American Public Health Association Meeting & Expo, Washington, DC, November 2-5, 2025
Xirui Liu presented her work to quantify the impact of death certificate cause-of-death errors on the conclusions of health risk models at the 2025 Annual American Public Health Association Meeting & Expo. Her presentation was the third in a series of presentations during a special session titled, “Advanced low-dose radiation health effects research: insights from the Million Person Study and key U.S. radiation cohorts”. This special session, which was organized by individuals associated with the Million Person Study, included three additional presentations: “U.S. Million Person Study of low-level and low-dose radiation health effects: importance, information, and innovation” (presented by Lawrence Dauer); “50 million person year radiation dosimetry for the U.S. Million Person Study radiation epidemiology study” (Michael Bellamy); and “Colossus open-source science: bridging the gap between big data and radiation epidemiology for the Million Person Study and beyond” (Amir Bahadori).
Sensitivity analysis to quantify the impact of outcome misclassification on health risk models: a simulation approach
Xirui Liu (USTUR), Stacey L. McComish (USTUR), Joey Y. Zhou (DOE), Sergey Y. Tolmachev (USTUR)
Background: The United States Transuranium and Uranium Registries (USTUR) investigates the deposition of uranium, plutonium, and americium in humans who had documented occupational exposures to these elements. Our examination of autopsy reports from these workers revealed cause-of-death discrepancies between death certificates and autopsy reports, where death certificates are commonly used in epidemiological studies and autopsy reports are widely recognized as the diagnostic gold standard. Traditional epidemiological assumptions suggest such misclassification would bias risk estimates toward the null.
Objectives: This simulation study quantified whether cause-of-death misclassification can (1) move odds ratios away from the null, and (2) shift non-significant dose-response associations to significance. Furthermore, this study aimed to determine the likelihood of these effects in a borderline significant scenario.
Methods: Simulation data included doses and outcomes, which were either derived from USTUR records or randomly generated. Initial datasets were generated with borderline statistical significance (p ≈ 0.05). Various over- and under-misclassification rates were used for simulation. 20,000 iterations were simulated for each combination of over- and under-classification. The logistic regression model was used to evaluate dose-outcome associations.
Results: Over- and under-misclassification rates, varying from 0 to 30%, cause-of-death misclassification moved odds ratios away from the null in 4%–47% of the simulations. In 8%–42% of cases, non-significant associations shifted to significant.
Conclusion: The traditional understanding that misclassification always biases estimates toward the null is more likely to be true when misclassification rates are high and the risk is high. [USTUR-0716-25A]