Causal inference methods for estimating effects of treatments or other interventions on health outcomes have seen rapid and extensive developments in recent years. Tasks for making causal inferences range from estimating average treatment effects, to conditional average treatment effects, to obtaining individualised predictions under different treatment choices. The types of treatment strategies we may be interested in range from giving a one-time treatment, to assigning a treatment to be sustained over a long period, to dynamic strategies that involve initiating a treatment when the patient reaches a certain health state.

This talk will attempt to provide a guide to the challenges we face in investigating causal effects of treatments on time-to-event outcomes using longitudinal observational healthcare data, and how they can be tackled. I will discuss the importance of clear specification of the research question and what we aim to estimate (the estimand), which in the time-to-event context often involves consideration of competing events.

I will give an overview of some of the statistical methods at our disposal for answering questions about the effects of treatments on time-to-event outcomes, which include marginal structural models, inverse probability weighting, censoring-and-weighting, g-formula, and extensions that offer double-robustness. Traditional methods for investigations into effects of exposures or treatments on survival, such as Cox regression, are sometimes stated as not being suitable for use in causal inference. However, I will discuss how these traditional techniques still form a fundamental part of the toolkit for causal inference for time-to-event outcomes.

A running example in the context of type 2 diabetes will be used, making use of an open access data set designed to mimic longitudinal observational data such as from longitudinal cohort studies, patient registries and electronic health records.