[R-sig-ME] MCMCglmm output 1) on right scale? 2) produces huge deviance spread? (Jarrod Hadfield)

[Ryan King]

For posterity / google: It turns out the key issues were
1) The R component specified in the model is not the latent
gaussian(0,1)  noise inherent in a probit regression, but in addition
to that quantity. Setting it to near zero results in the answers I
expected, but very poor mixing of the chain. Fixing it at some value,
and doing the deviance calculation with residual variance 1 +
mcmcoutput$VCV[,"units"] does what I want.

2) The mcmcoutput$Dev and $DIC are the deviance treating the R
component as a real piece of the model and not a computational device.
That is, there are effectively hundreds of additional parameters. If
you want the deviance treating R as a computational device, you have
to calculate it by hand as above.

3) I had a typo in my prior specification, G1=list(V=1, nu=1,
alpha.mu=0, alpha.v=number) does not trigger parameter expansion or an
error; it has to be alpha.V=number.

4) Near-singular variance components are possible in parameter
expanded models (ie a cauchy prior and data supporting the null G=0)
and lead to unstable calculations; the mixed model equation solver in
the block-gibbs update of random effects computes G^-1. There are
other formulations of the MME which don't require inverting G, but
they tend to be less sparse.

Ryan King