Effects of systematic exposure assessment errors in partially ecologic case-control studies.

Abstract

BACKGROUND: In ecologic studies, group-level rather than individual-level exposure data are used. When using group-level exposure data, established by sufficiently large samples of individual exposure assessments, the bias of the effect estimate due to sampling errors or random assessment errors at the individual-level is generally negligible. In contrast, systematic assessment errors may produce more pronounced errors in the group-level exposure measures, leading to bias in ecologic analyses. METHODS: We focus on effects of systematic exposure assessment errors in partially ecologic case-control studies. Individual-level information on disease status, group membership, and covariates is obtained from registries, whereas the exposure is a group-level measure obtained from an established exposure database. Effects on bias and coverage of 95% CI in various error situations are investigated under the linear risk model, using both simulated and empirical ecologic data on exposures that are binary at the individual level. RESULTS: Our simulations suggest that the bias produced by systematic exposure assessment errors under the linear risk model is generally approximately equal to the ratio of the slope bias and the intercept bias in ordinary linear regression with measurement errors in the independent variable. Consequently, bias in either direction can occur. Exposure assessment errors that systematically distort the group-level exposure measures have more pronounced effects on bias and coverage than errors producing random fluctuations of the group-level measures, which imply bias towards the null. CONCLUSIONS: The results indicate the need for careful consideration of potential effects of systematic distortions of the group-level exposure measures when constructing and applying group-level exposure databases, such as probabilistic job exposure matrices.