Predictive model of determinants of under-five child mortality in Uganda: A multilevel analysis approach

Abstract

Introduction: One of the primary priorities and targets of the Sustainable Development Goals (SDGs) for 2030 is to reduce under-five mortality to 25 deaths per 1000 live births in sub-Saharan countries. However, Uganda continues to experience a high under-five child mortality rate of 40 deaths per 1000 live births, which is almost double the SDG target. Although interventions to address under-five child mortality have largely focused on the individual child level, potential variations at the household and community levels have received limited statistical exploration. Objectives: This study aims to investigate the unaccounted-for variability in under-five child mortality, considering factors at the household and community levels. Additionally, the study aims to develop a predictive model for under-five child mortality in Uganda. Methodology: Using the 2016 Uganda Demographic and Health Survey (UDHS) dataset, a cross-sectional design was employed. The analysis included a sample of 15,522 live births that occurred within the five years prior to the survey. The outcome variable was binary, indicating whether a child died before the age of five (1) or survived (0). Standard logistic regression was used to examine individual-level factors associated with under-five child mortality, while mixed-effect logistic regression was utilized to account for nested data levels. A random forest model was developed to predict under-five child mortality in Uganda. Univariate analysis assessed variable distribution, and variables with a p-value <20% at the bivariate level were considered for the multivariable analysis. Adjusted Odds Ratios (AOR) with a p-value less than 0.05 were deemed significant determinants of under-five child mortality. Model selection was guided by the AIC and BIC criteria, and Intra-Class Correlation (ICC) and Likelihood Ratio Test (LRT) were used to evaluate cluster-level variation. Results: The findings revealed an under-five child mortality rate of 52 deaths per 1000 live births. Significant determinants of under-five child mortality included being a male child (AOR = 0.74), birth order (BO), BO 2-4: AOR=0.72, BO 5-7: AOR = 1.05, BO 8+: AOR = 1.65), mother's education (primary: AOR = 0.74; secondary: AOR = 0.66; higher level: AOR = 0.58), antenatal visitation (1-4 times: AOR = 0.32; 5+ times: AOR = 0.27), and contraceptive use compared to non-use (AOR = 0.63). The Random Forest model exhibited superior predictive performance with accuracy (95%), sensitivity (95%), precision (99%), F-score (98%), and an AUROC of 0.8490 compared to the statistical model. Variation in under-five child mortality was observed at the community level, accounting for 16.9% of the variance (ICC = 0.049). Conclusion: This study highlighted the existence of community-level variation in under-five child mortality. Significant determinants include maternal education, contraceptive use, child sex, and antenatal visitation. The Random Forest model emerged as the optimal predictive tool for under-five child mortality