Re: st: Mixed effects model with zero-inflated negative binomial outcome for repeated measures data

["JVerkuilen (Gmail)" <jvverkuilen@gmail.com>]

On Mon, Jan 14, 2013 at 5:18 AM, Robert Long <W.R.Long@leeds.ac.uk> wrote:
> Thanks again ! I have Skrondal. & Rabe-Hesketh in the office so I will check it today.
>
> As for which parameter estimates are needed, I don't /need/ the NG dispersion parameter and the zero inflation parameter - only as a means for fitting ZINB as the outcome, but as far as I understand they must be estimated in order to fit ZINB. As for clustering, the data are repeated measures on around 100 subjects. The random intercept and slope variance are not of primary interest in themselves - interest is centered on the interactions between time and 2 covariates - but of course it would not be right to ignore the clustering by subject. I think it is possible to leave out the random slopes, but not the random intercepts.>

Well yes, you would need to estimate them to estimate the ZINB, but
you could also estimate a mixed ZIP, which is a boundary case of ZINB,
or the mixed NB. These parameters are likely to trade off in odd ways
with a very complex model.

One of the reasons to fit the NB is to deal with unmodeled (or
unmodelable) overdispersion created by clustering, so modeling the
clustering may make the NB parameter drop.



> For completeness, here is the model current specification again in R
>
> glmmadmb(Y ~ time*X1 + time*X2 + (time | Subject),
>         data=final,family="nbinom2", zeroInflation=TRUE)

Does the ADMB estimate seem to converge well? If so that's a sign that
the model would work (and maybe you should just use them), but it
sounds like gllamm can't do what you want easily.

One alternative would be to fit the ZIP model and use robust standard errors.

*
*   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/