can be interpreted as the impact of fostering on enrollment under the assumption that, without
the fostering episode, the change in enrollment for the two groups would not systematically differ.
After the fostering, the change in enrollment for foster children is 1.8 percent higher than that of
host family siblings. The result is not statistically significant, but making full use of the panel
dimension of the data in the following sections yields more precise estimates.
To incorporate all available information, I use a household fixed effects regression which is com-
parable to the difference in differences estimator. In the simplest household fixed effects specification
(additional age and gender controls are added later), I estimate the following:
Sijt = βo + 7j + βι(EverFostered^ * AfterFosteringjt) + β%(EverFosteredij) + δt + εtjt (1)
where Sijt is the school enrollment status for child i in household j at time t, where household j
refers to either the host or biological household, 7j is the household fixed effect, EverFosteredij *
AfterFosteringjt indicates the years after the fostering for the foster child, EverFosteredij indi-
cates if the child is a foster child, δt are time dummies intended to capture any secular time effects
in school enrollment, and εtjt is a random, idiosyncratic error term.10 The coefficient β1 is the
effect of fostering on school enrollment for the foster child compared to the host siblings in the
same household. The main identification assumption for the estimate of βι to be consistent is that
any factors that influence why certain households send and receive children are captured by the 7j
household fixed effect term, and these factors do not vary over time. The household fixed effects
specification is identified by variation across children within the same household over time.
In addition to controlling for unobservables within the household that might be correlated with
fostering and school enrollment, a related exercise would be to control for a given child’s unobserved
10The secular time effects could also be captured by including an AfterFostering main effect, although that is
more restrictive than including unrestricted time dummies as in the text. Both approaches yield similar results.
12