Measurement Error

The term measurement error is different from exposure misclassification because the former implies a continuous variable, while the latter suggests a dichotomous one. Exposure to environmental toxicants is a continuous variable in the real world, and one of the most important improvements in exposure assessment is for studies to move from dichotomizing exposure into continuous, or at least multilevel, measurements.

For example, consider a study relating air pollution to respiratory illness. The outdoor ambient-air pollution concentration is available from community monitoring. The analysis of these data will seek to correlate variations in air pollution with variations in respiratory outcome. Variations in recorded levels of air pollution may be thought of as having the following components: measurement error associated with the monitoring instrument, variation in the amount of time individuals spent outdoors, geographic variation in the outdoor concentration of the pollutant in the vicinity of the monitor, variations in the indoor/outdoor ratio, and individual variations in delivered dose. Critical issues include the size of each error component and the cost of reducing each component.


One of the most important advantages of improved exposure assessment derives from its impact on misclassification. Small errors in exposure assignment may have dramatic results on estimation of effect. Because of the limited scope of exposure assessments in most environmental epidemiology, misclassification is likely to be a substantial problem. In general, the internal validity of an epidemiologic investigation can be reduced by misclassification of study subjects. Copeland et al. (1977) emphasize that bias from misclassification will be ''a function of the sensitivity and specificity of the classification procedure, the disease frequency, and exposure frequency." And in all case-control studies the bias depends on whether misclassification is the same or different in cases and controls, that is, nondifferential or differential in cases and controls.

In general, nondifferential misclassification causes measures of effect to be biased toward the null value. Such misclassification produces an underestimate of the effect, whereas differential misclassification can result in bias either toward or away from the null value. Copeland et al. (1977) argue that classification errors cannot be ignored and that investigators should attempt to estimate the magnitude of the errors. Dosemici et al. (1990) have shown that the predominant view that "nondifferential misclassification of exposure can only bias an estimate of a true positive odds ratio downward and not away from or beyond the null value" may

The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement