I don't think it is that easy. Adding individual dummy-variables will increase
the number of coefficients with the number of observations. I am not a
statistician but I somehow remember to read that this violates some basic
assumption of Likelihood estimation---and that this is the reason why we use
the conditional logit model for the binary fixed-effects model.
Uli
Joseph Coveney wrote:
> James Shaw wrote:
> > I was wondering if there is such a thing as fixed effects ordinal probit
> > regression. If so, could one simply add dummy variables for the panel
> > indicator (e.g., subject id) to the ordinal probit model to obtain fixed
> > effects estimates? Also, when estimating a fixed effects regression
> > model with a subject-level effect, how problematic is it if there are
> > missing observations on the dependent variable for some subjects (i.e.,
> > unbalanced panels)?
>
> ---------------------------------------------------------------------------
>-
>
> By analogy to -areg , absorb()-, it seems feasible to create dummy
> (indicator) variables for the panel identifier with -oprobit-, but doing
> this in an ordered categorical regression risks having "note: [X]
> observations completely determined. Standard errors questionable." at the
> bottom of the -oprobit- output. This would raise suspicions about Wald
> tests, although in a test case that I tried out where this happens (see
> do-file below), the Wald test agrees well with the corresponding likelihood
> ratio test. If panels are dropped due to collinearity in fitting the full
> model, then likelihood-ratio testing with the reduced (nested) models is
> problematic unless the same panels are fortuitously dropped in the latter.
>
> When observations are missing, handling the panel as a fixed effect seems
> to be mechanically possible--in the test case, -oprobit- attained
> convergence and didn't seem to drop any panels with a missing value--but it
> might be worthwhile to perform Monte Carlo simulations in order to
> determine whether hypothesis testing and parameter estimates behave as
> expected in such a circumstance before using -oprobit- on an unknown
> dataset with missing observations.
>
> -oprobit , cluster()- might serve as an alternative in some circumstances.
> With enough panels, another alternative would be to consider the panel as a
> random effect, and use -reoprob- or -gllamm-.
>
> Joseph Coveney
>
>
> ---------------------------------------------------------------------------
>-
>
> clear
> set more off
> set seed 20030906
> set obs 6
> forvalues i = 1/6 {
> generate float var`i' = 0.7
> quietly replace var`i' = 1.0 in `i'
> }
> mkmat var*, matrix(A)
> local means m1
> forvalues i = 2/6 {
> local means = "`means'" + " m`i'"
> }
> drawnorm `means', n(40) corr(A) clear
> generate byte pid = _n
> forvalues i = 1/6 {
> generate byte res`i' = 1 + int(norm(m`i') / 0.2)
> }
> matrix drop A
> drop m*
> reshape long res, i(pid) j(tim)
> xi: oprobit res i.tim i.pid, nolog
> estimates store A
> test _Itim_2 _Itim_3 _Itim_4 _Itim_5 _Itim_6
> xi: oprobit res i.pid, nolog
> lrtest A, stats
> xi: oprobit res i.tim, cluster(pid) nolog
> xi: reoprob res i.tim, i(pid) quad(30) nolog
> // consider -quadchk- here
> drop if uniform() > 0.85
> xi: oprobit res i.tim i.pid, nolog
> estimates store A
> test _Itim_2 _Itim_3 _Itim_4 _Itim_5 _Itim_6
> xi: oprobit res i.pid, nolog
> lrtest A, stats
> xi: oprobit res i.tim, cluster(pid) nolog
> xi: reoprob res i.tim, i(pid) quad(30) nolog
> exit
>
> ---------------------------------------------------------------------------
>-
>
>
>
>
> *
> * 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/
--
[email protected]
*
* 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/
