5.5 Specication checks
One of the identifying assumptions is that the health shock is uncorrelated to the error term in
equation (3), conditional on the included covariates in the model. As mentioned previously, health
shocks may be correlated with unobserved changes to household resources, even with the rst-
dierenced specication of the model. To assess whether correlated changes in household economic
conditions are contributing to the estimated eect of the health shock, I examine the correlation
between lagged income changes and levels, and the current health shock. If past economic shocks
(such as job loss) aect current health, one would expect a negative correlation between the health
shock and lagged income changes. Alternatively, if past levels of income aect investment in health
dierently between the rich and the poor, one would also expect a negative correlation between the
current health shock and the lagged income levels.
In Table 11, I show that there is little correlation between past income changes or levels and the
current health shock, indicating that this potential source of bias may be less of a concern in the
sample of study. For instance, in columns (1){(2), the coecients on lagged level (column 1) and
lagged growth (column 2) of earnings of the head and spouse are small and insignicantly dierent
from zero. The inclusion of county-by-year xed eects potentially accounts for many confounding
factors aecting the variations in health shocks.
Even if the eect of the health shock is contaminated by the eect of other types of shocks to
household resources, the estimated coecient on the interaction between the health shock and the
reform (1) may be unbiased as long as the omitted variable bias in the coecient of 4hijt does not
change with the reform. This is more of a concern if the reform changes the unobserved composition
among households suering from health shocks. Columns (3) reports the coecient estimate on