Is save_psis = T when calling loo from cmdstanr?

Interfaces CmdStan
,
1

Hi there,

May I ask the default value for save_psis when calling loo with a fitted model like fit$loo()? It seems in loo wiki Efficient approximate leave-one-out cross-validation (LOO) — loo • loo, save_psis is not enabled. So I guess it is the same case for cmdstanr Leave-one-out cross-validation (LOO-CV) — fit-method-loo • cmdstanr? The difference is cmdstanr set r_eff = T by default. Could you please help me to confirm that if my goal is to perform model comparison and minimize loo resulted object size, should I do this fit$loo(r_eff = T, save_psis = F) or fit$loo(r_eff = NULL, save_psis = F)

Thanks,
ZC

2

The default is save_psis = FALSE. I think this

seems like the right option based on what you described you want. It will result in calculating the r_eff values (which is good if you care about PSIS effective sample sizes and MCSE estimates) but it won’t save the psis object and should result in a smaller object compared to setting save_psis = TRUE.

2 Likes