The About page says what Unequal World is and why it exists. This page says exactly how each value is produced, when two numbers may be compared, and where the method breaks. The rule throughout: measured data is shown as-is and source-linked, anything we derive is labelled derived, and known artifacts are flagged or suppressed rather than presented as findings.
Unequal World combines three kinds of object: (a) country indicators served from the World Bank's Data360 / WDI and its federated IMF and UNICEF series; (b) within-city surfaces we built ourselves from national census and statistics offices; and (c) satellite-derived environmental layers. It is an atlas for seeing inequality and verifying its national figures, not a statistical model of its causes.
It is not: a causal model (cross-indicator relationships shown in the country panel are correlational and labelled as such); a nowcast (every value carries the source vintage, and several within-city surfaces rest on the last available census); a clinical instrument (the Brain age / exposome layer is a directional population-level index, not an individual prediction); or a single-vintage, single-method dataset that can be compared cell-for-cell across all cities (see §4).
Every city carries one of three tiers, shown on its card and in its source line. The tier governs whether a value may be read as a number or only as a pattern.
Only the 25 cities measured in a real currency income (US ACS dollars, IBGE reais, StatsSA rand, CSO/Ireland euro) at a fine spatial unit are flagged comparable: true and used in any cross-city ratio. A literacy rate and a household income are never placed on the same axis. Everything below the currency-income line is shown for its own internal gradient only.
Comparable here means narrow: measured in a real currency at fine resolution. It does not mean the cities measure the same construct. Each country defines household income differently (see §4), so even within this set a cross-city number is a contrast, not a like-for-like league table.
The comparable set is built from primary census micro-geography. We download the official small-area table, join median household income to the matching boundary, and render the choropleth at the finest unit the source publishes. Representative sources:
| Region / cities | Source | Spatial unit | Vintage |
|---|---|---|---|
| US (Detroit, Chicago, NYC, LA, +8) · 12 cities | US Census Bureau ACS 5-year (B19013, continuous median) | Block group | 2022 |
| Brazil (São Paulo, Rio, Brasília, +3) · 6 cities | IBGE Censo Demográfico, setores censitários | Census sector | 2022 |
| South Africa (Cape Town, Joburg, Durban, +3) · 6 cities | StatsSA Census, banded household income, via DataFirst (UCT) | Small Area Layer | 2011 |
| Ireland (Dublin) · 1 city | CSO Geographical Profiles of Income, table GPIIA01 (Revenue + DSP administrative income) | Small area | 2022 |
The income construct is not identical across countries. US ACS is continuous pre-tax money income; the Irish figure is modelled from Revenue/DSP administrative records; the South African 2011 figure is collected in income bands, so the rand values are bracket midpoints (imputations), not measured amounts. A banded measure compresses the extremes, which directly affects any ratio drawn from it.
The within-city "Gini" is a spatial, between-area statistic. Where the platform shows a city Gini, it is computed over area medians, so it captures inequality between neighbourhoods and discards inequality within each one. It is therefore systematically lower than, and not comparable to, the national household Gini shown in the country panel. The two are different measurements that happen to share a name.
Cape Town illustrates the resolution: 5,246 Small Area Layers carrying median annual household income across a steep gradient (roughly two orders of magnitude, ratio ~16×). Because the 2011 income is banded, that ratio is sensitive to how the open-ended top bracket is imputed, so it is best read as "a very large gap," not a precise multiple. The gradient, not a single city number, is the unit of analysis.
Two well-sourced numbers can still be incomparable. Three systematic effects are disclosed rather than hidden:
The city "development burden" is the one figure on the platform we compute ourselves. It is presented as a stacked breakdown by domain, used to show which domain weighs on a place. The four domains are:
Each domain is normalized to a common scale before weighting. When a domain is missing, its weight is redistributed across the present domains, and the card states whether the underlying data is weak, medium or strong. Domains that would rest only on placeholder values are suppressed, not displayed.
It is not a measurement, not an official index, and not the output of the Legaz et al. (2026, Nature Medicine) brain-aging model. The weights are an editorial judgement, informed by that study's direction of effect but chosen by us, and the picture would shift under different defensible weights.
Read the breakdown, not a rank. Because the weights are re-normalized whenever a domain is missing, two cities with different domain coverage are scored on different effective weight vectors. The composite total is therefore not a cross-city ranking; it is a within-city statement of which domain dominates. We deliberately do not publish a single 0 to 100 league table from it.
Environmental layers come from ESA Sentinel-2 (vegetation / NDVI), Sentinel-5P (NO₂), NASA VIIRS (night lights) and the EU JRC GHSL (built-up and population). Satellite estimates are convenient and global but carry systematic biases we correct or flag:
Every national value is served from the World Bank's published data and carries a one-click "Verify" chip back to the authoritative series (Data360 / WDI); the panel also pulls federated IMF (World Economic Outlook) and UNICEF values. The country panel's peer-comparison and divergence features are deterministic statistics, not machine learning: a country is flagged where its value departs from the robust median of its income-level and region peers by more than a fixed multiple of the median absolute deviation, and the peer sample size is disclosed. Inflection markers flag where a series measurably bent (a rolling-slope change beyond a fixed threshold); 187 carry a hand-researched, source-linked candidate explanation. All such relationships are correlational, not causal.
The scrollytelling stories are reported pieces built on a named data spine, not illustrations. Each names its source in-story:
The Health tab's Brain age layer is our platform-derived estimate of how a place's environment (air, water, green space, inequality, infrastructure) is associated with accelerated brain aging. It is grounded in the science of Legaz et al. (2026, Nature Medicine) and developed as a research collaboration with the Global Brain Health Institute at Trinity College Dublin, with support from the Max Planck Institute for Human Development and other partners. This is a collaboration, not a formal endorsement, and the layer is our composite, not the study's model output. It is a directional, population-level index and must not be read as a clinical or individual diagnosis.
It describes environmental exposure, not residents. The layer measures the pressure of a place's surroundings; it makes no claim about the brains, capacities or worth of the people who live there. It must never be read as "this neighbourhood's residents have older brains." We are alert to the risk that a "brain aging" frame laid over already-marginalised neighbourhoods can stigmatise them, so the layer is framed as exposure, not outcome, carries no neighbourhood-level individual claims, and is suppressed where the underlying environmental data is weak.
Reproducibility differs by layer, and we are precise about it. The country layer is one-click verifiable by anyone via the verify chips back to the World Bank's published series. The within-city surfaces are reproducible from their primary sources via the pipelines (download, join, render) in the repository, with per-city provenance machine-readable in cities-viz.json and a validation script that checks tier, comparability and source completeness before each build. The derived layers (the burden composite, Brain age, GHSL corrections and ground overrides) are inspectable, their inputs and code are in the repo, but they are not one-click reproducible; they encode editorial choices, which is exactly why this page documents them.