counties jointly is not possible, unless the assumption is made that doctors in both
counties react in the same manner to a lawsuit.
In order for the parameter of interest (al) to be identified, the time trend must be the
same for the untreated group (i.e., doctors who are not sued)15. Because it appears that
lawsuits are random, it is assumed that this condition is met. After controlling for doctor,
year, and hospital effects, aL is the estimate of the treatment effect of being sued. If
doctors react to lawsuits by increasing the number of c-sections, then aL should be
positive. The same will be true for a2, a3 and a4 if the response is not negligible for
additional lawsuits. Because the variation in lawsuits occurs at the doctor level, the
standard errors were clustered at the doctor level. The variables of interest in these
regressions are the sued variables. If the hypothesis that at doctor only responds to a first
lawsuit is true, then the coefficients on sued2 sued4 should not be significantly different
from zero.
With a binary dependent variable, a logit or probit model might be assumed. In
actuality, the combination of doctor level fixed effects as well as a doctor level treatment
effect implies that the dependent variable is the average of the doctor's procedure use in a
given year. The same model could have been estimated by collapsing the data to doctor
and year means and then applying ordinary least squares. I chose to estimate these
models at the individual patient level to ensure the most precise estimates possible. The
estimates are not qualitatively different when averages are used.
15 Because of the strong time trend and the fact that there are not a large number of lawsuits in a given year,
this is difficult to show.