A statistical method for imputation of uranium-235 lung burden using urinalysis data

Epidemiologic studies of groups of radiation workers are being performed by the Oak Ridge Associated Universities' Center for Epidemiologic Research in order to identify radiation-induced health effects and to quantify potential dose-response relationships. The historical radiological monitoring data sets available for a number of groups under study do not directly provide sufficient information to permit calculation of workers' doses. An example of this problem has been encountered with the Fernald Feed Materials Production Center studies. Biological monitoring data for workers exposed to airborne uranium dust in operations there consists of a large set of urinalysis records reporting mass quantities of uranium and a small set of lung count records reporting quantities of Uranium-235 and Uranium-238 present in the lung, estimated by gamma spectroscopy. Bayesian linear regression and prediction techniques have been applied to this data to enable the prediction of lung burdens of Uranium-235 in workers based on their urinalysis results and a "calibration" data set of urinalysis results and measurements of Uranium-235 lung activity. The output of the procedure is a Uranium-235 lung burden estimate provided as a probability density. Computer programs have been written to facilitate large-scale application of the procedure. As the precision of the input urinalysis data was not known, an exploration was made of the effects of uncertainty in this data set on regression parameters and prediction . A computer program was written to simulate the effect of uncertainty in the urinalysis measurements. A random value was chosen from an assumed distribution about each urinalysis result in the calibration data set and a representative urinalysis result for which a prediction was desired. The regression and prediction procedures were then performed, and this process was repeated a number of times. Distributions of results were obtained which indicate the effects of uncertain urinalysis values on regression parameters and prediction.