A systematic review and meta-analysis of the evidence on learning during the COVID-19 pandemic

42 studies across 15 countries yielding 291 effect estimates, screened from 5,153 de-duplicated hits; a further 19 studies rated at critical risk of bias were excluded from every analysis

School-aged children, predominantly primary and lower-secondary, in high-income countries plus four middle-income countries (Brazil, Colombia, Mexico, South Africa); no studies from low-income countries were found

Nature Human Behaviour 7, 375-385. Pre-registered systematic review with formal ROBINS-I risk-of-bias appraisal — the appraisal is the reason this grades B rather than C, because the 19 critical-risk studies are dropped rather than pooled. Estimates are inverse-variance weighted and expressed in Cohen's d against pre-pandemic trends. The identifying variation is a simultaneous, involuntary, externally imposed reduction in in-person schooling — the largest instructional-time shock ever measured — but it is NOT a clean subtraction of instructional hours: closure substituted remote schooling for in-person schooling and came bundled with household disruption, illness and economic shock, so the estimate is the effect of the closure package, not of hours per se. The authors themselves name the field's recurring biases: failure to account for secular time trends, non-representative samples, and imbalance between treatment and comparison groups.

Pooled learning deficit d = -0.14 (95% CI -0.17 to -0.10), t(41) = -7.30, p < 0.001. The authors' own conversion is the useful one: since students normally gain about 0.4 SD per school year, -0.14 SD "implies that students lost out on about 35% of a normal school year's worth of learning." The deficit arose early and then stopped moving — it "neither closed nor widened substantially" through two and a half years, which cuts against BOTH the fear of runaway accumulation and the hope of spontaneous catch-up. Maths was hit roughly twice as hard as reading (median -0.18 versus -0.09; mean difference delta = -0.07, t(41) = -4.02), the same subject asymmetry the summer-loss literature reports and for the same proposed reason: reading is practised outside school and maths mostly is not. Deficits are consistently larger for children from low socio-economic backgrounds, across timepoints, countries, grades and subjects. Middle-income countries fared appreciably worse than high-income ones (difference significant, t(41) = -2.78, p = 0.008, 95% CI -0.50 to -0.08), and the three largest deficits in the sample are all from middle-income countries.

Low for the shock itself — closures were imposed on entire jurisdictions and not chosen by families, so exposure does not express child or parental traits. The socio-economic gradient in the DEFICIT is a different matter and is not identified: families differ in home resources, parental availability and parental ability, all correlated with the same heritable traits that predict achievement, so "disadvantaged children lost more" is a description of who was hurt, not a demonstration that home environment caused the gap to widen.

The pooled estimate is itself a synthesis of 42 independent measurements in 15 countries, and the constituent national studies replicate each other's sign and rough magnitude (Engzell, Frey and Verhagen's Dutch estimate of ~0.08 SD for an eight-week closure sits inside the distribution). Later multi-country updates and national assessment programmes have continued to report deficits of this order with little spontaneous recovery, so the finding is treated as replicated. What is NOT replicated is any decomposition of how much of the deficit is missing instructional time as against everything else the pandemic did to children.