UC Berkeley

Survival analysis, i.e. the collection of statistical methods for analyzing data on the time to an event, is widely used in the applied sciences for studying outcomes including disease onset, device failure, and death. The results of these analyses are used to verify product safety and evaluate the viability of medical therapies, guiding decisions with ramificationsthat can extend around the globe and over decades. Commensurate with the importance of drawing accurate conclusions in these settings, many theoretical and methodological advancements have been developed to improve the reliability and efficiency of survival analyses. Modern causal inference frameworks allow questions about medical treatment efficacy to be formalized with mathematical rigor, and enable a detailed understanding of the necessary conditions for practitioners to be able to estimate causal effects from observed data. Robust and efficient semi-parametric estimators have been developed, capable of incorporating the flexible machine-learning algorithms which have been made practical by the increasing avail- ability of high performance computing. This dissertation is focused on a vision of modern survival analysis, guided by the Causal Roadmap and employing state-of-the-art estimators. In Chapter 2 we review the shortcomings of traditional survival analyses and demonstrate a Targeted Learning primary analysis of the LEADER cardiovascular outcome trial. In Chapter 3 we describe a new R package, concrete, which implements a recently developed continuous-time one-step targeted maximum likelihood estimator (TMLE) for time-to-event estimands with or without competing risks. Lastly in Chapter 4 we apply concrete to three analyses of the data from the SUSTAIN-6 cardiovascular outcome trial, demonstrating the potential of this package and estimator for answering commonly asked causal questions in time-to-event trial analyses.