Occupational epidemiologists frequently encounter large errors in measurement of exposures of interest, particularly in studies in which exposure must be retrospectively assessed. Even in studies where exposure assessment is concurrent with the collection of outcome data, measurement errors in occupational studies are often believed to be considerable. In general, these errors lead to bias of effect estimates toward the null and underestimation of the width of confidence intervals. Although statisticians have proposed numerous methods for treating this problem, these methods have not yet made their way into standard practice, because a) the proper use of these methods requires a detailed understanding of the statistical literature, and few examples exist of practical applications, b) the use of most of these methods requires custom software and time- consuming calculations, and c) they are often not appropriate for typical features of occupational studies. The work proposed in this grant will address each of these three obstacles. New and existing measurement-error methods will be applied to important occupational data sets. The GM/UAW study of the risk of respiratory cancer from machining fluids exposure is a retrospective cohort study of 40,000 workers. The New Mexico uranium miners study is also a retrospective design, investigating the relationship between radon decay products and lung cancer risk by following 3,469 miners. The ACE study is a cross- sectional design investigating the health effects of occupational exposure to anti-cancer drugs in 3550 nurses, pharmacists and nurses' aides. In all of these studies, more accurate exposure assessment has occurred in a subsample of work environments within some of the time periods during which the subjects are followed. Measurement error models will be developed from these more detailed data and used to produce estimates of relative risk which will be corrected for bias due to measurement error. Confidence bounds around these estimates will incorporate the additional variability due to uncertainty about """"""""true"""""""" exposure values. The analysis of retrospective cohort, case-control and cross-sectional data will be considered using logistic regression, Cox regression, and other methods for """"""""failure time"""""""" data, as required. When published, the analyses will provide other investigators with sound examples of the use of these techniques. The primary goal of this grant is to develop methods suitable for the types of studies and data most frequently encountered in occupational epidemiology.
| Spiegelman, D; Valanis, B (1998) Correcting for bias in relative risk estimates due to exposure measurement error: a case study of occupational exposure to antineoplastics in pharmacists. Am J Public Health 88:406-12 |
