Multilevel Zero inflated negative binomial regression?

[Danielle Rohret]

Hi All, I am running a negative binomial regression within a multilevel framework using the meglm command. The dependent variable is a count variable with an exposure variable specified. Is there a way to specify this to be a ZERO-INLFATED negative binomial and still maintain the multilevel framework? And, yes I suspect that the excess zero's are generated from a different underlying process. In other words, I have theoretical justification for wanting to do this. I have tried to work-in zinb and related options, but doesn't seem to work inside meglm (mixed-effects general linearized model). Thanks for any advice or suggestions, maybe there is an add-on for this?

[Tom Scott]

It looks like it did not exist in 2013 and based on a scan it does not look like it exists today: https://www.stata.com/statalist/arch.../msg01407.html

[Joseph Coveney]

You can find code here to fit such a model using SAS's PROC NLMIXED, which you can then convert into corresponding Stata menl code.

For the reasons given by Jay Verkuilen in the post that Tom linked to, I'm guessing that fitting such a model is a nightmare, and would be uncomfortable trusting coefficients without extensive exploration of the likelihood, which is liable to be fairly time-consuming.

[Joseph Coveney]

Oh, forgot to mention: have you considered a marginal model as a fallback option, such as what you'd get with, say, the following?

zinb <varlist>, vce(cluster <varname>)

[Danielle Rohret]

Thanks for this information. I think I will look into these suggestions.