricultural shocks and network quality as instruments for fostering. These
instruments have strong explanatory power in explaining why a household sends out a child in a given year (Akresh,
2004), but they have low power as instruments for the EverFosteredij * AfterFosteringjt variable which implicitly
is also measuring the duration of the fostering.
13