This study examines the spatially and temporally varying impact of public capital on regional economic growth in Morocco over the period 2011–2023. By specifying a flexible translog production function and estimating it using Geographically and Temporally Weighted Regression (GTWR), we account for both spatial non-stationarity and temporal dynamics across the twelve Moroccan regions. Our empirical findings uncover pronounced heterogeneity in the productivity of public investment. Coastal and economically diversified regions such as Casablanca-Settat, Rabat-Salé-Kénitra, and Tanger-Tétouan-Al Hoceïma demonstrate high and increasing output elasticities of public capital, reflecting robust complementarities with private capital accumulation and skilled labor. Conversely, interior and peripheral regions—including Drâa-Tafilalet and Beni Mellal-Khénifra—exhibit weak or insignificant effects, highlighting the limitations of a uniform infrastructure investment policy in structurally constrained territories. These results provide strong evidence of emerging spatial polarization and reinforce the need for territorially differentiated investment strategies. Our approach also demonstrates the methodological relevance of GTWR for evaluating public policy effects under spatial heterogeneity. The study contributes to the literature on regional growth, spatial econometrics, and infrastructure policy in developing economies.

Abstract Statistical procedures can be significantly affected by outliers, and spatial econometric models are no exception. While Maximum Likelihood (ML) estimators are common, computationally efficient Generalized Method of Moments (GM) estimators–such as Spatial Two-Stage Least Squares (STSLS) for lag models, GMerrorsar for error models, and Generalized Spatial Two-Stage Least Squares (GSTSLS) for combined (SARAR) models–are increasingly preferred for large datasets. However, systematic evidence on the comparative robustness of these diverse estimation techniques (ML versus GM) to local outliers remains limited, particularly concerning their performance when models are correctly specified versus potentially misspecified. This paper aims to fill this gap by employing the Local Influence Function (LIF), an extension of the classical influence function to spatial datasets. First, we conduct Monte Carlo experiments to systematically evaluate the robustness of Maximum Likelihood (ML) versus Generalized Method of Moments (GM) estimators for Spatial Lag (SLM), Spatial Error (SEM), and Spatial Autoregressive with Autoregressive Disturbances (SARAR) models. For each model type, estimators are assessed using data generated from their respective true Data Generating Processes (DGPs) under local perturbations of the dependent variable. These simulations indicate that, under correct model specification, ML estimators generally demonstrate superior robustness for key parameters ( $$\beta$$ , $$\rho$$ , $$\lambda$$ ) compared to their GM counterparts. Second, we demonstrate the practical utility of the LIF methodology using a large-scale dataset of approximately 7,900 Italian municipalities to analyze the relationship between average equivalent income and unemployment rates. Our application reveals that key parameters of the GM estimator for the most comprehensive model (SARAR/SAC) can exhibit considerable local sensitivity to data perturbations, and the LIF effectively identifies specific municipalities exerting disproportionate influence, allowing for the mapping of these spatial patterns of sensitivity.