Marie Skov Breum PhD Defence

Causal inference and mediation analysis for longitudinal data

Date Friday, January 26, 2024, at 14.00
Venue Øster Farimagsgade 5, 1353 Copenhagen K, room 1.1.18
After the defence, a reception will be held in 5-2-46.

Academic advisors:
Professor Torben Martinussen, Section of Biostatistics, Department of Public Health, University of Copenhagen
Professor Thomas Gerds, Section of Biostatistics, Department of Public Health, University of Copenhagen

Assessment committee:
Theis Lange, Department of Public Health, University of Copenhagen (chairman)
Oliver Dukes, Ghent University
Klaus Holst, Novo Nordisk

SUMMARY


This thesis is broadly concerned with developing statistical methodology for causal inference based on longitudinal data. In particular it focuses on causal mediation analysis for longitudinal data, and on the application of causal inference methodology to time-to-event analysis.


The thesis consists of a synopsis containing five chapters, followed by three manuscripts. Chapters 1-3 of the synopsis provide the necessary background knowledge for better understanding the methodological contributions of the manuscripts. Chapter 4 contains a summary of the manuscripts. Chapter 5 discusses limitations of the proposed methods and outlines possible directions for future research. The contributions of the manuscripts can be summarized as follows:

  • Manuscript I proposes a method for estimating the extent to which the effect of a baseline exposure on a terminal time-to-event outcome (e.g. death) is mediated through a non-terminal time-to-event outcome (e.g. onset of disease). The method extends the concept of ‘separable’ direct and indirect effects to the illness-death setting.
  • Manuscript II is motivated by a data application from the NASH clinical trial conducted by Novo Nordisk. We propose a method for estimating the extent to which the effects of Semaglutide on NASH is mediated through weight loss which is a repeatedly measured covariate. The proposed method builds upon work on ‘randomized interventional’ direct and indirect effects.
  • Manuscript III is concerned with the estimation of concordance measures for right censored time-to-event data. Many widely used estimators will depend on the censoring distribution under model misspecification. Our contribution is that we view the concordance measures as model-free estimands and propose non-parametric estimators.