How to verify that uncertainty (estimated from pymc3) is accurate?

  1. I haven’t used SBC in practice but the histogram looks reasonable. I’d recommend plotting it with the gray band like they have in Figure 1 and 2 to determine if you’re observing an acceptable amount of variance in the bins. The QQ plot looks pretty good, though you can tell there’s a super slight S-curve to the orange line indicating slight miscalibration (either the posterior is too tight or too disperse, depending on how you assigned the X and Y axes.

  2. I don’t think a single KS statistic alone is enough to determine anything. Are there any baselines you know should be perfectly calibrated you can compare against? E.g. in the privacy paper I linked above I compare the private methods (which have to deal with a lot of noise) to the non-private method which has exact data and should therefore always be calibrated.

  3. Agreed. There’s an interesting blip at 1, indicating the posterior’s mass is too often to the left of the true parameter, so it’s biased for some reason. Whereas the Half-Gaussian prior seems centered properly but either too wide or too narrow (depending on how you plotted it). As @chartl noted, looking at the trace would be helpful, as would a histograms of the posterior samples over multiple trials.

1 Like