Calculate LOO-CV for a Because Model — because

Calculates Leave-One-Out Cross-Validation using Pareto Smoothed Importance Sampling (PSIS-LOO) for a fitted Because model.

Usage

because_loo(model, ...)

Arguments

model: A fitted model object of class "because" returned by because. Note: If the model was not fitted with WAIC = TRUE (so log_lik is missing), this function will automatically refit the model (using a short MCMC run) to calculate the likelihoods.
...: Additional arguments passed to loo::loo().

Value

A loo object containing:

estimates: Table with ELPD (expected log pointwise predictive density), LOO-IC, and p_loo
diagnostics: Pareto k diagnostic values for each observation
pointwise: Pointwise contributions to LOO-IC

Details

LOO-CV (Leave-One-Out Cross-Validation) uses Pareto Smoothed Importance Sampling to approximate leave-one-out predictive performance without refitting the model N times. This is particularly useful for:

Model comparison when models have different numbers of latent variables
Identifying influential observations (via Pareto k diagnostics)
Robust predictive performance assessment

Pareto k diagnostics:

k < 0.5: Excellent (all estimates reliable)
0.5 < k < 0.7: Good (estimates okay)
0.7 < k < 1: Problematic (estimates unreliable)
k > 1: Very problematic (refit model excluding these observations)

Note on implementation: This function extracts the pointwise log-likelihoods calculated by the JAGS model (monitored as log_lik[i]) when WAIC = TRUE. It does not re-compute likelihoods from posterior samples in R, ensuring consistency with the fitted model structure.

Examples

if (FALSE) { # \dontrun{
  fit <- because(data, tree, equations, WAIC = TRUE)
  loo_result <- because_loo(fit)
  print(loo_result)

  # Check for problematic observations
  plot(loo_result)

  # Compare models
  loo_compare(loo_result1, loo_result2)
} # }