Methodology and utility of a job‐exposure matrix

Abstract

We have previously reported a study in which a job‐exposure matrix was applied to census data, identifying, e.g., polychlorinated biphenyls (PCBs) and creosote as increasing the risk of urothelial cancer. In this article, we expand on some theoretical issues, and present detailed accounts of constructed linkages for PCBs, creosote, and phenols. For agents of interest, one should emphasize the positive predictive value rather than the sensitivity in the construction of the matrix. The reverse is true for confounding factors; to avoid residual confounding after restriction to subjects unexposed for the confounding factors, one should emphasize sensitivity, possibly compromising the positive predictive value. This discrepancy between agents of interest and confounding factors may limit the application of a general matrix for studying several different diseases. The construction of the matrix is much harder, if sensitivity rather than positive predictive value is emphasized for an agent. Confounding from industry‐related agents arises due to a true mixed exposure in certain work tasks, but also due to a gross classification of occupations in the census. One should not confuse different levels of the positive predictive value with exposure dose. A “dose‐response” with different levels of positive predictive value reflects an accuracy of the matrix, not a biological phenomenon. Studies with exposure information from a job‐exposure matrix applied to registers with scant information on occupation and industry may be warranted for exposures and diseases for which previous studies with a detailed documentation of exposure have low precision.