Who arrives
is changing.

Fewer children, more prime-age and tertiary-educated arrivals

Across 179 destinations with comparable composition estimates and usable flow weights at both endpoints, the tertiary proxy rose from 12.5% to 20.5% of all arrivals. The share aged 25–54 increased from 42.7% to 47.4%. The child share fell from 25.4% to 20.3%, while the share aged 55 or older increased from 8.0% to 9.9%.

These are flow-weighted estimates for a balanced set of destinations, covering about 97% of the independent estimated world flow in each endpoint period. They are not simple averages of national proportions. Larger destinations receive larger weights, using a separate total-flow series.

The estimated female share declined from 49.0% to 47.2%. That result comes from the external sex-specific flow estimates. It cannot be calculated from the composition CSV alone, because every female profile and every male profile separately sums to one.

The terminology also matters. “Prime working age” identifies ages 25–54; it does not establish employment. The 55+ group includes people who may still work. “Tertiary” is shorthand here for the source’s post-secondary category at ages 20+, mapped in the underlying work from completed university education. It is a proxy, not a freshly harmonized OECD qualification measure. Among arrivals aged 25+, the corresponding share rose from 21.1% to 30.5%.

The educational direction is shared; the age pattern is less uniform

Canada’s tertiary share increased by 13.7 percentage points, Germany’s by 7.7 points and Japan’s by 10.8 points. Yet Canada’s share aged 25–54 changed by only 0.5 points, Germany’s rose by 1.8 points and Japan’s fell by 2.8 points. “More educated immigration” does not imply a single age trajectory.

The pyramids make those differences visible. Each side is normalized within sex, allowing the shapes of women’s and men’s arriving age distributions to be compared without confusing composition with the number of arrivals. The outline shows 1990–95 and the solid bars show the selected endpoint. These are immigration profiles, not pyramids of the entire foreign-born resident population.

The two regional panels pool country profiles with independent flow weights. The Gulf panel covers the six GCC states. The Southern European panel covers Italy, Spain, Greece and Portugal; it is deliberately not presented as the whole statistical region. Select any other destination in the atlas to append its own pyramid, including individual Gulf and Southern European states. All pyramids in that view share a common horizontal scale.

Large destinations differ in more than one dimension

A country can become more educated, younger in one part of its age distribution and older in another. The heatmap therefore preserves separate measures. It orders the 30 largest destinations by independently estimated immigration in 2015–20, then shows how their composition changed from 1990–95.

The ordering does not use the composition proportions. A country with a 30% tertiary share is not necessarily a larger immigration destination than a country with a 10% share. Sudan’s missing baseline remains blank: a geographical mismatch between current and former states is not repaired by inventing an early-period composition.

At the default threshold, 87 destinations had at least 100,000 estimated arrivals in both endpoint periods. Smaller destinations remain in the data, but do not enter this default ranking. The controls allow thresholds of 50,000, 250,000 and 500,000. This is a scale safeguard, not a confidence interval: a large modeled flow can still have an uncertain demographic composition.

Read a Sankey as a partition, not a life history

The first Sankey organizes one destination’s arrivals by sex, then by age, then by education. Its ribbons conserve the estimated share of that single cohort. Children form their own structural education category; they are not coded as adults with no schooling. The source label “No Education” refers to less than completed primary education, which is more precise than taking the label literally.

A ribbon running through two columns does not establish that the same identified people changed educational attainment. It is an aggregate accounting view. The dataset also does not supply an origin-by-age-by-education breakdown, so no destination profile has been assigned to individual migration corridors.

Two circular views answer a different question

The circular Sankey separates departure and arrival roles for six broad regions. A region may appear on both sides because it both sends and receives migrants. Each ribbon conserves the independently estimated number of moves from an origin region to a destination region. The two copies are not a claim that regional immigration equals regional emigration.

The directed chord chart retains the source’s 11 regions. It exposes reciprocity and connections within regions, using arrowheads to identify destinations. Internal regional ribbons are still international: they aggregate moves between countries in the same region. The smaller number of sectors makes the main connections readable; the exact matrix remains available for values hidden by overlapping ribbons.

A modest volume change can accompany a large composition change

Portugal’s estimated immigration total fell by 1.7% between the endpoint periods. Its tertiary proxy nevertheless rose by 8.3 percentage points. In the United States, the corresponding changes were +9.9% and +11.2 points. These two quantities have different denominators: the first measures a change in the number of arrivals, the second a change in their composition.

A country can experience little change in total immigration while undergoing a profound change in who arrives.

The scatterplot makes the distinction explicit. Its horizontal axis uses a logarithmic ratio so that proportional increases and decreases can be compared without allowing a few fast-growing flows to compress the rest. Its vertical axis remains in ordinary percentage points. It is descriptive evidence; it does not isolate immigration policy, destination demand or changing educational attainment in origin countries.

More education can reduce educational diversity

The normalized age-diversity index for the pooled balanced sample increased from 0.916 to 0.932. Educational diversity among arrivals aged 25+ fell from 0.983 to 0.948, even as the tertiary share rose. There is no contradiction. Entropy measures the evenness of category shares. If arrivals become more concentrated in secondary and tertiary categories, their educational distribution can become less even while average attainment rises.

The index runs from zero, when everyone occupies one category, to one, when the categories have equal shares. Education uses four attainment groups after restricting the population to age 25+. Age uses all 16 published groups, including the open-ended 75+ category. The level therefore depends on those bins. It should not be interpreted as diversity of origin, a measure of integration or an evaluation of migrants’ contributions.

A multivariate view preserves the composition constraint

Age and education shares cannot vary independently: together they exhaust an arriving population. For the optional multivariate view, the analysis retains 48 cells formed by 12 adult age groups, starting at age 20, and four education categories. Each adult composition is closed to one and transformed using the centered log ratio: the logarithm of each share relative to the geometric mean of the 48 shares.

Principal-component analysis then summarizes the transformed profiles, giving every country-period the same weight. The first two components account for 81.9% and 7.8% of the variance. No artificial positive constant is needed: all retained adult cells are strictly positive after the negligible floating-point corrections elsewhere in the source.

Distances in this view describe relative age–education compositions. They do not measure migration volume, a policy score or progress toward a desirable destination. The component loadings and all scores are supplied with the data so that the statistical summary can be inspected rather than accepted as a label.

Most of the global education shift happens within destinations

A rising global tertiary share could reflect changing profiles within countries, a shift of migration toward more highly educated destinations, or both. An exact symmetric decomposition separates these contributions. It attributes each destination’s change in share using its average endpoint flow weight, and each change in destination weight using its average endpoint share.

Of the 8.05-percentage-point increase in the pooled tertiary proxy, 6.65 points come from changes within destinations and 1.40 points from changes in their relative immigration weights. The within-destination component accounts for about 83% of the increase. This is an accounting result. It does not establish the causal mechanism behind either term.

A complete grid of predictions is not a complete grid of observations

The source combines census information with a model designed to estimate countries and periods where direct information is sparse. The published supplement lists 143 census samples across 74 countries. The atlas distinguishes destinations represented in that list from destinations without a listed local sample. All displayed composition estimates remain modeled values, including those for census contributors.

The United States has five listed samples and Canada two. Germany, Japan and the six GCC states have none in this list. That is a meaningful warning about evidence coverage, but it is not a calibrated uncertainty measure. Country-level sample presence also does not guarantee direct observations in either endpoint period.

The source CSV contains 123,090 rows. Every present country-period-sex group has the expected 55 age–education cells, and its shares sum to one within a maximum error of 3.11 × 10⁻¹⁵. There are no missing values inside existing rows and no duplicate keys. The 274 tiny negative entries are floating-point residues, no lower than −7.14 × 10⁻¹⁷; clipping them to zero and renormalizing changes no substantive result.

Missing groups still matter. Fifteen destinations appear only in the final period, leaving 150 missing country-period-sex groups, equivalent to 8,250 missing joint cells. Five further destinations lack a usable matched early flow total under consistent geographical identifiers. These cases are excluded from the balanced comparison, not assigned zero composition.

The source includes a very small amount of post-secondary mass at ages 15–19. The main tertiary proxy excludes it to follow the paper’s age-20 interpretation. Its largest country-period effect is 0.0343 percentage points. The partition Sankey preserves the published categories, while the exact data retain both original and adjusted shares.

The late-period direction is steadier than the country ordering

Dropping 2015–20 and instead ending in 2010–15 leaves the sign of the tertiary-share change unchanged in 97.7% of the 87 default eligible destinations. The correlation between the two absolute-change rankings is 0.922. For the 55+ share, direction agreement is lower, at 81.6%, and the ranking correlation is 0.784. These are sensitivity checks, not validations against ground truth.

The exercise also substitutes a demographic-accounting minimum-flow sex series for the preferred pseudo-Bayesian sex weights. The largest effect on the measured tertiary-share change is 1.32 percentage points among eligible destinations. The female-share results are much more sensitive because the entire measure comes from those sex weights. A precise female-share league table would overstate what this exercise establishes.

External statistics show a specific model weakness

Eurostat’s recorded immigration composition offers a useful, incomplete benchmark. Pooling annual 2015–19 observations provides five years of immigration by age and sex for Germany, Italy, Spain and Denmark. Germany, Italy and Spain use completed age; Denmark’s available series uses age reached during the year. These definitions and the relationship between annual recorded moves and five-year modeled transitions prevent an exact like-for-like validation.

Even with that qualification, the discrepancies are informative. For women arriving in Germany, the modeled 55+ share is 10.2%, against 5.1% in the pooled Eurostat series. For women arriving in Denmark, it is 11.5% against 3.8%. The model therefore overstates older arrivals by about 5.0 and 7.7 percentage points in those comparisons. The direction of a broad modeled age shift should not be mistaken for precise knowledge of every destination’s age profile.

The implication is a better question

An immigration total answers how many people are estimated to have arrived. It does not tell us their ages, qualifications or sex composition. Those characteristics can move considerably even where the total changes little.

The defensible reading of this exercise is that the modeled global arriving population became more educated and shifted away from childhood toward adult ages between 1990–95 and 2015–20. The national patterns differ, and their precision depends on uneven source evidence and external flow weights. A serious account of immigration needs to ask both how many arrive and who arrives.

Definitions that travel with the charts

Conditional denominators

For country c, period t and sex s, the published age–education cells sum to one. Combined shares multiply those cells by independent female or male flow fractions. Total-flow magnitudes come from a separately estimated total series.

Independent flow vintage

Abel–Cohen pseudo-Bayesian closed estimates, frozen from the authors’ IIASA explorer repository at commit 8ffca255dc15d149d1fc804963d519d29d47d693. The flow files were last committed on 28 April 2023. The newer Figshare download was inaccessible during this exercise; these are not advertised as the latest UN input vintage.

Global weights and ranks

Global results use 179 balanced destinations. Rankings require positive matched estimates and at least 100,000 arrivals in both endpoints by default. The heatmap uses the independently ranked 2015–20 top 30. Missing baselines remain unavailable.

Diversity and PCA

H = −Σp log(p) / log(K). Age: K=16. Education: K=4 among 25+. PCA uses CLR-transformed 20+ age–education profiles, 48 cells, all 1,074 balanced country-period profiles, equal weights and no extra variance scaling.

The source does not identify the age or education of migrants from each origin to each destination. The world map combines two independently sourced layers, joined only at destination and period. Curved map paths are geographic connectors, not observed travel itineraries. Boundaries are cartographic reference features, not territorial claims.

Downloads contain the full share audit, missing-cell report, country rankings for every measure, flow reconciliation, census coverage, sensitivity summaries, PCA loadings and exact chart tables. The supplied analysis rebuilds results from the published prediction file. It does not refit the source model, whose private census inputs are not redistributed.

Sources

• Yildiz & Abel: immigration composition and R scripts, Zenodo record 20814220, version 2. Exact file: Immigration_composition_by_RCSL_20260520.csv. Retrieved 8 September 2026. CC BY 4.0.

• Yildiz & Abel (2026), source methodology, Scientific Data. DOI: 10.1038/s41597-026-07969-8. Includes the published census-sample supplement.

• Abel & Cohen (2019), Bilateral international migration flow estimates for 200 countries; Abel & Cohen (2022), updated and refined by sex. Independent flow methodology based on demographic accounting and UN migrant-stock inputs.

• Authors’ IIASA migration explorer repository, frozen commit. Complete country matrices extracted once; aggregate region nodes excluded from country totals. Data files last changed 28 April 2023.

• Eurostat migr_imm8: immigration by age and sex. Annual 2015–19; retrieved 8 September 2026. Definition differences retained in the benchmark table.

• Natural Earth, public-domain cartographic geometry. D3.js 7.9.0, ISC licence, embedded for offline use. Inter font under the SIL Open Font License.