Five misconceptions about categorizing exposure variables measured with error in epidemiological research

Abstract In the analysis of epidemiological studies, categorizing continuous exposures, using either data-driven quantiles or predefined thresholds, is a common practice. Although an abundant literature describes pitfalls of categorization, substantial concerns which have not been adequately addressed arise when the original continuous exposure is measured with error. Measurement error in a continuous exposure leads to misclassification following categorization. This paper aims to dispel five misconceptions regarding the impact of measurement error in a continuous exposure that is categorized. First, that categorization could help infer the functional form of an exposure–outcome association. Second, that misclassification resulting from categorizing a continuous exposure with non-differential measurement error will also be non-differential (ie, independent from the outcome). Third, that categorization will necessarily mitigate the impact of measurement error compared to the continuous case. Fourth, that comparing extreme quantiles instead of comparing adjacent quantiles may reduce bias. Finally, that an estimated association is necessarily attenuated (toward the null), which we argue only holds in regressions with one mismeasured exposure (and may even be reversed in the presence of error-prone confounders). Consequently, epidemiologists who categorize continuous error-prone exposures should be aware of those misconceptions and appropriately discuss the expected impact of categorization on the estimated exposure–outcome associations.