Impact of biofortification on chronic child undernutrition
Estimating the effect of biofortified maize on child stunting through two distinct analytical pathways that converge on the same result.
Attributing an improvement in child growth to a change in diet is one of the most fragile inferences there is. A single method is not enough to defend it
The question was how much biofortified maize reduces stunting. The difficulty is not calculating a number; it is making that number withstand scrutiny.
Child stunting is driven by many intertwined factors: nutrition, illness, household conditions, maternal characteristics and access to services. Isolating the contribution of a single dietary change—replacing conventional maize with a biofortified variety richer in iron, zinc and quality protein—means separating that signal from everything else, and doing so with observational data, not a controlled trial.
An estimate like this, presented to an international donor with a single method behind it, is vulnerable: if the method is biased, the whole result is shaken. The challenge was not to produce an impact figure but a defensible one, able to withstand the objection that the result depends on the technique chosen.
2 complementary pathways
one based on the observed data, the other on evidence from the literature, converging on the same effect
6-59 months
target child population, aligned with WHO recommendations
5.8 points
reduction in stunting on which both pathways converge, under 100% coverage of the child population
Two distinct analytical paths to the same effect: the association observed in the data and the effect size from the literature
Endogenous pathway: nutritional index and GAM on observed growth
The first pathway estimates the effect from the association observed in the very population that has anthropometric data. A composite nutritional index is built from the child's bioavailable nutrients, and its effect on the height-for-age Z-score is modelled with a generalised additive model (GAM), which captures non-linear relationships that a linear regression would not represent. Preparation follows a seven-step, audit-ready variable selection pipeline—data dictionary audit, missing data handling, near-zero variance, Random Forest imputation, dimension reduction by conceptual blocks using PCA, correlation filtering and Boruta selection—so that every variable that enters the model does so for a documented reason.
Exogenous pathway: meta-analysis effect size modulated by individual adequacy
The second pathway starts not from the local data but from published evidence: the effect size from a meta-analysis of quality protein maize (QPM), a 9% increase in height growth velocity. That effect is not applied flatly but modulated child by child by a protein adequacy factor that acts as a dose-response curve—a logistic sigmoid that attenuates the effect in children with severe protein deficit, because QPM improves protein quality, not quantity—after which each child's height and Z-score are recalculated with the WHO LMS growth tables. It is the rigorous translation of a population-level finding to individual heterogeneity.
Bioavailable nutrition, not ingested nutrition
Both pathways share a carefully built nutritional base: what matters for growth is not the nutrient ingested but the nutrient actually absorbed. Supplementation with micronutrient powders is incorporated through a three-level adherence cascade (access, administration, dosing), absorbed zinc is estimated with the Miller equation and iron with an absorption factor suited to a plant-based diet, and each child's adequacy is classified against age-specific requirement thresholds. The biofortification factors are then applied to the maize-derived nutrients on that base.
An impact estimate that does not depend on a single method
Effect on stunting quantified at individual level
Every child in the population has a stunting profile under two conditions, conventional maize and biofortified maize, so the intervention's effect on stunting can be quantified child by child and aggregated to any geographical level, not just as a national average. That granularity is what later makes it possible to analyse the impact department by department.
Convergence between two methods as evidence of consistency
Each pathway separately estimates a 5.8 percentage point reduction in stunting under the scenario of 100% coverage of the child population, linked to the seed availability projected for 2028. When a pathway anchored in local data and another anchored in the international literature converge on the same figure, reliance on a single approach is reduced. The agreement strengthens the consistency of the estimate, although it does not remove the uncertainty or turn the projection into an observed effect.
An individual-level base ready for coverage scenario simulation
The full nutritional recalculation under the biofortified condition is precomputed for every child, together with a priority ranking for access to maize based on the household's farming profile. This leaves the ground prepared so that the programme's scenario simulation only has to select which children each penetration level reaches, without redoing the nutritional calculation.
What this case shows about the way we work
Methodological triangulation, not a single method
Estimating the same effect through two epistemologically distinct pathways—one endogenous to the data, the other exogenous, from the literature—is what turns a figure into a defensible result. It is a form of quality control that a single model, however sophisticated, cannot provide.
Audit-ready variable selection
The seven-step pipeline, from the data dictionary to the final model, is designed so that an external auditor can reconstruct why each variable is in or out. In a field where decisions have public health consequences, traceable variable selection is part of the rigour, not a formality.
Absorbed nutrition, not just reported intake
Modelling absorbed zinc with the Miller equation, iron with a factor consistent with a plant-based diet, and supplementation with a realistic adherence cascade avoids the common mistake of treating the nutrient ingested as if it were the nutrient used. The biology of absorption is in the model, not assumed.
From population finding to the individual child
Translating a meta-analysis effect size to each child through dose-response modulation, instead of applying it as a flat average, respects individual heterogeneity in response and connects the international evidence with the study's specific population.
I've been working with Rubén and Julia on a complex modeling project for over a year now, and they have been exceptional partners throughout. As our project has evolved, they have approached every shift collaboratively, receptively, and with flexibility. They have consistently gone above and beyond to ensure we develop both a robust model and a user interface that translates model predictions into clear insights. What we have now is a fully documented, auditable methodology we can put in front of funders and external reviewers with confidence.
Katie Adams
Director of Impact and Cost Effectiveness, New Seed
Capabilities deployed in this project
Shall we talk?
Tell us what you are working on and we will give you a straight answer on how we can help.