Large errors in measurement of the exposure variable are frequently encountered in occupational cancer studies. In general, these errors lead to bias of effect estimates towards the null and underestimation of the width of confidence intervals. Although statisticians have proposed numerous methods for treating this problem, most of these methods are not appropriate for occupational cancer studies, because; a) they make unrealistic assumption about the process generating error in the measurement of the exposure variable, b) they are not applicable to the survival data methods required for the analysis of cohort studies, and c) they do not consider interval estimation, but only address point estimation. In addition, the few developments which do not have the above limitations, have not made their way into standard practice because; a) they have not considered the particular issues which arise when a cumulative exposure variable is of primary interest, and b) the proper use of these methods requires a detailed understanding of the statistical literature and few realistic examples exist of practical applications. The work proposed in this grant, will develop new methods and adapt existing ones for use in occupational cancer studies with cumulative exposure as the variable of interest with none of the above limitations. Fully parametric likelihood methods, as originally proposed by Prentice in (1982), and new, semi-parametric methods, as elaborated by Robins et al. (1994, 1995), will both be considered. These methods will be applied to important data sets taken from occupational epidemiology: Samet et al. s 1991 study of the effects of occupational exposure to radon gas in relation to lung cancer mortality among New Mexico uranium miners, and Savitz et al. s 1995 study of the effects of occupational exposure to magnetic fields in relation to mortality from leukemia and brain cancer among workers in five electric power plants. Both of these studies have detailed validation data from which parametric and non-parametric measurement error models will be constructed, and used to obtain point and interval estimates of the exposure effect which are not contaminated by measurement error. In addition to contributing to the advancement of statistical methods available in this setting, the investigators will advance scientific understanding of the extent of the exposure-disease relationships investigated in these two studies, and provide a more realistic quantification of the uncertainty around this understanding.
