ネットワークメタ回帰を用いたSidesplitting

Abstract

Consistency is an important assumption to justify evidence synthesis in network metaanalysis. Sidesplitting is a representative method used to evaluate inconsistency; it decomposes the overall estimate of network meta-analysis on a specific treatment pair to those of direct and indirect comparisons and assesses their concordance. A relevant issue in sidesplitting is that adequate adjustments are needed in multi-arm trials (≥3 arms) to prevent biases. In existing methods, sidesplitting requires several restrictions on model parameters or additional parameter modeling and the computations are complicated. In this article, we show that sidesplitting involving the adjustments of multi-arm trials can be uniformly treated within a network meta-regression framework, especially via the modeling method of Noma et al. (2017; Stat Med 36:917-927), which introduces additional free parameters to adjust the biases caused by multi-arm trials.The proposed approach can be interpreted as a specific version of the design-bytreatment interaction model, and any inference methods for the network metaregression can be applied involving higher-order asymptotic approximations. The proposed method is applied to two network meta-analyses of hypertensive drugs.