I think what we are seeking is the marginal variance-covariance matrix of the parameter estimators (marginal with respect to the random effects random variable, B), which would have the form of the inverse of the crossproduct of a $(q+p)$ by $p$ matrix composed of the vertical concatenation of $-L{-1}RZXRX{-1}$ and $RX{-1}$. (Note: You do not want to calculate the first term by inverting $L$, use solve(L, RZX, system = "L")
[...] don't even think about using solve(L)
don't!, don't!, don't! - have I made myself clear?
don't do that (and we all know that someone will do exactly that for a very large $L$ and then send out messages about "R is SOOOOO SLOOOOW!!!!" :-) )
--- Douglas Bates R-SIG-Mixed-Models (March 2010)
不太会从邮件列表里搜这段话的出处?求助