Hi,

 

Thanks to Chris Jackson and Martyn Plummer who have helped me with this issue.

 

Chris stated that he didn’t have the same problem, running the Mice eg with the following changes to the code/inits given in the manual:

 

      t[i, j] ~ dgamma(r, mu[i])I(t.cen[i, j],)

      r ~ dunif(0, 100)

      Inits  list(beta = c(-1, -1, -1, -1), r = 3)

 

There is still a bimodal deviance, although it is much less pronounced than in my case.  I also did run into Trap errors on running the chain for longer, and the model appeared quite sensitive to initial values.  However, I want to parameterise the model with the mean as the log-linked predictor (see changes to code below):

 

t[i, j] ~ dgamma(r, mu[i])I(t.cen[i, j],)

mu[i] <-  r / mean[i]

mean[i] <- exp(beta[i])

 

I struggle to get this to run at all, and so appears even more sensitive.  When I have got it to run the deviance is much more pronounced bimodal.  I am looking for a parameterisation that is insensitive to sensible default priors (any advice on sensible default prior distribution & parameters for r, or maybe the Coefficient of Variation =root( 1/r)?) and initial values – which is possible for Weibull and Log-Normal error model.

 

Martyn ran this through JAGS and reported no strange behaviour at all.

 

My conclusion is that there is some feature of how the Gamma dist is coded in WinBUGS, apparently different to JAGS, that makes it more sensitive in this case.  When WinBUGS crashes at least this provides a clear alert to the user, otherwise it appears that the Deviance node should be inspected for strange behaviour.  If anyone is able to shed further light on this, this information would be gratefully received.

 

Many thanks

 

Phil

 

     LEGAL NOTICE
    
     Unless expressly stated otherwise, this message is confidential and
     may be privileged. It is intended for the addressee(s) only. Access to
     this e-mail by anyone else is unauthorised. If you are not an
     addressee, any disclosure or copying of the contents of this e-mail or
     any action taken (or not taken) in reliance on it is unauthorised and
     may be unlawful. If you are not an addressee, please inform the sender
     immediately.
    
     Pfizer Limited is registered in England under No. 526209 with its
     registered office at Ramsgate Road, Sandwich, Kent CT13 9NJ