Background: Network meta-analysis (NMA) can be conducted to compare the relative efficacy of multiple treatments. It is currently unclear whether access to individual participant data (IPD) is beneficial in NMA.
Methods: We discuss four key issues in a NMA and present several statistical methods to address these issues. In particular, we describe methods to overcome baseline imbalance (e.g. due to unlucky randomization), to investigate between-study heterogeneity in treatment effect, to investigate within-study heterogeneity in treatment effect and to minimize network inconsistency. For all methods, we highlight the role of IPD, and illustate their implementation in an empirical example dataset.
Results: Results indicate that when NMA do not incorporate patient- or study-level covariates, one-stage and two-stage models lead to similar estimates of relative treatment effect. However, two-stage models yielded slightly larger standard errors and substantially larger estimates of between-study heterogeneity. For all models, estimates of treatment effect gained precision and network consistency improved when adjusting for covariates. Most importantly, NMA that were based on IPD showed greater power in investigating the presence of effect modification, and helped to improve overall network consistency.
Conclusions: When investigating the relative efficacy of multiple treatments, NMA that are (partially) based on IPD may help to overcome several key issues hampering the validity and interpretability of traditional NMA analyses that are solely based on published aggregate data.