Inference under outcome misclassification in health risk models using a simulation study with a validation dataset

Death certificates are commonly used in epidemiological studies investigating the relationship between exposure and health outcomes. It is known that death certificates may misclassify the underlying causes of death, and it is commonly understood that if misclassification is non-differential, it will bias dose-response relationships toward the null or underestimate the association. This simulation study explores the probability that results of an individual study may contradict the general understanding by addressing two key questions: (1) what is the probability that misclassification of disease mortality moves measures of dose-response associations away from the null? and (2) what is the probability that misclassification moves measures of dose-response associations away from the null sufficiently to change the conclusion of a study from statistically non-significant to significant? As the starting point, this simulation study used a small group of radiation-exposed nuclear workers for whom both death certificates and autopsy reports were available. Results suggest that nominally non-differential misclassification can lead to an odds ratio that moves away from the null. For datasets where the initial p-values were slightly non-significant, the percentage of odds ratios that moved away from the null generally decreased with higher levels of misclassification, and the probability that the p-values associated with these odds ratios would change to significant decreased with increasing misclassification rates. The traditional heuristic is more likely to be true when: (1) there is a larger misclassification rate, and (2) there is a high association between dose and disease mortality. This has implications for environmental epidemiology, such as low-dose radiation epidemiology, where estimated effects are often small and conclusions may hinge on marginal statistical significance. As another implication, these findings apply broadly to various health outcomes, even if the outcome misclassification rate is low.