Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
| From | brendan.halpin@ul.ie (Brendan Halpin) |
| To | statalist <statalist@hsphsun2.harvard.edu> |
| Subject | st: non-convergence |
| Date | Tue, 07 Feb 2012 23:34:48 +0000 |
I'm having trouble with convergence, fitting a moderately complex multinomial logit. I have a substantively important interaction (categorical, 4*4) that is sparse in the data. For some subsets of the data it is properly estimated and is very significant by the LR test. For others it complains "not concave" during the ML iterations and reports "Warning: convergence not achieved" at the end. Typically, for one parameter estimate, the SE is not reported and for some others the parameter estimate is large and the SE enormous. I'm fitting these models programmatically on a reasonably large number of subsets of the data (for multiple imputation), and I'm primarily interested in the predicted probability. Two questions: First, is there a way to use the factor-variable notation to suppress a particular parameter in the interaction (or should I just use a recoded copy of the problematic variable)? Second, are there any circumstances under which the predicted probabilities might be usable after all? Regards, Brendan -- Brendan Halpin, Department of Sociology, University of Limerick, Ireland Tel: w +353-61-213147 f +353-61-202569 h +353-61-338562; Room F1-009 x 3147 mailto:brendan.halpin@ul.ie ULSociology on Facebook: http://on.fb.me/fjIK9t http://teaching.sociology.ul.ie/bhalpin/wordpress twitter:@ULSociology * * 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/