Bayesian Meta-Analysis with a Mixture of Mean Differences and Odds Ratios or Relative Risks to Achieve a Threshold

• Activity Number: 233
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Robert Grant*+
• Companies: St George's, University of London & Kingston University
• Keywords: Bayesian statistics ; meta-analysis ; coarse data ; latent variable

Abstract

Meta-analysis, pooling statistics from several studies to obtain a more precise estimate, is a literally vital tool in medical research. Methods to pool studies that report mean differences, or that report metrics for dichotomous outcomes such as odds ratios and relative risks, are widely used, but the published evidence frequently contains a mixture of these, for which there is no single flexible meta-analytic technique. The dichotomous metrics can be viewed as a coarse measure of an underlying latent continuous variable.

A Bayesian model for this form of meta-analysis is presented, which can incorporate absolute and relative rules for dichotomy. A simulation study with Stan software compares bias and coverage with classic meta-analysis, with particular regard to the coarsening mechanism in relation to reporting bias and Heitjan and Rubin's taxonomy of coarse data. A recent Cochrane review on tricyclic antidepressants in children is used as an applied example, which previously had separate analyses of studies reporting mean differences and studies reporting relative risks. The natural role of informative priors in the evidence-based healthcare setting is discussed.