Why you should avoid using multiple Fine–Gray models: insights from (attempts at) simulating proportional subdistribution hazards data

Authors: Edouard F. Bonneville, Liesbeth C. de Wreede, and Hein Putter

Abstract

Studies considering competing risks will often aim to estimate the cumulative incidence functions conditional on an individual’s baseline characteristics. While the Fine–Gray subdistribution hazard model is tailor-made for analysing only one of the competing events, it may still be used in settings where multiple competing events are of scientific interest, where it is specified for each cause in turn. In this work, we provide an overview of data-generating mechanisms where proportional subdistribution hazards hold for at least one cause. We use these to motivate why the use of multiple Fine–Gray models should be avoided in favour of better alternatives such as cause-specific hazard models.

Usage

Two main files are of interest:

• manuscript-figures.R - R script to reproducte the figures in the manuscript.
• helpers.R - helper functions that manuscript-figures.R relies on to compute true cumulative incidence function, cause-specific hazards, and subdistribution hazards, for both events.

The R environment can be reproduced using {renv}, by calling

renv::restore()

Ideas for further extension

• Generalize functions to accommodate user-defined hazard or cumulative distribution functions, and vectorize covariate input.

• Create a shiny app to automate plot creation for different parametrizations, and show when the total failure probability exceeds one/cause-specific hazards become negative.

• Make functions allowing user to actually simulate data (instead of just checking the true values).