A Comparative Evaluation of Statistical Methods in Hybrid Controlled Trials

TL;DR

This paper compares various statistical methods for hybrid-controlled trials, emphasizing the importance of accounting for heterogeneity and confounding, and recommends exploring multiple approaches through simulations for robust results.

Contribution

It provides a comprehensive evaluation of statistical approaches for hybrid-controlled trials, highlighting the need to consider heterogeneity and unmeasured confounding in method selection.

Findings

No single method outperforms others across all scenarios.

Methods must account for heterogeneity from confounding.

Multiple methods should be explored via simulations.

Abstract

Randomized clinical trials (RCTs) are widely considered the gold standard for evaluating the effectiveness of new treatments or interventions in drug development. Still, they may not be feasible in certain cases, such as with rare diseases where randomization to a control group is ethically challenging. In such scenarios, external data can complement either a single-arm trial or a hybrid-controlled trial. The hybrid-control design involves enrolling fewer concurrent control patients and then supplementing the control arm using external or historical data. Various statistical approaches, including frequentist methods (e.g., propensity score methods), Bayesian borrowing approaches (e.g., meta-analytic-predictive prior), and their integration, have been utilized to incorporate external information in hybrid-controlled trials. We evaluate several accessible methods for their robustness to between-study heterogeneity and unmeasured confounding utilizing a case study based on data from the DAPA-HF trial, along with comprehensive simulation studies. Our findings indicate that the optimal methods must take into account the heterogeneities from both measured and unmeasured confounding. Since no single method consistently outperforms all others, researchers should explore multiple methods through extensive simulations to evaluate their effectiveness across various scenarios.