Comparison and Selection of Network Meta‐Analysis Models for Fitting and Extrapolating Cancer Survival Data

Vollständiger Abstract

Worum geht es in dieser Arbeit?

ABSTRACT Aim To assess network meta‐analysis (NMA) model biases under different proportional hazards (PHs) scenarios, compare fitting and extrapolation performance for hazard ratios (HRs) using cancer survival data, identify metrics guiding model selection, and compare extrapolation strategies using parametric versus constant‐tail HRs. Methods We conducted 30 NMAs (five three‐arm trials, with each arm anchoring pairwise comparisons for overall survival and progression‐free survival), subgroup analyses assessed the PH assumption and curve types. Models included Cox‐PH, Fractional Polynomial, Royston–Parmar (RP), Piecewise Exponential (PWE), and Parametric Survival Model (PSM). For extrapolation, we categorized models into those using parametric‐extrapolation HRs and those assuming constant‐tail HRs. Main indicator was sum of squared errors (SSE), alongside Bias, assessed for observed data (SSE‐O, Bias‐O), fitting (SSE‐F, Bias‐F), and fitting‐extrapolation (SSE‐FE, Bias‐FE). Results Model choice introduced much more uncertainty when the PH assumption was violated than when held. RP‐2knot had the best HR fitting, comparable to RP‐1knot; errors were 0.5‐ to 33.7‐fold higher for other models. RP‐2knot and RP‐1knot demonstrated the best fitting‐extrapolation performance. Friedman tests showed RP‐2knot significantly outperformed other models in fitting, except RP‐1knot, and in fitting‐extrapolation, except RP‐1knot and Cox‐PH. SSE‐O and Bias‐O were highly correlated with model fitting ( ρ = 0.682–0.713) and moderately with fitting‐extrapolation. Within 2–3 years extrapolation window, constant‐tail HR models outperformed parametric‐extrapolation HR. Conclusions Model selection required more caution when the PH assumption failed. RP model showed promising fitting and extrapolation performance. SSE‐O and Bias‐O may guide model selection in HR‐fitting; extrapolation utility needs further validation. Non‐PH models using constant‐tail HRs showed better short‐term extrapolation.