Date: Thu, 1 Sep 2011 11:58:29 -0500
Reply-To: "firstname.lastname@example.org" <email@example.com>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Erin McMullen Jonaitis <jonaitis@WISC.EDU>
Subject: PROC MIXED: Correlation between fitted values & residuals
Content-type: text/plain; CHARSET=US-ASCII
With the help of Jason Schoeneberger & Bethany Bell's lovely macro Mixed_DX, I've been examining Level 1 residuals from some mixed models I've run, and have noticed a positive correlation between my fitted values and my residuals (ranging from 0.2 to 0.5 depending on the model). I'm looking for resources that can help me understand what's going on.
I have read the sections on diagnostics in a few introductory/applied regression/mixed-model texts I have lying around (Gelman & Hill; Fox; Singer & Willett; Cohen, Cohen, West, & Aiken) to no avail. I've done a few Google searches; the only practical suggestion I saw was to consider adding polynomial terms on some of my predictors, which I've tried, to no great effect.
The outcome is pretty close to normally distributed and the predictors are close to multivariate normal. I don't have serious problems with heterogeneity of variance. I have some outliers, but nothing really wild or influential. There's just this one hitch...
Advice/pointers to other resources much appreciated.
Erin Jonaitis, Ph.D.
Assistant Scientist, Wisconsin Alzheimer's Institute
7818 Big Sky Drive
Madison, WI 53719