The Effects of Air Pollution on Students’ Cognitive Performance: Evidence from Brazilian University Entrance Tests
Abstract
We examine the contemporaneous causal relationship between outdoor air pollution levels and student cognitive performance in Brazil’s nationwide university entrance examinations. Our analysis relies upon a unique and previously unexplored student-level data set allowing us to examine the effect of particulate matter (PM10) on students’ scores. In our main specification we construct individual-level panel data for the 2 days of exams across 3 years and apply student fixed effects to address potential endogeneity concerns. In addition, we take advantage of plausibly exogenous spatial and temporal variation in PM10 across municipalities in the states of Rio de Janeiro and São Paulo and utilize an instrumental variable approach based on wind direction. Our results suggest that air pollution negatively impacts the cognitive performance of students. We find suggestive evidence that boys may be more affected than girls, and less well-off exam takers at the bottom of the score distribution are more affected than their more privileged counterparts.
The World Health Organization (WHO) estimates that globally as many as 4.2 million premature deaths each year are linked to exposure to ambient air pollution.1 The WHO also points out that the health impacts associated with pollution exposure are not equitably distributed, with the poorest individuals in the poorest countries being disproportionately affected. The manner in which many common local air pollutants adversely affect health, particularly via the respiratory and cardiovascular systems, is now well known (Dockery et al. 1993; Hansen and Selte 2000; Weinhold 2008; Currie et al. 2009; Ling and van Eeden 2009; Neidell 2009; US EPA 2016). However, a more recent literature has examined the possibility that exposure to air pollution may also affect cognitive performance and, in turn, potentially human capital formation, productivity, and ultimately economic growth (Zweig et al. 2009; Currie et al. 2014; Ebenstein et al. 2016). If true, this raises the prospect of a “cycle of disadvantage” in which exposure to pollution constrains the development prospects of those who need it most.
The literature examining the impact of pollution exposure on human capital formation has found compelling evidence that air pollution, through its impact on health, increases school absenteeism (Currie et al. 2009; Ransom and Pope 2013; Chen et al. 2018; Liu and Salvo 2018), negatively affects school pupils’ scores (Almond et al. 2009; Nilsson 2009; Zweig et al. 2009; Reyes 2011; Sanders 2012; Bharadwaj et al. 2017), and decreases low- and high-skilled workers’ productivity (Chang et al. 2016; Heyes et al. 2016; Archsmith et al. 2018; Chang et al. 2019; Kahn and Li 2019). Other studies examine the accumulated impact of air pollution within a period of time on pupils’ scores at school (Zweig et al. 2009). Overall, these studies look at the long-run impact of fetal (Almond et al. 2009; Sanders 2012; Bharadwaj et al. 2017) or around birth (Nilsson 2009; Reyes 2011) exposure to air pollution on students’ scores and do so typically in developed country settings.
While many studies focus on the cumulative or long-run impacts of air pollution on pupils’ academic performance, few pay heed to the contemporaneous effect of pollution exposure on student performance. One exception is Ebenstein et al. (2016), who find negative effects of air pollution exposure on the day of high-stakes examinations in Israel, with negative long-run effects on individuals’ earnings. Another relevant study is Graff Zivin et al. (2020), who examine the impact of agricultural fires on Chinese students’ exam scores. They differentiate between upwind and downwind fires and find the difference between the two to have a statistically significant effect on exam performance. With regard to Brazil, Bedi et al. (2021) conduct an experiment at the University of São Paulo and suggest that exposure to high levels of PM2.5 reduces students’ performance in a reasoning test. Here we add to this literature by investigating the impacts of particulate matter on students’ scores in high-stakes exams by using a unique geographic and demographically representative data set. Despite the findings of Graff Zivin et al. (2020) and Bedi et al. (2021), there remains very limited evidence on the impacts of urban air pollution on students’ cognitive performance in the most highly polluted cities and regions of the developing world.
This paper examines the causal relationship between outdoor air pollution levels on the day of the nationwide university entry examinations and students’ cognitive performance in Brazil, and specifically in Rio de Janeiro and São Paulo, its most industrialized states. More precisely, we use Brazilian data on concentrations of ozone (O3) and particulate matter (PM10) and a rich administrative data set on students’ scores to examine the impact of air pollution on academic performance in national examinations (Exame Nacional do Ensino Médio; ENEM).2 This educational data set is combined with air pollution and weather monitoring station data to build a unique panel of data for the period 2015–17. We also test the effects of air pollution on the exam performance of male and female students separately since the previous literature has indicated that fetal exposure to pollution can have different long-term impacts on males and females (Jayachandran 2009; Sanders 2012). Similarly Ebenstein et al. (2016) found boys’ exam performance to be more affected by air pollution than that of girls. In addition, we utilize students’ socioeconomic characteristics to investigate the heterogeneous effects of air pollution on students’ scores per different economic strata.
Our analysis contrasts with Ebenstein et al. (2016), who consider the impact of an overall air quality index on the Israeli national examination composite score, and with Graff Zivin et al. (2020). The latter work could neither establish the direct link between fires and air pollution nor assess the direct effect of air pollutants on students’ scores due to lack of data on air pollution for the period of study. Unlike the previous literature, besides using student fixed effects, we also implement an instrumental variable (IV) approach to mitigate endogeneity concerns related to air pollution. Following Deryugina et al. (2019) and Bondy et al. (2020), we use wind direction as an exogenous shock capable of producing variation in air pollution levels on a particular day and locality. Our assumption relies on the fact that municipalities’ daily wind direction is not directly related to students’ scores in the national examinations except through its influence on air pollution. We take advantage of the geographic and temporal variation in wind direction in our data set to explore the exogeneity of our instrumental variable, that is, the variation in daily wind direction in the municipality in which the examination took place.
Our focus on Brazil is motivated by the fact that, by global standards, it experiences relatively high levels of air pollution and relatively low levels of academic performance. In terms of pollution, for instance, in 2015 the Environmental Company of São Paulo (Companhia Ambiental do Estado de São Paulo; CETESB) reported 972 occurrences of ozone levels being above WHO safe limits in the metropolitan area of that state.3 Regarding academic achievements, in the OECD’s Program for International Student Assessment (PISA) Brazil was ranked 39th out of 40 countries in 2019, revealing the relative underperformance of Brazilian students in comparison with their international counterparts, including most of Brazil’s Latin American neighbors. Furthermore, OECD research by Zapata et al. (2015) shows that only approximately 15% of 25–34-year-olds in Brazil completed tertiary education in 2014, compared to the OECD average of 41%.
Our findings suggest that an increase of 10 micrograms per cubic meter (μg/m3) of PM10 on the day of the examination decreases students’ scores by 6.1 points (8% SD). We find evidence that the students most affected by pollution are male, those with weaker school performance, and those from poorer households, although the differences that we find across these subsamples are not statistically significant. Finally, we run placebo tests using levels of pollution 1 day after the exams, with results reinforcing our main findings that it is indeed the poor air quality on the day of the exams that reduces exam scores. That is, pollution levels on the day after the tests have no significant impact on students’ cognitive performance. We therefore believe that our study provides policy makers with important evidence of a contemporaneous link between air pollution and exam performance.
The remainder of this paper is organized as follows. In section 1, we describe our data. We present our empirical strategy in section 2. In section 3, we present our results and conduct sensitivity analysis. In section 4, we present our conclusions.
1. Data and Summary Statistics
Our data set on daily air quality, weather conditions, and students’ ENEM scores in São Paulo and Rio de Janeiro between 2015 and 2017 are collected from three main sources as detailed below. As our study comprises 3 years of examinations with students taking exams in 2 days (Saturdays and Sundays), we build a panel data at the student level and control for the day of examinations. We describe our data set and present summary statistics below.
1.1. Air Pollution Data
Air pollution data from São Paulo and Rio de Janeiro are collected from their respective environmental state agencies: the Environmental Company of São Paulo (Companhia Ambiental do Estado de São Paulo; CETESB) and the State Institute for the Environment (Intituto Estadual do Meio Ambiente; Inea). Both agencies are responsible for monitoring the air quality of the two states, using ground-level automatic and nonautomatic monitoring stations.4 São Paulo has 63 automatic monitoring stations and 26 nonautomatic stations, while Rio de Janeiro has 58 automatic and 116 nonautomatic stations (see fig. A3). Inea’s and CETESB’s websites provide the addresses, latitude, and longitude of their monitoring stations, which we used to merge this data set with the national examination scores, by municipality centroid location.
Unfortunately, for confidentiality reasons, the Ministry of Education in Brazil does not provide the exact address where each student took their examinations, only the municipality. Hence, we are not able to identify the closest monitoring station to the exact exam location. Nevertheless, following Graff Zivin et al. (2020), we proxy each exam venue’s location using the centroid of the municipality in which the test took place. We then identify up to three pollution monitoring stations within 10 kilometers (km) of the municipality centroid and average hourly readings across those to form an average daily measure of pollution (Table 1 reports summary statistics of PM10 and O3).5 We then calculate the inverted weighted distance (IWD) for PM10 and O3 (in μg/m3) following Currie and Neidell (2005), Sager (2019), and Bondy et al. (2020).6
| Variable: Exam-Level Data | All | Girls | Boys |
|---|---|---|---|
| Total score (0–1,000 points) | 526 | 516 | 537 |
| (76.2) | (71.3) | (78) | |
| Sciences (0–1,000 points) | 502 | 493 | 515 |
| (78.9) | (76.4) | (80.2) | |
| Humanities (0–1,000 points) | 558 | 550 | 569 |
| (77.3) | (76.2) | (77.5) | |
| Languages (0–1,000 points) | 532 | 532 | 534 |
| (68.4) | (67.1) | (70) | |
| Mathematics (0–1,000 points) | 512 | 494 | 535 |
| (116) | (106) | (125) | |
| PM10 (μg/m3) | 17.5 | 17.5 | 17.5 |
| (6.11) | (6.13) | (6.08) | |
| O3 (μg/m3) | 41.5 | 41.5 | 41.6 |
| (12.3) | (12.3) | (12.1) | |
| Temperature (°C) | 20.6 | 20.6 | 20.6 |
| (3.2) | (3.25) | (3.24) | |
| Humidity (%) | 81.4 | 81.4 | 81.5 |
| (13.6) | (13.6) | (13.6) | |
| Wind speed (m/s) | 2.25 | 2.25 | 2.25 |
| (1.25) | (1.24) | (1.25) | |
| Observations | 2,931,368 | 1,671,910 | 1,259,458 |
| N municipalities | 47 | 47 | 47 |
1.2. Weather Data
To control for atmospheric climate we collected data on weather conditions from CETESB and Inea monitoring stations. This provides us with measures of temperature (in degrees Celsius, °C), relative humidity (in percentage, %), wind speed (in meters per second, m/s), and wind direction (in degrees).7 The latter is used as an instrumental variable for PM10. CETESB provides wind direction in cardinal mode; thus, to use it as a numerical variable, we transform it into degree direction. Moreover, as wind direction is an angle (360° circle) and covaries with wind speed (i.e., there are two components that need to be averaged together) to calculate its daily average we need to split out the east/west vector and the north/south vector, and then recombine them to find the speed and direction (Itron 2019; Webmet 2019).8 The remaining climate measures are also registered on an hourly basis, with the daily means calculated and assigned to the examinations’ municipalities using the IWD in a similar manner to which the pollution measure is built (Currie and Neidell 2005; Sager 2019; Bondy et al. 2020). Table 1 presents the summary statistics for weather conditions.
1.3. University Entrance Examination Scores and Other ENEM Data
The subject-specific ENEM scores are collected from the National Institute for Educational Research Instituto Nacional de Estudos e Pesquisas Educacionais Anísio Teixeira (INEP), an agency from the Brazilian Ministry of Education (INEP 2017). This is a rich administrative data set with detailed information on the students’ scores per exam subject at the student and school level, and demographic information, such as household income, parental level of education, age, race, gender, and special needs (tables 1 and 2 summarize this information).
| Variable: Demographic Information: Student-Level Data | Girls (%) | Boys (%) | All |
|---|---|---|---|
| Race: | |||
| White | 56.5 | 43.5 | 676,034 |
| Black | 56.4 | 43.6 | 209,759 |
| Mixed race | 58.1 | 41.9 | 512,356 |
| Asian | 60.2 | 39.8 | 31,787 |
| Indigenous | 51.5 | 48.5 | 6,285 |
| Nondeclared | 52.4 | 47.6 | 29,463 |
| Type of school: | |||
| Federal | 51.3 | 48.7 | 9,096 |
| State | 58.9 | 41.1 | 390,906 |
| Municipal | 57.8 | 42.2 | 10,234 |
| Private | 54.4 | 45.5 | 175,079 |
| Nondeclared | 43.2 | 56.8 | 880,369 |
| Mother’s education: | |||
| Never studied | 63.8 | 36.3 | 40,175 |
| Incomplete first elementary | 61.9 | 38.1 | 192,868 |
| Incomplete second elementary | 59.3 | 40.7 | 192,269 |
| Elementary | 57.5 | 42.5 | 217,829 |
| High school | 56.2 | 43.8 | 519,969 |
| Undergraduate | 46.8 | 53.2 | 175,996 |
| Postgraduate | 47.4 | 52.6 | 87,369 |
| Nondeclared | 51.4 | 48.6 | 43,209 |
| Father’s education: | |||
| Never studied | 62.7 | 37.3 | 47,118 |
| Incomplete first elementary | 61.4 | 38.6 | 235,784 |
| Incomplete second elementary | 59 | 41 | 197,293 |
| Elementary | 56.7 | 43.3 | 197,121 |
| High school | 55.4 | 44.6 | 441,797 |
| Undergraduate | 52.7 | 47.4 | 146,295 |
| Postgraduate | 51.4 | 48.6 | 73,888 |
| Nondeclared | 58.3 | 41.7 | 126,388 |
| Household income: | |||
| No salary | 59.7 | 40.3 | 25,797 |
| 1–7 min. wages | 58.9 | 41.1 | 1,113,334 |
| >7 min. wages | 50.4 | 49.6 | 326,553 |
| Number of students | 835,955 | 629,729 | 1,465,684 |
The data for 2015 indicate that exams took place in 261 municipalities: 208 municipalities in the state of São Paulo and 53 in Rio de Janeiro state. In 2016, the data set indicates 209 municipalities in the state of São Paulo and 51 in Rio de Janeiro state, making a total of 260 municipalities where the ENEM took place. In 2017, the state of São Paulo held examinations in 209 municipalities, and the state of Rio de Janeiro in 52. Note that since we use monitoring stations as far as 10 km from the municipality centroid to construct our air pollution and meteorological measures, we end up with 47 municipalities within our final sample.9 This reduction in the number of municipalities and hence number of spatial observations of pollution measures in our final sample consists of a trade-off decision that we must make. Had we opted for using stations located, for instance, 50 km from the municipality centroid, we would have experienced attenuation bias in our estimates.10 As, in a given year, every student takes the test on the same day, the variation in air pollution comes from the fact that they take exams in different locations. Nonetheless, the reduction in the number of observations in our final panel does not affect the power of our study since we still end up with a very large sample size.
As for demographic information, the INEP data provide the highest level of education achieved by each student’s parents. As such, we build categorical variables indicating each of the seven educational levels as follows: (i) never studied, (ii) incomplete first elementary if the parent attended school for a while but did not complete the first elementary, (iii) incomplete second elementary if the parent attended school for a while but did not complete the second elementary, (iv) elementary if the parent only completed elementary school, (v) high school if the parent completed high school, (vi) undergraduate if the parent completed undergraduate school, (vii) postgraduate if the parent completeda postgraduate level, and (viii) nondeclared if the students declared that they were unaware of their parents’ level of education (Zweig et al. 2009).
Students’ socioeconomic background is also likely to be an important predictor of their academic performance, particularly in developing countries where the degree of inequality is much higher than in developed nations. In Brazil it has been found that students’ socioeconomic characteristics can determine up to 85% of the variation in exam scores (Estadao 2019). In light of this, in our models that use municipality fixed effects, we control for socioeconomic background using parental household income. The original data do not display the income in figures, preventing us from using that data as a continuous variable. As a result, we created dummies for each level of parental household income. According to the INEP’s microdata information, there are 17 income ranges—monthly earnings from zero minimum wage until more than 20 times the minimum wage. As Table 2 shows, 76% of students have a household income between one and seven times the minimum wage, 1.8% belong to families who do not receive a salary, and around 22.2% of students’ parents received more than seven times the minimum wage.
1.3.1. The ENEM
The ENEM is a nonmandatory and standardized national exam used to evaluate high school students in Brazil and is the second largest in the world after the National Higher Education Entrance Examination in China. Since 2009, the exam has been used by federal and state universities as their admission test. The exam is also used to rank deprived students to receive points in the federal scholarship Universidade para Todos Program (or ProUni), as well as for certification for a high school degree.
The ENEM comprises four groups of multiple choice exams that take place on two days: humanities, sciences, languages, and mathematics, each of them with scores ranging from 0 to 1,000. The humanities exam includes history, geography, philosophy, and sociology; the science tests encompass chemistry, physics, and biology; and the language tests include Portuguese language, literature, foreign language (English or Spanish), arts, physical education, and information and communication technologies (INEP 2017). In 2015, the examinations occurred on October 24 and 25; in 2016, on November 5 and 6; and in 2017, on November 5 and 12. In each year the exams start at 13:30 and finish at 18:30 on the first day, and start at 13:30 and finish at 19:00 on the second day. In 2015 and 2016 on the first day, students took exams in humanities and sciences and on the second day languages and mathematics. In 2017 the order was slightly different, with students taking tests on languages and humanities on the first day and sciences and mathematics on the second day.
Although we have hourly data for both air pollution and weather variables, which would enable us to assign the measures of air pollution and weather condition at the exact moment of the exams, we follow the literature on the impact of air pollution on human capital and use the daily average of O3, PM10, humidity, temperature, wind speed, and wind direction (Ebenstein et al. 2016; Sager 2019; Bondy et al. 2020).
2. Empirical Strategy
We examine the partial correlation between PM10 exposure and test scores at the individual level for the years 2015–17. Nonetheless, there are a few identification challenges to overcome in order to establish the causal relation between air pollution and students’ performance due to the possible presence of omitted correlated elements. For instance, one concern is that students’ exam scores may be correlated with air pollution due to the fact that students from higher socioeconomic strata could live in municipalities with lower levels of air pollution (the parents could choose to live in less polluted areas, which are usually more expensive) (Banzhaf and Walsh 2008; Currie et al. 2014). Hence, although the candidates do not choose the exam venue, it is determined according to the student’s household postcode (INEP tries to assign the examination venue as close as possible [up to 30 km] to each student’s address). Hence, students who live in less polluted municipalities would end up taking exams in cities with cleaner air than other candidates who live in more deprived areas, which may be more polluted (Currie et al. 2014). This possibility of address selection is also highlighted by Neidell (2004, 2009), who characterizes it as an avoidance behavior by parents.
A final concern relates to the fact that the level of pollution and weather conditions in the days and venues where the exams are marked may influence the final marks, as the examiners could also have their productivity affected by these circumstances. However, all the tests are multiple choice, and the grades are given by a machine, which clearly is not affected by air pollution or weather determinants. Thus, to tackle those econometric concerns we apply two specification strategies as detailed below.
2.1. Panel Fixed Effects Model
As students take exams on 2 days, we employ student fixed effects and control for the exact day of the examinations. The main model thus takes the following form:
2.2. The Instrumental Variable Approach
After using student fixed effects and controlling for weather conditions, there may still remain concerns about the existence of classical measurement error. For example, the students’ actual exposure to outdoor air pollution may differ from the monitoring stations’ readings, which may bias our estimates downward due to attenuation bias (Moretti and Neidell 2011; Schlenker and Walker 2015; Chen et al. 2018). Furthermore, as explained above, PM10 is not randomly assigned, leading to the potential presence of unobserved time-varying effects that we cannot account for. Therefore, we also experiment with an instrumental variable approach to estimate our models. In this regard we follow previous works, as, for example, Deryugina et al. (2019), Anderson (2020), and Bondy et al. (2020), and adopt wind direction as our instrument for air pollution, which can be considered a natural experiment that results in exogenous shocks to local air pollution in a specific area due to wind shocks (Currie et al. 2014).
Following Deryugina et al. (2019) and Bondy et al. (2020), we use k-means cluster to create three groups for all the pollution monitors in our sample. Clustering is largely used to assign pollution monitors into a predetermined number of groups according to the groups’ latitude and longitude. Hence, each group represents a geographic area.13 To choose the optimal number of regional clusters, we use the k-means cluster algorithm (Makles 2012). First, we follow the official Instituto Brasileiro de Geografia e Estatistica (IBGE) regional division of the São Paulo and Rio de Janeiro states to define as 11 the number of maximum possible clusters the municipalities in our final sample could belong to. The IBGE adopts several forms to classify the territory into smaller regions, for example, macro-regions, micro-regions, and meso-regions. We adopted the latter classification in line with CETESB’s and Inea’s monitoring stations’ organization to define the municipalities from our sample into those regional groups. We found 11 regional levels, which we inputted in the k-means algorithm to find the optimal number of clusters per latitude and longitude. Figure 1 shows the within sum of squares (WSS), log(WSS), η2, which is similar to the R2, and the proportional reduction of error (PRE) coefficient (PRE) for all k-cluster solutions. Accordingly, clustering with appears to be the optimal solution. More specifically, at we see a kink in the WSS and log(WSS). The term η2 points to a reduction of the WSS by 88% and PRE to a reduction of about 79% compared with the solution. Yet, the reduction in WSS is negligible for .14 The within sum of squares (WSS) (a), log(WSS) (b), η2 (c), and PRE (d) for all K cluster solutions (optimal number of clusters). The figure indicates clustering with to be the optimal solution. At , there is a kink in the WSS and log(WSS), respectively; η2 points to a reduction of the WSS by 88% and PRE to a reduction of about 79% compared with the k = 2 solution. However, the reduction in WSS is negligible for .
Note that the number of groups is somewhat (by latitude and longitude) arbitrarily chosen (Deryugina et al. 2019; Bondy et al. 2020). As a result, varies across geographic groups (Deryugina et al. 2019). It follows that, in using the interaction between the geographic groups with the wind direction dummies, we aim to mitigate the concern that a municipality’s pollution monitors may not properly measure the average pollution each student within each municipality is affected by since the pollution station locations are sparse throughout the two states Rio de Janeiro and São Paulo. Furthermore, we also do not know the precise location of the exam venues within each municipality. As such, our identification strategy limits the effect of municipality wind direction on pollution to be the same for all stations within each of the three geographic groups, which comprise various municipalities. As explained by Deryugina et al. (2019) and by Bondy et al. (2020), this confinement shortens the impact of pollution variation from different sources, reducing measurement error in our models. The concern here is that the 47 municipalities in our sample engulf a large area, and wind may carry particulate matter from different sources.15 That is, if the station is close to the point source, its measures are likely to be higher than another station located far from the point source, which may cause measurement error in our estimations. On the other hand, nonpoint sources tend to present more similar influence on the monitors in the same group and, hence, are more likely to be responsible for the variation on pollution that we expect to capture from the variables (Deryugina et al. 2019).
To identify the causal relationship between pollution and the students’ scores we follow the strategy employed by Deryugina et al. (2019) and Bondy et al. (2020) to build our instrumental variable. To this end we rely on the assumption that municipalities’ daily wind direction is not directly related to the students’ scores in the national examinations, except through its influence on air pollution. Therefore, as argued by Deryugina et al. (2019), we do not need to know the exact source of pollution since the use of wind direction—which can be considered random to the pollution-exam relationship—to predict air quality eliminates the concern of potential correlation between what generates pollution and other factors affecting student performance. We take advantage of the geographic and temporal variation in wind direction in our data set as the source of exogeneity of our instrumental variable. That is, we use the variation in daily wind direction in the municipality where the examination took place as an instrument for pollution. Our model can be formally described as:
Figure 2 shows the structure of our instruments by depicting the daily average PM10 across the 47 municipalities where the exams took place in our final sample by wind direction measured by its daily mean. It illustrates the relationship from our first stage, suggesting that days with wind blowing from the north, northeast, and east have on average higher levels of PM10.16 The effect of wind direction on air pollution. Daily average of PM10 concentrations for the 47 municipalities in the final sample per direction from where the wind blew during the exams days. 95% confidence intervals. Details on the data described in the text.
Our instrumental variable approach confirms our fixed effects results. Primarily, our valid set of instruments helps us to address potential unobserved factors that could remain in our baseline findings. Suffice it to say that our estimates are roughly the same in both empirical strategies. In addition, by using interactions between wind direction and geographical clusters, we are capable of demonstrating plausibly exogenous variation in air pollution measures across different regions constituting our sample.
As for the validity of our instrument, wind direction fulfills the relevance requirement since it varies along with the concentration of PM10 as demonstrated by the F-test statistics in tables A1 and 4. Moreover, with regard to the exogeneity requirement, the main assumption is that wind direction only affects students’ scores via air pollution.
Although the previous literature has shown that O3 and PM10 are not highly correlated, we use the daily average of pollution measures of PM10 as our main pollutant measure (Zweig et al. 2009; Ebenstein et al. 2016; Sager 2019). This separate equations approach, that is, the focus on one particular pollutant, has been used by, for example, Currie (2009) and Arceo et al. (2016). Other work prefers to use a composite measure of air pollution.17 Nonetheless, we estimate equations with both pollutants jointly as a robustness check, and, as Table 7 shows, the estimate signs and magnitudes of PM10 remain stable as when estimated separately.
For the temperature variables we build six bins of the size 3°C, with the first bin comprising temperatures between 16°C and 19°C, and the last bin for temperatures between 28°C and 31°C. This approach allows temperature to have a nonlinear impact. In all specifications, we use linear and quadratic terms for temperature (or bins) and humidity and their interactions (Ebenstein et al. 2016).
3. Empirical Results
3.1. Main Results
Table 3 provides our estimations’ results. For completeness we start by using municipality fixed effects. From column 1 we can observe that an additional 10 μg/m3 of PM10 corresponds to a reduction in students’ scores by 1%, though not significant. On the other hand, when it comes to utilizing student fixed effects and controlling for day fixed effects, the results in column 3—our preferred specification—suggest that an increase of 10 μg/m3 of PM10 on the day of examinations leads to a highly significant reduction of 8% of a standard deviation in students’ scores.
| Municipality Fixed Effects | Student Fixed Effects | |||
|---|---|---|---|---|
| Variables | (1) | (2) | (3) | (4) |
| PM10 (10 μg/m3) | −.01 | −.08*** | ||
| (.05) | (.03) | |||
| 1 if PM10 > 20 | −.04 | −.11*** | ||
| (.03) | (.04) | |||
| Student FE | N | N | Y | Y |
| Student controls | Y | Y | N | N |
| Day FE | Y | Y | Y | Y |
| Weather controls | Y | Y | Y | Y |
| Observations | 2,944,911 | 2,944,911 | 2,931,368 | 2,931,368 |
| N municipalities | 47 | 47 | 47 | 47 |
| R-squared | .17 | .17 | .07 | .07 |
To explore possible nonlinearity between air pollution and students’ cognitive performance, we present results from a complementary specification using a dummy variable for air pollution that equals 1 if the level of PM10 is above 20 μg/m3—WHO’s acceptable threshold for this pollutant. After controlling for student and day fixed effects, our results reported in column 4 suggest that students face a significant reduction of 11% SD in their scores.
While our main specification using student fixed effects illustrates the relationship between bad air quality and students’ cognitive performance, there might remain potential unobserved time-varying circumstances which that strategy may not be capturing. As such, we turn to our instrumental variable findings. First, our IV first-stage F-test statistic informs us that wind direction is indeed a strong predictor of air pollution and hence a strong instrument.18 Second, our findings in Table 4 suggest that an additional 10 μg/m3 of PM10 on exam days caused the students’ marks to drop by 8% SD (col. 3). If we compare this with the baseline fixed effects results presented in column 1, we notice that the results remain broadly the same even though there is a slight reduction in the sample size.
| Baseline | With Instruments | |||
|---|---|---|---|---|
| Variables | (1) | (2) | (3) | (4) |
| PM10 (10 μg/m3) | −.08*** | −.08*** | ||
| (.03) | (.01) | |||
| 1 if PM10 > 20 | −.11*** | −.13*** | ||
| (.04) | (.01) | |||
| F first stage | 55.4 | 71 | ||
| Observations | 2,931,368 | 2,931,368 | 2,405,938 | 2,405,938 |
| N municipalities | 47 | 47 | 47 | 47 |
When it comes to nonlinear PM10 for the instrumental variable specification, column 4 in Table 4 displays the same pattern of results as the one using linear pollution measure (col. 2). That is, students taking exams in venues and days above the WHO’s threshold have their scores reduced by 13% SD compared to their counterparts exposed to air pollution levels equal to or below to 20 μg/m3. Also, regarding the first stage for the nonlinear approach, the F-test statistic still reveals that wind direction is a strong predictor of air pollution.
3.2. Heterogeneity across Subpopulations
In this section we analyze the heterogeneous effect of the treatment (air pollution) for our preferred specification reported in columns 3 and 4 from Table 3: (i) by gender, (ii) across the distribution of scores per municipality, and (iii) in terms of different levels of income. Through this heterogeneity inspection, we aim to unveil how different subpopulations have their cognitive performance affected by the exposure to air pollution during high-stakes examinations.
With regard to gender differentials, our findings in Table 5 suggest that boys are more negatively affected by poor air quality during exams than girls. To put this into perspective, an increase of 10 μg/m3 of PM10 causes males’ scores to drop by 11% SD, while girls suffer a reduction of 6% SD. This heterogeneity is also observed when using a nonlinear pollution variable; the results persist, with boys’ scores being reduced by 15% SD and girls’ by 8% SD, meaning that boys who sat the examinations in days with air pollution levels above the WHO’s acceptable threshold faced a reduction in scores 7 percentage points greater than girls (cols. 5 and 6). However, as Table 5 indicates, Z-tests of the difference in the estimated coefficients for boys and girls present p-values > .1.
| Variables | All (1) | Girls (2) | Boys (3) | All (4) | Girls (5) | Boys (6) |
|---|---|---|---|---|---|---|
| PM10 (10 μg/m3) | −.08*** | −.06*** | −.11*** | |||
| (.03) | (.03) | (.03) | ||||
| 1 if PM10 > 20 | −.11*** | −.08** | −.15*** | |||
| (.04) | (.03) | (.05) | ||||
| P-value Z-test (girls − boys) | p = .88 | p = .88 | ||||
| Observations | 2,931,368 | 1,671,910 | 1,259,458 | 2,931,368 | 1,671,910 | 1,259,458 |
| N municipalities | 47 | 47 | 47 | 47 | 47 | 47 |
| R-squared | .07 | .09 | .08 | .07 | .09 | .08 |
Nevertheless, these suggestive findings are consistent with previous studies, including Jayachandran (2009), Sanders (2012), and Ebenstein et al. (2016). Even though the role of gender in the effects of air pollution on respiratory health remains unclear, some epidemiology studies point out that males are more affected than females. For example, Abbey et al. (1998) link PM10 exposure to reduced lung functioning among males but not females, Galizia and Kinney (1999) find similar results regarding the exposure to O3, and Chen et al. (2017) identify that contemporaneous and cumulative exposure to air pollution has a stronger negative effect for men than for women. Similarly, Ebenstein et al. (2016) find the effect of pollution on exam performance to differ by gender and cite the increased incidence of asthma among boys as a possible cause.
Next we assess whether students at different points of the performance distribution are differently affected by poor air quality. To this end we split our final sample into subsets of low and high scores within each municipality. One can note from panel A of Table 6 that students below the average have their scores 1 percentage point more impacted than pupils that rank above the average. Stated differently, stronger students exposed to an additional 10 μg/m3 of PM10 during the exams have their performance reduced by 6% SD (col. 3), while weaker students are more negatively affected, presenting a drop of 7% SD (col. 2). When we compare this to the baseline estimation (col. 1) using the whole sample, the average reduction is 8% SD. Another comparison is made between the top 5% scores and bottom 5%. Column 4 shows that students who rank very low in the exam suffer a reduction of 5% on their marks, while students on the top of the distribution seem not to be significantly affected by an increase of pollution levels during examinations (col. 5). Finally, columns 6–10 depict results from the same exercises utilizing the nonlinear pollution variable. These findings suggest that students above the median and who took the exams in days with PM10 above the WHO’S threshold are more negatively impacted, with scores dropping by 12% SD, while students in the top 5% of the ranking are again not significantly affected by taking exams in days with pollution levels higher than 20 μg/m3. Again, we note that the differences in our results across the performance distribution are not statistically significant.
| Variables | Baseline (1) | Below Median (2) | Above Median (3) | Bottom 5 Percentile (4) | Top 5 Percentile (5) | Baseline (6) | Below Median (7) | Above Median (8) | Bottom 5 Percentile (9) | Top 5 Percentile (10) |
|---|---|---|---|---|---|---|---|---|---|---|
| A. The Distribution of Examination Scores | ||||||||||
| PM10 (10 μg/m3) | −.08*** | −.07*** | −.06** | −.05*** | −.03 | |||||
| (.03) | (.02) | (.03) | (.01) | (.03) | ||||||
| 1 if PM10 > 20 | −.11*** | −.07* | −.09*** | −.05** | −.04 | |||||
| (.04) | (.04) | (.03) | (.02) | (.03) | ||||||
| P-value Z-test (diff. ability) | p = .41 | p = .26 | p = .66 | p = .39 | ||||||
| B. Low and High Socioeconomic Status | ||||||||||
| PM10 (10 μg/m3) | −.08*** | −.09*** | −.08** | −.09*** | −.07 | |||||
| (.03) | (.02) | (.04) | (.02) | (.06) | ||||||
| 1 if PM10 > 20 | −.11*** | −.10* | −.12*** | −.09** | −.10** | |||||
| (.04) | (.05) | (.04) | (.04) | (.05) | ||||||
| P-value Z-test (diff. income) | p = .41 | p = .37 | p = .62 | p = .56 | ||||||
In panel B of Table 6 we slice our sample to look into how students from different socioeconomic strata may be disproportionately affected by air pollution during the tests (Hsiang et al. 2019). For example, Neidell (2004) finds that less well-off households are more impacted by air pollution given their reduced capability of compensating behavior. Students from poorer backgrounds seem to be more affected by poor air quality. It may also be due to the fact that weaker students and low income are correlated (Ebenstein et al. 2016). We therefore subsampled the individuals according to their household incomes and compared them using the mean of the municipality average income. While the differences are again not statistically significant, from panel B of Table 6 one can notice that students below the median or below the bottom 5% face a reduction of 9% SD in their scores for every extra 10 μg/m3 in the day of the tests (cols. 2 and 4), whereas wealthier students are not significantly impacted by a deterioration in air quality.
3.3. Sensitivity Analysis
To assess the robustness of our results, we experiment with a number of alternative specifications. So far we have reported results in which we include only one pollutant at a time in our estimations. Given the generally high correlation between many common local air pollutants, there continues to be debate around whether or not different pollutants have independent health effects (see, e.g., Lipfert et al. 2009; Cao et al. 2011; Gan et al. 2011; Katanoda et al. 2011; Chen et al. 2012). In our sample, O3 and PM10 have a pairwise correlation of 0.10, significant at 5%. Nevertheless, Table 7 provides the results from estimations in which both pollutants are included together. For ease of comparison, we report the baseline models with student fixed effects in columns 1 and 2 and with instrumental variables approach in columns 3 and 4. Columns 5–8 show the results of equations with both pollutants.
| Baseline Model | With Copollutant | |||||||
|---|---|---|---|---|---|---|---|---|
| Variables | FE (1) | FE (2) | IV (3) | IV (4) | FE (5) | FE (6) | IV (7) | IV (8) |
| PM10 (10 μg/m3) | −.08*** | −.08*** | −.08*** | −.15* | ||||
| (.03) | (.01) | (.03) | (.09) | |||||
| 1 if PM10 > 20 | −.11*** | −.13*** | −.05 | −.14** | ||||
| (.04) | (.01) | (.04) | (.06) | |||||
| O3 (10 μg/m3) | .01 | .01 | −.08 | −.07 | ||||
| (.02) | (.02) | (.06) | (.06) | |||||
| F first stage | 55.4 | 71 | 106 | 118 | ||||
| Observations | 2,931,368 | 2,931,368 | 2,405,938 | 2,405,938 | 2,931,368 | 2,931,368 | 2,388,034 | 2,388,034 |
| N municipalities | 47 | 47 | 47 | 47 | 47 | 47 | 47 | 47 |
| R-squared | .07 | .07 | .08 | .08 | .07 | .07 | .08 | .08 |
First, the estimation using student fixed effects presents the same coefficient for the linear form of PM10 with O3, negative but not significant. That is, the effect of an increase of 10% of particulate matter on the day of the exams reduces students’ scores by 8% SD as seen in the baseline model (col. 1). For our instrumental variable specification with copollutant, similarly to our finding using only particulate matter to capture air quality (col. 4), column 8 demonstrates that students passing exams in venues and days with levels of PM10 higher than the WHO’s threshold suffer a reduction of 14% SD in their scores compared to others who are subjected to lower levels of that pollutant. Furthermore, the O3 coefficient seems to continue to suggest no effect on students’ grades. Therefore, in line with Bedi et al.’s (2021) work on cognitive performance and air pollution in São Paulo, we see that PM10 continues to have overall the same impact on test takers, while O3 does not present a significant effect on the latter. These are expected results since PM10 penetrates indoors more easily than O3; hence one would indeed expect a smaller impact, or no impact at all, of the latter pollutant in a copollutant estimation.
In a second exercise we include a 1-day lead in our main specification with the linear form of PM10 to capture common situations, for example, political events and weather shocks. To do so we use data on air pollution from 1 day ahead of the actual date. Table 8 shows that air pollution levels in the day after the exams do not significantly affect students’ performance. Accordingly, an additional 10 μg/m3 on the day of the examinations is still responsible to a decrease of 5% SD in students’ scores, while PM10 measures from the following day have no impact on pupils’ marks. These results suggest that our main findings on students’ cognitive performance are driven not by other factors that started before the test day and correlated with pollution but solely by exposure to poor air quality on the day of the test, hence, a contemporaneous effect.19
We next conduct a third exercise to check the robustness of our results. As we do not know the precise location where the students took their exams, we have so far used the municipality centroid as a proxy for the exam venues and measured pollution using the average of up to three pollution monitoring stations within 10 km of that centroid. However, if a school is located toward the periphery of a municipality, then the monitoring stations within 10 km of the centroid may not be the best measure of pollution levels at that school’s location and, if there are fewer than three monitors within 10 km, we are omitting monitoring stations from our analysis that may be closer to the school’s location. To see whether this is affecting our results we here extend the 10 km limit to 50 km and plot the results in figure 3. We can see that PM10 continues to negatively affect scores, but the impacts fade and the standard errors inflate as the distance rises, with the coefficients corresponding to the 50 km estimation not being statistically significant. This would seem to suggest that the 10 km cut-off more accurately captures pollution concentrations at exam venues and that the pollutants’ levels decay insofar as they travel.20 Cognitive impact of PM10 per distance from the station to the municipality centroid. This figure reports the cognitive impact of air pollution within different distances; 10 km is the baseline model. All the specifications follow the main model with student fixed effects, day controls, and flexible weather condition as in column 3 of Table 3. We note an increase of standard errors, suggesting that the further away the stations, the less precise and significant the estimates.
Finally, to demonstrate the sensitiveness of our main specifications with binned weather controls, table A2 shows that the pollutant’s coefficient becomes positive and significant when we do not include those variables (cols. 1 and 2), corroborating the largely known fact that controlling for meteorological conditions is important. Furthermore, columns 3 and 4 present our specifications using linear weather controls and columns 5 and 6 with their squares and interactions. These results suggest that our main estimates are being driven not by unobserved meteorological conditions correlated with wind and students’ performance but solely by the effects of air pollution.
4. Conclusion
In this paper we take advantage of the individual-level panel data structure of a Brazilian student exam data set to estimate the causal effect of contemporaneous exposure to air pollution on students’ performance in university entrance exams from 2015 to 2017. To identify this relationship we use student fixed effects as our main specification strategy but, as robustness checks, we also experiment with a wide range of control variables such as individual characteristics, parents’ level of education, household income, weather controls (temperature, humidity, and wind speed), and day and municipality fixed effects. To address possible biases in our estimations stemming from classical measurement error and omitted variables, we also perform an instrumental variable approach where we rely on the temporal and geographical variation in wind direction to build our instrument. Our main assumption is that this variation is exogenous since it does not directly affect the students’ scores except through its effect on air pollution.
Our results consistently suggest a harmful effect of air pollution on exam performance. Indeed, we find that an increase of 10 μg/m3 of PM10 on the day of the examination decreases students’ scores by around 6 points (8% SD). Even when including a more flexible measure of our treatment that is utilizing a dummy variable to account for the days in which PM10 exceeded the WHO’s acceptable threshold, our findings still point to negative effects of air pollution on cognitive performance during examinations. We run several sensitivity checks and falsification tests which suggest that our results are not driven by spurious correlations.
We also find suggestive evidence that the effect of air pollution on exam performance appears to differ for males and females, with the performance of the former more adversely affected. This finding is consistent with previous studies, including Jayachandran (2009), Sanders (2012), and Ebenstein et al. (2016). Although the role of gender in the effects of air pollution on respiratory health remains unclear, some epidemiology studies point out that males are more affected than females. For example, Abbey et al. (1998) link PM10 exposure to reduced lung function among males but not females, and Galizia and Kinney (1999) find similar results regarding the exposure to O3. Similarly, Ebenstein et al. (2016) find the effect of pollution on exam performance to differ by gender and cite the increased incidence of asthma among boys as a possible cause. Finally, our results also suggest that poorer students may be more susceptible to air pollution than wealthier exam takers. These findings are also in accordance with previous studies (see, e.g., Neidell 2004; Ebenstein et al. 2016). All in all, our findings provide plausible evidence to suggest that cognitive performance may be hindered by poor air quality, but unequally so.
Notes
Juliana Carneiro is at the University of Bern. Matthew A. Cole (corresponding author) is at the University of Birmingham ([email protected]). Eric Strobl is at the University of Bern. We would like to thank Matthew Neidell and two anonymous referees for their constructive comments. We are also grateful to Tatyana Deryugina for her generous comments.
1. See http://www9.who.int/airpollution/ambient/health-impacts/en/.
2. PM10 represents particulate matter with a diameter of 10 micrometers or less.
3. The information can be found in the CETESB website: https://cetesb.sp.gov.br/.
4. The automatic stations are directly linked to a central computer that registers the pollution concentrations. In the nonautomatic stations, the samples are manually collected at the site to be analyzed in the laboratories of CETESB and Inea.
5. To be clear, if there are three monitors within 10 km we average across those, if there are two we average across those, and if there is only one then we use the daily average of that one monitor. If a municipality has no monitoring stations within 10 km of its centroid, then data for that municipality are dropped from our analysis. In our sensitivity analysis we increase the 10 km limit to 50 km.
6. Figures A1 and A2 show the daily average of both pollutants.
7. Unfortunately, we do not include rainfall as most stations do not present complete data for this variable for the period analyzed in this research.
8. As in Deryugina et al. (2019), wind direction refers to the direction the wind is blowing from.
9. Figure A4 shows the maps of Rio and São Paulo with the total amount of municipalities from the raw data that hosted examinations and the municipalities within our final sample
10. Table A3 presents the number of observations and municipalities for the five samples using the different distances, as well as the ratio of municipalities used in the final sample and the raw data set.
11. We use standardized test scores to account for the different subjects of the examination. Thus, standardized scores = (scores − )/SD; where is the mean and SD is the standard deviation. That is, we normalize the official scale scores to a mean equal to zero and standard deviation equal to 1 (Rosa et al. 2019; Garg et al. 2020).
12. In our econometric models, we add linear and quadratic terms for both temperature and relative humidity, and their linear and quadratic interaction.
13. K-means clustering aims to split n observations into k clusters in which each observation belongs to the cluster with the nearest mean. K-means algorithm minimizes within-cluster variances (squared Euclidean distances) and tends to find clusters of comparable spatial extent, i.e., by latitude and longitude in our case.
14. Our results do not leave doubt concerning the choice. Nonetheless, the k-means response depends on the initial cluster centers chosen by the researchers. Thus, we repeated the process 50 times and found to be the optimal answer 83% of the time. Furthermore, for more details on how to calculate the WSS, η2, and PRE, consult Makles (2012).
15. Note that we consider the direction the wind blows from, instead of the direction that the wind blows to, as in Anderson (2020).
16. Table A1 reports the first-stage estimations and the F-test statistics for the strength of our instruments.
17. For instance, Ebenstein et al. (2016), Chen et al. (2018), and Bondy et al. (2020) use the air quality index (AQI).
18. Figure 2 and table A1 provide the first-stage estimations.
19. Similar results can be found, e.g., in Bedi et al. (2021).
20. Similar results can be found, e.g., in Lai et al. (2018).
References
Abbey, David E., Raoul J. Burchette, Synnøve F. Knutsen, William F. McDonnell, Michael D. Lebowitz, and Paul L. Enright. 1998. Long-term particulate and other air pollutants and lung function in nonsmokers. American Journal of Respiratory and Critical Care Medicine 158 (1): 289–98. Almond, Douglas, Lena Edlund, and Mårten Palme. 2009. Chernobyl’s subclinical legacy: Prenatal exposure to radioactive fallout and school outcomes in Sweden. Quarterly Journal of Economics 124 (4): 1729–72. Anderson, Michael L. 2020. As the wind blows: The effects of long-term exposure to air pollution on mortality. Journal of the European Economic Association 18 (4): 1886–927. Arceo, Eva, Rema Hanna, and Paulina Oliva. 2016. Does the effect of pollution on infant mortality differ between developing and developed countries? Evidence from Mexico City. Economic Journal 126 (591): 257–80. Archsmith, James, Anthony Heyes, and Soodeh Saberian. 2018. Air quality and error quantity: Pollution and performance in a high-skilled, quality-focused occupation. Journal of the Association of Environmental and Resource Economists 5 (4): 827–63. Banzhaf, H. Spencer, and Randall P. Walsh. 2008. Do people vote with their feet? An empirical test of Tiebout. American Economic Review 98 (3): 843–63. Bedi, Arjun S., Marcos Y. Nakaguma, Brandon J. Restrepo, and Matthias Rieger. 2021. Particle pollution and cognition: Evidence from sensitive cognitive tests in Brazil. Journal of the Association of Environmental and Resource Economists 8 (3): 443–74. Bharadwaj, Prashant, Matthew Gibson, Joshua Graff Zivin, and Christopher Neilson. 2017. Gray matters: Fetal pollution exposure and human capital formation. Journal of the Association of Environmental and Resource Economists 4 (2): 505–42. Bondy, Malvina, Sefi Roth, and Lutz Sager. 2020. Crime is in the air: The contemporaneous relationship between air pollution and crime. Journal of the Association of Environmental and Resource Economists 7 (3): 555–85. Cao, Jie, Chunxue Yang, Jianxinn Li, Renjie Chen, Bingheng Chen, Dongfeng Gu, and Haidong Kan. 2011. Association between long-term exposure to outdoor air pollution and mortality in China: A cohort study. Journal of Hazardous Materials 86 (2–3): 1594–1600. Chang, Ton, Joshua Graff Zivin, Tal Gross, and Matthew Neidell. 2016. Particulate pollution and the productivity of pear packers. American Economic Journal: Economic Policy 8 (3): 141–69. ———. 2019. The effect of pollution on worker productivity: Evidence from call center workers in China. American Economic Journal: Applied Economics 11 (1): 151–72. Chen, Siyu, Chongshan Guo, and Xinfei Huang. 2018. Air pollution, student health, and school absences: Evidence from China. Journal of Environmental Economics and Management 92:465–97. Chen, Wei Huang, Chit-Ming Wong, Zongshuang Wang, Thuan Quoc Thach, Bingheng Chen, Haidong Kan, CAPES Collaborative Group, et al. 2012. Short-term exposure to sulfur dioxide and daily mortality in 17 Chinese cities: The China air pollution and health effects study (CAPES). Environmental Research 118:101–6. Chen, Xi, Xiaobo Zhang, and Xin Zhang. 2017. Smog in our brains: Gender differences in the impact of exposure to air pollution on cognitive performance. IZA Discussion paper 10628, Institute of Labor Economics, Bonn. Currie, Janet. 2009. Healthy, wealthy, and wise: Socioeconomic status, poor health in childhood, and human capital development. Journal of Economic Literature 47 (1): 87–122. Currie, Janet, Joshua Graff Zivin, Jamie Mullins, and Matthew Neidell. 2014. What do we know about short- and long-term effects of early-life exposure to pollution? Annual Review of Resource Economics 6 (1): 217–47. Currie, Janet, Eric A. Hanushek, E. Megan Kahn, Matthew Neidell, and Steven G. Rivkin. 2009. Does pollution increase school absences? Review of Economics and Statistics 91 (4): 682–94. Currie, Janet, and Matthew Neidell. 2005. Air pollution and infant health: What can we learn from California’s recent experience? Quarterly Journal of Economics 120 (3): 1003–30. Deryugina, Tatyana, Garth Heutel, Nolan H. Miller, David Molitor, and Julian Reif. 2019. The mortality and medical costs of air pollution: Evidence from changes in wind direction. American Economic Review 109 (12): 4178–219. Dockery, Douglas W., C. Arden Pope, Xiping Xu, John D. Spengler, James H. Ware, Martha E. Fay, Benjamin G. Ferris Jr., and Frank E. Speizer. 1993. An association between air pollution and mortality in six US cities. New England Journal of Medicine 329 (24): 1753–59. Ebenstein, Avraham, Victor Lavy, and Sefi Roth. 2016. The long-run economic consequences of high-stakes examinations: Evidence from transitory variation in pollution. American Economic Journal: Applied Economics 8 (4): 36–65. Estadao. 2019. Infográficos. https://www.estadao.com.br/infograficos/educacao,no-enem-1-a-cada-4-alunos-de-classe-media-triunfa-pobres-sao-1-a-cada-600,953041 (accessed November 12, 2019).Galizia, Audrey, and Patrick L. Kinney. 1999. Long-term residence in areas of high ozone: Associations with respiratory health in a nationwide sample of nonsmoking young adults [DSEE comments]. Environmental Health Perspectives 107 (8): 675–79. Gan, Wen Qi, Mieke Koehoorn, Hugh W. Davies, Paul A. Demers, Lillian Tamburic, and Michael Brauer. 2011. Long-term exposure to traffic-related air pollution and the risk of coronary heart disease hospitalization and mortality. Environmental Health Perspectives 119 (4): 501–7. Garg, Teevrat, Maulik Jagnani, and Vis Taraz. 2020. Temperature and human capital in India. Journal of the Association of Environmental and Resource Economists 7 (6): 1113–50. Graff Zivin, Joshua, Tong Liu, Yingquan Song, Qu Tang, and Peng Zhang. 2020. The unintended impacts of agricultural fires: Human capital in China. Journal of Development Economics 147:102560. Hansen, Anett C., and Harald K. Selte. 2000. Air pollution and sick-leaves. Environmental and Resource Economics 16 (1): 31–50. Heyes, Anthony, Matthew Neidell, and Soodeh Saberian. 2016. The effect of air pollution on investor behavior: Evidence from the S&P 500. Working paper 22753, National Bureau of Economic Research, Cambridge, MA. Hsiang, Solomon, Paulina Oliva, and Reed Walker. 2019. The distribution of environmental damages. Review of Environmental Economics and Policy 13 (1): 83–103. INEP. 2017. Microdados do enem. http://portal.inep.gov.br/web/guest/microdados (accessed August 19, 2019).Itron. 2019. Blogs itron. https://blogs.itron.com/computing-a-weighted-average-wind-speed-and-wind-direction-across-multiple-weather-stations (accessed November 19, 2019).Jayachandran, Seema. 2009. Air quality and early-life mortality evidence from Indonesia’s wildfires. Journal of Human Resources 44 (4): 916–54. Kahn, Matthew E., and Pei Li. 2019. The effect of pollution and heat on high skill public sector worker productivity in China. Working paper 25594, National Bureau of Economic Research, Cambridge, MA. Katanoda, Kota, Tomotaka Sobue, Hiroshi Satoh, Kazuo Tajima, Takaichiro Suzuki, Haruo Nakatsuka, Toshiro Takezaki, et al. 2011. An association between long-term exposure to ambient air pollution and mortality from lung cancer and respiratory diseases in Japan. Journal of Epidemiology 21 (2): 132–43. Lai, Wangyang, Y. Li, X. Tian, and S. Li. 2018. Agricultural fires and cognitive function: Evidence from crop production cycles. Available at SSRN 3039935. Ling, Sean H., and Stephan F. van Eeden. 2009. Particulate matter air pollution exposure: Role in the development and exacerbation of chronic obstructive pulmonary disease. International Journal of Chronic Obstructive Pulmonary Disease 4:233. Lipfert, Frederick W., Ronald E. Wyzga, Jack D. Baty, and J. Philip Miller. 2009. Air pollution and survival within the Washington University–EPRI veterans cohort: Risks based on modeled estimates of ambient levels of hazardous and criteria air pollutants. Journal of the Air and Waste Management Association 59 (4): 473–89. Liu, Haoming, and Alberto Salvo. 2018. Severe air pollution and child absences when schools and parents respond. Journal of Environmental Economics and Management 92:300–330. Makles, Anna. 2012. Stata tip 110: How to get the optimal k-means cluster solution. Stata Journal 12 (2): 347–51. Moretti, Enrico, and Matthew Neidell. 2011. Pollution, health, and avoidance behavior evidence from the ports of Los Angeles. Journal of Human Resources 46 (1): 154–75. Neidell, Matthew. 2004. Air pollution, health, and socio-economic status: The effect of outdoor air quality on childhood asthma. Journal of Health Economics 23 (6): 1209–36. ———. 2009. Information, avoidance behavior, and health: The effect of ozone on asthma hospitalizations. Journal of Human Resources 44 (2): 450–78. Nilsson, J. Peter. 2009. The long-term effects of early childhood lead exposure: Evidence from the phase-out of leaded gasoline. Institute for Labour Market Policy Evaluation (IFAU) Working paper. Ransom, Michael, and C. A. Pope. 2013. Air pollution and school absenteeism: Results from a natural experiment. Unpublished manuscript, Brigham Young University, Department of Economics. Reyes, Jessica Wolpaw. 2011. Childhood lead and academic performance in Massachusetts. New England Public Policy Center Working paper 11-3, Federal Reserve Bank of Boston. Rosa, Leonardo, Marcelo Martins, and Martin Carnoy. 2019. Achievement gains from reconfiguring early schooling: The case of Brazil’s primary education reform. Economics of Education Review 68:1–12. Sager, Lutz. 2019. Estimating the effect of air pollution on road safety using atmospheric temperature inversions. Journal of Environmental Economics and Management 98:102250. Sanders, Nicholas J. 2012. What doesn’t kill you makes you weaker: Prenatal pollution exposure and educational outcomes. Journal of Human Resources 47 (3): 826–50. Schlenker, W., and W. R. Walker. 2015. Airports, air pollution, and contemporaneous health. Review of Economic Studies 83 (2): 768–809. US EPA. 2016. Health effects of ozone in the general population. EPA Course Developers, US Environmental Protection Agency, Washington, DC. Webmet. 2019. The meteorological resource centre. http://www.webmet.com/met_monitoring/622.html (accessed November 11, 2019).Weinhold, Bob. 2008. Ozone nation: EPA standard panned by the people. https://ehp.niehs.nih.gov/doi/full/10.1289/ehp.116-a302 (accessed August 19, 2019).Zapata, Juliana, Beatriz Pont, Diana Toledo Figueroa, Judith Peterka, and Sylvain Fraccola. 2015. Education policy outlook: Brazil. Paris: OECD. Zweig, Jacqueline S., John C. Ham, and Edward L. Avol. 2009. Air pollution and academic performance: Evidence from California schools. University of Maryland, Department of Economics.





