False Indications of Dose-Response Nonlinearity in Large Epidemiologic Cancer Radiation Cohort Studies; A Simulation Exercise
This study explores the likely prevalence of false indications of dose-response nonlinearity in large epidemiologic cancer radiation cohort studies (A-bomb survivors, INWORKS, Techa River). Reasons: Increasing numbers of tests of nonlinearity are being made in studies. Hypothesized nonlinear dose-response models have been justified to policy makers by analyses that rely in part on isolated findings that could be statistical fluctuations. After removing dose nonlinearity (linearization) by adjusting person-years of observation at each dose category, indications of nonlinearity, necessarily false, were counted in 5,000 randomized replications of six datasets. The average frequency of any false positive for five indicators of nonlinearity tested against a linear null was roughly 25% in Monte Carlo simulations per study, consistent with binomial calculations, increasing to ~50% within 6 studies assessed. Comparable frequencies were found using Akaike's information criterion (AIC) for model selection or multi-model averaging. False above-zero threshold doses were found more than 50% of the time, averaging to 0.05 Gy, consistent with findings in the 6 studies. Such bias, uncorrected, could distort meta-analyses of multiple studies, because meta-analyses can incorporate high P value findings. AIC-based correction for the extra threshold parameter lowered these false occurrences to 8 to 19%. Given the simulation rates, the possibility of false positives might be noted when isolated findings of nonlinearity are discussed in a regulatory context. When reporting a threshold dose with a P value > 0.05, it would be informative to note the expected high false prevalence rate due to bias.