Methodological extension of the measurement error model which adjusts dosimetric errors

William F Anderson, BB/DCEG/NCI 
*Kazutaka Doi, National Institute of Radiological Sciences 

Keywords: measurement error, dose uncertainty, classical error, Berkson error

In analyses of radiation epidemiology data, regression methods are used to investigate the dose-response relationship. Errors in the outcome variable is assumed in usual regression models, and this may resulted in wider confidence intervals, not in biased estimates. Although errors in covariates are not assumed in the usual models, and this may be resulted in biased estimates. To decrease the bias, the measurement error models have been developed, which decrease the bias by quantifying the magnitude of covariate errors. The most common measurement error in radiation epidemiology is dose uncertainty, and the distinguishing features of dose uncertainty are the following: (1) dose errors are more homogeneous on a multiplicative than on an additives scale; (2) Two kinds of covariate errors (Berkson and classical) are known to exist. In analyses of A-bomb survivors, the regression calibration method proposed by Pierce et al. has been used because it is easy to be applied for various purpose of studies. Although Berkson errors are not fully handled in their method because their magnitude is known to be relatively small. We extended their method to handle Berkson errors appropriately, and enable to be applied to other studies. From the result of simulation studies, our proposed method works fairly well in the situation where the magnitude of Berkson errors are large.