Network Meta-Analysis of Multiple Outcomes: A Simulation Study and Application
The usefulness of a multivariate approach to compare treatments in the context of pairwise meta-analysis has been widely demonstrated in the literature. However, this approach has not yet been considered for multiple treatment comparisons. We believe that extending such methodology to network meta-a...
| المؤلفون الرئيسيون: | , , , , , |
|---|---|
| التنسيق: | Conference item |
| اللغة: | English |
| منشور في: |
Elsevier
2013
|
| الملخص: | The usefulness of a multivariate approach to compare treatments in the context of pairwise meta-analysis has been widely demonstrated in the literature. However, this approach has not yet been considered for multiple treatment comparisons. We believe that extending such methodology to network meta-analysis (NMA) will increase the primary evidence base allowing us to compare more interventions across multiple outcomes measures. Borrowing strength between outcome measures using multivariate NMA can also potentially increase the precision of relative treatment effect estimates and reduce the impact of outcome reporting bias.
Objectives
To extend standard NMA to incorporate multiple outcomes of interest and evaluate the use of multivariate NMA models through simulated and real datasets.
Methods
We developed a random effects multivariate NMA model to account for the correlation between multiple outcome measures. The potential benefits of this method were demonstrated in a simulated example comparing univariate and bivariate NMAs for continuous outcome measures. We further explored the application of our multivariate NMA model using a case study comparing anti-obesity pharmacological interventions for waist circumference, weight change and BMI change from baseline.
Results
The simulation study showed that through use of multivariate NMA the precision in mean relative treatment effects increased compared to a standard univariate NMA. This held true under multiple scenarios testing model parameters including both within- and between-outcome correlations. Similar findings were obtained from the application to the example dataset in obesity.
Conclusions
Our method proves particularly useful in reducing uncertainty around relative effectiveness estimates when the outcomes included for analysis are highly correlated. However, the advantages of the multivariate NMA are limited where there is little correlation between outcome measures. Further work will explore the applicability of multivariate NMA methods to different types of outcomes such as binary outcome measures. |
|---|