Iterative allocation of partially classified data in occupational epidemiologic studies

To allocate partially classified person-years in an occupational mortality study, a model-based iterative allocation (IA) algorithm is developed using the EM algorithm (Dempster AP, Laird NM, Rubin DB, JRSS, 1977). The IA is an adaptation and extension of a method proposed by Gong et al. (JASA, 1990). Gong et al. (JASA, 1990) considered misclassification involving a categorical non-time-dependent covariate in a Poisson regression setting to estimate the distribution of events and person-time for correctly classified individuals. We have implemented the methods of Gong et al. (JASA, 1990) in S-PLUS (1993) and have adapted and extended them in several directions for the purpose of classifying deaths and person-years in the presence of misclassification. Specifically, the IA incorporates two-time-dependent covariates and partial misclassification in one non-time-dependent covariate. The IA utilizes individual-level data and incorporates either internal cohort rates or external population-based rates into the individual-level imputation. Person-years generated using OCMAP-PLUS (Marsh et al., 1995) are allocated according to predicted probabilities that an individual worker is white. Variance estimates of the estimated person-years are computed using the bootstrap (Efron, 1982). The IA is compared to a non-parametric proportional allocation method (PAM) proposed by Marsh et al. (JOHS-Aust NZ, 1996). Both methods of allocating race-specific person-years gave similar estimates when applied to an occupational cohort in which racial information is unknown for a substantial number of workers.;A simulation study was implemented to assess assumptions and to assess the validity and the reliability of the IA. Ten datasets were simulated based on known data for a large cohort, by assigning racial data to be missing completely at random for some workers. Multiple imputation variances were computed for the mean estimated person-years. The IA was validated by comparing total age-time-race-specific person-year contributions to the corresponding quantities using the known data. The IA produces reliable estimates of age-time-race-specific person-years with a relative variation of the estimated mean age-time-race-specific person-years of about 5% and 10% for white and nonwhite workers, respectively.