A new bayesian framework for dose-response modelling and benchmark dose determination

Abstract

Introduction A new Bayesian framework for dose-response modelling and benchmark dose determination is introduced, following guidelines from the World Health Organization and the European Food Safety Authority. It incorporates model averaging across a family of candidate models and defines prior distributions at the level of response distributions, dose-response median models, and model parameters. In addition to providing a fully probabilistic inferential framework, the Bayesian paradigm offers some methodological advantages over frequentist approaches.Methods The proposed framework employs regularising default priors that are adaptive to study design and biologically plausible constraints. Informative priors can also be incorporated. The performance was evaluated against frequentist model averaging across four case studies using both real and simulated datasets.Results In simulations based on the Bisphenol A study, only 3% of BMD estimates obtained using frequentist inference met the EFSA accuracy criterion (BMDU/BMDL < 50), compared with 83% obtained using Bayesian inference. Other analyses showed that design-adaptive and biology-respecting regularising priors improved estimation relative to frequentist hard constraints. In a small-sample case study with poorly informative experimental design, frequentist quantile bootstrap confidence intervals for the BMD were highly sensitive to the number of bootstrap replicates, whereas Bayesian MCMC-based inference remained stable and reliable. A final case study demonstrated that constructing an informative overarching BMD prior from multiple historical studies improved estimation performance.Discussion The case studies illustrate improved estimation and the ability to incorporate historical information. These findings support the use of Bayesian methods as a powerful alternative for regulatory toxicology and risk assessment applications.