st: Fwd: xtologit, initial values not feasible r(1400);

[Patrick Krueger <pmkrueger@gmail.com>]

Dear Statalist,

I'm using Stata 13.0 to estimate a random effects ordered logistic
regression model via xtologit. I have two-level data (i.e.,
individuals nested within countries); there are 67 countries and
nearly 300,000 respondents across the countries. The minimum number of
individuals in a country is 929.

When trying to estimate an xtologit model for an ordinal outcome
variable, I get the error message: "initial values not feasible
r(1400);"

For example:

. xtologit srhealth, intmethod(mv) intpoints(12) trace

Fitting comparison model:

Iteration 0: log likelihood = -388090.55
Iteration 1: log likelihood = -388090.55

Refining starting values:

Grid node 0: log likelihood = -387924.06
Fitting full model:
---------------------------------------------------------------------

Iteration 0:
Parameter vector:
c1 c2 c3 c4 c5
r1 -4.398648 -2.403305 -.584515 1.234322 1

initial values not feasible
r(1400);

I have tried the following methods to troubleshoot the problem:

1. I fit the model in gllamm without a problem, so it seems that a
solution is achievable. But I was hoping the xtologit command would
run more quickly, especially once I start adding more covariates.

2. I used starting values (both from an ologit command and from the
gllamm results), but received the same error message.

3. I used the "startgrid" option to provide a better starting value
(i.e., the value from the gllamm model) for the random intercept
parameter. Again, I received the same error.

4. I added additional covariates into the model, but still received
the error message.

5. I tried different integration methods in the "intmethod" option,
but still received the error.

Can anyone advise how I might get the model to estimate with xtologit?

As an aside, does anyone know why xtologit is not supported by "mi
estimate"? Is there something that I should be aware of that makes
combining multiply imputed estimates from xtologit models work
differently from combining estimates from other kinds of random
effects models? (Pardon my ignorance on this front, please).

Best,

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