Documentation of pp_average "loo2"

loo

#1

Hi,

i ran pp_average() with “loo2”. For me, its not clear which exact method is represented by loo2. In the package brms, loo2 refers to loo_model_weights in the package loo. Taking a look at this package, several choices are listed e.g. “stacking”, “pseudobma” named in the paper of Yao et. al. (2018)*. The authors further refer to Pseudo-BMA+ (AIC type weighting using Bayesian Bootstrap and the expected log predictive density).

Which method does loo2 stand for? Do I have to specify this further in the pp_average() function?

  • Yao, Vehtari, Simpson & Gelman (2018): Using Stacking to Average Bayesian Predictive Distributions

#2

Unless you pass other arguments to loo_model_weights, "loo2" implies whatever the default of loo_model_weights is.


#3

Thank you for your reply. But i guess thats exactly the problem. In loo_model_weights, i do not see a default option. In the loo package, the method for loo_model_weights is described as: “Either “stacking” or “pseudobma”, indicating which method to use for obtaining the weights. “stacking” refers to stacking of predictive distributions and “pseudobma” refers to pseudo-BMA+ weighting (or plain pseudo-BMA weighting if BB is FALSE).” Which of them is the default?


#4

The first one, which is “stacking”. As a general rule in R, if multiple options are given, the first one will be the default unless otherwise specified.