Skip to content
Prev 5687 / 20628 Next
Original Raw

Covariance between two traits in MCMCglmm

Jarrod Hadfield 
Hi Ned,

You should probably interact your fixed effects with trait, because at  
the moment you are assuming things like Sex have the same effect on  
P.TimeFront and M.TimeFront  - I'm not sure if this in intentional?

Regarding priors - the ones you have used are quite informative with  
respect to some parameters. For example, specifying  nu=2 as a prior  
for a 2x2 covariance matrix means that the marginal prior for each of  
the variances is inverse-Wishart with nu=1. This can have a strong  
effect if there are not much data.

 From experience I find

list(V=diag(2), nu=2, alpha.mu=c(0,0), alpha.V=diag(2)*a))

where a is something large (e.g. 1000, depending on the scale of the  
data) works well for the two standard deviations and the correlation,  
in terms of informativeness.

You can't use parameter expanded priors for the residual term yet, so  
you will have to stick with the standard inverse-Wishart (or use  
another program). Generally, data are highly informative for the  
residual part so often the posterior is not very sensitive to the  
prior specification. Nevertheless, you should check alternatives:

V=diag(2), nu=1.002 gives the inverse-gamma prior for the variances  
with shape=scale=0.001
V=diag(2)*1e-6, nu=3 is flat for the correlation from -1 to 1

Cheers,

Jarrod
On 23 Mar 2011, at 19:59, Ned Dochtermann wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


  


  
  

    

  

    
      
    
    
Thread (2 messages)

  

  

    
            
        Ned Dochtermann
        
        Covariance between two traits in MCMCglmm
      
      Mar 23
    
    
      
            
        Jarrod Hadfield
        
        Covariance between two traits in MCMCglmm
      
      Mar 24