Capturing Interactions between Network Data Drivers with Global Sensitivity Analysis: Application to Refugee Flows

Abstract:

The amount of observed network data increases every year. These large datasets promise to help address important problems if the questions are correctly formulated. As network data availability increases, we have new opportunities to address important problems. This requires employing novel methods to achieve a better understanding of network systems. This new knowledge would also help modelers to build better predictive tools. Global Sensitivity Analysis (GSA) is a model evaluation technique that apportions the variance of the output onto to individual inputs (direct effects) or their combination (interactions). Could GSA inform the driver dynamics of observed data in a real network despite being a model analysis technique? For higher-dimensional output variance decomposition, GSA typically employs multivariate structured sampling techniques of the model input space. The method is assumption-free in that it does not assume additivity, linearity or monotonicity of the system outputs. To adapt GSA to the analysis of large observed network data we propose coupling with the network analysis technique Quadratic Assignment Procedure (QAP), which serves to inform the robustness of the GSA results. As a study case we analyze refugee flows between countries at a global scale. We compare the results of our novel GSA-QAP to those obtained by the conventional network analysis method Multiple Regression Quadratic Assignment Procedure (MR-QAP), explaining the agreements and disagreements and advantages of these two approaches. The work highlights the importance of some of the local and global factors driving refuge flows network data.