A comparison of multivariate and univariate meta-analysis

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Multivariate meta-analysis (MVMA) synthesizes correlated outcomes but relies on accurate within-study correlation specification.
• Previous research on MVMA versus univariate meta-analysis (UMA) performance has yielded inconsistent findings, particularly when correlations are misspecified.

Purpose of the Study:

• To clarify the comparative performance of MVMA and UMA under various conditions, including correct and misspecified within-study correlations.
• To investigate the impact of misspecified correlations on effect size and between-study variance estimates.

Main Methods:

• Two simulation studies were conducted, varying key parameters such as the number of outcomes/studies, sample size, heterogeneity, effect sizes, and correlation structures.
• The proportion of missing data in one outcome was also manipulated to assess its effect on both UMA and MVMA.

Main Results:

• UMA outperformed MVMA in more scenarios than vice versa, even when within-study correlations were correctly specified, contradicting some prior conclusions.
• Misspecified correlations had a greater impact on between-study variance estimates than on overall effect sizes, with UMA demonstrating better performance.
• Missing data in outcomes degraded estimation and testing for those outcomes in both UMA and MVMA.

Conclusions:

• UMA may be preferable to MVMA in many practical situations, especially when within-study correlations are uncertain or misspecified.
• Careful consideration of correlation specification is crucial when applying MVMA.
• Missing data poses challenges for both UMA and MVMA, necessitating robust handling strategies.