Andrew M. Jones 595have to be uncorrelated with the individual effect. In this case the orthogonality
conditions for the systems approach are rejected. The hypothesis of a unit root is
not rejected and models are also estimated for differences in log-market share as
a function of levels of the regressors. The estimated short-run price elasticities are
small but, due to the dynamics, the long-run impact of prices is substantial.
Other applications of the GMM and systems-GMM approaches include Baltagi
et al.(2005), who estimate dynamic equations for the log of hours worked by
Norwegian doctors, using panel data from the personnel register of the Norwegian
Association of Local and Regional Authorities. Clark and Etilé (2002) use seven
waves of BHPS to estimate dynamic models for cigarette consumption. This creates
some problems when the data are first differenced as there is considerable heaping
of the self-reported data around focal values, such as 20 cigarettes per day. Health
shocks are shown to influence levels of smoking. Hauck and Rice (2004) use 11
waves of the BHPS to estimate models for mental health, measured by the GHQ-
12 score. They compare static variance components models and dynamic GMM
estimators and find greater persistence of mental health problems among those
with lower socioeconomic status. Windmeijeret al.(2005) use the systems-GMM
estimator in panel data models of the demand for outpatient visits by GP practices
in England.
12.5.2 Applications with categorical outcomes
12.5.2.1 Pooled and random effects specifications
Contoyanniset al.(2003) consider the determinants of a binary indicator for func-
tional limitations using seven waves (1991–97) of the BHPS. Their models allow for
persistence in the observed outcomes due to state dependence (a direct effect of pre-
vious health status), unobservable individual effects (heterogeneity which is due to
unobserved factors that are fixed over time) and persistence in the transitory error
component. Allowing for persistence is important: a comparison of the observed
outcomes with those predicted by a simple binomial model shows that persistence
is substantial in the data. They estimate models for the repeatedly observed binary
health indicator, with and without state dependence, using panel probit models.
These are estimated by MSL using the GHK simulator with antithetic acceleration.
They also implement a test for the existence of asymptotic bias due to simulation
which is used to select the number of replications required for use in MSL.
In related work Contoyanniset al.(2004b) explore the dynamics of SAH in the
BHPS. The variable of interest is an ordered measure and the BHPS reveals evidence
of considerable persistence in individual’s health status. As SAH is measured at
each wave of the panel there are repeated measurements for a sample of individu-
als. SAH is modeled using a latent variable specification, which is estimated using
pooled ordered probits (with robust inference) and random effects ordered probit
models. The presence of lagged health is designed to capture state dependence, the
influence of previous health history on current health. The error term is split into
two components; the first captures time invariant individual heterogeneity; the
second is a time varying idiosyncratic component.