AR-k correlation structure "xtgee" - (*) All panels must have at least k+1 obs.

[Francis Clark]

Hello everyone,

I have a question regarding fitting an AR-k correlation structure to longitudinal data. I understand the caveat that "(*) All panels must have at least k+1 obs." In this case, if you set AR-k, all participants with a single visit (only) are dropped from the model. However, I am not sure of the rationale for doing this... given that models do not drop the baseline visit (if it's the only visit that exists for that participant) when using an exchangeable correlation structure, for example.

Could anyone explain why Stata does not allow you to fit AR-k on all participants - including those with a single visit? I realize that there is no correlation within a participant if they only have one study visit, but this is also the case for other correlation structures as well (e.g., exchangeable) - and the same restriction is not applied.

Thanks!

[Francis Clark]

Anyone have any insights into the statistical/substantive reasoning behind the xtgee package dropping single unit clusters (i.e., participants with a single study visit) when using AR-k? Again, I do not see why AR-k would require that "all panels must have at least k+1 obs" when other correlation structures (such as exchangeable) do not have such a requirement. My longitudinal data would follow a AR-1 correlation structure... but I can't use this if the models drop all of my participants with a single study visit/single row of observations.

[Francis Clark]

Any insights?

[Francis Clark]

One more bump - for good measure! A huge thanks to anyone who may have an answer to the above question.