Multilevel Modelling of the Factors Associated with Stunting among Under-five Children in Tanzania: A Bayesian Approach

Abstract

Introduction: In Tanzania, prevalence exceeds 30% and is unevenly distributed across geographical areas. The observed distribution may reflect unobserved contextual factors; however, little literature applies Bayesian multilevel generalised linear mixed (B-MGLM) models, which can account for them. This study aimed to model the determinants of stunting, controlling for contextual factors at both the region and enumeration area levels.

Methods: This was a secondary data analysis using four recent Tanzania Demographic and Health Surveys (TDHS) datasets. We studied 7,492, 6,806, 8,929, and 4,797 under-five children in TDHS 2004/5, 2010, 2015/16, and 2022, respectively. We used B-MGLM models to evaluate the determinants of stunting. We used the deviance information criterion to compare the models.

Results: About 41.4%, 34.1%, and 28.96% of under-five children were stunted in 2010, 2015/16, and 2022, respectively. Controlling for other factors in the model, a unit increase in MPI increased the odds of stunting by 10% [AOR = 1.1; 95% CI: 1.1, 1.2], while a unit increase in WAMI reduced the odds of stunting by 64% [AOR=0.36; 95% CI: 0.3, 0.4] in 2015/16 and 2010, respectively. Other interesting results were that children with multiple births had 17% [AOR=0.83; 95% CI: 0.7, 0.9] lower odds and 68% [AOR=1.68; 95% CI: 1.5, 1.9] higher odds of stunting in 2004/5 and 2010, respectively.

Conclusion: Stunting is associated with individual and contextual factors. Critical efforts should focus on the availability of items that form WAMI and MPI, with greater emphasis on elements that form WAMI because of its consistent negative relationship with stunting across the two surveys. Nutrition education should target working mothers and households with food insecurity.