pjmiller_57@YAHOO.COM wrote: > >Hello Everyone, > > I'm running some GEE models using PROC GENMOD. My criterion variables >consist of repeated lipid measures (cholesterol, triglycrides, total >cholesterol to hdl ratio) from a sample of HIV patients. Values for each >criterion are coded as 0 for normal and 1 for elevated. Lipids in HIV >patients are often affected by the antiretroviral drugs they are taking. So >I'm using a variety of arvs to predict whether or not patients had elevated >lipids. > > At this point, I'm wondering how I should specify the "working >correlation matrix." I know that "exchangeable" is the most common >structure, but I'm finding that my models sometimes crash ("Error in >computing the variance function") when it is used. In contrast, the models >seem to be better behaved when I use an "autoregressive" structure. Also, I >was thinking that this structure might make more sense for repeated >observtion within subjects as opposed to data where subjects are nested >within groups. Is my thinking right here? Or apriori would people have been >inclined to recommend the more commonly used exchangeable structure? > > Thanks, > > Paul >

I like your thinking. The "exchangeable" covariance structure was popular when it was the only thing that stat packages could compute. That's not true now.

I usually recommend that people look at the entire covariance matrix and see if the structure suggested by the situation makes sense. AR(1) or ARH(1) may best reflect the behavior of the data for your process. It's pretty hard to tell from way over here. But assuming any one covariance structure before looking at the data is bound to get you in trouble eventually... :-)

HTH, David -- David L. Cassell mathematical statistician Design Pathways 3115 NW Norwood Pl. Corvallis OR 97330

_________________________________________________________________ Puzzles, trivia teasers, word scrambles and more. Play for your chance to win! http://club.live.com/home.aspx?icid=CLUB_hotmailtextlink