Short Chains Implementation in Stan

Dear Stan Community

Currently I am doing a bayesian quantile regression and I am faccing some issues about the computation time to reach an R hat of 1. Reading some possible solution I found this idea of short chains, that was developed by phD Charles Margossian @charlesm93 (add some links of paper and thesis as reference)

https://arxiv.org/pdf/2110.13017

One of the main ideas he proposed there is not to do a few chains with a lof iterations but have a lot of short chains with short iterations, because of that the R hat is no longer useful but he proposed the nested R hat to evaluate the convergence of the model.

Is this possible to do this in Stan? If it is, does anyone have any example of this?

Here I found some code but I am not sure if there is a way to do this in Stan, mainly because of the need of that nested r hat

If you are using R, you can use nested-\widehat{R} implementation in posterior package: Nested Rhat convergence diagnostic — rhat_nested • posterior

5 Likes

Thank you for the info!!