Skip to content

v3.3.0

Latest
Latest

Choose a tag to compare

View all tags
@gcol33 gcol33 released this 22 Aug 17:07
· 15 commits to main since this release
v3.3.0
3cb4995

Breaking Changes

  • assocSelect() / corrPrune(): numeric-categorical pairs are now thresholded on eta (the correlation ratio) rather than eta-squared. Every pair type is now compared on one scale, so re-encoding a binary column between 0/1 numeric and a two-level factor no longer changes which variables survive. Mixed-type results may differ from 3.2.3 (#129).

Bug Fixes

  • modelPrune(): criterion = "condition_number" paired each predictor with the singular value at its column position, so pruning always stripped terms from the end of the formula. Predictors are now scored through the Belsley-Kuh-Welsch variance decomposition (#126).
  • modelPrune(): VIF for a multi-level factor averaged its dummy columns instead of computing a generalized VIF; now uses the Fox & Monette GVIF determinant ratio, verified against car::vif().
  • assocSelect() / corrPrune(): eta returned NA when a factor carried an unobserved level, aborting corrPrune(by = ...) on small groups. It is now defined over observed levels (#127).
  • C++ backend: subset ranking and corrPrune()'s exact-mode tie-break are now platform- and locale-independent (#128).
  • corrPrune(): kept columns are subset from the caller's original data, so a column's type no longer changes silently.
  • corrPrune(): grouped (by =) aggregation fixes -- associations are abs()-clamped before quantiling, and four silent data-loss paths in group_q handling are closed.
  • modelPrune(): single random-effect terms are no longer reinterpreted as random slopes on refit; added non-finite and zero-variance guards.

Full details in NEWS.md.

CRAN: https://cran.r-project.org/package=corrselect

Assets 3
Loading