Abstract
An evaluation was undertaken of historical industrial hygiene measurement data used in a retrospective occupational epidemiologic study of acrylonitrile workers to determine if the presence, direction, and magnitude of measurement error could be identified. The data included over 12,000 measurements taken from 1977 to as late as 1987 in eight facilities. Because there was no gold standard, subsets of the data were examined for internal consistencies. Measurement error was probably random, as the companies used recognized sampling and analytic techniques. There was little difference between the summer and winter means, suggesting that any error due to season is likely to be small. Error due to shift assignment could not be evaluated; therefore, no action was taken by the study investigators to compensate for this type of error. Two other sources that were considered to have small errors were the duration of the measurement and the presence of censored data (measurements below the limit of detection). For the former, there was little difference between the means based on longer-(≥360 minutes) and shorter-duration (<360 minutes) measurements. Therefore, means based on shorter-duration measurements were used where full-shift measurements were not available, but the estimates based on these means were distinguished from those based on the full-shift means. Censored data were treated mathematically using a nonbiased method. Comparison of means of measurements taken during typical and atypical work conditions (as defined by the plant) found that atypical means tended to be higher than means measured under typical conditions. Because use of measurements of atypical conditions in the calculation of means could result in a moderate error, the study investigators weighted the atypical measurements by the frequency of the atypical occurrences, where possible. Finally, the largest and most likely source of error was from monitoring of jobs based on their exposure concentration. This was not a concern in the epidemiologic study, however, because exposure estimates were developed for individual jobs. No evaluation could be made as to the reliability of the measurements or possible error due to nonrandom sampling. An example of the effect of error on exposure-response relationships is presented.