Sensitivity analysis using Morris: Just screening or an effective ranking method?

Abstract

Abstract Sensitivity analysis (SA) is a fundamental practice for analyzing model behavior under different conditions of application. A number of SA techniques were proposed, ranging from simple screening methods to computationally expensive variance-based ones. In this study, we compared the Morris and E-FAST methods by applying them to three widely used generic crop models largely differing for complexity and for the approaches used to formalize knowledge on crop physiology, i.e., STICS, CropSyst and WOFOST. SA experiments were carried out at sub-model level on rice crops grown under different environmental conditions. Results highlighted the lack of linearity between the total-order sensitivity estimates provided by E-FAST and Morris, although the concordance (TDCC) between the parameter rankings obtained with the two methods was always significant at the 0.05 level for parameters involved with crop growth and for those involved with phenological development for STICS, whereas it was significant at the 0.10 level for the phenology parameters of CropSyst and WOFOST. Given Morris required less than 3% of the model executions needed by E-FAST, our results allow considering Morris as a suitable alternative to more demanding SA methods when ranking parameters or discriminating between influential and non-influential model factors are the SA goals, especially in computationally expensive SA studies.