Andrew M. Jones 613here is those exempt from these charges, made up of those with private insurance,
co-insured children and low-income households. A DD strategy is adopted, with
the effect of the co-payments on doctor visits identified by the interaction between
the timing of the reform and exemption from charges. Winkelmann argues that the
assumption of a common trend, which is essential for identification, is plausible
in this context.
The use of a control group in DD methods help avoids the spurious inferences
that can arise in a simple before and after comparison. For example, in Wagstaff
and Yu’s (2007) evaluation of the impact of the World Bank’s Health VIII project
in Gansu province in China they find only a small reduction in out-of-pocket
spending on health care in counties exposed to the program. A before-and-after
comparison would suggest that there was no significant improvement in this
outcome. However, the trend in the control group of counties shows a rise in
out-of-pocket payments, so the DD estimate reveals better outcomes among the
treated relative to the controls. An important caveat is that the validity of the
DD estimates relies on the comparability of the treated and controls in terms of
the underlying trend in the outcomes. The comparability of treatment and control
groups can be enhanced by combining the DD approach with matching estimators
(as in Dawsonet al.,2007; Galianiet al.,2005; Wagstaff and Yu, 2007). In Wagstaff
and Yu (2007), the unmatched DD estimate for the impact of Health VIII on the
availability of medical equipment in township health centers does not show a sig-
nificant effect, but when the controls are matched with treated counties an effect
is revealed. This is because the counties selected for the project tended to be poorer
than the average among the controls so that, on average, the funds available to
invest in equipment in the control counties lead to higher rates of increase. Match-
ing with control counties that face the same sort of financial constraints allows a
reliable comparison to be made.
The careful use of matching estimators, which should include tests of whether
the treated cases and selected controls are balanced in terms of their observed
characteristics, provides a link to strategies for testing the robustness of the iden-
tification assumptions that are built into the DD approach. The comparability of
the treatment and control groups can be assessed by comparing their observed
characteristics prior to treatment and, in particular, by testing the parallel trend
assumption prior to treatment. A good example of this is Galianiet al.(2005). They
use a DD strategy applied to panel data on municipalities in Argentina. The treat-
ment of interest is the privatization of local water services that took place in the
1990s and the outcomes are measures of general and cause-specific child mortal-
ity. There is sufficient data for the pre-treatment period to do graphical analysis of
the trends in the treatment and control groups. More importantly, it is possible to
estimate the two-way fixed effects specification for mortality rates only using the
data from the pre-treatment period, but including an indicator of which munici-
palities would go on to be treated. Evidence that this indicator is significant would
undermine the parallel trends assumption and mean that areas that privatized their
water supply were systematically different in terms of (trends in) mortality. In fact,
Galianiet al.find that the common trend assumption is not rejected. Their DD