I’m sharing a case study which demonstrates how to use Stan’s newly released embedded Laplace approximation. This is an update of an old case study we wrote when we had a prototype out. Open to feedback on this draft.
Thanks, Charles. I think this is super important for getting people to actually use th emethod.
Do you want to publish that with the other Stan case studies on mc-stan.org?
Yes, good idea, I’ll submit to Stan’s case study page.
Cool, thanks Charles. I agree, it would be nice to have this with the case studies on the Stan website
Des this method allow for posterior predictor of OOS observations? Or do I need to pass all observations that I want to predict to the model?
Currently, Stan doesn’t provide a dedicated function for out-of-sample predictions but it’s on our radar ( Add embedded laplace RNGs for out-of-sample prediciton · Issue #3179 · stan-dev/math · GitHub ).
With what Stan currently offers, you need to pass all your observations (in-sample and out-of-sample) in generated quantities. This is of course inefficient, especially for something like a GP model, but in a Stan model the computation tends to be dominated by operations in the transformed parameters and model blocks (which only deal with in-sample observations), so hopefully the additional cost is manageable.