Bookmark and Share

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

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

Re: st: Convergence Problems With Zero Truncated Negative Binomial Regression

From   Muhammad Anees <[email protected]>
To   [email protected]
Subject   Re: st: Convergence Problems With Zero Truncated Negative Binomial Regression
Date   Sat, 3 Mar 2012 16:18:23 +0500

Yes, I would further to suggest you how to detect dispersion in count
data models. In one discussion, Joseph Hilbe email me the following
suggestion. I owe to Jo Hilbe for his kind guidance in this regard. I
think it might be helpful to you while the same has been discussed on
this list earlier.

glm panelcount x1 x2 x3, fam(poi)

glm panelcount x1 x2 x3, fam(poi) cluster(panelvar)

glm panelcount x1 x2 x3, fam(nb ml)

glm panelcount x1 x2 x3, fam(nb ml) cluster(panelvar)

where panelvar is your panel variable, eg id and x1-x3 are predictors.
You'll have your own of course, these are just fillers.

Check the Pearson dispersion statistic for the Poisson model. If it is
under 1.0 the model is Poisson underdispersed. If it is, you cannot
use a negative binomial program. If it is over 1, then try the negati
model, checking the same dispersion statistic. See if it is under or
overdispersed. Then apply the cluster option as i show above to the
model. See if the standard errors change much. If not, then there is
not underdispersion.

Following the above steps would let you know the nature of what you
need to do with your modelling.

Hope this helps.

On Sat, Mar 3, 2012 at 4:26 AM, Michael Lebenbaum <[email protected]> wrote:
> Hi,
> I am attempting to do an analysis of zero truncated counts for two outcomes
> (both are zero truncated count distributions). So far I have been using
> Stata 10, with ztp and ztnb commands with svy. The zero truncated poisson
> model converges properly for both, however the zero truncated negative
> binomial model does not converge (It gets stuck at the fitting a constant
> only model (not concave)). The sample size is large 10K+, so I do not
> believe this is the problem. I think it has to do with sparse data in parts
> of the distribution as the vast majority (~95%+) of the data is at the low
> end of the range with much of the range having 0 or few observations
> (~80-90% of the range). In addition, when I delete a small percentage of the
> top users (2%), the zero-truncated negative binomial model converges. Does
> anyone know of alternative solutions? I would rather not have to alter the
> sample if it is not necessary. Thank you.
> M


Muhammad Anees
Assistant Professor/Programme Coordinator
COMSATS Institute of Information Technology
Attock 43600, Pakistan

*   For searches and help try:

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