search

wbalmer / backtracks

Python package to fit relative astrometry with background star motion tracks.
https://backtracks.readthedocs.io/
BSD 3-Clause "New" or "Revised" License
6 stars 2 forks source link

Posterior quantiles #15

Closed tomasstolker closed 10 months ago

tomasstolker commented 10 months ago

Should we change this to 16th and 84th percentiles? Which are more common to quote as uncertainties?

https://github.com/wbalmer/backtrack/blob/930c57124dea87df3f73430150265ce07949746a/backtrack/backtrack.py#L453

gotten commented 10 months ago

I'm not too familiar with dynesty. After looking at the code it seems it calculates quantiles for each of the parameters and stores it in the pkl results files. It also uses them as parameters for the trackplots (labeled 1 sigma). So 0.16, 0.84 sound like the correct quantiles as you suggest. However taking the lower and upper 1 sigma value of each of the parameters combined will not be the same as plotting the 1 sigma contour for where we can find our target (our parameters are not independent). Which is a different problem highlighted by https://github.com/wbalmer/backtrack/pull/13 . So: 0.16, 0.84 is better as it gives correct values in the pkl files, but we probably need to rewrite part of the code in the plotting routine to show actual 1 sigma probabilities in the trackplots.

wbalmer commented 10 months ago

for some reason while i wrote that line of code i thought that 0.5+/-0.18 = [0.32,0.68] was the 1 sigma interval, but that must have been late night coding (probably while up observing). this should be corrected in commit de0e8f6 and folded into PR #18