Inferring spatial causations from spatial associations considering spatial lags: a case study of geohazards in Yunnan, China

The inference of causation is essential for interpretable scientific research and evidence-based policymaking. Existing causal methods – including those based on intervention, forecasting, and graph models – have achieved considerable success in non-spatial domains such as medicine, economics, and sociology. However, significant challenges arise when these methods are applied to geographic phenomena, which constitute complex spatiotemporal processes integrating geo-events and geo-environments. Current causal methods, which predominantly infer relationships from temporal information, often fail to fully utilize spatial information, such as the locations and distributions of geo-events. Furthermore, they typically overlook spatial lags between events and how these lags vary across different geo-environments. These limitations result in potentially incorrect causal conclusions. To address these issues, we propose a novel causal framework for geographic phenomena. This framework develops an Adaptive Gaussian Field (AGF) model to capture spatial lags. It first extracts spatial association patterns from geo-event locations and distributions, then infers causal relationships from these associations. The integration of spatial lags into both association mining and causal inference is implemented through two novel methods: a spatial-lagged co-distribution mining method and a spatial-lagged Peter-Clark (spatial-PC) method. This proposed framework offers a robust approach for uncovering the underlying mechanisms of geohazards in Yunnan, China.