Beyond the Hazard Ratio: Causal Inference from Time-to-Event Data with Dependent Censoring, Confounding, and Competing Risks.

In epidemiological research, time-to-event outcomes, commonly referred to as survival outcomes, are a common subject of investigation. Here, we first describe the circumstances under which methods of survival analysis are necessary, review the roles of hazard functions, and discuss the limitations of hazard ratios for causal inference. Second, we explain how confounding and dependent censoring, which are common in observational studies, can be addressed by inverse probability weighting and parametric g-formula estimators.

We emphasize that hazards serve as building blocks for estimating counterfactual risks. Third, we summarize recent developments in defining causal estimands in the presence of competing risks, including risk without eliminating competing events, net risk, and cumulative incidence under modified treatment. To illustrate how these challenges are addressed in practice, we revisit a recent clinical study on pharmacological interventions for the onset of dementia.

We further empirically compare various estimands in the presence of competing events, including separable effects, through simulations. Our overall aim is to elucidate the need to move beyond routinely used methods of survival analysis, particularly the mere estimation of hazard ratios, if the goal is to draw causal inferences. This paper provides an overview of causal survival analysis, focusing on how confounding, dependent censoring, and competing risks can be addressed to estimate causal parameters of interest (counterfactual survival functions or counterfactual risks), which are interpretable and often meaningful for investigators.