[Limdep Nlogit List] convergence & random parameters logit

Andrew Cooper andrew.cooper at unh.edu
Tue Sep 4 22:45:38 EST 2007

I'm currently trying to analyze data from an observational study using 
random parameters multinomial logit.  I have 81 subjects with between 75 
and 400 observations of each subject choosing between the same 9 chioces.

When I run this as a standard multinomial logit, the model converges 
without a problem.  When I switch to RPL, I can only get the model to 
converge if: 1) I set the convergence criteria loose to the point of 
absurdity (e.g., Tlg=Tlf=Tlb=0.1 - and I have my doubts as to whether it's 
really converged) or 2) fix all the random parameters (e.g., set the 
distribution to "C").

Otherwise, I get an error:

Initial iterations cannot improve function.Status=3
   Error   805: Initial iterations cannot improve function.Status=3
Function=  .39524066282D+04, at entry,  .39515005851D+04 at exit
   Error  1025: Failed to fit model. See earlier diagnostic.

I've tried subsampling my data so that I have approximately equal number 
of observations per subject (ie. between 75 and 100).  I've tried 
restricting the choice set to remove the rarely chosen options (e.g., 
chosen less than 1% of the time).  I've also subsampled to remove those 
individuals who always chose the same option. I've tried different fitting 
algorithms (Newton, BHHH, etc.) and different numbers of Pts.  I've both 
included and not included correlation in the random parameters.

None of this, alone or in combination, helped with convergence.

The listserv archives suggested trying a latent variable model, but when I 
do that, I got an error message saying:

At least one parameter did not leave start value.
Normal exit from iterations. Exit status=0.
   Error  1027: Models - estimated variance matrix of estimates is singular

Part of my problem is that I'm not sure why RPL models fail to converge, 
so I'm not sure what else I should be looking for. Any suggestions as to 
what I should try next?  Or, should I just stick with a fixed-effects-only 
repeated-measures multinomial logit by setting  all the random parameter 
distributions to "C"?

Thanks for your advice!!


"What if the Hokey Pokey is all it really is about?" - Jimmy Buffett

Andrew B. Cooper, Ph.D.
Department of Natural Resources
Institute for the Study of Earth, Oceans and Space
Morse Hall 142
University of New Hampshire
Durham, NH  03824

andrew.cooper at unh.edu
603.862.4254, 603.862.0243 (FAX)


More information about the Limdep mailing list