How Much Did Schoolchildren Lose During COVID-19 School Closures — and How Much Have They Recovered?

The question

Between March 2020 and 2022, schools worldwide shut their doors to more than a billion children at once — the largest disruption to formal education in recorded history. Some systems closed for weeks; others, particularly across Latin America and parts of South Asia, stayed remote or hybrid for a year or more. The immediate fear was learning loss: students falling behind the trajectory they would have followed in an ordinary classroom. Five years on, that fear can be assessed against actual data. Two sub-questions matter and are often conflated:

1. How large were the losses when schools were closed? This is the closure-period question, answered by pre/post assessments, difference-in-differences designs against historical trends, and meta-analyses of both.
2. How much of the deficit remained once schools reopened — and what, if anything, closed it? This is the recovery question, answered by later assessment rounds, within-student longitudinal tracking, and evaluations of remediation programmes.

This brief synthesizes the best current evidence on both, with special attention to where the two literatures disagree and why.

How the evidence is built

Three design families dominate, and their weaknesses differ:

• Cohort comparisons contrast a pandemic-exposed grade cohort against an earlier, pre-pandemic cohort on comparable assessments (TIMSS, PIRLS, national tests). They are cheap and global but confound cohort composition, test-format changes, and any secular trend not captured by the comparison years.
• Within-student or administrative-panel designs track the same students, or near-universal administrative records, across the closure boundary. The Dutch national-testing study (Engzell, Frey & Verhagen, PNAS, 2021) is the archetype: biannual standardized tests for roughly 350,000 pupils, with difference-in-differences against the three prior years. São Paulo State's universe-of-students administrative data (Lichand et al., Nature Human Behaviour, 2022; Lichand & Dória, PNAS, 2024) is the largest such system in a middle-income setting. These designs identify causal closure effects far better but are rare outside a handful of countries.
• Meta-analyses pool both, inheriting the weaknesses of their inputs. Three large ones now exist — Betthäuser, Bach-Mortensen & Engzell (Nature Human Behaviour, 2023); Di Pietro (Educational Research Review, 2023); and Wisenöcker et al. (Learning and Instruction, 2025) — with a combined scope exceeding 150 primary studies.

The post-pandemic international assessments (TIMSS 2023, PIRLS 2021, PISA 2022) add a fourth family: deviations from each country's own long-run trend line across repeated cross-sections, which is how Gajderowicz et al. (Learning and Instruction, 2026) estimate the global picture. Each family answers a slightly different question, and the honest synthesis keeps them apart rather than averaging them into a single number.

How large were the losses

The meta-analytic consensus is a pooled loss of roughly 0.14 to 0.20 standard deviations — about a third of a typical school year. Betthäuser et al. (2023) pool to d ≈ −0.14; Di Pietro (2023), with a deliberately more geographically balanced sample of 39 studies across 19 countries, gets −0.19; Wisenöcker et al. (2025), the largest compilation (763 effect sizes from 103 studies in 45 countries), gets −0.20 (95% CI −0.28 to −0.12). Against a benchmark of roughly 0.4 SD gained in a normal school year, the Wisenöcker point estimate corresponds to losing about half a year's worth of learning. Publication-bias corrections in that meta-analysis (PET-PEESE) push the estimate slightly further from zero, not closer — one of the few literatures where that pattern holds.

Four regularities survive across all three syntheses:

• Mathematics suffers about twice as much as reading. Wisenöcker et al. find d = −0.26 for mathematics versus −0.12 for reading; Di Pietro finds the same asymmetry. The mechanism is plausible: math learning is more scaffolded and school-dependent, while reading draws more on home environments.
• Disadvantaged students lose 40–60% more. In the Dutch study, pupils from the least-educated households lost 40–60% more than average. TIMSS 2023 finds top performers statistically unaffected while low performers (10th percentile) lost 0.14–0.23 SD. The distributional effect is the most consistent finding in the entire literature.
• Closure duration has a dose-response relationship with loss. In Gajderowicz et al.'s (2026) analysis of TIMSS 2023 — more than two million students across 71 education systems — achievement sat 0.11 SD below pre-pandemic trends on average, but the deviation grew with every additional week of full or partial closure: roughly 0.07–0.21 SD of extra loss for a full school year of closure, depending on grade and subject. Grade 8 students lost more than Grade 4 students.
• Effects are small in Cohen's terms but real in schooling terms. A d of −0.15 reads as "small" by generic effect-size conventions; it reads as "roughly eight weeks of instruction" when converted against annual learning rates. The framing matters: these are not statistical artifacts but curriculum-sequencing gaps that compound.

How much has been recovered

Here the evidence genuinely diverges, and the divergence is informative rather than embarrassing.

TIMSS 2023 shows the deficit had not closed at the system level by 2023. Gajderowicz et al. find achievement still 0.11 SD below the long-run trend across 71 systems — three to four years after the closures. If recovery had been rapid and universal, this international assessment should have shown a return to trend; it did not. The losses remain concentrated exactly where the closure-era evidence predicted: low achievers, girls in some subjects, linguistic minorities.

The United States shows a large but partial recovery that stalled. The Education Recovery Scorecard project (reported by Stanford's Institute for Economic Policy Research in 2025), which tracks district-level assessment data, found U.S. students roughly half a grade behind in mathematics as of spring 2024 — recovered from the pandemic peak but far from baseline, with recovery rates up to four times likelier in financially healthier districts and chronic absenteeism consistently associated with slower catch-up. Wright et al. (Child Development, 2026), following a longitudinal cohort of disadvantaged U.S. children, documents continued academic stagnation relative to pre-pandemic trajectories — the micro-level counterpart to the district-level picture.

Brazil shows partial, remediation-dependent recovery. The São Paulo State administrative data provide the sharpest causal picture anywhere. Lichand & Dória (PNAS, 2024) estimate that secondary students under extended remote learning in 2020 learned roughly 28% of the in-person equivalent; by the fourth quarter of 2021 — a full year back in classrooms — they had recovered only about 37% of the accumulated loss. Critically, the gaps between municipalities that reopened early and those that stayed remote did not fade with time; if anything they grew (a persistence ratio of 215% by late 2021), which is the opposite of mechanical catch-up. Where recovery exceeded the state average, it tracked specific remedial policies: tutoring sessions, extended full-time schooling, improved management practice, and family communication, each associated with up to 18 percentage points of extra recovery relative to the state-wide rate. Elsewhere in Brazil the picture is mixed: Miranda & Baum (EPAA, 2024) find essentially no recovery in Brasília public schools between 2020 and 2021, while Starling-Alves, Hirata & Oliveira (International Journal of Educational Development, 2023), following the same students longitudinally in a well-resourced municipal system, find that most children kept gaining reading skills during closure — slower than normal, producing a roughly one-year cohort gap rather than an outright loss.

Some settings did fully recover. The strongest countercase is Singh, Romero & Muralidharan (ADBI Working Paper 1536, 2026), who use near-universal administrative test data for Tamil Nadu, India — a state of tens of millions of schoolchildren — and find complete recovery of pandemic-era learning losses within roughly 20 months of reopening. German large-scale assessments point the same direction at the top: Schult et al. (Learning and Instruction, 2026), analyzing more than 200,000 students across all public schools in one German state, find that cohort-level recovery from grade 5 to grade 8 proceeded at pre-pandemic rates regardless of a school's social composition or its number of lost school days — on average, German schools rebounded. Hang, Aurino & Wolf (International Journal of Educational Development, 2026) add an important temporal twist from Ghana: remote-learning engagement during closure predicted better achievement in 2021 and 2022 but no longer in 2024 — associations that looked like protection faded with enough time, and below-average students did not benefit at all.

What actually closed the gap

Three evidence streams converge on the same conclusion: recovery follows active remediation, not the mere return to in-person classes.

First, the São Paulo quasi-experiments quantify the dose. Municipalities and schools that adopted tutoring, extra instructional time, or structured management support recovered measurably faster than the state average; reopening timing alone produced small, slow effects (each additional week of in-person classes worth about 0.005 SD of recovery — roughly 5% of a normal year's learning rate). Second, the international pattern separates effort from exposure: Tamil Nadu's full recovery coincided with an intensive state remediation push, while systems that reopened without remediation (early-pandemic Brasília; U.S. districts with high absenteeism and uneven spending of federal recovery funds) recovered least. Third, and most decisively, randomized evidence now shows what specific remediation can do. Angrist et al. (Nature, 2026) report five randomized trials across India, Kenya, Nepal, the Philippines, and Uganda during COVID-era closures: weekly 20-minute phone-call tutorials delivering targeted numeracy instruction to each child's level, combined with SMS content, produced pooled gains of 0.32 SD versus controls, at about US$11 per child for eight weeks — roughly 3.6 quality-adjusted years of instruction per US$100 spent, placing it among the most cost-effective education interventions ever measured. In Uganda, where control-group children lost ground (the share who could divide fell from 17% to 10%), treated children reached 48% — full recovery and more. Gains were equally large when delivered by government teachers as by NGO instructors, and SMS alone was far weaker than live targeted calls.

The common ingredient across the successful cases is targeting: instruction pitched at the child's actual learning level rather than the grade-level curriculum, with enough contact time to move each child up a step. That is the one recovery mechanism with consistent support from quasi-experimental, administrative, and randomized evidence alike.

Mechanisms, competing explanations, and what the numbers miss

Why doesn't reopening automatically restore the trajectory? The São Paulo persistence result — gaps that widened rather than closed after in-person classes resumed — points away from a simple "missed content" model toward at least three channels: motivational scarring (students disengaged from schooling during remote learning and did not fully re-engage), household scarring (parental job loss, bereavement, and mental-health effects during the pandemic depressed the home learning environment), and structural mismatch (grade-level curriculum continues while the missing prerequisite content is never retaught). The Starling-Alves longitudinal finding — children still gaining during closure, just more slowly — fits a dosage model: remote instruction delivered a fraction of normal learning productivity, and whether that fraction sums to a "loss" or merely a "gap" depends on how long the exposure lasted and what counterfactual growth rate one assumes.

Two competing explanations deserve serious treatment. The first: the measured losses are partly artifacts of test anxiety, changed test formats, or pandemic-era confounds (illness, bereavement, family stress), not lost learning per se. Engzell et al. address this directly: a placebo outcome — oral reading fluency, a test with no curricular content and no preparation effect — shrank by two-thirds relative to the main curricular assessment, meaning the large majority of the measured drop reflects lost knowledge rather than exam-day conditions. The second: recovery claims are inflated by regression to the mean and selective attrition of the lowest performers from testing systems. This is a real concern in systems with high dropout during the pandemic (São Paulo's remote period raised dropout risk by 365%), and it is why within-student designs and universal administrative data — not cross-sectional cohort averages — deserve the most weight.

Finally, the measured outcomes are narrow. Standardized math and reading scores are what the literature tracks best; they miss socio-emotional development, non-cognitive skills, school dropout (itself a large and irreversible cost, especially in lower-income systems), and the long-run wage and health consequences that will take a decade to measure. The World Bank has warned that the closure cohort could lose trillions in lifetime earnings; those projections are modelled extrapolations from test-score elasticities, not observed facts, and should be cited as scenarios.

Evidence quality, caveats, and what would change the picture

• The loss estimates are robust; the recovery estimates are not. Meta-analytic loss figures vary by a factor of two (−0.14 to −0.20) depending on sample composition, but never cross into "no effect." Recovery estimates range from "full" (Tamil Nadu) to "none" (Brasília, São Paulo without remediation, U.S. low-income cohorts) and are highly sensitive to design, outcome measure, and time horizon. Any single recovery number should be distrusted; the conditional statement — recovery where remediation was intensive, stagnation where it was not — is what the evidence actually supports.
• Publication and language bias exist but do not overturn the loss consensus. Wisenöcker et al. document funnel asymmetry and correct for it; the corrected estimate is slightly larger, not smaller, than the raw pooled one.
• The closure-duration gradient is correlational at the cross-country level. Countries that closed longer differ from those that closed briefly in income, governance, pandemic severity, and baseline system quality; Gajderowicz et al. control for demographics and country fixed effects, but the dose-response should not be read as a clean causal parameter.
• TIMSS 2023 measures mathematics and science in grades 4 and 8. Language, humanities, early-childhood, and tertiary outcomes are largely outside the post-pandemic international picture.
• What would update this most: PISA 2025 results (due December 2026) on 15-year-olds; longer-run follow-ups of the Tamil Nadu full-recovery cohort; and randomized evidence on remediation delivered after reopening at scale (the phone-tutoring trials ran during closure; the São Paulo remediation evidence is quasi-experimental).

What this brief does not claim

It does not claim the losses are permanent — several settings have demonstrated full recovery, and the closure cohort is still early in its schooling. It does not claim closures were a policy error; their health and mortality benefits are outside this brief's scope, and the education question is about costs, not net welfare. It does not treat remote instruction as worthless: most studies find some learning during closure (São Paulo: about 28% of the in-person rate; Starling-Alves: continued reading gains), just far below normal productivity. And it does not extrapolate test-score losses directly into adult outcomes; the earnings and productivity projections remain modelled scenarios.

Verdict and calibrated confidence

That school closures caused real, measurable learning losses on the order of 0.14–0.20 SD — roughly a third of a school year: ~95%. Three independent meta-analyses, dozens of quasi-experimental studies, and the post-pandemic international assessments all converge, and the publication-bias corrections do not weaken the effect.

That the losses disproportionately harmed disadvantaged and low-achieving students, widening inequality: ~90%. The single most replicated pattern in the literature, from the Dutch family-level analysis to TIMSS 2023's quantile regressions to the U.S. district-level recovery gradients.

That mathematics was damaged about twice as much as reading: ~85%. Consistent across both major meta-analyses that tested the contrast, with a plausible schooling-dependence mechanism.

That as of 2023–2024, the average school system had not fully recovered: ~80%. TIMSS 2023's continued 0.11 SD deviation across 71 systems is the anchor; the U.S. district data agree. The residual uncertainty reflects the genuine full-recovery findings from Tamil Nadu and the German state, which show that non-recovery is not a law of nature.

That reopening alone was insufficient and targeted remediation drove the recovery that did happen: ~85%. Supported by the São Paulo persistence-and-remediation design, the cross-country effort-versus-outcome pattern, and five randomized trials of targeted phone tutoring. Lower than 90% because the randomized evidence for post-closure remediation at system scale is still thin, and Tamil Nadu's full recovery blends state effort with cohort composition in ways not fully decomposed.

What would move me most: the December 2026 PISA 2025 release (the first assessment of the oldest closure-exposed cohort at scale); registered replications of the Tamil Nadu result in other large middle-income systems; and long-run tracking of whether the low-achiever losses documented in TIMSS 2023 concentrate into dropout, track-placement, and labor-market scarring over the next decade. On current evidence, the defensible summary is: the world lost roughly a third of a school year's worth of learning, the loss compounded inequality, the average system carries a measurable deficit into the mid-2020s, and every documented recovery succeeded on purpose.

References

1. Engzell P, Frey A, Verhagen MD. Learning loss due to school closures during the COVID-19 pandemic. Proceedings of the National Academy of Sciences. 2021;118(17):e2022376118. doi:10.1073/pnas.2022376118.
2. Betthäuser BA, Bach-Mortensen M, Engzell P. A systematic review and meta-analysis of the impact of the COVID-19 pandemic on learning. Nature Human Behaviour. 2023;7:1078–1090. doi:10.1038/s41562-022-01506-4.
3. Di Pietro G. The impact of Covid-19 on student achievement: Evidence from a recent meta-analysis. Educational Research Review. 2023;40:100530. doi:10.1016/j.edurev.2023.100530.
4. Wisenöcker AS, Helm C, Große CS, Hübner N, Zitzmann S. A meta-analysis of students' academic learning losses over the course of the COVID-19 pandemic. Learning and Instruction. 2025;98:102111. doi:10.1016/j.learninstruc.2025.102111.
5. Gajderowicz T, Jakubowski M, Kennedy A, Kjeldsen CC, Patrinos HA, Strietholt R. The learning crisis: Three years after COVID-19. Learning and Instruction. 2026. doi:10.1016/j.learninstruc.2026.102373.
6. Dela Cruz J, Jerrard J, Kim J, et al. Learning loss and recovery from the COVID-19 pandemic. ADB Working Paper No. 653. Asian Development Bank; 2025. https://www.adb.org/publications/learning-loss-recovery-covid19-pandemic
7. Singh A, Romero J, Muralidharan K. Full recovery from the COVID-19 pandemic learning loss: Evidence from India. ADBI Working Paper No. 1536. Asian Development Bank Institute; 2026. https://www.adb.org/adbi/publications/full-recovery-from-covid-19-pandemic-learning-loss-evidence-from-india
8. Lichand G, Dória CA, Leal-Neto O, Fernandes JPC. The impacts of remote learning in secondary education during the pandemic in Brazil. Nature Human Behaviour. 2022;6:1079–1086. doi:10.1038/s41562-022-01350-6.
9. Lichand G, Dória CA. The lasting impacts of remote learning in the absence of remedial policies: Evidence from Brazil. Proceedings of the National Academy of Sciences. 2024;121(21):e2316300121. doi:10.1073/pnas.2316300121.
10. Miranda EM, Baum DR. COVID-19 learning loss and recovery in Brazil: Assessing gaps across social groups. Education Policy Analysis Archives. 2024;32. doi:10.14507/epaa.32.8082.
11. Starling-Alves I, Hirata G, Oliveira JBA. Covid-19 school closures negatively impacted elementary-school students' reading comprehension and reading fluency skills. International Journal of Educational Development. 2023;99:102753. doi:10.1016/j.ijedudev.2023.102753.
12. Wright AM, Martin A, Pollak SD, Phillips DA, Livas G, Johnson A. COVID-19-induced school closures and disadvantaged children's post-COVID academic growth: A longitudinal cohort study. Child Development. 2026. doi:10.1093/chidev/aacaf030.
13. Angrist N, Ainomugisha M, Bathena SP, et al. Mobile education builds resilience during shocks in five countries. Nature. 2026. doi:10.1038/s41586-026-10990-x.
14. Schult J, Fauth B, Schneider R, Lindner MA. How schools rebound from pandemic learning loss: Longitudinal findings from mandatory large-scale assessments. Learning and Instruction. 2026;101:102263. doi:10.1016/j.learninstruc.2025.102263.
15. Hang H, Aurino E, Wolf S. Child development post-COVID-19 school closures in Ghana: The roles of remote learning experiences. International Journal of Educational Development. 2026. doi:10.1016/j.ijedudev.2026.103650.
16. Stanford Institute for Economic Policy Research. New research reveals striking variations in pandemic recovery among U.S. school districts (Education Recovery Scorecard). Stanford University; 2025. https://siepr.stanford.edu/news/new-research-reveals-striking-variations-pandemic-recovery-among-us-school-districts