Compute the Widely Applicable Information Criterion (WAIC) for brma model objects and store the result in the object.
Details
WAIC is an alternative to LOO-CV for estimating out-of-sample predictive accuracy. Like LOO, it evaluates expected predictive performance for new observations.
In most cases, LOO-PSIS (via add_loo) is preferred over WAIC
because it provides better estimates and includes diagnostics (Pareto k
values) that indicate when the approximation may be unreliable.
For Gaussian multilevel and known-V brma.mv() models,
estimate-unit WAIC uses the same integrated conditional log-likelihood matrix
as estimate-unit LOO, \(p(y_i \mid y_{-i}, \theta)\). Estimate deletion
retains the fitted grouping and dependency structure. Weighted Gaussian
cluster models integrate the powered row likelihood, with its normalizing
constant, over the retained-data cluster-effect distribution. Earlier stored
weighted estimate-unit WAIC results require recomputation with add_waic().
For brma.mv() models with known random-effect group covariance
R, the known R matrix and fitted random-effect covariance enter
through BayesTools' metadata-defined marginal \(ZGZ'\) covariance. The same
four-field comparison policy applies across sampled, marginalized, and mixed
local-effect representations.