roc_auc() now handles tied scores correctly. The previous version accumulated one edge at a time in sorted order, so tied edges (e.g. PIPs of exactly 0 or 1) were ordered by matrix position. That made the AUC depend on variable order and could bias it by several hundredths. The curve now steps only at distinct scores, and the AUC equals the Mann-Whitney statistic. With no tied scores, the AUC is unchanged.
roc_auc() returns NA for auc (instead of an error) when the truth has no edges or no non-edges.
confusion_at_threshold() returns a list (was a named numeric vector) with precision and F1 added, and its third argument is now called threshold (was thr), with default 0.5. Use cm$TPR rather than cm["TPR"]. plot_recovery() returns these lists.
plot_roc() now uses base graphics and returns the ROC list invisibly (it previously returned a ggplot object). It gains add = TRUE to overlay curves.
precision_to_pcor() returns a matrix of NA with a warning (instead of an error) for a non-positive diagonal, and keeps dimnames.
New methods for the ggmtools generics pip_table() (PIPs as a long edge table) and pcor_estimate() (posterior-mean partial correlations, one matrix per group). Scripts written with these work on spikeyglass fits too.
Bug fixes
SSVS and multi-platform samplers: the spike-and-slab update of the graph similarity theta (inherited from the Matlab reference) drew its proposal with beta_prop as a scale but used it as a rate in the acceptance ratio, and had the sign of the (alpha - alpha_prop) * log(theta) term reversed for spike-to-slab moves. The platform similarity Phi had the same sign error and mixed scale and rate in its normalizing terms. The ratios are now computed from explicit gamma densities (as the G-Wishart sampler already did). Default SSVS fits are essentially unchanged (beta_prop = 1, w = 0.9); fits with smaller w, beta_prop != 1, or the multi-platform sampler will differ. beta_prop is now a rate for every sampler.
A partial hyper list no longer wipes out the other defaults for the G-Wishart and SSVS samplers. Unknown hyperparameter names are now an error.
simulate_multiggm(K = 1) returned two groups.
simulate_multiggm_platform() now shares one base graph across platforms (it drew an independent graph per platform); the data-generating process is documented, and the base and platform graphs are returned.
method = "ssvs_platform" now stops when nchains > 1 and warns that parallel is ignored, instead of silently running one chain.
Removed ineffective EIGEN_USE_BLAS / EIGEN_USE_LAPACKE defines from the mGHS sources (no change to results).
Repository and CRAN hygiene: compiled objects untracked, old demo moved to dev/, Depends: R (>= 4.0.0), BugReports, ASCII-only sources.
New tests for each fix, plus reproducibility tests (same seed, parallel versus serial, R versus C++ engine).
multiGGM 0.4.0
New features
multiggm_mcmc(method = "mghs") fits the multiple graphical horseshoe (Busatto & Stingo 2025), adapted from github.com/cbusatto/mGHS. It returns a multiggm_fit, so coef(), fitted(), posterior_pcor(), posterior_ci(), diff_prob_pcor(), plot_trace("theta") and multi-chain fitting work. It reproduces upstream’s draws exactly for the same seed (inst/mghs/VALIDATION.md). Local fixes: rejection samplers that could hang, stuck checks that missed NaN, and a cut-model indexing bug (inst/mghs/UPSTREAM.md). New options: thin, save_draws = FALSE (keep only running summaries), automatic restart of stuck chains.