Statistical issues in radiation dose-response analysis of employees of the nuclear industry in Oak Ridge, Tennessee

Poisson regression methods are used to describe dose-response relations for cancer mortality for a subcohort of 28,347 white male radiation workers. Age specific baseline rates are described using both internal and external (US white male) rates. Regression analyses are based on an analytic data structure (ADS) that consists of a table of observed deaths, expected deaths, and person-years at risk for each combination of levels of seven risk factors. The factors are socioeconomic status, length of employment, birth cohort, age at risk, facility, internal exposure, and external exposure. Each observation in the ADS consists of the index value of each of the stratifying factors, the observed deaths, the expected deaths, the person-years, and the ten year lagged average cumulative dose. Regression diagnostics show that a linear exponential relative risk model is not appropriate for these data. Results are presented using a main effects model for factors other than external radiation, and an excess relative risk term for cumulative external radiation dose