Deletion/Substitution/Addition Algorithm for Partitioning the Covariate Space in Prediction

Annette Molinaro, National Cancer Institute, National Institutes of Health
Mark J. van der Laan, Division of Biostatistics, School of Public Health, University of California, Berkeley

Abstract

We propose a new method for predicting censored (and non-censored) clinical outcomes from a highly-complex covariate space. Previously we suggested a unified strategy for predictor construction, selection, and performance assessment. Here we introduce a new algorithm which generates a piecewise constant estimation sieve of candidate predictors based on an intensive and comprehensive search over the entire covariate space. This algorithm allows us to elucidate interactions and correlation patterns in addition to main effects.