st: RE: Re: xtmelogit convergence issues and log transforming IVs

[William Hauser <whauseriii@gmail.com>]

Nick,
Excellent point regarding reviewers.  Regarding your point 1, I found
it very strange that the model would not converge with the crime
seriousness variable in it because, as you say, normality of the
predictors isn't supposed to matter.  But it does, or at least it
seems to.  I can't provide you with a mathematical proof but I can say
that the model won't converge with either or both of the
non-transformed predictors in the model (even as the sole predictor).
At first I thought it was a property of the hierarchical model but the
non-convergence also occurs in the ordinary non-nested logit model
("logit prison_sentence crime_seriousness, or").  The problem
magically goes away when the variable is transformed.

I thought about asking about this because, well, it runs contrary to
everything I've been told but then I decided the issue was tangential
to the bigger problem of how to transform and interpret the variables
afterwards.  It's a interesting question and one I'd love to hear
speculation about but I can just side step the whole issue by
transforming the variable.

In case you're interested, both variables prevent convergence either
singly or together.  Prior record points has a skew of 1.8 and
kurtosis of 7, crime seriousness points is more seriously non-normal
and has a skew of 10.5 and a kurtosis of 520 (again, these are
"points" assigned under a arbitrary scoring system).  It looks to me
like it's the overdispersion and not the skew that's creating problems
but I have not a clue why.

Will Hauser
*
*   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/