Collaborative Research: Combining Expert Judgments for Environmental Risk Analysis.

点击翻译按钮获取中文摘要

This collaborative project with Clemen (0084383) and Hammitt (0084372) will develop and demonstrate improved methods for using expert judgment in environmental risk analysis. Risk-analytic models rely heavily on parameter estimates obtained using either informal or formal expert judgments. Such methods can be sensitive to two problems: biases in experts coding their knowledge into probability distributions (such as overconfidence bias), and the typically uncertain degree of dependence in judgments between experts (e.g., each expert's judgment combines his reading of a common scientific literature and his own experience and interpretations). The project will analyze the properties of two state-of-the-art methods for combining expert judgments: the "classical" method developed by Roger Cooke and colleagues at Delft University of Technology and the "copula" method developed by Robert Clemen and colleagues at Duke University. It will compare the methods mathematically and by evaluating their performance on synthetic and actual expert judgment data sets. In addition, the project will develop improved methods for combining distributions based on fully Bayesian methods for incorporating overconfidence and dependence among experts, and compare them with the two existing methods.The project will lead to better methods and an improved understanding of alternative methods for combining expert judgment in environmental risk analysis. Many parameters in a risk-analysis model cannot be measured directly, for physical or ethical reasons (e.g., one typically cannot trace pollutants far from their source nor conduct toxicity studies on humans). As a result, parameter estimates are often based on expert judgment. In most cases, the expert judgment is applied informally, as when the model builders use their own "best guess" estimates of parameter values. In some cases (such as risks associated with nuclear power), judgments (in the form of probability distributions) are elicited from a panel of experts. However, there is at present no standard method for combining judgments from multiple experts, and limited understanding of the properties of alternative methods. This research will lead to better understanding of alternative methods and, ultimately, to better use of expert judgment.