Group-level vs. individual-level model comparison

Modeling
,
3

Thanks for the answer!

  1. Pooling vs. no pooling
    I agree. It’s just something that came up again and again when talking to psychiatry researchers. I have also had no-pooling models tested explicitly against multilevel ones and faring much worse, but still the critique came. But yes, posterior predictive checks by participants is something to work on.

  2. Extracting individual loo scores from multilevel model
    I’d also be happy to hear more from @avehtari as to how the loo scores of individual participants in the multilevel model could be interpreted and what they could be used for.

  3. Mixture model
    One of the models at stake is pretty hard to fit (divergences and chains stalling). After discussions here (e.g. [Clarification] Divergences during sample_prior = "only" model fitting) we managed to fit it with no divergences by using a simplex prior on the btw subject variance across parameters. Adding that model to a mixture model has so far just generated endless divergences.

  4. Qualitative model checking: yes! We had Danielle in Aarhus a couple of weeks ago (talk here: https://youtu.be/tNkmsAOn7aU) and I am converted :-)