Average IQ by Country

Everyone took the same free online reasoning test on MyIQtested: 33 multiple-choice puzzles with four options each, 20 of them matrix puzzles based on the ICAR framework, plus 6 series and logic, 4 verbal and 3 numerical. There is no timer. The unit is one person’s first completed attempt, identified by a first-party browser cookie, with all 33 answered, taken from 20 July to 29 September 2026, in 3 to 60 minutes, by someone who reported an age of 18 or over. Retakes never stand in for a first attempt. The test itself is described on the methodology page.

Every query was run against our analytics store in the last days of September 2026, using only what the privacy policy discloses: completion events, scores, timing, approximate country, device, language version and the age band people chose to give. Country comes from the connection, not from anything the person told us.

A country is placed on the list only when the data can pin it down, not when it crosses a round number of takers. Four things must hold. Its adults must cover at least 90% of the age mix of everyone who took the test, so its average can be adjusted for age. Its index must be reliable, meaning at least 80% of what the number says is a real difference between countries rather than noise, judged against the roughly 3-point gap between typical countries. Its 95% range must be no wider than plus or minus 3 points, one test question. And it must give the same answer, within normal drift, in the first and second half of the 72-day window. In practice that takes about 110 to 210 adult first attempts; 76 countries qualify.

The ranges include more than which people happened to take the test. They also allow for how much the mix of takers changes from week to week, so even the biggest countries carry about plus or minus 2 points. Every range is a 95% confidence interval, and every difference we point out holds at 95% confidence after correcting for the number of comparisons (Holm method). Countries that cannot be told apart share a tier, and we publish tiers, not ranks.

Countries left out are marked "not enough data yet". That says nothing about how they would score: with more takers they could land anywhere, so their figures are not shown.

The four guards

• Age coverage. Adults in age bands that make up at least 90% of the standard age mix (a band counts with 5 or more takers), so the average can be adjusted for age.
• Reliability of at least 0.8. Reliability is how much of the number is a real difference between countries rather than noise: τ² / (τ² + SE²), where τ = 2.5 points is how far apart typical countries really are, read from the 35 anchor countries with 500 or more adults.
• A 95% range of no more than ±3 points. That is one test question.
• Stability. The country’s figure in the second half of the window, minus the first, is within what the typical country’s drift (0.95 points per half) and its own sampling error allow.

Adults needed

Reliability is the guard that binds: the ±3-point range alone would need only 85 to 170 adults. The columns are how unevenly the takers spread across age bands (1 is an even spread).

Stability

No placed country failed the half-to-half test; 5 (BO, GP, IR, IS, MT) were too small to test it and rest on the other three guards. The test is Bonferroni-corrected across countries, so a country is not dropped for one unlucky half.

The test counts correct answers and reports 51 plus three points for each one, with a floor of 55 and a ceiling of 150, so scores move in steps of 3. The reference point is fixed rather than estimated from takers, so absolute averages run well above 100 and are never shown here. The ceiling also compresses the top: in the five countries at the top of the list that have the measure, 10% to 16% of takers answer 30 or more of the 33 correctly.

For each country, the figure shown is the difference between its age-standardised mean and the UK sample’s (12,457 adults), in test points: +2.9 means 2.9 points above the UK sample, and the UK sample itself is the reference (0). Three points is one question. Age standardisation re-weights each country to the pooled adult age mix of the window (18 to 24 50.1%, 25 to 34 25.4%, 35 to 44 10.6%, 45 to 54 6.5%, 55 to 64 3.7%, 65 and over 3.6%), so a country of older takers is not compared with a country of students. Measuring from the UK sample follows the convention of the published national tables; it carries no other meaning. The 95% range is the figure plus or minus 1.96 standard errors, and the standard error carries the sampling error and the week-to-week change in who takes the test, which is why no range is narrower than about ±2 even for a country with 200,000 takers. The 10th and 90th percentiles are measured from the same UK reference, and the middle 80% of takers is the distance between them. The CSV gives each figure as gap_vs_uk with its 95% range, and also as index_uk_100, the same number plus 100, so files written against earlier releases keep working.

Countries are sorted by that figure and grouped into tiers. No two countries in a tier differ at 95% confidence once the sample, the week-to-week change in who takes the test and the number of comparisons are all allowed for (Holm method, every pair of countries at once). Each line between two neighbouring tiers is crossed by at least one pair that does differ at that standard. That does not make every country in a tier higher than every country in the next one: near a line, some countries on either side still cannot be told apart. That gives 4 tiers. How they are built: every pair of placed countries is compared with a z test on the difference of their figures, and all pairs are corrected together with the Holm method, which holds the chance of even one false difference anywhere in the list to 5%; neighbouring tiers merge, smallest gap first, while no pair across them is declared different. A tier is not a rank and no country has a rank number.

We show the raw age-standardised difference with its range. Shrinkage is at most 19.6% of the distance to the centre for any placed country and would pull countries that really are far from the rest (the Indonesian-language communities) toward it. An empirical-Bayes estimate is in the CSV (shrunk_index_uk_100, on the index scale) for reference only.

• Everyone under 18. Of first-time takers who gave an age, about 51% were under 18, and none are in any figure. Among the adults who remain, about 86% took the test on a phone.
• Every retake. Only a person’s first completed attempt counts, so practice on the same questions does not move a country’s number.
• Attempts under 3 or over 60 minutes.
• Countries the data cannot pin down. 53 more countries have 30 or more adults but fail a guard. They are “not enough data yet”: no figure, no tier, and never drawn as low, because with more takers they could land anywhere. This page lists the 2 closest to being placed, and the CSV lists all of them with the reason and a blank figure. Countries with fewer than 30 adults are not listed.

Does the order survive other samples?

The order of countries was recomputed under every cut below. A rank correlation of 1.00 would mean nothing moved. Language version and device moved it most, which is why the page makes no claim about fine differences.

Does it match outside data?

Loosely. Countries whose pupils score higher in the World Bank’s harmonised learning data, which combines international school tests for 164 countries, also score higher here. The agreement cannot show what drives it: internet access, for example, may track the same order, though its correlation with our country figures does not reach 95% confidence. Holding internet use constant, the partial correlation with learning outcomes is 0.38 (p = 0.11, 20 countries).

These three rows use the countries with 1,000 or more adults. Across 68 placed countries with a learning score, leaving out Indonesia, Malaysia and the countries where the Indonesian version is common, the rank correlation with learning scores is 0.66. We chose that cut after seeing the data, so treat it as a lead, not a result.

Device, weekdays and school terms

Among adults, people on a computer outscored people on a phone in every country where both groups pass 300 takers (Figure 2). Weekdays against weekends showed nothing: the largest difference in any country was 0.5 points.

Indonesian school holidays are why the window opens on 20 July. In the peak week, at the start of July, Indonesian takers made 1.07 million attempts, 81.4% of them by under-18s. As term resumed, the under-18 share fell to 49.7% and the weekly average rose by 3.5 points. Nothing changed about Indonesians over those weeks; a different group of people was taking the test. Including the wave barely changes the order of countries (rank correlation 0.97), but it would have swamped the averages.

• Nobody was sampled. People who search for an IQ test are not a random slice of any country, and nothing here is weighted to a population.
• Who is online differs. Where internet access is patchy, takers skew young, urban and connected, so a country’s takers are rarely like its population.
• Language versions differ. Translations use four separately written verbal questions. We could not test whether the versions are equally hard.
• Age is self-reported. Age is asked after the last question, and nobody is checked. About half of those who answer say they are under 18 and are excluded, so an under-18 who says otherwise stays in.
• Location comes from the connection. A VPN, a trip abroad or a job in another country puts people in the wrong row. Where most takers used the Indonesian version (Indonesia, Taiwan, Singapore, South Korea and others), a row describes an Indonesian-language community living there more than the country.
• Reach is uneven. Indonesia supplies 59.3% of the first attempts in the countries with 1,000 or more adults, largely because the test ranks at the top of Indonesian search results. Elsewhere our takers are a small, curious slice of people who went looking for an IQ test in English.
• Effort is invisible. We drop attempts under three minutes, but a careless answer can take longer than that.
• One short test. Thirty-three questions, twenty of them matrices, with a ceiling at the top. A different test would reorder some countries.
• A snapshot. Ten weeks of one summer. School calendars shift who takes the test, and the figures will move with them.
• A person can count twice. Someone who tests on a second device or clears their cookies is counted again, and people in one city on one day are slightly alike. The week-to-week term in every range is what carries this.

Average IQs, rank numbers, a figure for any country not placed, trends within the window, claims that one language version is easier, and any comparison of children with adults across countries. Each looked tempting and each failed at least one check.

Dataset version

3.1

Adults in placed countries

427,076

Responses

20 July to 29 September 2026

Licence

CC BY 4.0

The numbers are those of version 3.1. If the data or the method changes, the version changes and this line says what moved. Cite the version you used: myiqtested.com/average-iq-by-country.

References

• Angrist, N., Djankov, S., Goldberg, P. K., & Patrinos, H. A. (2021). Measuring human capital using global learning data. Nature, 592(7854), 403–408. doi:10.1038/s41586-021-03323-7
• Flynn, J. R. (1987). Massive IQ gains in 14 nations: what IQ tests really measure. Psychological Bulletin, 101(2), 171–191. doi:10.1037/0033-2909.101.2.171
• Lynn, R., & Vanhanen, T. (2002). IQ and the Wealth of Nations. Praeger.
• OECD (2023). PISA 2022 Results (Volume I): The State of Learning and Equity in Education. OECD Publishing. doi:10.1787/53f23881-en
• Pietschnig, J., & Voracek, M. (2015). One century of global IQ gains: a formal meta-analysis of the Flynn effect (1909–2013). Perspectives on Psychological Science, 10(3), 282–306. doi:10.1177/1745691615577701
• Richardson, K. (2004). IQ and the Wealth of Nations [Book review]. Heredity, 92(4), 359–360. doi:10.1038/sj.hdy.6800418
• Ritchie, S. J., & Tucker-Drob, E. M. (2018). How much does education improve intelligence? A meta-analysis. Psychological Science, 29(8), 1358–1369. doi:10.1177/0956797618774253
• Sear, R. (2022). ‘National IQ’ datasets do not provide accurate, unbiased or comparable measures of cognitive ability worldwide. PsyArXiv. doi:10.31234/osf.io/26vfb
• Wicherts, J. M., Dolan, C. V., & van der Maas, H. L. J. (2010). A systematic literature review of the average IQ of sub-Saharan Africans. Intelligence, 38(1), 1–20. doi:10.1016/j.intell.2009.05.002
• World Bank (2020). Harmonized Test Scores, Human Capital Index (indicator HD.HCI.HLOS). Accessed 15 September 2026. data.worldbank.org/indicator/HD.HCI.HLOS