Asian Journal of Probability and Statistics

Aim: To develop and evaluate a hierarchical multilevel logistic regression model for predicting maternal mortality in Kenya.
Methodology: Secondary data from the Kenya Demographic and Health Survey (KDHS) (2022), Kenya Population and Housing Census (KPHC) (2019), and national health registries for 2018–2023 were combined, yielding a sample of 32,156 women aged 15–49 years. Random Forest was used for variable selection, followed by two-level logistic regression modelling. Model performance was assessed using the Akaike information criterion (AIC), the Bayesian information criterion (BIC), likelihood-ratio tests (LRT), and repeated 10-fold cross-validation.
Results: The null model indicated significant between-county variation (ICC = 8.2%), supporting the use of a multilevel approach. The final model showed good fit (AIC = 3,498.1; marginal R^{2}=16.6\%; conditional R^{2}=24.7\%). Higher education reduced the odds of maternal mortality by up to 75%, while each additional antenatal care (ANC) visit reduced the odds by 15%. In contrast, high parity (OR = 1.26), greater distance to health facilities (OR = 1.02 per km), and county poverty rate (OR = 1.03) significantly increased mortality risk. The model demonstrated acceptable predictive performance (AUC = 0.78), with specificity of 88.9%. The confusion matrix is based on pooled out-of-fold predictions across all five repetitions of 10-fold cross-validation. In the internal validation set, the model showed limited sensitivity for maternal death (36.4%; 4/11) and very low PPV (1.1%), indicating limited case-detection performance at the selected threshold in this rare-outcome dataset.
Application: The preliminarily developed and internally assessed multilevel predictive model provides a framework for population-level risk stratification. It may inform county health management teams about the equitable allocation of maternal healthcare resources, such as targeted deployment of community health promoters and emergency transport subsidies to high-poverty counties.
Conclusion: These findings highlight the importance of accounting for both individual- and county-level factors in maternal mortality research and support targeted interventions focused on improving female education, ANC, and equitable resource allocation in high-poverty counties. The consistent reductions in information criteria, coupled with significant LRTs at each modeling stage, substantiate that maternal mortality in Kenya is best understood through a multilevel analytical framework. The model’s acceptable discriminative performance (AUC = 0.78) and satisfactory calibration indicate that it distinguishes between high-risk and low-risk profiles while maintaining probabilistic accuracy across the predicted risk spectrum. However, the model requires external validation before clinical or operational use.