Presentation (CTI)

OBJECTIVES: Real-world evidence is increasingly used to inform clinical and policy decisions, but causal estimates derived from observational data remain vulnerable to unmeasured confounding. Quantitative bias analysis (QBA) can assess this vulnerability. However, QBA requires specification of bias parameters that may be informed by expert knowledge or external evidence, but selecting plausible ranges remains challenging in many real-world evidence settings. We propose and evaluate a data-informed sensitivity analysis approach that leverages observed data structure to define plausible ranges for unmeasured confounding in real-world evidence.
METHODS: We simulated observational health data that included non-random exposure assignment, measured confounding, and residual unmeasured confounding. The exposure-outcome association was estimated using regression adjustment for measured covariates, yielding a risk ratio (RR). Conventional deterministic and probabilistic QBA were applied using investigator-specified bias parameters. A data-informed approach was then implemented, where random forest models were used to characterize the strength of association between measured confounders and the exposure and outcome. These estimates were used to define plausible parameter ranges for the association between an unmeasured confounder and the outcome.
RESULTS: Adjusted RR was 2.13 (95% onfidence interval: 1.85-2.46), compared to the true data-generating RR=1.73. In deterministic QBA, weak confounding assumptions (prevalence of the unmeasured confounder among exposed/unexposed: 0.30/0.10; RR between unmeasured confounder and outcome=1.25) minimally attenuated the estimate (RR=2.04). Stronger confounder-outcome association and larger prevalence imbalances (RR=2-3), produced adjusted RRs=1.8 to 1.1. Probabilistic QBA produced a median adjusted RR=1.95 (95% simulation interval: 1.59-2.13), failing to recover the true effect under the assumed parameter distributions. The data-informed approach yielded adjusted estimates ranging from 2.10 to 1.73, with unmeasured confounder-outcome associations closer to the strongest observed confounders, closely approximating the true effect.
CONCLUSIONS: Benchmarking bias parameters against observed confounder structures provides a practical framework for sensitivity analyses that address unmeasured confounding in causal estimates derived from real-world evidence.