Presentation (CTI)

OBJECTIVES: Multilevel network meta-regression (ML-NMR) enhances network meta-analysis by statistically adjusting for effect modification (EM) in connected networks. Existing methods for disconnected networks (MAIC and STC), however, are limited to pairwise comparisons and cannot transport estimates beyond the comparator population. To address this gap, we introduce ML-UMR, a novel extension of ML-NMR for unanchored comparisons, and assess its performance via simulation.
METHODS: Our simulation study indirectly compared binary outcomes for treatment A (index) vs. B (comparator) from two single-arm studies. Patient-level data are available for A and only aggregate-level information for B. Population imbalance was induced across studies by generating correlated prognostic factors (PFs) with different means, and outcomes were simulated: 1) assuming PFs have the same effect on outcomes for A and B (shared PF assumption [SPFA]), which implies no EM; 2) relaxing SPFA, thereby inducing moderate/strong EM. ML-UMR models invoking and relaxing SPFA were fitted to assess the bias and coverage of 95% credible intervals of predicted marginal log odds ratios (LORs) in the comparator and index populations.
RESULTS: ML-UMR models invoking SPFA accurately predicted LORs in the comparator population regardless of the true EM strength (|bias|<0.009 [<6.4%] and 94.8%-96.0% coverage). The predicted LORs in the index population were unbiased in the absence of EM and relatively robust to moderate EM (bias=-0.06 [-11%]; coverage=93%); a high degree of bias (-0.29 [-67%]) was observed, however, when EM was strong. Relaxing SPFA in the ML-UMR model resulted in accurate LORs in both the index and comparator populations across all scenarios: bias<0.007 [<1.5%] and 93.4%-95.6% coverage.
CONCLUSIONS: This study demonstrates that the ML-NMR framework can be extended for unanchored indirect comparisons. Unlike MAIC and STC, ML-UMR allows transporting relative effect estimates to any target population under certain assumptions and can compare any number of treatments, but it remains subject to limitations inherent to unanchored comparisons.