[Joao Santos Silva]

Dear Lida Metallinou,

Please see

Wooldridge, Jeffrey, (1999), Distribution-free estimation of some nonlinear panel data models, Journal of Econometrics, 90, issue 1, p. 77-97.

Best wishes,

Joao
PS: Actually, with Poisson regression with FE you generally cannot compute probabilities.

[Lida Metallinou]

Dear Professor Silva,

Thank you very much for your quick reply,

Kind Regards,
Lida

[Lida Metallinou]

Hello everyone,

Apologies for coming back on the same topic but I have some more questions.

Dear Professor Joao Santos Silva, thank you for answering all my previous questions.

To remind my problem and nature of data, I am dealing I have a three-dimensional panel data set structure model where the dependent variable is in country, year, industry form and has a lot of zeros. My main explanatory variables vary by industry, by year, by country the one and the other by industry, by year.

I have a read a lot of STATAlist posts and everyone seems to suggest the FE poisson estimator(Wooldridge, 1999a)- xtpoisson with FE - is the best approach when there the proportion of zeros is large and when you want to account for heterogeneity. However, because I need to write my analysis I am still a bit confused. From reading a lot of STATA lists posts,I understood that is better not to use zero-inflated models. However, is there any reference or any other explanation that I could mention in order to understand why zero-inflated models are not good in my context since I have a lot of zeros?I am thinking to present that model as well for comparison reasons.

Also, I understood that overdispersion is not an issue under the xtpoisson with fe-approach. Is there any other reference that I could use for that argument.
At the same time, I found that -xtnbreg with fe- is not so much preferred (Alison & Waterman, 2002). Is there any other reference that I could use of why -xtnbreg with fe- is not such a good approach as Poisson with FE?

Any advice on reference would be highly appreciated.

Thank you very much,

Kind regards,
Lida

[Joao Santos Silva]

Dear Lida Metallinou,

Let me see if I can answer all the questions:

1 - Zero inflated models are interesting when we have two types of zero; in particular we need that for some observations y is always equal to zero for any value of the regressors. If you think that is the case you have, it is worth trying such a model. Note however, that you cannot gauge whether a zero-inflated model is needed just from the proportion of zeros.

2 - The reference you need about overdispersion in the FE Poisson regression is Wooldridge, 1999.

3 - About the NB with FE, see also

Guimarães, Paulo, 2008. "The fixed effects negative binomial model revisited," Economics Letters, Elsevier, vol. 99(1), pages 63-66, April.

Best wishes,

Joao

[Lida Metallinou]

Dear Professor Joao,

Thank you so much for all your answers to your questions. I believe that I understood what you mean about zero-inflation.

One last thing that I would like to mention, from reading a lot of posts, it seems that zero-inflation in the panel data structure is not used.And I thought that this is why it's not the preferred method to go. So, I understood that if I want to test the zero-inflation models just for comparison reasons, I need to ignore the panel data set structure of my data, which I presume it's not the best(so, I can't xtset the data before running the regression).

However, I tried to add manually the dummies to control for FE, which basically doesn't mean that ignore the panel structure.

Thank you very much for your help,
Kind regards,
Lida