A semiparametric approach to evaluate the harm of low-dose exposures

While moderate to high levels of radiation exposure is known to cause adverse health effects, there is still controversy about the lowest dose that could be harmful. Given that epidemiological studies of practical sizes are unlikely to provide sufficient statistical power to detect a small risk in the low-dose range of concern, greater emphasis should be given to evaluating low-dose risk uncertainty. Using simulations under various dose-response relationships with a threshold, we show that a conventional approach based on simple parametric models (e.g. the linear model with or without a threshold) can be inefficient, biased and/or inaccurate in uncertainty evaluations at low doses. Alternatively, we consider a Bayesian semiparametric model of a connected piecewise-linear function allowing for autocorrelations between adjacent line sections. With no specific assumption, this can describe various plausible dose-response curves while appropriately handling the risk uncertainty. In particular, it can relatively accurately evaluate the dose range in which a threshold might exist, while retaining statistical power for a small risk increase after the threshold. As an illustration, we analyse cancer incidence data of Japanese atomic bomb survivors, a primary epidemiological source of quantitative risk estimates for health effects from radiation exposure.