Presentation (CTI)

OBJECTIVES: Health Technology Assessment decisions require reliable estimates of relative treatment effects between multiple interventions. Population adjustment methods are used to obtain estimates adjusting for differences between study populations. Increasingly, these methods are used to make “unanchored” comparisons with disconnected networks and single-arm studies, relying on the stringent, as-yet untestable assumption of conditional constancy of absolute effects. Multilevel Network Meta-Regression (ML-NMR) is a population adjustment method that coherently synthesises individual participant data and aggregate data from multiple studies. Unlike other methods, ML-NMR provides estimates for any target population, scales to networks of any size, and allows key assumptions to be tested. However, ML-NMR has not yet been proposed for disconnected networks.
METHODS: We extend the ML-NMR framework to unanchored scenarios. Networks are reconnected using either a fixed baseline model that combines similar studies, or a two-stage random baseline model that captures uncertainty from unobserved prognostic factors. To guide anchor selection, we assess population similarity using propensity score overlap and clinical expertise. We propose methods to evaluate the conditional constancy assumption, using cross-validation to assess predictive accuracy and the random baseline variance to quantify unobserved differences between studies. We illustrate different bridging strategies using a manually disconnected network of plaque psoriasis treatments, benchmarking against the fully connected network.
RESULTS: Bridging comparisons between disconnected studies produced unbiased estimates when populations were highly similar but introduced 6-18% bias when populations differed substantially. The random baseline model incorporated additional uncertainty due to unobserved differences which here was small (𝜏; = 0.10 [0.03, 0.32]). Careful anchor selection, based on population similarity and clinical judgement, was essential for valid treatment comparisons.
CONCLUSIONS: Extending ML-NMR to disconnected networks and single-arm studies brings the advantages of this approach to unanchored settings. Practical methods to assess the strong assumptions are provided, and methods are implemented in the multinma R package.