On 28 Jan 2003 16:43:17 -0800, [EMAIL PROTECTED] (SR Millis) wrote: > I'm interested in doing an informal survey of preferred approaches to > analyzing pretest-posttest data. Patients have been randomized to either > a placebo group (coded 0) or medication group (coded 1). T1 represents > performance on a cognitive measure at baseline before placebo/drug. T2 > is the post-test performance on the same cognitive measure. > > One could perform an ANCOVA (with t1 as covariate), analyze change > scores (eg, paired t-test), use AUC, or use GEE. Which do you choose?
For paired t-tests, I see two paired t-tests; and that does not comprise a test of Group. The t-test of those simple changes is what you test as the interaction with GEE, isn't it? If AUC stands for "area under the Curve," I don't see the curve. After testing simple changes (AOV) or regressed changes (ANCOVA), the third alternative is to ignore the Pre (usually because there is almost no variation) and test the Post by itself. > Are there preferred alternatives? For randomized studies where variance has not changed, the ANCOVA is usually superior - in power, and in describing the hypothesis. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
