Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


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

Re: st: How to deal with the quasi-complete separation problem in the logit part of ZINB analysis


From   Nick Cox <njcoxstata@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: How to deal with the quasi-complete separation problem in the logit part of ZINB analysis
Date   Sun, 30 Sep 2012 15:42:03 +0100

That is a Jeffreys prior, or a Jeffreys' prior, as it was the idea of
Sir Harold Jeffreys.

In a letter to me in 1976 he wrote that he preferred the form
Jeffreys's, but this is rarely seen.

Nick

On Sun, Sep 30, 2012 at 3:14 PM, JVerkuilen (Gmail)
<jvverkuilen@gmail.com> wrote:
> On Sun, Sep 30, 2012 at 2:18 AM,  <smztsmzt@163.com> wrote:
>>
>> I encountered the quasi-complete separation problem in my data when I doing the ZINB analysis in Stata.
>>
>> I have found and installed the firthlogit ADO file which is very useful to deal with the separation effect by using the PML method, but how could I use the same way in the ZINB analysis? Does any one know that there might be some similar package, command, parameters or ADO files that
>> could take the same effect in the ZINB analysis?
>
> In my experience the ZINB is quite challenging to fit. You have both
> negative binomial overdispersion and excess zeros. I'd try fitting a
> ZIP or simplify the ZI component of the model, assuming you are using
> some predictors for it.
>
> I'm sure with some programming the ideas in -firthlogit- could be
> extended to other models. Essentially from the math it looks like an
> approximation to a Jeffrey's prior in a fully Bayesian analysis.
> (Update: Went and looked at the paper by Firth and that's exactly what
> it is!) Other priors could similarly help. Often you can use a
> "pseudo-data" approach and add a few fake data cases to approximate a
> prior. Any recommendations beyond that would require some information
> about the dataset. If all your predictors are discrete you can often
> simply add a few observations to the cells created by the table. It
> might be to add a few cases with average covariate values and use MI
> to predict observations.
>
> Note that all these *tricks* are just that, tricks or devices that are
> not well-founded theoretically.
>
> Jay
>
> *
> *   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index