Modeling Area Deprivation Index Using Non-Gaussian Fixed Rank Kriging

Socioeconomic disparities shape health outcomes across the U.S., with the Area Deprivation Index (ADI) serving as a key measure of community-level disadvantage. Predicting ADI using Social Determinants of Health (SDOH) like income and health-care access allows for targeted interventions. However, traditional models often ignore spatial patterns, limiting accuracy. This study compares conventional and spatial models to improve ADI prediction. Tract-level ADI and SDOH data were analyzed using six predictors selected via stepwise AIC: median income and distances to five healthcare facility types. Five models were tested: Linear Regression, GAM, GAMLSS, Gaussian FRK, and Poisson FRK. FRK models use low-rank basis functions to efficiently capture complex spatial dependencies in large datasets. Model performance was evaluated using R², RMSE, MAE, AIC, and cross-validation. Linear regression performed worst; GAM and GAMLSS improved results by modeling non-linearity. Gaussian FRK enhanced spatial prediction but oversmoothed local detail. Poisson FRK delivered the best accuracy, capturing both regional and local deprivation patterns. Based on our findings spatial models are essential for analyzing geographically structured data. They capture spatial dependence and distributional complexity, improving prediction and interpretation, unlike traditional methods. Their use supports more accurate location-specific insights in public health and beyond.