Article
A comparison of different approaches to binary random-effects network meta-analysis in sparse networks
Search Medline for
Authors
-
Svenja Seide
- Universität Heidelberg, Heidelberg, Germany
-
Katrin Jensen
- Universität Heidelberg, Heidelberg, Germany
-
Meinhard Kieser
- Universität Heidelberg, Heidelberg, Germany
| Published: | September 6, 2019 |
|---|
Outline
Text
Network meta-analysis allows evaluating multiple treatments for one condition by combining direct and indirect evidence. Networks of clinical trials involving between-study heterogeneity and multiple arms per trial are highly relevant in practice. However, they are often sparse with respect to the available direct evidence and the number of trials per contrast, which complicates modelling. Outcome data of event-type poses additional challenges in many applications. With respect to available methods, it is still unclear which one performs best in such scenarios. Current comparisons are commonly based on empirical data that do not allow a systematic investigation of the performance characteristics. We shortly introduce a data-generating model suitable for the above-mentioned challenges, which is a modification of an existing data-generating model in the pairwise case [1]. Based on this data-generating model, we perform a simulation study of network meta-analyses with binary endpoint, including between-study heterogeneity and multi-arm trials as well as varying degrees of sparsity. We evaluate random-effects network meta-analysis methods using such synthetic data with known theoretical values focusing on their performance under varying degree of sparsity with respect to directly observed contrasts and the number of trials in the network. Our simulation study evaluates frequentist (netmeta and mvmeta) and Bayesian network meta-analytical methods. Based on an empirical network of eight anti-tumor necrosis factor treatments [2], we simulate and evaluate coverage, width of confidence or credible intervals, RMSE, and ranking of treatments. Two different network geometries, scenarios with and without multi-arm trials, as well as different heterogeneities and numbers of trials per contrast are simulated. When a sufficient number of trials per contrasts is available, especially when heterogeneity is low, all methods perform well. In other scenarios, a trade-off between precision of the estimates and coverage of the true effect is observed which we discuss with respect to the performance of the above-mentioned methods.
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
This contribution has already been published: [3]
References
- 1.
- Pateras K, Nikolakopoulos S, Roes K. Data-generating models of dichotomous outcomes: Heterogeneity in simulation studies for a random-effects meta-analysis. Statistics in medicine. 2018 Mar 30;37(7):1115-24. DOI: 10.1002/sim.7569
- 2.
- Warren FC, Abrams KR, Sutton AJ. Hierarchical network meta-analysis models to address sparsity of events and differing treatment classifications with regard to adverse outcomes. Statistics in medicine. 2014 Jun 30;33(14):2449-66. DOI: 10.1002/sim.6131
- 3.
- Seide S, Jensen K, Kieser M. Comparing methods for random-effects network meta-analysis with binary outcomes in sparse networks. In: 40th Annual Conference of the International Society for Clinical Biostatistics; 2019 Jul 17; Leuven, Belgium. 2019.
