Expand description
§solow-bayes
Bayesian generalized linear mixed models fit by mean-field variational
Bayes (deterministic; no MCMC). Two families are provided with a
random-effects design (exog_vc): BinomialBayesMixedGLM (logit link) and
PoissonBayesMixedGLM (log link). Both are fit by BayesMixedGlm::fit_vb,
which maximizes the evidence lower bound (ELBO) of a factored Gaussian
variational posterior over the fixed effects, the variance-component
parameters (log standard deviations of the random effects), and the random
effect realizations.
The parameterization, ELBO and its gradient mirror the canonical
statistical-computing reference’s genmod.bayes_mixed_glm module exactly, so
the posterior means (BayesMixedGlmResults::fe_mean,
BayesMixedGlmResults::vcp_mean, BayesMixedGlmResults::vc_mean) and the
ELBO (BayesMixedGlmResults::elbo / BayesMixedGlmResults::llf) agree to
optimizer tolerance.
§Model
For observation i the linear predictor is
eta_i = x_i · fe + z_i · vc, where x is exog (fixed effects) and z is
exog_vc (random effects). Each random effect realization vc_j is
Gaussian with mean zero and standard deviation exp(vcp[ident[j]]); the
vcp parameters (log standard deviations) have a N(0, vcp_p²) prior and the
fixed effects a N(0, fe_p²) prior.
Structs§
- Bayes
Mixed Glm - A Bayesian mixed GLM with a random-effects design, fit by variational Bayes.
- Bayes
Mixed GlmResults - Posterior summary returned by
BayesMixedGlm::fit_vb. - Binomial
Bayes MixedGLM - A binomial (logit-link) Bayesian mixed GLM. Thin constructor over
BayesMixedGlm. - MapResult
- Result of a MAP (posterior-mode) fit, returned by
BayesMixedGlm::fit_map. - Poisson
Bayes MixedGLM - A Poisson (log-link) Bayesian mixed GLM. Thin constructor over
BayesMixedGlm.
Enums§
- Family
- The GLM family for the Bayesian mixed model.