Convergence and uncertainty analyses in Monte-Carlo based sensitivity analysis

Abstract

Sensitivity analysis plays an important role in model development, calibration, uncertainty analysis, scenario analysis, and, hence, decision making. With the availability of different sensitivity analysis techniques, selecting an appropriate technique, monitoring the convergence and estimating the uncertainty of the sensitivity indices are very crucial for environmental modelling, especially for distributed models due to their high non-linearity, non-monotonicity, highly correlated parameters, and intensive computational requirements. It would be useful to identify whether some techniques outperform others with respect to computational requirements, reliability, and other criteria. This paper proposes two methods to monitor the convergence and estimate the uncertainty of sensitivity analysis techniques. One is based on the central limit theorem and the other on the bootstrap technique. These two methods are implemented to assess five different sensitivity analysis techniques applied to an environmental model. These techniques are: the Sobol' method, the Morris method, Linear Regression (LR), Regionalized Sensitivity Analysis (RSA), and non-parametric smoothing. The results show that: (i) the Sobol' method is very robust in quantifying sensitivities and ranking parameters despite a large number of model evaluations; (ii) the Morris method is efficient to rank out unimportant parameters at a medium cost; (iii) the non-parametric smoothing is reliable and robust in quantifying the main effects and low-order interactions while requiring a small number of model evaluations; finally (iv) the other two techniques, that is, LR and RSA, should be used with care.