Skip to content
Prev 4570 / 20628 Next

priors for a multi-response model (MCMCglmm)

Hi Ned,

I think you haven't specified the model that you want to fit.

1) you have "random=~us(trait):units, rcov=~us(trait):units" which is  
fitting the same terms for the random effects and the residuals. I  
think you want something like  "random=~us(trait):individual,  
rcov=~us(trait):units"? At the moment any separation of these effects  
is coming entirely from the prior.

2) depending on what your responses are "~trait+period+day" for the  
fixed terms is probably not appropriate. You probably want to interact  
period and day with trait, so that the effect of these two predictors  
can have independent effects on each response

3) zero-inflated poisson are treated as multi-response. The first  
latent variable is for the poisson counts, and the second is for the  
zero-inflation. The residual variance for the zero-inflated part  
cannot be estimated from the data (for the same reasons that the  
residual variance in a binary model is not identified) and neither can  
the residual covariance between zero-inflation and the poisson counts  
(you only see one of the processes in any one observation).  I  
generally place strong priors on them by fixing the residual variance  
to one and the residual covariance to zero. I achieve this by fitting  
an idh residual structure idh(trait):units which fixes the residual  
covariance to zero, and fix the 2nd diagonal element of the residual  
matrix (the zero-inflation variance) to one in the prior:  
"prior=list(R=diag(2), nu=...,  fix=2)"

4)  If you have 3 ZIP responses you have 6 "traits" to worry about.   
It is not possible to place the appropriate constraints on this  
matrix: the sub-diagonals cannot be estimated (corresponding to the  
covariance between the poisson counts and the zero-inflation within  
traits) and neither can the even elements of the diagonal (the 3 zero- 
inflation terms). You may be able to specify a prior which has a  
desirable marginal effect on the things you want to calculate, but I  
think this would be hard and a lot of work.

5) These are very parameter rich models, and I would avoid them unless  
you have a lot of individuals measured many times.

Cheers,

Jarrod
On 11 Oct 2010, at 21:06, Ned Dochtermann wrote: