Estimating optimal surrogate endpoints by machine learning and targeted minimum loss-based estimation in two-phase sampling studies

Abstract

This dissertation provides contributions in two areas: the application of TMLE in estimation of an optimal surrogate and implementation of inverse probability of censoring weighted targeted minimum loss-based estimation (IPCW-TMLE). In Chapter 1 we develop methodology for the estimation of optimal surrogates in randomized trials using targeted minimum loss-based estimation (TMLE), first in the setting of complete data, and then in Chapter 2, extended to the setting of two-phase data, seeking to make the methodology more applicable to real randomized trials. In Chapter 3 we present a comparison of IPCW-TMLE to a commonly used method of Breslow and Holubkov for parameter estimation in two-phase studies. The simulation study presented assesses the comparative differences in bias and efficiency of estimates obtained by both methods. In Chapter 4, IPCW-TMLE is elaborated for estimation of causal parameters of interest in right-censored two-phase studies. The methods developed in this dissertation have broad application to randomized clinical trials with two-phase designs for measuring biomarkers. Many of the methods described in this dissertation are illustrated with application to two dengue phase 3 vaccine efficacy trials.