Instrumental Variables in Randomized Trials

Many randomized clinical trials fail to play out as intended: some participants assigned to the treatment group remain untreated, while others assigned to the control group cross over and receive treatment. In such settings, intention-to-treat analyses that compare participants by treatment assignment are diluted by noncompliance, while per-protocol analyses that compare participants by treatment received are contaminated by selection bias. Instrumental variables methods can address both problems. We explain the rationale for instrumental variables estimation in clinical trials and illustrate instrumental variables methods through an analysis of the effect of revascularization on quality of life. We argue that instrumental variables analysis should be central to pragmatic trials of all kinds, strategy trials in particular, and emerging "nudge trials" that encourage specific health-related behaviors in large populations.