Nonparametric methods for integration of survival analysis and machine learning

Abstract

This dissertation develops practical methodology incorporating modern machine learning techniques into statistical inference, with a particular focus on the analysis of time-to-event data. Time-to-event data are commonly encountered in biomedical studies, where incomplete follow-up and truncation-induced sampling bias may preclude the use of standard analysis procedures. The primary intended application of this work is variable importance, although the methods developed here are appropriate for a wider range of problems. Chapter 1 serves as an introduction to the dissertation. The three methodological chapters overlap but function as distinct, standalone units. In Chapter 2, we propose an algorithm-agnostic, nonparametric procedure for assessing variable importance for right-censored time-to-event outcomes. In the Chapter 3, we develop a framework in which arbitrary machine learning algorithms can be applied to estimate personalized survival curves from data subject to both censoring and truncation. Chapter 4 addresses the use of sample splitting to provide inference on variable importance when the true importance lies on the boundary of the parameter space.