Modeling with Robust Kernel Nonparametric Regression on Childhood Stunting in Kalimantan

Abstract

This study models stunting prevalence across 56 regencies/cities in Kalimantan using robust kernel nonparametric regression. This approach addresses the nonlinear relationship between stunting and four predictors: access to improved sanitation, low birth weight, population density, and poverty rate. An examination of influential observations using DFFITS identified five regencies as outliers; thus, the robust MM-estimator approach was applied to mitigate the influence of these extreme observations on the estimation results. Optimal bandwidth selection was performed using the Cross-Validation (CV) method across several kernel functions, namely Epanechnikov, Gaussian, and Uniform. The results demonstrated that the Uniform kernel function yielded the smallest CV value with a bandwidth combination of h1=0.6, h2=0.2, h3=0.6, and h4=0.2. The Robust Uniform Kernel model delivered the best performance, with an MSE of 2.0556, RMSE of 1.4337, MAE of 0.6581, and  of 0.9510. This study indicates that robust MM-estimator kernel nonparametric regression can produce stunting prevalence estimates that are more accurate, flexible, and stable in the presence of outliers.