Presentation (CTI)

OBJECTIVES: Real-world external control arm (RW-ECAs) are increasingly used to inform regulatory and HTA submissions. Due to reduced oversight and different criteria to define outcomes in clinical practice vs trials, differential outcome misclassification is a potential source of bias when comparing trial data to RW-ECA studies. Quantitative bias analysis (QBA) is increasingly used to assess the impact of unobserved confounding, but its broader use is limited. This study uses simulation to evaluate the impact of differential outcome misclassification on treatment effect estimates and applies QBA approaches to adjust for misclassification bias.
METHODS: We simulated a hypothetical single-arm trial and an RW-ECA with binary outcome data. The trial arm was assumed to have no outcome misclassification, while three observed control datasets were generated with varying levels of misclassification, based on pre-specified sensitivity and specificity parameters. Treatment effects were reported as odds ratios (ORs). We applied (1) deterministic QBA, correcting control outcome counts using known sensitivity and specificity values, and (2) probabilistic QBA, where sensitivity and specificity were treated as random variables with defined distributions (e.g., uniform) and adjusted estimates were obtained via Monte Carlo simulation (10,000 replications). The performance of each method was assessed by comparing bias-corrected estimates to the true effect.
RESULTS: The true treatment effect estimate was OR 0.53 (95% CI, 0.30, 0.94). Before correction, OR estimates ranged from 0.60 to 0.84, after deterministic correction the OR estimates ranged from 0.43 to 0.57 across misclassification scenarios. Probabilistic QBA produced median OR estimates ranging between 0.43 and 0.57.
CONCLUSIONS: Outcome misclassification in studies with an RW-ECA can meaningfully bias treatment effect estimates. Deterministic and probabilistic QBA methods provide feasible approaches to adjust for misclassification, with probabilistic methods offering a more flexible framework to capture parameter uncertainty.