TSP
θi
= 1
—
= 1
—
LIMDEP
STATA
Although the three estimators are asymptotically equivalent, they show different small-sample
properties.21
The three software packages have been used to estimate 5 possible model specifications adopting
both fixed and random effects. The execution times (on a Risk6000 990 platform) and program sizes are
compared in the following table:
Package |
Execution Time (sec.) |
Lines Code |
TSP |
5 |
30 |
LIMDEP |
12 |
28 |
STATA |
38 |
46 |
On the issue concerning the numerical differences of the random effects estimates, we have analyzed
the comparative efficiency of the small-sample properties of the three estimators. This task has been
accomplished by running a parametric bootstrap procedure starting from the OLS estimation residuals in
one selected specification22. The parametric bootstrap used in this exercise is not a bootstrap in the usual
sense of the word, it consists in the Monte Carlo exercise described in the following seven steps:
21 On this subject see Swamy and Arora (1972).
22 See Bellmann, Breitung and Wagner (1989).
12