Anders Munch PhD Defense - Targeted learning with right-censored data
Invitation to PhD defense
Anders Munch
Targeted learning with right-censored data
Date Friday, 1 December 2023, at 14.00
Venue Øster Farimagsgade 5, 1353 Copenhagen K, room 2.1.12 (CSS)
The defense will be followed by a reception in room 5.2.46 (CSS)
Academic Advisors
Professor Thomas Gerds
Section of Biostatistics, Department of Public Health, University of Copenhagen, Denmark
Professor Claus Ekstrøm
Section of Biostatistics, Department of Public Health, University of Copenhagen, Denmark
Assessment Committee
Professor Thomas Scheike (chair)
Section of Biostatistics, Department of Public Health, University of Copenhagen, Denmark
Associate professor Morten Overgaard
Department of Public Health – Biostatistics, Aarhus University
Professor Jan Beyersmann
Institute of Statistics, Ulm University
Summary
The aim of this thesis is to contribute to the advancement of statistically rigorous methods that enable the utilization of data-adaptive estimators based on continuous-time observations that may be right-censored. Right-censoring often occurs when subjects are observed over a period of time, which is a typical situation in biostatistics. Conventional statistical methods for handling this type of data are based on (semi-) parametric models or simple nonparametric models. Importantly, for these approaches to provide valid statistical inference, the models have to be pre-specified. An appealing alternative is to use data-adaptive methods like machine learning, which provide more flexible models and tools for adapting the models to the observed data. The challenge with machine learning-based estimation strategies is to conduct valid statistical inference. Targeted learning addresses this challenge using semi-parametric efficiency theory. Although extensively studied for causal inference, the adaptation of targeted learning to right-censored problems in continuous-time data is less mature. The thesis is comprised of a synopsis with eight chapters and three manuscripts. The synopsis gives an introduction to the central theoretical concepts that underpin the topics covered by the manuscripts. The three manuscripts extend the framework and tools of targeted learning to settings with right-censored data in three different directions. The first manuscript applies the general framework of targeted learning to the illness-death model. We construct a class of estimators of the state occupation probabilities that can leverage data-adaptive estimators of the state transitions in the model. The second manuscript discusses the challenges facing the statistician who wants to construct a super learner from right-censored data. In this manuscript we also propose a new super learner and compare it to existing methods. The third manuscript formally extends the highly-adaptive lasso to settings that include conditional density and hazard function estimation.