Heat, Test Scores, and Learning: What the Evidence Shows

Two teams, one exam, opposite answers

Brazil's national high-school exam, the ENEM, is one of the largest standardized tests in the world. It is also a natural laboratory: millions of students, one country, fixed exam dates, and weather that ranges from mild to brutal.

Two teams of economists used it to ask the same question: does temperature affect scores?

Li and Patel, publishing in Economics Letters in 2021, found that a 1°C increase in dry-bulb temperature improved the average z-score by 0.003 standard deviations. They called the effect negligible, robust across specifications, and empty of meaningful heterogeneity by subject, gender, score quartile, or race [1].

Melo and Suzuki, in a working paper circulated in 2021, found that a one-standard-deviation increase in exam-time temperature — 3.68°C — lowered scores by 0.036 standard deviations. The per-degree coefficient was -0.0097, roughly three times the size of Li and Patel's estimate and opposite in sign. They also found that the effect shrank as the exam became more consequential: about 58% smaller at higher stakes, and more than 80% smaller at the highest [2].

Same exam, overlapping years, opposite conclusions.

If heat lowers scores, air-conditioning schools and moving exam dates become investments with measurable returns. If it does not, those policies are expensive placebos. And if the effect depends on stakes, the same hot day can be harmless in a low-stakes practice test and decisive in a college entrance exam.

The contradiction is the story. It tells us that "the effect of heat on learning" is not a single number. It is a family of effects that depend on what is measured, when, where, and under what incentives.

The question, and why it matters

Narrow the question first. "Heat and learning" can mean three different things:

1. Acute performance. Does a hotter exam day lower the score a student produces from the knowledge they already have?
2. Cumulative learning. Does sustained heat during the school year reduce how much knowledge a student accumulates?
3. Attainment. Does heat change whether a student graduates, passes a threshold, enters college, or drops out?

They are related, but evidence for one does not automatically transfer to another.

The policy stakes are concrete. Tests allocate scarce seats, trigger accountability sanctions, determine teacher bonuses, and gate access to higher education. A small shift can be decisive near a cutoff. Heat is routine.

Brazil makes the stakes visible. According to the 2025 Anuário Brasileiro da Educação Básica, drawing on the 2024 school census, only 38.7% of public-school classrooms had any climate-control equipment — air conditioning, a heater, or a climatizer — and 42.1% of all classrooms did. "Any equipment" is a generous definition; actual air conditioning is rarer [22]. In a country that runs one of the world's largest high-stakes exams, most public-school students sit in rooms that cannot reliably control temperature.

That is the background for the ENEM disagreement.

Three claims that are not the same

Arguments about heat and schooling often slide between three claims.

Claim 1: Heat lowers scores on hot exam days. This is the best-established claim. The designs are clean: compare the same student across exam days, or the same school across years, and ask whether the hotter day produces a lower score. The answer is usually yes, especially in math and especially without cooling.

Claim 2: Heat reduces how much students learn over a school year. This is harder. A score is a noisy measure of knowledge, and a student who underperforms on a hot day may recover later. To isolate learning, researchers use cumulative exposure over the school year, control for test-day weather, and check whether weekends and summers — when school is out — show the same effect. If heat only disrupts the test, summer heat should not matter. If heat disrupts schooling, it should.

Claim 3: Heat changes life outcomes. This is hardest. A score effect becomes an attainment effect only when it crosses a threshold: a pass/fail line, a college cutoff, a graduation requirement. Small average effects can produce large allocation effects near those thresholds. They can also be offset by institutions: teachers adjust grades, students retake exams, schools reschedule.

Most of the evidence supports Claim 1. The evidence for Claim 2 is real but more contingent. Claim 3 is plausible and increasingly documented, but it depends on institutional details.

The mechanisms differ too. Acute heat affects attention, working memory, and physiological comfort; cumulative heat can also affect sleep, attendance, and instructional time; attainment effects add institutional responses such as grade manipulation and dropout. A single "heat effect" cannot carry all of that.

What the best-identified studies find

The literature has expanded rapidly since 2020. The strongest studies share a strategy: exploit plausibly exogenous variation in temperature, hold the person or school constant, and measure outcomes that matter.

In the United States, Park, Goodman, Hurwitz, and Smith used 10 million students who retook the PSAT. They found that hotter school days in the years before the test reduced scores, with extreme heat especially damaging. Weekend and summer temperatures had little effect, which points to school-time disruption rather than a general climate effect. Without air conditioning, a 1°F hotter school year reduced that year's learning by about 1%. Air conditioning largely offset the effect, and hot school days accounted for roughly 5% of the racial achievement gap [3].

Park's follow-up used New York City's Regents exams, which are high-stakes and inflexibly scheduled. A one-standard-deviation increase in exam-time temperature (+6.2°F) reduced performance by about 5.5% of a standard deviation. Taking an exam at 90°F rather than 75°F reduced performance by about 13% of a standard deviation. A 90°F exam day lowered the probability of passing a subject by roughly 10%, and a one-standard-deviation increase in average exam-time temperature reduced on-time graduation by about 3 percentage points. Teachers also marked grades upward after hot exams, softening the attainment damage [4].

The pattern travels. Park, Behrer, and Goodman combined PISA data from 58 countries with 12,000 U.S. school districts. Each additional day above 26.7°C in the three years before the exam lowered PISA scores by about 0.18% of a standard deviation. The effect was driven by school days, not weekends or summers, and was concentrated in poorer countries. In the U.S. district data, each hot school day lowered achievement by about 0.07% of a standard deviation [5].

In China, Graff Zivin, Song, Tang, and Zhang used the national college entrance exam, with more than 14 million observations. A 1°C increase in exam-period temperature reduced total scores by about 0.34%. A one-standard-deviation increase within counties (2°C) reduced scores by about 0.68%, or 5.83% of a standard deviation, and lowered the probability of admission to a first-tier university by about 1.2% [7].

In Mexico, Cruz and Fabián used 60 million test scores from 18 million students. They separated chronic exposure over the school year from acute exposure on exam day. In the full sample, a 1°C increase in average school-year maximum temperature reduced scores by 0.0133 standard deviations; exam-day temperature reduced them by 0.004. Chronic exposure accounted for roughly three-quarters of the total. Chronic exposure dominated in warm, humid municipalities; acute exposure dominated in temperate, dry ones. The chronic effect was larger for disadvantaged and rural students; the acute effect was larger for advantaged students [10].

In Italy, Negrisolo and coauthors used 3.4 million fifth-grade students across seven school years. Each additional day above 26.7°C reduced math scores by 0.0044 standard deviations — about four times the comparable U.S. estimate. Scaled to national variation in heat exposure, the penalty was 0.064 standard deviations, roughly 25% of the parental-education gap [11].

In Tanzania, Melo, Osendorf, and Sandholtz used the universe of Primary School Leaving Exam records from 2016–2021. Each additional degree Celsius reduced the probability of passing a subject exam by about 0.3 percentage points, from a base of 69%. A one-standard-deviation increase (3.89°C) reduced pass rates by about 1.2 percentage points; temperatures above 28°C reduced them by about 2 percentage points. Effects were larger for girls, lower-achieving students, and students in low-electrification areas. Hotter exam periods also reduced the probability of passing the overall exam and progressing to secondary school [12].

In India, Agarwal, Ghosh, Scarazzato, and Waltl linked upper-primary exam results to local weather. Both cumulative heat during the school year and hotter exam periods reduced performance. Even under an optimistic warming scenario, a constant temperature increase of 0.64°C would lower pass rates by about 3%. Effects peaked above 40°C and were largest for the urban poor; vegetation helped but did not offset the risk [13].

In Brazil, Costa and Goldemberg used data from 30,000 schools and found that a one-standard-deviation increase in hot days raised dropout rates in public schools by 5.1%, particularly in urban areas. Private schools showed no effect, consistent with a role for infrastructure and adaptation [14].

Colombia shows how much design matters. Hoffmann, Pulido, and Vera-Cossio found that extreme heat in the week before the exam reduced scores: 10 additional hours above 32°C reduced scores by 0.01 standard deviations, and 10 additional hours above 35°C reduced them by 0.03. Mean temperature over the week had no significant effect. The negative effect was larger for public-school and low-income students [9]. Martínez, Zuluaga, and Buriticá found that a 1°C increase in average maximum temperature over the year before the exam reduced scores by at least 2% of a standard deviation, with urban students hit hardest and rural students slightly positive. When a national scholarship program raised the stakes, the heat effect became 11.9% larger, not smaller [8].

That last result directly challenges the idea that higher stakes always buffer heat through effort. In Brazil, higher stakes appeared to shrink the effect. In Colombia, higher stakes amplified it. The difference may lie in whether the binding constraint is motivation or cognitive capacity — and in whether institutions can compensate.

Table 1 summarizes the core estimates.

How large is the effect, really?

Per-degree estimates are small.

In hot, poorly cooled settings, acute estimates cluster roughly between 0.004 and 0.03 standard deviations per degree Celsius. A plausible central range is 0.01 to 0.02. A 5°C hotter exam day therefore shifts scores by roughly 0.05 to 0.10 standard deviations; an 8°C difference shifts them by roughly 0.08 to 0.16. Those are modest by education-research standards, but near a pass/fail cutoff they can change who advances.

Park's "1% of a year's learning per 1°F" sounds small, and it is, but it is an average; the losses concentrate at extremes and in uncooled schools. The Italian estimate — 0.064 SD from national heat variability — is about a quarter of the parental-education gap. The Mexican chronic estimate implies that a 5°C hotter school year reduces scores by about 0.067 SD.

The distribution matters more than the average. Heat effects are nonlinear: temperatures above roughly 26–30°C do most of the damage, and extremes above 35–40°C do more. Averages hide that. Colombia is the cleanest illustration: mean temperature over the week before the exam had no significant effect, but hours above 32°C and 35°C did. The same coefficient can imply very different policy consequences in a temperate, air-conditioned country and a hot, uncooled one.

Why mathematics, and why the gaps widen

A consistent pattern across several studies is that math scores are more heat-sensitive than reading scores. Graff Zivin, Hsiang, and Neidell found short-run heat reduced math performance but not reading, and the NWEA analysis found the same [6][16]. Italy finds effects in both math and language, but the math result is the headline [11].

It is not universal. Melo and Suzuki use a total score, and Li and Patel find no subject heterogeneity [2][1]. So "math is more sensitive" is a robust tendency, not a law.

Why might it be true? Math tests lean more on working memory, fluid reasoning, and sequential problem-solving, which are sensitive to physiological stress. Reading tests lean more on vocabulary and accumulated knowledge, which may be more resilient. Heat may also change effort allocation. These are hypotheses, not settled mechanisms.

The equity pattern is more consistent than the subject pattern, but it is not uniform either. Heat tends to widen gaps when exposure and cooling are unequal. In the U.S., hot school days account for about 5% of the racial achievement gap [3]. In Italy, the heat penalty scales with socioeconomic vulnerability [11]. In Tanzania, girls, low achievers, and students in low-electrification areas are hit hardest [12]. In Mexico, the chronic channel is concentrated among disadvantaged and rural students [10]. In Brazil, public schools show dropout effects while private schools do not [14].

But there are exceptions that matter. Martínez and coauthors find larger heat effects among urban and higher-socioeconomic-status students in Colombia, not the poorest [8]. Hoffmann and coauthors find larger effects for public-school and low-income students in the same country [9]. The difference may reflect different outcomes, exposure windows, and comparison groups. The honest summary is that disadvantage often amplifies heat effects, but the mechanism is not automatic: it depends on who is exposed, who has cooling, and who can compensate.

The air-conditioning result

If there is a policy lever with strong evidence behind it, it is cooling.

Park et al. exploit school-level variation in air-conditioning penetration and find that it largely offsets the effect of a hotter school year [3]. The Italian study notes that only about 10% of municipalities have school cooling, while heating is legally mandated — a regulatory asymmetry [11]. The India study finds that vegetation around schools helps, but not enough [13]. The Brazil dropout study finds no effect in private schools, consistent with better infrastructure [14]. The Tanzania study finds larger effects where school electrification is lower [12].

Air conditioning is not a complete solution. It is expensive, consumes energy, can fail, and may not reach the poorest schools without targeted funding. It also does not address exposure at home or during sleep. But the evidence that cooling mitigates the test-score effect is among the most robust findings in this literature.

There is a deeper question: does air conditioning improve learning, or only test-taking? If heat only impairs performance on the day, cooling exam rooms raises scores without raising knowledge. If heat disrupts learning over the year, cooling classrooms raises both. The cumulative designs suggest the latter is real, but the long-run evidence is mixed.

Graff Zivin, Hsiang, and Neidell found that short-run heat reduced math performance, but long-run climate exposure had no statistically significant relationship with human capital. The authors interpret this as evidence of adaptation and compensatory behavior: families and schools offset short-run losses over time [6]. That is an important caution. It does not mean heat is harmless; it means the long-run effect depends on how much compensation is possible. In poor, hot, poorly resourced settings, compensation is harder.

The counterevidence

A serious synthesis has to take the nulls seriously.

Li and Patel find a small positive effect of temperature on ENEM scores, with no meaningful heterogeneity [1]. Carlson and Shepardson, using six years of Tulsa data, find no significant test-day effect in fall or winter, a negative effect in spring math, and a positive effect of summer heat on fall scores [15]. The summer result is a reminder that the relationship is not mechanical.

The NWEA analysis finds a negative effect only at very high temperatures: above 100°F, math scores fall by about 0.06 SD, with high-poverty schools at about 0.09 SD [16]. That is a narrower trigger than "any hot day."

The long-run null in Graff Zivin et al. is the most important counterweight [6]. If adaptation can offset short-run losses, projections that extrapolate from test-day estimates to lifetime human capital may overstate the damage. The authors are explicit that their estimates are imprecise and that the absence of a long-run effect is not proof of zero effect. But the burden of proof shifts toward those who claim large persistent losses.

There is also a measurement problem. Most studies use outdoor temperature as a proxy for the temperature students actually experience. Indoor temperature depends on building materials, ventilation, shade, occupancy, and cooling. A study that finds a null may be measuring the wrong thing; a study that finds an effect may be capturing correlated shocks such as air pollution, humidity, or disease. The best studies control for some of these, but not all.

And there is publication bias. The review by Prentice and coauthors found 34 studies, 76% with significant effects, and a pooled estimate of -0.025 SD per 1°C — larger than most single-study estimates. The review is a preprint, and its pooled sample is small (eight studies for the main estimate). It notes that few studies come from low-income settings [17]. A pooled estimate that is larger than the underlying studies is a warning sign, not a settled fact.

The review literature is consistent with a real negative effect but not with a single magnitude. Umerani and coauthors, reviewing 14 studies, find that hotter temperatures are generally associated with worse outcomes, with low-income and racial-minority students more vulnerable [18]. Vasilakopoulou and Santamouris, reviewing seven studies covering 14.5 million students in 61 countries, find that chronic heat impairs learning, complex tasks are more affected than simple ones, and acclimatization and air conditioning are protective [19]. These reviews support the direction of the effect; they cannot settle its size.

What would change the answer

Several findings would materially update this synthesis.

First, a pre-registered randomized trial that randomizes exam rooms, exam timing, or cooling, measures indoor temperature, and reports learning outcomes — not just test scores — would be decisive. The current evidence is quasi-experimental; a well-powered field experiment would test the causal chain directly.

Second, a harmonized reanalysis of Brazil's ENEM, using the same sample, temperature measure, and specification, would resolve the sign disagreement. That is feasible with public microdata and would show whether the contradiction is a data artifact or a real context effect.

Third, a long-run individual panel linking heat exposure to earnings, education, and health would tell us whether test-score effects persist. The current long-run evidence is thin and mostly U.S.-based.

Fourth, a staggered rollout of school cooling — in Brazil, India, Mexico, or elsewhere — could provide a natural experiment. If cooling raises scores and learning, the case for investment is strong. If it raises scores but not learning, the policy is narrower than it looks.

Fifth, better measurement of sleep, attendance, and instructional time would clarify the mechanisms. Minor and coauthors show that sleep loss rises steeply above 10°C nights, especially for low-income and older people [20]. Nennstiel and König show that home air conditioning is stratified by socioeconomic status in PISA data, and that in hotter climates it is associated with higher scores [21]. These are plausible pathways, but they are not yet integrated into the test-score literature.

Calibrated confidence

My current probabilities:

• 90% that hot exam days lower math and cognitive scores in settings without effective cooling.
• 80% that the effect is large enough to matter at high-stakes thresholds — roughly 0.05–0.15 SD on a very hot day — even if the average effect is modest.
• 65% that sustained heat during the school year reduces cumulative learning in hot, poorly cooled schools, beyond any test-day effect.
• 80% that air conditioning and ventilation substantially mitigate the effect.
• 75% that heat widens socioeconomic or racial achievement gaps where exposure and cooling are unequal.
• 70% that test-day heat changes attainment through pass/fail, graduation, or college-admission thresholds.
• 60% that learning losses persist into adulthood in settings without compensation.
• 50/50 on the sign of the average ENEM effect in Brazil; the published and working-paper estimates are too far apart to choose a winner without a harmonized reanalysis.
• 70% that math is more heat-sensitive than reading, with substantial variation by test and context.

I would revise the cumulative-learning estimate downward if the long-run null replicates in low-income settings with good measurement. I would revise it upward if randomized cooling trials show effects on learning that persist after the exam.

Verdict

Heat is a real but conditional obstacle to learning.

The best evidence says that hot exam days lower scores, especially in math, especially in schools without cooling, and especially for students who are already disadvantaged. The effect is small per degree but large enough to matter at thresholds, and it is nonlinear: extremes do most of the damage.

The evidence that heat reduces cumulative learning is credible but less settled. Cumulative-exposure designs find effects; long-run adaptation studies find nulls; the difference is plausibly about compensation. In rich, cooled settings, the long-run damage may be modest. In hot, poor, uncooled settings, it is more likely to persist.

The Brazil contradiction is not a reason to ignore the question. It is a reason to be precise about it. Two credible analyses of the same exam disagree on the sign, which means the effect is sensitive to sample, specification, and context. The right response is not to pick the more congenial estimate. It is to harmonize the data and find out.

The policy implications are less ambiguous than the science. Cooling schools, improving ventilation and shade, scheduling high-stakes exams away from the hottest periods, and protecting sleep are all defensible. They are also unevenly distributed: the students most exposed to heat are often the ones least able to compensate. That is the equity case, and it does not depend on the exact coefficient.

The cost of heat, in other words, is not just a lower score on a hot day. It is the accumulation of small disadvantages in places where the infrastructure to absorb them is missing.

References

1. Li, X., & Patel, P. C. (2021). Weather and high-stakes exam performance: Evidence from student-level administrative data in Brazil. Economics Letters, 199, 109698. https://doi.org/10.1016/j.econlet.2020.109698
2. Melo, A. P., & Suzuki, M. (2021). Temperature, effort, and achievement: Evidence from a large-scale standardized exam in Brazil [Working paper]. https://mizuhirosuzuki.github.io/assets/pdf/exam_brazil_paper.pdf
3. Park, R. J., Goodman, J., Hurwitz, M., & Smith, J. (2020). Heat and learning. American Economic Journal: Economic Policy, 12(2), 306–339. https://doi.org/10.1257/pol.20180612
4. Park, R. J. (2022). Hot temperature and high-stakes performance. Journal of Human Resources, 57(2), 400–434. https://doi.org/10.3368/jhr.57.2.0618-9535r3
5. Park, R. J., Behrer, A. P., & Goodman, J. (2021). Learning is inhibited by heat exposure, both internationally and within the United States. Nature Human Behaviour, 5(1), 19–27. https://doi.org/10.1038/s41562-020-00959-9
6. Graff Zivin, J., Hsiang, S. M., & Neidell, M. (2018). Temperature and human capital in the short and long run. Journal of the Association of Environmental and Resource Economists, 5(1), 77–105. https://doi.org/10.1086/694177
7. Graff Zivin, J., Song, Y., Tang, Q., & Zhang, P. (2020). Temperature and high-stakes cognitive performance: Evidence from the national college entrance examination in China. Journal of Environmental Economics and Management, 104, 102365. https://doi.org/10.1016/j.jeem.2020.102365
8. Martínez, J. M., Zuluaga, V., & Buriticá, A. (2025). Sweating bullets: Heat, high-stakes evaluations, and the role of incentives. Environmental and Resource Economics, 88, 2429–2467. https://doi.org/10.1007/s10640-025-01011-y
9. Hoffmann, B., Pulido, X., & Vera-Cossio, D. (2023). The unequal effect of temperature on test scores: Evidence from Colombia (IDB Working Paper). https://doi.org/10.18235/0004832
10. Cruz, G., & Fabián, M. (2026). The unequal burden of heat on learning [Preprint]. https://doi.org/10.2139/ssrn.7115097
11. Negrisolo, G., Brunori, P., Fedeli, E., Lopresti, I., Triventi, M., & Valentini, A. (2026). Heat and learning inequality: Evidence from Italy (III Working Paper 173). https://doi.org/10.21953/researchonline.lse.ac.uk.00138899
12. Melo, A. P., Osendorf, A., & Sandholtz, W. A. (2026). The immediate and lasting effects of heat on human capital: Evidence from high-stakes primary school exams (UNU-WIDER Working Paper 2026/793). https://doi.org/10.35188/unu-wider/2026/793-0
13. Agarwal, S., Ghosh, P., Scarazzato, F., & Waltl, S. R. (2025). Adverse hot-humid conditions and educational outcomes in India: Impact measurement and mitigation pathways (Cambridge Working Paper 2025-06). https://doi.org/10.17863/cam.122260
14. Costa, F., & Goldemberg, D. (2025). Too hot to learn? Evidence from high school dropouts in Brazil. Economics Letters, 247, 112157. https://doi.org/10.1016/j.econlet.2024.112157
15. Carlson, D., & Shepardson, A. (2024). Under the weather? The effects of temperature on student test performance (EdWorkingPaper 24-910). Annenberg Institute at Brown University. https://eric.ed.gov/?id=ED672402
16. Postell, A., Kuhfeld, M., Kowalski, A., & Isaacs, J. (2025). Hot test days, lower math scores [Research brief]. NWEA. https://www.nwea.org/research/publication/heat-and-learning-loss/
17. Prentice, C., Meneveau, M., Wu, J., Girard, L.-C., Zheng, Y., & Vergunst, F. (2025). Heat effects on children's education outcomes: A systematic review and meta-analysis [Preprint]. https://doi.org/10.31234/osf.io/6jrx3_v1
18. Umerani, A., Sachs, A. L., Karsan, S., Saeed, A., Rogers, H., Sprague, N., & Ekenga, C. C. (2026). Climate, classrooms, and children: A systematic review of temperature-related impacts on academic performance. Journal of School Health, 96(8). https://doi.org/10.1111/josh.70197
19. Vasilakopoulou, K., & Santamouris, M. (2025). Cumulative exposure to urban heat can affect the learning capacity of students and penalize the vulnerable and low-income young population: A systematic review. PLOS Climate, 4(7), e0000618. https://doi.org/10.1371/journal.pclm.0000618
20. Minor, K., Bjerre-Nielsen, A., Jonasdottir, S. S., Lehmann, S., & Obradovich, N. (2022). Rising temperatures erode human sleep globally. One Earth, 5(5), 534–549. https://doi.org/10.1016/j.oneear.2022.04.008
21. Nennstiel, R., & König, C. (2026). The thermostat of opportunity: Home air conditioning, its social gradient, and educational consequences. Energy Research & Social Science, 138, 104785. https://doi.org/10.1016/j.erss.2026.104785
22. Todos Pela Educação. (2025). Anuário Brasileiro da Educação Básica 2025 — Infraestrutura. https://www.anuario.todospelaeducacao.org.br/2025/capitulo-12-infraestrutura.html