Assessing the commonly used assumptions in estimating the principal causal effect in clinical trials

In addition to the average treatment effect (ATE) for all randomized patients, sometimes it is important to understand the ATE for a principal stratum, a subset of patients defined by one or more post-baseline variables. For example, what is the ATE for those patients who could be compliant with the experimental treatment? Commonly used assumptions include monotonicity, principal ignorability, and cross-world assumptions of principal ignorability and principal strata independence. Most of these assumptions cannot be evaluated in clinical trials with parallel treatment arms. In this article, we evaluate these assumptions through a 2x2 cross-over study in which the potential outcomes under both treatments can be observed, provided there are no carry-over and study period effects. From this example, it seemed the monotonicity assumption and the within-treatment principal ignorability assumptions did not hold well. On the other hand, the assumptions of cross-world principal ignorability and cross-world principal stratum independence conditional on baseline covaraites seemed to hold well. With the latter assumptions, we estimated the ATE for principal strata, defined by whether the blood glucose standard deviation increased in each treatment period, without relying on the cross-over feature. These estimates were very close to the ATE estimate when exploiting the cross-over feature of the trial. To the best of our knowledge, this article is the first attempt to evaluate the plausibility of commonly used assumptions for estimating ATE for principal strata using the setting of a cross-over trial.