Spatial Heterogeneity of Poverty Determinants in Indonesia A Hierarchical Geographically Weighted Regression Approach

Abstract

This study employs the Hierarchical Geographically Weighted Regression (HGWR) model to analyze poverty determinants in Indonesia, addressing spatial heterogeneity and hierarchical data structures simultaneously. Using data across 34 provinces and 508 regencies/cities, the HGWR model with a Gaussian kernel (bandwidth = 15) substantially outperforms the global Ordinary Least Squares (OLS) regression, increasing the R2 value from 0.502 to 0.754. The Moran's I test on the global model residuals (0.2216, p < 0.001) justifies the urgency of accounting for spatial nonstationarity, while the HGWR post-estimation residuals show that spatial autocorrelation is successfully eliminated (-0.000641, p = 0.416). At the regency/city level, adjusted per capita expenditure and the poverty line significantly reduce the poverty headcount rate, whereas the average years of schooling shows no significant localized effect. At the provincial level, the Human Development Index (HDI) consistently reduces poverty (mean coefficient of -0.8082) but exhibits substantial spatial variation, where the impact is strongest in eastern Indonesia (coefficients < -0.95) and weakest in Java (coefficients > -0.65). Conversely, expected years of schooling exhibits a positive mean coefficient (3.4755), with its positive effects highly concentrated in Java. These findings conclude that poverty reduction strategies in Indonesia must be place-based rather than uniform, prioritizing provincial HDI improvements in eastern Indonesia where the marginal returns of development policy are highest.