Survey data: Make a decision between Poisson regression and Negative binomial regression

[Piccolo Wang]

Dear Stata users,

I hold a question here: I am using survey data and deciding which regression model to use for a count variable. Right now I narrow the options to Poisson regression and Negative binomial regression. However, I don't know the Stata code to test the equality/inequality of mean and variance of the observed counts. Can anyone let me the code? Thanks a lot!

Piccolo

[Carlo Lazzaro]

Piccolo:
see -estat gof- in -help poisson postestimation-.
Almost all the Poisson regressions are overdispersed (ie, variance/mean>1); the usual fit is to go -nbreg-, which -alpha- parameter allows for modelling the overdispersion detected under -poisson-.

[daniel klein]

Carlo is probably right about overdispersion. However, it seems that this assumption is only relevant for the standard errors. Thus, using robust standard errors in a poisson model could be an alternative.

Best
Daniel

[Carlo Lazzaro]

Daniel's helpful reply make me think about another issue: overdispersion can be apparent or real (and these features require different methods for dealing with overdispersion).
The deeply missed Joe Hilbe's textbook https://www.stata.com/bookstore/modeling-count-data/ is, in my opinion, one of the best in the field of count data regression models.