Role of model parameterization in risk-based decision support: An empirical exploration

The degree with which to parameterize a computer model that is to be used for risk-based resource management decision support has been a topic of much discussion in the environmental modeling industry, and remains a difficult choice facing practitioners. High-dimensional parameterization schemes allow for a more robust expression of model input uncertainty over traditional lower-dimensional schemes, but often incur a higher computational burden and require greater understanding of inverse problem theory to implement effectively. However, a number of significant questions remain, such as: What level of parameterization is needed to adequately express uncertainty for a given decision-relevant simulated output?, and To what extent can a simplified parameterization be adopted while maintaining the ability of the model to serve as a decision-support tool?. This study addresses these questions, among others, by using empirical paired complex-simple model analyses to investigate the consequences of reduced parameterization on decision-relevant simulated outputs in terms of bias incursion and underestimation of uncertainty. A Bayesian decision analysis approach is adopted to facilitate evaluation of parameterization reduction outcomes, not only in terms of the prior and posterior probability density functions of decision-relevant simulated outputs, but also in terms of the management decisions that would be made on their basis. Two integrated surface water/groundwater model case study examples are presented; the first is a complex synthetic model used to forecast groundwater abstraction-induced changes in ecologically-sensitive streamflow characteristics, and the second is a real-world regional-scale model (Hauraki Plains, New Zealand) used to simulate nitrate-loading impacts on water quality. It is shown empirically that, for some decision-relevant simulated outputs, even relatively high-dimensional parameterization schemes ( > 2,000 adjustable parameters) display significant bias in simulated outputs as a result of improper parameter compensation induced through history matching, relative to complex parameterization schemes ( > 100,000 adjustable parameters)-ultimately leading to incorrect decisions and resource management action. For other decision-relevant simulated outputs, however, reduced parameterization schemes may be appropriate for resource management decision making, especially when considering a prior uncertainty stance (i.e., without undertaking history matching) and when considering differences between simulated outputs that do not depend on local-scale heterogeneity.