mfp2

Overview

mfp2 implements multivariable fractional polynomial (MFP) models and their extensions. It performs variable selection and functional form selection when modeling the relationship between a covariate matrix x and an outcome y. The function currently supports generalized linear models with families "gaussian", "binomial", and "poisson", as well as Cox proportional hazards models ("cox").

In addition to the standard MFP, it allows modeling of a sigmoid-shaped covariate effects using the approximate cumulative distribution (ACD) transformation, which cannot be achieved with standard fractional polynomial functions. It also supports semi-continuous covariates with a “spike at zero” through a dedicated two-stage selection procedure.

Compatibility with existing software packages

mfp2 closely emulates the functionality of the mfp and mfpa packages in Stata. It extends the functionality of the existing mfp package in R by providing:

• both matrix and formula interfaces for input data,
• sigmoid transformations via the approximate cumulative distribution (ACD) transformation,
• support for covariates with a spike at zero,
• estimation and plotting of contrasts and partial linear predictors to investigate and visualize non-linear effects, and
• various computational optimizations to improve speed and usability.

Installation

# Install the development version from GitHub
# install.packages("pak")
# pak::pak("EdwinKipruto/mfp2")
install.packages("mfp2", repos = c('https://edwinkipruto.r-universe.dev', 'https://cloud.r-project.org'))

# or 
# install.packages("remotes")
remotes::install_github("EdwinKipruto/mfp2")

References

To learn more about the MFP algorithm, see Royston and Sauerbrei (2008), Multivariable Model-Building: A Pragmatic Approach to Regression Analysis based on Fractional Polynomials for Modelling Continuous Variables. John Wiley & Sons.

For details on the ACD transformation, see Royston (2014), A smooth covariate rank transformation for use in regression models with a sigmoid dose–response function. The Stata Journal.

For the spike-at-zero algorithm, see Becher et al. (2012), Analysing covariates with spike at zero: a modified FP procedure and conceptual issues. Biometrical Journal, 54(5), 686–700.