topipa / iwmm Goto Github PK

I think it would be good to have a moment_match.brmsfit method. It should probably return a modified brmsfit object with the draws replaced with the moment matched draws, so that further functions from brms would work (e.g. posterior_predict, see n-kall/priorsense#18). I'm not sure where the weights would be stored though. At some stage I think brms will store the posterior in draws format (rather than the current stanfit object), so that would help, but that will likely only happen in the next major release.

Currently loo::psis is used to calculate the pareto-k diagnostics, but there is work to have a pareto_khat diagnostic function in posterior. Using this instead would remove the dependency on loo and allow for loo to suggest iwmm (as discussed here).

I can make a branch implementing this change and create a PR once the corresponding functions are merged into posterior

Now that we've changed over to posterior rather than loo for the psis code, we could think about submitting to CRAN so that other packages could use the iwmm functionality. Specifically I'm thinking of priorsense and potentially loo itself. I think that iwmm could be used to fix stan-dev/loo/issues/209

I'm opening this issue mainly to keep track of the tasks we would need to do to get iwmm on CRAN.

There's a PR to change the output of unconstrain_draws to a draws object (stan-dev/cmdstanr/pull/886) so we can simplify the code below when that gets merged