I am estimating a Gaussian copula with beta marginals using zero-trick. The beta distributions has regression, and I am using the "pbeta" function for the probability integral transformation of the marginals. The model successfully completes the adaptation phase with 5000 iterations, but even after the adaptation phase, the step sizes are very small (about 0.00001) for the actual MCMC. This signals that the chain is not efficient and it needs large number of burn-in, which takes too much time. I suspect that the zero-trick makes the chain inefficient. Is there any way that I can find the reason for the inefficiency of the chain and how can I make it more efficient? Any help is appreciated.
I suspect that JAGS is falling back on a random walk Metropolis-Hastings algorithm, which is just too inefficient to work in your problem. You can check this (See ?list.samplers in rjags or "samplers to" in the JAGS manual).
Hamiltonian Monte Carlo would probably work wonders on this problem. You might want to look into Stan to see if that can help.
Is there anyway that I could change the updater for some nodes? I guess for the regression coefficients JAGS is using block updater, and that might be the reason. Can I change it to uni-variate MH or maybe slice update?