Same in other sources: an arc joins two indicators from different catalogs whose titles and descriptions are almost identical, for example the same quantity published by two institutions. Tick two or more catalogs to see them. The slider sets how similar they must be (0.90 or more: nearly the same text).
Shared words: an arc joins two indicators whose titles share rare words (not “rate”, “total”…). A brighter arc means more shared words. These arcs can be long: two indicators can share a word and still sit far apart on the map.
The overall brightness goes down automatically when there are many arcs. While a search is active, the arcs instead link the matching indicators to each other.
Search matches whole words (and their endings): “oil” finds “oil”, “oils”, not “soil”.
Thousands of indicators describe how societies function: health, education, income, democracy, energy, climate. They are produced by many institutions, each with its own vocabulary and categories, and are scattered across as many databases. This map places them in a single three-dimensional space so that the structure of the available data becomes visible: which domains are densely documented, which sources measure the same things, and where the gaps are.
How to read it
Each point is an indicator. Its size reflects the declared coverage: number of countries times number of years.
Position reflects meaning. Indicators whose titles and descriptions are about similar things sit close together.
Colour can show the source catalog, the category each catalog gives to its indicators, or the domain (health, energy, education, labour…). Domains use one list for every source, so the same topic measured by different institutions gets the same colour; categories are the sources’ own labels and differ from one catalog to another.
Arcs (see Links) join indicators that are the same thing in different sources, or that share rare words in their titles.
The indicator card links to the producer’s page (click the title) and lists the same indicator in other sources when there is one.
Search a word, such as “dollar” or “gini”, to light up every matching indicator. Click a point for its source, unit, coverage and description.
The three maps
The Map buttons change how the positions are computed. The indicators are the same; only the layout differs.
By meaning. A language model reads the title and description of each indicator and turns them into numbers. Indicators that mean similar things end up close, even when they use different words (for example “life expectancy” and “mortality”).
By words. No AI model: indicators are placed according to the words their titles and descriptions have in common, rare words counting more. It is simple and transparent, but two different words for the same idea are not related. It tends to group indicators by source, because a source often reuses the same phrasing.
By domain. Each indicator is first assigned to a domain (health, energy, education…) shared by all sources, and the positions are then computed so that each domain forms its own group, after removing the typical style of each source. It shows best which topics are covered and by whom. The slider Space between domains pulls the groups apart or brings them closer.
Sources
Our World in Data, the World Bank (WDI), the OECD, the ILO, Eurostat, the IMF, the European Central Bank, UNICEF, the UN Sustainable Development Goals database, Google Data Commons, the World Inequality Database, V-Dem, the Comparative Political Data Set, the Quality of Government dataset and the Penn World Table. Each indicator keeps the name of its original source.
Method and limits
Titles and descriptions are converted into numerical vectors with a sentence-embedding model (or, for “By words”, a word-frequency method), and the three-dimensional positions are computed with a dimension-reduction method (UMAP). For “By domain”, the typical writing style of each source is subtracted first, so that topics rather than sources drive proximity, and the layout is guided by the domain. Domains are assigned automatically by comparing each indicator with a short description of each domain, then smoothed using the indicator’s original category.
This is an exploratory aid, not a reference classification: indicators with ambiguous or code-only titles can be misplaced or left unclassified, and coverage is the declared one, not the actual number of observations.