A comparative study of GWR, HLM, and HGWR for modeling childhood stunting in Indonesia

Abstract

Childhood stunting is an important biological and public health problem in Indonesia, with far-reaching consequences for cognitive development, disease risk and economic productivity. The prevalence of stunting shows considerable spatial and contextual differences, which are influenced by local socioeconomic and environmental conditions. We compare three regression models, Geographically Weighted Regression (GWR), Hierarchical Linear Model (HLM), and Hierarchical Geographically Weighted Regression (HGWR), to represent this complexity with the observations of 514 districts nested in 34 provinces in Indonesia. The outcome variable is the prevalence of stunting (%), with district-level predictors of poverty rate, education, sanitation, immunisation, and access to clean water and province-level predictors of HDI, health budget per capita, and prevalence of malnutrition. All Models capture Different dimensions of structure: GWR for spatial non-stationarity, HLM for hierarchical nesting, and HGWR for both. Model assessment was based on AIC, RMSE and adj. R² values. The best model for overall goodness of fit was HGWR (AIC = 3196.563; RMSE = 5.159; Adjusted R² = 65.73%), in which it showed high performance to model spatially-structured and hierarchically-nested health themes. This research highlights the value of spatial-hierarchical models and encourages support for the soundness of HGWR as an effective paradigm to support targeted public health interventions in geospatial epidemiology.