Is there any way to update model posteriors and predicted values after moment_match = TRUE? Otherwise, apart from getting K <0.7 how can it be useful in the results we obtain from the model?
Thanks for your guidance.
Currently, no, although it would be a nice feature. LOO predictions are useful for model checking, for example by examining coverage of LOO predictive intervals or uniformity of LOO-PIT values, and if moment matching is used it would be good to use it for those LOO predictions, too.
The reason why it’s not currently there is that it requires some work to add.
For each moment matched LOO-posterior we would need to store the transformed draws or the parameters of the affine transformation, or we would need to rerun the moment matching at the time of making predictions. We could reduce the memory or computation time by doing this only for the folds that originally had high Pareto-k. Any such implementation adds complexity to the code which means it’s not quick to add.
I added an issue Support moment matching LOO for LOO predictions · Issue #1674 · paul-buerkner/brms · GitHub so that we don’t forget this, but it may take some time this gets implemented.
@avehtari Thank you. I am looking forward to it. On a different note, applying moment_match=TRUE on the resulting model from brms is extremely slow. I wonder if that will change in the future.
Thanks for all.