credsubs: Multiplicity-Adjusted Subset Identification

Abstract

Subset identification methods are used to select the subset of a covariate space over which the conditional distribution of a response has certain properties - for example, identifying types of patients whose conditional treatment effect is positive. An often critical requirement of subset identification methods is multiplicity control, by which the family-wise Type I error rate is controlled, rather than the Type I error rate of each covariate-determined hypothesis separately. The credible subset (or credible subgroup) method provides a multiplicity-controlled estimate of the target subset in the form of two bounding subsets: one which entirely contains the target subset, and one which is entirely contained by it. We introduce a new R package, credsubs, which constructs credible subset estimates using a sample from the joint posterior distribution of any regression model, a description of the covariate space, and a function mapping the parameters and covariates to the subset criterion. We demonstrate parametric and nonparametric applications of the package to a clinical trial dataset and a neuroimaging dataset, respectively.