Which diagnostics are included?

[jgabry]

At a minimum I’d say

• Effective sample sizes (new robust version)
• R-hat (new robust version)
• Monte Carlo standard errors

Questions:

• Which other general MCMC diagnostics are worth including?
• Do we include algorithm-specific diagnostics? For example for hmc and variants.
• For things like ESS and R-hat that have new versions, do we include the old versions for comparison?

[paul-buerkner]

Let's implement the minimum first and then see what is missing.

I would like to eventually have the algorithm-specific diagnostics (HMC) as well, especially as we need them in bayesplot, for instance.

I would argue we should also support some old diagnostics for comparison and research purposes. The code overhead should in any case be minimal.

[jgabry]

Sounds good to me.

[paul-buerkner]

Same question as in #3: What diagnostics should be included in the default output of summarise_draws?

[paul-buerkner]

The old/basic versions of (split) Rhat and ESS are now exported as well via Rhat_basic and ess_basic. I think that the currently implemented diagnostics are sufficient for the beta release. All the sampler specific diagnostics may still be implemented at a later stage. But let's discuss this on monday.

[paul-buerkner]

We decided so show Rhat and bulk/tail ESS in the summary by default leaving out the MCSEs in order not to clutter the default output too much. Of course, they can easily be added to the summary output.