I wouldn't think -gllamm- cares whether your data count or continuous.
But as soon as you start fitting something more complicated than plain
GLM in a multilevel context, you need to start thinking: is my
overdispersion described by a fixed effect or a random effect? Do I
need to model it as correlated with the term for the mean? Besides,
the standard approaches to overdispersion try to model extra zeroes
rather than outliers.
One thing you might be able to trick -gllamm- into is a mixture model
with say two components corresponding to "low" Poisson counts and
"high" Poisson counts. I've never had any luck figuring out how to do
this in -gllamm-, but it should be described either in the manual on
Berkeley page or in the Longitudional Modeling book. This is a tricky
problem as the model is not identified when there is only one
component, and the likelihood ratio between the models with and
without that dispersion is not a chi-square.
Finally, -xtmepoisson- will likely be quite a bit faster than
-gllamm-, so you might want to use it at least for starting values.
On 7/16/08, Johan Mesterton <[email protected]> wrote:
> Dear all,
>
> I'm analyzing several different outcomes using a multilevel model
> (GLLAMM) on my clustered data. One of the outcomes is count data with
> some outliers (that I do not want to remove) and when fitting the
> model using family(poisson) overdispersion is a problem. As
> family(nbinomial) cannot be used within the GLLAMM framework, can
> anyone guide me as to how I could solve this problem?
>
> The program does not accept me using the family(gamma) for the model -
> possibly because the program recognizes that it is count data.
>
> Best regards,
> Johan
--
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: Please do not reply to my Gmail address as I don't check
it regularly.
*
* 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/
