The choice of the risk to dose–response model that best describes radiation epidemiological data is a main step in radiation-related health risk assessment. Often, different models are published which all fit the data similarly well and are all deemed plausible by various groups of the scientific community. With the diversity of available models that have a comparable goodness of fit to the data, a source uncertainty arises when assessing radiation health risks with only one model: the uncertainty arising from model choice. One technique which may be applied to address this source of uncertainty is multi-model inference (MMI), which allows a composite or averaged model based on several plausible models to be built. For this purpose, different plausible models are fitted to the same dataset, and their goodness of fit is quantified via a statistical measure (e.g. AIC or BIC). The composite model is then built as a weighted mean of all the considered models, where the value of the measure is used to calculate the weight for each model. A review of several articles applying MMI in radiation-related risk assessment for different outcomes is presented here. Additionally, a new approach to overcome an inherent problem of the MMI approach, which clearly penalises excess relative risk models with stratified baseline models due to the high number of parameters compared to risk models with parametric baselines, is illustrated. Finally, the advantages of the different statistical measures to quantify the goodness of fit are elucidated, and results obtained with a newly proposed multi-method–multi-model inference (M 4 I) approach are presented. This M 4 I approach offers a possibility to generate a single risk estimate based on MMI risk estimates calculated with different statistical measures. Generally, it is recommended to consider the uncertainty of model choice, by applying MMI and considering M 4 I, in radiation risk analyses.
Hafner et al. (2026) studied this question.