Comparing line of therapy selection methods in advanced non–small cell lung cancer (aNSCLC): A simulation study.

Abstract

e23357 Background: External control arms utilizing real-world patients are increasingly used with single-arm clinical trials to demonstrate drug effectiveness in oncology. These patients often have multiple eligible lines of therapy (LoTs), the choice of which can meaningfully impact conclusions about effectiveness. Current LoT selection approaches include using the first eligible LoT (FEL), a random eligible LoT (REL), all eligible LoTs (AEL), and a novel method, stratified random LoT (SRL). We compare SRL to current selection approaches to identify an optimal approach in aNSCLC. Methods: Method performance was evaluated in a simulation study emulating trial-like EGFR+ aNSCLC patients and external controls for overall survival. Primary analyses were based on synthetic cohorts with typical prior LoT distribution between arm (Δ≈2), moderate treatment effect (HR≈0.73), and robust sample size (n≈750 per arm). We investigated the impact of varying LoT overlap, sample size, and effect size. Methods were evaluated on their ability to estimate effectiveness in naïve and weighted analyses using root mean square error, bias, and coverage. Results: Across primary analyses, SRL outperformed FEL with lower error and higher coverage, and performed comparably to or better than AEL. SRL more often overestimated treatment effects, whereas AEL tended to underestimate treatment effects. REL showed high error and bias, and low coverage, insufficiently addressed by weighting. SRL remained robust with smaller samples, null effects, and limited LoT overlap (Δ≈3); FEL and AEL degraded as overlap decreased. Conclusions: SRL demonstrated robust performance and unique value in disease settings where LoT strongly confounds clinical outcomes. Optimal LoT selection depends on therapeutic area and study context. Evaluation of LoT selection methods should incorporate multiple performance metrics; low bias alone may obscure poor inferential performance. Performance metrics. Primary Smaller Sample Size No Treatment Effect Less LoT Overlap Weighting approach Method HR RMSE Bias Coverage HR RMSE Bias Coverage HR RMSE Bias Coverage HR RMSE Bias Coverage SMR FEL 0.75 0.14 0.01 47% 0.81 0.19 0.08 44% 1.01 0.19 0.01 48% 0.98 0.91 0.24 17% REL 0.63 0.11 -0.10 32% 0.69 0.08 -0.05 73% 0.86 0.16 -0.15 26% 0.72 0.09 -0.02 67% AEL 0.77 0.06 0.03 85% 0.79 0.08 0.05 78% 1.05 0.08 0.04 86% 0.79 0.08 0.05 76% SRL 0.71 0.06 -0.03 84% 0.68 0.08 -0.05 72% 0.98 0.07 -0.02 88% 0.71 0.06 -0.03 82% sIPTW FEL 0.72 0.09 -0.01 66% 0.72 0.10 -0.02 61% 0.99 0.12 -0.01 66% 0.83 0.46 0.09 27% REL 0.60 0.14 -0.13 11% 0.59 0.15 -0.14 10% 0.82 0.19 -0.18 10% 0.66 0.10 -0.08 50% AEL 0.77 0.06 0.04 77% 0.74 0.05 0.01 86% 1.05 0.08 0.05 77% 0.80 0.08 0.06 66% SRL 0.69 0.06 -0.04 83% 0.67 0.08 -0.06 72% 0.96 0.08 -0.04 89% 0.70 0.07 -0.04 85% HR: Hazard Ratio; RMSE: Root Mean Square Error; sIPTW: Stabilized Inverse Probability of Treatment Weighting; SMR: Standardized Mortality Ratio.