How the numbers are calculated

What a percentile means here

A percentile is the share of a comparison group that sits below your value. "Taller than 84% of Norwegian men" means: in my estimate of the height distribution of adult men in Norway, 84% of them are shorter than you and 16% are taller.

The comparison group matters more than the number. Each result names its group (your country and gender where the data allows, otherwise the world). 84% among Norwegian men is about 98% among all men worldwide, because Norwegians are tall. Both are true; they answer different questions.

"1 in 8 people" is the same number written the other way: if 87.5% are below you, 12.5% are at or above, and 100 / 12.5 = 8.

The normal-distribution estimate

Most modules do not have a list of eight billion measurements. They have two numbers from a published study: a mean and a standard deviation (SD) for a group. From those two numbers I assume the trait follows a normal distribution, the bell curve, and compute:

1. How many SDs your value is from the mean: (your value minus the mean) divided by the SD. This is the z-score.
2. The share of a normal curve that lies below that z-score. This is the percentile.

This works well for traits that really are roughly bell-shaped: height, foot length, reaction time within an age band, resting heart rate, digit span. IQ fits perfectly because it is defined that way (mean 100, SD 15).

It works badly for skewed traits. Income is the clearest case: it is log-normal with a long right tail, so a bell curve would say nobody earns ten times the median. For income I use World Bank percentile breakpoints directly and interpolate between them. Carbon footprint, screen time and savings rate are also right-skewed and bounded at zero; there I still use a normal curve, so the extreme ends are less reliable. Those pages say so.

Some traits are not curves at all. Eye colour and handedness are categories, so the page reports the share of people in your category and nothing more. The following modules use a table or category lookup instead of a curve: Name Popularity, Income & Wealth, Eye Color, Handedness, Birth Month, You in History, Language & Vocabulary, Hearing Range.

Measured versus self-reported

Six modules measure something in your browser: Reaction Time, Typing Speed, Reading Speed, Voice & Speaking Pace, Hearing Range, Working Memory. The input is what the browser recorded, with the usual caveats about hardware. A 60 Hz screen adds up to 16 ms to a reaction time; cheap earphones cut off well below 20 kHz.

The rest take what you type: Height, Age, Body Weight & BMI, Penis Size, Name Popularity, Shoe Size, Income & Wealth, Eye Color, Handedness, Birth Month, Sleep Duration, Carbon Footprint, Screen Time, Social Connections, Political Position, Food & Diet, IQ & Cognitive Score, You in History, Financial Trajectory, Sleep Quality, Language & Vocabulary, Resting Heart Rate, Introversion / Extraversion, Risk Appetite. Self-report has known biases. In surveys, men overstate their height by 1 to 2 cm on average, self-reported IQ runs 10 to 15 points high, and people under-report screen time and over-report sleep. The reference data for those modules is mostly measured (NCD-RisC heights come from measurement surveys, not questionnaires), so an honest answer gives an honest percentile and a generous one gives a generous percentile.

The questionnaire modules (diet, risk appetite, introversion, sleep quality, political position) use short scales I adapted from published instruments. They are shorter than the originals, and the means and SDs come from studies of the full instruments. Treat those percentiles as approximate.

Country coverage

The country picker offers 40 countries. The by-country pages cover more:

• Height: 186 countries plus a global figure (NCD-RisC, adults born in 1996)
• BMI: 186 countries plus a global figure (NCD-RisC, 2016)
• Life expectancy: 200 countries plus a global figure (UN WPP 2024, estimates for 2023)
• Median age: 30 countries plus a global figure
• Sleep duration and screen time: 20 countries each
• Everything else: one global distribution

When your country is missing from a dataset, the module falls back to the global figure and the result says "globally" instead of naming the country. Country figures also come from different years; the source line on each module gives the year. NCD-RisC does not publish standard deviations. SD for countries added in 2026-10 uses the global SD.

Rounding, and why you never see 100%

Percentiles are shown as whole numbers. Anything that would round to 100 is shown as >99%, and anything that would round to 0 as <1%. A normal curve never reaches 100%, and no finite sample can support the claim that nobody on Earth is taller than you. "1 in X" is rounded to the nearest whole number, and internally the percentile is clamped to 0.01 to 99.99 before that division so the number stays finite.

The rarity label on your card (Common, Unusual, Rare and so on) is the average distance from the 50th percentile across your three most extreme traits. It is a descriptive score from 0 to 49, not a probability. I deliberately do not multiply percentiles together: traits are correlated, so "one in a billion" claims from multiplying them are wrong.

Sources

Each module page links its own source. This is the complete list. The "kind" column is there so you can weigh them: a journal paper with clinician measurements is not the same as a benchmark site aggregating whoever turned up.

Known limitations

• The bell-curve assumption is an approximation everywhere and a poor one at the tails of skewed traits.
• Benchmark samples are self-selected. Human Benchmark visitors skew young, on desktop, and trying hard. Sleep Cycle users own a smartphone and care about sleep. Both make "average" look better than the real population.
• Most datasets split people into men and women only, so the site does too for those modules.
• Country means are for adults and from different survey years. Age-specific norms exist only for reaction time and hearing.
• The shortened questionnaires are not validated instruments. They are my adaptations.
• The historical module uses rounded, region-level estimates. "Life expectancy 29 in 1800" hides enormous variation.
• Name data is United States only; names outside that record get a rough estimate and the page says so.

Report an error

If a number is wrong, a source is misattributed, or a page claims more than its data supports, I want to know. Open an issue at github.com/rastomatti/howrare.me/issues or email hello@howrare.me. Include the module and, if you can, the source you think is better.