Added Gamma, inverse-Gaussian, and negative-binomial regression. GLM families can now be supplied as names, functions, or family objects, including alternative links.
Added ordinal regression through ordinal_family() and multinomial regression through multinomial_family(), with prediction and plotting support.
Added parametric survival regression through survreg_family() and Fine–Gray competing-risks regression through finegray_family().
Added generalized estimating equations through gee_family(), with cluster identifiers supplied through id and optional measurement order through waves. GEE models default to QICu-based selection.
Added mfpi() for interaction analysis with a categorical grouping variable, including prediction and plotting methods for group-specific functions and differences.
Added spike-at-zero modelling and options for modelling positive predictor values with or without a separate zero indicator.
Added support for assessing categorical predictors as complete terms, with term_groups available in the default interface.
Added force_max_fp_vars to retain specified predictors at their maximum permitted FP degree.
Added plot.mfp2(). fracplot() is deprecated and will be removed in a future
release; use plot() with a fitted mfp2 object instead.
Added the fitter argument to choose between stats::glm.fit() ("base")
and the optional fastglm backend for GLM families.
The following low-level helper functions are no longer exported and are now internal: apply_shift_scale(), assign_df(), center_matrix(), create_dummy_variables(), find_scale_factor(), find_shift_factor(), fit_acd(), generate_powers_acd(), generate_powers_fp(), transform_matrix(), and transform_vector_acd(). Shifting, scaling, centring, and ACD transformations are controlled through the shift, scale, center, and acd_vars arguments of mfp2(), or through shift, scale, center, and acd in fp() and fp2(). transform_vector_fp() remains exported.
mfp2() now warns when it is given arguments it does not use, such as a misspelled argument name.
In the formula interface, named values for top-level df, alpha, select, shift, scale, or center are now an error; previously a value such as shift = c(age = 5) was silently applied to every predictor.
Variable-specific df, select, alpha, shift, and scale settings in the default methods now use named vectors matched to predictor columns. Partial named vectors are supported; unnamed vectors containing more than one value are rejected.
Cox models now default to ties = "efron" rather than "breslow", and to nocenter = c(-1, 0, 1) rather than NULL. The "exact" tie option is no longer accepted.
Observation weights must be finite and strictly positive; zero weights are no longer accepted.
ACD settings now use acd_vars in the default interface and acd = TRUE within fp() or fp2(). The former acdx spelling remains available as a deprecated alias.
The direct strata argument to mfp2.formula() is deprecated in favour of strata(...) in the formula. For formula-stratified models, prediction data must contain the original stratification variables.
Fixed binomial response predictions for new observations.
Corrected formula evaluation during subsetting and improved handling of unused factor levels. Removing levels from a factor with custom contrasts now produces an informative error.
Improved support for namespace-qualified terms such as mfp2::fp() and mfp2::fp2().
Updated the final regression coefficients so that they are reported on the original scale of the data.
Added the nseq argument to predict.mfp2().
Improved the package documentation and added references concerning the influence of influential observations on fractional polynomial functions.
Fixed an issue in predict.mfp2() when type = "terms" was used.
Fixed an issue in mfp2() where the keep argument was not applied when criterion = "aic" or criterion = "bic".
Corrected the BIC calculation for Cox models so that the effective sample size is based on the number of observed events rather than the total number of observations.