Collaborative Research on Dependence Assessment for Decision and Risk Analysis

专家判断是重要决策问题的关键输入，这些判断的引出和建模在决策分析、风险分析和专家系统等方法中形式化。随着问题越来越复杂（例如，区域和全球环境变化的多系统建模），有许多感兴趣的变量，评估和建模关于变量之间关系的专家知识变得更加重要。到目前为止，条件分布提供了唯一正式、严格和完全通用的方式来表示专家对一组相互关联的随机变量的知识。然而，除非变量的数量非常小，否则这种方法会变得笨拙和不可行的，并且它可能会造成认知问题。另一种方法是用copula表示联合分布，copula是边际分布的函数，并对变量之间的依赖进行编码。本研究的总体目标是开发在实际的专家知识编码情境下的依赖评估程序。在这个为期一年的试点项目中，研究人员将主要关注数学发展和依赖性评估。他们还将开始调查通过边际分布和依赖度量（与边际分布和条件分布相比）获得专家知识所涉及的行为问题。这项初步工作将为今后制定健全的评估程序和可执行的协议以及使用这些评估的分析框架奠定基础。

Expert judgments are key inputs in important decision-making problems, and the elicitation and modeling of such judgments are formalized in methodologies such as decision analysis, risk analysis, and expert systems. As problems grow in complexity (e.g., multi-systems modeling for regional and global environmental change), with many variables of interest, assessing and modeling expert knowledge about relationships among variables becomes more crucial. Up to this point, conditional distributions have provided the only formal, rigorous, and fully general way to represent an expert's knowledge about a set of interrelated random variables. However, this approach becomes unwieldy and infeasible unless the number of variables is very small, and it may pose cognitive problems. An alternative approach involves representing joint distributions in terms of copulas, which are functions of the marginal distributions and encode the dependence among the variables. The overall goal of the proposed research is to develop procedures for dependence assessment in practical expert-knowledge encoding situations. In this one-year pilot project, the researchers will focus primarily on the mathematical development and dependence assessments. They will also begin an investigation into the behavioral issues involved in the elicitation of expert knowledge via marginal distributions and dependence measures (as compared to marginal and conditional distributions). This initial work will provide the basis for the future development of sound assessment procedures and implementable protocols as well as an analytical framework for using those assessments.