Food insecurity, caloric intake and nutritional status among children under 5 years old: a predictive modelling analysis of the MAL-ED multi-country cohort

Problem Statement
In humanitarian and crisis settings, decision-making regarding resource allocation and service design for preventing child acute malnutrition relies on robust estimates of malnutrition prevalence and trends. While mechanistic models exist to project weight trajectories and malnutrition evolution, they critically require caloric intake as an input. However, caloric intake is notoriously difficult to measure in real-time, resource-constrained environments, often requiring lengthy 24-hour dietary recalls. The authors investigated whether statistical predictive models could serve as a "shortcut," accurately forecasting caloric intake or anthropometric changes (specifically weight-for-height Z-score, WHZ) based on more readily available predictors such as household food insecurity, infectious disease prevalence, and demographic factors. The central question was whether statistical models could offer accurate predictions along the causal pathway from food insecurity to caloric intake to nutritional status, potentially replacing or supplementing ground data collection.

Methodology
The study conducted a secondary analysis of the Malnutrition and Enteric Disease (MAL-ED) multi-country birth cohort (2009–2014), which included children from Bangladesh, Brazil, India, Nepal, Pakistan, Peru, South Africa, and Tanzania. The dataset provided longitudinal data on household food insecurity (using a 9-item questionnaire aligned with the FAO Food Insecurity Experience Scale), dietary intake (24-hour recalls), anthropometry, and infectious disease symptoms for children aged 9 to 35 months.

The authors pre-processed the data to align variables around anthropometric measurement instances, calculating the absolute daily mean change in WHZ. They excluded observations with missing key covariates (country, residence, birth weight, breastfeeding status, disease prevalence) and those where measurements fell outside specific time windows (e.g., caloric intake or food insecurity data more than 7 days from the anthropometric measure).

Three predictive models were tested:

• M1: Predicting change in WHZ as a function of household food insecurity.
• M2: Predicting change in WHZ as a function of caloric intake.
• M3: Predicting caloric intake as a function of household food insecurity.

All models included additional predictors: age, sex, birth weight, urban/rural residence, breastfeeding status, and the longitudinal prevalence of diarrhoea, acute respiratory infection (ARI), and fever. Four machine learning and statistical methods were evaluated for each model: Random Forests, Lasso regression, Generalised Additive Models (GAMs), and Generalised Boosted Regression (GBM). Performance was assessed using 100-fold cross-validation, with metrics including Root Mean Square Error (RMSE), absolute bias, and prediction line dilution (regression dilution). Sensitivity analyses were performed by fitting models to individual country datasets.

Key Results
The analysis yielded negative findings regarding the feasibility of statistical prediction in this context:

• Low Correlation: There was low correlation among the key variables of interest (food insecurity, caloric intake, and WHZ) at both country and individual levels.
• Poor Model Performance: All three models (M1, M2, and M3) demonstrated low predictive performance across all tested methods.
  • M1 (Food insecurity → WHZ change): Showed extreme regression dilution (slopes near zero) and high relative bias (e.g., -87.3% to -58.4% depending on the method).
  • M2 (Caloric intake → WHZ change): Also exhibited significant regression dilution (71.4% to 95.0%) and variable bias.
  • M3 (Food insecurity → Caloric intake): While this model showed the most promising performance with lower relative error (5.3% to 11.6%), it still suffered from substantial regression dilution (48.9% to 66.3%), indicating that food insecurity was a weak predictor of actual caloric intake.
• Predictor Importance: Feature importance analysis revealed that food insecurity was less important than birth weight and age for predicting WHZ changes (M1) and significantly less important than breastfeeding status for predicting caloric intake (M3). Conversely, caloric intake was the most important predictor for WHZ changes (M2), though the model still failed to predict the outcome accurately.
• Sensitivity: Low performance persisted even when models were fitted to individual country cohorts or when excluding specific countries like Bangladesh.

Significance and Conclusions
The paper concludes that statistical prediction of key variables along the causal pathway to childhood acute malnutrition—specifically predicting anthropometric changes from food insecurity or caloric intake from food insecurity—is likely not feasible using the tested approaches and datasets. The authors attribute these negative findings to several factors:

1. Measurement Error: Inaccuracies in predictor measurement (e.g., caregiver-reported illness, 24-hour dietary recall limitations, and FIES questionnaire biases) likely caused significant regression dilution.
2. Data Range: The MAL-ED cohort, while rigorous, did not encompass the extreme scenarios of food insecurity and starvation common in acute crisis settings. The narrow range of predictor and outcome values in the study limits the model's ability to capture the severe dynamics seen in humanitarian emergencies.
3. Complexity: The relationship between food insecurity, coping mechanisms, and malnutrition is likely non-linear and complex, defying simple statistical modeling.

The authors state that these findings imply that mechanistic models requiring caloric intake as an input cannot rely on statistical shortcuts derived from food insecurity data. Instead, such models must depend on empirical data (resource-intensive dietary recalls) or scenario assumptions. The paper suggests that future efforts to improve prediction would require ad-hoc, prospective data collection in settings with extreme variability, accompanied by repeat measurements to correct for regression dilution. Until then, mechanistic models or expert elicitation, potentially combined with limited ground measurement, remain the more promising approaches for anticipating malnutrition evolution in crisis settings.