Added support for the Cox proportional hazards model via
family = "cox". Feature screening for this family uses a
new (iterative) sure independence screening routine driven by the
extended BIC of the marginal Cox fits.
y may now be supplied for the Cox family either as a
survival::Surv object or as a two-column
(time, status) matrix or data frame.
Added a robust Median-of-Means gradient-descent (MoM-GD)
estimator, selected with robust = "mom", for heavy-tailed
responses. It supports the Gaussian and Negative-Binomial
(family = "negative.binomial") families and reports basic
bootstrap confidence intervals.
New arguments accompanying the robust estimator:
nblock — number of Median-of-Means blocks (default
5).nboot — number of bootstrap resamples for the
confidence interval (default 200).theta — fixed Negative-Binomial dispersion; estimated
from the screened design when NULL.New argument CI controls whether confidence
intervals are computed. The default depends on the estimator:
TRUE for robust = "none" (closed-form
intervals are cheap) and FALSE for
robust = "mom" (the bootstrap is not run unless requested).
When CI = FALSE, only point estimates are returned and the
interval bounds are NA.
Added two example datasets: data_cox_model and
data_nb_model.
Expanded the documentation for mglm.test() to
describe the accepted forms and constraints of every argument, and added
runnable examples for the Cox model and for robust MoM
estimation.
The package description no longer advertises p-values; the function reports information-gain point estimates and confidence intervals.
Removed the dependency on tidyverse, which was imported in full but not necessary. This considerably reduces the installation footprint.
Added explicit imports of stats::BIC(),
stats::sd(), and survival::Surv().
Added testthat (>= 3.0.0) to
Suggests and enabled testthat edition 3.
Substantially expanded input validation with informative error messages. X, y, mod.idx, family, robust, penalty, tune, CI.type, iter, trace, conf.level, lambda, nlambda, nboot, nblock, and theta are all checked for type, shape, finiteness, and admissible ranges. mod.idx must now be a non-empty list of disjoint, in-range column indices.
Fixed the handling of a user-supplied lambda sequence or confidence levels.
Initial release.
mglm.test() estimates the information gain
contributed by individual modalities in multimodal regression models,
measured by Expected Relative Entropy (ERE) or a pseudo-R-squared, with
confidence intervals. Supports the Gaussian and Binomial families with
canonical links.
summary() method for objects of class
mglm.test.
Example datasets data_linear_model and
data_logistic_model.