In my prior post in this subject, I erroneously wrote that the intraclass
correlation coefficient was 70% for the simulation; I believe that it was
actually 49% (0.7^2). Bias (and difficulty in getting convergence in the first
place) should increase with increasing intraclass correlation coefficient of
the latent variable. Even with a panel length as long as 36, with a high
enough correlation, the magnitude of the bias could very well be unacceptible:
Stata Corp's posting regarding fixed-effects ordered probit two weeks ago
mentioned Vince Wiggens's simulations with unconditional fixed-effects logistic
regression, which they noted performed poorly until panel length increased to
50, or so.
Note that an analogous concern would apply to a random-effects model that
included a subject-related covariate (fatigue level, for instance) that might
correlate with the random effect and render estimates inconsistent with this
approach, too.
Joseph Coveney
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
