Compilation of NTP and published dose‐response data: study of informative priors in the EFSA platform for Bayesian benchmark dose analysis

Abstract In an update in 2020, the WHO proposed the use of Bayesian models expressed in terms of “natural” parameters and “technical” parameters. In its guidance document (EFSA, 2022), the Bayesian paradigm was also recommended by EFSA. It allows a fully probabilistic approach with distributions on all parameters and models. The paradigm formalizes the combination of (new) data with (historical) prior knowledge. However, more guidance is needed to allow users to fully exploit the potential of using informative priors. A repository of prior distributions is created. Based on criteria for screening the NTP database and EFSA journal, data were retrieved for 228 substances from the NTP database and 41 substances from the EFSA journal. Compounds were classified on different levels, e.g., chemical structure and composition, toxicological endpoints, and mode(s) of action. Methodology and R functions were developed for constructing informative PERT priors, including different strategies to combine multiple historical studies. No method for combining multiple studies outperformed the others in all scenarios of a simulation study, but the approach of mixing posteriors is recommended. The most challenging part was the development of a methodology to create an overarching prior distribution for groups of chemicals. After the selection of the toxicological criteria for grouping, studies underwent a further selection based on statistical criteria. A key component of the methodology is the standardisation of the dose scale. In general, the use of informative priors resulted in an increase in precision of the BMD estimate. The analysis of all the historical datasets underlying the repository provided further insights into the technical parameter d. Attempts to optimise the default prior for the d parameter did not result in a consistent and structural improvement of the BMD estimators. The project ends with prospective views on future pathways to further enhance and optimise EFSA's Bayesian model averaging approach.