812 Computational Considerations in Microeconometrics
Table 15.6 Endogenous NB, MSL withS=50 or 100(1) S= 50 (2) S= 100
Coef. Std. error Coef. Std. errorHMO
AGE 0.173 0.018 0.172 0.0184
FIRMSIZE 0.0224 0.00120 0.0224 0.00121
Intercept −1.449 0.0735 −1.446 0.0737
OMC
AGE 0.229 0.0301 0.231 0.0301
FIRMSIZE 0.0190 0.00194 0.0190 0.00193
Intercept −3.505 0.125 −3.514 0.125
DOCVIS
HMO 1.161 0.0573 1.080 0.254
OMC 0.511 0.173 0.850 0.424
AGE 0.225 0.0125 0.224 0.0123
Intercept −0.519 0.0760 −0.529 0.103
LNALPHA
Intercept 0.197 0.0625 0.163 0.0716
λHMO
−0.976 0.0498 −0.866 0.306
λOMC
−0.107 0.180 −0.473 0.474
log-like − 41725 − 41722
Comp. time 932.11 seconds 1165.27 seconds15.6.3 Example: MSL estimation
To illustrate the method, we use pooled data from the Medical Expenditure
Panel Surveys 1996–2003. The sample consists of 13,469 observations on persons
between the ages of 19 and 64. The outcome variable is the number of doctor
visits in a year (DOCVIS) and the multinomial treatment variable describes the
type of health insurance plan (INSTYPE) and takes three values: (i) fee-for-service
(FFS)–the control; (ii) health maintenance organizations (HMO); (iii) other man-
aged care organizations (OMC). Exogenous covariates are AGE and FIRMSIZE; the
latter serves as an exclusion restriction, i.e., as an instrument. The specification is
deliberately (over)simplified by excluding many variables that would appear both
in the choice and outcome models. Our objective is to demonstrate the feasibility
of computation. MSL estimates, obtained using Halton sequences withS=50 and
100, are given in Table 15.6.
The results show that even for such a simplified example the computational
time is nontrivial. The sample size in this example is fairly large, which is expected
to improve the precision of the estimation. Although the log-likelihood values are
similar forS=50 andS=100, we see some differences in the coefficient estimates.