Presentation (CTI)

OBJECTIVES: Evidence base on the use of multi-level network meta-regression (ML-NMR) to analyze time-to-event (TTE) outcomes is sparse. We assessed the performance of ML-NMR in estimating the relative treatment effects using parametric and flexible parametric (spline) models, comparing it with a Bayesian indirect treatment comparison (bITC) using a simulated TTE dataset.
METHODS: We simulated a hypothetical individual patient dataset for Trial-AB (treatment B vs A) and an aggregated dataset (AgD) for Trial-AC (treatment C vs A), including both continuous and dichotomous covariates. Following the methods from Phillippo et al. (2020), survival times were modeled using a full range of parametric and monotonic spline (Mspline) models within a Bayesian framework using Stan. We introduced treatment effects onto the coefficients to account for non-proportional hazards (NPH) and produced effect estimates (hazard ratios [HRs] and 95% credible intervals [CrIs]) in AgD population. The predictive performance of each model was evaluated using the leave-one-out information criterion (LOOIC), with lower values indicating better predictive performance.
RESULTS: The Weibull model demonstrated the best fit (LOOIC: -261.8), followed by the monotonic spline knot-1 model (LOOIC: -252.3). Relative effect estimates for C vs B from Weibull bITC, Weibull ML-NMR-PH, Weibull ML-NMR-NPH, and spline knot-1 ML-NMR were 2.39 (95%CrI: 1.46-4.01), 2.48 (95%CrI:1.39-4.48), 2.59 (95%CrI: 1.35-4.90), and 2.68 (95%CrI: 1.46-4.89), respectively. Variations in point estimates were attributable to covariate interaction and relaxation of PH assumption in the models. Although both bITC and ML-NMR models provided comparable estimates for treatment C vs B, ML-NMR accounted for both baseline covariates and NPH, potentially improving the validity of the estimates in the AgD population.
CONCLUSIONS: ML-NMR enables prediction of estimands between any pair of treatments for a target population. Mspline models in ML-NMR offer enhanced flexibility in capturing complex hazard functions, thus improving long-term survival extrapolation.