Generalized Poisson with Fixed Effects

[Debbie Watkins]

Hello all,
I am trying to run a poisson regression based on the the skew of the dependent variable. However, this variable also seems to exhibit under-dispersion so I am considering running a generalized poisson estimation. My concern is that my model will require a number of fixed effects and I am worried about the incidental parameters problem. I know that xtpoisson , fe would not be subject to this problem. And I was wondering if gnpoisson will subject to incidental parameters problem in the presence of a lot of dummies and if so if there is a work around? Otherwise, would it be more appropriate to run xtpoisson, fe in place of gnpoisson despite the under-dispersion of the underlying data?

Thank you very much in advance.

Best,
Debbie

[Carlo Lazzaro]

Debbie:
welcome to the list.
Please post what you typed and what Stata gave you back (as per FAQ). Thanks.
As an aside, please note that -xtpoisson,fe- will provide you with conditional fixed effects.
Eventually, these threads migh be useful:
http://www.stata.com/statalist/archi.../msg00488.html
http://www.stata.com/statalist/archi.../msg00844.html

[Joao Santos Silva]

Dear Debbie,

In addition to Carlo's helpful advice, I would had that I am not aware of any method to deal with the incidental parameter problem in the context of the generalized Poisson. Unless you want to compute probabilities of events, your best bet is to go with a fixed-effects Poisson regression.

Best wishes,

Joao

[Debbie Watkins]

Dear Carlo and Joao,

Thank you both so much for your very quick and helpful responses. Carlo - I haven't yet begun entering this regression in Stata so I don't yet have anything to post. I am still at the beginning stage where I am thinking more generally about which estimation might make more sense given the features of my data. But for future questions I will make sure to post the relevant code. Based on both of your helpful advice I will begin with xtpoisson, fe. It seems that avoiding the incidental parameters problem may be more important than applying a specialized poisson for underdispersned data.

Just as an aside, if the problem had been over-dispersion instead of under-dispersion of the data in the presence of many fixed effects ( xtnbreg , fe versus xtpoisson, fe). What would you recommend? My understanding is that xtnbreg, fe does not fully fix the incidental parameters problem. Would your recommendation still be to opt for xtpoisson, fe over xtnbreg , fe as long as I don't need to compute probabilities of events?

Thank you again. I really should have signed up for this forum sooner.

Best,
Debbie

[Joao Santos Silva]

Dear Debbie,

Indeed, I would still go for Poisson with FE :-)

Best wishes,

Joao

[Debbie Watkins]

Thank you!

Best,
Debbie