A probability network model for TMDL development in the Neuse River watershed, NC

Abstract

For the purposes of providing the scientific basis for Total Maximum Daily Load (TMDL) decisions, probability network models represent a potential improvement over the current reliance on deterministic simulation models. Also called Bayesian networks, or “Bayes nets,” these models make appropriate use of existing data, facilitate the extension of predictions to publiclymeaningful endpoints, and allow for a thorough analysis of uncertainty. Additionally, they are easily communicated to stakeholders and can be updated quickly as new information is obtained. We illustrate the probability network approach using an application to the management of eutrophication in the Neuse River Estuary, North Carolina. Management endpoints were linked to external nutrient loading using a set of assertions concerning the probabilistic relationships among the variables. These relationships, expressed as conditional distributions, were then quantified using either 1) mathematical representations of large-scale processes, 2) statistical associations based on historical data, or 3) probabilistic relationships elicited from scientific experts. Probabilistic predictions of endpoints are based on the entire linked set of conditional probabilities. Not only does this network structure provide a more integrated approach to uncertainty analysis, but it also allows for easy updating of predictions and inference when new scenarios are tested or additional observations are made, thus facilitating the process of adaptive management. The probabilistic predictions generated by the model are also consistent with the risk assessment paradigm and allow management decisions to be made based on expected value theory. In addition, because the assessment endpoints were chosen to be of vital interest to stakeholders and decisionmakers, they can be easily conceived in terms of utility for use in a formal decision analysis.