Abstract:
|
Exposure measurement error causes bias in estimated relative risks (RRs) and loss of statistical power. The effect of nondifferential measurement error is rather well-known, and techniques for adjusting for measurement error, such as regression calibration approach, have become rather popular in epidemiologic research. Epidemiologists often categorize continuous exposure into quantiles. Since categorization of exposure measured with nondifferential error generally leads to differential misclassification, it is often suggested that observed RRs between quantiles might be biased in any direction. There is also a widespread belief that the comparison of extreme quantiles involves little misclassification and therefore results in smaller bias. Under a general risk model, we prove that, in spite of nondifferential misclassification, the log RR between any exposure quantiles is always attenuated with the attenuation factor equal to the correlation coefficient between true and measured exposure on the continuous scale. This result suggests a simple procedure for adjusting observed RRs and also demonstrates that the belief about the comparison of the extreme quantiles is fallacious.
|