Use Split R-hat #477

ahartikainen · 2018-12-31T08:52:03Z

Update R-hat calculation to use Split R-hat described in BDA3.

Follows calculation done in https://github.com/stan-dev/pystan/blob/develop/pystan/_chains.pyx

ahartikainen · 2018-12-31T10:38:06Z

mc = "parameters { real y;} model { y ~ normal(0,3);}"
sm = pystan.StanModel(model_code=mc)
fit = sm.sampling()
idata = az.from_pystan(posterior=fit)

Will yield

pystan._chains.split_potential_scale_reduction(fit.sim, 0)
# 1.0033341784610692
az.stats.diagnostics._get_split_rhat(idata.posterior.y.values, round_to=16)
# 1.0033341784610692

canyon289 · 2018-12-31T17:07:34Z

arviz/stats/diagnostics.py

 shape = values.shape
 if len(shape) != 2:
 raise TypeError("Effective sample size calculation requires 2 dimensional arrays.")
 _, num_samples = shape
-
+ num_split = num_samples // 2
+ # Calculate split chain mean


Thanks for comments!

ahartikainen added 7 commits December 31, 2018 10:01

rename rhats

8f07b09

split rhat

f1fd5d1

fix forestplot

960eb4e

fix rhats

c756cfd

fix tests

a61768c

fix missing 2

cf6afa7

fix rhat docstring and test for ndarray

3bd185c

canyon289 approved these changes Dec 31, 2018

View reviewed changes

canyon289 merged commit 7f189c5 into master Dec 31, 2018

canyon289 mentioned this pull request Dec 31, 2018

Rename arviz.gelman_rubin -> arviz.rhat and use the same Rhat as in Stan #361

Closed

canyon289 deleted the bugfix/fix_rhat branch January 4, 2019 07:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use Split R-hat #477

Use Split R-hat #477

ahartikainen commented Dec 31, 2018 •

edited

Loading

ahartikainen commented Dec 31, 2018

canyon289 Dec 31, 2018

Use Split R-hat #477

Use Split R-hat #477

Conversation

ahartikainen commented Dec 31, 2018 • edited Loading

ahartikainen commented Dec 31, 2018

canyon289 Dec 31, 2018

Choose a reason for hiding this comment

ahartikainen commented Dec 31, 2018 •

edited

Loading