autocorrelation values, autocorrelation time, Geyer's truncation lag

stan-dev / posterior

The posterior R package

https://mc-stan.org/posterior/

Other

167 stars 23 forks source link

Open avehtari opened 4 years ago

avehtari commented 4 years ago

It would be good to have for diagnostics

[ ] autocorrelation values up to some lag (to be checked as numbers or as a plot)
[ ] autocorrelation time (this is related to ESS, but can be useful separately to choose thin value)
[ ] Geyer's truncation lag (this is a safer option for thin value)
[ ] lag which has 95% (or user defined) cumulative autocorrelation time (produces smaller but almost as good thin value as Geyer's rule)

And make a thin function using subsetting method.

paul-buerkner commented 4 years ago

How do we output the autocorrelation when we have multiple chains as each chain will emit its own autocorrelation?

avehtari commented 4 years ago

How do we output the autocorrelation when we have multiple chains as each chain will emit its own autocorrelation?

Combine using Equation (16) in the new Rhat paper.

paul-buerkner commented 4 years ago

Thanks! Thinning is now possible via the thin method.

ASKurz commented 8 months ago

Since it looks like this has been open for a while, I'd like to express my user's enthusiasm for a nice autocorrelation() function.