Jeff writes,
Can you suggest some (simple) material on the differences between ML MLM and Bayesian MLM? Since we are using LMER, and not winbugs etc., then Justin and I are NOT doing Bayesian MLM?
My reply: The only real difference is when the unexplained group-level variance is hard to estimate, which typically occurs when this variance is low and the number of groups is small (in which case, classical point estimates won’t work well, and even a small amount of prior information will help; see Sections 5.1 and 5.2 of this paper). I think we discuss this issue in the new book, also see this from 1996, for example.
If I recall correctly from your class this spring, another difference is that Bayesian MLM has biased estimates of its parameters. Of course, this bias buys you extra precision and the ability to integrate prior information. Is this correct?
We talk extensively about comparing the two in our paper on multilevel estimators for TSCS data.
http://pan.oxfordjournals.org/cgi/content/abstrac…