Association between multidimensional poverty and urban spatial network design: Comparison between theory-driven and data-driven lenses

Abstract

Poverty is increasingly identified as an urban phenomenon despite the promise that urban areas hold as the centres of economic and social progress. There is a need for knowledge on how the spatial network design, a signature of urban areas, is associated with poverty. This paper addresses this need by combining conventional theory-driven and emerging data-driven methods. We computed a multidimensional poverty index -MPI using geocoded household survey data in Kenya, which was treated as the dependent variable (target). The spatial Design Network Analysis (DNA) plug-in in ArcGIS Pro was used to quantify metrics of the spatial network design from a road network dataset of the study area, which was treated as the independent variables (features). We used the capability approach to provide a theoretical basis linking the social and physical network attributes. We then applied logistic regression and a machine learning algorithm, XGBoost, to analyse the network predictors of multidimensional poverty while controlling for confounders. The results of the logistic regression suggested that network density had the largest magnitude of margins (−1.004), which is significant at (p < 0.01) and negatively associated with multidimensional poverty. In contrast, results from the XGBoost algorithm suggested that network efficiency was the most important feature of the road network, with an impact of 16 percentage points. Severance and betweenness were among the top five important features of the network in both logistic regression and XGBoost. The situation of a household in either a formal or informal settlement was the most important confounder in both models. The results suggest that theory-driven logistic regression outperforms the machine learning algorithm based on our data and method. The logistic regression had an AUC of 0.794 compared to 0.692 in XGBoost. Our paper contributes to the knowledge about the association between spatial network design and multidimensional poverty, which helps improve our hypothesis and informs our theory. In addition, the results reveal the spatial design features that planners and policymakers should pay attention to in urban areas. We propose further research considering spatial heterogeneity and spatial dependence in the analysis.