When the denominator becomes a model

#Question

What changes when the population at risk is estimated dynamically by a model rather than treated as a fixed count?

#Answer in brief

The final indicator becomes partly model-dependent.

A public-health rate is often written as if both parts were observed:

rate = observed events / population at risk

In practice, the denominator is frequently an estimate. Census projections, registry counts, administrative lists, mobility and eligibility rules already make populations less fixed than the formula suggests. WHO explicitly treats target-population quality as a central problem in vaccination coverage measurement.^[1]

AI and modern data integration can push this further. A denominator may increasingly be inferred from multiple imperfect sources, updated over time and stratified at finer geographic or demographic resolution.

That can improve measurement.

It also means the denominator is no longer just a number in the table. It is an output of a model whose assumptions, uncertainty and version belong to the indicator.

#Denominators are population models even before AI

The denominator in a rate answers a substantive question:

Who was actually at risk, eligible or under observation during this period?

That question is rarely identical to:

How many people were listed in one administrative file?

A vaccination denominator may depend on age, residence, schedule, migration and eligibility.

A disease-incidence denominator may need to reflect a population that changes seasonally or moves across administrative boundaries.

A service-utilization denominator may depend on the population that actually uses a catchment area rather than on legal residence alone.

Research using mobile-phone data has already shown that seasonally varying population numbers can materially change estimated disease incidence when static denominators are replaced with dynamic ones.^[2]

AI does not invent the denominator problem.

It makes much more ambitious denominator models feasible.

#From projected population to inferred population

Consider a city that wants a weekly estimate of the population present in each district.

Potential signals could include:

• civil-registration and resident-registry data;
• census projections;
• school enrolment;
• health-service utilization;
• public-transport use;
• mobile-device mobility aggregates;
• tourism data;
• housing occupancy;
• employment location;
• recent migration indicators.

A statistical or AI system could combine these signals into a continuously updated estimate.

That estimate may be more realistic than an annual projection.

But it is still an estimate.

The final incidence rate now depends on:

1. observed or classified events;
2. the population model;
3. the data sources feeding the model;
4. the assumptions used to reconcile conflicting sources;
5. the update schedule;
6. the uncertainty of the estimate.

The denominator has become an analytical object.

#Model-dependent public-health indicators

This article proposes the term model-dependent public-health indicator for an indicator whose reported value materially depends on a model-generated numerator, denominator or both.

The concept is deliberately broader than AI.

A Bayesian population model, a statistical small-area estimator and an AI-driven multisource estimator can all create model dependence.

The important distinction is not whether the system is labelled AI.

It is whether the final indicator can be interpreted without understanding the model that created one of its components.

For a model-dependent denominator:

indicator_t = observed numerator_t / estimated population_t(model, data, assumptions)

The uncertainty of the denominator is part of the uncertainty of the indicator.

#Precision can hide denominator uncertainty

Suppose 8,340 eligible people are recorded as vaccinated and a model estimates 10,000 eligible residents.

Coverage is reported as:

83.4%

That decimal looks precise.

But imagine the plausible denominator is 9,500–10,700 depending on migration and eligibility assumptions.

The corresponding coverage range is approximately 78%–88%.

A single decimal place can therefore communicate numerical precision while hiding population uncertainty.

WHO's target-population guidance explicitly identifies implausibly high coverage, erratic year-to-year changes and outbreaks despite high reported coverage as signals that denominator estimates may be inaccurate.^[1]

The same principle applies when the denominator is produced by a more sophisticated model: more complexity does not remove the need to communicate uncertainty.

#An AI denominator can change while the population does not

Suppose a dynamic population estimator is updated from version 2.1 to 2.2.

The new model gives greater weight to recent mobility data.

No one moves overnight.

No vaccination records change.

But the estimated eligible population for one district increases by 6%.

Reported coverage falls.

A time-series user might interpret that fall as deterioration in programme performance.

In fact, the measurement system changed.

This is the denominator equivalent of changing a diagnostic assay in the middle of surveillance.

Model version is therefore epidemiological metadata.

#Numerator and denominator can become coupled

A deeper problem appears when the same data source influences both parts of the rate.

Suppose health-service utilization helps estimate the local population denominator.

The numerator is also derived from health-service records.

An outbreak increases health-care attendance.

The denominator model sees more activity and infers a larger present population.

The numerator and denominator have now become statistically and causally coupled through the same observation system.

This can dampen or amplify the rate in ways that a conventional formula does not reveal.

The solution is not necessarily to avoid multisource estimation.

It is to document dependencies explicitly.

#A minimum denominator provenance record

If a public-health indicator uses a model-derived denominator, the publication should preserve enough information to reconstruct what the population estimate represented.

A minimum record could include:

• target-population definition;
• geographic and temporal scope;
• data sources;
• model or estimation method;
• model version;
• update date;
• treatment of migration and temporary presence;
• eligibility rules;
• uncertainty interval or uncertainty category;
• validation data;
• known population groups likely to be undercounted or overcounted.

For high-consequence indicators, this information should be linked to the published rate rather than stored only in technical documentation.

#Static and dynamic denominators answer different questions

A resident population can be the correct denominator for one question.

A present population may be better for another.

For example:

• long-term disease burden may appropriately use resident-population estimates;
• short-term emergency demand may be influenced by people physically present;
• vaccination programme responsibility may follow legal eligibility or residency;
• outbreak attack rates may require the actual exposed population.

There is no universally superior denominator.

The denominator should follow the estimand.

AI can help estimate a complicated population, but it cannot decide which population the question should be about.

#WHO's own coverage methods illustrate the principle

WHO and UNICEF do not treat one administrative coverage fraction as unquestioned truth.

WUENIC integrates country-reported administrative data, survey evidence and contextual information to estimate the most likely coverage level.^[3]

That process is not the same as the AI-native denominator proposed here, but it demonstrates the broader principle: important population-health indicators can already be estimates built from multiple evidence streams rather than direct counts.

WHO's 2026 guidance on modelling for immunization decision-making also emphasizes appraisal of model assumptions and outputs rather than treating model estimates as self-validating.^[4]

A future AI-enabled denominator should meet at least the same standard of interpretability.

#Equity can be hidden inside denominator quality

Denominator uncertainty is rarely evenly distributed.

Population groups with weaker administrative visibility may include:

• recent migrants;
• people without stable housing;
• highly mobile workers;
• people using services outside their area of residence;
• populations with incomplete registration;
• people living in rapidly changing informal settlements.

A model may improve average accuracy while remaining systematically less certain for precisely the populations public health most needs to see.

A denominator model should therefore report subgroup and geographic uncertainty, not only one global performance statistic.

Otherwise AI can create a more sophisticated denominator while preserving an old invisibility.

#Validation should target the rate, not only the denominator model

A denominator model can have low average population error and still create important bias in a specific indicator.

Validation should ask:

1. How accurate is the population estimate?
2. Where and when is it inaccurate?
3. How does that error propagate into the public-health rate?
4. Would programme decisions change under plausible denominator alternatives?

The fourth question is crucial.

The purpose of a denominator is not to win a population-estimation benchmark.

It is to support a public-health inference or decision.

#A practical reporting pattern

For important model-dependent indicators, a compact reporting format could look like:

Vaccination coverage: 83.4%
Numerator: 8,340 vaccinated eligible people
Denominator: 10,000 estimated eligible residents
Denominator model: PopulationNow v2.2
Model update: 2026-10-01
Plausible denominator interval: 9,500–10,700
Coverage under denominator interval: 78.0%–87.8%
Primary uncertainty: recent migration and temporary residence

The specific format is a proposal, not an established standard.

The principle is that the denominator should no longer disappear behind the percentage.

#The deeper shift

Public health has long used modeled estimates.

AI changes the scale, frequency and granularity at which such estimates can be produced.

That creates an opportunity to replace stale denominators with populations that better reflect reality.

But it also risks turning familiar indicators into opaque composites.

The correct response is not to reject model-derived denominators.

It is to make the model part of the indicator's provenance.

#Limitations and uncertainty

Dynamic denominators are not inherently AI-based, and the strongest empirical evidence cited here predates current generative AI.^[2] This article extrapolates from established denominator problems and contemporary modelling capacity to propose a framework for future model-dependent indicators. The proposal has not been validated as a reporting standard. Population estimation methods should be selected for the jurisdiction, data environment and decision at hand, and privacy constraints may appropriately limit the use of mobility or other granular data. WHO immunization guidance is used here to illustrate denominator and model-appraisal principles; it should not be generalized mechanically to every public-health indicator.^[1][4]