Closed samueldnj closed 7 years ago
Okay, I was incorrect in using the mixed effects model from Pinheiro and Bates, at least in the form that I was using it, as I was jerry rigging an observation error model to use process error. After digging through Jim Thorson's mixed effects repo and looking at Russell Millar's MLE book I found that just adding the the pre-transformed independent REs to the likelihood appear to work fine.
Posterior modes for variance terms are around the right magnitude on well behaved data sets, so I think we can close this for now.
Problem
2 error components make estimating MLEs difficult. There's confounding all over the place
Possible solution
Treat like a multilevel linear mixed effects model
Progress:
I've made a first pass at this today (December 13) using the likelihood in Pinheiro and Bates Ch 2.4 (p64) but this doesn't seem to have improved things much. There was one run that had improved posteriors for each variance term, but then the observation model was whack, and I can't seem to reproduce.