Aki Vehtari
avehtari.bsky.social
Aki Vehtari
@avehtari.bsky.social
Professor in computational Bayesian modeling at Aalto University, Finland. Bayesian Data Analysis 3rd ed, Regression and Other Stories, and Active Statistics co-author. #mcmc_stan and #arviz developer.

Web page https://users.aalto.fi/~ave/
Yes
November 11, 2025 at 5:46 PM
It looked much more impressive for the naked eye. The photo is with a phone and daylight setting. The night mode photo was too bright, losing contrast. The photo also misses the movement and immersion
November 2, 2025 at 2:06 PM
Thanks for the feedback! Can you add your comments to the PR discussion? It's easier to discuss there without message length limit.

Shortly:

p_worse is expected to be too high when |elpd_diff|<4

Reporting all diagnostics would take too much space, and thus they are in order of dominance
October 29, 2025 at 3:20 PM
We have a PR for `loo` package github.com/stan-dev/loo..., adding new columns to `loo_compare()` output. Check out how the new output looks in a variant of that case study users.aalto.fi/~ave/casestu...

Please, check out the case study and possibly try also that PR, and provide feedback!

3/3
add diff diagnostics by avehtari · Pull Request #300 · stan-dev/loo
The corresponding issue #299 This PR is not ready for merge, but needs some thought
github.com
October 28, 2025 at 1:39 PM
We provide conditions when the predictive performance difference and associated SE based normal approximation is well calibrated and can be used to compute a probability that one model has better performance

The case study with code is available at users.aalto.fi/~ave/casestu...

2/
Uncertainty in Bayesian LOO-CV Model Comparison
users.aalto.fi
October 28, 2025 at 1:39 PM