Addressing Model Uncertainty in Microbial Risk Assessment

• Activity Number: 325
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Risk Analysis
• Author(s): Ravi Varadhan*+
• Companies: Johns Hopkins University
• Keywords: model averaging ; objective priors ; empirical Bayes ; low dose extrapolation

Abstract

Microbial risk assessment involves assessing the risk (probability) of an adverse health outcome, typically illness or infection, when humans are exposed to low doses of microorganisms via food or water supply. Risk assessment is primarily based on fitting dose-response models to binary response (ill/well) data obtained from healthy adult volunteers exposed to high infectious doses of pathogens. Models are used to predict the risk at low doses at which no data are typically available. This extrapolation is highly sensitive to the statistical model used to fit the observed dose-response data. Bayesian model averaging is presented here as a principled approach to addressing this model uncertainty. We evaluate various objective Bayesian methods of computing the posterior model probabilities, such as Bayesian information criteria, Jeffreys priors, and an empirical Bayes approach. The performance of these methods will be evaluated using both simulated data and data from an actual dose-response study. We will also demonstrate an exact (simulation-free) approach to compute the quantiles of the posterior predictive distribution of the quantity of interest.