LoginSign Up
← Back to all sparks
B

brglm2

ANALYTICS
Velocity0.0

R package brglm2 by ikosmidis — release notes from GitHub.

github.com/ikosmidis/brglm2 ↗

A bias-reduction package reaches 1.0 by adding an estimator built for high-dimensional logistic regression

r-packageregressionbias-reductionhigh-dimensionalestimation-methods
◆Current state
brglm2 fits generalized linear models using mean and median bias reduction rather than plain maximum likelihood, which matters most when ML estimates are infinite or badly biased. The 0.7-0.9 line broadened coverage — negative binomial via brnb(), ordinal superiority measures, the expo() method for exponentiated parameters, add1()/drop1() so step() stops silently producing nonsense. Version 1.0.0 in August 2025 added mdyplFit(), estimating logistic regression by maximum Diaconis-Ylvisaker prior penalized likelihood with optional high-dimensional corrections. The two releases since have tuned that new path.
◆Where it's heading
The package's older work assumed the classical regime where observations comfortably outnumber parameters. mdyplFit() and its hd_correction argument target the opposite case, and the follow-up releases are almost entirely about it — Pearson residuals on original responses, aliased parameter handling, the sloe() signal-strength estimator ignoring leverage-one observations. Meanwhile the older surface gets graceful-failure work: brglm_fit() now returns its latest estimates with warnings rather than aborting.
◆Prediction
Given that 1.0.1 and 1.1.0 are both dominated by mdyplFit follow-ups while the classical path receives only robustness fixes, further work on high-dimensional corrections is the likeliest direction.

◆Recent moves

  1. 3mo ago

    brglm2 v1.1.0

    Fixes ordinal superiority measures under interactions, extends ordinal_superiority.bracl() to adjacent category logit models with or without proportional odds, and makes brglm_fit() fail gracefully by returning its latest estimates with warnings. Released the same day as sibling package detectseparation v0.4, part of a coordinated pass across the maintainer's packages.

    View source ↗
  2. 8mo ago

    brglm2 v1.0.1

    A consolidation release for the mdyplFit path: Pearson residuals computed on original responses, a fix for high-dimensional corrections with aliased parameters, and sloe() ignoring leverage-one observations when estimating contaminated signal strength. Almost every item concerns the estimator introduced one version earlier.

    View source ↗
  3. 11mo ago

    1.0.0 adds maximum DY-prior penalized likelihood for logistic regression

    ⚡ SPARK

    Adds mdyplFit() as a glm() method, estimating logistic regression by maximum Diaconis-Ylvisaker prior penalized likelihood with optional high-dimensional corrections to estimates, standard errors and z statistics. It marks the point where the package stops being solely about classical bias reduction, and everything released since has been follow-up work on this path.

    View source ↗
  4. 1y ago

    brglm2 v0.9.3

    Adds add1() and drop1() methods for brglmFit objects so step() no longer silently falls back to glm methods and returns meaningless results, and improves brnb() behavior when ML estimates are infinite. A quiet correctness fix with real consequences for anyone doing automated model selection.

    View source ↗
  5. 1y ago

    brglm2 v0.9.2

    Switches the convergence criterion for brglm_fit iterations from the L1 to the L-infinity norm of the step size, so the epsilon tolerance directly determines the precision of reported estimates instead of scaling with parameter count. A small change that makes a control parameter mean what users assume it means.

    View source ↗
  6. 3y ago

    brglm2 v0.9.1

    Bundles 0.9.0 and 0.9.1, whose centrepiece is the expo() method for estimating exponentiated GLM parameters with bias corrections — odds ratios and multiplicative Poisson effects estimated directly rather than by transforming a corrected estimate. Datasets are added to support the examples.

    View source ↗