fbrglm
A formula-based wrapper around 'glmnet' that brings the 'glm()'-compatible modeling workflow to regularized generalized linear models. Training-time 'terms', 'xlevels', and 'contrasts' are stored on the fit object and reused at predict time, so the design matrix is reconstructed consistently across sessions. Complete-case bookkeeping is exposed via 'nobs_info', and linearly dependent columns are detected by a QR pivot and reported as 'NA' in 'coef()' and 'summary()' (the 'stats::glm()' convention), distinguishing "not identifiable" from "shrunk to zero by the penalty". Novel factor levels at predict time raise the same error 'stats::predict.glm()' does by default, with 'on_new_levels = "na"' as a production-style opt-in. Accepts character family strings ('gaussian', 'binomial', 'poisson', 'cox', 'multinomial', 'mgaussian') and any 'glm' family object the underlying 'glmnet' itself accepts, including 'Gamma' and fixed-theta negative binomial via 'MASS::negative.binomial'.
This page was built for software: fbrglm
