You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Since the title of #178 is centered around the iso_probability_contours functionality which is being implemented in #188 (it might be good to also have a KDE-based version alongside @Stefan-Heimersheim's kde-less estimates), I would like to collect the other ideas for 2d plots in a separate issue.
The basic idea is that '67% iso-probability contour' and 'minimum area contour containing 67% of the probability mass' are equivalent. The latter definition in theory means one could produce credibility regions without estimating the probability density directly, i.e. credibility regions without KDEs.
My first thoughts on this would be to use a triangulation/voronoi diagram (possibly smoothed/estimated) to get area estimates for each sample, and construct an algorithm for assembling the minimum area containing 67% probability mass (which sounds on the face of it like a relatively straightforward constrained optimisation problem). Likely issues will arise with noisiness of area estimation from random samples (as one finds when doing single-point histograms).
The text was updated successfully, but these errors were encountered:
Since the title of #178 is centered around the iso_probability_contours functionality which is being implemented in #188 (it might be good to also have a KDE-based version alongside @Stefan-Heimersheim's kde-less estimates), I would like to collect the other ideas for 2d plots in a separate issue.
critical idea from @Stefan-Heimersheim
proof for D-dimensions
The basic idea is that '67% iso-probability contour' and 'minimum area contour containing 67% of the probability mass' are equivalent. The latter definition in theory means one could produce credibility regions without estimating the probability density directly, i.e. credibility regions without KDEs.
My first thoughts on this would be to use a triangulation/voronoi diagram (possibly smoothed/estimated) to get area estimates for each sample, and construct an algorithm for assembling the minimum area containing 67% probability mass (which sounds on the face of it like a relatively straightforward constrained optimisation problem). Likely issues will arise with noisiness of area estimation from random samples (as one finds when doing single-point histograms).
The text was updated successfully, but these errors were encountered: