BAYESIAN NETWORK MODELS FOR SUPPORTING WATER QUALITY MANAGEMENT

Bayesian network (BN) modeling is a graphical modeling method commonly used to build causal models representing existing knowledge on a specific subject. A BN model starts with a graphical representation of causal links and uses conditional probability for establishing quantitative relationships among the nodes in the causal diagram. Because BN was initially developed to represent the thought process of human experts in performing certain specific tasks such as medical diagnoses, a BN model can explicitly incorporate human knowledge using conditional probability distribution tables. As a result, a BN model is capable of using information from data and expert elicitation, a combination that makes BN ideal for supporting decision-making under uncertainty. Applications of BN in environmental and ecological studies are mostly focused on modeling for supporting environmental management, as the computational complexity limits BN to use only categorical variables. Compared to traditional statistical modeling approach, BN is unique in that it quantifies links between two variables using conditional probability distribution. Graphical presentations of a BN model often resemble other “network-based” models such as the structure equation model (SEM), a topic introduced in a companion presentation. This similarity in presentation illustrates a common feature of network based modeling approaches in their intuitive representation of the underlying causal relationships of interest. This talk will introduce the basic structure of a BN model and its construction, using recently BN models for supporting stream water quality management. These models are based both from existing NAWQA data and expert opinions. The author is interested in exploring the roles of various network-based modeling approaches in both science and management, as well as future development of network based modeling.