Mixed Models with R/lme4/mcmcsamp
Problems with mcmcsamp:
mcmcsamp is broken
Date: 2007-11-09 22:37 Priority: 3 State: Open Submitted by: Assigned to: Douglas Bates (dmbates) Summary: mcmcsamp is broken
Detailed description I think I must have missed an inverse in the update of the variance components when switching from the Omega representation to the ST representation. The current chains are nonsense. FollowupsAttachmentsCommitsChanges Message Date: 2008-06-23 14:39 Sender: Douglas Bates
Starting with release 0.999375-16 the mcmcsamp function has switched to a Gaussian update of the redundant (i.e. possibly negative) representation of variance components. This gives the desired behavior of not getting stuck near zero but the results for the variance components seem to be systematically too small. Also sampling of covariance components needs to be added (the MCMC_T function in lme4/pkg/src/lmer.c). Date: 2008-04-04 17:56 Sender: Douglas Bates
The mcmcsamp method does work for LMMs in version 0.999375-10 and later but can still get stuck when values of a variance component approach zero. Redundant parameters (a la Gelman et all, JCGS, 2008) may help - however I am still concerned about whether they make sense because the posterior density is improper in those cases.