[R-sig-ME] observation level random effects/kinship model

[Ryan King]

> Ryan, have you got some links for what you talk about ? I'm quite surprised
> that there are already packages for doing that because a lot of geneticists
> are using ASREML even if they have to pay for it. I really hope that there
> is a good reason for that !

Sorry, I missed this the first time. MCMCglmm allows an arbitrary
correlation matrix for random effects (ginverse options) and has a
built-in for numerator-relatedness-matrix given pedigree. AnimalINLA
also has a built-in for numerator-relatedness-matrix given pedigree.

If that parameterization is awkward/slow, you can also use the
decomposition trick in either.  That is, let K be your matrix, and U
%*% D %*%  t(U) its cholesky factorization. Then you can set the RE
design matrix Z = U %*% sqrt(D) with an identity covariance matrix; in
MCMCglmm that's idv(Z) and in inla a  f( ..., model="z") .

Both these packages rely for speed on Z and or COV(RE) or its inverse
being sparse, so I sometimes play with using the PMA package to
compute a sparse approximate SVD. Presumably an incomplete cholesky
factorization could do the same thing.

ASREML is probably worth the (NIH's) money; my understanding is that
it's fast, flexible, and robust. I don't know if the above have been
designed with very large datasets in mind.

Ryan King