Poster: A Case Study on Model Averaging for Risk Assessment

Model averaging (MA) uses an average of multiple model curves to estimate the benchmark dose (BMD).  MA is more flexible in accounting for model uncertainty than single model selection and has been shown to have superior statistical properties to single model selection.  MA is included in U.S. EPA’s Benchmark Modeling Software (BMDS) for dichotomous data. It is included in the R package ToxicR for both continuous and dichotomous data.  Case study for chloroform:  MA was applied to the non-nested dichotomous and continuous endpoints. A sensitivity analysis was also conducted for the two prior sets using parameter priors of varying variability. 

Impact/Purpose

Model averaging has been shown to have superior statistical properties to single model selection, as it is more flexible in accounting for model uncertainty.To demonstrate the utility of the method, Bayesian model averaging (BMA) was applied to non-nested Chloroform datasets using a Laplace-based method.The case study demonstrates that MA is a viable alternative to the traditional single model selection approach that yields practical results even in cases where single model selection did not produce a usable BMD, and it can be utilized by general dose response modelers.

Citation

Blessinger, T., Christine Cai, H. Ju, AND Colin Peterson. Poster: A Case Study on Model Averaging for Risk Assessment. Joint Statistical Meeting 2024, Portland, OR, August 03 - 08, 2024.