A potential pitfall in control of covariates in epidemiologic studies

Control of covariates is essential in nonexperimental epidemiologic studies. Important covariates, such as smoking or alcohol consumption, often are crudely categorized in epidemiologic analyses. In this paper, I illustrate by both hypothetical and empirical examples that control of crudely categorized covariates can yield strongly misleading results. In particular, I show that, under certain conditions, control for crudely classified covariates can even be worse than not controlling for such covariates at all. I conclude that covariate specification is an issue that requires much more care than it commonly receives in epidemiologic analyses.