JAGS and WinBUGS giving differing DIC -

i'm doing network meta-analysis including several clinical trials. response binomial. each trial contains several treatments.

when random effects model, output jags , winbugs similar. when fixed effects model, dic , pd components way out, though posteriors of parameters i'm interested in similar.

i've got similar models have gaussian response, not binomial, , jags , winbugs in agreement.

the bugs/jags code fixed effects model lifted page 61 of this , appears below. however, same code runs , produces similar posteriors using winbugs , jags, it's dic , pd differ markedly. don't think code problem.

for(i in 1:ns){  # loop on studies   mu[i] ~ dnorm(0, .0001)     # vague priors trial baselines   (k in 1:na[i]) {  # loop on arms     r[i, k] ~ dbin(p[i, k], n[i, k])       # binomial likelihood     logit(p[i, k]) <- mu[i] + d[t[i, k]] - d[t[i, 1]]       # model linear predictor     rhat[i, k] <- p[i, k] * n[i, k]       # expected value of numerators     dev[i, k] <-        2 * (r[i, k] * (log(r[i, k]) - log(rhat[i, k])) +            (n[i, k] - r[i, k]) * (log(n[i, k] - r[i, k]) +       - log(n[i, k] - rhat[i, k])                         ))       # deviance contribution   }   resdev[i] <- sum(dev[i, 1:na[i]])     # summed residual deviance contribution trial } totresdev <- sum(resdev[])   # total residual deviance  d[1] <- 0   # treatment effect 0 reference treatment (k in 2:nt){   d[k] ~ dnorm(0, .0001) }  # vague priors treatment effects 

i found old post describing known issue, that's old me think it's same problem.

are there known issues jags reporting wrong dic , pd? (searching "jags bugs" isn't helpful.)

i'd grateful of pointers.

there number of different ways calculate pd, , method used jags differs used winbugs. see details section of following file:

?rjags::dic 

specifically:

dic (spiegelhalter et al 2002) calculated adding “effective number of parameters” (pd) expected deviance. definition of pd used dic.samples 1 proposed plummer (2002) , requires 2 or more parallel chains in model.

the details in (lengthy, worth reading) discussion of following paper:

spiegelhalter, d., n. best, b. carlin, , a. van der linde (2002), bayesian measures of model complexity , fit (with discussion). journal of royal statistical society series b 64, 583-639.

in general: methods calculate pd (of aware) approximations, , if 2 such methods disagree because assumptions behind 1 (or both) approximation(s) not being met.

matt