[R-sig-ME] MCMCglmm : Bivariate model and prior

Jarrod Hadfield j.hadfield at ed.ac.uk
Tue Jul 28 20:40:27 CEST 2015


Hi,

The answer to your first question is field work.

The prior for the residual term is list(V=V, nu=nu, fix=2) where V is  
some covariance matrix with V[2,2]=1, and nu is the degree of belief  
parameter. You could try a flat prior initially i.e V=diag(2) and nu=0.

I would advise using family="threshold" rather than family="ordinal"  
for this analysis. If the estimated residual covariance matrix is VR  
then then the residual correlation on the latent scale is:

VR[1,2]/sqrt(VR[1,1])

with family="ordinal" it is:

VR[1,2]/sqrt(VR[1,1]*2)

and so this constrains the correlation to be less in magnitude than  
the theoretical limits of -1 and 1 because VR is constrained to be  
positive definite.

Cheers,

Jarrod





Quoting Grégory DANIEL <daniel.gregory3 at gmail.com> on Tue, 28 Jul 2015  
20:01:54 +0200:

> Dear MCMCglmm users,
>
> Few weeks ago, I asked a question concerning some issues that I have  
> to parametrize a bivariate model in MCMCglmm. So I'm sorry to ask  
> again. But I want to be sure that I had no response because no one  
> says or because it's impossible, and not because of an simple  
> oversight, a field work or a vacation. ^^
>
> I wish to run a bivariate model with MCMCglmm package. But one  
> response variable is a continuous one (family = gaussian) and the  
> other is a binary one (family = ordinal). I have sought a way to  
> parameterize properly the priors (in the archives of this list too),  
> I tried some solutions that I thought, but in vain. I think I have  
> to fix to 1 the residual variance for the binary variable and leave  
> it free for the continuous variable.
> Is it possible in MCMCglmm ? And if yes : how do I parameterize the  
> priors for the residual variance ?
>
> I'll very grateful if someone could answer me.
>
> Grégory DANIEL
> PhD Student - LBBE
> University of Lyon 1
> 69100 Villeurbanne - FRANCE
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.



More information about the R-sig-mixed-models mailing list