RE: st: asclogit not converging

["Jacobs, David" <jacobs.184@sociology.osu.edu>]

One possible culprit that you might not have already checked concerns disparities in the range of explanatory variables.  

If you have a variable like U.S. city population, which may run from (say) 25,000 to almost 9 million, divide that variable by 10,000 or 100,000 or whatever it takes to get the variable in question down to a range like that of the rest of the variables in your model.  I find this simple trick often works when zero inflated or other count models won't converge.  

It may work for your models.

Dave Jacobs

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Nick Cox
Sent: Tuesday, February 19, 2013 8:38 AM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: asclogit not converging

This kind of question arises frequently, and for understandable reasons. But it is usually really hard to say anything that is useful and not banal.

This is not a help line, but a discussion list, so for us to be able to help you have to have a problem that can be discussed.

At least some of us might feel that there needs to be a generic FAQ.

1. We don't know your research problem and are not well placed to advise on better models matching your project aims and constraints.

2. We can't see your data.

3. If a model fit doesn't converge, Stata is quite possibly telling you that the model is not a good idea. You may need a different idea, not another version of the same idea.

4. (Personal opinion) If a model didn't converge, results from early iterations are at best dubious and at worst useless. As you can't tell which, go for useless.

Specifically, in this case, how did you "normalise" a count varying between 1 and 40? (Standardisation in the sense of any linear re-expression I would expect to be cosmetic here.)

My only concrete guess here is that you have one or more outliers in your extra variable -- and that even under transformation they remain outliers. That's often highly problematic. Also, some crazy transformations that people try often make things worse not better.

Nick

On Tue, Feb 19, 2013 at 12:55 PM, Subramanian, Hemang <Hemang.Subramanian@scheller.gatech.edu> wrote:

> I am using Stata MP 12.1.
>
> 1.I use an alternative specific conditional logit choice model with the following command.
> asclogit achoice ratio_simulcast buyer_distance supply  meanpv  
> price_variability, case(caseid) alternatives(auction_code) There are about 115962 cases with an average of 9 observations/units per case.
>
> 2.The moment I introduce another variable "number_of_centers" into this model as follows:
> asclogit achoice ratio_simulcast buyer_distance supply  meanpv  
> price_variability number_of_centers, case(caseid) 
> alternatives(auction_code)
>
> The model stops converging. The number_of_centers variable is the total number of auction centers, and is a number between 1 and 40, different for each observation.
> I tried various combinations, including standardizing the variable, 
> normalizing the variable, etc, but the results never converge - I need 
> help here--
>
> 3.---with iter(15) or iter(n) --- option however, I am able to retrieve the corresponding coefficients of all other variables. Is this the correct usage??
>
*
*   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/

*
*   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/