Bias and Random Errors in Historical Data of Exposure to Organic Solvents

This is the authors' abstract. We don't add key points for this paper.

The biases and random errors in data from industrial hygiene measurements have been evaluated. Such data are often used for exposure assessment in epidemiologic studies by people who are not professional industrial hygienists. The quality of such data is, therefore, of paramount importance. Data sets from two investigations of the exposure to organic solvents in the same industry have been analyzed. The study shows that the errors and analysis of field samples are larger than errors determined from laboratory data but smaller than errors from other sources. One exception, in which laboratory errors are large, is when data originate from different ways of determining recovery. Very large biases may arise for certain combinations of analyses and extraction liquid when the “equilibrium method” is used. The data sets from the two investigations had a bias, which was as large as a factor of 5 to 10. This bias is believed to be caused by the use of two different sampling strategies. In all industrial hygiene, data, large random errors are embedded because of day-to-day variations in exposure levels; therefore, the data are mostly used grouped into three classes of low, medium, and high exposure. To do so is meaningful, as a large amount of data will compensate for misclassifications. But a large amount of data cannot compensate for the bias introduced when merging data generated under different sampling strategies. Data describing the same object, but originating from different sampling strategies, should not be pooled without proper tests, which show that data are comparable. If industrial hygiene data are collected in databases and used for exposure assessments, a strict classification of data in groups having the same or almost the same bias should be established and enforced. In this paper it is shown that a large number of classes are necessary to cover the different sources of bias found in industrial hygiene measurements. It is proposed that a set of rules be established for classification of data before they are stored in databases. Such a set of rules would cause a huge amount of data to be disregarded for the use as historical data, but classification in more homogeneous classes may more than compensate for the loss of data.