On the problem of model validation for predictive exposure assessments

The development and use of models for predicting exposures are increasingly common and are essential for many risk assessments of the United States Environmental Protection Agency (EPA). Exposure assessments conducted by the EPA to assist regulatory or policy decisions are often challenged to demonstrate their ''scientific validity''. Model validation has thus inevitably become a major concern of both EPA officials and the regulated community, sufficiently so that the EPA's Risk Assessment Forum is considering guidance for model validation. The present paper seeks to codify the issues and extensive foregoing discussion of validation with special reference to the development and use of models for predicting the impact of novel chemicals on the environment. Its preparation has been part of the process in formulating a White Paper for the EPA's Risk Assessment Forum. Its subject matter has been drawn from a variety of fields, including ecosystem analysis, surface water quality management, the contamination of groundwaters from high-level nuclear waste, and the control of air quality. The philosophical and conceptual bases of model validation are reviewed, from which it is apparent that validation should be understood as a task of product (or tool) design, for which some form of protocol for quality assurance will ultimately be needed. The commonly used procedures and methods of model validation are also reviewed, including the analysis of uncertainty. Following a survey of past attempts at resolving the issue of model validation, we close by introducing the notion of a model having maximum relevance to the performance of a specific task, such as, for example, a predictive exposure assessment.