I thought of a procedure to generate high probability density regions with probability $1-\alpha$ from $n$ MCMC draws:
- Find the $\lfloor(1-\alpha)\cdot n\rfloor$ draws with the largest probability density;
- Take the confidence hull of these draws as an approximate HPD region.
Is this procedure correct? I believe it only works when the distribution is convex (or maybe log-convex, I'm not sure). I think this is a more flexible approach than using Mahalanobis distance, since the resulting region is convex but not ellipsoidal, and requires less tuning than kernel density based methods.
Thanks in advance!