Comparison of two sets of Monte Carlo estimators of Sobol' indices. (October 2021)

• Record Type: Journal Article
• Authors: Azzini, Ivano; Mara, Thierry A.; Rosati, Rossana
• Is Part Of: Environmental modelling & software. Volume 144(2021)
• Publication Date: 2021-10
• Subjects: Global sensitivity analysis -- Variance-based sensitivity indices -- First-order Sobol' index -- Total-order Sobol' index -- Asymptotic normality -- Radiative forcing model; Environmental monitoring -- Computer programs -- Periodicals; Ecology -- Computer simulation -- Periodicals; Digital computer simulation -- Periodicals; Computer software -- Periodicals; Environmental Monitoring -- Periodicals; Computer Simulation -- Periodicals; Environnement -- Surveillance -- Logiciels -- Périodiques; Écologie -- Simulation, Méthodes de -- Périodiques; Simulation par ordinateur -- Périodiques; Logiciels -- Périodiques; Computer software; Digital computer simulation; Ecology -- Computer simulation; Environmental monitoring -- Computer programs; Periodicals; Electronic journals; 363.70015118
• DOI: 10.1016/j.envsoft.2021.105167
• Languages: English
• ISSNs: 1364-8152

Abstract

Abstract: This study compares the performances of two sampling-based strategies for the simultaneous estimation of the first- and total-order variance-based sensitivity indices (a.k.a. Sobol' indices). The first strategy corresponds to the current approach employed by practitioners and recommended in the literature. The second one was only recently introduced by the first and last authors of the present article. Both strategies rely on different estimators of first- and total-order Sobol' indices. The asymptotic normal variances of the two sets of estimators are established and their accuracies are compared theoretically and numerically. The results show that the new strategy outperforms the current one. The global sensitivity analysis of the radiative forcing model of sulfur aerosols is performed with the new strategy. The results confirm that in this model interactions are important and only one input variable is irrelevant. Highlights: We consider two sets of Monte Carlo estimators of first- and total-order Sobol' indices. We derive their variances under the asymptotic normality assumption. The new estimators provide first-order sensitivity indices always smaller or equal to the total-order sensitivity indices. We find that the new set of estimators outperforms the current one. Numerical exercises confirm our theoretical findings.