I think this one might actually be the winner. First, the results. I expanded the vertical scale around the true answer to better show the convergence. The 95% bounds contain the real answer 97% of the time. (Bounds are colored red when they do not contain the answer). The only extended period outside the bounds is after a somewhat fluky block was sampled early on, throwing one of the stratum estimates off by quite a bit. But the time anybody would actually be accepting the result, any deviations from the bounds is quite minor.
So, what did I actually do? Well, as the Russians appear to be reading this blog again (really, a bunch of hits lately), I'll have to be a little vague; especially since my adviser doesn't seem to be in any hurry to get this published. But, basically, it's just subbing in the sample moments with a prior bias towards extremely variable results. It works, it's fast, and it parallelizes. Still some writing to do, but I'm callin' this one done.
No comments:
Post a Comment