Using xtnbreg fe with heteroskedasticity

[Chris Remijn]

Dear Statalisters,

I am currently working on my Master's thesis and trying to finalize my paper using Stata 16. I have panel data with around 1000 firms for 15 years. My dependent variable is a count variable (between 0 and 850) and is highly over-dispersed (mean = 28.24978, variance = 2699.362). Therefore, I use a negative binomial regression with fixed effects, as advised by my supervisor, although not fully recommended by some Statalisters. However, as heteroskedasticity is present, I would like to use the vce(robust) option, which is not an option for xtnbreg. Does anyone have any advice on how to deal with this issue? Except for US_ceo and year_num, all variables are continuous. My code is as follows:

xtnbreg ten_k_total_counts std_ceo_age std_ceo_formal_education std_ceo_fin_position ceo_age_x_dynamism ceo_formal_education_x_dynamism ceo_fin_position_x_dynamism std_employees_num std_return_on_assets std_tobinsq US_ceo i.year_num, fe

As you can see, I have also used the standardized values of my IVs and control variables, as advised by my supervisor. However, I do not really get why this is necessary for interpreting my interaction terms. Any advice will be greatly appreciated.

Best,

Chris

[Joao Santos Silva]

Dear Chris Remijn,

I am one of the "Statalisters" who advises against the use of the xtnbreg with FE, so may main advice is that you switch to Poisson regression. Now, xtnbreg already accounts for some form of heteroskedasticity and may not be valid if other form of heteroskedasticity is present; I guess that is why no robust option is offered and this highlights the drawbacks of this estimator.

Best wishes,

Joao

[Chris Remijn]

Dear Joao Santos Silva

Thanks for your reply! As I am totally new to using count models and am not capable of making sense of Jeffrey Wooldridge's paper, would you be so kind to briefly tell me why (although the estimated g.o.f. and likelihood ratio test assume otherwise) xtpoisson if preferred over xtnbreg? I have already compared xtnbreg against the xtpoisson command with fe vce(robust) and almost all significant results disappeared. Moreover, what are your thought on using standardized values?

Thanks again for your help!

Best,

Chris

[Joao Santos Silva]

Dear Chris Remijn,

Simply put, the Poisson FE estimator is consistent under very mild conditions, whereas the NB estimator is only consistent under very restrictive assumptions that are unlikely to be satisfied. Moreover, the NB fixed effects model is not a fixed effects model in the usual sense.
On the standardized variables, I am not sure they help, but I do not think they hurt, so you might as well use them.

Best wishes,

Joao