Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Re: st: -mlogit- iterations not concave if not omitting intercept


From   "Filippo Maria D'Arcangelo" <filippo.darcangelo@unibocconi.it>
To   statalist@hsphsun2.harvard.edu
Subject   Re: Re: st: -mlogit- iterations not concave if not omitting intercept
Date   Fri, 12 Jul 2013 17:32:17 +0200 (CEST)

I finally decided to trim the categories from 16 to 10.
The model converges pretty well now. This way, I also hope to reduce the IIA. I know that 10 categories are probably still too many but I couldn't go with less.
Moreover, I took Maarten's advice not to remove the constant as there was no reason to do that (except for having the log likelihoods to converge, which is fairly not a good reason).
I still don't have completely clear why it worked, though.
Thank you,

filippo


On Fri, Jul 12, 2013 at 10:13 AM, Filippo Maria D'Arcangelo wrote:
>> I am using the -mlogit- command in order to estimate probabilities over 16 categories of a dependent variable.
>> It is performing pretty decently when i omit the intercept (with the -, nocons- option).
>> I have 8 dependent variables plus a categorical variable that i wish to include. This categorical variable has 10 categories and i use the -i.- prefix when i run the regression. Two of the categories are omitted (not one): i think this is due to the few observations in one of them.
>>
>> As soon as I regress without the -, nocons- option, my regression suffers from non-concavity. I think that the problem is in the categorical covariate.
> 
> You are asking a lot from your data, probably too much. I would not
> expect an -mlogit- with a 16 category depedent variable and a 10
> category independent variable to converge in most datasets I usually
> work with (approximately 5,000 - 200.000 observations). It might
> happen, but it would surprise me when it does. You are just spreading
> your data too thin. It may work in Census like datasets, but even
> there it could easily be that the variables are such that data just
> does not contain the information necessary to estimate that model. The
> fact that your model seems to work with the -nocons- option makes me
> very worried. Unless you manually entered that constant back in in
> some way (e.g. M.L. Buis (2012) "Stata tip 106: With or without
> reference", The Stata Journal, 12(1), pp. 162-164.) you should never
> exclude the constant in a model like -mlogit-.
> 
> -- Maarten
> 
> ---------------------------------
> Maarten L. Buis
> WZB
> Reichpietschufer 50
> 10785 Berlin
> Germany
> 
> http://www.maartenbuis.nl
> ---------------------------------
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index