Code and data for HDSTAR-GNN: A heterogeneous dynamic spatiotemporal graph neural network model for quantifying spatially cascading impacts of extreme precipitation on economic growth

Abstract

Dynamic spatiotemporal autoregressive models explicitly represent spatial dependency and temporal dynamics underlying an evolving spatial system, providing a natural framework for investigating how climate shocks propagate, persist, and amplify. Yet their application to fine-resolution climate-impact assessment is critically constrained by intensive computational demand and an inability to capture heterogeneous spatial cascades. Here we address these challenges by reformulating these models as a specialized graph convolution neural network architecture. By establishing their mathematical equivalence, we develop a flexible heterogeneous dynamic spatiotemporal autoregressive graph neural network (HDSTAR-GNN) model that allows spatial and temporal autocorrelation parameters to vary locally. The model reduces computational complexity to approximately O(n) while retaining the transparent parameter interpretation of statistical models. Monte Carlo simulations demonstrate that HDSTAR-GNN can accurately recover true parameters under both homogeneous and heterogeneous spatiotemporal processes. Applied to global 0.25° grids, HDSTAR-GNN reveals pronounced geographical heterogeneity in the spatial cascading effects of consecutive extreme precipitation (CEP) on economic growth. On average, one additional CEP day is associated with a total reduction of 0.17% of mean growth rate, with indirect effects accounting for approximately 45% of the total impact. The proposed framework provides a basis for combining statistical interpretability with the computational flexibility of deep learning.Python and R codes for model estimation and Monte Carlo simulation, data used in the empirical analysis, and for implementing the Monte Carlo simulation experiments are publicly available at here.model.py Core model module containing the neural-network architecture, spatial autoregressive components, loss functions, and utility functions shared by the simulation and empirical analysis.simulation.py Monte Carlo simulation script for evaluating the estimation performance of the proposed model under controlled data-generating processes.case_study.py Empirical analysis script implementing the proposed model using the gridded panel dataset, including data preparation, model estimation, and result extraction.panel_mod.parquet Processed panel dataset used in the empirical case study, containing the variables and spatial-temporal identifiers required for model estimation.Simu_SDPD.R R script implementing the conventional spatial dynamic panel data estimator used as a benchmark in the Monte Carlo simulations.Ar6_Region.zip Spatial data representing IPCC AR6 reference regions, used for partitioning the global grid in simulation experiments.ERA5_grid.nc ERA5 gridded data, spatially interpolated to construct global 2° gridded data for simulation.gadm_global_l1.zip Global first-level administrative boundary data from GADM, used for spatial visualization and regional analysis.