Andrew M. Jones 603the unobservables is specified and the full model is estimated by ML. In this case
the outcomes of interest are binary and count measures of health care utilization.
The treatment variables reflect the individual’s choice of insurance plan, which is
modeled using a random utility framework. The options are categorized as HMOs,
other-managed care and non-managed care plans, with the choice of options
specified using a multinomial logit specification. In Deb and Trivedi (2006) the
data are taken from the 1996 US Medical Expenditure Panel Survey (MEPS), while in
Debet al.(2006a) the MEPS data are augmented by the 1996 Community Tracking
Survey (CTS) as a test for the external validity of the findings. Utilization and insur-
ance plan are modeled simultaneously to take account of the possibility of selection
by patients, insurers and providers. Identification of the latent factor model rests on
the assumption that each of the multinomial choices depends on a unique latent
factor, based on independent and identically distributed (i.i.d.) normal draws, and
these are allowed to be freely correlated with the error in the outcome equation. The
parametric specification is identified by functional form, but exclusion restrictions
are also imposed, using employment status and occupational sector as predictors
of insurance plans, while excluding them from the outcome equations. Simple
linear models are used to provide informal checks for the validity of these instru-
ments. Estimation uses MSL, accelerated by using quasi-random draws from Halton
sequences. The results do not show evidence of favorable selection into HMO plans,
but the average treatment effects on the use of care are much larger when selection
is taken into account and there is considerable heterogeneity in the effects. Monte
Carlo simulation is also used to compute the standard errors of these treatment
effects.
Other applications of MSL include Lindeboomet al.(2002), who use the LASA
panel to estimate a five-equation model for the use of long-term care services
among elderly residents of Amsterdam. They take draws from a multivariate nor-
mal distribution and use antithetics to accelerate the estimation. The results show
strong effects of health status, sex, socioeconomic status and prices on the use of
institutional care. Two papers by Pudney and Shields (2000a, 2000b) use MSL to
estimate a system of equations comprised of a generalized ordered probit model
for British nurses’ pay grades along with auxiliary equations for training, career
breaks, work outside the NHS, and part-time work. A common factor structure
leads to a log-likelihood based on the multivariate normal and hence the need
for simulation estimation. The MSL estimation only uses 50 replications, without
acceleration, which is relatively few for this kind of application. Pudney and Shields
(2000b) make a case for identification based on functional form rather than exclu-
sion restrictions. Vera-Hernandez (2003) makes use of MSL to estimate a structural
model of insurance coverage and health care use applied to data from the RAND
Health Insurance Experiment.
12.6.2 Applications using Bayesian MCMC
Patient selection makes it hard to identify the impact of the size and characteris-
tics of hospitals on the quality of their outcomes. The problem arises when there is