1
$\begingroup$

I am fitting a multi-level state-space model and am running into a situation where the Gelman-Rubin diagnostic shows acceptable convergence (R-hat < 1.01), but when I look at the trace plots of the MCMC iterations, they show a considerable lack of convergence. An example for one parameter from my model is shown below, although I am getting similar results for all of my parameters. I am wondering if this is a problem that is already known, and if this should point me to something wrong with my model. Obviously, I know that I can try adding more iterations, but I am still curious as to why the G-R statistics indicates acceptable convergence in this situation.

MCMC convergence diagnostics

$\endgroup$
2
$\begingroup$

Basically, the Gelman-Rubin diagnostic looks at how different your repeated chains are. They are extremely similar in your case, giving you a relatively small GR diagnostic. However, at the same time your chains are obviously not converged and you need a longer burn-in (warm-up) period until you get actually useful posterior densities.

$\endgroup$
2
  • 2
    $\begingroup$ It looks as though all the chains were started at similar values that are a long way from the high-probability regions of the posterior, so that they're all taking the same path slowly in the same direction. I think between-chain diagnostics will be more useful when the starting points of the chains are more spread out across the prior. $\endgroup$ – Thomas Lumley May 27 at 1:13
  • 1
    $\begingroup$ This makes sense and I figured my situation must be due to convergence between the chains, and not convergence around a steady parameter space. I must say though that I am surprised by how often I see the Gelman-Rubin diagnostic used to assess 'convergence' generally, when relying on this alone can obviously be highly problematic. $\endgroup$ – darkness May 27 at 2:46
1
$\begingroup$

When describing the Gelman-Rubin statistic (R hat) in Bayesian Data Analysis (3rd Ed, ch. 11.4, pp. 283-285), Gelman and his coauthors say that the multiple chains used in calculating R hat should be simulated with overdispersed starting points and further that each chain should be split at the middle into two parts.

For example, even if only two chains are simulated, the diagnostic would be computed across the four half chains.

If you make either of these changes (diffuse starting points and chain splitting) to your diagnostic procedure, I think you will find the Gelman-Rubin statistic would identify that your chains have failed to converge.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.