A simple principal stratum estimator for failure to initiate treatment

A common intercurrent event affecting many trials is when some participants do not begin their assigned treatment. For example, in a trial comparing two different methods for fluid delivery during surgery, some participants may have their surgery cancelled. Similarly, in a double-blind drug trial, some participants may not receive any dose of study medication. The commonly used intention-to-treat analysis preserves the randomisation structure, thus protecting against biases from post-randomisation exclusions. However, it estimates a treatment policy effect (i.e. addresses the question "what is the effect of the intervention, regardless of whether the participant actually begins treatment?"), which may not be the most clinically relevant estimand. A principal stratum approach, estimating the treatment effect in the subpopulation of participants who would initiate treatment (regardless of treatment arm), may be a more clinically relevant estimand for many trials. We show that a simple principal stratum estimator based on a "modified intention-to-treat" population, where participants who experience the intercurrent event are excluded, is unbiased for the principal stratum estimand under certain assumptions that are likely to be plausible in many trials, namely that participants who initiate the intervention under one treatment condition would also do so under the other treatment condition. We provide several examples of trials where this assumption is plausible, and several instances where it is not. We conclude that this simple principal stratum estimator can be a useful strategy for handling failure to initiate treatment.