A Bayesian Bivariate Spatial Analysis of the Shared and Distinct Determinants of Stunting and Wasting Among Children in Ethiopia: Evidence from the 2019 Mini DHS

Using a Bayesian bivariate spatial analysis of 2019 Mini DHS data, this study reveals that while stunting and wasting in Ethiopian children share common socio-economic and biological risk factors, they exhibit largely independent geographic patterns, necessitating tailored interventions that address both chronic and acute malnutrition.

The Gist

Imagine Ethiopia as a giant, diverse garden. In this garden, there are two types of "weeds" that threaten the health of the young plants (the children):

1. Stunting: This is like a plant that is stunted in height. It's been growing slowly for a long time because it hasn't had enough good soil or water over the years. It's a chronic, long-term problem.
2. Wasting: This is like a plant that suddenly wilts and shrinks. It happened quickly, maybe because of a sudden drought or a pest attack. It's an acute, short-term emergency.

For a long time, gardeners (policymakers) looked at these two problems separately. They thought, "If we fix the water, we fix both," or they treated them as totally unrelated.

This paper is like a high-tech, 3D map that looks at both weeds at the same time.

Here is the story of what the researchers found, explained simply:

1. The New Tool: A "Double-Lens" Camera

Instead of taking two separate photos of the garden (one for stunting, one for wasting), the researchers used a special Bayesian Bivariate Spatial Model.

• The Analogy: Imagine wearing glasses with two lenses. One lens focuses on the slow-growing plants, and the other on the wilting ones. But the magic is that the lenses are connected. They talk to each other. If the "stunting lens" sees a pattern, it helps the "wasting lens" understand the picture better, even if the data for wasting is a bit fuzzy.
• The Result: They used a super-smart computer method (called SPDE-INLA) to draw a map that shows exactly where these problems are hiding, smoothing out the "noise" of small sample sizes.

2. The Big Surprise: They Live in Different Neighborhoods

The researchers expected the two problems to be everywhere together. They thought, "Where there is poverty, there is both stunting and wasting."

• The Reality: The map showed they are actually living in different neighborhoods.
  • Stunting (The Long-Term Problem): This is a big, heavy blanket covering the Northern and Central highlands. It's like a slow-moving fog. It's caused by deep, long-term issues like poor education and poverty that have been there for generations.
  • Wasting (The Emergency): This is like a sudden fire. It's concentrated in the Northeastern desert regions (Afar and Somali). These are areas where people herd animals (pastoralists). When the rain fails or a drought hits, these children get sick and lose weight quickly.
• The Lesson: You can't fight a fire with a long-term education plan, and you can't fix a slow-growing plant with a quick band-aid. The solutions need to be different for each region.

3. The "Household Head" Twist

The study found a funny and important pattern about who is running the house.

• The Analogy: Think of a Female-Headed Household as a family that is very good at planning and saving (like a wise gardener who rotates crops and saves seeds).
  • Result: Their children are less likely to be stunted (short) because the long-term care is good.
  • The Catch: But, these families are often poorer and more vulnerable to sudden shocks. If a drought hits or a family member gets sick, they don't have a safety net. So, their children are more likely to be wasted (suddenly sick).
• The Lesson: Female-headed households are doing a great job at long-term health but need emergency safety nets to handle sudden crises.

4. The "First 1,000 Days" Rule

The study confirmed what we suspected about age, but with a twist.

• Stunting: It's like building a house. If you don't lay a good foundation in the first two years (the first 1,000 days), the house will never reach its full height. The risk of being stunted skyrockets in the first 24 months and then levels off.
• Wasting: This is more like a sudden storm. It can happen at any age, but it's less tied to a specific "critical window" and more tied to immediate events like illness or lack of food right now.

5. The "Hidden" Factors

Even after looking at money, education, and where people live, the map showed there are still "ghosts" in the data.

• The Analogy: Imagine you've accounted for the soil and the rain, but some plants are still sick. There must be something else—maybe a hidden disease in the water, or a specific type of pest.
• The Finding: The researchers found that the "ghosts" (unseen factors) operate over a distance of about 158 kilometers. This means the problems aren't just about one village; they are about entire regions, likely due to things like local climate patterns, market access, or regional health systems that the survey didn't ask about.

The Bottom Line: One Size Does Not Fit All

The most important message of this paper is: Stop treating all malnutrition the same.

• For the North/Central Highlands: We need long-term structural changes. Build schools, improve farming, and lift families out of poverty. This is a marathon, not a sprint.
• For the Northeast Deserts: We need emergency response teams. We need mobile clinics, drought-resistant food supplies, and rapid aid when the weather turns bad. This is a sprint.

By using this "double-lens" map, Ethiopia can stop throwing resources at the wrong problems and start giving the right help to the right places, ensuring every child in the garden gets a chance to grow tall and strong.

Technical Summary

1. Problem Statement

Childhood malnutrition in Ethiopia presents a dual burden of stunting (chronic undernutrition) and wasting (acute undernutrition). While these conditions often co-exist, existing literature frequently treats them as independent outcomes or analyzes them using non-spatial models. This approach leads to two critical gaps:

• Siloed Analysis: Ignoring the biological and behavioral interdependence between stunting and wasting fails to capture shared risk factors and distinct etiological pathways.
• Spatial Neglect: Standard models often ignore spatial autocorrelation (the "neighborhood effect"), where nearby areas share similar risks due to unobserved environmental, climatic, or infrastructural factors. This can result in biased parameter estimates and the masking of critical sub-regional hotspots.

The study aims to disentangle the determinants of stunting and wasting by jointly modeling them as correlated binary outcomes while accounting for complex spatial dependencies and non-linear covariate effects.

2. Methodology

The study employs a Bayesian Bivariate Hierarchical Geostatistical Model using the Stochastic Partial Differential Equation (SPDE) approach combined with Integrated Nested Laplace Approximation (INLA).

• Data Source: Secondary data from the 2019 Ethiopia Mini Demographic and Health Survey (EMDHS).
  • Sample: 5,405 children (aged 0–59 months) from 305 clusters (93 urban, 212 rural).
  • Outcomes: Binary indicators for Stunting (HAZ < -2.0) and Wasting (WHZ < -2.0).
  • Covariates: Child-level (age, sex, birth plurality, breastfeeding), Maternal (education, age, parity), Household (wealth index, sex of head, number of under-5s), and Environmental (residence, GPS coordinates).
• Model Specification:
  • Joint Modeling: Outcomes $y_{ij}^{(k)}$ (where $k=1$ for stunting, $k=2$ for wasting) are modeled as Bernoulli trials with a logit link.
  • Linear Predictor: Includes outcome-specific fixed effects ($\beta$), non-linear smooth terms for child age and maternal age (using second-order random walk priors, RW2), and random effects.
  • Spatial Structure:
    • SPDE Approach: Represents the spatial random field as a Gaussian Markov Random Field (GMRF) over a triangulated mesh, utilizing a Matérn covariance function.
    • Bivariate Correlation: Models an exchangeable correlation structure between the spatial fields of stunting and wasting to estimate the residual spatial correlation ($\rho$).
    • Cluster Effects: Includes independent and identically distributed (i.i.d.) Gaussian random effects for clusters to account for sampling design.
  • Inference: Bayesian inference via INLA (R-INLA package) with weakly informative Penalized Complexity (PC) priors. Survey weights were incorporated via a weighted likelihood formulation.
• Model Selection: Four models were compared using Deviance Information Criterion (DIC), Watanabe–Akaike Information Criterion (WAIC), and Marginal Log-Likelihood. The final model included bivariate spatial effects and non-linear age terms.

3. Key Contributions

• Methodological Advancement: This is one of the first studies in Ethiopia to apply a bivariate SPDE-INLA framework to child malnutrition. It demonstrates how joint modeling "borrows strength" from the more stable indicator (stunting) to improve the precision of estimates for the more volatile indicator (wasting).
• Decoupling Drivers: The study successfully distinguishes between shared socio-economic drivers and outcome-specific geographic patterns, moving beyond national averages to provide high-resolution, region-specific risk maps.
• Non-Linear Age Effects: By using RW2 priors, the study captures the non-monotonic, U-shaped, and plateauing effects of child and maternal age on malnutrition, which linear models often miss.

4. Key Results

A. Model Performance

• Model 4 (Bivariate spatial + non-linear child and maternal age) provided the best fit (lowest DIC: 8753.63; highest Marginal Log-Likelihood: -4645.46), confirming the necessity of modeling non-linear age effects and distinct spatial processes.

B. Spatial Correlation

• Residual Correlation ($\rho$): The estimated spatial correlation between stunting and wasting was weak and negative ($\rho = -0.12$, 95% CI: -0.41 to 0.19).
• Interpretation: After adjusting for observed covariates (wealth, education, etc.), the unexplained geographic patterns of stunting and wasting are largely independent. This implies that the drivers of acute shocks (wasting) are distinct from the drivers of chronic deprivation (stunting) at the residual level.

C. Determinants (Fixed Effects)

• Shared Risk Factors:
  • Male Sex: Higher risk for both stunting (OR: 0.71 for females) and wasting (OR: 0.59 for females).
  • Multiple Births: Significantly increased risk for both (Stunting OR: 1.82; Wasting OR: 2.36).
  • Maternal Education & Wealth: Higher education and wealth are strong protective factors, particularly for stunting.
  • Household Size: More under-5 children increases stunting risk.
• Divergent Effects:
  • Female-Headed Households: Associated with lower odds of stunting (OR: 0.73) but higher odds of wasting (OR: 1.40). This suggests these households manage long-term nutrition well but are more vulnerable to acute economic/environmental shocks.
  • Place of Residence: After spatial adjustment, the urban/rural divide was not statistically significant for either outcome, suggesting that spatial clustering explains much of the rural disparity.

D. Spatial Patterns (Random Effects)

• Stunting Hotspots: Concentrated in the Northern and Central Highlands (Amhara and Tigray regions), indicating persistent chronic deprivation.
• Wasting Hotspots: Concentrated in the Northeastern Pastoralist regions (Afar and Somali regions), indicating vulnerability to climatic shocks and food insecurity.
• Spatial Range: The estimated spatial range is 158 km, suggesting unobserved drivers operate at a broad regional scale rather than just the village level.

E. Age Effects

• Child Age: Stunting risk rises sharply in the first 24 months (the "first 1,000 days") and then plateaus. Wasting risk is more stable across age but shows slight increases in early infancy.
• Maternal Age: A U-shaped relationship was observed, with higher risks for children of adolescent mothers and a slight increase for older mothers.

5. Significance and Policy Implications

• Targeted Interventions: The findings argue against a "one-size-fits-all" approach.
  • Highlands (Stunting): Require long-term structural interventions (maternal education, poverty alleviation, agricultural diversification).
  • Pastoralist Regions (Wasting): Require rapid-response, shock-responsive programs (drought-resilient food systems, mobile health services).
• Resource Allocation: The exceedance probability maps provide a statistically rigorous basis for prioritizing zones where malnutrition prevalence is highly likely to exceed WHO thresholds.
• Female-Headed Households: Policies should include safety nets specifically designed to buffer these households against acute shocks, even if their long-term nutritional stability is relatively better.

In conclusion, the study demonstrates that while stunting and wasting share common socio-economic roots, their geographic expressions are distinct. Effective policy in Ethiopia requires a dual strategy that addresses chronic structural deprivation in the highlands while simultaneously building resilience against acute shocks in pastoralist regions.