16(27)
Since the households are observed for more than one year, we use the
clustered sandwich estimator of standard errors, which allows for
intragroup correlations, i.e., observations are independent across
households but not necessarily in a household that is observed for several
years. This DD specification could be written as
,
x
(t)
(t)
it
i
)
exp(
0
Where:
)(
0
t
is the baseline hazard function,
x
it
is a vector of covariates
summarizing observed differences between households at time
t
, and the
interaction term,
T d
97
, which is the DID estimator of the reform effect
and variables summarizing the duration dependence of the hazard rate.
Vector
x
it
includes both time-constant and time-varying observable
explanatory variables;
is a vector of parameters to be estimated.
Since this study uses grouped data, where the time intervals are of equal
length, model (3) will be estimated using discrete-time proportional hazard
models and any change in explanatory variables will multiply the hazard
function by a scale function and leave the shape of the baseline hazard
unaffected. The model will first be estimated using the complementary log-
log (cloglog) model (Prentice and Gloeckler, 1978), without including any
frailty term. We will then assume both normally distributed (Bhat, 1996)
and gamma-distributed (Meyer, 1990) unobserved heterogeneity.
Furthermore, since cloglog models assume duration independence (i.e., the
probability of failing at any particular time is always the same), we will take
account of the time dependence by testing both “log” time (i.e., a
continuous time specification) and quadratic time specifications for duration
dependence. These variables are entered as covariates in the regression.
