Presentation (CTI)

OBJECTIVES: Indirect treatment comparisons (ITCs) are essential when head-to-head trials are unavailable, but can be biased by between-trial population differences. Matching-Adjusted Indirect Comparison (MAIC), Simulated Treatment Comparison (STC) and multi-level network-meta-regression (ML-NMR) adjust for between-trial differences in observed effect-modifiers, but differ in assumptions and interpretation. Choice of method should consider the target estimand (whether the treatment effect is defined at the population or individual covariate level), the target population (MAIC and STC are restricted to the comparator population; ML-NMR may specify target population) and suitability of statistical approach to available data. We explored how these methods influenced estimated treatment effects and interpretation for an example in psoriasis.
METHODS: We compared MAIC, plug-in-means STC (frequentist) and ML-NMR (Bayesian) for PASI 75 response (a 75% response threshold) in psoriasis. Individual patient data were available from UNCOVER-2 (ixekizumab) and aggregate data from FIXTURE (secukinumab), with etanercept as common comparator. Adjusted covariates were duration of psoriasis, previous systemic treatment, body surface area affected, weight, and psoriatic arthritis. We estimated odds ratios (ORs) and corresponding confidence or credible intervals (CI/CrI).
RESULTS: The ORs for PASI 75 for ixekizumab versus secukinumab when matching to the pooled aggregate characteristics of FIXTURE were: 3.12 (95% CI: 1.74, 5.59) for MAIC; 3.34 (95% CI: 1.70, 6.57) for STC and 3.44 (95% CrI: 1.76, 6.92) for ML-NMR. The unadjusted OR was 3.81 (95% CI: 2.21, 6.59).
CONCLUSIONS: The MAIC OR was lower than ORs estimated by ML-NMR and STC. The OR using plug-in-means STC, representing the estimand for the average patient, was between the MAIC and ML-NMR ORs, both representing the marginal estimand, suggesting that factors besides estimand may drive differences in results across approaches. Our results are consistent prior research suggesting that regression-based methods (ML-NMR, STC) perform more similarly to each other than to MAIC (Phillippo et al., 2020)