News

Release notes for bgms and for this documentation site, newest first. The package’s complete change log, including breaking changes and bug fixes, is NEWS.md in the package repository.

2026-08-04: bgms 0.2.0.0 on CRAN

This release widens the model class. bgm() now fits Gaussian graphical models with variable_type = "continuous", and mixed networks of binary, ordinal, and continuous variables from a per-variable type vector. Fitted models gained predict() and simulate() methods. The C++ backend was rewritten around a shared NUTS sampler running its chains in parallel; inclusion probabilities are Rao-Blackwellized by default, and extract_inclusion_bf() reports inclusion Bayes factors that stay finite where the raw average saturates. Priors are now built with constructor functions such as normal_prior() and beta_bernoulli_prior(). New as well: prior_sensitivity_check(), which reports which edge verdicts depend on the slab scale you chose.

Several defaults changed, so a script written against 0.1.6.3 can run and give different numbers. Pairwise parameters are now reported on the association scale, about half their old values. The interaction slab defaults to normal_prior(scale = 1) in both bgm() and bgmCompare(), which also takes difference_family = "Normal" for group differences; on continuous and mixed models precision_graph_prior defaults to "hierarchical", so the prior over structures is exactly the edge prior you wrote down. iter and warmup default to 2000 and the NUTS target_accept to 0.80. The package’s NEWS.md lists every one of them with the argument that restores the old behavior.

2026-08-04: Documentation site launched

This site replaces the reference-only pages that used to live at this address. It adds a User’s Guide that explains the modelling and prior choices behind a fit, a Reference grouped by topic rather than by help file, and a Manual that documents the C++ internals for contributors.

Links to the old pages redirect here, so existing bookmarks and citations keep working.