Quality Assurance Testing of Version 1.3 of U.S. EPA Benchmark Dose Software (Presentation)
EPA benchmark dose software (BMDS) issued to evaluate chemical dose-response data in support of Agency risk assessments, and must therefore be dependable. Quality assurance testing methods developed for BMDS were designed to assess model dependability with respect to curve-fitting, rate of convergence on a benchmark dose solution, and reliability of benchmark dose estimates. Each model in BMDS version 1.3 was evaluated comparing them to an equivalent non-EPA model, as well as the previous version of the BMDS model, using data sets obtained from EPA cancer and non-cancer data bases. Different configurations (e.g., various response levels and parameter restrictions) for each of the EPA and non-EPA models were applied to 100 dichotomous or continuous dose-response data sets, resulting in well over 1,000 runs per model. Curve-fitting results were assessed by comparing maximum log-likelihood values among the models. Results obtained to data indicate that the BMDS dichotomous models are as dependable and reliable as the non-EPA models employed, with BMD and BMDL convergence rates ranging from 96 to 100%. BMDS continuous model results were plausible and convergence rates ranging from 96 to 100%. BMDS continuous model results were plausible and internally consistent, with high convergence rates (85% and 100%), but were not as reproducible by the non-EPA models. Given that none of the non-EPA continuous models employed have received the type of public and peer review that the BMDS models have received, it is reasonable to assume that most of the observed differences between these models were due to errors in the non-EPA models. One of the three BMDS nested models )logistic) performed at least as well as its equivalent, non-EPA model. The other two BMDS nested models (NCTR and Rai and Van Ryzin) were found to be adequate, but inferior to their non-EPA models with respect to how often they provided results and the plausibility of the results. Thus, these two BMDS nested models are considered less reliable than the BMDS nested logistic model.