Re: st: mlogit iterations not concave

[Partha Deb <partha.deb@hunter.cuny.edu>]

Jessica,

I suggested you reduce the number of categories for your dependent 
variable.  It should help you identify sources of singularity or 
near-singularity, which is almost certainly the reason for the "non 
concavity".  I though you were including county-indicators but now it 
seems that you have county-level continuous variables.  If so, keep them 
as continuous variables.  Categorizing those makes is more likely you 
run into problems with "empty" cells.  Your overall problem is, almost 
surely, that your model is not well specified.

To get a feel for where your problem is add the -iter(20) option to your 
-mlogit- command.  -mlogit- will then stop iterating after 20 iterations 
and produce a table of coefficients / standard errors.  Look for 
coefficients that are too big / small and/or standard errors that are 
too big / small or "." .

Hope this helps.

Partha


Jessica Bishop-Royse wrote:
> Dear StataListers,
> I am working on a project where I am trying to estimate a mlogit for a 9
> category nominal dv (cause of death), using both county level predictors (67
> counties) and individual level predictors (continuous vars like education
> and age as well as dummied vars like not married). When I run county level
> ivs by themselves with the dv, I get significant results.  When I run
> individual level ivs by themselves with the dv I get signficant results.
> When I try to incorporate both, county level predictors don't seem to be
> significant.
>
> I originally was trying to estimate multilevel models (gllamm) when it was
> suggested that I try mlogit for the timebeing.  I also dummied the county
> level variables (so that they are no longer continuious, but rather
> representative of established cutpoints).  It was also suggested last week
> that I collapse my dv into fewer categories.  However, when I do this, Stata
> runs and runs and runs and gives me the message that iterations are "not
> concave".  However, no such problems modeling with the 9 category dependent
> variable (just not significant results for county level predictors).
>
> My question is this:  what is this "not concave" message?  And is this an
> indicator that I need to abandon mlogit? Opinion?
>
> . mlogit cause6death black maternaled maternalage real_parity for_born
> pncint pncadqp pncindq crisisbirth multiple if birth1980==1, rrr
>
> Iteration 0:   log likelihood =  -10390.13
> Iteration 1:   log likelihood = -10242.229
> Iteration 2:   log likelihood = -9337.9062
> Iteration 3:   log likelihood = -9238.5535
> Iteration 4:   log likelihood = -9220.1816
> Iteration 5:   log likelihood = -9217.2186
> Iteration 6:   log likelihood = -9216.5682
> Iteration 7:   log likelihood = -9216.4613
> Iteration 8:   log likelihood = -9216.4493
> Iteration 9:   log likelihood = -9216.4468
> Iteration 10:  log likelihood = -9216.4463
> Iteration 11:  log likelihood = -9216.4461
> Iteration 12:  log likelihood = -9216.4461
> Iteration 13:  log likelihood = -9216.4461  (not concave)
> Iteration 14:  log likelihood = -9216.4461  (not concave)
> Iteration 15:  log likelihood = -9216.4461  (not concave)
> Iteration 16:  log likelihood = -9216.4461  (not concave)
> Iteration 17:  log likelihood = -9216.4461  (not concave)
> --Break--
> r(1);
>
> end of do-file
>
> --Break--
> r(1);
>
> --
> Jessi Bishop-Royse
>
> http://sites.google.com/site/wakullagirlssoccer/
> http://sites.google.com/site/jessicacbishoproyse/Home
>
> "Without a struggle, there can be no progress."
> Fredrick Douglass
>
> *
> *   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/
>   

--
Partha Deb
Professor of Economics
Hunter College
ph:  (212) 772-5435
fax: (212) 772-5398
http://urban.hunter.cuny.edu/~deb/

Emancipate yourselves from mental slavery
None but ourselves can free our minds.
	- Bob Marley

*
*   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/