Does Railway Accessibility Cause Peri-Urban Expansion? A Spatial Graph Double Machine Learning Framework for Station Area Planning in Tripura, Northeast India

Railway investment is widely expected to drive peri-urban expansion around small and medium towns, yet station siting is endogenous to existing development potential, so conventional land change models recover association rather than cause. This study develops Spatial Graph Double Machine Learning (SG-DML), which embeds graph neural networks within a Neyman-orthogonal estimator and applies it to the 264 km Churaibari–Sabroom railway corridor and its 27 stations in Tripura, Northeast India. Land parcels form the nodes of a graph whose edges combine boundary adjacency with road–rail network connectivity; Sentinel-2 land cover maps (2015–2023) supply the outcome, and a tri-zonal model of 1, 2 and 3 km station neighbourhoods structures the covariates. A graph attention network estimates the propensity and a graph convolutional network the outcome, under buffered spectral cross-fitting, which prevents information leakage between folds. On synthetic benchmarks with injected spatial confounding, SG-DML recovered an average treatment effect of 0.148 against a true value of 0.150; outperformed a spatial Durbin model, propensity score matching, and standard double machine learning; and passed placebo tests. Stress tests that deliberately violated the identifying assumptions showed that SG-DML estimates can remain biased under endogenous station siting and that nominal cluster-robust intervals under-covered the known effect in the smaller synthetic analogue. The corridor-level average treatment effect and its confidence interval cannot be recovered from the supplied manuscript because parcel-level estimation outputs are unavailable. The network distance dose–response and an empirically supported spatial reach also remain unquantified. Descriptive station-level patterns motivate differentiated station area planning but do not by themselves establish the magnitude of a railway-induced causal effect.