A Structured Spatio-Temporal Neural Architecture for Panel Microdata with Explicit Space–Time Lag Decomposition

Spatial panel econometrics has usually included contemporaneous spatial lags, temporal persistence, and space-time diffusion terms within parametric models. These models often assume that the observation unit matches the spatial interaction matrix, requiring outcomes to be defined or aggregated at the spatial level. This can be limiting in micro-survey panels, where data are at the individual or household level, but spatial interaction is defined over larger geographic areas. Meanwhile, spatio-temporal neural networks in machine learning model spatial and temporal dependence on graph data but often mix space and time without keeping the clear separation found in spatial econometrics. This study introduces a structured multi-branch neural network for panel microdata that mirrors the decomposition of dynamic spatial panel models while keeping micro-level observations. Contemporaneous spatial lags, temporal lags, and spatially lagged temporal components are designed as separate feature blocks and embedded as distinct channels in a nonlinear model. The main contribution is enforcing the econometric lag structure within a flexible predictive framework.