Wildlife biologist Wayne Hallstrom writes,

I [Wayne] have been running some mixed models and am trying to determine whether the mcmcsamp(R) mean and CIs are more appropriate for presentation of results versus traditional fixef(R) coefficient estimates. Trying to figure this out has led me to ask this question of you, because you are recommended as one of the main people in the Bayesian statistics world who is familiar with this newer idea. Any advice or recommended reading you could suggest would be greatly appreciated.

My short answer is that if you’re using mcmcsamp, you should be using mcsamp, which is a wrapper for mcmcsamp that’s in our “arm” package in R. My longer answer is that se’s from lmer() should be a bit too small, but probably no big deal if the number of groups is large (so that the group-level variance components are estimated accurately).

Wayne then asks:

What do consider to be a large number of groups? There are 8 groups with 2 subgroups in some of those 8, leading to a total of 13 subgroups in this particular dataset.

I think 13 typically is a lot–there must be some more formal calculation based on estimated variance components and chi-squared distributions…

P.S. Just to be clear: mcsamp() is in the “arm” package in R, and mcmcsamp() (which is called by mcsamp()) is in the “lme4” package. If you install and load the “arm” package, it’s all there.