Three institutions look at the same underlying phenomenon—how many people live where, how long they live, and why they move—and produce numbers that do not agree, are not supposed to agree, and answer different questions. A national census counts. The UN Population Division projects. A migration survey asks why. Treating any one of these as “the” population number is the most common error in reporting on demographic change, and it is avoidable once the three approaches are compared on the dimensions that actually distinguish them: what they measure, how current the number can be, and how much uncertainty rides along with it.
This article works through that comparison in the domain where the differences matter most in daily life—fertility and mortality trends, aging populations, migration, housing markets, education pipelines, health systems, and the newer, contested category of climate-linked mobility. The goal is not to crown a winner. Complementary approaches measuring different things is the correct outcome, not a flaw to be resolved.
Three approaches, three questions
Census and vital registration answers “how many people, of what age and sex, were actually enumerated on a given date, and how many births and deaths were registered in the interim.” It is a headcount with legal and administrative weight—apportionment, funding formulas, service planning all key off it. Its unit is the individual record.
UN cohort-component demographic modeling, embodied in the UN Population Division’s World Population Prospects (WPP), answers “given today’s age structure and current trends in fertility, mortality, and migration, what will the population’s age and sex composition look like in five, twenty, or seventy-five years.” The 2024 revision advanced a base population for each of 237 countries and areas from 1 January 1950 forward through single-year intervals, applying age- and sex-specific fertility, mortality, and migration rates in each step—the cohort-component method for projecting population, CCMPP [1]. Its unit is a birth cohort moving through time, not an individual.
Survey-based migration research answers a different question again: not how many people crossed a border, but why, under what conditions, and with what intention—the motives, constraints, and decision structures behind a move. The IOM’s World Migration Report compiles exactly this kind of contextual, often survey-derived material alongside migrant stock counts, covering displacement, remittances, and the social and economic drivers behind mobility [4].
None of the three can substitute for the other two. A census cannot tell you why someone migrated. A cohort-component projection cannot tell you how many people were actually alive last Tuesday—it is a model run forward from an estimated base, not a count. A migration survey cannot give you the fertility rate of a country. They are complementary instruments pointed at different facets of the same underlying reality, and the comparison below is organized around three axes: coverage, timeliness, and uncertainty.
Coverage: who gets counted, and who gets missed
A census aims for a complete enumeration, and the honest way to evaluate one is not “how many people did it count” but “how far off was that count, and for whom.” The United States’ 2020 Census used a Post-Enumeration Survey (PES)—an independent follow-up sample compared against the original count—to estimate coverage error. At the national level, the PES found no statistically significant net coverage error in the total population count, but that null result concealed offsetting errors by group: Black or African American and Hispanic respondents, young children, and renters were undercounted, while non-Hispanic White respondents, adults over 50, and homeowners were overcounted [5]. State-level results diverged even more: six states showed a statistically significant undercount, eight showed an overcount, and the rest showed no detectable difference [5]. Pew’s summary of the same results underscored that a “good” national total is compatible with meaningfully biased subnational and demographic detail [6]. The Government Accountability Office’s review of the same cycle flagged the coverage errors as a direct input into planning for the 2030 Census, including proposed changes to address the recurring undercount of renters and young children [9].
This is coverage error as a fact, directly measured by an independent survey built for that purpose—not a modeling assumption. It is the strongest kind of evidence a demographic method can produce about itself, and it is also the method’s chief limitation: a census only tells you about coverage where you built a matching survey to check it. Most countries do not run a PES-equivalent every cycle, and many of the poorest and most rapidly urbanizing places—precisely where population change is fastest—have the weakest vital registration and the least frequent censuses.
Cohort-component projections inherit whatever coverage gaps exist in their inputs, plus a coverage question of their own kind: which countries have data recent and complete enough to anchor a base population credibly. WPP2024 drew on 1,910 censuses, vital registration data for 169 countries, and demographic indicators from 3,189 surveys [1]—an enormous and heterogeneous input base, precision-weighted rather than uniformly trusted. Where vital registration is thin, the Population Division leans more heavily on survey-derived indirect estimation techniques, which carry wider error bands by construction. The projection layer does not close the coverage gap in the underlying data; it propagates it forward, with the gap showing up as wider uncertainty intervals rather than as a bias correction.
Migration surveys have a coverage problem of a third kind: they can only interview people they can find, and the hardest-to-count migrants—undocumented crossers, people in transit, displaced populations moving informally—are structurally undersampled by the same logic that makes them hard to count in a census. The IOM report’s figure of 281 million international migrants globally is itself a stock estimate built from national administrative and census sources rather than a survey headcount, and the report is explicit that irregular and forcibly displaced populations are the categories where confidence is lowest even as the report also finds displacement estimates reached a record high of roughly 117 million by the end of 2022 [4]. Surveys add depth—why someone left, what would make them return—precisely in the population segment where basic counting is least reliable, which means qualitative richness and quantitative confidence trade off against each other in exactly the group policymakers most need numbers on.
Timeliness: how current is the number when you read it
A census is a snapshot with a long shelf life and a long production lag. It happens once a decade in most countries that run one at all, and the detailed data products—the kind used for school-district planning or transit forecasting—can take one to three years to fully release after enumeration. Vital registration, where it functions well, is far faster: births and deaths can be tabulated within a calendar year, which is why countries with strong civil registration systems can update population estimates annually even between census years.
UN cohort-component projections are updated on a roughly two-year cycle (WPP2022, WPP2024), and each revision is a snapshot forecast, not a live feed—the moment a new fertility or mortality data point becomes available in some country, it sits unincorporated until the next full revision. This is a genuine tradeoff: the biennial cadence buys internal consistency (every country’s projection uses the same vintage of methodology and input data, so comparisons across countries are apples-to-apples) at the cost of currency. A country whose fertility rate moves sharply between revisions will not show that shift in WPP data until the next release cycle.
Migration surveys sit at the opposite end: they can be fielded quickly in response to an emerging situation—a new displacement event, a policy change—and can capture motive and intention data no administrative system collects. But a fast, targeted survey usually trades away the population-representativeness that makes a census or a national vital-registration system comparable across time and place. The IOM report itself functions as a biennial synthesis rather than a real-time feed, pulling together administrative migrant-stock counts, national survey data of varying vintage, and displacement tracking from agencies like UNHCR into a single flagship publication released every two years [4].
Uncertainty: how the three approaches quantify (or don’t) their own error
This is the axis with the sharpest methodological contrast. A census, in principle, is not supposed to carry sampling uncertainty—it’s meant to be a complete count—so its error has to be estimated through a separate, independent instrument (the Post-Enumeration Survey model used in the US and comparable dual-system estimation approaches used in other countries). That is a strength: coverage error is measured, not assumed, wherever a PES-equivalent is run. It is also a limitation: without that independent check, a census’s error is simply unknown, and the practice of running one at all is uneven across countries and cycles.
UN population projections went through a specific methodological transition on this exact point. Long-run projections had historically been reported as a small number of named scenarios—commonly “high,” “medium,” and “low” fertility variants—which implicitly bundled a huge and unstated range of possible futures into three lines. In July 2014, the UN Population Division issued fully probabilistic population projections for all countries for the first time, replacing the deterministic variants with Bayesian hierarchical models of future fertility and life expectancy that generate genuine prediction intervals rather than three arbitrarily labeled paths [3]. This is a case where the method itself became more honest about what it doesn’t know: instead of a single “medium variant” line implying false precision, the current approach reports a full distribution, and the width of that distribution is itself information—it tends to be narrow for near-term totals in large, low-mortality countries and very wide for long-horizon fertility in countries with volatile recent trends.
Migration research has the least standardized uncertainty language of the three. Some of the sharpest numbers in circulation—remittance flows, migrant stock totals—are administrative tallies with implicit rather than stated confidence intervals (the IOM report’s finding that international remittances rose from 128billion to 831 billion dollars between 2000 and 2022 is drawn from balance-of-payments reporting, not a sampled survey [4]), while other figures, particularly on irregular or forced migration, are explicitly modeled estimates with wide and sometimes only qualitatively described uncertainty. The result is a literature where a reader has to check, source by source, whether a migration number is a counted total, an administrative estimate, or a modeled projection—the three are frequently presented in the same paragraph without being distinguished.
What this means for aging, housing, education, and health planning
Here the three approaches feed different, complementary parts of the same planning problem, and confusing which approach produced which number causes real errors in each domain.
Aging. The fact that WPP2024 projects a rising global median age and a growing share of the population over 65 is a model output, not a census fact—it depends on assumed future fertility and mortality paths, quantified now with explicit uncertainty bands rather than a single deterministic number [3, 1]. The current count of people over 65 in a given country, by contrast, is a census/vital-registration fact, current only as of the last enumeration or registration update. Pension-system planning needs both: the current count to size present obligations, the projection (with its uncertainty band) to size future ones. Reporting a UN median-variant figure as if it were a measured count is the single most common conflation in coverage of population aging.
Housing. Local housing demand forecasting depends on household formation rates, which are themselves derived from age-structure projections crossed with migration assumptions—two of the three approaches stacked on top of each other, compounding uncertainty at each step. A housing agency that treats a twenty-year household-formation forecast as a fact rather than a scenario conditioned on migration assumptions is importing model uncertainty without labeling it as such.
Education. School-capacity planning is one of the few places where the census/vital-registration approach dominates almost to the exclusion of the other two, because the relevant horizon (five to twelve years, following an already-born cohort through the school system) is short enough that current birth registration data is a better guide than a long-run demographic model, and the population in question (a specific, geographically fixed cohort) is exactly what a census is built to count well.
Health systems. Health planning needs both the mortality and morbidity detail that vital registration and specialized health surveys provide and the age-structure projections that determine future disease burden by age—cardiovascular and dementia-linked care, for instance, scale with the projected size of older cohorts, which is a modeled figure, not a counted one.
Climate-linked mobility. This is the domain where the distinctness of the three approaches, and the risk of blurring them, is starkest. The World Bank’s Groundswell modeling exercise projected that climate change could produce up to 216 million internal migrants across six regions by 2050 under a pessimistic, high-emissions, low-development scenario, with Sub-Saharan Africa accounting for the largest share at up to 86 million [7]. This number is neither a census count nor a UN cohort-component demographic projection: it is a scenario output from an integrated assessment model that combines climate, socioeconomic, and gravity-style migration modeling, built around explicit, stated assumptions about future emissions and development pathways. The same report states that immediate, concerted emissions reduction and inclusive development could cut the projected number by as much as 80 percent [7]—which is the report’s own way of flagging that 216 million is a conditional scenario, not a forecast in the sense the UN population projections use the term, and specifically not something anyone has counted or will count as such, because “climate migrant” is not a legal or administrative category that any census or vital-registration system enumerates directly.
Separating fact, claim, analysis, scenario, and prediction
To make the distinctions above concrete:
Fact, directly measured. The US 2020 Census PES found no statistically significant national net coverage error but significant undercounts of Black, Hispanic, young-child, and renter populations, and significant state-level over- and undercounts in fourteen states combined [5, 6]. WPP2024 was built from 1,910 censuses, vital registration for 169 countries, and 3,189 surveys [1]. IOM’s count of 281 million international migrants and roughly 117 million displaced people by end-2022 are administrative/institutional tallies [4].
Institutional/vendor-style claim, attributed as such. The IOM’s characterization of migration as “part of the solution” to global challenges is the report’s own framing choice, not a measured finding; it should be read as the institution’s stated position rather than a neutral fact.
Analysis. The transition from deterministic “high/medium/low” variants to Bayesian probabilistic projections in 2014 was a genuine methodological improvement in how uncertainty is communicated, not merely a technical footnote—the earlier three-line convention systematically understated the plausible range of outcomes by presenting only three named paths, and there is no serious methodological dispute that the probabilistic replacement is an improvement, though experts continue to debate the choice of priors and historical training windows the Bayesian hierarchical models use [3].
Scenario, explicitly conditional. The Groundswell 216 million figure is a scenario under stated high-emissions, low-development assumptions, with an explicit alternative pathway (aggressive mitigation plus inclusive development) that the same model reduces by roughly 80 percent [7]. It is not a prediction of what will happen; it is a bracketing exercise showing how much the outcome depends on policy choices not yet made.
Prediction, with horizon and disconfirmation condition. If global median age continues rising roughly on the trajectory embedded in WPP2024’s medium-fertility path through 2050, that is a conditional prediction whose disconfirmation condition is observable well before 2050: a sustained upward break in fertility rates in the large countries driving the global median (rather than continued decline or stabilization at current sub-replacement levels) within the next one to two WPP revision cycles would be the specific signal that the medium-variant path is wrong, and analysts revise expectations at each biennial WPP release precisely because such breaks do sometimes occur regionally even when the global aggregate holds steady.
Where the approaches disagree without either being wrong
The most instructive disagreements are not measurement errors but definitional ones. A country’s vital-registration birth count and the WPP’s assumed future fertility rate for that same country are not in tension when they diverge for a given year—the first is a fact about the past, the second is a modeled assumption about the future, and a model is allowed to assume something other than last year’s exact rate if there’s reason (a policy change, a structural trend) to expect a shift. Similarly, an IOM migrant-stock count and a national census’s count of foreign-born residents can differ legitimately because they use different reference definitions—de jure residence versus a broader migrant-stock concept that includes people in transit or awaiting status determination.
A practical rule follows from this: when a demographic figure appears in a report or a news story, ask which of the three production processes generated it before asking whether it is “accurate.” A number described as a “count” or “enumeration” should trace back to a census or a vital-registration system, and its error, if quantified at all, should come from an independent check like a Post-Enumeration Survey rather than from the number’s own internal consistency. A number described as a “projection” or attached to a future year should trace back to a cohort-component model, and it should carry an explicit uncertainty interval or scenario label — a bare single figure for 2050 with no stated range is a sign the original probabilistic output has been flattened somewhere in transmission. A number attached to motive, intention, or a category like “climate migrant” that has no administrative registration process behind it should trace back to a survey or a modeled scenario, and it should be reported with the same scenario language the source used, not upgraded into a forecast the source never claimed to make.
This also clarifies where genuine expert disagreement lives, as distinct from mere numerical mismatch. Demographers disagree, sometimes sharply, about which fertility model best extrapolates a country’s future path, about how much weight to give recent survey data versus longer historical vital-registration series, and about how the Bayesian hierarchical models underlying the UN’s probabilistic projections should treat countries with unusually volatile recent history — Bongaarts and others have written at length about the cohort-component method’s known sensitivities to these choices, describing it as a workhorse technique whose outputs are only as good as the fertility and mortality assumptions fed into it [8]. That is a live methodological dispute among people using the same approach. It is a different thing entirely from the surface-level disagreement between a census total and a WPP projection for the same country and year, which is not a dispute at all — it is two different questions, correctly given two different answers.
The discipline this article has tried to model throughout—coverage, timeliness, uncertainty, each stated on its own terms—is the same discipline a reader needs when a number crosses from one of these three literatures into a news story or a policy brief. A census count, a UN projection, and a migration-survey estimate are not three attempts at the same number that disagree; they are three different numbers, built for three different purposes, and the comparison that matters is not which one is right but which one answers the question actually being asked.
