Estimating child dietary intake at national scale

Integrating two unlinked national surveys to estimate dietary intake in the child population, combining consumption data, anthropometry and vital records through statistical transfer.

Sector

Nutrition and food health

Client

New Seed

Scope

Children aged 6-59 months · national coverage

Deliverable

Individual nutritional profiles estimated through transfer between surveys

Challenge

Two national surveys describe the same population, but neither measures everything the analysis needs at once

The nutrition question required combining nutrient consumption with children's anthropometric status. Neither survey collected both.

Estimating iron, zinc and quality protein intake in a country's child population requires two types of information that rarely sit in the same source: household food consumption and the child's anthropometric measurements. In Guatemala, the ENCOVI survey collects household consumption and socioeconomic characteristics, while SIVESNU collects child anthropometry. They share neither individuals nor identifiers, so they cannot be linked directly.

Two technical limitations add to that disconnect. Consumption is recorded at household level, not individual level, so attributing intake to each child requires a principled allocation rather than a simple division. And children's age is coded in whole years, which is too coarse for growth tables: they work in months and have steep slopes at the youngest ages.

4 data sources

harmonised and integrated: two national surveys, the INCAP food composition tables and the vital registry of Guatemala's National Statistics Institute (INE)

2 national surveys

unlinked, connected through statistical transfer between populations

6 intake variables

modelled independently: energy, iron, zinc, digestible protein, lysine and tryptophan

From household to individual with Adult Male Equivalents

Intake is first calculated at household level, combining the INCAP food composition table (energy, protein, iron, zinc, amino acids, PDCAAS) with reported consumption and the corresponding unit conversions. To allocate that consumption among household members, the Adult Male Equivalent method is applied: it assigns each individual a share of intake according to their relative requirements, instead of an equal split that would ignore differences in age and sex.

Synthetic age in months anchored to the vital registry

To get past age being coded in whole years, an age in months is generated for each child with a quota-based allocation algorithm, calibrated against the empirical birth distributions in the INE vital registry by department and sex. The allocation is weighted by the survey's expansion factors, so that the distribution represented at population level reproduces the proportions observed in the registry, and it respects biological plausibility between siblings in the same household.

Transfer between surveys: gamma regression and distribution adjustment by quantile mapping

Nutritional profiles are transferred from one survey to the other in two stages. The first uses six gamma regression models with a log link, one per variable, suited to positive quantities with a skewed distribution. They are trained on the demographic, socioeconomic and geographical predictors common to both surveys, with Yeo-Johnson transformations of the continuous predictors, variable selection by Wald tests and the complex sampling design built in. A model's conditional mean reproduces the expected level of each variable but not the shape of its distribution in the target population, so a second, fine-tuning stage uses quantile mapping to align the full distribution of the predictions with the reference population distribution. The result preserves not only the central value but also the spread and the tails, which matters when the downstream analysis depends on thresholds of nutrient inadequacy.

Impact

Individual nutritional profiles for a population that previously had only anthropometry

Dietary intake estimated at individual level

Every child in the anthropometric survey now has an estimated intake profile for iron, zinc, digestible protein, lysine, tryptophan and energy, with the contribution of maize also separated from the rest of the diet. This enables the joint analysis of intake and nutritional status that neither survey allowed on its own.

The full age range, compatible with growth references

Synthetically recovering age in months makes it possible to represent the 6 to 11-month band specifically and improves compatibility with growth tables, reducing discretisation error at the ages where the curves are steepest.

A reusable population base for the rest of the programme

The transferred individual profiles and the age in months form the base for the programme's later analyses of nutritional impact and scenario simulation, with a documented, reproducible methodological chain open to external audit.

Integrating sources that share no individuals

Statistical transfer between independent surveys when there is no common identifier, preserving population distributions instead of forcing a match that the data do not support.

Complex sampling design built into population estimation

Expansion factors and survey structure built into the population allocation and the model estimation, so that the results represent the population and not just the sample.

Solutions anchored in official sources

Every decision rests on official sources, not on arbitrary assumptions: age in months is calibrated against the INE vital registry, and food composition comes from the INCAP tables.

A reproducible, auditable chain

The whole process, from raw source to individual profile, is documented in reproducible R code, ready for independent verification by auditors or donors.

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.