Approximate Bayesian Computation for Chemical Screening Model Selection

Regulatory screening models for chemical exposure estimation often present conflicting goals. One objective is to minimize the probability that model predictions underestimate dose for individuals in the target population. This can be accomplished in a trivial manner by ratcheting up the degree of conservatism in the model structure and parameterization to produce high exposure estimates. However, to be useful in the screening process, a second contrasting objective is to minimize the degree of over-prediction so that high exposure estimates do not forward lower priority chemicals for additional analyses. We employ a likelihood-free approach, approximate Bayesian computation, to select and parameterize terrestrial dermal exposure models for amphibians exposed to pesticides. We compare model predictions to a data set that contains eight studies and 798 individual post-exposure body burdens across 11 amphibian species and 12 pesticides. Our objective function combines a binomial classification approach and a distance approach. The classification approach characterizes false negatives at the individual level by estimating the proportion of body burdens with a model prediction less that the observed level. The distance approach minimizes the overall degree of conservatism by estimating the aggregate amount of over-prediction across all the observations. We present the technical implementation, the advantages of this approach in a regulatory screening context and the results of the model selection exercise for estimating pesticide exposure in terrestrial amphibians.