I'm using the geepack package in R to fit a variety of models to repeated measures data, albeit with varying cluster sizes. However, for some correlation structures ("exchangeable", for instance), I obtain unreasonably large estimates for my parameters, usually on the order of $10^{13}$ to $10^{16}$, with standard errors a magnitude or two smaller. Specifying the "independence" correlation structure produces, as you'd expect, more reasonable estimates which are also in line with previous models, including ones where an "exchangeable" correlation structure was specified. These data that're giving me trouble come from the same sources; in all respects, they're the same data. What gives? :-(

  • $\begingroup$ Can you tell us more about the data? Is it plausible that the effects are legitimately that large? If not-- in some situations (e.g. some kinds of case-control data) non-independence GEEs give inconsistent parameter estimates. Also, is it possible that complete/semi-complete separation (e.g. see the answers to this question stats.stackexchange.com/questions/27442/…) is an issue? This produces unstable regression coefficient estimates $\endgroup$ – Macro May 2 '12 at 1:43
  • $\begingroup$ It sounds like the sort of problem illustrated here; with an exchangeable working correlation, it's possible, albeit rare, that the mean model estimates and correlation matrix estimates are not compatible. NR Chaganty and co-authors have published many papers on this topic. $\endgroup$ – guest May 2 '12 at 4:49

Your Answer

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

Browse other questions tagged or ask your own question.