Re: st: RE: Backed Up Iteration

[William Hauser <whauseriii@gmail.com>]

Anna-Leigh,
I don't know about fractional models but I doubt the size of the
dataset has anything to do with your difficulties.  I'm working with a
dataset of over 1.4 million cases and aside from the time it takes to
compute some math intensive models everything seems to work just fine.

It's probably a consequence of some pattern in the data such as
(multi)collinearity or maybe a lack of variability.  Definitely check
and recheck your variables for miss-coded or unusual/extreme cases. I
was having trouble getting a hierarchical logit model to converge and
traced the problem to some extreme overdispersion and outliers on two
of my predictor variables (which was in itself unexpected since the
distribution of the predictors shouldn't matter).

You say adding a 4th variable in, no matter what it is, prevents
convergence.  Have you tried starting with a different first 3
variables and then seeing if a 4th can be added? With 15 predictors
you've got a lot of potential combinations to choose from.
Objectively, the number of variables isn't as problematic as the
strength of association among them and the degree to which that's
shared with the variation in the outcome.  It's hard to make more than
generic recommendations without understanding a bit more about the
specifics of the model you're trying to estimate.

Caveat emptor, this advice is probably worth about what you paid for it.

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/