RE -xtgee- Equivalence with GLM?

[Sally Dee]

Hi!

I was looking around for a panel version of GLM to run a log-gamma model on. From what I read in Cameron & Trivedi - although in linear models a pooled estimator is consistent but inefficient compared to a RE estimator, this is not the case for nonlinear models. In fact, pooled estimators will be inconsistent in nonlinear models. I have a panel model with a specification that includes time dummies that I would like to estimate in -xtgee-, but I'm confused as to whether time dummies fits with the population average interpretation of the GEE. Could I just get away with using GLM? Any help would be greatly appreciated.

[Clyde Schechter]

If I understand what you want, see -help meglm-.

[Joseph Coveney]

From what I read in Cameron & Trivedi . . . pooled estimators will be inconsistent in nonlinear models.

Could you elaborate on that?

That differs from my understanding of GEE models, if that's what you mean by "pooled estimators".

Are you sure that they weren't referring to the incidental parameters problem (panel variable in the model as indicator variables)?

[Sally Dee]

Hi! In Cameron & Trivedi's 'Microeconometrics Using Stata' on p.603 - it says "The RE model has a different conditional mean than that for pooled and PA models, unless the random individual effects are additive or multiplicative. So, unlike the linear case, pooled estimation in nonlinear models leads to inconsistent parameter estimates if the assumed RE model is appropriate and vice versa'. And Clyde - thank you for the recommendation for -meglm- however I ran into a similar problem that I've seen in other forum posts that -meglm- will not model zero's in the way 'glm' will.

[Sally Dee]

When I say 'pooled estimators' - I may be using that incorrectly but I've generally seen GLM models used with cross-sectional data. Thus, I had rationalised that trying to force panel data into a GLM using (T-1) time dummies would be a 'pooled' process.