Hi gurus!

On another list that I'm "listening" to, first a message came advocating the SAS-macro GLIMMIX and then another message came warning against some problems with the approach used (part of that message attached below). Now I wonder if there's someone out there, that understands these estimation approaches better than me (should be plenty :-), that could tell if this is really the case and then where the possible limits are for creating the problem?

TIA!

Ulf Emanuelson Unit of Epidemiology Swedish Univ. Agric. Sci.

----------original message------------------------------------------------------- It's been mentioned that there is a SAS macro for fitting random effects logistic regression models to longitudinal or clustered binary data. Unfortunately, it has been shown (e.g., Breslow and Lin, Biometrika, 1995) that the estimation approach that the macro uses (iterative fits of a weighted linear mixed effects model) is *biased* for small to moderate cluster sizes (e.g., numbers of observations per subject) or small to moderately sized variance components (i.e., within-cluster or -subject correlation).

---snip

Tom TenHave Center for Biostatistics and Epidemiology Hershey Medical Center Penn State University email ttenhave@biostats.hmc.psu.edu