Model selection for component network meta-analysis in connected and disconnected networks: a simulation study

TL;DR

This study evaluates model selection strategies for component network meta-analysis (CNMA) in both connected and disconnected networks, highlighting the effectiveness of CNMA under certain conditions and recommending separate analyses for disconnected networks.

Contribution

Introduces a forward model selection strategy for CNMA and assesses its performance in connected and disconnected networks through simulations and real data.

Findings

CNMA performs well in connected networks when additivity holds.

Model selection struggles in disconnected networks, suggesting separate subnetwork analyses.

Simulations show the importance of network connectivity for CNMA effectiveness.

Abstract

Network meta-analysis (NMA) is widely used in evidence synthesis to estimate the effects of several competing interventions for a given clinical condition. One of the challenges is that it is not possible in disconnected networks. Component network meta-analysis (CNMA) allows technically 'reconnecting' a disconnected network with multicomponent interventions. The additive CNMA model assumes that the effect of any multicomponent intervention is the additive sum of its components. This assumption can be relaxed by adding interaction component terms, which improves the goodness of fit but decreases the network connectivity. Model selection aims at finding the model with a reasonable balance between the goodness of fit and connectivity (selected CNMA model). We aim to introduce a forward model selection strategy for CNMA models and to investigate the performance of CNMA models for connected and disconnected networks. We applied the methods to a real Cochrane review dataset and simulated data with additive, mildly, or strongly violated intervention effects. We started with connected networks, and we artificially constructed disconnected networks. We compared the results of the additive and the selected CNMAs from each connected and disconnected network with the NMA using the mean squared error and coverage probability. CNMA models provide good performance for connected networks and can be an alternative to standard NMA if additivity holds. On the contrary, model selection does not perform well for disconnected networks, and we recommend conducting separate analyses of subnetworks.