Spatio-Temporal Inequalities and Predictive Models for Determinants of Undernutrition Among Women and Children in Ethiopia
Women's and children's undernutrition is a major concern in many sub-Saharan African countries including Ethiopia; however, little information exists on its spatial and temporal variation within a single national context. In this study, the spatial and temporal patterns of women and childhood undernutrition from 2000 to 2016 were examined, and the demographic and socioeconomic factors affecting were it. The study also developed predictive models using Machine Learning (ML) algorithms to predict the sociodemographic risks for undernutrition among children in Ethiopia. The regional and temporal trends in women and childhood undernutrition were mapped using data from the 2000, 2005, 2011, and 2016 Ethiopian Demographic and Health Surveys (EDHS). Bayesian multilevel logistic regression models were used to estimate the effects of individual-level and regional effects of undernutrition among women and children. The study also used three ML algorithms (the generalized linear model (GLM), random forest (RF), and eXtreme Gradient Boosting (XGBoost)) models to predict sociodemographic risks of undernutrition. The results show substantial regional variations in women's and children's undernutrition in Ethiopia, which are significantly affected by individual level factors, household level factors and community level characteristics. Undernutrition risks among women are considerably higher for teenage women, no education, never-married, unemployed, residing in rural areas, as well as women from poor households than their counterparts. Additionally, in regions with lower levels of women's education, the risk of undernutrition was higher. On the other hand, the prevalence of stunting and underweight among under-5 children was highest in Amhara and Afar regions, while wasting was highest in Somali. From 2000-2016, all regions of the country except Dire Dawa showed a decrease in stunting and underweight. The greatest reduction was observed in SNNPR. Higher risks of childhood undernutrition was observed among male children, who were delivered at home, having smaller size at birth, and from mothers with lower status (underweight, no education and poor households). New machine learning algorithms such as RFM and XGBoost provide superior predictive accuracy for predicting childhood undernutrition determinants compared with the traditional logistic regression model (GLM). The new ML algorithms also revealed new factors that are associated to childhood undernutrition which are not observed in extant literature. Policy decisions should focus on bridging the regional disparities in the country. Special emphasis are required in regions that have made little progress such as Amhara, Afar and Dire Dawa for stunting and underweight (both women and children) and Somali for wasting. This study reinforces the need to improve child spacing, poverty alleviation programs, maternal nutrition and educational attainment to promote child wellbeing in Ethiopia.