Arkadipta Ghosh wrote:
I am trying to run a Zero inflated Poisson model (both with and without fixed
effects), and the ml estimation process does not converge even after hundreds
of iterations. It keeps saying "not concave". I understand that in such cases
one can write an ml program or try setting initial values to ensure a quicker
convergence. I have tried both but something seems to be going wrong with
those. In other words, I might be making some basic mistake somewhere or doing
something wrong. Is there a better procedure or a good resource that I can look
up (on the web, may be)? Sorry for this basic question but I am relatively new
to ML estimation and thought you could help. Thanks,
--------------------------------------------------------------------------------
By "both with and without fixed effect," do you mean that it won't converge
even with a constant-only model, i.e., -zip response_variable, inflate(_cons)-?
Or does it mean that you stll keep a lot of variables in the equation predictng
zero counts, i.e., in the -inflate()- option? If it's the latter, then some of
the variables might be collinear or nearly collinear, or one or more might have
a coefficient with maximum likelihood when it's essentially at infinity, or
there might be a problem with scaling between the variables causing numerical
problems. (I don't have any particular insight into ZIP modeling, and these
possibilities are just generalization of those given in a FAQ at StataCorp's
website as to what might cause not-concave messages at at least the last
iteration for -ml-, in general. See www.stata.com/support/faqs/stat/ml.html
and, in particular, the passages at the end of the section titled, Problem 2:
Nonconcave log-likelihoods.)
Joseph Coveney
*
* 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/
