We introduce a novel statistical concept, called a supervised distance matrix, which quantifies pairwise similarity between variables in terms of their association with an outcome. Supervised distance matrices are derived in two stages. First, the observed data is transformed based on particular working models for association. Examples of transformations include residuals or influence curves from regression models. In the second stage, a choice of distance measure is used to compute all pairwise distances between variables in the transformed data. We present consistent estimators of the resulting distance matrix, including an inverse probability of censoring weighted estimator for use with right-censored outcomes. Supervised distance matrices can be used with standard (unsupervised) clustering algorithms to identify groups of similarly predictive variables and to discover subpopulations of related samples. This approach is illustrated using simulations and an analysis of gene expression data with a censored survival outcome. The proposed methods are widely applicable in genomics and other fields where high-dimensional data is collected on each subject.
