Count model selection

[Liang Wang Jack]

Dear scientists,

Code

countfit

shows us tests and fit statistics for poisson, negative binomial, zero-inflated poisson and zero-inflated negative binomial model. What's is the code that I can compare
1, poisson hurdle model with negative binomial hurdle model, 2, poisson hurdle model with poisson model, 3 poisson hurdle model with zero-inflated poisson model?

Thanks!

Liang

[Marcos Almeida]

To make comparisons between such models, AIC as well as BIC may be a good starting point.

[Phil Bromiley]

As an alternative or in addition to AIC/BIC, if you can do an encompassing general model, then you can test down from there. For example, there should be a way in zip to test the zero inflated portion. If you can't support the zip portion, then you're in (2) rather than (3). Similarly, negative binomial estimates often include what you need to test versus Poisson (although several commentators on this list have made strong cases for Poisson over NB even with over dispersion)..

Following the FAQ on asking questions will also help you get more useful answers - provide Stata code in code delimiters, readable Stata output, and sample data using dataex. The ability to do tests among models would also depend on what routines you use to estimate the models. For example, if you were doing this all in GSEM, then you'd have a different testing problem than if you used CMP or a set of different routines for each model.

[Liang Wang Jack]

To make comparisons between such models, AIC as well as BIC may be a good starting point.

Thank you Marcos, can we do Vuong test for Poisson hurdle model and standard hurdle model by Stata?

[Marcos Almeida]

You unfortunately didn't say much about the commands you used, as already remarked by Phil.

That said, you may check - postestimations - with regards to the command or the user-written program, if this is the case.

I have not used hurdle models so far. I gather the Voung test aims fundamentally to compare zero-inflated versus non zero-inflated models, hence hurdle models may not fit in this situation.

To end, shall the commands you've used be - tpoisson - or - tnbreg - , as far as I'm concerned, the postestimations point basically to AIC/BIC, as previously remarked.

Hopefully that helps.