Compute LOO-CV — loo_fit • bayesim

Compute leave-one-out cross-validation using Pareto-smoothed importance sampling (PSIS-LOO). Named loo_fit to avoid clashing with loo::loo().

Usage

loo_fit(fitter, fit_result, log_lik = NULL, save_psis = FALSE)

fitter: An S7 Fitter object
fit_result: A bayesim_fit_result object from fit_model()
log_lik: Optional pointwise log-likelihood matrix (S x N, draws x observations) for the training set, as returned by log_lik_matrix(). When supplied, methods should use it instead of recomputing their own; build_loo_context() passes the matrix it already computed so the weighted-prediction (PSIS) path pays for it once per task (#73). NULL (the default, and for standalone calls) means the method computes its own.
save_psis: Logical; when TRUE, methods that run the PSIS machinery for the summary should keep the fitted PSIS object and return it as psis_object (see below). build_loo_context() sets it on the weighted-prediction path so the tail smoothing runs once per task instead of twice (#76); the default FALSE discards it, as standalone callers have no use for it.

A list containing:

elpd: Expected log predictive density (scalar)
p_loo: Effective number of parameters (scalar)
elpd_se: Standard error of ELPD (scalar)
pareto_k: Pareto k diagnostic values (vector of length N)
r_eff: Chain-aware relative efficiencies used for the summary (vector of length N), or NULL when none were computed (e.g. i.i.d. draws). build_loo_context() reuses it for the PSIS object (#73).
psis_object: The fitted PSIS (importance-sampling) object the summary was derived from, or NULL. Methods that compute the summary via loo::loo(ll, r_eff, save_psis = TRUE) return its $psis_object; it is identical to loo::psis(-ll, r_eff) on the same matrix, so build_loo_context() reuses it for loo::E_loo()-weighted predictions instead of smoothing the tails a second time (#76). NULL when save_psis = FALSE or the fitter does not fit a PSIS object at all.
Additional loo-specific diagnostics