Independent analysis of the radiation risk for leukaemia in children and adults with mortality data (1950-2003) of Japanese A-bomb survivors.

A recent analysis of leukaemia mortality in Japanese A-bomb survivors has applied descriptive models, collected together from previous studies, to derive a joint excess relative risk estimate (ERR) by multi-model inference (MMI) (Walsh and Kaiser in Radiat Environ Biophys 50:21-35, 2011). The models use a linear-quadratic dose response with differing dose effect modifiers. In the present study, a set of more than 40 models has been submitted to a rigorous statistical selection procedure which fosters the parsimonious deployment of model parameters based on pairwise likelihood ratio tests. Nested models were consequently excluded from risk assessment. The set comprises models of the excess absolute risk (EAR) and two types of non-standard ERR models with sigmoidal responses or two line spline functions with a changing slope at a break point. Due to clearly higher values of the Akaike Information Criterion, none of the EAR models has been selected, but two non-standard ERR models qualified for MMI. The preferred ERR model applies a purely quadratic dose response which is slightly damped by an exponential factor at high doses and modified by a power function for attained age. Compared to the previous analysis, the present study reports similar point estimates and confidence intervals (CI) of the ERR from MMI for doses between 0.5 and 2.5 Sv. However, at lower doses, the point estimates are markedly reduced by factors between two and five, although the reduction was not statistically significant. The 2.5 % percentiles of the ERR from the preferred quadratic-exponential model did not fall below zero risk in exposure scenarios for children, adolescents and adults at very low doses down to 10 mSv. Yet, MMI produced risk estimates with a positive 2.5 % percentile only above doses of some 300 mSv. Compared to CI from a single model of choice, CI from MMI are broadened in cohort strata with low statistical power by a combination of risk extrapolations from several models. Reverting to MMI can relieve the dilemma of needing to choose between models with largely different consequences for risk assessment in public health.