Not understood [🇷 for BE/BA]

Thanks PharmCat,

funny, I read the ref. you mention (Lindstrom & Bates) yesterday and I did not understand much of it. Will try again.

❝ Problem1: when you have big matrix invercing is slow as hell, but if you inverce by blocks (you can do it because it block-diagonal) it can be done faster.

Sounds neat, will check if I can somehow do such a thing.

❝ Problem2: In BE case you have many equal Zi matrices, so in logREML calculation (if you calc by subject) you inverce one matrix many times (because V = ZGZ'+R, Z can be different, G and R not changing subject by subject). If you cache result - you can seriously increase performance.

My approach is very simple: I make a case for saying that when T is not replicated while R is replicated, then it is not possible to make a meaningful model when V=ZGZt+R (as long as we count on R having the withins and G having the betweens). If we insest on ZGZt+R then we get irrelevant variance components associated with T whichever way we parameterise the covariance matrix. Therefore, I am working in V directly, meaning I am writing the blocks into it directly.
This may still be compatible with your proposal, though.

Pass or fail!
ElMaestro