Causal Inference with Post-Treatment Complications
SIZHU LU – UNIVERSITY OF CALIFORNIA, BERKELEY
ABSTRACT
The analysis of treatment effects is often complicated by events that occur after treatment initiation and before outcome measurement. Examples include treatment discontinuation in clinical trials, intervention cross-over or contamination in social and behavioral experiments, and changes in background policy in policy evaluation experiments. In clinical trials, such events are known as intercurrent events, and recent regulatory guidance, most notably ICH E9(R1), has emphasized the importance of handling them at the estimand level. Existing approaches, however, remain largely descriptive and are particularly limited in the practically common setting with competing intercurrent events: the occurrence of one intercurrent event censors the observation of subsequent ones.
In this talk, I will present our proposed causal framework for treatment effect estimation with competing intercurrent events. The framework formalizes causal estimands, establishes nonparametric identification under transparent assumptions, and introduces robust and efficient estimators. I will also demonstrate the approach using randomized trials in systemic lupus erythematosus. In addition, we extend the framework to accommodate alternative strategies for handling different types of intercurrent events that require distinct causal estimands.