fibr: Prior-Fraction Diagnostics for Hierarchical Models
Computes the prior fraction, the per-group pooling or shrinkage
factor, for hierarchical models, including directly from 'brms' fits. For
each group-level coefficient the prior fraction is the share of the
posterior precision contributed by the shrinkage prior relative to the
likelihood; values near one indicate a coefficient that is prior-dominated
(the centring/non-centring funnel regime), values near zero indicate a
likelihood-dominated coefficient that is well identified from the data.
These quantities are invisible to standard convergence diagnostics such as
R-hat and effective sample size, and they indicate where a non-centred
reparameterisation is likely to help. A companion advisor reports the same
decomposition for changepoint random effects fitted with 'smoothbp'. The
underlying geometry (the Fisher-metric connection on the base-fiber split,
for which this connection is flat so the obstruction is statistical rather
than geometric) is described in Bindoff (2026)
<doi:10.5281/zenodo.20724550>; code reproducing the paper is in the
package's source repository.
Documentation:
Downloads:
Linking:
Please use the canonical form
https://CRAN.R-project.org/package=fibr
to link to this page.
[aut, cre]