A Comparison for Non-Specialists of Workflow Steps and Similarity of Factor Rankings for Several Global Sensitivity Analysis Methods

TL;DR

This paper compares various global sensitivity analysis methods to help non-experts understand their workflows, interpret outputs, and assess similarities in factor rankings across different models.

Contribution

It provides a comparative analysis of multiple GSA methods, highlighting their workflows, interpretability, and similarity in factor rankings for non-specialists.

Findings

High similarity in factor rankings across methods (Kendall's W)

Sobol' indices are easy to interpret and informative

Regression trees offer additional insights into interactions

Abstract

Global sensitivity analysis (GSA) is a recommended step in the use of computer simulation models. GSA quantifies the relative importance of model inputs on outputs (Factor Ranking), identifies inputs that could be fixed, thus simplifying model calibration (Factor Fixing), and pinpointing areas for future data collection (Factor Prioritization). Given the wide variety of GSA methods, choosing between methods can be challenging for non-GSA experts. Issues include workflow steps and complexity, interpretation of GSA outputs, and the degree of similarity between methods in Factor Ranking. We conducted a study of both widely and less commonly used GSA methods applied to three simulators of differing complexity. All methods share common issues around implementation with specification of parameter ranges particularly critical. Similarities in Factor Rankings were generally high based on Kendall's W. Sobol' first order and total sensitivity indices were easy to interpret and informative with regression trees providing additional insight into interactions.