Adherence Adjusted Estimates in Randomized Clinical Trials

Randomized clinical trials (RCTs) are considered the gold standard for establishing the efficacy of an intervention. This is because randomizing patients to the different arms of an intervention ensures that the distributions of observed and unobserved confounders in the arms are the same. As a result, any differences in outcomes can be attributed to the difference in interventions. However, this advantage of randomization is lost when participants drop out, cross over to another treatment, or do not comply with the regimen. The traditional approach to the issue is to act as if randomization was maintained and analyze the data as intended in the design (intent to treat [ITT] analysis). However, the inference based on the ITT provides information on the prescription of the intervention rather than its use. Other approaches, such as looking at differences in outcome in those treated per protocol, are problematic because of the high risk of imbalance in measured and unmeasured confounders whose distribution cannot be assumed to be similar across arms because the randomization does not hold. Using the potential outcome concept first published by Neyman, Rubin developed what has come to be known as Rubin’s causal model (RCM) to address these issues. This chapter is a brief overview of the application of the RCM to the problem of poor adherence to treatment in RCTs. It provides existing solutions from the literature for most types of data that a researcher will encounter. It also looks at the validity of the assumptions of the RCM.