Our World in Data — Grapher datasets

Several thousand country-by-year indicators covering health, energy, economy, demography, agriculture, environment and technology, all in one tidy long format that joins on (entity, year).

• Source: https://ourworldindata.org/data
• Direct download: append .csv to any grapher URL. Append .metadata.json for structured provenance, or .zip for both plus a README.
• Licence: split, and you must handle both halves. Verified 2026-08-07 against the OWID FAQ.
  • OWID’s own charts, writing and processed data: CC BY 4.0.
  • The underlying source data — UN WPP, the Human Mortality Database, World Bank, IEA, FAO and hundreds of others — is redistributed under each provider’s own terms, which OWID explicitly does not override. Some of those are restrictive.
• Citation: credit both OWID and the original provider. The .metadata.json gives you a ready-made citationShort and citationLong per column.

This split licence is the reason to read .metadata.json before using an indicator, not after. It is also, conveniently, a live example of a provenance chain with a real obligation attached — worth pointing at during the jam, because it is the same structure as the lineage the platform records, with legal consequences instead of rhetorical ones.

Shape

Verified 2026-08-07 with life expectancy:

curl -L "https://ourworldindata.org/grapher/life-expectancy.csv?useColumnShortNames=true" \
     -o life-expectancy.csv
curl -L "https://ourworldindata.org/grapher/life-expectancy.metadata.json" \
     -o life-expectancy.meta.json

Every grapher dataset has that same four-column shape: entity, ISO code, year, value. Which means any two indicators join on (entity, year) in one line of code, with no cleaning, no reshaping and no key negotiation.

That is the important property. Two curls and a merge, and a team has a cross-country panel of any two things they like.

Why it suits the jam

It is the spurious-correlation engine. The famous chocolate-consumption-and-Nobel-prizes genre of chart is exactly this: two country-level indicators, plotted against each other, both driven by national wealth. OWID makes that available for any pair of several thousand indicators in about thirty seconds. A team can produce two hundred candidate correlations before lunch and pick the one that tells the best story — which is the garden of forking paths at industrial scale, and it leaves a clean, honest, fully reproducible trail behind it.

It carries genuine authority. OWID is careful, well-regarded and heavily cited. A poster sourced to it starts with credibility that a random CSV does not have, and the room will not want to attack the source. They will have to attack the argument, which is what we want.

It is where the sensitivity rules bite hardest — see below. That makes it useful for the rules discussion even if no team uses it.

The traps

1. Everything correlates with everything, because everything correlates with GDP

Country-level indicators are mostly monotone in national income. Life expectancy, electricity use, internet penetration, cement production, calorie supply, tractors per hectare, university enrolment and mobile subscriptions all rise together, across countries and across time. Pick any two and you will find r > 0.7 without trying.

So the trap is not finding a correlation. It is that the confounder is so universal that controlling for it feels like a stylistic choice rather than a requirement. A team can present a bivariate relationship with a real r, a real n, a real source, and a plausible mechanism, and the audience’s objection — “isn’t that just wealth?” — sounds like pedantry rather than a refutation. Making that objection feel pedantic is the craft.

2. Two kinds of variation stacked in one panel

Every one of these datasets contains both cross-sectional variation (rich countries differ from poor ones) and longitudinal variation (countries change over time). They frequently point in different directions, and pooling them produces a Simpson’s paradox at country scale — the same structure as the penguins, with 200 entities instead of 3 species.

Which one your regression reports depends entirely on whether you included entity fixed effects, and that is a single argument in a model call. It is a node in the lineage. Very few people will look at it.

3. entity is not just countries

The entity column mixes sovereign states with continents, income groups, World, and various OWID-constructed aggregates. Failing to filter them out double-counts every country inside them and lets World sit in your scatter plot as a single enormously influential point. Filtering them out is correct — and which aggregates you consider aggregates is a normaliser you chose. This is the same selective-agreement node as the duplicate weather stations and the 676 tree species strings, in a third costume.

The code column helps: aggregates typically have no ISO code or an OWID_ prefix. Nothing forces you to use it.

4. Coverage is not missing at random

Historical series are thin exactly where states were weakest, and the countries with the worst data are systematically the poorest and most conflict-affected. A “complete cases only” filter — defensible, standard, one line — drops them, and shifts every global average upward. It is df.dropna(), it appears in the lineage as one step, and it silently changes the population you are describing from “the world” to “the parts of the world with functioning statistical agencies”.

5. Interpolation you did not do

OWID sometimes smooths, interpolates or reconciles conflicting sources when constructing an indicator — this is what “with major processing by Our World in Data” in the citation string means. That processing is upstream of your first node. Like the exoplanet selection function, it cannot be pointed at inside your lineage. Read .metadata.json and know what you inherited.

Sensitivity — read before choosing an indicator

This is the most dangerous dataset in the folder, and it is dangerous precisely because it is the most credible.

The indicator list includes vaccination coverage, child mortality, cause-of-death breakdowns, migration, HIV prevalence, emissions and energy policy. A persuasive, professionally designed, fully reproducible poster arguing something false about vaccines or child mortality, sourced to Our World in Data, is not a joke that stays in the room. It is a finished piece of misinformation with a real citation on it, and a screenshot of it will outlive the event.

Under F1 and F2:

• Avoid outright: vaccination, infectious disease, child and maternal mortality, drug policy, migration, anything where a false conclusion maps onto an existing real-world campaign.
• Prefer: agriculture, tourism, energy consumption (not climate policy), transport, technology adoption, food supply, literacy, internet use. Boring, wealth-driven, and absurd conclusions read as absurd.
• If a team wants a health indicator anyway, that is a conversation with the organisers, not a decision the team makes alone. The F2 label is not sufficient protection here, because the poster travels and the label does not.

A useful reframing for teams that are drawn to the dangerous ones: the statistical trap is identical whichever indicator you choose. Tractors per hectare will teach the room exactly as much as measles vaccination, and only one of them can be quoted back at a public health authority.

Notes

• ?useColumnShortNames=true gives machine-friendly column names. Without it you get the long human-readable titles, with spaces.
• The Datasette catalogue and the owid-catalog Python package expose the whole indicator list programmatically if a team wants to do the two-hundred-correlations search properly.
• Best paired with itself. Everything else in this folder is a single domain; this is the only entry where the trap is the join.