Panel data and count dependent variable

[Andreas Psarras]

Hi all,
I have a panel data set of 18 years and 50 regions. My dependent variable is a count one, with over-dispersion. Hausman test indicates the use of fixed effects. Initially I thought that I should use Negative Binomial Fixed Effects (NBFE), but after reading some posts in the Forum I decided to use Poisson. So I use -xtpoisson [depvar] [indvar] , fe- command. My dependent variable is victims of car accidents. As I can understand I cannot use rates (for example victims per 1000 inhabits) because then I cannot use Poisson. However, all regions have different population. Is it right to have just the number of car accident victims? If not how can I engage victims and population of each region in analysis, in order to have proper and strong results?
Thank you in advance.

[Clyde Schechter]

This is precisely what the -exposure()- option of -poisson- and -xtpoisson- is for.

[Jeff Wooldridge]

Andreas: The key is that if you use a rate then the exponential functional form might not be the best fit. I would use the accident counts and then include log(population) as an explanatory variable in xtpoisson. If youuse the -exposure()- option then it forces the coefficient to be one. But I can imagine -- based on congestion arguments, for example -- that the elasticity of accidents with respect to population is different from one. In any case, it's very easy to estimate the elasticity and see how different the coefficient is from one. You might even test that (but I'm not going to suggest that because Clyde might be looking in ;-)).

You definitely should use the robust variance matrix to account for violations of the Poisson assumption and serial correlation.

xtset region_id year
gen lpop = log(pop)
xtpoisson accidents lpop x1 ... xK i.year, vce(robust)

[Andreas Psarras]

Thank you very much for your help. I will try both (lpop and exposure) and see the difference. I used a yeardummy for years (1,2,3...). I guess its quite the same with i.year.