Exploratory-Analysis-of-Bayesian-Models

Exploratory-Analysis-of-Bayesian-Models copied to clipboard

plot_compare, WAIC. Bayes Factor

i think we should use this section to explain how to handle multiple observations and ic calculation for hierarchical models (i.e. similarly to these two discourse answers: one, two).

ArviZ docs on waic, loo and compare should then add this to See also/References section. Eventually, that would close https://github.com/arviz-devs/arviz/issues/987 and https://github.com/arviz-devs/arviz/issues/998.

Does the rugby model sound as a good example for this? I don't like that leave one match out is so similar to leave half match/goal recording out. A couple options are to use football league data (premier or spanish league have 20 teams instead of 8, but still not sure would work) or using a similar model but on some other field where the equivalent to match has 3 or 4 components instead of only 2.

Good idea! Working on a football model would be fun 😉 Are there some good data available though? We can also write a first version with the rugby model (highlighting that the two different CVs are not expected to always be the same), and update with a new, better football example later?

Are there some good data available though?

It is easy to copy from wikipedia :wink: and do some pandas magic.

I support everything said here

We should also include an explanation on the different methods of computing model's weights (mostly stacking vs pseudo-BMA) and more important the effects this have on the expected results. See https://github.com/arviz-devs/arviz/issues/2077 for a common misconception we need to address