1

I run Bayesian models with Runjags and then convert the output in MCMC.list with the coda package. I check convergence with the Gelman-Rubin diagnostic (univariate).

Sometimes, the PSRF is large just because a chain sampled a large value at some point (PsRF goes from ~1 to 1,2). On the other side, the PSRF is close to 1 with Runjags. Sometimes Runjags has larger PSRF than Coda.

I didn't find what is the difference in calculation. Do you know it? Is-it ok to think that the parameter converged from plots even if the PSRF is 1,2?

merv
  • 67,214
  • 13
  • 180
  • 245
user3192383
  • 63
  • 3
  • Generally, I would think that if a single large-value observation in low likelihood space is throwing off your R_hat (PSRF) that much, then you aren't sampling enough. – merv Mar 04 '19 at 16:36

0 Answers0