Empirical Likelihood Quantile Regression for Right-Censored Data

Abstract

Quantile estimation of time-to-event data plays a key role in many medical applications, especially conditional on covariates of interest. In such settings, bias due to model misspecification is an important concern. As such, Empirical Likelihood (EL) is a particularly attractive estimation approach, making minimal parametric modeling assumptions without unduly compromising statistical efficiency. However, observed survival times are typically subject to right-censoring, in which case most EL approaches cannot be applied directly. In this thesis, we revisit a widely-applicable Expectation-Maximization (EM) algorithm for right-censored EL. As the covariate-free EL function becomes discontinuous in the conditional setting, we propose a continuity correction for which the computational properties of EM are retained. Several approaches to obtaining confidence intervals are explored. We provide an implementation of our method and related algorithms in the R package flexEL. The source code is written in C++ for high computational performance, and a straightforward interface allows users to fit arbitrary EL models with little programming effort.