What the data can tell you, how sure we are, and what they cannot answer.
Every number is computed from the PISA 2022 and 2025 public files, with PISA's standard errors, and checked against the OECD's published tables. The same analyses are run for every country. Only the United States, Israel and Latvia have so far been compared with their national reports. For other countries, check the OECD's notes on the country before quoting.
Subject
Check before you quote
Three claims that national reports and news stories often make, checked for this country with all students and with PISA's margins of error.
Who fell: weak or strong students?
Did boys and girls change differently?
Use a rangeRank
Compared with the U.S. national report (NCES)
The National Center for Education Statistics published Highlights of U.S. PISA 2025 Results (NCES 2026-032). We checked its main findings against the analysis on this page. The numbers and the conclusions are the same. One headline sentence goes further than the data, though NCES's own detail does not. And NCES is more careful than a simple reading of the data about results by background.
Read all U.S. results with caution
The U.S. sample did not meet PISA's standards: 45% of sampled schools took part before replacement (54% after), and 76% of sampled students. The OECD's Adjudication Group concluded that "it is not possible to exclude the possibility of bias on performance indicators, nor to determine its most likely direction" (PISA 2025 Technical Report, Chapter 18). Under adverse but hypothetical scenarios, the distortion could reach about 15% of a standard deviation.
Where to quote carefully
Headline stronger than the detail"The largest gap of any education system"
Where both agree the data cannot answer
Not reportedResults by family background and by race or ethnicity
More than 40% of U.S. students did not receive the student questionnaire: about half because 9 of the 50 states chose not to administer it, and about half because their schools tested early in the testing window. The OECD recommended "limited reporting of questionnaire results", and NCES did not report results by socio-economic status or by race or ethnicity. On this page, U.S. results that rely on the student questionnaire (family background, immigrant background and students' answers) carry a warning: they leave out those 9 states entirely, and no statistical adjustment can bring them back.
Where the results are the same
Mean scores (reading 490, mathematics 463, science 502) and the change since 2022 (reading down 14 points, significant; mathematics and science not significantly different), the number of education systems scoring higher, similar and lower, the shares of low and top performers, the 90th-to-10th percentile gap in reading (312 points), and the gender gap: in mathematics it widened from 13 to 24 points in favour of boys, a significant change, while in reading and science it did not change significantly. NCES compares with 89 other education systems, including Cyprus, which is not in the public file.
Latvia: what changed, and for whom
Latvia's national presentation of the results (University of Latvia, 8 September 2026) was not accessible to us, so this section compares with the OECD's country note for Latvia and adds an analysis of the public data.
The language of the test changed
Between 2023 and 2025 Latvia completed a transition to Latvian as the sole language of instruction in compulsory education, and "starting with 2023, all centralized tests (including PISA) were held in Latvian for all students". "In 2022, some 28% of students had taken the PISA test in Russian" (OECD country note). Changes since 2022 therefore mix two things: what students learned, and the language in which they were tested.
Read with the test language in mindWhere the decline is concentrated
Change since 2022 for students who speak Latvian at home: reading −20 points (±7), mathematics −12 (±6), science −3 (±8, not significant). For students who speak another language at home, mostly Russian: reading −95 (±10), mathematics −43 (±9), science −69 (±11). In 2022 most of these students were tested in Russian, in 2025 in Latvian. Part of their drop may reflect taking the test in a second language rather than a loss of knowledge, and the data cannot separate the two. "Latvian at home" means tested in Latvian and speaking the test language at home, a definition that is comparable across the two cycles. An independent analysis of the same public data (J. Ošlejs, nra.lv, September 2026) reports the same 2025 results for students who speak Latvian at home.
Agrees with the OECDThe gap between strong and weak students widened
Depends on who is countedAdvantaged and disadvantaged students
Where the results are the same
Mean scores fell in all three subjects and were "the lowest ever observed in PISA in Latvia", and the shares of low and top performers match the OECD's tables. The OECD states that "in Latvia, all data met the quality standards set by PISA".
Compared with Israel's national report (RAMA)
RAMA's report on PISA 2025 (September 2026) is Israel's official publication. We checked its main findings against the analysis on this page. Most numbers are the same. In a few places, this analysis leads to a different conclusion.
Where the conclusion differs
Different conclusion"The strong weakened"
Different conclusion"Boys fell more than girls"
Different precisionA single rank
Different approachLinks to scores
RAMA shows score gaps between students by factors such as AI use, family support, phone distraction and teacher shortage, with a caution that they are not causal. This page does not show such gaps, because they mostly reflect other differences between students (see the class-size example).
Where the results are the same
Means and changes (reading 436, down 38; mathematics 433, down 25; science 442, down 23), shares of low and top performers, the trends for Hebrew and Arabic speakers (down 44 and 21 points in reading), low performers by language, and the principals' and students' reports (teacher shortage 50% against 39% in the OECD, "school has been a waste of time" 37% against 24%, phone ban 30% against 49%). As in RAMA's report, Haredi education is shown for girls only, and their scores match the report (reading 503, mathematics 455, science 466).
Only in RAMA's report
Trends by sector between 2022 and 2025 (for example, a 52-point drop in reading in state education) and by sector and gender. The public 2022 file does not identify sector, so they cannot be computed from open data.
How much did scores change?
The data support this
Each row is one country, from the largest decline (top) to the largest gain. Dots are the change, lines the 95% margin including the linking error. The black line is the median country.Show as a table"Similar" means the difference in 2025 scores is not statistically significant. A single rank would overstate how precisely a country's place is known.
Weak and strong students
20222025
The 10th percentile is the score below which the weakest tenth of students fall; the 90th, the score above which the strongest tenth fall. Labels show the change with its 95% margin.Points between the 90th and the 10th percentile, 2025. Each row is one country, widest gap at the top. The black line is the median country. Lines are 95% margins.Show as a table
For whom?
Partly: check each kind
Gaps and changes between groups
Supported, check the cell sizes
20222025
± is the 95% margin. A dash marks a group with fewer than 30 students or 5 schools, which PISA does not report. Changes include the linking error.
Change by family background
Students with background data onlyAll students
Change 2022 to 2025 by family-background third, within the country. Lines are 95% margins. Students without background data are placed by their predicted background.
Whether an effect differs between groups
Understated by the data
Questions such as "does family background matter more in some schools than in others?" ask how an effect differs between groups. The model that produces PISA scores does not include such differences, so the released data show them smaller than they are.
Large difference: the extra background gap in Arabic-language schools in Israel, science25 → 20 pointsAbout a fifth smaller in released-style data. Still clear.
Modest difference: whether background matters more for the weakest students, mathematics16 → 7 of 21Simulations in which it was found: full model against released-style data.
Simulation on Israel's PISA 2025 design, 21 runs. The same model is used in every country. Report a small or missing difference of this kind as inconclusive, not as "the same for everyone".
What students and principals report
How common, for whom, and change
PISA asks students and principals about school life. These answers are facts worth reporting in their own right. This section shows how common each answer is. It does not link answers to scores.
Among all countries, 2025Each row is one country, sorted by share. The black line is the OECD average. Lines are 95% margins.Show as a table
Change since 2022
20222025
Groups with fewer than 30 students or 5 schools answering are not reported. Lines are 95% margins.
Why did scores change?
PISA cannot answer this
PISA tests each student once and does not assign schools or policies. It shows what goes together, not what causes what. A gap in scores between students who differ on one question usually reflects many other differences between them.
A worked example from Israel: class size
In Israel, students in larger classes score higher in mathematics. This is not because crowding helps. Larger classes are more common in schools with more advantaged students. Taking only school background and sector into account cuts the gap by about 40%, and other differences remain unmeasured.
What PISA can do for a "why" claim is check its premise. If a claim says a decline is due to a teacher shortage, the section above shows whether principals reported more shortage in 2025 than in 2022. A factor that did not change cannot by itself explain a change.