This repository contains code to reproduce the analyses in "A general framework for inference on algorithm-agnostic variable importance" by Williamson, Gilbert, Simon, and Carone (Journal of the American Statistical Association (Theory & Methods), 2021). All analyses were implemented in the freely available R programming language; specifically, version 4.0.2. All analyses use the R package vimp version 2.2.3.
This README file provides an overview of the code available in the repository.
We have separated our code further into two sub-directories based on the two main objectives of the manuscript:
- Numerical experiments to evaluate the operating characteristics of our proposed method under varying data-generating mechanisms (
sims). - An analysis of an antibody against HIV-1 infection (
data_analysis).
All analyses were performed on a Linux cluster using the Slurm batch scheduling system. The head node of the batch scheduler allows the shorthand "ml" in place of "module load". If you use a different batch scheduling system, the individual code files are flagged with the line where you can change batch variables. If you prefer to run the analyses locally, you may -- however, these analyses will then take a large amount of time.
If you encounter any bugs or have any specific questions about the analysis, please file an issue.