Bioequivalence and Bioavailability Forum

Hi MaxPerry,

❝ ElMaestro: This certainly seems rational. On the other hand, consider a simple ANOVA with multiple measurements of a particular feature on each subject in a group. The test of significance between subjects has to rely upon a within-subject error estimate.

To me this certainly sounds like a mixed model and thus not a simple ANOVA because the feature response is correlated within subjects. Proc Mixed, not Glm.

❝ But the thread ultimately seems to settle on the explanation that the important distinction was whether subject effects were considered fixed or random. I'll read about this in a book and try to get my head around it.

Uh oh - the Bogus Statement Warning Light is now flashing ("BSWL" for future reference).
In GLM subjects are treated as fixed; SAS then has defined the bogus statement (search the forum to dig deeper into this term) which pretends to involve setting subjects as random, but the actual fit is still done with subjects as fixed (yeah, it has to be so because it is a GLM).
My proposed bottom line: In a GLM (that is to say, in a 2,2,2-BE study) subjects are fixed; post-fit bogus gymnastics ascertains p-value with the between-subject variability in the F-test denominator for Seq. In a study with replication, within-subject responses to the same treatment are (can be) correlated and therefore a mixed model is used.