Causal Inference
Evaluating Comparative Effectiveness Under Real-World Adherence Patterns Nicholas Williams* Nicholas Williams Williams Williams Williams Williams Williams Williams University of California, Berkeley
Deviations from treatment protocols are common in drug studies. For example, persons with HIV may experience interruptions in their antiretroviral therapy (ART), and certain ART regimens may be more robust than others with respect to patients maintaining virologic suppression. Thus, prescribers and patients need to understand the effectiveness of alternative regimens under imperfect adherence, often referred to as “forgiveness.” The standard approach for assessing forgiveness is to compare the outcome across strata of observed adherence levels. For example, this could correspond to comparing the proportion with viral suppression by ART regimen within strata of observed adherence levels. Adherence, however, is often a mediator; it is affected by the treatment regimen (e.g., daily oral regimens may be harder to maintain than long-acting injectables) and influences the outcome, and conditioning on the observed adherence level will generally yield biased estimates of effectiveness. Following the Causal Roadmap, we propose several alternative estimands, including joint stochastic interventions on treatment and adherence. These estimands capture dynamic treatment regimes and the cumulative effects of longitudinal adherence patterns. In simulations, we demonstrate how doubly robust estimators can target these estimands under minimal assumptions and contrast with the standard approach. We first illustrate shortcomings of the standard approaches and then extend to more realistic settings. We conclude with practical recommendations for generating evidence on comparative effectiveness under imperfect adherence.