Initial CRAN release.
Formula / data interface to glmnet covering the full
glmnet family surface: the six character strings (gaussian,
binomial, poisson, cox,
multinomial, mgaussian) plus arbitrary GLM
family objects such as stats::Gamma(link = "log") and
MASS::negative.binomial(theta = ...).
Predict-time design-matrix reconstruction from the training-time
terms, xlevels, and contrasts
stored on the fit object, independent of session-level
options("contrasts"). Absent factor-level columns are
zero-padded; both main effects and interaction terms are aligned in one
pass.
Rank-deficient designs are handled in the spirit of
stats::glm(): a column-pivoted QR check on the
column-centred design matrix identifies linearly dependent columns, the
underlying glmnet fit only sees the independent subset, and
the dropped columns surface as NA in coef()
and summary() so they can be told apart from coefficients
the L1 penalty shrunk to zero.
predict() follows stats::predict.glm()
on novel factor levels in newdata by default (an error
message matching glm’s “factor g has new levels …” string). A new
predict(fit, newdata, on_new_levels = "na") opt-in is
provided for batch / production scoring pipelines: rows with unseen
levels are returned as NA and a warning naming the affected
row count is emitted, while the rest score normally.
Complete-case filtering happens through
model.frame(..., na.action = na.omit); the dropped and used
counts are exposed as fit$nobs_info and announced via a
one-line message at fit time.
A single lambda argument covers the three common
selection rules — "cv_min", "cv_1se", and
"fix" (paired with lambda_value). The chosen
numeric lambda is stored on fit$lambda_value
and reused by predict(), coef(),
summary(), and plot() so there is a single
source of truth.
S3 methods (print, summary,
predict, coef, nobs,
plot) mirror the glm() surface.
summary() follows stats::summary.glm()’s
coefficient table layout, including a glm-style header “(N not defined
because of singularities: …)” when the design was rank-deficient, and a
permanent footer explaining why no standard errors, z-values, or
p-values are reported under the current infer = "none" mode
(shrinkage bias, data-driven lambda selection, and active-set
conditioning).
Accessors as_glmnet() and
as_cv_glmnet() expose the underlying glmnet
and cv.glmnet objects for downstream tooling that consumes
them directly.
Two vignettes: vignette("fbrglm") walks the
formula/data interface, nobs_info, factor narrowing, and
offsets; vignette("fbrglm-families") walks the
family-by-family worked examples plus the piecewise-exponential survival
formulation.