Re: st: Nested zip

[Stas Kolenikov <skolenik@gmail.com>]

On Tue, Dec 6, 2011 at 11:13 AM, Kris Anderson <andersok@reed.edu> wrote:
> I am trying to model a count variable (lifetime use of drugs) nested in
> schools. Given its distribution, a zip model would be ideal (lots of
> zeros). However, it seems Stata will run an xtmepoisson but not zip. Does
> anyone know why?

-zip- is essentially a mixture model. First, you have to model the
latent class of a given individual: whether that person is in the
identically-zero or in the non-trivial Poisson component. Second, you
have the -glm- model for the Poisson part. To generalize this to a
nested structure, you need to introduce the level-2 effects somewhere:
either in the mixture part, or in the Poisson part, or in both (in
which case, the effects may or don't have to be correlated). So there
are at least four possible random effect extensions of the -zip- model
(all nested in the most general model with correlated level-2 effects,
although the likelihood ratios have non-regular, non-chi2
distributions). Whether the fixed effect/conditional extensions of
-zip- are feasible, I don't know, but highly doubt it (given the
complexity of the model).

Would it make sense to model the ever-use of drugs via -xtmelogit- instead?

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   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/