Advancing hierarchical model and variable Selection methods: evaluating performance, Interpretability and implications

Abstract

Healthcare utilization continues to be a relevant issue for the public health sector in developing situations. However, conventional statistical frameworks tend to understate the contextual and hierarchical nature of the health data. Prior studies have primarily used classical regression approaches that treat observations as autonomous entities, neglecting the relevance of group- or area-context. This disregard for clustering amplifiers limits the health utilization research scope, precision, and interpretability. This research employs a Bayesian hierarchical modeling approach to contextual, socioeconomic, and maternal variables concerning the use of maternal health services in Kenya. The model was applied to hierarchical data using the Least Absolute Shrinkage and Selection Operator (LASSO) technique for variable selection and the Hierarchical Bayesian Information Criterion (HBIC) for model selection. The modified model encompasses fixed effects (population-level predictors), random effects (county heterogeneity), and contextual effects. The HBIC results indicate that Age and Religion were the most influencial predictors. The results highlight the importance of maternal education, health insurance, marital status, household income, religion, and other socioeconomic variables in explaining the inequitable and, at times, regionally differentiated utilization of health services. Incorporating prior information and treating parameter uncertainty are ways the Bayesian approach generates more accurate posterior estimates. This research illustrates the value of Bayesian hierarchical modelling for policy-related inference and small-area estimation in closing the methodological void in public health analytics. The results provide the foundation for more equitable, evidence-based maternal health intervention strategies that account for individual and contextual differences.