Statistical Challenges for Causal Inference Using Time-to-Event Real-World Data

Real-world data (RWD) have been increasingly used in drug development, e.g., for indirect comparisons of treatments in real-world settings and internal trials, or to augment a small internal control arm in a phase II trial. Using RWD for causal inference is a difficult task due to the lack of treatment randomization. RWD with time-to-event (henceforth referred to as TTE RWD) outcomes present extra challenges since causal effect estimands for TTE outcomes may also be difficult to determine. We briefly review the issues of hazard ratio as a causal estimand, Neyman–Rubin’s causal framework, and alternatives to hazard ratio such as the restricted mean survival time, followed by a description of several causal inference methods, with emphasis on their use for TTE data. The selection of time zero for TTE RWD and its impact on confounding adjustment are also discussed, including the use of trial emulation approaches. A short introduction to pseudo-observation-based approaches shows how they can facilitate the use of advanced confounding adjustment methods such as the doubly robust estimators and Bayesian approaches. Bayesian causal inference can take the advantage of the Bayesian framework for some tasks such as augmenting internal control arm by RWD. Some other topics such as using aggregated RWD are also discussed. This chapter ends with a summary and discussion of some remaining issues for future methodology research.