Digital Spillovers and Regional Growth in the European Union: Spatial Panel Econometrics with Machine-Learning Clustering and Validation for 234 NUTS-2 Regions, 2017–2024

Does digital adoption raise regional prosperity in the region where it occurs, or does it work through neighbouring regions? We assemble a georeferenced panel of 234 EU NUTS-2 regions observed from 2017 to 2024, linking Eurostat data on GDP per capita in purchasing power standards, household broadband and internet use, employment and social exclusion to NUTS 2024 boundaries. Global Moran's I is positive and significant for every variable and year (0.39–0.78), and LISA maps reveal persistent high-income and low-income clusters. Non-spatial estimators give a positive digital coefficient: 0.21% higher GDP per capita per percentage point of adoption under two-way fixed effects, and 1.83% when lagged adoption instruments a noisy, survey-based regressor. However, fixed-effects residuals remain spatially autocorrelated, and a Spatial Durbin Model (ρ = 0.70) shows that the own-region (direct) effect is not significant, while the spillover (indirect) effect is positive (0.20% per point, p = 0.039) and robust across three spatial weight matrices. Of six clustering algorithms compared on eleven validity indices, Ward hierarchical clustering performs best and identifies seven spatially contiguous regional profiles. Of six machine-learning regressors, a random forest validated by region-grouped cross-validation performs best (out-of-sample R² = 0.57). Its SHAP decomposition reproduces the econometric signs: own adoption dominates in levels, while neighbours' adoption and employment dominate within regions. Digital policy in the EU should therefore be targeted at the level of functional macro-regions, not individual regions.