On the Diagnose tab in Shinystan there is a histogram of the Monte Carlo standard error divided by the posterior standard deviation, i.e., “mcse/sd”. Is this simply the ratio of “mcse_mean” and “sd” in the summary statistics one gets from summarise_draws? Can someone point me to a reference that discusses this diagnostic in more detail? Thanks.
1 Like
I would guess you are correct, but I don’t have a good reference, @jonah (who develops shinystan) might know more. I think the appeal of using “mcse/sd” as a diagnostics is that it tells you how much variability you expect to be there from the Monte Carlo process relative to the scale of the posterior. As such it is not perfect - one would generally want to interpret MCSE withour rescaling, considering the exact role of the parameter in the model - but as a quick check it is better than unscaled MCSE which is hard to interpret without reference to a specific model.