Approaches to harmonize mortality data sets in three diverse radiation worker cohorts

While there is a well-established link between ionizing radiation and cancer, there are uncertainties with effects following low doses delivered at low dose rates. To address these gaps, the ongoing Million Person Study of Radiation Workers and Veterans (MPS) is investigating the likelihood of a variety of cancer and non-cancer effects following chronic exposure to low dose-rate ionizing radiation. One challenge is and will be combining and harmonizing diverse cohorts with widely different measures of socio-economic status, birth cohorts, dose distributions and sex ratios. Herein, we have evaluated non-cancer mortality in three cohorts for which dose reconstructions have been completed: Rocketdyne (Atomics International, California, 1948-2008), Mound (Dayton, Ohio, 1944-2009) and nuclear weapons test participants (Atomic Veterans, 1945-2012). These three cohorts represent a small fraction of the overall MPS but provide valuable insight into methods of combining and harmonizing data from multiple diverse cohorts that can later be considered for all MPS cohorts. Heart disease mortality, including both underlying and contributing causes of death, was chosen for illustrating the statistical approaches. In all three cohorts, radiation dose estimates were distributed very differently by different measures of socio-economic status. Further, the effect of birth cohort was significantly different for heart disease mortality in all three cohorts, with all studies showing that later birth cohorts have lower rates of heart disease mortality than the earlier. The goal of this paper is not to quantify radiation effects based on these combined cohorts and it would be inappropriate to do so. Rather these cohorts are used to illustrate approaches for combining multiple data sets that incorporate the full set of individual confounder and cofactor information available from each cohort, though widely different. We identified five different methods to combine the results of these three datasets: the simple pooled analysis (PA), PA including study interactions, traditional stratified analysis, and both fixed and random effects meta-analysis. We describe the similarities and differences between the combined results using these approaches.