Pragmatic clinical trials provide the opportunity to conduct patient-centered outcomes research (PCOR) and comparative effectiveness research (CER) in real-world medical settings. An increasingly attractive design for pragmatic clinical trials is a cluster-randomized design, in which interventions are assigned at the cluster level and outcomes are measured at the individual level. One important but under-investigated challenge in pragmatic cluster-randomized trials (CRTs) is post-randomization selection bias. Our motivating example is the PRimary Care Opioid Use Disorders Treatment (PROUD) Trial, a pragmatic CRT that tests the implementation of a program for increasing medication-based treatment for primary care patients with opioid use disorders (OUDs). Because of the PROUD intervention, the intervention clinics are more likely than the nonintervention clinics to diagnose and treat OUD patients; thus, the OUD patients who were diagnosed post-randomization in the intervention arm may be systematically different from those in the control arm, and directly comparing the outcomes of these patients between the two groups would be biased. However, ignoring all the OUD patients diagnosed post-randomization would lead to significant loss of information and fail to capture patient heterogeneity. Currently, there is a lack of statistical methods for addressing post-randomization selection bias in the context of complex pragmatic CRTs. The proposed research aims to fill this methodological gap by developing and disseminating innovative and generalizable causal inference methods to address post-randomization selection bias in the design and analysis of CRTs. The methodological underpinning is principal stratification, a powerful general framework for addressing post-assignment variables in causal inference.
First, based on principal stratification, we will define a class of causal estimands (i.e., target of estimation in the analysis) to capture patient heterogeneity in the presence of post-randomization selection bias in CRTs, and provide assumptions and formulas for identifying these estimands. Second, we will develop several novel statistical methods and models to estimate these estimands, including
- a robust frequentist estimation method based on generalized estimating equations and
- a flexible Bayesian model to leverage rich individual characteristics for more precision.
Finally, we will apply the proposed methods to the PROUD Trial to evaluate the causal effects of increasing medication-based treatment of OUD patients in primary care on patient-level acute care utilization, and conduct extensive simulations based on PROUD to examine the operating characteristics of the proposed methods and compare them with standard methods. The application to PROUD will provide timely, accurate, and comprehensive PCOR/CER results about OUD treatments, which are of particular clinical and policy relevance amid the ongoing opioid crisis in the United States. More generally, the proposed methods are applicable to any CRT that shares similar structures as PROUD, regardless of the targeted medical conditions or patient outcomes. Through improved design and analysis of pragmatic CRTs, this project will provide more reliable comparative effectiveness information for clinicians and policy makers, and it will ultimately benefit patients with a wide range of conditions.