The impacts of remote learning in secondary education during the pandemic in Brazil

Before COVID-19, middle-income countries had expanded access to basic education but still faced a learning crisis: many students could not read age-appropriate texts and large shares finished high school without minimum skills. COVID-19 school closures threatened to worsen learning and disrupt enrolment, especially in settings with limited internet access, constrained home study environments, and reduced parental support. With over 1.6 billion children affected globally, quantifying learning losses in primary and secondary education is urgent to inform decisions on reopening schools, even in contexts with rising vaccination coverage. Prior evidence largely comes from high-income countries or from settings where remote learning expanded access relative to no schooling, limiting its relevance for middle-income contexts like Brazil. This study asks: What were the causal impacts of remote learning on dropout risk and standardized test performance among secondary students in São Paulo State during the pandemic, and how did partial school reopening affect outcomes? The study is important to separate the effects of remote instruction from other pandemic shocks and to inform policies on safe school reopening and remedial strategies.

Existing studies often focus on high-income countries, higher education, or selected student populations (e.g., charter or online schools), limiting generalizability. Evidence from middle-income countries commonly evaluates remote instruction against a counterfactual of no schooling, not against in-person classes. Pandemic-era analyses frequently rely on simulations or non-comparable tests and cohorts, conflating remote learning with other COVID-19 effects. Even credible strategies (e.g., variation in recess length or prior epidemics) do not directly map to COVID-19 instruction mode changes. Prior credible evidence for secondary education during COVID-19 mainly covers high-income contexts, leaving gaps on dropout effects and heterogeneity by age, gender, race, and socioeconomic status. This paper contributes by leveraging natural experiments in São Paulo State with consistent testing scales, exploiting the within-year switch from in-person to remote classes while keeping exams remote across quarters in 2020, and by using staggered reopening to estimate ITT effects. It also parses out local COVID-19 disease activity and documents heterogeneity (larger losses among girls, non-white students, poorer neighborhoods, and schools without prior online activity).

Setting and data: São Paulo State administered quarterly standardized tests (AAPs) throughout 2020 on the same scale as prior years. Classes were in-person in Q1 2020 and remote in Q2–Q4; exams were remote in all 2020 quarters. Data include quarterly attendance, math and Portuguese scorecard grades, and standardized test scores for all 6th–12th graders in 2018–2020. Focusing on 2019–2020 yields 4,719,696 observations for middle school and 3,791,024 for high school; 83.3% have standardized test scores. Outcome definitions: High dropout risk equals 1 if a student had no math and no Portuguese grades recorded in a quarter, 0 otherwise; this proxy is validated against pre-pandemic dropouts and predicts non-attendance in early 2021. Standardized test scores average math and Portuguese test scores each quarter (Q4 2020 provides only overall standardized scores). Empirical strategies: (1) Differences-in-differences (DiD) to estimate remote learning impacts by contrasting within-year changes between Q1 and Q4 in 2020 versus 2019, absorbing grade fixed effects and holding exam mode constant within 2020. Naive cross-year comparisons (Q4 2020 vs Q4 2019; and a DiD using 2018–2019 as baseline) are shown for benchmark but conflate other 2020 changes. OLS is used with standard errors clustered at the school level. (2) Selection adjustment for test-score analysis through propensity score methods: within each grade and quarter, estimate the propensity to take the exam based on student and school characteristics; include a cubic polynomial of the propensity in regressions and re-weight observations by the inverse propensity to reflect the universe of students. (3) Heterogeneity: allow treatment effects to vary non-parametrically with municipal per-capita COVID-19 cases/deaths after residualizing outcomes and cases for student/school covariates; estimate heterogeneity by age, gender, race, neighborhood income, and prior online activity. (4) School reopening effects: use municipal decrees authorizing reopening (from October 2020; high school in-person classes from November; middle school remained remote) to estimate ITT effects via a triple-differences design: differences between middle- and high-school students within municipalities that authorized reopening versus those that did not, before and after reopening. OLS is used with standard errors clustered at the municipality level for these analyses. All hypothesis tests are two-sided. Ethics: approved by the University of Zurich IRB (2020-079). Analyses were conducted within the education secretariat’s secure environment; only aggregate outputs were exported.

- Remote learning markedly increased dropout risk and reduced learning: In DiD estimates comparing Q1 to Q4 within year (2020 vs 2019), high dropout risk rose by about 0.0621 (SE 0.0002), a 365% increase relative to the Q4 2019 mean of 0.017 (P < 0.001). This proxy suggests actual dropouts could have risen from ~10% to ~35% during remote learning. - Standardized test scores fell substantially under remote learning: Naive cross-year comparisons misleadingly suggest increases (e.g., +0.652 s.d.), driven by selection and exam-mode differences. The DiD approach reveals losses of approximately −0.32 s.d. (SE 0.0001; P < 0.001), implying students learned only 27.5% of the in-person equivalent (a 72.5% setback). Given ~35 weeks of closures, losses average ~0.009 s.d. per week. When expressed in percentile terms (relative to baseline), losses correspond to ~22–25 percentile points over a year. - Heterogeneity: Losses in test scores were broadly uniform across grades (≥60% of in-person learning lost across all grades). Dropout risk increased by ≥300% across grades (percentage terms), except for high-school seniors where the relative increase was smaller due to higher baseline risk. Losses were larger for girls and non-white students, schools in poorer neighborhoods, and schools without prior online academic activities. Subject-specific effects were worse in math (students learned ~20% of in-person equivalent by Q3 2020) than in Portuguese (~40%). - Role of local disease activity: Allowing effects to vary with municipal COVID-19 cases/deaths does not materially change conclusions. Learning losses did not systematically vary with disease activity; dropout risk showed slight, non-significant increases with higher local activity, but effects remained very large even at the low end (≥247% increase). - Effects of reopening: Municipal authorization to reopen in Q4 2020 led to positive ITT effects on high-school standardized test scores (approximately 20% improvement relative to control), while optional in-person activities for middle school did not improve their outcomes. Reopening did not mitigate the substantial increase in dropout risk.

The within-cohort, within-year DiD design isolates the impact of remote instruction from other pandemic-related factors and exam-mode differences. Results show remote learning caused dramatic increases in dropout risk and sizable learning losses, much larger than suggested by naive cross-year comparisons. The lack of systematic variation in losses with local COVID-19 activity supports the interpretation that instructional mode is the primary driver of educational impacts rather than local health shocks. Heterogeneity patterns indicate that remote learning exacerbated pre-existing inequalities, disproportionately affecting girls, non-white students, and those in poorer areas or schools lacking prior online capacity. The triple-differences ITT evidence shows that resuming in-person high-school classes, even partially and late in the year, improved learning outcomes relative to continued remote instruction, underscoring the educational benefits of reopening. However, reopening did not offset heightened dropout risk, suggesting the need for targeted engagement and retention strategies alongside instructional restoration.

This study provides causal evidence from a large middle-income setting that remote learning during COVID-19 substantially increased dropout risk and produced large learning losses in secondary education. By leveraging natural experiments within São Paulo State and robust identification strategies with selection adjustments, the paper quantifies the magnitude of losses, documents equity-relevant heterogeneity, and shows that even partial resumption of in-person high-school classes yields measurable learning gains. Policy implications include prioritizing safe school reopening, implementing intensive remedial education (especially in math), and targeted interventions to re-engage at-risk students. Future research should examine longer-term effects on attainment and labor-market outcomes, the effectiveness of specific remedial and dropout-prevention interventions, and heterogeneity by family background variables such as parental education, and extend analyses to private schools and other regions.

- Dropout risk is measured via a proxy (missing math and Portuguese grades), which, while validated and predictive of later non-attendance, introduces classification uncertainty and complicates direct comparisons with studies using actual re-enrolment. - Standardized tests in 2020 differed in mode and curriculum coverage (simplified curriculum; longer completion time), necessitating within-year DiD to control for these differences; residual unobserved differences might remain. - Selection into taking standardized tests during remote learning required propensity score adjustments; although results are robust, unmeasured factors could still bias estimates. - Reopening analyses estimate intention-to-treat effects based on municipal authorization; actual school-level reopening and individual attendance are unobserved, likely attenuating estimated effects. - Data cover public schools in São Paulo State; findings may not generalize to private schools or other contexts. Student-level variables like parental education are unavailable, limiting exploration of some heterogeneities. - The dataset cannot be shared publicly due to identifiers, constraining external replication outside the secure environment.