Combining Several Pairwise Comparisons in Meta-Analysis for Joint Test of Effect Size

• Activity Number: 194
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Shaheena Bashir*+ and Celia M.T. Greenwood
• Companies: University Health Network, Toronto and Department of Epidemiology, Biostatistics and Occupational Health, McGill University
• Keywords: multivariate meta analysis ; random effects model ; heterogeneity statistics ; moment estimators

Abstract

Multivariate meta-analysis combines estimates of several related parameters across several studies. This work investigates the feasibility of obtaining, in a meta-analysis context, an omnibus F-statistic that can test for differences among more than two groups by combining several pairwise comparisons. The question arose in a meta-analysis of gene expression studies, where each study compared the same four treatment groups, leading to six possible pairwise comparisons from each study. We have developed a test statistic and evaluated its performance in simulations, when compared to standard random effect meta-analysis approaches of pairwise contrasts. The simulations showed that our meta-analysis omnibus statistic under a random effect model has a distribution that matches the expected F-distribution for small quantiles, but has a shorter right tail. Nevertheless, our statistic has better control of type I error than the minimum pairwise contrast p-value, even after empirical correction. Finally, as could be expected, power is better than "all-pair" power but not as high as "any-pair" power.