Extract posterior pointwise log-likelihood draws from a fitted
brma object. The result is an \(S \times K\) or \(S \times G\)
matrix, where \(S\) is the number of posterior draws, \(K\) is the
number of estimates, and \(G\) is the number of clusters.
Arguments
- object
a brma model object.
- ...
currently unused.
- unit
output unit. See
add_loo.
Value
An \(S \times K\) or \(S \times G\) matrix of pointwise
log-likelihood draws with RoBMA_target metadata.
Details
This function returns posterior log-likelihood draws, not a scalar maximized
log-likelihood. The logLik.brma method returns the same matrix
for the stats::logLik() generic; AIC() and BIC() are not
defined for it.
The log-likelihood is computed for each target at each posterior sample.
For binomial and Poisson models, each observation consists of a pair of
counts (ai/ci or x1i/x2i) that together define a single effect size estimate.
Estimate-unit columns condition on the retained observations in each
declared dependency block. For Gaussian targets after integrating their
declared local Gaussian sources, including correlated known-V
brma.mv() and specialized multilevel normal models, these are
Schur conditional scores \(p(y_i \mid y_{-i}, \theta)\); their row sum
generally is not the full joint likelihood. Likelihood powers and
their normalizing constants are retained inside the local-effect integral
for weighted Gaussian cluster models. The observation CDF used by residuals
remains unpowered. Numerically unresolved retained moments raise
RoBMA_weighted_gaussian_unavailable.
Likelihood powers and
retained-sampling context follow the declared target. Independent
observations reduce to ordinary pointwise contributions. Cluster-unit
columns delete each complete cluster jointly.
Sampling-conditioned selection models retain the declared remaining sampling
errors and random context for estimate deletion, including ordinary component
branches and constant selection weights. Cluster deletion keeps its distinct
joint target. Earlier affected estimate scores must be recomputed.
Estimate-unit GLMM draws integrate the nuisance base rate or log-rate over
its prior: point priors and untruncated beta (base rate) and normal
(log-rate) priors adaptively, other priors on a refined prior-CDF grid; a
uniform(0, 1) base-rate prior is a Beta(1, 1) prior and is integrated
adaptively. The adaptive Gauss-Hermite rules raise the quadrature order until
successive orders agree within the tolerance; this agreement is an empirical
refinement check, not an error bound. The grid is accepted only when its log-likelihood changes by less
than 1e-6 between refinements. Any nuisance prior without an adaptive route
(other than an untruncated beta base rate or an untruncated normal log-rate;
truncated ones take the grid too) can fail to reach this, especially
with few events or very large arms, and then stops with the error class
RoBMA_glmm_grid_nonconvergence described below; gamma, lognormal, and
inverse-gamma log-rate priors and lognormal base-rate priors are examples. A
beta base-rate prior or a normal log-rate prior avoids it. The automatic
computation after fitting (the autocompute.loo and autocompute.waic
options) warns instead and returns the fit without LOO/WAIC.
Numerical quadrature failures share class RoBMA_glmm_numerical_error.
AGHQ failures have class RoBMA_glmm_aghq_nonconvergence; prior-CDF refinement
failures retain RoBMA_glmm_grid_nonconvergence. Their fields identify the
outcome family, observation, and available sample/refinement diagnostics.
A beta baserate prior whose shapes are too extreme for the prior-CDF grid to
be represented in double precision is unavailable and raises
RoBMA_glmm_prior_unavailable, also of class RoBMA_glmm_numerical_error.
Cluster-unit GLMM draws are unavailable until certified nested
adaptive quadrature is implemented.