Skip to main content

Research Repository

Advanced Search

Development and application of multivariate meta-analysis in medical research: borrowing strength across multiple correlated outcomes

MacDonald Hattle, Miriam Susanna

Development and application of multivariate meta-analysis in medical research: borrowing strength across multiple correlated outcomes Thumbnail


Authors

Miriam Susanna MacDonald Hattle



Abstract

Multivariate meta-analysis methods combine effect estimates for multiple correlated outcomes (such as systolic and diastolic blood pressure) from independent studies, utilising their between-study and within-study correlations. In contrast, a univariate meta-analysis pools effect estimates for each outcome independently. By using the multivariate over the univariate approach, there is a potential gain in information toward summary meta-analysis results, quantified by the Borrowing of Strength (BoS) statistic, a percentage reduction in the variance for the summary effect between the two approaches. This thesis examines BoS and multivariate meta-analysis applications in detail.
Firstly, multivariate meta-analysis is applied to an individual participant data metaanalysis examining the effect of diet and exercise interventions during pregnancy. This shows results from the univariate and multivariate meta-analyses are similar. However, a review of 43 Cochrane reviews concludes that although results between the univariate and multivariate are often similar, a few multivariate meta-analyses do give important differences to results from univariate meta analyses, and these are shown to have a larger magnitude of BoS.
This motivates research to identify predictors of BoS and to develop a model to predict BoS (in advance of analysis) to flag when researchers should consider multivariate meta-analysis. Additionally, an interactive tool is developed to investigate the relationship between the various characteristics and the magnitude of BoS. The magnitude of BoS is shown mathematically to be approximately bounded by the percentage of missing data for the outcome of interest.
A novel application of bivariate meta-analysis is then proposed for trials with continuous outcomes analysed with final score or ANCOVA models, with examination in real examples and a simulation study.
In conclusion, multivariate meta-analysis may provide important differences to univariate meta-analysis when BoS is large, and so researchers should consider the approach when BoS is expected to be large.

Publicly Available Date Mar 29, 2024

Files




Downloadable Citations