Forum: developers
Monitor Forum | 
Start New Thread
| Design of mcmcsamp [ Reply ] By: Douglas Bates on 2008-01-07 14:28 | [forum:347] |
|
In previous versions of lme4 the mcmcsamp was tuned to the linear mixed model and updated both the fixed effects and the random effects simultaneously. For GLMMs or NLMMs it would be better to separate the updating of the fixed effects from the updating of the random effects. We already have the update_u function to determine the conditional modes of the random effects. For GLMMs and NLMMs update_u also calls update_L so the conditional modes and an approximation to the conditional variance-covariance matrix are available. These can be used in a Metropolist-Hastings update operation. A similar operation should work for the update of the fixed effects. An IRLS algorithm can be used to determine the conditional modes of the fixed effects. Updating the variance-covariance of the random effects can, I think, be done entirely on the u scale. If saveb is FALSE (and in most cases it is) then there will be no need to even create the random effects on the b scale. |
|
