Presentation (CTI)

OBJECTIVES: Unanchored indirect comparisons are often required when randomized evidence is unavailable. Common approaches (MAIC, STC) are limited to pairwise settings and typically restrict inference to the comparator population. Multilevel unanchored meta-regression (ML-UMR), an extension of ML-NMR, enables population-adjusted unanchored comparisons of two or more studies and transport of treatment effects across populations. We introduce ML-UMR for time-to-event outcomes and evaluate its performance under varying degrees of effect modification (EM).
METHODS: A simulation study compared treatments A (index) and B (comparator) using two single-arm studies with substantial population imbalance in a binary prognostic factor (PF). Survival times followed Weibull models with a common shape for A and B. ML-UMR was fit assuming a shared PF effect across treatments (no EM). Four scenarios assessed performance under misspecification, inducing EM by shifting the PF effect for B by +/-10% (weak) or +/-50% (strong). Bias and coverage of 95% credible intervals were evaluated for marginal time-varying log hazard ratios (HRs) and restricted mean survival time differences and log ratios, transported to both index and comparator populations.
RESULTS: In the absence of EM, ML-UMR produced unbiased estimates in both populations (bias <4%) with near-nominal coverage (93%-96%). Under misspecification, performance patterns differed by population and EM magnitude and direction. With weak EM (+/-10%), marginal effects remained relatively robust, with bias <12% and coverage ranging from 87%-95%. Under strong positive EM (+50%), bias for log HRs commonly exceeded 30%, with coverage frequently below 60% and approaching 0% at later times. Strong negative EM (-50%) also resulted in large bias, although coverage declined more gradually.
CONCLUSIONS: ML-UMR can be extended to unanchored comparisons of survival outcomes and enables estimation of effects in any population. While robust when PFs are shared or weakly modified, ML-UMR is sensitive to misspecification under strong EM. Methods to relax the shared PF assumption require further research.