checking for convergence of MCMC?

[jbloom]

I was curious if there was some easy way to check for convergence of the MCMC, since the example doesn't show any convergence plots.

As best I can tell, the PosteriorHandler returned by running fit tracks the value of the model parameters (eg, growth advantages) at each timepoint, so in principle those could all be plotted as a function of MCMC step to check for convergence.

However, I could not figure out how to access the loss function (log likelihood) at each step. I would think ideally plotting that (which is just a one dimensional variable) might be the easiest way to check for convergence.

Is there a way to access that after the fitting to make such plots, or is there some other suggested way to check for convergence?

I especially was unsure what represents a "reasonable" number of steps as the example in the evofr documentation uses 1000 warmup followed by 1000 steps (which actually takes a while), but the Bedford lab pipeline for forecasts-ncov only uses 200 warmup and 200 samples.