What Causal Inference Teaches Us About the Limitations of Indirect Treatment Comparisons for Health Technology Assessment

PURPOSE: Indirect treatment comparisons (ITCs) are essential in HTA when direct head-to-head trials are unavailable. However, ITCs are susceptible to biases threatening their validity. This workshop will introduce attendees to the key principles of causal inference and treatment effect heterogeneity and explain how these principles can improve our understanding of the limitations of ITCs within the framework of HTA. Participants will learn why ITCs are “essentially observational findings across trials” (Cochrane Handbook) and how to critically evaluate their validity. DESCRIPTION: Dr. Siebert will introduce the session giving an overview on the key principles of causality and causal diagrams (8min) followed by presentations from the three speakers and discussion/questions (10min). The audience will be asked (real-time polling) to consider when, and to what extent, evidence from ITCs analyses should be considered by HTA agencies. Dr. Chatton will begin by establishing a formal definition of causal effects using the potential-outcomes framework, emphasizing counterfactuality and the estimand as central concepts. He will review the crucial assumptions necessary for valid causal inference (exchangeability, positivity, consistency and noninterference) and the main types of causal estimators available within the context of an externally controlled single-arm trial (14min). It’s often acknowledged that “the choice of effect measure may have a considerable impact on [one’s] analysis, and also on the degree of observed heterogeneity” (EUnetHTA). Through the lens of causal inference, Dr. Webster-Clark will explain how the choice of effect measure also determines the set of variables upon which one must adjust to maintain external validity (14min). Dr. Campbell will demonstrate how these principles translate to ITCs, clarifying the concepts of transportability, non-collapsibility, the relevance of marginal, conditional and population-average estimands, and the impact of failing to adjust for prognostic variables (14min).