Statistics Seminar

Abstract: Across the sciences, statistical methods for meta-analysis are used to integrate findings and explore variation in results from multiple sources of evidence, such as multiple primary studies conducted on the same topic. A basic but critical decision in meta-analysis is the choice of metric on which to quantify the findings from difference sources. In practice, the choice of effect metric is strongly influenced by disciplinary conventions and heuristics, with little attention to model fit. I propose using a predictive fit metric (predictive log density) to inform the choice of effect metric in meta-analysis. I demonstrate this approach across several applications where different sets of effect metrics may be relevant, some of which require specification of auxiliary models to directly compare metrics.