PURPOSE:Several methodological articles on bivariate meta-analysis have been published, but this technique remains rarely used although many trials include several correlated outcomes. Taking into account the correlation between outcomes in a meta-analysis offers potential to improve estimates by borrowing strength across outcomes and reducing potential bias when one outcome is not reported in all studies, e.g. in presence of selective reporting. This workshop will present methods for bivariate meta-analysis and review criteria to take into consideration when choosing a meta-analysis method.
DESCRIPTION:We will provide an overview of published models for meta-analysis with two endpoints, including fixed-effect and random-effect model. Different ways of specifying the correlation between outcomes have been proposed. The empirical correlation between outcomes will be presented for several published meta-analyses, in different therapeutic areas. The correlation between progression-free survival and overall survival in oncology will be studied in particular. The concepts of between-studies correlation and within-studies correlation will be explained. The relationship between variances for the two endpoints will be described using some examples and the role of this relationship in bivariate meta-analysis will be presented. We will then introduce a bivariate method using an imputation of the variance for studies for which one endpoint is not reported. The results of bivariate meta-analysis models and standard univariate models will be presented and compared in several cases, for different numbers of studies, different levels of heterogeneity between studies, different values of correlation between endpoints, and with missing values at random or not at random. Finally, we will discuss ways to incorporate results of bivariate meta-analysis in cost-effectiveness models. Participants will be encouraged to comment on applicability and usefulness of these methods in different therapeutic areas.