ABSTRACT
Multivariate meta-analysis may be appropriate when some or all studies in a meta-analysis report more than one outcome. This chapter describes the multivariate meta-analysis model and compares various methods for its estimation, focusing on two-stage methods. The standard model requires the within-study correlations between estimates for different outcomes to be known; various approaches when they are not known are discussed. Advantages and disadvantages of the multivariate approach compared with multiple univariate meta-analyses are then described. Advantages are that multivariate meta-analysis can yield more precise and/or less biased estimates for single outcomes through “borrowing of strength” and that it enables inferences to be drawn across multiple outcomes. Disadvantages are its greater complexity, with potential borrowing of strength often not being achieved, possible estimation problems, and extra modeling assumptions. Applications are discussed to confounder adjustment in observational studies, diagnostic test accuracy (where a one-stage approach is preferred), longitudinal data, and multinomial data.
