Accounting for model error in risk assessments: Alternatives to adopting a bias towards conservative risk estimates in decision models

Abstract

Two recently developed approaches to quantification of model (conceptual) error in a single groundwater model, a per-datum calibration methodology and a Bayesian model error analysis, were applied to a problem of 90Sr migration to water wells at Chernobyl, Ukraine. The intent of this composition is to demonstrate their utility to accounting for the uncertainty due to model error in estimating risks (or costs) in decision models. Bayesian model error analysis resulted in a more conservative estimate of the probability of the Pripyat Town well field contamination than did the per-datum calibration approach. This difference in risk estimates is a result of the conceptual differences between the two methods. Per-datum calibration relies primarily on information on model error contained in the measurements of the dependent variables to quantify its effect on model predictions. The Bayesian model error analysis assigns equal importance to prior information on the parameters and measurements of the dependent variable, thus allowing the incorporation of a more informative description of parameter distributions, as well as subjective judgement into a risk analysis. The suitability of either of the two methods, when applied to a specific problem, may be determined based on the nature and quantity of available data.