Model Averaging Methods for the Evaluation of Dose-Response Model Uncertainty When Assessing the Suitability of Studies for Estimating Risk
A December 2013 report from the National Research Council (NRC) made a number of recommendations pertaining to evidence evaluation, systematic review, and dose-response (including meta-analysis) to the U.S. Environmental Protection Agency (EPA) for the development of an inorganic arsenic risk assessment, one of which was that EPA should focus on high-quality epidemiologic studies with low risk of bias that assess low-to-moderate arsenic exposures. After a systematic review of the literature and a thorough risk of bias analysis, EPA identified two studies of bladder and lung cancer risk in a Taiwanese population that met the NRC criteria and would support dose-response analysis. Their utility for risk assessment purposes is supported by the fact that they formed the basis of arsenic risk assessments performed by the U.S. Food and Drug Administration (FDA) and World Health Organization (WHO). The present analysis was performed to determine the suitability of these Taiwanese studies for estimating bladder and lung cancer risk to U.S. populations at levels of arsenic exposure commonly experienced in the U.S. This analysis builds upon innovative model averaging dose-response methods developed by the FDA. Reported number of cases were adjusted to account for confounders and less-than-lifetime exposure, bootstrap methods were used to estimate the µg/kg-day inorganic arsenic dose from drinking water and dietary exposures, and multiple models weighted by Bayesian Information Criterion were fit to the adjusted incidence and dose data to generate dose-specific risk estimates. One key aspect of the present analysis is that, unlike previous FDA and WHO analyses, we investigated the use of “unrestricted” models in which no constraint was placed on the slope or power parameters in order to restrict model shape. Although no substantial evidence has been found to suggest that use of the Taiwanese studies would result in systematic bias regarding differences in susceptibility factors or estimates of risk in the range of the data, substantial model uncertainty surrounds the use of restricted vs. unrestricted models and risk extrapolations to lower doses more relevant to US populations.