Extending multilevel network meta-regression to disconnected networks and single-arm studies: a case study on plaque psoriasis treatments

Introduction
Single-arm studies and disconnected evidence networks are becoming common in health technology assessments. This has led to increasing use of unanchored population-adjustment methods for indirect treatment comparisons, which rely on strong assumptions that all prognostic factors and effect modifiers have been accounted for. Current methods have limitations: they only apply to pairwise indirect comparisons (one individual patient data (IPD) and one aggregate data (AgD) study), and can only produce estimates for the AgD population which may not represent the decision target population. Multilevel network meta-regression (ML-NMR) is an extension of network meta-analysis that overcomes some of these limitations to coherently incorporate both IPD and AgD, whilst adjusting for prognostic factors and effect modifiers in anchored population-adjusted analyses. However, ML-NMR has not yet been extended to unanchored analyses.
Methods
We extend ML-NMR to incorporate single-arm studies and disconnected networks, by predicting baseline response on a reference treatment for these disconnected studies using either other studies in the network or external evidence. We demonstrate with a network of treatments for moderate-to-severe plaque psoriasis, removing treatment arms to artificially disconnect the network, then reconnecting using different sources of evidence to predict reference treatment response, including external AgD from the PROSPECT and Chiricozzi 2019 cohort studies. We compare estimates of relative treatment effects and absolute outcome probabilities from the reconnected networks against those derived from the original connected network, treating the original connected analysis as the ‘truth’. Analyses were performed in a Bayesian framework in R and Stan using the multinma R package.
Results
Analysis across multiple reconnected networks indicated variability in estimated treatment effects and absolute outcomes across study populations, depending on the evidence source used for reconnection. The PROSPECT reconnected network generally provided estimates close to those from the original connected network, whereas estimates from the Chiricozzi reconnected network deviated more substantially. This can be attributed to the similarity of the PROSPECT population to the disconnected study population in terms of prognostic and effect modifying factors, whereas the Chiricozzi population was more different.
Conclusion
We have demonstrated, for the first time, how ML-NMR may be extended to single-arm studies and disconnected networks. Performance of the reconnected network varied by the study population used for reconnection, emphasizing the need for careful evidence selection methods to assess bias. Future research should continue to explore new methods for network reconnection and methods for validating the strong assumptions required.