We’re pleased to share that our methodological work on variable selection with random forests has now been published in BMC Bioinformatics.
The paper introduces shadowVIMP, a novel variable selection method based on random forest importance measures and permutation testing, designed with high-dimensional data settings in mind. The method addresses known challenges of permutation-based random forest variable importance measures, particularly biases caused by correlated and categorical variables.
shadowVIMP was developed by our colleagues Tim Müller, Dr. Hannes Buchner, Oktawia Miluch, and Maria Blanco in cooperation with Dr. Roman Hornung (Ludwig Maximilian University of Munich) and Prof. Dr. Silke Szymczak (University of Lübeck).
By combining random forest variable importance with a multiple testing framework and preserving the correlation structure between variables, the method enables more reliable and interpretable variable selection results.
In extensive simulation studies, shadowVIMP showed advantages over competing approaches, particularly in high-dimensional settings, improving sensitivity while maintaining multiple testing control. The method was also illustrated using a real-world application on Alzheimer’s disease data.
The accompanying open-source R package shadowVIMP is available on CRAN and allows users to apply the method directly in practice.
This work is one example of how we invest in methodological development to address challenges arising in real-world data analysis projects.
Link to R-package on CRAN: https://cran.r-project.org/web/packages/shadowVIMP/index.html
