This release focuses on optimization, code hardening, and improved handling of beta-binomial models.
Major Features & Behavioral Changes
- Model Averaging: Introduced an
averagingargument inbestModel()to compute final coefficients as a BIC-weighted average across consistency- screened models. This reduces model selection variance and guarantees consistent models. - Analytic Consistency Checks: Replaced the discrete grid check with an
exact analytical monotonicity check using
polyroot()(now the default incheckConsistency()). This detects narrow violations that a discrete grid can miss and is faster. - Consistency Screening: Flat candidate models (independent of L) are now correctly flagged as inconsistent, and screening evaluates 8 age points (up from 2) to catch intersecting percentile curves.
- Plotting: Discrete beta-binomial quantiles are now correctly rendered as
step functions in
plot()andcompare().
Deprecations
- Deprecated
subsample_lm()and thesubsamplingargument inbestModel(). Subsampling OLS coefficients adds Monte-Carlo noise without improving the fit and is entirely replaced by the newaveragingmethod.
Bug Fixes & Performance Optimizations
- Fixed issues causing
NaNRMSE values (switched from.lm.fit()tolm.fit()) and crashes during consistency screening with non-Taylor predictors. - Fixed beta-binomial probability normalization over truncated supports and resolved L-BFGS-B optimization convergence failures.
- Fixed an edge case failure in
checkConsistency()for conventional norming (minA1 == maxA1). - Significant performance improvements in beta-binomial modeling via
precomputation of
lchoose()and grouped-data optimizations inpredict().
- local WIN11, 64Bit install, R 4.6.0
- winbuilder win release, win old release, win development
- Automatic checks on GitHub: Ubuntu (old-rel1, devel, release), MacOS latest, Windows latest
There were no ERRORs, WARNINGs or NOTEs.
There are currently no downstream dependencies for this package.