New Algorithms and Software for Significance Controlled Variable Selection

Abstract

Stepwise regression algorithms have been widely used for a variety of applications and continue to be a fundamental tool in variable selection. Most functions available in statistical software packages deliver models that may contain insignificant predictors because of the criterion of the optimization at each step. Here we introduce an R package that provides the user with several measures of the prospective model at each step of the algorithm. These prospective models are checked with multiple testing p-value corrections such as Bonferroni and False Discovery Rate and hence the algorithm’s final model includes only predictors that have their significance controlled by the choice of correction type and alpha level. Moreover, the steps forward or backward can have an entry or drop criterion that is a combination of the p-values of prospective models. We illustrate the functionality of the package with examples and simulations.