Specifying Weakly Informative Priors in Bayesian Meta-Analysis

• In Event: AERA Poster Session Thursday 7:45am
  In Poster Session: Poster Session: Statistical Theory, Quantitative Methods and Applications

Abstract

In small-sample meta-analysis, Bayesian random-effects models often yield credible intervals for parameters of interest that are overly conservative. This study introduces a data-driven approach to specify a weakly informative prior distribution for the heterogeneity parameter aiming to alleviate the wide intervals obtained from Bayesian random-effects models. The findings from empirical data analysis and simulation study show that an inverse-gamma prior with the mode set to the maximum likelihood estimate of the between-study variance yields credible intervals that are close to the nominal 95% coverage. The near-nominal coverage rate remains consistent across various simulation settings, suggesting that this alternative approach offers a plausible range of values for the heterogeneity parameter in small-sample situations.

Authors

• Junok Kim, University of California - Los Angeles

• Michael H. Seltzer, University of California - Los Angeles