Income inequality in Southern Africa: Evidence from interpretable machine learning and panel econometric models

This study analyses the determinants of income inequality in Southern Africa by combining interpretable ensemble machine learning with panel econometric models. The study uses an un-balanced panel dataset covering eleven Southern African countries over the period 1991–2019. The analysis is anchored in an adapted Social Determinants of Health framework, which helps organise the selected variables into structural, institutional, and intermediary socio-economic domains. Four ensemble models, namely Categorical Boosting, Extreme Gradient Boosting, Gra-dient Boosting Machine, and Adaptive Boosting, are used to capture non-linear relationships and interaction effects, while random effects and two-way fixed effects models provide econometric benchmarks. The machine learning models show strong internal predictive performance, with an out-of-sample R 2 0.98, although the training and testing design is based on a stratified random split across pooled country-year observations. Therefore, the reported testing results are inter-preted as internal validation within the observed Southern African panel rather than evidence of full generalisation to new countries or future periods. SHapley Additive exPlanations results show that unemployment is the most influential predictor of income inequality across all four ensemble models, followed by gross domestic product per capita and institutional quality. The comparison with panel econometric models shows partial convergence, particularly for gross domestic product per capita, years of schooling, and natural resource rents, while differences in unemployment and democracy suggest that their effects may be non-linear or interaction-dependent. The findings indicate that inequality in Southern Africa is shaped by labour market exclusion, uneven eco-nomic growth, institutional conditions, education-related disparities, and resource dependence. The results support policies that promote inclusive employment, improve the quality and equity of education, strengthen accountable institutions, and ensure transparent management of nat-ural resource revenues. The study contributes by showing that interpretable machine learning can complement, rather than replace, conventional panel econometrics in explaining complex inequality dynamics.