I also have some *real* experience with the ASREML GLMM. I even did a small
test on a data set in which I only had progeny data and did not include any
relationships between sires. Then, animal model and a sire&dam model should
give the exact same answers (after fidling around with some quarters,
halves, etc.). Without the GLMM link, it indeed did. However, with a GLMM,
answers came out differently! This supports Kim's concerns. I also had
larger trust in the sire&dam model, and use that now (or a sire & mgs
dependent on the data), and also apply it to "complete pedigree analyses"
(which becomes an approximation, but not a bad one, if parents have records
too). Of course, a "reduced animal model" would be most perfect, but the
sire- and dam model is easily fitted in ASREML from an ordinary pedigree
file and by "overlaying" dam with sire.
Also with Gibbs sampling, the combination threshold model and animal model
seems to be "unstable", to put it mildly (it "blows up"). I have heard
people say that the combination also is theoretically not valid (one quote I
remember was from Wang at WCGALP Armidale). Here also sire & dam models (or
reduced animal models) appear the solution.
Dr. L.L.G. (Luc) Janss, Scientist Quantitative Genetics
ID-Lelystad "Institute for Animal Science and Health"
P.O. Box 65 Tel. +31.3220.127.116.11
8200 AB Lelystad, The Netherlands Fax. +31.318.104.22.168
Asreml mailinglist archive: http://www.chiswick.anprod.csiro.au/lists/asreml