[SciPy-user] computing Bayesian credible intervals

Anne Archibald peridot.faceted@gmail....
Tue May 6 04:09:44 CDT 2008


2008/5/6 Johann Cohen-Tanugi <cohen@slac.stanford.edu>:
> hi Anne, thanks! I will need some time to digest your suggestions and
>  try to implement them. I think that the condition p(a)=p(b) where p is
>  the actual pdf and not an integral of it, is the sufficient condition
>  for unimodal distributions so that the credible interval is uniquely
>  defined as the interval encompassing the mode.

Yes, this should work; I would have gone with an interval that
represented (say) 5th, 50th, and 95th percentiles, bracketing the
median instead of the mode, but that might be peculiar if the
distribution has large tails or you want to compare with a
maximum-likelihood method (which returns the mode, I guess). My actual
experience with Bayesian statistics is pretty limited; it's probably
best to go with whatever's most standard (if the numerics can be made
to behave).


Anne


More information about the SciPy-user mailing list