Distributional sensitivity analysis

Abstract

Among the uses for global sensitivity analysis is factor prioritization. A key assumption for this is that a given factor can, through further research, be fixed to some point on its domain. For factors containing epistemic uncertainty, this is an optimistic assumption, which can lead to inappropriate resource allocation. Thus, this research develops an original method, referred to as distributional sensitivity analysis, that considers which factors would on average cause the greatest reduction in output variance, given that the portion of a particular factor's variance that can be reduced is a random variable. A key aspect of the method is that the analysis is performed directly on the samples that were generated during a global sensitivity analysis using acceptance/rejection sampling. In general, if for each factor, N model runs are required for a global sensitivity analysis, then those same N model runs are sufficient for a distributional sensitivity analysis.