NEWS

---

IBGS 1.1.1

Bug fixes

• A batch-scored acceptance whose exact re-fit is singular now recovers its own one-step fit. The batched scorer for the binomial and poisson families derives a proposal's criterion from the bordered panel without fitting coefficients. When such a proposal was accepted and its exact re-fit turned out singular, the sampler committed the coefficient vector left over from the last successfully fitted model — a different model's coefficients, mapped positionally onto the accepted model's columns — and every subsequent one-step score warm-started from that state until the next successful re-fit. The sampler now recovers the accepted model's own one-step fit through the same call the per-proposal path makes, keeping the convention that an iterative family falls back to its one-step score and coefficients; if even the one-step system is singular the move is rejected. The condition requires a candidate to pass the batch's relative pivot margin while the enlarged model fails the absolute pivot floor — for example a genuinely predictive column of very small magnitude — and cannot arise for the gaussian family, whose batched acceptances were already revoked on a singular re-fit. Output is unchanged on every design where the condition never fires. A regression test (tests/test-refit.R) pins the corrected behaviour.

IBGS 1.1.0

Performance

An optimisation pass over the Metropolis-within-Gibbs sweep. No interface changes and no change to what any criterion reports: the samplers do the same search without repeating work that a single-coordinate proposal cannot have changed.

• The active-column list is spliced rather than rebuilt. The sweep used to re-derive its sorted list of active columns from the inclusion vector on every proposal, a scan over all p candidates, although a single-coordinate flip changes that list in one position. The coordinate is now spliced in or out at its sorted position, making the per-proposal bookkeeping linear in the model size instead of in p. Applied to the gaussian/binomial/poisson sweep and to the linear-mixed sweep, which carried a duplicate of the same code.
• The gaussian sampler scores every candidate addition in one batch. Adding a column borders the current Gram sub-block, so its Cholesky factor extends the current one by a single row and column and the only per-candidate work is one triangular solve; stacking all candidates turns that into a single BLAS-3 dtrsm. A sweep visits every candidate but changes the model only when a move is accepted, which for a criterion that fits the data is well under once per sweep, so one factorisation serves the whole sweep.
• The binomial and poisson samplers share the first Newton step's weights. A proposed column warm-starts at zero, so the step's linear predictor is the current model's whatever column is being added, and the mean, IRLS weight and working response derived from it are the same for every candidate. They are now computed once per accepted move rather than once per proposal.
• Their normal equations are batched on the same principle. In those fixed weights, adding a column borders the current model's weighted system exactly as in the gaussian case, so one Cholesky and two triangular solves give every candidate's coefficients, and each candidate's linear predictor follows as a rank-one update of the current model's. Unlike the gaussian case the criterion does not fall out of the algebra: the deviance still needs the proposal's own linear predictor and a pass over the observations.

Each batch is used only where it repays its own construction and is well conditioned: for models at most half the size cap, for at most 64 coefficients, for candidate columns not already explained to within 1e-3 by the current model, and -- for the iterative families, whose panel is more expensive to build -- only while the squared model size stays below the candidate count, which keeps it out of the short block-screening sweeps of glmIBGS(). Every removal, the Cox sampler, and any run gated out by those conditions keep their existing path.

Reproducibility of this release

The spliced list and the shared IRLS weights are bit-identical by construction. The spliced list is in the same ascending order the rescan produced, so every candidate fit sees the same matrix, and the cached weights are the values the per-proposal code computed.

The two batched scorers compute the same quantities by a different arithmetic path and so are not guaranteed bit-identical, although in testing they were. Accepted models are re-fitted directly, as the iterative families already were, so every recorded criterion and every collinearity verdict still comes from the same factorisation as before and only an accept/reject decision can move -- and only where the two agree to within rounding of a Metropolis threshold, after which the chain is a different realisation of the same target.

Internal

• Fixed a latent double fit in the accept path: a proposal both scored by one IRLS step and scored from a batch would have been re-fitted twice. The two re-fit branches are now one, keeping each family's convention -- the gaussian batch rejects a model the direct fit will not take, since it needs a usable factor to continue, while an iterative family keeps the one-step score as it always has. No released version could reach this, batched scoring having been gaussian-only and gaussian never taking the one-step path.

IBGS 1.0.1

Model selection criteria

• ebic.gamma now defaults to 1, the value Chen & Chen (2012) show to be selection-consistent for any p = O(n^kappa); it was 0.5, which qualifies only while p < n and so was below the threshold in the ultrahigh-dimensional problems the package targets.
• A supplied ebic.gamma is held in [max(0, 1 - log(n)/(2*log(p))), 1] — a value below that consistency floor is raised to it, one above 1 lowered, with a message reporting the change. The floor uses the event count for the Cox samplers, matching the log(d) in their own penalty.
• The exBIC documentation now cites Chen & Chen (2012), Extended BIC for small-n-large-P sparse GLM, whose 2 nu(s) gamma log P form is the one implemented; the 2008 Biometrika paper it previously cited uses the binomial-coefficient form. Both are listed in the package references.

Sampler controls

• start = c("null", "full") removed. Every chain starts from the empty model, which was already the default and avoids the ill-conditioned saturated start where the criterion (notably AICc) can be degenerate.
• Fixed: the final long run of coxIBGS() and lmeIBGS() ignored start and began from the full candidate set. Under heavy censoring that set can reach the event count d, where the AICc correction 2*npred*(npred+1)/(d - npred - 1) changes sign and a model saturated in the events scores far below any honest one, trapping the chain. Those runs now start empty like every other chain.
• fast removed from glmGibbs()/glmIBGS(). Binomial and poisson proposals are always scored with one warm-started IRLS step, with accepted models re-fitted to full convergence — the former default, and strictly faster than the alternative for the same reported criteria.

IBGS 1.0.0 (2026-07-04)

Initial CRAN release.

Variable selection

• Iterated block Gibbs samplers with screen–select–threshold refinement (glmIBGS(), coxIBGS(), lmeIBGS()) and matching plain block Gibbs samplers (glmGibbs(), coxGibbs(), lmeGibbs()) for ultrahigh-dimensional problems.
• Families: gaussian, binomial and poisson generalized linear models (least squares / iteratively reweighted least squares), the Cox proportional-hazards model (Efron partial likelihood), and linear mixed models.
• Model selection by AIC, BIC, AICc or extended BIC (exBIC).

Sampler controls

• start = c("null", "full") sets the initial model of the Gibbs chain(s); the default "null" starts from the empty model and grows, avoiding the ill-conditioned full-model start.
• permute = TRUE (the default) draws a fresh random coordinate permutation (Fisher–Yates) each Gibbs sweep, so every coordinate is updated exactly once per sweep; permute = FALSE restores a fixed in-order sweep.
• glmGibbs()/glmIBGS() accept an opt-in fast = TRUE for the binomial/poisson families, scoring each single-coordinate proposal with one warm-started IRLS step and re-fitting only accepted models to full convergence (reported criteria and coefficients stay exact).
• Parallel block screening through 'OpenMP' (n.cores) and an optional near-collinearity guard (cor.check).

Model averaging

• Each fit retains the best n.models models, summarized in C so the returned object stays compact even for thousands of predictors.
• predict(), fitted() and coef() average over the retained models with smooth-SIC (BMA-style) weights, on the link or response scale; a single retained model can be selected with average = FALSE. Conditional prediction with random-effect BLUPs is available for lme fits.

Object methods and diagnostics

• print() and summary() for the fit, with selected-variable and top-model tables and a convergence-diagnostics block.
• plot() dispatches to the diagnostics, also exported individually: plotICtrace() (criterion trace), plotMargProb() (marginal inclusion probabilities as a dot-and-whisker plot), plotModelFreq() (top-model visit frequencies), plotGelman() and plotAutocorr(). plotMargProb() and plotModelFreq() accept horizontal = TRUE for a horizontal layout.
• Convergence diagnostics computed in C: the Gelman–Rubin shrink-factor evolution, the Geweke z-statistic, the effective sample size, and the autocorrelation of the criterion trace.

C backend

• The sampler is implemented in C with parallel block screening through 'OpenMP'.