Fixed Effects Poisson or Negative Binomial?

[John Robinson]

Hi,

I am trying to choose between a Fixed-Effects Poisson regression model and a Fixed-Effects Negative Binominal regression but my results are quite different under both models and I am not sure which to choose.

I am trying to estimate how the number of supermarkets in a country is affected by a particular policy (REW), across Europe, and have count data which is not affected by zero values (min = 20 and max = 6600) and so considered the FE Poisson model (which I believe Professor Wooldridge recommends for such a situation). Under such a model (with robust SE) I get a coefficient of 0.401 which is highly statistically significant (p=0.000) [using bootstrap SE p=0.065] but if I use a FE Negative Binomial model the coefficient is lower, at 0.127, and highly insignificant (p=0.508). How do I choose between these two models? Is the FE Poisson appropriate here?

My understanding is that so long as I use vce(robust) overdispersion doesn't really matter in the Poisson model, although because I am using fixed-effects perhaps I should bootstrap my standard errors to deal with clustering?
I also read somewhere about the NBreg model only being conditional FE and so it doesn't properly deal with fixed effects in the same manner that xtreg, fe deals with fixed effects?

Thanks in advance for any help.
John

[Joao Santos Silva]

Dear John,

Indeed, the NBreg is not a proper FE model and it is not robust, so just go for Poisson FE with clustered standard errors; no need to bootstrap.

Best wishes,

Joao

[John Robinson]

Dear Joao,

Many thanks for your prompt response and your help in this matter, it is much appreciated!

Best,
John

[Felix Bittmann]

Sorry for my ignorance, but why is xtnbreg, fe not a proper FE model, say, in comparison to xtpoisson, fe?

[Joao Santos Silva]

Dear Felix Bittmann,

If you search this forum, you will find discussions and references on that. See, e.g., here.

Best wishes,

Joao