5% seems quite a high acceptance rate when there are so many simulations. Using 0.1% gives more sensible looking results for the scale parameter, but the location still has very large posterior variance. Regression post-processing would probably help. Alternatively an idea I’ve been looking at recently is to replace “mads” with “norm” where:
res=lm(summ~prior)$residuals

norm=apply(res, 2, sd)

(Also in the code the true parameter values are (0,1) rather than (3,2) as in the diagram.)

Anyway, interesting example!

]]>