OBJECTIVES: Matching-adjusted indirect comparisons (MAICs) can estimate treatment effects in an aggregate data (target) population. However, when population-average conditional effect estimates of both interventions in both populations are needed, multi-level network meta-regression (ML-NMR) is more appropriate, as it allows for selecting any target population. We consider an MLNMR comparing efficacy of midostaurin (RATIFY:NCT00651261) and quizartinib (QuANTUM-First:NCTNCT02668653) versus placebo for patients with newly diagnosed FLT3-internal-tandem-duplication-positive (FLT3-ITD+) acute myeloid leukaemia.

METHODS: Individual patient-level data was available for QuANTUM-First, while aggregate data for complete response (CR), relapse after CR (CIR) and overall survival (OS) were available for RATIFY. Clinical experts identified potential treatment effect modifiers (TEMs) and prognostic variables (PVs) for adjustment in the ML-NMR. Both fixed and random effects models were fitted. For survival outcomes, proportional hazard Weibull and M-spline models were fitted. The target population was QuANTUM-First. Model selection was based on the Pareto-smoothed importance sampling leave-one-out cross-validation information criterion (PSIS-LOO CV).

RESULTS: Identified TEMs/PVs were sex, NPM1 mutation status, age and platelet count. Fixed-effect models were selected for all outcomes by PSIS-LOO CV. For survival outcomes, the M-spline model was selected. When comparing quizartinib to placebo in the target population, ML-NMR results closely mirrored observed trial results apart from CIR, which produced a more favourable quizartinib outcome. When comparing midostaurin and placebo, extrapolating midostaurin’s trial results to an older population (similar to QuANTUM-First) produced a numerically more favourable outcome for CR, while the numerical advantage of midostaurin over placebo for CIR diminished.

CONCLUSIONS: The ML-NMR approach successfully modeled CR and OS but resulted in an inconsistent model fit for CIR. ML-NMR offers several advantages over traditional population adjustment methods. However, caution is advised when interpreting results and applying them to economic models as ML-NMR typically targets a population-average conditional treatment effect, which differs from the population-average marginal treatment effect.