Methods for Causal Inference in Randomized Trials with Multiple Versions of Control and Noncompliance, with an Application to Behavioral Intervention Trials

Abstract

Behavioral therapies are a class of interventions with a wide array of applications.Because of the complicated nature of these interventions, however, conducting randomized controlled trials of these interventions poses unique challenges compared to the classical blinded, placebo-controlled RCT. The primary issue is that RCTs of behavioral interventions often use treatment-as-usual (TAU) control groups, due to the lack of a feasible ”placebo” equivalent to the active intervention. As a result,control groups in these trials are typically heterogeneous with respect to the form of treatment received, making causal inference under the standard assumption of ”no multiple versions of treatment” no longer applicable. In this dissertation, we develop frameworks for causal inference in single-site and multi-site RCTs with multiple ver-sions of control due to the use of a TAU control group. We define causal estimands of interest based on a principle stratification approach. We show that these causal estimands are only partially identified with data from a single-site RCT, but can be identified under certain assumptions with data from a multi-site RCT. We then propose methods for performing inference for these causal estimands, either through bounding (in the case of partial identifiability) or point estimation (in the case of identifiability). Finally, we apply these methods to an RCT of a behavioral therapy intervention for children with autism. Additional work in this dissertation includes an examination of identifiability issues with methods for causal inference in RCTs with partial compliance, a tutorial for a Bayesian approach to binary non-compliance in RCTs, and a systematic review of behavioral interventions for children with autism.