PROC GENMOD with cluster and modified poisson -what working correlation structure should I choose

[lis]

I am working on a cross-sectional study and I will be using the modified poisson regression analysis. My data has a cluster variable.

Thus, I am running the code showed below. However, I do not know what type of working correlation should I use.  Independent is the default but not sure how to decide this.  Any help would be appreciate it.  Thanks.

proc genmod data=data
class cluster
model y = x1 x2/dist=binomial link=log
repeated subject=cluster
run;

[Ksharp]

usually TYPE=UN . But I am not sure it is the best covariance structure.

[StatDave]

The unstructured matrix (TYPE=UN) is the most general, allowing all pairs in the clusters to have separate correlations.  However, it is also the most complex since it has the most correlations to estimate.  As a result, using this structure often results in model fitting problems, particularly as the clusters get large. The QIC statistics can help to determine an appropriate structure as discussed in this note. But remember that the GEE method implemented by the REPEATED statement is robust to incorrect specification of the structure, so it is common to use fairly simple structures such as the independence (TYPE=IND) or exchangable (TYPE=EXCH) structures. 

[lis]

Thank you very much for the information. This is very useful.

##- Please type your reply above this line. Simple formatting, no
attachments. -##