Robust Subsampling ANOVA Methods for Sensitivity Analysis of Water Resource and Environmental Models

Sensitivity analysis is an important component for modelling water resource and environmental processes. Analysis of Variance (ANOVA), has been widely used for global sensitivity analysis for various models. However, the applicability of ANOVA is restricted by this biased variance estimator. To address this issue, the subsampling based ANOVA method are developed in this study, in which multiple subsampling(single-, multiple- and full-subsampling) techniques are proposed to diminish the effect of the biased variance estimator of ANOVA. Two case studies including one simplified regression model and one hydrological model are used to illustrate the applicability of the proposed approaches. Results indicate that: (1) the subsampling procedures effectively diminish the biases resulting from traditional ANOVA method; (2) among the proposed subsampling approaches, the full-subsampling ANOVA has the most robust performance; (3) compared with Sobol's method, the subsampling ANOVA methods can significantly reduce the calculation requirements while achieve similar sensitivity characterization for model parameters. This study serves as a first basis for the application of subsampling ANOVA methods to sensitivity analysis for water resource and environmental models.