Causal Inference in Randomized Encouragement Design Studies with Non-compliance and Non-ignorable Missing Outcomes
Encouragement design studies are particularly useful for estimating the effect of an intervention that cannot itself be randomly administered to some and not to others. They require a randomly selected group receive extra encouragement to undertake the treatment of interest, where the encouragement typically takes the form of additional information or incentives. Noncompliance and missing data are particular problems in encouragement design studies, where encouragement to take the treatment, rather than the treatment itself, is randomized. We consider a clustered encouragement design (CED) where the randomization is at the level of the cluster. The motivating study looks at whether computer-based care suggestions can improve physician adherence and patient outcomes in veterans with chronic heart failure. Veterans Affairs (VA) treatment guidelines in this population have been created to improve quality of health care, lessen clinical practice variation, and reduce costs (Audet et al., 1990; Institute of Medicine, 1992). Since physician adherence to these guidelines has been inadequate, the original study focused on methods to improve physician adherence, although an equally important question is whether increased physician adherence improves patient outcomes. Naïve methods can produce biased estimates of the effect of physician adherence on patient outcomes since physician adherence is measured post-randomization. So here we re-analyze the data to determine the effect of physician adherence on patient outcomes using causal inference methods that focus on the treatment effect among a subpopulation of patients defined by their physician's potential compliance status.