Regression Calibration for Classical Exposure Measurement Error in Environmental Epidemiology Studies Using Multiple Local Surrogate Exposures

• Published in: American journal of epidemiology Vol. 172; no. 3; pp. 344 - 352
• Main Authors: Bateson, Thomas F., Wright, J. Michael
• Format: Journal Article
• Language: English
• Published: Cary, NC Oxford University Press 01.08.2010; Oxford Publishing Limited (England)
• Subjects: bias (epidemiology); Biological and medical sciences; Calibration; Data Interpretation, Statistical; Environment. Living conditions; environmental exposure; Environmental Exposure - statistics & numerical data; Environmental studies; epidemiologic methods; Epidemiologic Studies; Epidemiology; Exposure; General aspects; Human exposure; measurement error; Measurement errors; Medical sciences; Miscellaneous; misclassification; Performance assessment; Public health. Hygiene; Public health. Hygiene-occupational medicine; Regression analysis; surrogate marker; water pollution; Multiple; Biological marker; Regression; Exposure; Regression analysis; Calibration; Epidemiology; bias (epidemiology); Health and environment; epidemiologic methods; surrogate marker; environmental exposure; Environment; Methodological bias; Water pollution; Measurement error; Public health; misclassification
• ISSN: 0002-9262; 1476-6256; 1476-6256
• DOI: 10.1093/aje/kwq123

Environmental epidemiologic studies are often hierarchical in nature if they estimate individuals’ personal exposures using ambient metrics. Local samples are indirect surrogate measures of true local pollutant concentrations which estimate true personal exposures. These ambient metrics include classical-type nondifferential measurement error. The authors simulated subjects’ true exposures and their corresponding surrogate exposures as the mean of local samples and assessed the amount of bias attributable to classical and Berkson measurement error on odds ratios, assuming that the logit of risk depends on true individual-level exposure. The authors calibrated surrogate exposures using scalar transformation functions based on observed within- and between-locality variances and compared regression-calibrated results with naive results using surrogate exposures. The authors further assessed the performance of regression calibration in the presence of Berkson-type error. Following calibration, bias due to classical-type measurement error, resulting in as much as 50% attenuation in naive regression estimates, was eliminated. Berkson-type error appeared to attenuate logistic regression results less than 1%. This regression calibration method reduces effects of classical measurement error that are typical of epidemiologic studies using multiple local surrogate exposures as indirect surrogate exposures for unobserved individual exposures. Berkson-type error did not alter the performance of regression calibration. This regression calibration method does not require a supplemental validation study to compute an attenuation factor.