Main

The coronavirus disease 2019 (COVID-19) pandemic has led to one of the largest disruptions to learning in history. To a large extent, this is due to school closures, which are estimated to have affected 95% of the world’s student population1. But even when face-to-face teaching resumed, instruction has often been compromised by hybrid teaching, and by children or teachers having to quarantine and miss classes. The effect of limited face-to-face instruction is compounded by the pandemic’s consequences for children’s out-of-school learning environment, as well as their mental and physical health. Lockdowns have restricted children’s movement and their ability to play, meet other children and engage in extra-curricular activities. Children’s wellbeing and family relationships have also suffered due to economic uncertainties and conflicting demands of work, care and learning. These negative consequences can be expected to be most pronounced for children from low socio-economic family backgrounds, exacerbating pre-existing educational inequalities.

It is critical to understand the extent to which learning progress has changed since the onset of the COVID-19 pandemic. We use the term ‘learning deficit’ to encompass both a delay in expected learning progress, as well as a loss of skills and knowledge already gained. The COVID-19 learning deficit is likely to affect children’s life chances through their education and labour market prospects. At the societal level, it can have important implications for growth, prosperity and social cohesion. As policy-makers across the world are seeking to limit further learning deficits and to devise policies to recover learning deficits that have already been incurred, assessing the current state of learning is crucial. A careful assessment of the COVID-19 learning deficit is also necessary to weigh the true costs and benefits of school closures.

A number of narrative reviews have sought to summarize the emerging research on COVID-19 and learning, mostly focusing on learning progress relatively early in the pandemic2,3,4,5,6. Moreover, two reviews harmonized and synthesized existing estimates of learning deficits during the pandemic7,8. In line with the narrative reviews, these two reviews find a substantial reduction in learning progress during the pandemic. However, this finding is based on a relatively small number of studies (18 and 10 studies, respectively). The limited evidence that was available at the time these reviews were conducted also precluded them from meta-analysing variation in the magnitude of learning deficits over time and across subjects, different groups of students or country contexts.

In this Article, we conduct a systematic review and meta-analysis of the evidence on COVID-19 learning deficits 2.5 years into the pandemic. Our primary pre-registered research question was ‘What is the effect of the COVID-19 pandemic on learning progress amongst school-age children?’, and we address this question using evidence from studies examining changes in learning outcomes during the pandemic. Our second pre-registered research aim was ‘To examine whether the effect of the COVID-19 pandemic on learning differs across different social background groups, age groups, boys and girls, learning areas or subjects, national contexts’.

We contribute to the existing research in two ways. First, we describe and appraise the up-to-date body of evidence, including its geographic reach and quality. More specifically, we ask the following questions: (1) what is the state of the evidence, in terms of the available peer-reviewed research and grey literature, on learning progress of school-aged children during the COVID-19 pandemic?, (2) which countries are represented in the available evidence? and (3) what is the quality of the existing evidence?

Our second contribution is to harmonize, synthesize and meta-analyse the existing evidence, with special attention to variation across different subpopulations and country contexts. On the basis of the identified studies, we ask (4) to what extent has the learning progress of school-aged children changed since the onset of the pandemic?, (5) how has the magnitude of the learning deficit (if any) evolved since the beginning of the pandemic?, (6) to what extent has the pandemic reinforced inequalities between children from different socio-economic backgrounds?, (7) are there differences in the magnitude of learning deficits between subject domains (maths and reading) and between age groups (primary and secondary students)? and (8) to what extent does the magnitude of learning deficits vary across national contexts?

Below, we report our answers to each of these questions in turn. The questions correspond to the analysis plan set out in our pre-registered protocol (https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021249944), but we have adjusted the order and wording to aid readability. We had planned to examine gender differences in learning progress during the pandemic, but found there to be insufficient evidence to conduct this subgroup analysis, as the large majority of the identified studies do not provide evidence on learning deficits separately by gender. We also planned to examine how the magnitude of learning deficits differs across groups of students with varying exposures to school closures. This was not possible as the available data on school closures lack sufficient depth with respect to variation of school closures within countries, across grade levels and with respect to different modes of instruction, to meaningfully examine this association.

Results

The state of the evidence

Our systematic review identified 42 studies on learning progress during the COVID-19 pandemic that met our inclusion criteria. To be included in our systematic review and meta-analysis, studies had to use a measure of learning that can be standardized (using Cohen’s d) and base their estimates on empirical data collected since the onset of the COVID-19 pandemic (rather than making projections based on pre-COVID-19 data). As shown in Fig. 1, the initial literature search resulted in 5,153 hits after removal of duplicates. All studies were double screened by the first two authors. The formal database search process identified 15 eligible studies. We also hand searched relevant preprint repositories and policy databases. Further, to ensure that our study selection was as up to date as possible, we conducted two full forward and backward citation searches of all included studies on 15 February 2022, and on 8 August 2022. The citation and preprint hand searches allowed us to identify 27 additional eligible studies, resulting in a total of 42 studies. Most of these studies were published after the initial database search, which illustrates that the body of evidence continues to expand. Most studies provide multiple estimates of COVID-19 learning deficits, separately for maths and reading and for different school grades. The number of estimates (n = 291) is therefore larger than the number of included studies (n = 42).

Fig. 1
figure 1

Flow diagram of the study identification and selection process, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.

The geographic reach of evidence is limited

Table 1 presents all included studies and estimates of COVID-19 learning deficits (in brackets), grouped by the 15 countries represented: Australia, Belgium, Brazil, Colombia, Denmark, Germany, Italy, Mexico, the Netherlands, South Africa, Spain, Sweden, Switzerland, the UK and the United States. About half of the estimates (n = 149) are from the United States, 58 are from the UK, a further 70 are from other European countries and the remaining 14 estimates are from Australia, Brazil, Colombia, Mexico and South Africa. As this list shows, there is a strong over-representation of studies from high-income countries, a dearth of studies from middle-income countries and no studies from low-income countries. This skewed representation should be kept in mind when interpreting our synthesis of the existing evidence on COVID-19 learning deficits.

Table 1 Studies and estimates by country

The quality of evidence is mixed

We assessed the quality of the evidence using an adapted version of the Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) tool9. More specifically, we analysed the risk of bias of each estimate from confounding, sample selection, classification of treatments, missing data, the measurement of outcomes and the selection of reported results. A.M.B.-M. and B.A.B. performed the risk-of-bias assessments, which were independently checked by the respective other author. We then assigned each study an overall risk-of-bias rating (low, moderate, serious or critical) based on the estimate and domain with the highest risk of bias.

Figure 2a shows the distribution of all studies of COVID-19 learning deficits according to their risk-of-bias rating separately for each domain (top six rows), as well as the distribution of studies according to their overall risk of bias rating (bottom row). The overall risk of bias was considered ‘low’ for 15% of studies, ‘moderate’ for 30% of studies, ‘serious’ for 25% of studies and ‘critical’ for 30% of studies.

Fig. 2: Risk of bias and publication bias.
figure 2

a, Domain-specific and overall distribution of studies of COVID-19 learning deficits by risk of bias rating using ROBINS-I, including studies rated to be at critical risk of bias (n = 19 out of a total of n = 61 studies shown in this figure). In line with ROBINS-I guidance, studies rated to be at critical risk of bias were excluded from all analyses and other figures in this article and in the Supplementary Information (including b). b, z curve: distribution of the z scores of all estimates included in the meta-analysis (n = 291) to test for publication bias. The dotted line indicates z = 1.96 (P = 0.050), the conventional threshold for statistical significance. The overlaid curve shows a normal distribution. The absence of a spike in the distribution of the z scores just above the threshold for statistical significance and the absence of a slump just below it indicate the absence of evidence for publication bias.

In line with ROBINS-I guidance, we excluded studies rated to be at critical risk of bias (n = 19) from all of our analyses and figures, except for Fig. 2a, which visualizes the distribution of studies according to their risk of bias9. These are thus not part of the 42 studies included in our meta-analysis. Supplementary Table 2 provides an overview of these studies as well as the main potential sources of risk of bias. Moreover, in Supplementary Figs. 36, we replicate all our results excluding studies deemed to be at serious risk of bias.

As shown in Fig. 2a, common sources of potential bias were confounding, sample selection and missing data. Studies rated at risk of confounding typically compared only two timepoints, without accounting for longer time trends in learning progress. The main causes of selection bias were the use of convenience samples and insufficient consideration of self-selection by schools or students. Several studies found evidence of selection bias, often with students from a low socio-economic background or schools in deprived areas being under-represented after (as compared with before) the pandemic, but this was not always adjusted for. Some studies also reported a higher amount of missing data post-pandemic, again generally without adjustment, and several studies did not report any information on missing data. For an overview of the risk-of-bias ratings for each domain of each study, see Supplementary Fig. 1 and Supplementary Tables 1 and 2.

No evidence of publication bias

Publication bias can occur if authors self-censor to conform to theoretical expectations, or if journals favour statistically significant results. To mitigate this concern, we include not only published papers, but also preprints, working papers and policy reports.

Moreover, Fig. 2b tests for publication bias by showing the distribution of z-statistics for the effect size estimates of all identified studies. The dotted line indicates z = 1.96 (P = 0.050), the conventional threshold for statistical significance. The overlaid curve shows a normal distribution. If there was publication bias, we would expect a spike just above the threshold, and a slump just below it. There is no indication of this. Moreover, we do not find a left-skewed distribution of P values (see P curve in Supplementary Fig. 2a), or an association between estimates of learning deficits and their standard errors (see funnel plot in Supplementary Fig. 2b) that would suggest publication bias. Publication bias thus does not appear to be a major concern.

Having assessed the quality of the existing evidence, we now present the substantive results of our meta-analysis, focusing on the magnitude of COVID-19 learning deficits and on the variation in learning deficits over time, across different groups of students, and across country contexts.

Learning progress slowed substantially during the pandemic

Figure 3 shows the effect sizes that we extracted from each study (averaged across grades and learning subject) as well as the pooled effect size (red diamond). Effects are expressed in standard deviations, using Cohen’s d. Estimates are pooled using inverse variance weights. The pooled effect size across all studies is d = −0.14, t(41) = −7.30, two-tailed P = 0.000, 95% confidence interval (CI) −0.17 to −0.10. Under normal circumstances, students generally improve their performance by around 0.4 standard deviations per school year10,11,12. Thus, the overall effect of d = −0.14 suggests that students lost out on 0.14/0.4, or about 35%, of a school year’s worth of learning. On average, the learning progress of school-aged children has slowed substantially during the pandemic.

Fig. 3: Forest plot showing individual estimates by study (n = 42), averaged across subjects and grade levels, and the overall effect size estimate, pooled using inverse variance weights and a random-effects model.
figure 3

Effect sizes are expressed in standard deviations, using Cohen’s d, with 95% CI, and are sorted by magnitude.

Learning deficits arose early in the pandemic and persist

One may expect that children were able to recover learning that was lost early in the pandemic, after teachers and families had time to adjust to the new learning conditions and after structures for online learning and for recovering early learning deficits were set up. However, existing research on teacher strikes in Belgium13 and Argentina14, shortened school years in Germany15 and disruptions to education during World War II16 suggests that learning deficits are difficult to compensate and tend to persist in the long run.

Figure 4 plots the magnitude of estimated learning deficits (on the vertical axis) by the date of measurement (on the horizontal axis). The colour of the circles reflects the relevant country, the size of the circles indicates the sample size for a given estimate and the line displays a linear trend. The figure suggests that learning deficits opened up early in the pandemic and have neither closed nor substantially widened since then. We find no evidence that the slope coefficient is different from zero (βmonths = −0.00, t(41) = −7.30, two-tailed P = 0.097, 95% CI −0.01 to 0.00). This implies that efforts by children, parents, teachers and policy-makers to adjust to the changed circumstance have been successful in preventing further learning deficits but so far have been unable to reverse them. As shown in Supplementary Fig. 8, the pattern of persistent learning deficits also emerges within each of the three countries for which we have a relatively large number of estimates at different timepoints: the United States, the UK and the Netherlands. However, it is important to note that estimates of learning deficits are based on distinct samples of students. Future research should continue to follow the learning progress of cohorts of students in different countries to reveal how learning deficits of these cohorts have developed and continue to develop since the onset of the pandemic.

Fig. 4: Estimates of COVID-19 learning deficits (n = 291), by date of measurement.
figure 4

The horizontal axis displays the date on which learning progress was measured. The vertical axis displays estimated learning deficits, expressed in standard deviation (s.d.) using Cohen’s d. The colour of the circles reflects the respective country, the size of the circles indicates the sample size for a given estimate and the line displays a linear trend with a 95% CI. The trend line is estimated as a linear regression using ordinary least squares, with standard errors clustered at the study level (n = 42 clusters). βmonths = −0.00, t(41) = −7.30, two-tailed P = 0.097, 95% CI −0.01 to 0.00.

Socio-economic inequality in education increased

Existing research on the development of learning gaps during summer vacations17,18, disruptions to schooling during the Ebola outbreak in Sierra Leone and Guinea19, and the 2005 earthquake in Pakistan20 shows that the suspension of face-to-face teaching can increase educational inequality between children from different socio-economic backgrounds. Learning deficits during the COVID-19 pandemic are likely to have been particularly pronounced for children from low socio-economic backgrounds. These children have been more affected by school closures than children from more advantaged backgrounds21. Moreover, they are likely to be disadvantaged with respect to their access and ability to use digital learning technology, the quality of their home learning environment, the learning support they receive from teachers and parents, and their ability to study autonomously22,23,24.

Most studies we identify examine changes in socio-economic inequality during the pandemic, attesting to the importance of the issue. As studies use different measures of socio-economic background (for example, parental income, parental education, free school meal eligibility or neighbourhood disadvantage), pooling the estimates is not possible. Instead, we code all estimates according to whether they indicate a reduction, no change or an increase in learning inequality during the pandemic. Figure 5 displays this information. Estimates that indicate an increase in inequality are shown on the right, those that indicate a decrease on the left and those that suggest no change in the middle. Squares represent estimates of changes in inequality during the pandemic in reading performance, and circles represent estimates of changes in inequality in maths performance. The shading represents when in the pandemic educational inequality was measured, differentiating between the first, second and third year of the pandemic. Estimates are also arranged horizontally by grade level. A large majority of estimates indicate an increase in educational inequality between children from different socio-economic backgrounds. This holds for both maths and reading, across primary and secondary education, at each stage of the pandemic, and independently of how socio-economic background is measured.

Fig. 5: Harvest plot summarizing the evidence on changes in educational inequality between students from different socio-economic backgrounds during the pandemic.
figure 5

Each circle/square refers to one estimate of over-time change in inequality in maths/reading performance (n = 211). Estimates that find a decrease/no change/increase in inequality are grouped on the left/middle/right. Within these categories, estimates are ordered horizontally by school grade. The shading indicates when in the pandemic a given measure was taken.

Learning deficits are larger in maths than in reading

Available research on summer learning deficits17,25, student absenteeism26,27 and extreme weather events28 suggests that learning progress in mathematics is more dependent on formal instruction than in reading. This might be due to parents being better equipped to help their children with reading, and children advancing their reading skills (but not their maths skills) when reading for enjoyment outside of school. Figure 6a shows that, similarly to earlier disruptions to learning, the estimated learning deficits during the COVID-19 pandemic are larger for maths than for reading (mean difference δ = −0.07, t(41) = −4.02, two-tailed P = 0.000, 95% CI −0.11 to −0.04). This difference is statistically significant and robust to dropping estimates from individual countries (Supplementary Fig. 9).

Fig. 6: Variation in estimates of COVID-19 learning deficits (n = 291) across different characteristics.
figure 6

Each plot shows the distribution of COVID-19 learning deficit estimates for the respective subgroup, with the box marking the interquartile range and the white circle denoting the median. Whiskers mark upper and lower adjacent values: the furthest observation within 1.5 interquartile range of either side of the box. a, Learning subject (reading versus maths). Median: reading −0.09, maths −0.18. Interquartile range: reading −0.15 to −0.02, maths −0.23 to −0.09. b, Level of education (primary versus secondary). Median: primary −0.12, secondary −0.12. Interquartile range: primary −0.19 to −0.05, secondary −0.21 to −0.06. c, Country income level (high versus middle). Median: high −0.12, middle −0.37. Interquartile range: high −0.20 to −0.05, middle −0.65 to −0.30.

No evidence of variation across grade levels

One may expect learning deficits to be smaller for older than for younger children, as older children may be more autonomous in their learning and better able to cope with a sudden change in their learning environment. However, older students were subject to longer school closures in some countries, such as Denmark29, based partly on the assumption that they would be better able to learn from home. This may have offset any advantage that older children would otherwise have had in learning remotely.

Figure 6b shows the distribution of estimates of learning deficits for students at the primary and secondary level, respectively. Our analysis yields no evidence of variation in learning deficits across grade levels (mean difference δ = −0.01, t(41) = −0.59, two-tailed P = 0.556, 95% CI −0.06 to 0.03). Due to the limited number of available estimates of learning deficits, we cannot be certain about whether learning deficits differ between primary and secondary students or not.

Learning deficits are larger in poorer countries

Low- and middle-income countries were already struggling with a learning crisis before the pandemic. Despite large expansions of the proportion of children in school, children in low- and middle-income countries still perform poorly by international standards, and inequality in learning remains high30,31,32. The pandemic is likely to deepen this learning crisis and to undo past progress. Schools in low- and middle-income countries have not only been closed for longer, but have also had fewer resources to facilitate remote learning33,34. Moreover, the economic resources, availability of digital learning equipment and ability of children, parents, teachers and governments to support learning from home are likely to be lower in low- and middle-income countries35.

As discussed above, most evidence on COVID-19 learning deficits comes from high-income countries. We found no studies on low-income countries that met our inclusion criteria, and evidence from middle-income countries is limited to Brazil, Colombia, Mexico and South Africa. Figure 6c groups the estimates of COVID-19 learning deficits in these four middle-income countries together (on the right) and compares them with estimates from high-income countries (on the left). The learning deficit is appreciably larger in middle-income countries than in high-income countries (mean difference δ = −0.29, t(41) = −2.78, two-tailed P = 0.008, 95% CI −0.50 to −0.08). In fact, the three largest estimates of learning deficits in our sample are from middle-income countries (Fig. 3)36,37,38.

Discussion

Two years since the COVID-19 pandemic, there is a growing number of studies examining the learning progress of school-aged children during the pandemic. This paper first systematically reviews the existing literature on learning progress of school-aged children during the pandemic and appraises its geographic reach and quality. Second, it harmonizes, synthesizes and meta-analyses the existing evidence to examine the extent to which learning progress has changed since the onset of the pandemic, and how it varies across different groups of students and across country contexts.

Our meta-analysis suggests that learning progress has slowed substantially during the COVID-19 pandemic. The pooled effect size of d = −0.14, implies that students lost out on about 35% of a normal school year’s worth of learning. This confirms initial concerns that substantial learning deficits would arise during the pandemic10,39,40. But our results also suggest that fears of an accumulation of learning deficits as the pandemic continues have not materialized41,42. On average, learning deficits emerged early in the pandemic and have neither closed nor widened substantially. Future research should continue to follow the learning progress of cohorts of students in different countries to reveal how learning deficits of these cohorts have developed and continue to develop since the onset of the pandemic.

Most studies that we identify find that learning deficits have been largest for children from disadvantaged socio-economic backgrounds. This holds across different timepoints during the pandemic, countries, grade levels and learning subjects, and independently of how socio-economic background is measured. It suggests that the pandemic has exacerbated educational inequalities between children from different socio-economic backgrounds, which were already large before the pandemic43,44. Policy initiatives to compensate learning deficits need to prioritize support for children from low socio-economic backgrounds in order to allow them to recover the learning they lost during the pandemic.

There is a need for future research to assess how the COVID-19 pandemic has affected gender inequality in education. So far, there is very little evidence on this issue. The large majority of the studies that we identify do not examine learning deficits separately by gender.

Comparing estimates of learning deficits across subjects, we find that learning deficits tend to be larger in maths than in reading. As noted above, this may be due to the fact that parents and children have been in a better position to compensate school-based learning in reading by reading at home. Accordingly, there are grounds for policy initiatives to prioritize the compensation of learning deficits in maths and other science subjects.

A limitation of this study and the existing body of evidence on learning progress during the COVID-19 pandemic is that the existing studies primarily focus on high-income countries, while there is a dearth of evidence from low- and middle-income countries. This is particularly concerning because the small number of existing studies from middle-income countries suggest that learning deficits have been particularly severe in these countries. Learning deficits are likely to be even larger in low-income countries, considering that these countries already faced a learning crisis before the pandemic, generally implemented longer school closures, and were under-resourced and ill-equipped to facilitate remote learning32,33,34,35,45. It is critical that this evidence gap on low- and middle-income countries is addressed swiftly, and that the infrastructure to collect and share data on educational performance in middle- and low-income countries is strengthened. Collecting and making available these data is a key prerequisite for fully understanding how learning progress and related outcomes have changed since the onset of the pandemic46.

A further limitation is that about half of the studies that we identify are rated as having a serious or critical risk of bias. We seek to limit the risk of bias in our results by excluding all studies rated to be at critical risk of bias from all of our analyses. Moreover, in Supplementary Figs. 36, we show that our results are robust to further excluding studies deemed to be at serious risk of bias. Future studies should minimize risk of bias in estimating learning deficits by employing research designs that appropriately account for common sources of bias. These include a lack of accounting for secular time trends, non-representative samples and imbalances between treatment and comparison groups.

The persistence of learning deficits two and a half years into the pandemic highlights the need for well-designed, well-resourced and decisive policy initiatives to recover learning deficits. Policy-makers, schools and families will need to identify and realize opportunities to complement and expand on regular school-based learning. Experimental evidence from low- and middle-income countries suggests that even relatively low-tech and low-cost learning interventions can have substantial, positive effects on students’ learning progress in the context of remote learning. For example, sending SMS messages with numeracy problems accompanied by short phone calls was found to lead to substantial learning gains in numeracy in Botswana47. Sending motivational text messages successfully limited learning losses in maths and Portuguese in Brazil48.

More evidence is needed to assess the effectiveness of other interventions for limiting or recovering learning deficits. Potential avenues include the use of the often extensive summer holidays to offer summer schools and learning camps, extending school days and school weeks, and organizing and scaling up tutoring programmes. Further potential lies in developing, advertising and providing access to learning apps, online learning platforms or educational TV programmes that are free at the point of use. Many countries have already begun investing substantial resources to capitalize on some of these opportunities. If these interventions prove effective, and if the momentum of existing policy efforts is maintained and expanded, the disruptions to learning during the pandemic may be a window of opportunity to improve the education afforded to children.

Methods

Eligibility criteria

We consider all types of primary research, including peer-reviewed publications, preprints, working papers and reports, for inclusion. To be eligible for inclusion, studies have to measure learning progress using test scores that can be standardized across studies using Cohen’s d. Moreover, studies have to be in English, Danish, Dutch, French, German, Norwegian, Spanish or Swedish.

Search strategy and study identification

We identified relevant studies using the following steps. First, we developed a Boolean search string defining the population (school-aged children), exposure (the COVID-19 pandemic) and outcomes of interest (learning progress). The full search string can be found in Section 1.1 of Supplementary Information. Second, we used this string to search the following academic databases: Coronavirus Research Database, the Education Resources Information Centre, International Bibliography of the Social Sciences, Politics Collection (PAIS index, policy file index, political science database and worldwide political science abstracts), Social Science Database, Sociology Collection (applied social science index and abstracts, sociological abstracts and sociology database), Cumulative Index to Nursing and Allied Health Literature, and Web of Science. Second, we hand-searched multiple preprint and working paper repositories (Social Science Research Network, Munich Personal RePEc Archive, IZA, National Bureau of Economic Research, OSF Preprints, PsyArXiv, SocArXiv and EdArXiv) and relevant policy websites, including the websites of the Organization for Economic Co-operation and Development, the United Nations, the World Bank and the Education Endowment Foundation. Third, we periodically posted our protocol via Twitter in order to crowdsource additional relevant studies not identified through the search. All titles and abstracts identified in our search were double-screened using the Rayyan online application49. Our initial search was conducted on 27 April 2021, and we conducted two forward and backward citation searches of all eligible studies identified in the above steps, on 14 February 2022, and on 8 August 2022, to ensure that our analysis includes recent relevant research.

Data extraction

From the studies that meet our inclusion criteria we extracted all estimates of learning deficits during the pandemic, separately for maths and reading and for different school grades. We also extracted the corresponding sample size, standard error, date(s) of measurement, author name(s) and country. Last, we recorded whether studies differentiate between children’s socio-economic background, which measure is used to this end and whether studies find an increase, decrease or no change in learning inequality. We contacted study authors if any of the above information was missing in the study. Data extraction was performed by B.A.B. and validated independently by A.M.B.-M., with discrepancies resolved through discussion and by conferring with P.E.

Measurement and standardizationr

We standardize all estimates of learning deficits during the pandemic using Cohen’s d, which expresses effect sizes in terms of standard deviations. Cohen’s d is calculated as the difference in the mean learning gain in a given subject (maths or reading) over two comparable periods before and after the onset of the pandemic, divided by the pooled standard deviation of learning progress in this subject:

$$d=\frac{{\bar{x}}_{1}-{\bar{x}}_{2}}{s},$$

where

$$s=\sqrt{\frac{\left({s}_{1}^{2}+{s}_{2}^{2}\right)}{2}}.$$

Effect sizes expressed as β coefficients are converted to Cohen’s d:

$$d=\frac{\beta }{se}\times \sqrt{\frac{1}{{n}_{1}}+\frac{1}{{n}_{2}}}.$$

Subject

We use a binary indicator for whether the study outcome is maths or reading. One study does not differentiate the outcome but includes a composite of maths and reading scores50.

Level of education

We distinguish between primary and secondary education. We first consulted the original studies for this information. Where this was not stated in a given study, students’ age was used in conjunction with information about education systems from external sources to determine the level of education51.

Country income level

We follow the World Bank’s classification of countries into four income groups: low, lower-middle, upper-middle and high income. Four countries in our sample are in the upper-middle-income group: Brazil, Colombia, Mexico and South Africa. All other countries are in the high-income group.

Data synthesis

We synthesize our data using three synthesis techniques. First, we generate a forest plot, based on all available estimates of learning progress during the pandemic. We pool estimates using a random-effects restricted maximum likelihood model and inverse variance weights to calculate an overall effect size (Fig. 3)52. Second, we code all estimates of changes in educational inequality between children from different socio-economic backgrounds during the pandemic, according to whether they indicate an increase, a decrease or no change in educational inequality. We visualize the resulting distribution using a harvest plot (Fig. 5)53. Third, given that the limited amount of available evidence precludes multivariate or causal analyses, we examine the bivariate association between COVID-19 learning deficits and the months in which learning was measured using a scatter plot (Fig. 4), and the bivariate association between COVID-19 learning deficits and subject, grade level and countries’ income level, using a series of violin plots (Fig. 6). The reported estimates, CIs and statistical significance tests of these bivariate associations are based on common-effects models with standard errors clustered by study, and two-sided tests. With respect to statistical tests reported, the data distribution was assumed to be normal, but this was not formally tested. The distribution of estimates of learning deficits is shown separately for the different moderator categories in Fig. 6.

Pre-registration

We prospectively registered a protocol of our systematic review and meta-analysis in the International Prospective Register of Systematic Reviews (CRD42021249944) on 19 April 2021 (https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021249944).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.