Model Variable Augmentation (MVA) for Diagnostic Assessment of Sensitivity Analysis Results. Issue 4 (2nd April 2019)

• Record Type: Journal Article
• Authors: Mai, Juliane; Tolson, Bryan A.
• Is Part Of: Water resources research. Volume 55:Issue 4(2019)
• Page Start: 2631
• Page End: 2651
• Publication Date: 2019-04-02
• Subjects: sensitivity analysis -- Sobol' method -- PAWN method -- model variable ranking -- uncertainty quantification -- quality assessment; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2018WR023382
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Sensitivity analysis (SA) is a critical part in the construction of all models, including environmental and water resources simulation models. For example, SA functions to characterize which model inputs the model outputs are overly sensitive or insensitive to. However, the quality of SA results is rarely assessed. If assessed, bootstrapping of the sensitivity results is used to determine the reliability of the SA output. Bootstrapping, however, is known to be inappropriate with small sample sizes. In contrast, increasing model computational burdens continues to drive researchers to apply existing SA techniques and develop new ones, with smaller and smaller sample sizes. The new Model Variable Augmentation (MVA) approach is therefore introduced here to assess the quality of SA results without performing any additional model runs or requiring bootstrapping. MVA augments the original model input variables with additional variables of known properties. The sensitivities of the augmented model variables are used to draw conclusions on the reliability of the other "original" model parameters' sensitivities. The MVA method is applied to two global SA methods: the variance‐based Sobol' method and the moment‐independent PAWN method. MVA is scrutinized using analytical benchmark functions and then used to quality check sensitivity results of two hydrologic models. Results show the following: (1) MVA is a framework to quality check the implementation of a SA method; (2) for Sobol' and PAWN analyses, MVA‐assisted ranking of input sensitivity measures outperforms the standard ranking procedure without MVA; and (3) MVA provides reasonable estimation of the uncertainty of sensitivity estimates. Key Points: A new approach of Model Variable Augmentation (MVA) is proposed to assess the quality of sensitivity estimates MVA is independent of the sensitivity method used and the model analyzed while not requiring any further model runs MVA is shown to check consistency of SA methods, predicts uncertainty bounds of sensitivity indexes, and yields robust variable rankings