The Impact of COVID-19 on Education in Latin America: Long-Run Implications for Poverty and Inequality

3.1 Methodology and Data

We consider both enrollment and net enrollment in primary, secondary or tertiary education. While enrollment takes into account all of students enrolled in a particular level of education regardless of age, net enrollment focuses on the appropriate age group for each educational level, excluding students who are above or below the expected age. Also, we analyze enrollment at private schools—as opposed to public schools—conditional on being enrolled.

3.2 Change in Educational Enrollment

Figure 2 summarizes the results from estimating Equation (3.1) for the pooled sample of 12 countries, considering enrollment at the educational system without distinguishing educational levels. The figure shows by how much the estimated enrollment in 2020 differs from the trend in enrollment rates of the previous decade for the entire region.5 Since we do not have surveys for each year in each country, we group survey years such that in every time-period every country is included. The point estimates show the (population-unweighted) average estimate across all countries, controlling for sex, age, urban or rural place of residence, and parental education. The counterfactual estimate for the enrollment rates in 2020 is computed by extrapolation, following the trend of the previous periods. From Figure 2 it is clear that the actual enrollment rates in 2020 are substantially lower than the counterfactual.

Table 2 shows the coefficient estimates of the dummy for the year 2020—that is, the estimated $$$ \beta $$$ in Equation (3.1)—for each country, controlling for the characteristics mentioned above. The size of the coefficient indicates the deviation in percentage points from the overall trend, conditional on the included controls. The last row shows the average value of the dependent variable over the years—that is, 2009–2020. Panel A of the table shows the results for enrollment, Panel B for net enrollment, Panel C for enrollment at private schools.

In all countries enrollment rates negatively deviate from the trend in 2020. In most countries, this change is statistically significant. On average, measured over the 12 Latin American countries in our sample, enrollment of children and young people between 6 and 24 years old fell by around two percentage points. The decrease in net enrollment and enrollment at private schools is also in this order of magnitude in absolute terms, but higher in relative terms because of the lower average baseline level. A decrease in enrollment by two percentage points means that a large number of Latin American children and young people—around three million—dropped out of school or did not enroll in the higher track due to the pandemic in 2020. The strongest decline in enrollment is found in Chile and Peru—around 5 percentage points.

Figure 3 shows the results for different age ranges: 6–11, the typical primary-school age range; 12–17, the corresponding age range for secondary education; and 18–24, the age range in which individuals are usually enrolled in tertiary education.6 On average, in 2020 there is a decrease in enrollment rates for all three age-groups. The strongest impact occurred among young adults aged 18–24 (almost 3 percentage points on average). However, in eight of the 12 countries in our sample, enrollment decreases substantially also in primary education. The lowest decline is found in secondary education.7 Tables A2 and A3 in the Appendix show the coefficient estimates separately by age groups for net enrollment and enrollment in private education, respectively. Interestingly, the decrease in private education enrollment is particularly strong at the tertiary education level (age range 18–24), which indeed is the educational level with the highest concentration of students enrolled at private institutions in most countries. Since we do not observe a significant increase in college graduation in 2020 with respect to previous years (see Figure A1 in the Appendix), we conclude that the drop in enrollment observed during the pandemic is driven by an increase in dropout.

3.3 Change in Enrollment Across Subgroups

Figures A2–A4 in the Appendix show the effect of the pandemic on enrollment in Latin America for different population groups: men vs women; urban vs rural areas; low vs high parental education. For this analysis, we estimate Equation (3.1) for each group separately based on the pooled sample with normalized survey weights. In all cases, the estimates suggest a negative deviation from the trend in enrollment in 2020. However, in some cases, the effect of the pandemic on different sub-groups seems more heterogeneous. For instance, for households in rural areas we cannot reject that the change is not different from zero.

Enrollment rates in 2020 negatively deviate from the trend for both men and women. While the effect on enrollment and net enrollment is higher for men, the effect on enrollment in private education is higher, in relative terms, among women. However, these differences between genders are not statistically significant. A similar pattern emerges in the difference between urban and rural place of residence. The decrease in enrollment and net enrollment is sharper in urban areas, while not significantly different from zero in rural places. In contrast, the relative decrease in enrollment in private education is stronger in rural than in urban areas. In this case, the differences between groups are statistically significant.

We also assess heterogeneous effects on enrollment by parental education. We compare children of parents with a low level of education—that is, parents with a secondary education degree or less—and children of parents with a high level of education—that is, parents with more than a secondary education degree.8 In the two groups of children the decrease in enrollment (both gross and net) is similar, substantial, and statistically significant. Enrollment in private education is hereby an important exception: for this outcome, the relative decrease is stronger among the children of low-educated parents.

4.1 Methodology and Data

School closures imply a severe interruption in the learning process, potentially leading to profound implications for the formation of both cognitive and non-cognitive skills in children, with subsequent effects on their future productivity and earnings. Furthermore, in line with Cunha and Heckman (2007) and Cunha et al. (2010), these effects would be highly asymmetric depending on the age of children and their family's socioeconomic level, thereby affecting long-term poverty and income inequality.

Unfortunately, it is extremely difficult to estimate the long run impact of the school closures due to COVID-19 on the human capital stock given the proximity of the shock and the data at hand. NLT provide first estimates on the long run effect of school closures based on two main simplifying assumptions: (i) the human capital stock, which is relevant as a determinant of future earnings, is well approximated by years of formal education, and (ii) the loss of human capital can be approximated by the share of the school year in which the learning process was interrupted. While the first assumption is standard in part of the literature (e.g., Black & Devereux, 2011), the second follows the contributions of Abadzie et al. (Abadzi, 2009) and Adda (2016).9 The human capital loss could be lower if successful compensating measures are taken in the future, but on the other hand, they could be larger if interruptions make the process of learning more difficult to resume and lead to cumulative learning losses over time.10

In NLT, A is defined as the likelihood to have access to the internet—a key factor to be able to receive on-line education—of students in households with a given educational background. In contrast to NLT, where this likelihood is predicted for each individual based on the distribution of access to the internet by education groups estimated for each country from other data sources, in our specification A is the actual availability of internet in the household, a piece of information available in most national household surveys.11

4.2 Predicted Educational Losses

In this section we illustrate the results for the 12 Latin American countries analyzed in the previous section plus Uruguay.15 For the sake of simplicity, in most cases we present the unweighted cross-country average of the results—that is, we calculate the results for each country and then take the average across countries.16 The cross-country average represents the typical Latin American country, which we refer to as “Latin America.”

On average, schools in the region remained closed for a total of 269 days in 2020 and 2021 due to the COVID-19 pandemic. Considering that a standard school year in Latin America consists of 189 days, the cumulative school closures amounted to 1.42 years. The ratio of school closure days to the total number of days in a typical school year ranges from 0.84 in Uruguay to over 1.7 in Ecuador and Mexico (Figure 4).

Table 3 shows the summary statistics of k, that is, the share of a year of schooling lost due to the COVID-19 pandemic when adjusting for government and family reactions. The mean value for Latin America is 0.59: more than half of a year was lost due to the school closures even when considering compensatory measures. There is considerable heterogeneity across countries: from 0.32 in Uruguay to 0.85 in Ecuador (Figure 5).17

The value of k is not homogeneous as students differ in the access to internet, parental background, and household size. Table 3 reports the value of k by parental education and household per capita income deciles for the aggregate of Latin America. The education loss is strongly decreasing in the socioeconomic situation of the household. While the mean loss for a student in the bottom decile of the income distribution is estimated in 81% of a schooling year, the loss becomes 22 percent for a student in the top decile.18

4.3 Remote Learning with the World Bank's High-Frequency Phone Surveys

The World Bank's High-Frequency Phone Survey (HFPS) is one of the most ambitious data sets collected immediately after the onset of the pandemic. The surveys were carried out in over 100 countries, including 13 in Latin America. They offer real-time information on the situation and behavior of households during the first months of the pandemic, including information on remote learning strategies followed during school closures, such as online sessions and assigned homework.19

Using microdata from the first three rounds of these surveys—carried out between May and August 2020—, Bracco et al. (2022) show the estimated probability of children engaging in different remote learning activities by parental education. Unfortunately, due to the small samples, estimates are available only for the pool of Latin American countries.

5.1 Methodology

We start from an income distribution in baseline year t_o. We choose 2019, the most recent year with available microdata from a national household survey for most countries. Data for 2021 were not yet available for most countries at the time of writing this paper, and 2020 was a very unusual year to use it as baseline.

For simplicity, and to focus only on the impact of the pandemic through the educational channel, we assume that without the COVID-19 pandemic the educational structure and income distribution T years after t_o (t_T) would be identical to those in the baseline year. For simplicity in the explanation, we focus on year 2045, that is, 25 years after the COVID-19 shock. In Section 5.3 we extend the analysis to the dynamics of the impact over time. Also, in the interest of simplicity, we initially assume that the pandemic in 2020 affected human capital accumulation of children and youths aged 5–20. Consequently, in year 2045 the workers who were affected by the pandemic back in 2020 are those aged 30–45. This will be our “treatment group” G.20

According to the discussion in previous sections, we carry out two simulations: (1) we subtract months of instructional loss to all adults in G due to school closures, following the methodology explained in Section 4, and (2) we subtract years of education to dropouts associated to the pandemic according to the results in Section 3.

5.2 Results

The average results for Latin America are presented in Table 4. For each simulation the table shows average labor income (in PPP USD),23 average household per capita income (in PPP USD), the Gini coefficient for the distribution of per capita income, and three poverty measures computed using the 6.85-USD-a-day line: the poverty headcount ratio, the poverty gap and the severity index (FGT(2)). In most panels there are three sets of indicators computed for (i) the group pf dropouts, (ii) the cohort aged 30–45 in the base year, and (iii) the whole population.

The first column (“original”) shows values for 2045 without the COVID-19 shock, which, by construction, are those for the year of the latest available household survey (2019). The second column displays the results for the drop-out exercise. Since we randomly select the dropouts (within the groups), in column (ii) we show the mean results over 50 draws. The next two columns show the results for the exercise of instructional losses due to school closures under two extreme alternatives: assuming no adjustments to the loss of days of school (f=n=0 and α = 1), and assuming parental and government adjustments (f and n >0, and α < 1). Columns (v) and (vi) present the combined effects of the two exercises (dropouts and school closures). The second panel of the table—columns vii to xi—shows proportional changes in relation to the original situation for each indicator.24

The impact of leaving school is strong over the group of dropouts. Mean income falls by 17.6 percent, the incidence of poverty climbs by 22.7 percent and the severity of poverty jumps by 38.5 percent. Since the proportion of dropouts is relatively small (around 1 percent) the size of the impact becomes small for the treated cohort 30–45, and tiny for the whole population. The poverty headcount ratio increases by 0.6 percent in the cohort 30–45 and by 0.2 percent in the population due to the drop-out effect.

The impacts are larger due to the generalized instructional losses caused by the school closures. If reactions to school closures were inexistent or fully ineffective (columns iii and viii), the COVID-19 shock would imply a fall by around 15 percent in earnings in the treatment group in 2045. The impact would be a decline by 10.5 percent of average household per capita income. These income changes have significant effects over the poverty indicators within this group: the poverty headcount ratio would increase by 19.3 percent, the poverty gap by 20.2 percent and the severity index by 20.2 percent. Values are smaller (around 8 percent) when considering the impact on the entire population.

Interestingly, the Gini on the distribution of labor income for the treated group decreases. This is not an unexpected result. It is driven by the combination of (i) the subtraction of a similar number of days of schooling to all individuals, and (ii) the convexity of the earnings-education profile. This fall in inequality is the other side of the “paradox of progress” (Alejo et al., 2022; Bourguignon et al., 2004). Unlike inequality in labor income, the Gini for the distribution of household per capita income remains roughly unchanged, likely due to a compensating factor: labor income is a less important source of income in the top of the distribution than in the bottom, due to the role of capital income.

Column (iv) reveals that the government and parental reactions could have had a partial ameliorating effect on the impact of the COVID-19 shock. For instance, the poverty gap in the treated group increases only by 8.4 percent instead of 20.2 percent if both government and parents effectively reacted to the school closures. Consistent with the fact that reactions to school closures were asymmetric across households depending on the socioeconomic status, we find a (very minor) increase in inequality in the full adjustment alternative, over both the labor and the per capita income distributions.

The last columns in the table show the combined effects. The COVID-19 pandemic implies a substantial increase in poverty among school dropouts (between 22.7 percent and 38.5 percent depending on the indicator). For the treated cohort 30–45 and for the entire population the results are mostly driven by the instructional losses due to the school closures. Poverty would increase between 8.4 percent and 20.7 percent in the shocked cohort, depending on the indicator and on the effectiveness of the reactions of governments and families during the pandemic to compensate for the school closures. The increase in income poverty for the entire population would be in the range between 3.4 percent and 8.1 percent. The sign of the changes in income inequality depends on the parental adjustments: unequalizing with full adjustment and equalizing with no adjustments. In any case, changes in income inequality would be almost negligible.25

5.3 Dynamics

The previous exercises were focused on a particular year in the future (2045), when all the students in 2020 are already young adults. In this section we extend the analysis to the period 2021–2075 and examine the dynamics of the impact. The main results are presented in Figures 6-8. The figures show the trajectories over time of household per capita income, inequality, and poverty after the COVID-19 shock on education under two alternative assumptions: no adjustment and full adjustment by families and government.

Figure 6 reveals an initial larger drop in income in 2021 after the shock, as a result of the loss in human capital for the group of young workers affected by the pandemic. This group is, however, small since in 2021 individuals in the treatment group are aged 6–21, and hence few of them are active in the labor market. The impact of the shock on education during the pandemic grows over time as the treatment cohort becomes older and enters the labor market. However, at some point this generation starts to retire and then the effect vanishes away.26 The impact of the COVID-19 shock on incomes would reach its maximum value in 2045. Figure 7 shows similar patterns for two income poverty indicators: the headcount ratio (panel a) and the severity index (panel b).

The dynamics in the case of inequality are rather different (Figure 8). Initially, as the shock affects young workers, who are typically poorer than the rest, the impact is unequalizing. As times passes the treatment group becomes older and more affluent, and then the COVID-19 shock turns equalizing. In any case, changes are very small, and become even smaller in the case of full adjustment.

5.4 Comparison with Fuchs-Schündeln et al. (2022)

As explained above, to obtain these estimates we employed microsimulations and mainly relied on the approach developed by Neidhöfer et al. (2021) (NLT) to simulate the impact of the pandemic on educational attainments. A complementary analysis that estimates the long-term distributional impact of COVID-19 school closures is Fuchs-Schündeln et al. (2022) (FKLP), who employ a structural life-cycle model and calibrate their model using US data. Similar to our analysis and NLT, they consider two main inputs: parental and governmental investment in education. While we capture parental investments through their correlation with parental education without specifying different channels, FKLP consider monetary investment and time investment separately in the model. They show that both monetary and time investment are positively correlated with parental education. Another difference between the two approaches is that FKLP consider differential effects of school closures at different ages by incorporating dynamic complementarities in human capital production, while NLT assume that the strength of the impact does not vary with age and show, by robustness checks, that the results are not sensitive to this assumption.

FKLP initially assume school closures of 1 year and additionally analyse the effects of remote learning as a mitigation measure, assuming that it is able to mitigate half a year of schooling loss. Following NLT, we instead measure the length of school closures and compensatory measures empirically for each country in the sample based on various indicators and data sources. However, the results from both studies show a similar pattern: substantial long-term implications for affected children, with a stronger impact on disadvantaged children. FKLP highlight that all children leave the pandemic with reduced human capital, while in the approach of NLT highly educated parents completely substitute schooling during closures, which seems consistent with existing empirical evidence of minimal impacts on children of highly educated parents (e.g., Betthäuser et al., 2023). In other estimations, FKLP also assume perfect substitutability between parents and schooling, leading to weaker average long-run welfare losses but stronger distributional effects. In conclusion, although the models and methodologies employed in this paper and in FKLP differ, the main assumptions and findings on the long-run impacts of COVID-19 school closures are generally consistent.