A comparison of multivariate and univariate meta-analysis

Multivariate meta-analysis (MVMA) extends univariate meta-analysis (UMA) by jointly synthesizing correlated outcomes across studies, but it requires specification of within-study correlations that may be misspecified in practice. Although prior research suggested that MVMA should outperform UMA when correlations are correctly specified, existing findings have been inconsistent when correlations are misspecified. To clarify these issues, we conducted two simulation studies varying the number of outcomes, the number of studies, per-group sample size, between-study heterogeneity, overall effect sizes, true correlations at both the within- and between-study levels, and the specified within-study correlation, and the proportion of missing data in the second outcome. Results showed that when within-study correlations were correctly specified, UMA outperformed MVMA in more conditions than vice versa, differing from previous conclusions that favored MVMA. When correlations were misspecified, the impact was larger on between-study variance estimates than on overall effect sizes, and UMA again performed better. Finally, when some outcomes had missing data, the estimation and testing for the corresponding outcomes were degraded in both UMA and MVMA.