st: Re: Trouble with ml init and convergence

["Joseph Coveney" <jcoveney@bigplanet.com>]

Dana Chandler wrote (excerpted):

. . .

I'm estimating a model where I have a cross-section that includes the
population and the number of religious organizations within several
hundred towns. Each town has 0 to 5 organizations (top-coded) and I'd
like to estimate an ordered probit where the outcome is the number of
religious organizations and the predictors are population or some
transformation of the population variable.

In the simplest model, I estimate "oprobit outcome pop". This works
fine. However, for another specification, I want to allow population
to have a differing effect on the outcome variable depending on how
many organizations are already in the town. Hence, I create dummies
pop_w0-pop_w5 where I calculate DUMMY(outcome>=N)*pop. Next, I
estimate "oprobit outcome pop pop_w1 pop_w2 pop_w3 pop_w4 pop_w5".
This DOESN'T work which is why I start to use the oprobit that I
programmed with ml lf where hopefully I can set initial values to help
it converge. Unfortunately, I'm still having trouble with that.

. . .

--------------------------------------------------------------------------------

I don't think I can help you with choosing initial values, but I do have a
couple of observations that might help.

First, from your description of what it is that you're ultimately trying to do,
it seems that what you're looking for is essentially a relaxation of the
proportional odds assumption in an ordered logistic regression (this might have
already been suggested in reply to your earlier post in January).  If so, then
-gologit- or -gologit2- would be the way to go, instead of attempting to place
constraints on each of an exhaustive set of predictors (I assume that you've
tried the -collinear- option of -oprobit-) that are generated as a determinate
function of the response and remaining predictor.

Second, would it be beneficial to view the number of religious organizations in
town as a count variable rather than an ordered categorical variable?  If so,
then you'd be looking at Poisson or negative binomial regression with a
(quadratic, cubic, quartic, fractional, whatever) polynomial or (restricted
cubic or B) spline in population.

Joseph Coveney


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/