608 Panel Data Methods
exact matching and instead weight observations by their proximity, in terms of
the propensity score.
An important requirement is that the model for treatment, used to construct
the propensity score, should only include variables that are unaffected by par-
ticipation in the treatment, or the anticipation of participation. The matching
variables should be either time invariant characteristics or variables that are mea-
sured before participation in the treatment and that are not affected by anticipation
of participation. The crucial condition for identification of treatment effects using
the matching approach is that the selection into treatment should be ignorable,
conditional on the observed covariates.^3
Joneset al.(2007a) use the ECHP to estimate the impact of private health insur-
ance coverage on the use of specialist visits in four European countries that allow
supplementary coverage. The results show that the probability of having private
insurance increases with income and with better reported health. Private insurance
has a positive effect on the probability of specialist visits in all countries, although
the magnitude is sensitive to the choice of estimator. They match treated indi-
viduals with non-treated individuals inversely weighted for the distance in terms
of estimated propensity scores, with weights constructed using kernel smoothed
distance weighting. They ensure that all cases are supported by controls. The qual-
ity of the matching can be assessed by computing the reduction of the pseudo R^2
of the insurance regression before and after matching. To evaluate the extent to
which matching on propensity scores balances the distribution of thexs between
the insured and the uninsured group, they compute the bias reduction due to
matching for each of thexs.
Dano (2005) uses a 10% sample of the Danish population, drawn from register
data, to give a panel for 1981–2000. She estimates the impact of injuries sustained
in road traffic accidents on economic outcomes. Although these accidents are un-
anticipated health shocks, their incidence varies with observable and unobservable
characteristics that can be associated with the outcomes and the estimates of the
treatment effect need to be adjusted for this. Due to the large sample size, one-to-
one matching without replacement is used, with matching on the linear index from
the propensity score. The matching is combined with a difference-in-differences
approach to control for time invariant unobservables. The study finds an impact
of injuries on earned income for older and low-income individuals, but also shows
the compensating effects of public transfers in the Danish system.
García-Gómez and López-Nicolás (2006) adopt a matched panel data difference-
in-differences estimator. They use panel data from the Spanish component of the
ECHP to explore the impact of health shocks on employment and of employment
shocks on health. Their strategy is to match exactly on pre-treatment outcomes,
in order to control for time invariant observables. This means that controls are
restricted to those individuals who were identical to the treated in terms of their
pre-treatment outcomes. In order to define the treated and the controls, a three-
year window is adopted. In the case of health shocks, the treated are those who are
in good health in the first year and then move to bad health in the next two years,
and who are in employment for the first two years. The controls remain in good