Causal models for time-varying exposure and confounding in pharmacoepidemiology

Specialeforsvar: Martin Russek

Titel: Causal models for time-varying exposure and confounding in pharmacoepidemiology

Abstract: Time-varying confounding can occur in any study encompassing more than one time-point and has the potential to bias results if not properly accounted for. We give an introduction to two models that have been shown to properly adjust for time-varying confounding with time-to-event data, Cox Marginal Structural Models and Structural Nested Cumulative Failure Time Models. Both are known to have some drawbacks, as well as advantages with respect to each other. We aim to explore those through theory, looking at differences in assumptions and interpretation. Furthermore, we perform a simulation study with data resembling that of a typical pharmacoepidemiological study, to evaluate bias and uncertainty quantification of the two methods compared to a standard analysis approach. Finally, we apply the methods to a real data example using Danish registry data and compare their results. We conclude that Cox Marginal Structural Models are preferable in our setting, as long as the corresponding assumptions are met. The major drawback of Structural Nested Cumulative Failure Time Models is the lack of thorough software implementation.

Vejledere: Anders Tolver
Morten Andersen, SUND
Censor: Claus Dethlefsen, NOVO