[R-sig-ME] glmer vs. MCMCglmm

[Hervé CHAPUIS]

Hi everyone,

Hope my questionsare  not too stupid,  as I could easily be defined as a 
Bayesian dummy.
I intend to estimate the genetic variance for a binary trait (disease 
resistance) in a simulated population.
First, I have used lmer and glmer.  However, sometimes I have false 
convergences, and  I can't specify "nAGQ=5" when fitting two random 
effects (sire AND dam).
This is the reason why I have decided to give a glimpse at MCMCglmm.
Well, what can I do when I have this kind of error message :

"Erreur dans MCMCglmm(Y ~ 1, random = ~PERE, family = "categorical", 
data = PERF) :
  ill-conditioned G/R structure: use proper priors if you haven't or 
rescale data if you have". ???

How to specify proper priors ? If I have a binary trait, the residual 
variance can't be estimated, so that  it has to be fixed, isn't it ? 
unless I use a proper prior.

In another design, I obtain results, but the sampled sire variance far 
exceeds the parameter space boundaries, leading to  an abnormaly  high 
heritability coefficient (above 1). The glmer estimate, on the other 
hand, is much in adequation with the expected value.

I am still trying to implement an animal model as specified in the 
MCMCglmm manual, but I can't figure out how an heritability can be 
estimated so high without a big mistake. But I can't see it.
 
Any help will be greatly appreciated.

Thanks.

-- 
Cordialement, 


Hervé CHAPUIS
SYSAAF
Station de Recherches Avicoles
37380 NOUZILLY

tel : 02 47 42 76 77
fax : 02 47 42 76 46