The COVID learning loss studies, added up.

In the spring of 2020, schools closed for most of the world’s pupils. At the peak, roughly 1.6 billion children were out of class (Betthäuser, Bach-Mortensen & Engzell, 2023). It was the largest disruption to formal schooling on record. It was also, grimly, an experiment. Every school system improvised its own mix of closure, remote teaching and reopening. Every system generated data.

The question that mattered from the first week was blunt: how much learning would children lose, and who would lose most? The early answers were guesses. Then came projections built from older research. Then national measurements from exam registers and testing archives. Finally, in 2023, a meta-analysis — a study that pools many studies into one estimate — put the global picture on a single page (Betthäuser, Bach-Mortensen & Engzell, 2023).

This article follows that sequence. First the projections, then the cleanest single measurement, then the global average. After that come the two gradients that should worry policymakers more than any average: the one inside countries, and the one between them. Last is the recovery evidence — what has come back, what has not, and which repair tools carry causal support.

The projections came first

The first serious numbers were forecasts, not measurements. In 2020, Kuhfeld and colleagues asked a careful question: what if the closure behaved like a very long summer holiday? Summer learning loss had been studied for decades. Growth across the school year was well mapped in US testing data from millions of pupils. Splicing the two produced a projection of where children might stand when schools reopened (Kuhfeld et al., 2020).

The projection was sobering, and lopsided. Pupils returning in autumn 2020 were forecast to keep most of a typical year’s reading gains. In maths, the forecast retained gain was on the order of only 37 to 50 percent of a normal year (Kuhfeld et al., 2020). The lopsidedness had a plausible cause. Children read at home, at wildly varying rates. Almost nobody practises long division for fun. Maths lives at school, so maths had more to lose.

The paper’s second warning aged even better than its first. Whatever happened to the average, the spread would widen. Some children would sail through on books, broadband and parental time. Others would lose months. Schools were told to plan for wider gaps within every classroom, not just a lower mean (Kuhfeld et al., 2020). Both warnings held up when real data arrived.

The cleanest measurement

The first clean measurement came from the Netherlands, and it is still the single best study we have. Engzell, Frey and Verhagen used the national pupil registers: standardised tests taken twice a year by primary pupils across the country. They compared the 2020 cohort against the three cohorts before it, same tests, same schools, same time of year (Engzell, Frey & Verhagen, 2021).

The setting was close to a best case. The Dutch closure was short — about eight weeks. Broadband access was among the best in the world. School funding is comparatively equitable. If emergency remote schooling could substitute for classrooms anywhere, it was here.

It did not substitute. Pupils lost roughly 3 percentile points, about 0.08 standard deviations — a fifth of a school year, almost exactly the share of the year schools were shut (Engzell, Frey & Verhagen, 2021). The plain reading is stark. During two months of remote schooling, the average Dutch child learned little or nothing.

The loss was not evenly spread. Among children from homes with the least education, losses ran up to 60 percent larger than the national average (Engzell, Frey & Verhagen, 2021). That single result previews everything the rest of this article documents at scale: a real average loss, and a steeper one underneath it.

The global ledger

By 2023 there were enough national studies to add up. Betthäuser, Bach-Mortensen and Engzell pooled 42 studies across 15 countries — the first proper global accounting of pandemic learning deficits (Betthäuser, Bach-Mortensen & Engzell, 2023).

The pooled deficit was about 0.14 standard deviations, which the authors translate as roughly 35 percent of a normal school year’s progress. Just as important was the shape over time. Deficits appeared early in the pandemic and then persisted — through mid-2022 they neither closed nor grew much on average (Betthäuser, Bach-Mortensen & Engzell, 2023). Children kept learning, but from a lower rung. The hole travelled with them.

Inside the pooled number sat three patterns. Deficits were larger in maths than in reading, as the projections had guessed (Kuhfeld et al., 2020). They were larger for children from poorer families in most studies that could test it. And the evidence itself was skewed: nearly all of it came from high-income countries, while the middle-income countries that were measured showed clearly larger deficits (Betthäuser, Bach-Mortensen & Engzell, 2023).

The inequality gradient

Averages are what headlines quote. Gradients are what school systems should plan around, because every well-measured system shows the same tilt: the children who started with the least lost the most.

Flanders, in Belgium, measured it at the school-leaving line. Maldonado and De Witte used the standardised tests taken at the end of Flemish primary school. Against earlier cohorts, the 2020 pupils scored roughly 0.19 standard deviations lower in maths and 0.29 lower in Dutch. Inequality rose at the same time — the spread between pupils within schools grew by roughly a fifth (Maldonado & De Witte, 2022).

The United States measured it through testing archives. In NWEA’s national data, achievement fell furthest in high-poverty schools. By autumn 2021, maths scores in high-poverty schools sat roughly 0.27 standard deviations below pre-pandemic peers, against roughly 0.20 in low-poverty schools — and the reading gradient looked similar (Kuhfeld, Soland & Lewis, 2022).

The mechanism is not mysterious. Remote schooling converts home resources into learning: a quiet room, a device per child, an adult with time to help. Those inputs are exactly what income buys. In several systems, schools serving poorer pupils also stayed remote longer. So the same shock arrived with a larger dose, landing on thinner buffers (Betthäuser, Bach-Mortensen & Engzell, 2023).

Poorer countries lost more

The same gradient repeats between countries. In the global pool, the middle-income countries with usable data — Brazil, Mexico, Colombia, South Africa — showed markedly larger deficits than the high-income average (Betthäuser, Bach-Mortensen & Engzell, 2023). Longer closures, weaker connectivity and thinner household buffers all pushed the same way.

The sharpest documented case is São Paulo, Brazil. Lichand and colleagues compared secondary pupils taught remotely with earlier in-person cohorts, using the state’s own standardised tests. Scores under remote schooling came out roughly 0.32 standard deviations lower. Put differently, pupils learned only about a quarter of what an in-person year normally delivers. The risk of dropping out rose several-fold in the same data (Lichand et al., 2022).

That dropout signal deserves more attention than it gets. Test scores measure the children still being tested. Where the pandemic pushed adolescents out of school entirely — into work, care duties or disengagement — the score-based estimates quietly lose the very pupils who lost the most (Lichand et al., 2022).

Remote schooling was the dose

Was the damage done by the virus, or by the closures? The best within-country evidence treats schooling mode as the variable. US states reopened on wildly different schedules, so Jack, Halloran, Okun and Oster could compare districts that made different choices while facing the same pandemic (Jack et al., 2023).

The pattern was clean. Pass rates on state tests fell everywhere, but they fell far more where schooling stayed remote or hybrid for longer. Where in-person teaching was on offer, losses were markedly smaller. And the interaction was the cruel part: remote schooling did the most damage in high-poverty districts — often the same districts that stayed remote longest (Jack et al., 2023). Disadvantage picked the dose, and the dose hit disadvantage hardest.

Across countries, the numbers rhyme like a dose-response curve. Eight closed weeks in the Netherlands cost about 0.08 SD (Engzell, Frey & Verhagen, 2021). A mostly remote stretch in São Paulo cost roughly 0.32 (Lichand et al., 2022). Cross-country comparisons are rough — tests, ages and conditions all differ — but the direction never flips: more time out of the classroom, more loss.

What recovery looks like so far

The recovery record is best documented in the United States, where the NWEA testing archive tracks tens of millions of test events. Kuhfeld, Soland and Lewis followed achievement across three pandemic-affected school years. Their timing finding surprised many people: the largest slide came during the 2020–21 school year, not the chaotic spring of 2020 (Kuhfeld, Soland & Lewis, 2022).

By autumn 2021, US pupils sat roughly 0.20 to 0.27 standard deviations behind pre-pandemic peers in maths, and roughly 0.09 to 0.18 behind in reading, depending on grade (Kuhfeld, Soland & Lewis, 2022). Then the curve bent. Growth in 2021–22 began to outpace pre-pandemic norms, and later NWEA reports tracked a continuing, partial rebound. Partial is the operative word: gaps to the 2019 benchmark remained, and they closed slowest in high-poverty schools.

What actually closes the remaining gap? The tool with the strongest causal record is high-dosage tutoring. Pooling nearly a hundred randomised trials, Nickow, Oreopoulos and Quan estimate an average effect of roughly 0.37 standard deviations — among the largest reliable effects ever measured for a scalable schooling intervention (Nickow, Oreopoulos & Quan, 2020). The recipe behind that number is specific: several sessions a week, during the school day, tied to the class curriculum, delivered by a consistent tutor.

The catch is arithmetic. Tutoring at that dosage is expensive in money and people, and the pooled effects come mostly from well-run programmes of modest size. Systems that stretched the model thin — fewer sessions, after-school slots, rotating tutors — saw much weaker results. Recovery, in short, is a targeting problem and a dosage problem, and budgets solve neither automatically (Nickow, Oreopoulos & Quan, 2020).

What the evidence doesn’t show

The learning-loss literature is unusually strong for education research — registers, archives, whole cohorts. It still has edges, and honest policy should hold them alongside the headline.

• Closures are not the whole cause. The studies measure the pandemic period, not school closure in isolation. Illness, bereavement, family income shocks and stress travelled with the closures, and the designs cannot fully separate the strands (Betthäuser, Bach-Mortensen & Engzell, 2023).
• Averages hide enormous spread. The projections warned that variability would widen, and it did. Behind every mean deficit sit children who lost nothing and children who lost a year (Kuhfeld et al., 2020).
• Who took the tests changed. Score-based estimates only see tested pupils. Where absence and dropout rose, the most affected children quietly left the sample — which flatters the numbers (Lichand et al., 2022).
• The poorest countries are barely measured. The global pool draws overwhelmingly on high- and middle-income systems. For the countries with the longest closures and least connectivity, magnitude is still guesswork (Betthäuser, Bach-Mortensen & Engzell, 2023).
• Long-run costs are projections. Estimates of lifetime earnings losses are models built on the score deficits, not observed outcomes. They are useful for budgeting and unprovable for a generation.
• Recovery tools thin out at scale. The tutoring effect of ≈0.37 SD comes mostly from modest, well-run programmes; stretched versions deliver less. Scale-up evidence is still young (Nickow, Oreopoulos & Quan, 2020).

Where the evidence stops

1. 1Closures are not the whole cause
2. 2Averages hide enormous spread
3. 3Who took the tests changed
4. 4The poorest countries are barely measured
5. 5Long-run costs are projections
6. 6Recovery tools thin out at scale

© 2026 FUTURE PROOF™

Reading the loss like a school system

Read together, the studies condense into a short operating manual for schools, districts and ministries still carrying the deficit.

Diagnose level, not grade. The pandemic widened the spread inside every classroom, so the grade label now describes pupils less accurately than at any time on record (Kuhfeld et al., 2020). Recovery teaching that starts from the curriculum’s assumed level misses the children furthest behind. Start from measured levels, child by child.

Put maths first. The deficit is consistently larger in maths than reading, in the projections, the national studies and the global pool (Betthäuser, Bach-Mortensen & Engzell, 2023). Maths is also the subject least likely to repair itself at home. Weight catch-up time accordingly.

Weight money by the gradient, not the mean. Losses ran steepest for poorer pupils and poorer schools everywhere they were measured (Maldonado & De Witte, 2022). Flat per-pupil recovery funding rebuilds the old gaps on schedule. Progressive funding is not politics here; it is arithmetic.

Buy dosage, not licences. The strongest recovery tool is tutoring at real dosage — several sessions a week, in school time, tied to the curriculum (Nickow, Oreopoulos & Quan, 2020). A cheaper programme at a tenth of the dose is not a tenth as good; it is usually indistinguishable from nothing.

Count the missing pupils. Attendance and dropout are recovery metrics, not administrative ones. The children absent from the test are the loss estimate’s blind spot (Lichand et al., 2022).

Plan in years, and measure as you go. The deficit persisted for two years without intervention (Betthäuser, Bach-Mortensen & Engzell, 2023), and the US rebound is real but partial (Kuhfeld, Soland & Lewis, 2022). One catch-up term was never a plan. A multi-year budget with live measurement is.