Understanding Contributions of Indirect and Direct Evidence to Statistical Power in Bayesian Network Meta-Analysis: Simulation Studies and Real-World Applications

Abstract

Network meta-analysis (NMA) has become a popular tool for simultaneously evaluating multiple interventions in biomedical studies. However, performing power analysis in NMAs can be challenging because it depends on network structures and forms of hypotheses to be tested. In this study, we investigate how varying evidence structures within Bayesian NMAs influence the statistical power and bias of relative treatment effect estimates using simulations and real-world case studies with binary outcomes. We first conduct a comprehensive simulation study to examine the properties of power in NMA under various scenarios, including different effect sizes, between-study heterogeneity, evidence composition, and model parameterizations. We then provide case studies to illustrate the practical application of our methods for networks of different sizes and structures. Our results suggest that powers are notably sensitive to the type and strength of evidence for each hypothesis and can be improved by increasing sample sizes, reducing between-study heterogeneity, and adding more direct evidence. In addition, power behaviors of contrast-based and arm-based parameterizations largely agree. The findings provide insights into the empirical implications of power analysis in Bayesian NMA, optimizing future NMA design for more reliable and robust healthcare decision-making.