Causal discovery from city data, where urban scaling meets information theory

Causal discovery from urban data offers an unprecedented opportunity for research and practice in urban science. Whether they implement traditional regression analyses or more sophisticated tools for time series analysis, existing approaches to causal discovery consider cities to be statistically equivalent – an assumption that is seldom met in real urban systems, where cities dramatically differ in population size, land area, etc. In this study, we embrace the heterogeneity between cities within a novel approach to causal discovery that integrates urban scaling and information theory, towards the discovery of associations between urban processes. Our approach takes as input time series of salient urban variables for causal discovery, along with a set of static features that capture the heterogeneities within the system. Using a Cobb-Douglas function, we extract features-adjusted metropolitan indicators (FsAMIs) that mitigate spurious dependencies among variables due to the underlying variations in spatial features. Through the application of transfer entropy on FsAMIs, we ultimately perform model-free causal inference between urban processes. We validate our approach on synthetic datasets representative of a pair of time-varying urban mock variables and on two real-world open urban datasets related to climate change and infectious diseases. Results demonstrate that our framework outperforms the state of the art across diverse scenarios, minimizing false (positive and negative) inferences. Our methodology offers a powerful tool that advances data-driven urban policy and planning, leveraging open urban datasets to gain robust insight into relationships among urban variables.