My collaborator and I are finding that when we try to run coda::gelman.diag() on our model fit often leads to memory problems. We can obviously adjust this with thinning and burnin to some extent, but wondering what shortcuts might be available to us. We lose the "multivariate Rhat," but it looks like JAGS is creating an output (rjags and R2jags) my.model.fit$BUGSoutput$summary with an estimated Rhat for each monitored parameter.... is this likely sufficient?
If all of the single-variable Rhat estimates in the BUGSoutput are below our target value (say 1.1) are we being anti-conservative in any way? Is there a reason to call coda::gelman.diag() in addition to this (gelman.diag also estimates the upper confidence limit on Rhat)?
Are there good rules of thumb to allow us to both include a long enough chain for high effective sample size, but also avoid memory limitations with coda::gelman.diag*()?
