lmer and semi-definite covariance matrices for randomeffects
On Wed, 3 Apr 2013, Gabriel Baud-Bovy wrote:
Thank you for your suggestion. Still, I am wondering whether lmer could fit the model in principle and, if not, why. Here is my current understanding:
[SNIP]
What I am missing is a clear statement of why a covariance structure of the form sigma*CorA where CorA is semi-definite positive will not work (if it does not). Which step in the computation of the loglikelihood by lmer cannot be done in this case.
AIUI, the details of the algorithms tend to shift around in the development versions. My very simple minded understanding of these matters is that one encounters these kinds of problems when one preinverts the component matrices for efficiency (so CorA is SPD, but Va*CorA+Ve*I isn't). Cheers, David Duffy.