Exchangeable or Unstructured GEE

#1
Hi all,

I was wondering if someone can give me a clear understanding on the difference between exchangeable working correlation matrix and unstructured working correlation matrix and when is best to use one or the other?

Many thanks!
 
#2
Hi all,

I was wondering if someone can give me a clear understanding on the difference between exchangeable working correlation matrix and unstructured working correlation matrix and when is best to use one or the other?

Many thanks!
I think it depends largely on your research question. I'm in public health and when we're looking at patients clustered within physicians, we typically want to account for this dependency. In this case, we would use an exchangeable correlation structure because: (1) there will be several clusters (number of doctors) (2) there's no substantial reason to think that the correlation of patients among a given physician is different for another physicians patients. Again, we will have several clusters here and GEE averages over all of them.

I would use the unstructured approach if I believed there are genuine differences in the way the sub units are correlated between clusters. For example, If you might do this in an analysis that involves looking at 'site' or 'hospital location' because the relationship among patients and quality of care is probably different at each hospital. In this example, hospital would only have a few levels (few distinct hospitals) and I don't think the unstructured approach is unreasonable.

Cheers,