Article type
Year
Abstract
Background: Missing treatment effect estimates for particular outcomes in a study have the potential to affect the conclusions in a meta-analysis, especially if missingness is a result of outcome reporting bias (ORB). As well as missing treatment effect estimates at the study level, outcome data may also be missing within studies at the individual participant level. Multivariate meta-analysis of individual participant data (IPD) has the potential to overcome the impact of both these problems, by utilising the correlation between outcomes.
Objectives: To investigate, in a range of ORB and missing data scenarios, the magnitude of bias in pooled treatment effect estimates for multiple outcomes using standard (univariate) meta-analysis, and to quantify how much the ‘borrowing of strength’ (BoS) from multivariate meta-analysis reduces such bias and increases precision.
Methods: A simulation study was conducted where IPD was generated from an assumed multivariate fixed-effect model, and missing data were created at either the patient-level or the study-level, or both. In each simulation, the bias, precision and coverage of univariate and multivariate methods was compared, and the BoS quantified.
Results: Results show that the BoS in a multivariate model can substantially reduce the magnitude of bias and increase precision in the pooled estimates, especially when ORB is present and when correlation is modelled at both the patient-level and the study-level. In a missing at random scenario (0.8 w/s correlation) the BoS was 34.9% (outcome 2), meaning a reduction of 35% in the variance of the pooled estimate. Moreover BoS increases as the correlation and amount of missing data increases.
Conclusions: Meta-analysis results may be unreliable if there are missing outcome data. A multivariate meta-analysis approach is a potential statistical solution for reducing the impact of missing data, and is especially appealing when IPD are available to deal with missing data at both the patient and study levels.
Objectives: To investigate, in a range of ORB and missing data scenarios, the magnitude of bias in pooled treatment effect estimates for multiple outcomes using standard (univariate) meta-analysis, and to quantify how much the ‘borrowing of strength’ (BoS) from multivariate meta-analysis reduces such bias and increases precision.
Methods: A simulation study was conducted where IPD was generated from an assumed multivariate fixed-effect model, and missing data were created at either the patient-level or the study-level, or both. In each simulation, the bias, precision and coverage of univariate and multivariate methods was compared, and the BoS quantified.
Results: Results show that the BoS in a multivariate model can substantially reduce the magnitude of bias and increase precision in the pooled estimates, especially when ORB is present and when correlation is modelled at both the patient-level and the study-level. In a missing at random scenario (0.8 w/s correlation) the BoS was 34.9% (outcome 2), meaning a reduction of 35% in the variance of the pooled estimate. Moreover BoS increases as the correlation and amount of missing data increases.
Conclusions: Meta-analysis results may be unreliable if there are missing outcome data. A multivariate meta-analysis approach is a potential statistical solution for reducing the impact of missing data, and is especially appealing when IPD are available to deal with missing data at both the patient and study levels.