Incorporating deep uncertainty into the elementary effects method for robust global sensitivity analysis

Internally-consistent scenarios are increasingly used in social–ecological systems modelling to explore how a complex system might be influenced by deeply uncertain future conditions such as climate, population, and demand and supply of resources and energy. The presence of deep uncertainty requires model diagnostic techniques such as global sensitivity analysis to provide reliable diagnostic insights that are robust to highly uncertain future conditions. We extended the elementary effects method of Morris, which is widely used to screen important model input factors at low computational cost, by incorporating deep uncertainty via the use of scenarios, and evaluated its potential as a robust global sensitivity analysis approach. We applied this robust elementary effects (rEE) method to the highly-parameterised Australian continental Land Use Trade-Offs (LUTO) model—a complex, non-linear model with strong interactions between parameters. We compared rEE sensitivity indicators with robust global sensitivity analysis (RGSA) indicators based on the variance-based eFAST method that imposes relatively high computational demand. We found that the rEE method provided a good approximation of the main effects and was effective in screening the most influential model parameters under deep uncertainty at a greatly reduced computational cost. However, the rEE method was not able to match the accuracy of the eFAST-based method in identifying the most influential parameters in the complex LUTO model based on their total effects. We conclude that the rEE method is well-suited for screening complex models, and possibly for efficient RGSA of models with weak interaction effects, but not for RGSA of complex models. Despite its limitations, rEE is a valuable addition to the robust global sensitivity analysis toolbox, helping to provide insights into model performance under deep uncertainty.