Doctoral candidate in Biostatistics, Mengli Xiao, will present:
“Innovative Statistical Methods for Meta-Analyses With Between-Study Heterogeneity”
PhD Adviser: Haitao Chu
Abstract: To assess the benefits and harms of medical interventions, meta-analysis plays a vital role in combining results from multiple studies. While similarities between studies motivate the notion of combining independent results, the combined result may be biased by between-study heterogeneity. Specifically, heterogeneity originating from non-replicable studies (that is, study results are inconsistent) or varying study-specific characteristics can change the conclusion of a meta-analysis. However, current meta-analysis methods have limited ability to handle these problems. We first develop a statistical framework establishing whether studies are replicable from each other to give sufficiently similar results, a fundamental assumption of a meta-analysis. Violating such an assumption is often misconceived as high heterogeneity in practice, and we show that the proposed statistic effectively distinguishes heterogeneity from non-replicability. When replicable studies in a meta-analysis show substantial heterogeneity, the dependence of the relative treatment effect on unknown study-specific characteristics is commonly underestimated. We then propose using a bivariate meta-analysis method that estimates treatment effects conditioning on effect modifiers approximated by baseline risk, leading to the insight that no summary effect measure stays constant with varying study conditions. We conclude that understanding how to improve meta-analysis from replicability analysis and modeling unknown study characteristics is essential in patient-centered practice.