Validation in Statistics and Machine Learning - Abstract

Grömping, Ulrike

Variable Importance in Linear Models and Random Forests

In many regression contexts, with regression defined in its broadest sense, researchers are interested in variable importance, either for simplified comparison between regressors in a given model or for supporting variable selection. Variable importance metrics in linear models and random forests have a lot of common ground. The nature of variance-based variable importance metrics is investigated for the linear model, simulation comparisons to random forest variable importance are presented, and requirements for the adequacy of variable importance metrics are discussed.