Using "exposure prediction rules" for exposure assessment: An example on whole-body vibration in taxi drivers

Background: It is often difficult and expensive to make direct measurements of an individual's occupational or environmental exposures in large epidemiologic studies. Methods: In this study, we used information collected in validation studies to develop a prediction rule for assessing exposure in a study with no direct measurement. We established a prediction rule through mixed-effect modeling of direct measurement data and information on observable exposure predictors and their interactions. Specifically, we used 383 measures of whole-body vibration from 247 professional taxi drivers and attempted to quantify vibration exposures for individuals in a large study on low back pain. Results: Using the "jackknife method," we found that our prediction rule had an acceptably low relative prediction error of 11% (95% confidence interval-10-12%). Implementing the prediction rule would result in measurement errors independent of low back pain and of all identified and observable predictors of whole-body vibration. We applied the predicted levels to compute each person's daily exposure, and found a strong association between the predicted daily whole-body vibration exposure and prevalence of low back pain. This supported the construct validity of the exposure prediction rule. Conclusions: The predictive and construct validity of our prediction rule suggests that this general statistical approach can be useful in other occupational settings to improve the quality of exposure assessment.