Misinterpretation of categorical rate ratios and inappropriate exposure–response model fitting can lead to biased estimates of risk: Ethylene oxide case study

There are pitfalls associated with exposure–response modeling of human epidemiological data based on rate ratios (RRs). Exposure–response modeling is best based on individual data, when available, rather than being based on summary results of that data such as categorical RRs. Because the data for the controls (or the lowest exposure interval if there are not enough controls) are random and not known with certainty a priori, any exposure–response model fit to RRs should estimate the intercept rather than fixing it equal to one. Evaluation of a model’s goodness-of-fit to the individual data should not be based on the assumption that summary RRs describe the true underlying exposure–response relationship. These pitfalls are illustrated by Monte Carlo simulation examples with known underlying models. That these pitfalls are a practical concern is illustrated by the need for U.S. EPA to reconsider its most recent evaluation of ethylene oxide. If they had avoided these pitfalls, their exposure–response modeling would have been in better agreement with the log-linear model fit to the individual data.