Beyond the Environmental Kuznets Curve: A Machine Learning and Econometric Assessment of Growth–CO2 Decoupling Across Six Central American Economies, 1990–2023

Central America faces a defining energy-growth challenge: sustaining economic development while curbing carbon emissions. Existing decoupling and Environmental Kuznets Curve (EKC) studies for Latin America typically pool heterogeneous economies or rely on single-country cases. This study fills that gap with a country-year panel for six Central American economies spanning 1990–2023 for GDP per capita and CO2 emissions and 2000–2021 for the clustering, EKC and Random Forest analyses that additionally require energy intensity and renewable-share data, combining GDP per capita, CO2 emissions per capita, energy intensity and renewable electricity share from World Bank and Global Carbon Project data. We apply k-means clustering, PELT structural break detection, panel fixed-effects EKC estimation with Driscoll–Kraay standard errors, Dumitrescu–Hurlin Granger non-causality tests, Random Forest importance with rolling-origin cross-validation, and the Tapio decoupling index. Clustering separates higher-income, higher-renewable-share regimes (Costa Rica, Panama) from lower-income, higher-energy-intensity regimes (the other four countries). The panel EKC yields an insignificant turning point, and Granger tests find no robust bidirectional causality. Random Forest ranks renewable share and energy intensity above GDP per capita as CO2 predictors, and rolling-origin cross-validation (R-squared = 0.856) falls below naive five-fold validation (0.938). Tapio elasticities classify four countries under weak decoupling and Guatemala and El Salvador under expansive coupling. These findings indicate that Central American decoupling is structural, associated with the electricity generation mix, rather than an automatic byproduct of growth.