New win methods for addressing multiple and composite outcomes in cluster-randomized trials - Project Summary/Abstract Cluster-randomized trials (CRTs) offer a powerful experimental design to evaluate the impact of complex interventions in medicine and public health, and have become among the most common study designs for embedded pragmatic trials seeking to improve patient care. Unlike individual-randomized trials, CRTs randomize intact clusters of subjects to alternative conditions, resulting in correlated observations that require specialized inferential tools. To maximize the patient-centered information gleaned from CRTs, there is an increasing interest in investigating the net treatment effects defined through win estimands (net benefit, win ratio, win odds) on multiple—prioritized and non-prioritized—outcomes that capture benefit or risk. However, existing statistical methods in CRTs have been primarily developed to address a single outcome measure. Principled methods that leverage the power of win statistics to guide the analysis of multiple and composite outcomes in CRTs are sparse, and their formal development requires addressing several methodological challenges. First, CRTs present a multilevel data structure that necessitates clear definitions of estimands. However, the existing estimands in CRTs do not address pairwise comparisons with multiple outcomes. Second, while covariate adjustment is a promising technique to improve estimation precision with a single outcome, robust and efficient covariate adjustment methods when targeting win estimands in CRTs remain under-developed, leading to missed opportunities in optimizing trial efficiency. Third, there is a lack of regression tools that enable the association analyses between win parameters and covariates when multiple outcomes collected from different individuals are correlated. To overcome these challenges, we will develop a suite of novel methods using clustered win statistics by maximizing information extraction from both covariates and outcomes, and apply our new methods to several CRTs that assess electronic health record alerts interventions for patients with heart failure. In Specific Aim 1, we will leverage the semiparametric efficiency theory of U-statistics to develop model-robust and efficient covariate-adjusted win estimators in CRTs with multivariate outcomes. In Specific Aim 2, we will expand the covariate-adjusted win estimators to censored time-to-event outcomes, and capitalize on the dual role of covariates in simultaneously addressing censoring bias and improving efficiency. In Specific Aim 3, we will develop marginal win regression methods to study the associational effects of covariates on win fractions under cluster correlated data, when the outcomes are either quantitative, or censored time to event. In Specific Aim 4, we will develop open-source software packages, tutorial papers, case studies, and short courses for all proposed methods, and use the NIH Pragmatic Trials Collaboratory as a unique national channel for effective dissemination. This work promises to advance the knowledge gained from past and future CRTs with complex outcomes by substantially expanding the toolbox for addressing win estimands in the multilevel data setting.
