MCMCglmm R-structure problem in heteroskedastic model

Hi Paul,

When you fit an animal model, MCMCglmm introduces "phantom" parents  
with missing data. These phantom parents allow the inverse of A (the  
relationship matrix) to be a) calculated easily b) very sparse and c)  
permuted for solving the mixed model equations more quickly.  When an  
interaction is fitted between animal and a factor like asthma.family  
phantom records are set up for all asthma.family/animal  combinations,  
hence  6730 missing records in models 2 and 3. If the pedigree links  
do not exist between individuals phenotyped in the two groups, then  
there would be more efficient ways of setting up the model, but they  
are not implemented. When the number of missing records is large it  
can even be more efficient (in terms of effective sample size per  
second) to do away with phantom parents and use a denser inverse A  
(specify nodes="TIPS" in the call to MCMCglmm).
If some individuals in mfs.fam have not been phenotyped you can use  
prunePed to remove useless individuals (make sure make.base=TRUE).

In model3 the phantom parents have not been assigned an asthma.family  
because you have a simple random term ~animal.  When it comes to build  
the R-structure it does not "know" what to do with the individuals.  
These phantom parents could be assigned any asthma.family and it would  
make no difference to the posterior, but unfortunatley I did not think  
that far head when writing MCMCglmm. I will try and fix this in later  
versions. The work around is to add the phantom parents to mfs.fam,  
assigning them anything for the fixed effects but having NA for the  
response. It should then run as intended.

Cheers,

Jarrod

MCMCglmm R-structure problem in heteroskedastic model

Thread (4 messages)