Dietary survey data: avoiding fortification modeling errors

If day-to-day variation is left in the dataset, the modeled distribution becomes artificially wide. The result is simultaneous overestimation of deficiency risk and excessive-intake risk.

This is not a minor statistical defect. It changes the fortification dose, the food vehicle selected, the estimated number of inadequately supplied people, and the probability that the population approaches a Tolerable Upper Intake Level. Food fortification policy planning therefore depends on data processing before it depends on nutrient chemistry.

The relevant workflow is sequential:

1. Estimate usual intake rather than observed short-term intake.

2. Separate within-person and between-person variance.

3. Correct nutrient composition data for added and intrinsic sources.

4. Model food allocation when the input is household-level.

5. Test fortification scenarios against both inadequacy and upper-intake constraints.

6. Report uncertainty instead of presenting a single dose as a measured fact.

A dietary survey is an observation system. It becomes a policy instrument only after measurement error has been modeled.

The perils of unadjusted dietary recall data

A 24-hour dietary recall has a defined use. It captures reported consumption during a specified period. It does not directly measure the long-term distribution of usual intake in a population.

Vitamin D intake illustrates the problem. Intake from fortified foods can be intermittent. A person may consume a fortified dairy product or breakfast cereal on one survey day and none on the next. Fish consumption is also irregular. A short-term record can therefore place the same person at an artificially low or high position in the population distribution.

For population assessment, the relevant variable is usual intake:

• the long-term average intake of an individual;
• the distribution of those individual averages across the population;
• the proportion below an Estimated Average Requirement or another adequacy threshold;
• the proportion above a defined safety threshold.

The observed recall contains two components. One is the persistent difference between people. The other is random day-to-day fluctuation within the same person. The second component must be reduced or removed statistically.

Without that correction, low-intake days are treated as if they represent low usual intake. High-intake days are treated as if they represent high usual intake. The tails of the distribution expand. Both ends become unreliable.

This creates two predictable distortions.

Deficiency risk is inflated

Suppose a person has moderate average intake but consumes vitamin D-containing foods irregularly. A recall taken on a low-consumption day places that individual below the adequacy threshold. If many participants are processed in the same manner, the model counts temporary low days as persistent dietary inadequacy.

The estimated prevalence of insufficiency rises. A policy analyst may then select a higher fortification concentration, add more food vehicles, or classify a larger population as requiring intervention.

The error is especially material when:

• the nutrient is concentrated in a small number of foods;
• fortified products are not consumed daily;
• the survey includes only one recall per participant;
• the distribution is evaluated near a low threshold;
• the policy scenario depends on small changes in median or lower-percentile intake.

Excess-intake risk is also inflated

The same unadjusted data can produce an exaggerated upper tail. A participant who reports several fortified products on one day may appear to have a high usual intake. The model may classify that temporary observation as evidence of potential Tolerable Upper Intake Level exceedance.

This can lead to an unnecessarily conservative fortification limit. It can also produce an internally inconsistent result: a population is modeled as having substantial inadequacy and substantial excessive intake at the same time because the data contain unseparated within-person variance.

That combination is not impossible in a real population. It is often a measurement artifact when based on raw short-term recalls.

The mean is not the main problem

A common response is to compare the mean observed intake with the mean intake after fortification. This is insufficient. The mean can remain relatively stable while the distribution changes materially.

Fortification decisions depend on the distribution. The policy question is not only whether average vitamin D intake increases. It is also:

• how many people move above the adequacy threshold;
• whether vulnerable age or sex groups remain below it;
• how the 95th percentile changes;
• whether high consumers approach or exceed the UL;
• whether the modeled result is driven by a small number of irregular consumption records.

A population mean cannot answer these questions.

Statistical methods for estimating usual intake

The correction requires a statistical method that isolates within-person variance. Several established approaches are used in dietary intake analysis:

• the National Cancer Institute method;
• the Multiple Source Method;
• the Iowa State University method;
• SPADE.

These methods differ in implementation, assumptions, software environment, and treatment of episodically consumed foods. They share the same operational purpose: estimate the distribution of usual intake from repeated short-term observations.

The input structure normally includes repeated dietary measurements for at least a subset of participants. Covariates can include age, sex, survey day, weekend status, season, and relevant demographic variables. The method then separates systematic differences between individuals from random variation within individuals.

NCI method

The NCI method is designed to estimate usual intake distributions and the proportion of a population below or above specified thresholds. It can handle foods and nutrients consumed nearly every day as well as episodically consumed items.

For an almost-daily nutrient, the model focuses on the amount consumed. For an episodic food, it first addresses the probability of consumption and then the amount consumed on a consumption day. This distinction matters for vitamin D modeling because the nutrient may enter the diet through irregular food categories such as fish or selected fortified products.

The NCI approach is usually implemented with a two-part model for episodic consumption:

1. the probability that an individual consumes the food on a given day;

2. the amount consumed conditional on consumption.

The two components are then combined to estimate usual intake. Covariates can be used to account for structured differences in consumption patterns.

Multiple Source Method

The Multiple Source Method also estimates usual intake by combining repeated observations and covariate information. It is often used in population nutrition studies where the available survey structure does not fit a simple averaging procedure.

The method estimates usual consumption through a combination of consumption probability and consumed amount. This is useful when a food or nutrient has many zero-intake records. A zero on one recall day is not interpreted automatically as zero usual intake.

Iowa State University method

The Iowa State University method is another established approach for estimating usual intake distributions. Comparative simulation work has evaluated it alongside the NCI method, the Multiple Source Method, and SPADE.

The choice among methods should not be made by brand preference. It should follow the survey design and the target variable. A model for daily nutrient intake is not identical to a model for an episodically consumed food. The treatment of zeros, transformation of intake data, variance structure, and adjustment for covariates all affect the output.

SPADE

SPADE is used to estimate usual intake distributions and inadequacy prevalence, particularly in age- and sex-specific population assessments. It provides a structured route from observed intake data to estimated usual intake.

The method selection is only one part of the process. The analyst must also document:

• the number and timing of repeated recalls;
• the proportion of participants with repeat measurements;
• the nutrient database version;
• treatment of supplements;
• treatment of fortified products;
• age and sex stratification;
• the selected reference requirement and UL;
• the assumptions used for episodic consumption.

A technically correct method can still produce a poor policy estimate if the nutrient composition database is incomplete or the survey design is ignored.

Fortification dose should be selected from a modeled usual-intake distribution, not from the raw minimum and maximum of recall data.

Food composition databases are a second failure point

Statistical adjustment cannot correct a nutrient database that does not identify the source of the nutrient.

Food composition systems may report the total vitamin D content of a product without consistently distinguishing between:

• naturally occurring vitamin D;
• vitamin D added by the manufacturer;
• nutrient contributed by a premix;
• nutrient added under a voluntary fortification standard;
• nutrient present because of reformulation or product-specific enrichment.

This distinction affects both baseline intake and the modeled effect of a new policy.

If voluntarily fortified foods are under-recorded, the baseline population intake is underestimated. A proposed intervention may then appear to deliver a larger incremental benefit than it would in practice. The model can also overstate the number of people below an adequacy threshold.

The opposite problem is possible. If a food database assigns a generic fortified composition to products that are not consistently fortified, the baseline intake is overstated. The modeled benefit of mandatory fortification is then reduced artificially.

Product-level composition is not optional in a detailed model

A national survey often records foods using categories that are broader than the commercial market. The database may contain a single entry for breakfast cereal, milk substitute, spread, or dairy product. The market may contain multiple products with different fortification concentrations.

For vitamin D, this creates a direct problem. Two foods with similar macronutrient profiles can have materially different vitamin D concentrations. The nutrient difference is invisible if the survey coding system does not capture brand, product type, or fortification status.

At minimum, a policy model should define how it handles:

• branded versus generic food records;
• fortified and non-fortified variants;
• standard recipes and reformulated products;
• imported products with different fortification rules;
• fortified foods consumed outside the home;
• supplements and medicinal preparations;
• missing or uncertain fortification values.

Where exact brand-level concentrations are unavailable, the model should not manufacture them. It should use documented assumptions and run sensitivity analyses across plausible concentrations.

Added nutrient and intrinsic nutrient are not interchangeable

The policy mechanism is different depending on the source. A naturally occurring vitamin D contribution is part of the baseline food supply. A voluntarily added contribution is a market-dependent input. Mandatory fortification is a regulatory intervention.

Combining these sources into one undifferentiated nutrient value weakens the model. It prevents the analyst from identifying which portion of population intake is already controlled by policy, which portion depends on manufacturer behavior, and which portion could change if product composition changes.

This matters for compliance analysis. A country may have a fortification standard on paper, but the effective population exposure depends on the concentration actually present in products, the share of products covered, and consumer intake of those products.

The database must therefore represent the food system, not only the nutrient table.

Household expenditure surveys require allocation modeling

Household Consumption and Expenditure Survey data can provide broad information about food acquisition. They are useful where individual dietary surveys are limited or absent. They do not directly provide individual dietary intake.

The household is the measurement unit. The fortification model often requires the individual. The conversion between the two requires assumptions about allocation.

A household purchase of fortified flour, milk, or cooking oil does not specify:

• which household member consumed the product;
• how much each member consumed;
• whether the product was consumed inside or outside the home;
• how much was wasted;
• how much remained in storage;
• whether consumption differed by age, sex, work pattern, or health status.

An apparent individual intake estimate derived from household acquisition data is therefore conditional on an allocation model. It is not an observed individual measurement.

Allocation assumptions can change the policy conclusion

A simple per-capita division assumes equal distribution. That assumption is rarely neutral. Children, adults, and older people do not consume identical quantities. Household members may have different access to specific foods. Food may be preferentially served to some members or consumed outside the household.

If fortified food is allocated equally in the model, exposure among some groups may be overestimated and among others underestimated. The error is then carried into inadequacy and UL calculations.

The solution is not to discard HCES data. The solution is to state the allocation mechanism and test it.

Useful sensitivity scenarios can include:

• equal per-capita allocation;
• age-adjusted allocation;
• sex- and age-specific consumption coefficients;
• allocation based on observed dietary survey patterns;
• exclusion or separate treatment of food consumed away from home;
• alternative waste and storage assumptions.

The results should be triangulated with individual dietary surveys where available and with biomarker data where the nutrient and biomarker relationship is sufficiently interpretable.

HCES and dietary surveys answer different questions

An individual dietary survey estimates reported consumption by person. An HCES measures household acquisition and expenditure. They should not be merged as if they were interchangeable observations.

Their roles can be separated:

• individual dietary recalls provide person-level consumption patterns;
• HCES data describe market access, household acquisition, and food availability;
• food balance or supply data provide national-level availability;
• biomarkers provide physiological status, subject to timing and biological interpretation.

The strongest fortification models use these data sources as complementary layers. They do not allow one source to substitute silently for another.

Designing a fortification scenario around the distribution

Once usual intake and food composition have been estimated, the intervention can be modeled. The scenario should specify the food vehicle, coverage, concentration, consumption frequency, compliance, and population subgroup.

A generic scenario can be represented through several operational variables:

• baseline usual vitamin D intake;
• additional vitamin D per unit of fortified food;
• quantity of the food consumed;
• frequency of consumption;
• proportion of products complying with the standard;
• proportion of the population consuming the vehicle;
• nutrient degradation during processing and storage;
• contribution from supplements and other fortified foods.

The resulting intake is not simply the legal fortification concentration. It is the concentration multiplied by effective exposure.

Legal concentration is not delivered concentration

Industrial processing can reduce the vitamin content between premix addition and consumption. The relevant parameter is the degradation rate across the full production and distribution chain.

Losses can occur during:

• thermal processing;
• extrusion;
• drying;
• oxygen exposure;
• light exposure;
• storage;
• interaction with the food matrix;
• preparation by the consumer.

Matrix encapsulation can improve stability, but it can also change dispersion, release, analytical recovery, and cost. The premix specification must therefore be connected to the finished-food concentration, not treated as an independent laboratory value.

A model that uses the declared addition level without a degradation adjustment will overestimate bioavailability yield and population exposure. A model that assumes complete loss will produce an unnecessarily high addition target. Neither is a production-relevant estimate.

Compliance is a distribution, not a binary label

A food category may be legally covered by a fortification standard while actual products vary in concentration. Compliance can be affected by premix handling, dosing accuracy, batch variation, procurement quality, laboratory capacity, and enforcement.

Global modeling across 185 countries has estimated that existing food fortification programs prevent 7.0 billion inadequate person-nutrient intakes annually. The same analysis estimated that improved compliance could prevent an additional 13.1 billion inadequacies.

The operational implication is direct. Expanding the legal scope of fortification is not the only available intervention. Improving compliance with an existing standard may produce a larger effective change than increasing the nominal dose.

A scenario should therefore distinguish at least three conditions:

1. nominal fortification under the legal standard;

2. observed or estimated current compliance;

3. improved compliance with the same standard.

These are different policy states. Combining them conceals the implementation gap.

Safety evaluation requires the upper tail

Fortification modeling cannot stop after estimating the reduction in inadequacy. It must examine high intake.

A common safety baseline in population modeling is the 95th percentile of usual intake. That percentile is not the UL. It is a population-level screening point used to evaluate whether a proposed fortification scenario pushes high consumers toward an established upper limit.

The analysis should evaluate:

• the 95th percentile of baseline intake;
• the 95th percentile after fortification;
• intake among high consumers of multiple fortified products;
• supplement users where data permit;
• subgroup-specific distributions;
• uncertainty in food composition and compliance;
• the relationship between modeled intake and the applicable UL.

The upper tail is particularly sensitive to database errors. If added vitamin D is not distinguished from intrinsic vitamin D, or if multiple fortified products are collapsed into generic food categories, high consumers can be misclassified.

The 15 E% and 25 E% energy-intake assumptions evaluated in some fortification safety models illustrate another issue. Safety results can depend on how much energy is assumed to come from the food and beverage categories under consideration. A model that changes the energy contribution without documenting the assumption is not comparable across scenarios.

A production-grade workflow for national survey data

A practical analysis should move through defined stages. Each stage produces an input for the next.

1. Define the policy population

Specify age groups, sex groups, pregnancy status where relevant, geographic coverage, and whether institutionalized populations are included. Vitamin D intake patterns and supplement use are not uniform across these strata.

Do not produce a single national estimate first and apply subgroup conclusions later. Fortification policy may work adequately for one group and leave another below the target.

2. Audit the dietary instruments

Record the number of recalls, recall spacing, season, weekday distribution, reporting method, and missingness. Determine whether the survey captures recipes, brand information, fortified status, supplements, and food consumed outside the home.

The instrument defines the resolution of the final model. A database cannot recover product detail that the survey never collected.

3. Harmonize foods and nutrients

Map survey food codes to a consistent composition database. Track every transformation. Identify whether the nutrient value is intrinsic, added, assumed, or missing.

Dietary survey harmonization is not a cosmetic data-cleaning step. It determines whether a fortified food is visible to the model.

4. Estimate usual intake

Select NCI, MSM, ISU, SPADE, or another justified method based on the structure of the data. Separate daily nutrients from episodically consumed foods. Use repeated observations to estimate within-person variance.

Report uncertainty intervals or scenario ranges. A single point estimate implies a level of precision that the survey may not support.

5. Add the intervention

Apply the proposed concentration to the relevant food records. Adjust for expected degradation rates, product coverage, compliance, and consumption frequency. If the intervention targets a processed food, use the finished-food concentration rather than the premix input.

Where matrix encapsulation is used, treat stability and bioavailability yield as separate parameters. Chemical retention is not automatically equivalent to physiological availability.

6. Test adequacy and safety together

Calculate the change in inadequacy prevalence and the change in high-percentile intake. Examine both national and subgroup results.

A fortification scenario that reduces inadequacy but creates a material upper-tail concern requires reformulation, a narrower vehicle, lower concentration, or a targeted strategy. The model should identify that trade-off rather than hide it in an average.

7. Stress-test the assumptions

Run sensitivity analyses for:

• food composition values;
• voluntary fortification prevalence;
• compliance;
• degradation rates;
• supplement use;
• household allocation;
• energy share of the targeted food categories;
• missing consumption days;
• alternative statistical methods.

The purpose is not to make the model indecisive. It is to identify which parameters control the policy result.

Global scale does not reduce the need for local data quality

Global fortification models can quantify the scale of preventable inadequacy. They can compare countries and identify where improved program performance may have the largest effect. They cannot replace national measurement of the foods actually consumed and the products actually sold.

A global estimate may use harmonized assumptions across 185 countries. National implementation requires local detail:

• the legal fortification vehicle;
• the actual market share of covered products;
• the population’s consumption frequency;
• the local nutrient composition database;
• compliance with the standard;
• supplement availability;
• food preparation practices;
• the applicable dietary reference values and UL.

The more aggregated the model, the more carefully its limitations must be communicated. Global outputs are useful for prioritization. They are not finished dosing instructions for a national program.

The same rule applies to regulatory interpretation. EFSA opinions, WHO guidance, national dietary reference intakes, food safety regulations, nutrition labeling laws, and health claim requirements operate at different levels. A dietary model can estimate exposure. It does not, by itself, establish legal authorization or a permitted health claim.

The cost-benefit calculation is statistical before it is financial

Fortification cost is often calculated as premix cost per unit of food. That is incomplete. The real cost includes analytical testing, dosing equipment, quality control, reformulation, packaging, stability validation, monitoring, and enforcement.

The benefit is also not the nominal amount of vitamin D added. It is the number of inadequate intakes prevented after degradation, incomplete compliance, irregular consumption, and existing dietary sources are included.

A technically efficient intervention has four characteristics:

• the food vehicle is widely consumed by the target population;
• intake is sufficiently regular to produce a stable exposure;
• the nutrient remains stable through processing and storage;
• the compliance burden is proportionate to the expected public health gain.

A high concentration in a poorly consumed or poorly controlled vehicle may produce less benefit than a lower concentration in a widely consumed product with reliable compliance.

The model should therefore compare scenarios on effective delivered intake, not declared concentration. It should also show the cost of improving compliance separately from the cost of increasing the legal dose.

Conclusion

National dietary survey data fortification modeling fails when short-term observations are treated as usual intake, when food composition databases conceal added nutrients, or when household acquisition is presented as individual consumption.

The corrective route is defined:

1. isolate within-person variance with an appropriate usual-intake method;

2. harmonize food records with source-specific nutrient composition;

3. model voluntary fortification and compliance explicitly;

4. allocate household food data through tested assumptions;

5. incorporate degradation rates and finished-food concentrations;

6. evaluate inadequacy and upper-tail exposure in the same scenario;

7. report sensitivity ranges instead of false precision.

The cost-benefit conclusion is strict. Better data processing usually produces a more targeted fortification dose, a clearer estimate of prevented inadequacy, and fewer unnecessary safety constraints. The laboratory and the factory cannot compensate for an intake distribution built from the wrong statistical unit.