I am trying to fit a GEE Poisson model on a panel dataset consisting of T=360 and N=304 for a total of >108,000 observations in Stata. My response variable measures a count of people imprisoned, and I am interested in the effect of three dummy variables compared to a baseline scenario. I am new to this class of models so pardon me if the questions sounds dumb.

To exemplify, the syntax used is the following:

xtgee y i.policy n_jobs n_jailed, family(poisson) link(log) corr(exchangeable)`

My question is the following: I am achieving significant results only when using an exchangeable or independent correlation structure, while when choosing other (as stationary or AR(k)), results for my variables of interest suddenly became largely non significant. Even when I use robust standard errors through the


command, either results for i.policy are non significant or the model does not achieve convergence. Is there anyone that can help me in understanding why this is happening? Should I assume that the fact the results are significant in the exchangeable scenario mean that the models is correctly estimated and non-biased?


You should not use statistical significance of predictors as an indicator of whether the GEE is correctly specified. To select the appropriate working correlation structure for your data, you could use the QIC criterion. In R this is available in the geepack package.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.