Title: Targeted causal estimation in continuous time
Authors: Helene Charlotte Rytgaard - University of Copenhagen (Denmark) [presenting]
Abstract: The aim is to discuss aspects of a continuous time generalization of longitudinal targeted minimum loss-based estimation (TMLE). TMLE is a framework for estimation of causal parameters that combines data-adaptive estimation with a targeting procedure tailored to optimal estimation of a specific low-dimensional parameter of interest. The existing TMLE methods for longitudinal data rely on a discrete data structure where observations are made on the same grid points of time for all subjects. In most realistic settings, however, both exposure and outcome can happen at arbitrary points in time. We propose a unified methodology relying on counting process modeling to handle the data exactly as they are observed. This involves construction of hazard-based initial estimators of the components of the partial likelihood and an extension of the targeting procedure that combines updating steps of intensities and conditional expectations. Potential applications include analysis of large-scale registry databases with regularly updated measurements of all members of a population over large timespans.