Autocorrelations and differencing

Hi Stan community

I was wondering how one should treat posterior samples of a stanfit object regarding the autocorrelation of the samples.
Say we sample 10k samples after 10k warmup.

ESS, Rhat are okay / good.

Yet, when we look at the ACF of saying a scale parameter, ACF goes to 0 around the 15th observation for all 4 chains.

Should we then thin the series to every 15th value to have proper inferences and predictions?

Sorry for this likely very basic question… I could not find a satisfying answer with a couple of google searches…

I guess it would be easiest to simply re-run the model and increase thin to 15?

Just use all of your draws. (Unless there are memory issues, then thinning could help)

Yes, the reason is that you get better estimates with all the draws. The MCMC standard error being reported gives you standard errors on mean estimates.

1 Like