High-confidence predictors of childhood anemia in Ghana: a parallel multilevel and machine learning analysis of the 2022 demographic and health survey

Abstract Background Childhood anemia still remains a major public health concern in Ghana, with a national prevalence of 49.0%. Understanding this burden using a combined multilevel and machine learning framework is essential for designing effective interventions. Objective This study used a parallel multilevel and machine learning framework to identify high-confidence predictors of childhood anemia among children (6–59 months) in Ghana using the 2022 Demographic and Health Survey ( n = 3,858 children nested in 235 districts after excluding districts with fewer than two children). Methods We fitted a two-level random-intercept logistic regression (children nested within districts) and three machine learning classifiers (Random Forest, XGBoost, Elastic Net), using district-grouped 70/30 train–test splitting to avoid leakage across the level-2 unit of analysis. SHAP (SHapley Additive exPlanations) values were computed for all three classifiers; the values reported here are drawn primarily from Random Forest, with concordance against XGBoost and Elastic Net assessed as a robustness check (Spearman’s ρ = 0.79–0.87, all p < 0.001). For multi-category predictors, SHAP importance was aggregated across category dummies before ranking. Predictors significant in the multilevel model ( p < 0.05) and ranked among the top 15 by aggregated mean |SHAP| were classified as high-confidence; sensitivity of this classification to the top-15 cutoff is reported. Results The prevalence of childhood anemia was 49.0% (95% CI: 46.4%-51.6%), with significant district-level variation (ICC = 6.2%). Nine high-confidence predictors were identified; child age, mean temperature, child sex, maternal anemia, maternal age, birth order, population density, stunting and fever. Six district-level contextual variables, nightlights, elevation, rainfall, enhanced vegetation index, malaria prevalence, and travel time, were not significant in the random intercept model but ranked among the top 15 predictors in the SHAP analysis. This divergence remains an unresolved finding warranting further investigation rather than being attributed to a specific mechanism. Elastic Net achieved the highest AUC (0.727, 95% CI: 0.702–0.755), followed by Random Forest (0.707, 95% CI: 0.680–0.736) and XGBoost (0.685, 95% CI: 0.659–0.712). Ho Municipal in the Volta region, Ahanta West district in the Western region, Nkwanta South in the Oti region, and Wa West in the Upper West region were identified as the districts with the highest odds of anemia using the caterpillar plot. Conclusions One out of every two children (6–59 months) in Ghana suffers from Anemia. The hybrid approach identified child age maternal anemia, maternal age, stunting and mean temperature as key predictors, while revealing nonlinear contextual effects. High risk districts provide actionable evidence for geographically targeted, climate-resilient, and multisectoral interventions.