Decomposing Variance Decomposition for Stochastic Models: Application to a Proof-Of-Concept Human Migration Agent-Based Model

Agent-based models (ABMs) are promising tools for improving our understanding of complex natural-human systems and supporting decision-making processes. ABM bottom-up approach is increasingly employed to recreate emergent behaviors that mimic real complex system dynamics. However, often the knowledge and data available for building and testing the ABM and its parts are scarce. Due to ABM output complexity, exhaustive analysis methods are required to increase ABM transparency and ensure that the ABM behavior mimics the real system. Global sensitivity analysis (GSA) is one of the most used model analysis methods, as it identifies the most important model inputs and their interactions and can be used to explore model behaviors that occur in certain regions of the parameter space. However, due to ABM’s stochastic nature, GSA application to ABMs can result in misleading interpretations of the ABM input importance. Here, we review 3 alternative GSA approaches identified in the literature for ABMs and other stochastic models. Using two study cases, a benchmark non-linear function and a proof-of-concept migration ABM, we illustrate the differences in input importance and the shortcomings of current approaches and propose a new GSA approach for the evaluation of stochastic models which separates inputs importance for deterministic and stochastic uncertainties. The former is related to changes in the expected value of model realizations and the latter to changes in the variance of model realizations. Our analysis of the proof-of-concept migration ABM finds that how factors are weighed is more important than the values of the inputs and identifies what inputs are more important for the deterministic and stochastic uncertainties. The analysis also identifies outputs for which the deterministic uncertainty is small, being almost random. This information allows the modeler to evaluate the optimal degree of model complexity and choose among alternative model structures.