Projpred with mixed models

I’m developing a mixed model with 6 variables where the subjects are random effects. I would like to do variable selection. One way would be exploring all 2^6 combinations and running a LOOCV to select the model with the lowest ELPD. However that seems to be computationally daunting.

Instead, are the methods described in projpred appropriate for mixed models as well? I’m using rstanarm to develop these.

Thanks a ton!!

That isn’t implemented yet in the proppred R package, but the idea of it makes sense if you want to implement it yourself. Also, choosing among a huge number of models by which has the highest ELPD (no matter whether you use LOO or something else to estimate it) is quite biased, so the projection pursuit approach is better.

I predict we’ll have implementation available at latest before StanCon, but maybe before that.

that is great, thank you! I was waiting for that too

Hi Aki,
is the projpred package available for brms glmm models already?

Not yet, but I still assume we’ll have it ready (in some form) before StanCon

1 Like