Perhaps I misunderstand Rubin's missing data theory, and/or perhaps
its not relevant to Thierry's problem.
I was under the impression that if the probability of missingness
depends on the value observed for some other data (MAR), then by
including this data and structuring the likelihood correctly then
correct inferences (i.e. in the absence of missingness) could be made.
Given that the default na.action of lmer seems to deletes other data
(complete case analysis), it is hard to see how the other data can be
used to 'correct' for missingness. MCMCglmm uses augmentation for
missing data. Internally, this is often used just to simplify/speed up
the matrix operations using dummy data. ?However, I had presumed that
if users really did have MAR data then the augmentation would take
care of this. I know ASReml has an na.includeY argument so presumably
there is something to be gained by not reducing the problem to a
complete-case analysis, but perhaps this function is there just to
allow users to make predictions for missing data points. I know the
asreml team read this list, so perhaps they could comment?