Extract the normalized Pareto smoothed importance sampling (PSIS) weights from a brma model object.
Arguments
- object
a brma model object.
- unit
output/deletion unit. See
add_loo.- ...
currently unused.
Value
An S x K matrix for estimate-unit LOO, or S x G
matrix for cluster-unit LOO, with posterior samples in rows and LOO targets
in columns. Columns are normalized to sum to one.
Details
LOO must first be computed with
object <- add_loo(object, unit = unit). This method extracts the
stored PSIS object and does not compute LOO. It does not compute weights
across models. For predictive model weights, use
loo_model_weights(model_1, model_2); for a model comparison table,
use loo_compare(model_1, model_2).
When the score integrates local random effects, these weights reweight global-parameter draws. They must not be applied to full-data sampled local effects as though those were draws from the deletion posterior; local-effect calculations require the corresponding retained-data reconstruction.