Subject: Re: st: What should I do if stata can not show the result of nbreg

[jhilbe@aol.com]

Ye SUN/Maarten:

While it is likely that Ye SUN has a Poisson and not a NB model, this 
is indicated as a possibility by it's failure to converge. There could 
be other reasons for thie behavior of course. I would like to comment 
on Ye SUNs statement,

> I try -estat gof- to test whether there is overdispersion problem, it 
is
> insignificant indicating no overdispersion problem I think.

This is NOT the way to check for possible overdispersion. Ye SUN, run 
the Poisson model using the "glm" command, checking the value of the 
Pearson dispersion statistic.  It is in the upper right column "(1/df) 
Pearson  = " .  Values greater than 1.0 indicate possible 
overdispersion. HIgher values of this statistic indicate more 
correlation or overdispersion. The overdispersion may in fact be only 
apparent -- and there are ways to check for this. But if the 
overdispersion is real, then the model is not Poisson.  What you need 
to do is determine the likely cause of overdispersion, if possible. The 
outcome will likely determine the type of count model that is most 
appropriate for your data. For one thing, if you find that the 
dispersion is close to 1.0, you have a Poisson model. It will explain 
why you have the problems with nbreg.

Remember, the Poisson model can be regarded as a negative binomial with 
an ancillary or heterogeneity parameter value of zero. Depending on the 
data, models having a heterogeneity parameter (alpha) value close to 
zero are also Poisson.  The problem is that if the value is 0, or 
numerically extremely close to zero, the NB model will not converge due 
to division by zero.  The NB model becomes Poisson as the value of 
alpha approaches zero, but it cannot actually be at zero.

Joseph Hilbe





-----------------------------

Date: Sun, 8 Jan 2012 15:44:01 +0100
From: Maarten Buis <maartenlbuis@gmail.com>
Subject: Re: st: What should I do if stata can not show the result of 
nbreg

On Sun, Jan 8, 2012 at 9:24 AM, Ye SUN wrote:
> My dependent variable is count number: the number of boards that a
> director serves in (this kind of data seems not fit to poisson)
>
> I first try poisson regression and get the results and then I try
> -estat gof- to test whether there is overdispersion problem, it is
> insignificant indicating no overdispersion problem I think.
> I also try -zip- and get similar results, and vuong test indicates
> there is no excess zero problems I think.
>
> I also want to try negative binomial regression,( nbreg and zinb ),
> but stata run for quite a long time and still run without showing any
> results. What does that mean?