Online Program

A Bayesian Approach to Meta-analysis for Patient Outcomes in United States Teaching versus Nonteaching Hospitals

Keywords: Bayesian,meta-analysis,heterogeneity,between-study variance

Introduction: A fixed-effects estimate of between-study variance may underestimate the uncertainty of effect size in meta-analyses with high heterogeneity. Bayesian methods have the flexibility to model the between-study variance. We conducted a Bayesian meta-analysis in a study of mortality and length of stay (LOS) for teaching vs nonteaching hospitals and compared the results with traditional methods. Methods: Mortality was assessed in 94 studies and LOS in 46 studies. A hierarchical Bayesian scheme based on a random-effect model was set and Markov Chain Monte Carlo with Gibbs sampling was used for the posterior estimates. Results: Both methods showed that the odds ratio for mortality of the teaching hospital was slightly smaller and the LOS slightly longer. However, the heterogeneity was high for both mortality and LOS. Although the effect size estimates were almost the same, the Bayesian 95% Credible Intervals tended to be wider. Conclusions: In meta-analyses where heterogeneity is large, Bayesian methods may provide better variance estimations. This becomes very relevant as meta-analyses become more popular in the current era of comparative effectiveness research.