Abstract

We study the design and interpretation of randomized controlled trials (RCTs) that compare medical treatments already in wide use. These RCTs often compare treatments that vary in their curative effectiveness and side effects, plus they can involve dynamic decisions by patients about whether and when to switch to a more aggressive treatment. Although the goal of RCTs is often an estimate of the population average treatment effect (ATE), the above factors induce patient selection into RCTs that generates estimates of a volunteer average treatment effect (VATE) different from the ATE. We demonstrate how to address these issues using a dynamic decision model and empirical methods that combine population and RCT data. We apply our method to treatment strategies for prostate cancer.