ArticlePDF Available

The impact of exposure to air pollution on cognitive performance

Authors:

Abstract and Figures

Significance Most of the population in developing countries live in places with unsafe air. Utilizing variations in transitory and cumulative air pollution exposures for the same individuals over time in China, we provide evidence that polluted air may impede cognitive ability as people become older, especially for less educated men. Cutting annual mean concentration of particulate matter smaller than 10 μm (PM10) in China to the Environmental Protection Agency’s standard (50 μg/m ³ ) would move people from the median to the 63rd percentile (verbal test scores) and the 58th percentile (math test scores), respectively. The damage on the aging brain by air pollution likely imposes substantial health and economic costs, considering that cognitive functioning is critical for the elderly for both running daily errands and making high-stake decisions.
Content may be subject to copyright.
The impact of exposure to air pollution on
cognitive performance
Xin Zhang
a,1
, Xi Chen
b,c,1
, and Xiaobo Zhang
d,e,2
a
School of Statistics, Beijing Normal University, Beijing 100875, China;
b
Department of Health Policy and Management, Yale School of Public Health, New
Haven, CT 06520;
c
Department of Economics, Yale University, New Haven, CT 06511;
d
National School of Development, Peking University, Beijing 100871,
China; and
e
Division of Development Strategy and Governance, International Food Policy Research Institute, Washington, DC 20005
Edited by Robert M. Hauser, Center for Demography of Health and Aging, Madison, WI, and approved July 23, 2018 (received for review June 8, 2018)
This paper examines the effect of both cumulative and transitory
exposures to air pollution for the same individuals over time on
cognitive performance by matching a nationally representative
longitudinal survey and air quality data in China according to
the exact time and geographic locations of the cognitive tests.
We find that long-term exposure to air pollution impedes cogni-
tive performance in verbal and math tests. We provide evidence
that the effect of air pollution on verbal tests becomes more pro-
nounced as people age, especially for men and the less educated.
The damage on the aging brain by air pollution likely imposes
substantial health and economic costs, considering that cognitive
functioning is critical for the elderly for both running daily errands
and making high-stake decisions.
aging
|
cognitive decline
|
air pollution
|
gender difference
|
China
While a large body of literature has shown that air pollution
harms human health, in terms of life expectancy (1), ill-
ness and hospitalization (2), child health (3), health behavior (4),
and dementia (57), knowledge about the potential conse-
quences of air pollution on cognitive abilities is more limited. A
few existing studies on the impact of air pollution on cognition
have mainly focused on young students (811). It is unclear
whether their findings hold for the whole population or not, in
particular for older cohort. Our paper fills this knowledge gap by
examining the pollutioncognition relationship by age in China
based on a nationally representative longitudinal dataset at the
individual level.
We find that air pollution impairs verbal tests, and the effect
becomes stronger as people age, especially for less educated men.
Cognitive decline or impairment are risk factors of Alzheimers
disease and other forms of dementia for elderly persons. As the
most expensive form of cognitive decline, Alzheimersdiseasealone
costs $226 billion of health services and 18 billion labor hours of
unpaid caregiving in 2015 (6). Moreover, given that senior citizens
have to make a host of complex high-stake economic decisions,
such as purchasing health insurance and planning retirement, the
decay in cognitive ability induced by air pollution will likely impair
the quality of the important decisions (12). The damage on the
aging brain by air pollution likely imposes substantial health and
economic cost, which has been neglected in the policy discourse.
Therefore, the finding on the detrimental effect of air pollution on
the aging brain has important policy implications.
On the technical level, our paper has tried to overcome several
common challenges facing this strand of empirical studies. First,
we address the potential problem of omitted variables, which
may be correlated with both cognition and exposure to air pol-
lution, on estimation bias by using a panel data at the individual
level. Most studies, except for those of Ebenstein et al. (10) and
Marcotte (13), fail to account for individual-level heterogeneity
due to data limitation. For instance, Ham et al. (8) only control
for school-grade fixed effects; Bharadwaj et al. (14) include only
sibling fixed effects. In this study, because we have access to a
longitudinal dataset, the China Family Panel Studies (CFPS), we
can remove individual-level unobservable factors.
Second, we have matched exposure to local environmental
stressors with individual cognitive performance according to the
exact time of test taking. This is more precise than in previous
studies, for instance, that of Ham et al. (8), who match yearly air
pollution with average standardized test scores at the school-
grade level. Third, most existing studies consider either the ef-
fects of transitory or cumulative exposure to air pollution, but
rarely both effects simultaneously, except for Marcotte (13). For
example, Ham et al. (8) and Ebenstein et al. (10) focus on con-
temporaneous exposure; Bharadwaj et al. (14), Molina (15), and
Sanders (16) examine the effect of cumulative exposure. We are
among the first to examine the cognitive impact of cumulative
exposure to air pollution while controlling for contemporaneous
exposure. By controlling for the latter, we can evaluate the relative
importance of transitory and accumulative effects. We find that
the accumulative effect dominates.
Given that cognitive ability shapes human behavior and de-
cision making, our result provides supporting evidence on the
findings about the negative effect of air pollution on decision
making (7, 17), risk attitude (11), and behavior (11, 18). The
damage on cognitive ability by air pollution also likely impedes
the development of human capital. In fact, a few studies have
found that exposure to air pollution lowers educational attain-
ment (10, 16) and results in lower labor productivity (1922).
Air pollution is a ubiquitous problem in developing countries.
According to the global ambient air pollution database compiled
Significance
Most of the population in developing countries live in places
with unsafe air. Utilizing variations in transitory and cumula-
tive air pollution exposures for the same individuals over time
in China, we provide evidence that polluted air may impede
cognitive ability as people become older, especially for less
educated men. Cutting annual mean concentration of particu-
late matter smaller than 10 μm (PM10) in China to the Envi-
ronmental Protection Agencys standard (50 μg/m
3
) would
move people from the median to the 63rd percentile (verbal
test scores) and the 58th percentile (math test scores), re-
spectively. The damage on the aging brain by air pollution
likely imposes substantial health and economic costs, consid-
ering that cognitive functioning is critical for the elderly for
both running daily errands and making high-stake decisions.
Author contributions: X.C. and Xiaobo Zhang designed research; Xin Zhang, X.C., and
Xiaobo Zhang performed research; Xin Zhang analyzed data; and Xin Zhang, X.C., and
Xiaobo Zhang wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
Published under the PNAS license.
1
Xin Zhang and X.C. contributed equally to this work.
2
To whom correspondence should be addressed. Email: x.zhang@nsd.pku.edu.cn.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1809474115/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1809474115 PNAS Latest Articles
|
1of5
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
SUSTAINABILITY
SCIENCE
by the World Health Organization (www.who.int/phe/health_topics/
outdoorair/databases/cities/en/), the top 20 most polluted cities are
all in developing countries. Almost all of the cities (98%) in low-
and middle-income countries with more than 100,000 residents fail
to meet World Health Organization air quality guidelines. There-
fore, the research findings on China, the largest developing country
with severe air pollution, can also shed light on other developing
countries.
The remainder of the paper is organized as follows. Data
Sources describes the data, and Econometric Model lays out the
empirical strategy. Empirical Results presents our main findings.
Conclusions provides some conclusions. In SI Appendix, we also
discuss the scientific background of this study and potential
mechanisms in detail.
Data Sources
The dataset for this analysis is based on several sources. The
cognitive test scores come from the CFPS, a nationally repre-
sentative survey of Chinese families and individuals. The waves
2010 and 2014 contain the same cognitive ability module, that is,
24 standardized mathematics questions and 34 word-recognition
questions. All of these questions are sorted in ascending order of
difficulty, and the final test score is defined as the rank of the
hardest question that a respondent is able to answer correctly.
The survey also provides exact information about the geographic
locations and dates of interviews for all respondents, which en-
ables us to match test scores with local air quality data more
precisely.
Air quality is measured using the air pollution index (API),
which is calculated based on daily readings of three air pollut-
ants, namely sulfur dioxide (SO
2
), nitrogen dioxide (NO
2
), and
particulate matter smaller than 10 μm (PM10). The API ranges
from 0 to 500, with larger values indicating worse air quality. Daily
API observations are obtained from the city-level air quality re-
port published by the Chinese Ministry of Environmental Pro-
tection. The report includes 86 major cities in 2000 and covers
most of the cities in China in 2014.
Our analysis also includes rich weather measures on the in-
terview date, enabling us to separate the impact of air pollution
from general weather patterns. The weather data are derived
from the National Centers for Environmental Information of the
US National Oceanic and Atmospheric Administration. The
dataset contains daily records of rich weather conditions from
402 monitoring stations in China.
We match city-level API with CFPS samples in the following
way. If a CFPS county is within an API reporting city, we use the
citys API readings as the countys readings. If a CFPS county is
not located in any cities with API readings, we match it to the
nearest API reporting city within a radius of 40 km according to
the distance between the CFPS county centroid and the city
boundaries. In SI Appendix,Part 2: Description of Data, we show
the results are robust to a wide range of matching radiuses and
alternative matching strategies. The final dataset used in this
study includes 31,959 observations. SI Appendix describes the
data and the matching procedure in greater detail.
Econometric Model
Our baseline econometric specification is as follows:
Scoreijt =α1Pjt +α2·1
kX
k1
n=0
Pj,tn+X
ijtβ+W
jtϕ
+T
jtγ+λi+δj+ηt+fðtÞ+«ijt .[1]
The dependent variable Score
ijt
is the cognition test scores of
respondent iin county jat date t.P
jt
is the contemporaneous
air quality measure at date t. The key variable ð1=kÞPk1
n=0Pj,tnis
themeanAPIreadinginthepastkdays, which measures cumula-
tive exposure. X
ijt
is a set of the observable demographic correlates
of the respondents. We also control for a vector of contemporane-
ous weather conditions W
jt
and a vector of county-level character-
istics T
jt
to account for factors that are correlated with both test
scores and air quality. λ
i
denotes individual fixed effects. δ
j
repre-
sents county fixed effects, which cannot be wiped out by individual
fixed effects since some respondents do not live in the same coun-
ties across the two waves. η
t
indicates month, day of week, and
postmeridiem hour fixed effects. f(t) is the quadratic monthly time
trend that ranges from 1 (January 2010) to 60 (December 2014). «
ijt
is the error term. SEs are clustered at the county level.
By conditioning on the individual fixed effects, the key pa-
rameters are identified by making use of variations in exposure
to air pollution for the same respondent in the 2010 and 2014
surveys. SI Appendix, Fig. S1 displays the monthly distribution of
interview times in the two waves of the CFPS survey. Although a
majority of interviews were conducted in July and August when
college students were employed as numerators, the survey spans
all months and seasons, providing us with large temporal varia-
tions. There is a concern that the results are mainly driven by the
skewed sample distribution in the summer months, when air
pollution is not as serious as in winter. SI Appendix, Fig. S11 and
Table S12 also show that our findings still hold if giving an in-
terview in winter greater weight than that in nonwinter so that the
two periods share the same weight. The study was approved by the
institutional review board (IRB) at Peking University (Approval
IRB00001052-14010). All participants gave informed consent in
accordance with policies of the IRB at Peking University.
Empirical Results
Estimates of the Effect of Air Pollution on Cognitive Test Scores. SI
Appendix, Table S1 reports the results for estimation of Eq. 1
using seven windows of air pollution exposure, that is, 1-d, 7-d,
30-d, 90-d, 1-y, 2-y, and 3-y exposures. Panel A presents the es-
timates for the verbal test scores, while panel B displays the
results for the math test scores. Three findings are apparent from
the table. First, in general, air pollution inhibits respondentstest
performance. Except for the effects of 1-d and 7-d air pollution
exposure on math test scores (first and second columns in panel
B), all of the coefficients for mean APIs over a longer period are
negative and statistically significant. Second, the damage of air
pollution on cognitive performance is more sizable when using
longer window of exposure measure. As shown at the bottom in
panel A, an increase in the 7-d-mean API by 1 SD lowers verbal
test scores by 0.278 point (0.026 SD), while a 1 SD increase in
average API over 3 y before the interview is associated with 1.132
points (0.108 SD) drop in verbal test scores. Third, air pollution
exposure appears to exert a more negative effect on verbal test
performance than math test performance. The changes in SDs in
the parentheses presented at the bottom of panel A for verbal
test scores are more pronounced than the corresponding ones in
panel B for math test scores. SI Appendix, Table S11 further
confirms that the baseline results are robust to alternative speci-
fications without controlling for potentially endogenous variables
or individual fixed effects.
Fig. 1 visualizes our baseline results obtained from SI Appendix,
Tables S2a and S2b.Fig.1Arefers to the results for verbal tests,
while Fig. 1Bis for math test scores. Each figure presents the
estimated coefficients for different windows of the mean API
readings, together with their 95% and 99% confidence intervals,
for the male and female subsamples, respectively. As shown in Fig.
1A, exposure to air pollution is associated with lower verbal test
scores for both men and women regardless of the length of ex-
posure. In general, the effect becomes larger as the duration of
exposure to air pollution increases. Men are more vulnerable to
air pollution than women. The gender difference is statistically
significant, as shown by the asterisks in Fig. 1A. As illustrated in SI
2of5
|
www.pnas.org/cgi/doi/10.1073/pnas.1809474115 Zhang et al.
Appendix,Part 6: Scientific Backgrounds and Potential Mechanisms,
this gender gap is likely caused by their different sizes of white
matter activated, which can be reduced by air pollution.
As presented in Fig. 1B, the effect on math tests is more muted
than that on verbal tests. SI Appendix,Part 6: Scientific Back-
grounds and Potential Mechanisms speculates that the observed
patterns are probably associated with gender difference in white
matter and gray matter. Air pollution has a stronger effect on
white matter (required more by verbal tests) than on gray matter
(required more by math tests). Since men have a much smaller
amount of white matter activated during intelligence tests, their
cognitive performance, especially in the verbal domain, tends to
be more affected by exposure to air pollution.
Estimates of the Age-Cohort Effect of Air Pollution on Cognitive Test
Scores. To understand how air pollution affects cognition as
people age, we examine the accumulative effects for verbal and
math test scores, respectively, for different age cohorts. The age
cohort effects are measured by the interaction terms between
3-y-mean API and age cohort dummies 2534, 3544, 4554, 55
64, and 65+in 2014. The age band 1024 is the reference cate-
gory. Fig. 2 plots the estimated coefficients on the interaction
terms for the male and female subsamples coupled with 95% and
99% confidence intervals. The numerical results are shown in SI
Appendix, Table S3. In Fig. 2, A and B present results for verbal
and math tests, respectively. Compared with younger age co-
horts, the negative effect on the verbal test performance is more
Fig. 1. The figures plot the estimated coefficients on air pollution for the male and female subsamples with 95% and 99% confidence intervals based on the
estimates in SI Appendix, Tables S2a and S2b.Aand Brefer to verbal and math test scores, respectively. Air pollution data are matched between each CFPS
county centroid and its nearest API reporting city boundary within a radius of 40 km (i.e., 25 miles). The asterisks in the figure indicate the significance of the
malefemale difference denoting the results of Wald tests: *10% significance level; **5% significance level; ***1% significance level.
Fig. 2. The age cohort effects of air pollution on cognitive test scores include interaction terms between 3-y-mean API and age cohort dummies 2534,
3544, 4554, 5564, and 65+in 2014. The age band 1024 is the reference category. The figures plot the estimated coefficients on the interaction terms
for the male and female subsamples with 95% and 99% confidence intervals based on the estimates in SI Appendix,TableS3.Aand Brefer to verbal and
math test scores, respectively. Air pollution data are matched between each CFPS county centroid and its nearest API reporting city boundary within a
radius of 40 km (i.e., 25 miles). The asterisks in the figure indicate the significance of the malefemale difference denoting the results of Wald tests: *10%
significance level; **5% significance level; ***1% significance level.
Zhang et al. PNAS Latest Articles
|
3of5
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
SUSTAINABILITY
SCIENCE
pronounced for the older cohorts, especially among males. As a
result, the gender gap in the decline of verbal skills widens as
people age. Such pattern, however, is less noticeable for the
math tests.
Estimates of the Age-Cohort Effect of Air Pollution on Cognitive Test
Scores by Educational Attainment. We repeat the exercises in Fig. 2
to identify potential heterogeneous effects of air pollution on
verbal test scores by educational attainment, that is, primary
school or below versus middle school or above. Fig. 3 and SI
Appendix, Table S4a display the estimated coefficients on the
interaction terms for the male and female subsamples, together
with 95% and 99% confidence intervals across age cohorts. As
shown in Fig. 3A, the effect for men over 44 y old with primary
school education or below is highly negative. Among the more
educated subsample (Fig. 3B), the negative effect only shows up
for men aged 65 and above. Such pattern, however, is not evident
for older women regardless of their education. SI Appendix,
Table S4b further displays age cohort effects of air pollution on
math test scores by educational attainment. There is no clear
difference.
Falsification Tests. Some time-variant unobserved factors, such as
migrating to a different city with more ambient air and better-
paying job between 2012 and 2014, may affect both cognitive test
scores and exposure to air pollution even after controlling for
individual fixed effects. To address this concern, SI Appendix,
Fig. S10 reports a falsification test, examining whether API
readings on the days after cognitive tests affect test scores. If the
time series of API readings embody some unobserved factors
that are correlated with the outcome variables, using the API
readings after the test to replace the current and past API
readings in regressions would yield similar results. However, for
the whole sample as well as the male and female subsamples, all
of the coefficients are not statistically different from zero, largely
dismissing the concern about potential omitted variables. It is
worth noting that the coefficients in this falsification test are
smaller in size than the main effects.
Robustness. People may become more impatient or uncooperative
when exposed to more polluted air. Therefore, it is possible that
the observed negative effect on cognitive performance is due to
behavioral change rather than impaired cognition. To check this
possibility, we examine the impact of exposure to air pollution and
patience and cooperation during the interview in SI Appendix,
Table S13. None of the coefficients for API is significant, largely
dismissing this channel. Changes in the brain chemistry or com-
position are likely more plausible channels between air pollution
and cognition. It is beyond the scope of this paper to test the exact
mechanism, so we leave it as agenda for future research.
Our baseline results are also robust to a wide variety of
specification checks. SI Appendix, Fig. S12 and Table S14 doc-
ument that migration is unlikely to significantly bias our esti-
mates. SI Appendix, Fig. S13 and Table S15 show that the results
are qualitatively unchanged after excluding polluted occupations.
Furthermore, as revealed in SI Appendix, Table S16, the baseline
results are robust to controlling for province-by-year fixed effects
and clustering SEs at the province level.
Interpretation. SI Appendix,Part 5: Estimating Movement in the
Test Distribution Using the Coefficient Estimates calculates
movement in test distribution using the coefficient estimates.
Reducing the population-weighted annual mean concentration
of PM10 over 2014 in China to levels below the National Ambient
Air Quality Standards published by the US Environmental Pro-
tection Agency will on average lift verbal test scores by 2.41 points
(or the movement of people from the median to the 63rd per-
centile in the verbal test distribution) and math test scores by 0.39
point (or the movement of people from the median to the 58th
percentile in the math test distribution).
The effect is particularly large for less educated men older
than 64. A 1 SD decrease in 3-y-mean API leads to an increase in
verbal test scores by 9.18 points (or the movement of people
from the median to the 87th percentile in the verbal test distri-
bution) for this group relative to the cohort younger than 25. The
effect remains sizable for more educated older men. A 1 SD
decrease in 3-y-mean API is associated with an increase in verbal
test scores by 1.88 points (or the movement of people from the
Fig. 3. The age cohort effects of air pollution on verbal test scores by educational attainment include interaction terms between 3-y-mean API and age cohort
dummies 2534, 3544, 4554, 5564, and 65+in 2014. The age band 1024 is the reference category. The figures plot the estimated coefficients on the interaction
terms for the male and female subsamples with 95% and 99% confidence intervals based on the estimates in SI Appendix,TableS4a.Arefers to the subsample
with education level at the primary school or below, while Bincludes the subsample with middle school education or above. Air pollution data are matched
between each CFPS county centroid and its nearest API reporting city boundary within a radius of 40 km (i.e., 25 miles). The asterisks in the figure indicate the
significance of the malefemale difference denoting the results of Wald tests: *10% significance level; **5% significance level; ***1% significance level.
4of5
|
www.pnas.org/cgi/doi/10.1073/pnas.1809474115 Zhang et al.
median to the 69th percentile in the verbal test distribution) for
them relative to their younger counterpart.
Conclusions
This paper estimates the contemporaneous and cumulative im-
pacts of air pollution on cognition by matching the scores of
verbal and math tests given to people age 10 and above in a na-
tionally representative survey with local air quality data according
to the exact dates and locations of the interviews. We find that
accumulative exposure to air pollution impedes verbal test scores.
As people age, the negative effect becomes more pronounced,
especially for men. The gender gap is particularly large for the
less educated.
Our findings about the damaging effect of air pollution on
cognition, particularly on the aging brain, imply that the indirect
effect on social welfare could be much larger than previously
thought. A narrow focus on the negative effect on health may
underestimate the total cost of air pollution.
ACKNOWLEDGMENTS. We appreciate the Institute of Social Science Survey
at Peking University for providing us with the CFPS data, and the Qingyue
Open Environmental Data Center for the support on environmental data
processing. We are grateful for comments from participants at various
seminars and conferences: Institute of Labor Economics (IZA) (2016), Cornell
(2016), University of Pennsylvania (2016), Yale (2016), Peking University
(2016), Shanghai JiaoTong University (2016), University of Minnesota
(2016), Tsinghua University (2016), US Council on Foreign Relations (2016),
Keio University (2017), Shanghai University of Finance and Economics (2017),
and Renmin University of China (2018). This study is funded by the Yale
Macmillan Center Faculty Research Fund, the US Federal PEPPER Center
Scholar Award (P30AG021342), two NIH/National Institute on Aging Grants
(1 R03 AG048920 and K01AG053408), the China Postdoctoral Science Foun-
dation Grants (2017M620653 and 2018T110057), and the Fundamental Re-
search Funds for the Central Universities. The views expressed herein and
any remaining errors are the authorsand do not represent any
official agency.
1. Chay KY, Greenstone M (2003) The impact of air pollution on infant mortality: Evi-
dence from geographic variation in pollution shocks induced by a recession. Q J Econ
118:11211167.
2. Cohen AJ, et al. (2005) The global burden of disease due to outdoor air pollution.
J Toxicol Environ Health A 68:13011307.
3. Currie J, Graff Zivin J, Mullins J, Neidell M (2014) What do we know about short- and
long-term effects of early-life exposure to pollution? Annu Rev Resour Econ 6:
217247.
4. Sun C, Kahn ME, Zheng SQ (2017) Self-protection investment exacerbates air pollution
exposure inequality in urban China. Ecol Econ 131:468474.
5. Cacciottolo M, et al. (2017) Particulate air pollutants, APOE alleles and their contri-
butions to cognitive impairment in older women and to amyloidogenesis in experi-
mental models. Transl Psychiatry 7:e1022.
6. Alzheimers Association (2015) 2015 Alzheimers disease facts and figures. Alzheimers
Dement 11:332384.
7. Chen H, et al. (2017) Living near major roads and the incidence of dementia, Par-
kinsons disease, and multiple sclerosis: A population-based cohort study. Lancet 389:
718726.
8. Ham J, Zweig J, Avol E (2014) Pollution, test scores and distribution of academic
achievement: Evidence from California schools 20022008. Available at econweb.
umd.edu/ham/Pollution_all_accept_feb%202014a.pdf. Accessed December 18, 2016.
9. Stafford TM (2015) Indoor air quality and academic performance. J Environ Econ
Manage 70:3450.
10. Ebenstein A, Lavy V, Roth S (2016) The long-run economic consequences of high-stakes
examinations: Evidence from transitory variation in pollution. Am Econ J Appl Econ 8:3665.
11. Chew SH, Huang W, Li X (2018) Haze and decision making: A natural laboratory ex-
periment. Working paper (National University of Singapore, Singapore).
12. Bishop K, Ketcham J, Kuminoff N, Powers C (2018) Hazed and confused: Air pollution,
cognitive functioning, and financial decision making among the elderly. Working
paper (Arizona State University, Tempe, AZ).
13. Marcotte DE (2016) Something in the air? Pollution, allergens and childrens cognitive
functioning (IZAInstitute of Labor Economics, Bonn), IZA Discussion Paper 9689.
14. Bharadwaj P, Gibson M, Graff Zivin J, Neilson C (2017) Gray matters: Fetal pollution
exposure and human capital formation. J Assoc Environ Resour Econ 4:505542.
15. Molina T (2016) Pollution, ability, and gender-specific investment responses to
shocks. Available at https://dornsife.usc.edu/assets/sites/524/docs/Seminars_Fall_2016/
Molina_Pollution_Sep2016.pdf. Accessed December 18, 2016.
16. Sanders NJ (2012) What doesnt kill you makes you weaker: Prenatal pollution ex-
posure and educational outcomes. J Hum Resour 47:826850.
17. Block ML, et al. (2012) The outdoor air pollution and brain health workshop.
Neurotoxicology 33:972984.
18. Herrnstadt E, Muehlegger E (2016) Air pollution and criminal activity: Evidence from
chicago microdata. NBER working paper no. 21787 (National Bureau of Economic
Research, Cambridge, MA).
19. Zivin JG, Neidell M (2012) The impact of pollution on worker productivity. Am Econ
Rev 102:36523673.
20. Chang T, Graff Zivin J, Gross T, Neidell M (2014) The effect of pollution on office
workers: Evidence from call center workers in China. NBER working paper 22328
(National Bureau of Economic Research, Cambridge, MA).
21. Chang T, Graff Zivin J, Gross T, Neidell M (2016) Particulate pollution and the pro-
ductivity of pear packers. Am Econ J Econ Policy 8:141169.
22. He JX, Liu HM, Salvo A (March 3, 2018) Severe air pollution and labor productivity:
Evidence from industrial towns in China. Am Econ J Appl Econ, 10.2139/ssrn.2581311.
Zhang et al. PNAS Latest Articles
|
5of5
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
SUSTAINABILITY
SCIENCE
www.pnas.org/cgi/doi/10.1073/pnas. 115
1
SUPPORTING INFORMATION
The Impact of Exposure to Air Pollution on Cognitive Performance
PART 1: Description of Data
A. Cognitive Test Scores Data
We utilize cognitive test scores from the China Family Panel Studies (CFPS), a
nationally representative survey of Chinese families and individuals conducted in 2010 and
2014. The CFPS is funded by Peking University and carried out by the university’s Institute
of Social Science Survey (1). The survey uses multistage probability proportional to size
sampling with implicit stratification to better represent Chinese society. The 2010 CFPS
baseline sample is drawn through three stages (i.e. county, village, and household) from 25
provinces. The 162 randomly chosen counties largely represent Chinese society (2). The
CFPS includes questions on a wide range of topics for families and individuals, including
their economic activities, education outcomes, family dynamics and relationships, health,
and cognitive abilities.
The CFPS is suitable for our study for several reasons. First, the survey includes
several standardized cognitive tests. Second, exact information about the geographic
locations and dates of interviews is available to us for all respondents, enabling us to
precisely match individual test scores in the survey with local air-quality data. Third, the
longitudinal data allow us to remove unobserved individual factors that may bias estimates.
Further, the survey embodies rich information at multiple levels, allowing us to control for
a wide range of covariates. Finally, because the cognitive tests are administered to all age
cohorts older than 10, we can study the effects of air pollution on different age groups.
CFPS 2010 and CFPS 2014 contain the same cognitive ability module, i.e., 24
standardized mathematics questions and 34 word-recognition questions. All these
1809474
2
questions are obtained from standard textbooks and are sorted in ascending order of
difficulty. The starting question depends on the respondent’s education level. Specifically,
those whose education level is primary school or below start with the 1st question; those
who attended middle school begin with the 9th question in the verbal test and the 5th
question in the math test; and those who finished high school or above start with the 21st
question in the verbal test and the 13th question in the math test. The test ends when the
individual incorrectly answers three questions in succession. The final test score is defined
as the rank of the hardest question a respondent is able to answer correctly. If the respondent
fails to answer any questions during the test, his or her test score is assigned as the rank of
the starting question minus one. For example, a respondent with middle school education
begins with the 9th question in the verbal test. If the hardest question he can correctly
answer is the 14th question, then his verbal test scores would be 14. However, if he fails
the 9th, 10th, and 11th questions consecutively, his verbal test scores would be 8. The
respondents did not know the rules before they were interviewed, so they did not have the
incentive to fail the tests on purpose.
B. Air Pollution and Weather Data
We measure air quality using the air pollution index (API), which is aggregated based
on daily readings for three atmospheric pollutants, namely sulfur dioxide (SO2), nitrogen
dioxide (NO2), and particulate matter smaller than 10 micrometers (PM10).
1
Carbon
monoxide (CO), ozone, and particulate matter smaller than 2.5 micrometers (PM2.5) were
not added to the basket of the index until 2014. Because all the cognitive tests were
administered between 2010 and 2014, we transform the air quality index (AQI), which
includes six pollutants, to the API in 2014 and use the API in our paper. The API ranges
1
We use the Chinese Ministry of Environmental Protection’s (MEP’s) breakpoints table (see Table S5) and
the following formula to generate the API measurement: IP = ((IHI - ILO) / (BPHI - BPLO)) * (CP - BPLO) + ILO,
where IP is the index for pollutant P, CP is the rounded concentration of pollutant P, BPHI is the breakpoint
that is greater than or equal to CP, BPLO is the breakpoint that is less than or equal to CP, IHI is the API value
corresponding to BPHI, and ILO is the API value corresponding to BPLO. The API represents the highest index
value calculated for each pollutant.
3
from 0 to 500, with larger values indicating worse air quality. Daily API observations are
obtained from the city-level air-quality report published by the Chinese Ministry of
Environmental Protection (MEP) (3). The report includes 86 major cities in 2000 and
covers most of the cities in China in 2014. If the government indeed manipulates the API
data as suggested by Chen et al. (4) and Ghanem and Zhang (5), using the official API data
would underestimate the true impact of air pollution. In this case, our estimates would
represent a lower bound. Figure S2 plots the daily API in China from 2010 to 2014,
showing large temporal and regional variations.
The weather data are derived from the National Centers for Environmental
Information of the US National Oceanic and Atmospheric Administration (6). The dataset
contains daily records of rich weather conditions from 402 monitoring stations in China.
Graff Zivin, Hsiang, and Neidell (7) find that high temperature is associated with
significant decreases in cognitive performance on math in the short run. Hence, we control
for a set of temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total
precipitation, mean wind speed, and a dummy for bad weather to capture the effects. Bad
weather is defined as fog, rain/drizzle, snow/ice pellets, hail, thunder, and tornadoes/funnel
clouds in the data.
C. Other Covariates
As a complement to the CFPS, we also use the China county census data to control
for confounders that might vary across locations and years in China. For example, in the
empirical analysis, we include a vector of county-level characteristics, including GDP per
capita (deflated to 2010 yuan), population density, and industrial value share. These
additional covariates are taken from the China Data Center of the University of Michigan
and linked to the CFPS counties by year (8). Besides, the CFPS data also provide us with
rich demographic characteristics at the individual level. We control gender, age and its
square and cubic terms, log form of household per capita income and years of education in
the main regressions. Table S6 describes all the key variables and their summary statistics.
4
PART 2: Robustness of Results to Alternative Matching Radiuses and Strategies
We employ two matching methods. In each method, we test a wide range of matching
radiuses. The findings are robust to different matching radiuses and strategies. Section A
and Section B below describe them in detail.
A. Matching CFPS Counties to the Boundaries of API Reporting Cities
As the first alternative matching strategy, we calculate the distance between the
centroid of each CFPS county and the boundary of API reporting cities. The MEP of China
reports APIs at the city level. Figure S3 plots the boundaries of API cities.
2
In specific, if
a CFPS county is located within an API reporting city, we set the matching distance as zero
and treat the city’s API readings as the CFPS county’s readings. If a CFPS county is not
located in any cities with API readings, we match it to the nearest API reporting city within
40 kilometers of radius between the CFPS county centroid and city boundaries. We use the
radius of 40 km (i.e. 25 miles) in our analyses to keep consistency with the convention of
the literature (9).
There is a tradeoff between the precision of matching and sample attenuation. When
the radius becomes smaller, the matched counties becomes more precise at the cost of
losing samples. In order to evaluate the tradeoff, Table S7 and Table S8 report estimation
results based on nine different cutoff radiuses ranging from zero km to 80 km for the verbal
test scores and math test scores, respectively. For example, Panel A of Table S7 reports
regression results on a subsample including only respondents who live in API reporting
cities (matching distance equals to zero) for verbal test scores. While in Panel I of Table
S7, the sample is expanded to include residents living within 80 km to the nearest API cities.
The main findings hold no matter which radius is used, suggesting that measurement errors
and attenuation bias do not affect our key findings. In particular, as shown in Figure S5
2
Due to the confidential agreement, we are not allowed to plot the locations of CFPS counties on the map.
5
and Figure S6, the results for age-cohort effects are robust to the choice of radius.
B. Matching CFPS Counties to the Monitoring Stations
The city-level APIs are computed based on readings in multiple monitoring stations.
While we have access to the exact latitude and longitude information of all the monitoring
stations in each API reporting city in 2014, unfortunately the readings at the monitoring
station level are not available in 2010. Figure S4 displays the spatial distribution of air
quality monitoring stations.
3
To improve the matching precision, our second alternative
matching strategy assigns the reported city-level APIs in 2010 to all stations within the city
boundary and match CFPS counties to the nearest air quality monitoring station within a
specific radius. This alternative strategy assumes that the geographic locations of air quality
monitoring stations do not change much between 2010 and 2014, and variation in daily air
pollution across monitoring stations within each city is small. All the regressions are
weighted by the inverse distance to monitoring stations. As presented in Table S9 and
Table S10, our results using this new matching strategy are robust to various radiuses
between 40 km and 90 km. Besides, Figure S7 and Figure S8 plot the results for age-
cohort effects across various radiuses. Once again, the pattern is consistent: while there is
a negative effect on verbal tests for older men, the effect is much more muted on math tests.
C. Matching CFPS with Weather Data
The weather conditions are obtained as the inverse distance-weighted average of all
monitoring stations within a radius of 100 km of the county centroid. The matching radius
is comparable to those used in Deschenes, Greenstone and Guryan (10) and Deschenes and
Greenstone (11) for the weather data. The binary indicator for bad weather comes from the
nearest monitoring station.
D. Number of Observations in the Final Data
3
Still, due to the confidential agreement, we are not allowed to plot the locations of CFPS counties on the
map.
6
The CFPS surveyed a balanced panel of 25,486 individual respondents over age 10 in
2010 and 2014, for a total of 50,972 observations.
4
Of the individuals surveyed in both
waves, 282 are missing values for cognitive test scores. Among the remaining 50,789
observations, 37,918 observations could be matched to API and weather data.
5
Due to
some missing values for household demographics, the final dataset used in this study
includes 31,955 observations. Figure S9 displays the 24-hour time distribution of
respondents who took the cognitive tests and the hourly pollutant concentration. Most of
the cognition tests were conducted in the afternoon and evening. Among the three
pollutants, PM10 is a dominant one throughout the day.
PART 3: Falsification Tests
In the falsification test, we employ a strategy similar to Bensnes (12), which tests the
effects of API readings on the days after the interviews on cognitive test scores. Figure
S10 presents the estimated coefficients with their 95 and 99 percent confidence intervals
from a regression of test scores on API readings one to six days into the future by gender.
For both the whole sample and the subsamples, all the coefficients are statistically
indifferent from zero, largely dismissing the concern about potential omitted variables.
PART 4: Robustness of Results to Alternative Specifications and Samples
In this section, we examine the robustness of our results according to several different
4
The attrition rates for consecutive waves, that is, 2010–2012 and 2012–2014, are 19.3 percent and 13.9
percent, respectively. We compare the attrition rate of the CFPS with that of the UK Household Longitudinal
Survey (UKHLS). The two surveys were conducted during the same period and followed similar interview
methods, so the UKHLS serves as a good benchmark for the CFPS. Compared to the UKHLS, the CFPSs
attrition rate is reasonable. The key reason for using the 2010 and 2014 waves is that the two waves included
exactly the same test modules, whereas the short memory and logic tests employed in the 2012 wave are not
comparable with the tests used in the other two waves.
5
Counties unmatched to any API report cities within 40 km or weather stations within 100 km are dropped.
The matching rate of 74.7 percent (37,918 out of 50,789) is within a reasonable range compared with other
studies. For example, Levinson (2012) was able to maintain 52.3 percent of the observations when matching
the US General Social Survey with PM10 readings from the Environmental Protection Agency’s Air Quality
System.
7
tests to confirm whether our results were qualitatively affected by the decisions made in
our paper along several dimensions, such as model specification, sample selection,
mechanism tests, and weighted regressions.
A. Alternative Specifications of Table S1
In Table S11, we further display alternative specifications of Table S1 with and
without demographic controls and individual fixed effects. Panel A is for verbal test scores,
while Panel B is for math test scores. In each panel, there are three parts, which correspond
to 7-day, 90-day and 1-year windows, respectively. The last column in each part just keeps
the original column in Table S1 for ease of comparison. The first column in each part
addresses the concern that household income per capita and years of education may be bad
controls, i.e., they are endogenous. We re-estimate our models with income per capita and
years of education excluded. The results in the first column are qualitatively identical to
those in the last columns, suggesting they are not bad controls. The middle column in each
part further explores the difference between longitudinal and cross-sectional estimations.
The comparison between the middle column and the last column indicates that the
statistical significance is basically the same between these two specifications, but the size
of the effect of air pollution without individual fixed effects is only two thirds of that with
individual fixed effects controlled.
B. Giving Interviews in Winter Months Greater Weights
Most of the survey interviews were conducted in the summer months. There is a
concern that the results are driven by the overwhelmingly large sample in the summer. We
divide the sample into two groups. Respondents in Group 1 were interviewed at least once
in winter months (November, December and January), while respondents in Group 2 were
only interviewed in non-winter months (from February to October). Observations in Group
1 are reweighted by the ratio of the number of observations in Group 2 divided by the
number of observations in Group 1. In doing so, we give observations in the winter months
8
greater weights. As revealed in Table S12, the weighted regression indicates that the results
are robust, i.e., the size of the effects is similar to that estimated in the baseline results.
Besides, Figure S11 further shows that the pattern of age-cohort effects still holds for
verbal tests. Hence, the underestimation of the effect of air pollution due to over
representation of the summer months, if any, is small.
C. Excluding the Channels of Impatience and Noncooperation
The negative effect of air pollution on cognitive performance may be driven by
behavior change. People may become more impatient or uncooperative when exposed to
more polluted air, thereby hampering their cognitive tests. The CFPS includes evaluation
on intervieweesdegrees of impatience and cooperation rated by the interviewers. The
ratings for impatience and cooperation are both scaled from 1 (low) to 7 (high). We explore
the effects of exposure to air pollution on impatience and cooperation using a similar
specification in Equation (1). Table S13 displays the results. Panels A and B are for
impatience cooperation, respectively. The results indicate that there is no significant
association between air pollution and intervieweesimpatience and cooperation, ruling out
the behavioral channel.
D. Restricting the Sample to Non-Migrants Only
The CFPS tracked and interviewed individuals who moved. For respondents who
moved between waves 2010 and 2014, our analysis matches API measures according to
their places of residence by the 2014 survey. A measurement issue arises that individuals
may not stay in their counties of residence for the whole period of cumulative measure of
exposure used. To more precisely match air pollution exposure in wave 2014, as a robust
check we exclude 1.3% of respondents who migrated across counties between 2010 and
2014. Similarly, to ensure precise matching of air pollution exposure in wave 2010, we
further use information on the time of moving into the latest addresses by the 2010 survey
to exclude those respondents who moved within each time window of pollution exposure.
9
For example, we exclude respondents who moved into the latest addresses in 2010 for the
1-year time window, 2009-2010 for the 2-year time window, and 2008-2010 for the 3-year
time window. As reported in Table S14, the majority of respondents in CFPS were non-
migrants across all time windows. In addition, Figure S12 also reveals a larger effect on
verbal test for the older male cohorts. Therefore, our key findings are robust to using non-
migrants only.
E. Excluding Polluted Occupations
In Table S15 and Figure S13, we perform robustness checks by excluding polluted
occupations. Polluted occupations include “Geology and mineral industry workers”,
“Workers in metal smelting and refining industry”, “Chemical product manufacturing
personnel”, “Textile workers”, “Production workers (wood processing, artificial board,
wood products, pulp and paper industry)”, and “Production and processing worker
(construction materials)”. Our baseline results are qualitatively unchanged after excluding
polluted occupations.
F. Controlling for Province-by-year Fixed Effects and Clustering Standard Errors at the
Province Level
As revealed in Table S16, our baseline results are robust to controlling for province-
by-year fixed effects and clustering standard errors at the province level.
PART 5: Estimating Movement in the Test Distribution Using the Coefficient
Estimates
Figure S14 plots the percentile of scores for verbal and math tests, respectively. As
revealed in Figure S14, a one-point increase in verbal test scores corresponds to moving
people from the median (i.e. the 50th percentile) to the 55th percentile in the verbal test
distribution, while a one-point increase in math test scores is equivalent to moving people
from the median to the 68th percentile in the math test distribution.
10
The population-weighted annual mean concentration of PM10 over 2014 in China is
112 μ g/m3, much higher than the National Ambient Air Quality Standards (NAAQS)
published by the U.S. Environmental Protection Agency (EPA).
6
Reducing the annual
mean PM10 to levels below the standard, which amounts to 56 units in one-year-mean API
derived from Table S5, will lead to a sizable increase in verbal test scores by 2.41 points
(or moving people from the median to the 63rd percentile in the verbal test distribution)
and math test scores by 0.39 point (or moving people from the median to the 58th percentile
in the math test distribution) calculated from Table S1.
Besides, we also evaluate the effect of exposure to air pollution on the older cohort
using estimated coefficients in Panel A and Panel B of TableS4a separately by educational
attainment. A one standard deviation decrease in 3-year-mean API leads to an increase in
verbal test scores by 9.18 points (or moving people from the median to the 87th percentile
in the verbal test distribution) for less educated men above age 65 relative to their
counterparts below age 25, while the same decline in API is associated with an increase in
verbal test scores by 1.88 points (or moving people from the median to the 69th percentile
in the verbal test distribution) for more educated older men relative to their younger
counterparts. The negative effect on the less (more) educated older men provides an upper
(lower) bound of the detrimental impact of air pollution exposure on older persons.
PART 6: Scientific Backgrounds and Potential Mechanisms
Air pollution may affect cognition through both physiological and psychological
pathways. In this Appendix, we hypothesize that differences in brain composition may help
explain why men appear more sensitive to air pollution. It is beyond the scope of this paper
to formally test this mechanism. We leave it as a future research topic.
6
The annual mean PM10 data at the city level are obtained from the China Environmental Statistical
Yearbook 2015, and the population data (for the weighting purpose) come from China City Statistical
Yearbook 2015. The standard of annual mean PM10 published by the EPA is 50 μg/m3. Source:
https://www3.epa.gov/ttn/naaqs/standards/pm/s_pm_history.html.
11
A few of these physiological pathways have been documented in the literature (13).
First, multiple pollutants (or toxic compounds bonded to the pollutants) may directly affect
brain chemistry. For example, ozone in the air can react with body molecules to create
toxins, causing asthma and respiratory problems (14).
7
Particulate matter (PM), especially
fine particles, can carry toxins through small passageways and directly enter the brain.
Braniš, Řezáčová, and Domasová (15) show that exposure to high PM concentrations
compromises cognitive performance even for people working indoors.
8
Second, people breathing polluted air are more likely to be subject to oxygen
deficiency, which in turn impairs their cognitive abilities (16, 17). Carbon monoxide (CO),
one important element of air pollution, prevents the body from releasing adequate oxygen
to vital organs, in particular to the brain, which consume a large fraction of total oxygen
intake. Third, air pollution could also damage the immune system, hinder neurological
development, and impair neuron behavior, all of which contribute to long-term memory
formation (18). Fourth, long-term exposure to pollution leads to the growth of white-matter
lesions, potentially inhibiting cognition (19). Further, exposure to highly concentrated air
pollution can be linked to markers of neuroinflammation and neuropathology that are
associated with neurodegenerative conditions, such as Alzheimer’s disease (20, 21).
Finally, a recent study on healthy children living in polluted environment with APOE Ɛ4
allele (known to increase risk of developing Alzheimer’s) demonstrates compromised
cognitive responses compared with those carrying APOE gene with Ɛ3 allele (22).
However, this gene environment interaction is only verified for children, while our main
findings are towards elder persons.
In addition to physiological pathways, air pollution could also disrupt cognitive
functioning through some psychological pathways. For example, high concentrations of
CO and nitrogen dioxide (NO2) are significantly associated with headache, eye irritation,
7
Ozone is formed through a chemical reaction between nitrogen oxides, sunlight, and various gaseous
pollutants.
8
PM is generated by power plants, factories, vehicles, dust, pollen and forest fires.
12
and respiratory problems (23).
9
High levels of ozone and sulfur dioxide (SO2) have also
been found to cause psychiatric distress (24).
10
Exposure to high concentrations of CO,
NO2, SO2, ozone, and PM may also increase the risk of depression (25).
Our central nervous system has two important tissues: gray matter and white matter.
Gray matter represents information processing centers, and white matter represents the
networking of or connections between these processing centers. Mathematics abilities,
which require more local processing, mainly depend on gray matter. While language skills,
which require integrating and assimilating information from distributed gray-matter
regions in the brain, mainly rely on white matter.
11
A brain scanning study conducted by Haier et al. (26) reveals that men have
approximately 6.5 times the amount of gray matter activated during general intelligence
tests than women do, but women have nearly 10 times the amount of white matter activated
during general intelligence tests than men do. See the figure below for a front view of grey
and white matter activation during IQ tests. This finding may help explain why men tend
to excel in math tests, while women tend to excel in verbal tests.
Figure: Front view of grey and white matter activation during IQ tests
Source: Haier et al. (2005).
A large body of literature has proven that air pollution can reduce the density of white
matter in the brain (19, 27, 28), which may directly explain why air pollution appears to
9
NO2 and CO are emitted by coal-burning power plants and the burning of fossil fuels.
10
SO2 is mainly emitted by coal-burning power plants.
11
University of California, Irvine. "Intelligence in Men and Women Is a Gray and White Matter." Science
Daily. www.sciencedaily.com/releases/2005/01/050121100142.htm [accessed January 25, 2017].
13
have a larger effect on verbal test than on math test scores. Besides, since men have a much
smaller amount of white matter activated during intelligence tests, their cognitive
performance, especially in the verbal domain, tends to be more affected by exposure to air
pollution.
14
References
1. China Family Panel Studies (CFPS): 2010-2014. Institute of Social Science Survey of
Peking University, China. http://www.isss.pku.edu.cn/cfps/EN/Data/.
2. Xie, Y., and J. Hu. 2014. “An Introduction to the China Family Panel Studies (CFPS).”
Chinese Sociological Review 47 (1): 329.
3. Air Quality Daily Reports: 2007-2014. Chinese Ministry of Environmental Protection.
http://data.epmap.org/eia/air.
4. Chen, Y., G. Z. Jin, N. Kumar, and G. Shi. 2012. “Gaming in Air Pollution Data?
Lessons from China.B.E. Journal of Economic Analysis and Policy 13 (3): 1–43.
5. Ghanem, D. and J. Zhang. 2014. “Effortless Perfection”: Do Chinese Cities
Manipulate Air Pollution Data? Journal of Environmental Economics and
Management 68: 203–225.
6. Global Summary of the Day (GSOD): 2010-2014. National Centers for Environmental
Information of the US National Oceanic and Atmospheric Administration,
ftp://ftp.ncdc.noaa.gov/pub/data/gsod.
7. Graff Zivin, J. S., S. M. Hsiang, and M. J. Neidell. 2015. Temperature and Human
Capital in the Short- and Long-Run. NBER Working Paper 21157. Cambridge, MA,
US: National Bureau of Economic Research.
8. China County Statistics: 2010-2014. China Data Center of University of Michigan.
http://chinadataonline.org/member/county/.
9. Levinson, A. 2012. “Valuing Public Goods Using Happiness Data: The Case of Air
Quality.Journal of Public Economics 96: 869–880.
10. Deschenes, O., M. Greenstone, and J. Guryan. 2009. “Climate Change and Birth
Weight.American Economic Association 99 (2): 211–217.
11. Deschenes, O., and M. Greenstone. 2011. “Climate Change, Mortality, and Adaptation:
Evidence from Annual Fluctuations in Weather in the US.American Economic
Journal: Applied Economics 3 (October): 152–185.
12. Bensnes, S. S. 2016. “You Sneeze, You Lose: The Impact of Pollen Exposure on
Cognitive Performance During High-Stakes High School Exams.Journal of Health
Economics 49: 1–13.
13. Block, M. L., and L. Calderón-Garcidueñas. 2009. “Air Pollution: Mechanisms of
Neuroinflammation and CNS Disease.” Trends in Neurosciences 32 (9): 506516.
14. Sanders, N. J. 2012. “What Doesn’t Kill You Makes You Weaker: Prenatal Pollution
Exposure and Educational Outcomes.Journal of Human Resources 47: 826850.
15. Braniš, M., P. Řezáčová, and M. Domasová. 2005. “The Effect of Outdoor Air and
Indoor Human Activity on Mass Concentrations of PM10, PM2.5, and PM1 in a
Classroom.” Environmental Research 99 (2): 143149.
15
16. Amitai, Y., Z. Zlotogorski, V. Golan-Katzav, A. Wexler, and D. Gross. 1998.
“Neuropsychological Impairment from Acute Low-Level Exposure to Carbon
Monoxide.Archives of Neurology 55 (6): 845–848.
17. Kampa, M., and E. Castanas. 2007. “Human Health Effects of Air Pollution.”
Environmental Pollution 151: 362367.
18. Perera, F. P., Z. Li, R. Whyatt, L. Hoepner, S. Wang, D. Camann, and V. Rauh. 2009.
“Prenatal Airborne Polycyclic Aromatic Hydrocarbon Exposure and Child IQ at Age
5 Years.Pediatrics 124 (2): e195–e202.
19. Calderón-Garcidueñas, L., A. Mora-Tiscareño, E. Ontiveros, G. mez-Garza, G.
Barragán-Mejía, J. Broadway, S. Chapman, G. Valencia-Salazar, V. Jewells, R. R.
Maronpot, C. Henríquez-Roldán, B. rez-Guillé, R. Torres-Jardón, L. Herrit, D.
Brooks, N. Osnaya-Brizuela, M. Monroy, A. González-Maciel, R. Reynoso-Robles, R.
Villarreal-Calderon, A. Solt, and R. Engle. 2008. “Air Pollution, Cognitive Deficits
and Brain Abnormalities: A Pilot Study with Children and Dogs.” Brain and Cognition
68 (2): 117127.
20. Calderón-Garcidueñas, L., Reed, W., Maronpot, RR., Henriquez-Roldán, C., Delgado-
Chavez, R., Calderón-Garcidueñas, A., Swenberg, J.A. 2004. “Brain inflammation and
Alzheimer’s-like pathology inindividuals exposed to severe air pollution.” Toxicologic
Pathology 32: 650658.
21. Levesque, S., Surace, MJ., McDonald, J., and Block, ML. 2011. Air pollution and the
brain: Subchronic diesel exhaust exposure causes neuroinflammation and elevates
early markers of neurodegenerative disease. Journal of Neuroinflammation 8: 105.
22. Calderón-Garcidueñas, L. et al. 2015. Mexico City normal weight children exposed to
high concentrations of ambient PM2.5 show high blood leptin and endothelin-1,
vitamin D deficiency, and food reward hormone dysregulation versus low pollution
controls. Relevance for obesity and Alzheimer disease. Environmental Research 140:
579-592.
23. Nattero, G., and A. Enrico. 1996. “Outdoor Pollution and Headache.Headache 36:
243–245.
24. Rotton, J., and J. Frey. 1984. “Psychological Costs of Air Pollution: Atmospheric
Conditions, Seasonal Trends, and Psychiatric Emergencies.” Population and
Environment 7 (1): 316.
25. Szyszkowicz, M. 2007. “Air Pollution and Emergency Department Visits for
Depression in Edmonton, Canada.” International Journal of Occupational Medicine
and Environmental Health 20 (3): 241245.
26. Haier, Richard J., Rex E. Jung, Ronald A. Yeo, Kevin Head, and Michael T. Alkire.
2005. “The Neuroanatomy of General Intelligence: Sex Matters.NeuroImage 25:
320-332.
27. Calderón-Garcidueñas, L., R. Engle, A. Mora-Tiscareño, M. Styner, G. mez-Garza,
H. Zhu, V. Jewells, R. Torres-Jardón, L. Romero, M. Monroy-Acosta, C. Bryant, L.
González-González, H. Medina-Cortina, A. D’Angiulli. 2011. “Exposure to Severe
Urban Air Pollution Influences Cognitive Outcomes, Brain Volume and Systemic
16
Inflammation in Clinically Healthy Children.” Brain and Cognition 77: 345-355.
28. Wilker, Elissa H., Sarah R. Preis, Alexa S. Beiser, Philip A. Wolf, Rhoda Au, Itai
Kloog, Wenyuan Li, Joel Schwartz, Petros Koutrakis, Charles DeCarli, Sudha
Seshadri, Murray A. Mittleman. 2015. “Long-Term Exposure to Fine Particulate
Matter, Residential Proximity to Major Roads and Measures of Brain Structure.”
Stroke 46:1161-1166.
17
Figure S1: Distribution of interview months in 2010 and 2014
Source: CFPS survey 2010 and 2014.
0
5,000 10,000 15,000
Jan Feb Apr May June Jul Aug Sep Oct Nov Dec
18
Figure S2: Daily air pollution index (API) in China, 2010–2014
Source: Daily air-quality report, Ministry of Environmental Protection of the People’s Republic of China.
Note: The daily mean API is calculated by finding the weighted average of all the API report cities
within the region, where the weights are the yearly population in each city. The US National Ambient
Air Quality Standard for fine particulate matter smaller than 10 micrometers is 0.15 mg/m3, which
corresponds to API = 100 in China. Northeast China includes Heilongjiang, Jilin, and Liaoning. North
China includes Beijing, Hebei, Inner Mongolia, Shanxi, and Tianjin. East China includes Anhui, Fujian,
Jiangsu, Jiangxi, Shandong, Shanghai, and Zhejiang. Northwest China includes Gansu, Ningxia,
Qinghai, Shanxi, and Xinjiang. Southwest China includes Guizhou, Sichuan, Tibet, Yunnan, and
Chongqing. South China includes Guangdong, Guangxi, Hainan, Henan, Hubei, and Hunan.
050 100 150 200 250 300 350
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
Northeast China
050 100 150 200 250 300 350
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
North China
050 100 150 200 250 300 350
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
East China
050 100 150 200 250 300 350
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
Northwest China
050 100 150 200 250
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
Southwest China
050 100 150 200 250
daily mean API
01jan2010 01jan2011 01jan2012 01jan2013 01jan2014 01jan2015
South China
19
Figure S3: The distribution of API reporting cities
Source: The Ministry of Environmental Protection of China.
Note: The legend 2010 represents all 86 API reporting cities in 2010, and the legend 2014 indicates
newly added API reporting cities in 2014, which cover most of the cities in China by 2014. This figure
is plotted using ArcMap 10.3.1. API = air pollution index.
20
Figure S4: The distribution of monitoring stations
Source: The Ministry of Environmental Protection of China.
Note: This figure is plotted using ArcMap 10.3.1.
21
Figure S5: Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
-.4 -.2 0.2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=0 km
-.4 -.2 0.2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=10 km
-.4 -.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=20 km
-.4 -.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=30 km
22
Figure S5 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=40 km
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=50 km
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=60 km
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=70 km
23
Figure S5 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014.
The age band 10-24 is the reference category. The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the male
and female subsamples with 95% and 99% confidence intervals. As APIs are reported at the city level, we calculate the distance between the centroid of
each CFPS county and the boundary of API reporting cities. If a CFPS county is located within an API reporting city, we set the matching distance as zero
and treat the city’s API readings as the CFPS county’s readings. If a CFPS county is not located in any cities with API readings, we match it to the nearest
API reporting city within a specific radius between the CFPS county centroids and the city boundaries. The figure reports estimation results based on nine
cutoff radiuses ranging from zero km to 80 km.
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=80 km
24
Figure S6: Robustness checks — matching distance to the boundaries of API reporting cities (math test scores)
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=0 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=10 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=20 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=30 km
25
Figure S6 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (math test scores)
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=40 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=50 km
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=60 km
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=70 km
26
Figure S6 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (math test scores)
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014.
The age band 10-24 is the reference category. The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the male
and female subsamples with 95% and 99% confidence intervals. As APIs are reported at the city level, we calculate the distance between the centroid of
each CFPS county and the boundary of API reporting cities. If a CFPS county is located within an API reporting city, we set the matching distance as zero
and treat the city’s API readings as the CFPS county’s readings. If a CFPS county is not located in any cities with API readings, we match it to the nearest
API reporting city within a specific radius between the CFPS county centroids and the city boundaries. The figure reports estimation results based on nine
cutoff radiuses ranging from zero km to 80 km.
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=80 km
27
Figure S7: Robustness checks — matching distance to monitoring stations (verbal test scores)
-.4 -.2 0.2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=40 km
-.4 -.2 0.2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=50 km
-.4 -.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=60 km
-.4 -.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=70 km
28
Figure S7 (continued): Robustness checks — matching distance to monitoring stations (verbal test scores)
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014.
The age band 10-24 is the reference category. The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the male
and female subsamples with 95% and 99% confidence intervals. The city-level APIs are computed based on readings from multiple monitoring stations in
each API reporting city, but the readings from the monitoring stations along with their latitude and longitude information have been available since 2014.
We assign the reported city-level APIs in 2010 to all stations within the city boundary under the assumption that that the geographic locations of air quality
monitoring stations do not change much between 2010 and 2014 and variations in daily air pollution across monitoring stations within each city are small.
We match CFPS counties to the nearest air quality monitoring station within a specific radius between the CFPS county centroids and the monitoring
stations. All the results are weighted by the inverse distance to monitoring stations. The table reports estimation results based on six cutoff radiuses
ranging from 40 km to 90 km.
-.4 -.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=80 km
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=90 km
29
Figure S8: Robustness checks — matching distance to monitoring stations (math test scores)
-.2 -.1 0.1 .2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=40 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=50 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=60 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=70 km
30
Figure S8 (continued): Robustness checks — matching distance to monitoring stations (math test scores)
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014.
The age band 10-24 is the reference category. The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the male
and female subsamples with 95% and 99% confidence intervals. The city-level APIs are computed based on readings from multiple monitoring stations in
each API reporting city, but the readings from the monitoring stations along with their latitude and longitude information have been available since 2014.
We assign the reported city-level APIs in 2010 to all stations within the city boundary under the assumption that that the geographic locations of air quality
monitoring stations do not change much between 2010 and 2014 and variations in daily air pollution across monitoring stations within each city are small.
We match CFPS counties to the nearest air quality monitoring station within a specific radius between the CFPS county centroids and the monitoring
stations. All the results are weighted by the inverse distance to monitoring stations. The table reports estimation results based on six cutoff radiuses
ranging from 40 km to 90 km.
-.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=80 km
-.15 -.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
matching distance=90 km
31
Figure S9: PM10 API, SO2 API, and NO2 API during the day, 2014
Source: Hourly air-quality report, Ministry of Environmental Protection (MEP) of the People’s
Republic of China.
Note: The hourly mean pollution concentrations are calculated using the average values from all
the monitoring stations in China. The left axis indicates the pollutant API that converts the
corresponding pollutant measure in micrograms per cubic meter g/m3) into an API score ranging
from 0 to 500 using a formula devised by the MEP. The right axis indicates the distribution of
interview time (percent). As this detailed air pollution component dataset has only been available
since 2014, we cannot use it in our main empirical analysis. API = air pollution index; NO2 =
nitrogen dioxide; PM10 = particulate matter 10 micrometers or less in diameter; SO2 = sulfur
dioxide.
0.0
1.0
2.0
3.0
4.0
5.0
6.0
7.0
8.0
9.0
10.0
0
10
20
30
40
50
60
70
80
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
percent
API
hour
interview time distribution PM10 API SO2 API NO2 API
32
Figure S10: Falsification tests - Effects of air pollution on test scores in the days after the
interview
Panel A: Verbal test scores
Panel B: Math test scores
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: The figure plots the coefficients with 95% and 99% confidence intervals from a regression of test
scores on air pollution index (API) readings in the days after the interview. Other controls and fixed
effects are the same as those presented in Table S1.
-.02 -.01
0.01
1 2 3 4 5 6
all male female
-.005
0
.005
1 2 3 4 5 6
all male female
33
Figure S11: Robustness checks – giving interviews in winter months greater weights
A. Verbal test scores
B. Math test scores
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort
dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age band 10-24 is the reference category.
The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the
male and female subsamples with 95% and 99% confidence intervals. Air pollution data are matched
between each CFPS county centroid and its nearest API reporting city boundary within a radius of 40km
(i.e. 25miles). We divide the sample into two groups. Respondents in Group 1 were interviewed at least
once in winter months (November, December and January), while respondents in Group 2 were only
interviewed in non-winter months (from February to October). Observations in Group 1 are reweighted
by the ratio of the number of observations in Group 2 divided by the number of observations in Group
1.
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
34
Figure S12: Robustness checks – using non-migrants
A. Verbal test scores
B. Math test scores
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort
dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age band 10-24 is the reference category.
The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the
male and female subsamples with 95% and 99% confidence intervals. Air pollution data are matched
between each CFPS county centroid and its nearest API reporting city boundary within a radius of 40km
(i.e. 25miles). In all the regressions, we exclude the respondents who moved across counties between
2010 and 2014. In addition, we exclude respondents who moved into the latest addresses in 2008-2010
for the 3-year time window.
-.4 -.2 0.2
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
35
Figure S13: Robustness checks – polluted occupations excluded
A. Verbal test scores
B. Math test scores
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: The age-cohort effects include interaction terms between 3-year-mean API and age cohort
dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age band 10-24 is the reference category.
The figures plot the estimated coefficients on the interaction terms for the whole sample as well as the
male and female subsamples with 95% and 99% confidence intervals. Air pollution data are matched
between each CFPS county centroid and its nearest API reporting city boundary within a radius of 40km
(i.e. 25miles). The regressions exclude respondents with polluted jobs. Polluted occupations include
“Geology and mineral industry workers”, “Workers in metal smelting and refining industry”, “Chemical
product manufacturing personnel”, “Textile workers”, “Production workers (wood processing, artificial
board, wood products, pulp and paper industry)”, and Production and processing worker (construction
materials)”.
-.3 -.2 -.1 0.1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
-.1 -.05
0.05 .1
25_34 35_44 45_54 55_64 65+
Age in 2014
all male female
36
Figure S14: Percentiles of cognitive test scores
Source: CFPS survey 2010 and 2014.
Note: The figure plots the percentiles of scores for verbal and math tests,
respectively.
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34
percentile
verbal test scores
A. Verbal test scores
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18 20 22 24
percentile
math test scores
B. Math test scores
37
Table S1: Effects of air pollution on cognitive test scores
1-day
7-day
30-day
90-day
1-year
2-year
3-year
(1)
(2)
(3)
(4)
(5)
(6)
(7)
A. Verbal test scores
t
API
-0.004*
-0.001
0.000
-0.002
-0.002
-0.003
-0.003
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.013**
-0.035***
-0.044***
-0.043***
-0.060***
-0.086***
(0.005)
(0.008)
(0.011)
(0.012)
(0.016)
(0.021)
Observations
31,955
31,955
31,955
31,955
31,955
31,955
31,955
Overall R-squared
0.285
0.279
0.288
0.291
0.281
0.279
0.278
Impact of a one SD reduction
in mean API on test scores
(SDs of test scores)
0.131
(0.012)
0.278
(0.026)
0.599
(0.057)
0.712
(0.068)
0.895
(0.085)
0.942
(0.090)
1.132
(0.108)
B. Math test scores
t
API
-0.001
-0.001
-0.001
-0.001
-0.001
-0.001
-0.001
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
1
0
1k
ti
iAPI
k
=
-0.003
-0.004*
-0.009**
-0.007**
-0.010**
-0.016**
(0.002)
(0.002)
(0.003)
(0.003)
(0.005)
(0.007)
Observations
31,955
31,955
31,955
31,955
31,955
31,955
31,955
Overall R-squared
0.449
0.440
0.455
0.447
0.451
0.450
0.447
Impact of a one SD reduction
in mean API on test scores
(SDs of test scores)
0.033
(0.005)
0.064
(0.010)
0.068
(0.011)
0.146
(0.023)
0.146
(0.023)
0.157
(0.025)
0.211
(0.033)
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note:
1
0
1k
ti
iAPI
k
=
indicates the mean of API readings in the past k days, where k equals 1, 7, 30, 90, 365, 730, and 1,095, respectively. All the regressions
include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem hour fixed effects; and a quadratic monthly time trend.
Demographic controls include gender, age and its square and cubic terms, household per capita income, and years of education. Weather controls include
20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a dummy for bad
weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share. Robust standard
errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and its nearest API
reporting city boundary within a radius of 40km (i.e. 25miles). API = air pollution index; SD = standard deviation. *10% significance level; **5%
significance level; ***1% significance level.
38
Table S2a: Effects of air pollution on verbal test scores, by gender
Dependent variable
1-day
7-day
30-day
90-day
1-year
2-year
3-year
verbal test scores
(1)
(2)
(3)
(4)
(5)
(6)
(7)
A. Male subsample
t
API
-0.007**
-0.003
-0.001
-0.003
-0.004
-0.005
-0.005
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
1
0
1k
ti
iAPI
k
=
-0.014**
-0.047***
-0.053***
-0.052***
-0.072***
-0.103***
(0.007)
(0.010)
(0.014)
(0.014)
(0.020)
(0.026)
Observations
15,315
15,315
15,315
15,315
15,315
15,315
15,315
Overall R-squared
0.195
0.193
0.193
0.194
0.187
0.187
0.186
B. Female subsample
t
API
-0.002
0.001
0.000
-0.000
-0.001
-0.001
-0.001
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.011**
-0.024***
-0.035***
-0.035***
-0.049***
-0.069***
(0.005)
(0.007)
(0.010)
(0.011)
(0.014)
(0.018)
Observations
16,640
16,640
16,640
16,640
16,640
16,640
16,640
Overall R-squared
0.392
0.390
0.392
0.395
0.383
0.381
0.381
C. Gender differences
Gender differences
-0.005
-0.003
-0.023***
-0.018*
-0.017**
-0.023*
-0.034**
Wald test chi-square
2.38
0.22
11.25
3.64
4.67
3.50
4.37
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note:
1
0
1k
ti
iAPI
k
=
indicates the mean of API readings in the past k days, where k equals 1, 7, 30, 90, 365, 730, and 1,095, respectively. All the regressions
include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem hour fixed effects; and a quadratic monthly time trend.
Demographic controls include gender, age and its square and cubic terms, household per capita income, and years of education. Weather controls include
20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a dummy for bad
weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share. Robust standard
errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and its nearest API
reporting city boundary within a radius of 40km (i.e. 25miles). The significance of the male-female differences is derived from Wald tests. API = air
pollution index; SD = standard deviation. *10% significance level; **5% significance level; ***1% significance level.
39
Table S2b: Effects of air pollution on math test scores, by gender
Dependent variable
1-day
7-day
30-day
90-day
1-year
2-year
3-year
math test scores
(1)
(2)
(3)
(4)
(5)
(6)
(7)
A. Male subsample
t
API
-0.002
-0.002
-0.002
-0.002
-0.002
-0.002
-0.002
(0.001)
(0.002)
(0.002)
(0.001)
(0.001)
(0.001)
(0.001)
1
0
1k
ti
iAPI
k
=
-0.001
-0.004
-0.007*
-0.005
-0.006
-0.012
(0.003)
(0.003)
(0.004)
(0.004)
(0.006)
(0.008)
Observations
15,315
15,315
15,315
15,315
15,315
15,315
15,315
Overall R-squared
0.444
0.445
0.429
0.442
0.422
0.442
0.428
B. Female subsample
t
API
-0.001
0.000
-0.001
-0.000
-0.001
-0.001
-0.001
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
1
0
1k
ti
iAPI
k
=
-0.004
-0.004
-0.010**
-0.008**
-0.012*
-0.019**
(0.003)
(0.003)
(0.005)
(0.004)
(0.006)
(0.009)
Observations
16,640
16,640
16,640
16,640
16,640
16,640
16,640
Overall R-squared
0.482
0.487
0.489
0.485
0.482
0.480
0.480
C. Gender differences
Gender differences
-0.001
0.003
0.000
0.003
0.003
0.006
0.007
Wald test chi-square
1.15
0.75
0.01
0.17
0.52
0.65
0.54
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note:
1
0
1k
ti
iAPI
k
=
indicates the mean of API readings in the past k days, where k equals 1, 7, 30, 90, 365, 730, and 1,095, respectively. All the regressions
include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem hour fixed effects; and a quadratic monthly time trend.
Demographic controls include gender, age and its square and cubic terms, household per capita income, and years of education. Weather controls include
20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a dummy for bad
weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share. Robust standard
errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and its nearest API
reporting city boundary within a radius of 40km (i.e. 25miles). The significance of the male-female differences is derived from Wald tests. API = air
pollution index; SD = standard deviation. *10% significance level; **5% significance level; ***1% significance level.
40
Table S3: Age-cohort effects of air pollution on cognitive test scores, by gender
Dependent variable
Verbal test scores
Math test scores
All
Male
Female
Gender
differences
(chi-square)
All
Male
Female
Gender
differences
(chi-square)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
3-year-mean API
age25-34
-0.017
-0.028
-0.003
-0.025
-0.024
-0.004
-0.044**
0.040
(0.027)
(0.037)
(0.034)
(0.29)
(0.022)
(0.033)
(0.019)
(1.69)
3-year-mean API
age35-44
-0.015
-0.020
-0.007
-0.013
-0.004
0.012
-0.020
0.032
(0.038)
(0.040)
(0.048)
(0.08)
(0.024)
(0.035)
(0.020)
(1.22)
3-year-mean API
age45-54
-0.097***
-0.108***
-0.084
-0.024
-0.008
0.016
-0.032
0.048*
(0.036)
(0.033)
(0.055)
(0.18)
(0.024)
(0.030)
(0.024)
(3.48)
3-year-mean API
age55-64
-0.077
-0.120***
-0.030
-0.090*
0.017
0.035
0.001
0.034
(0.049)
(0.044)
(0.064)
(3.09)
(0.021)
(0.027)
(0.021)
(1.80)
3-year-mean API
age65+
-0.114**
-0.192***
-0.026
-0.166***
-0.007
-0.004
-0.010
0.006
(0.046)
(0.050)
(0.057)
(8.06)
(0.021)
(0.028)
(0.022)
(0.04)
Observations
31,955
15,315
16,640
31,955
15,315
16,640
Overall R-squared
0.644
0.644
0.644
0.644
0.644
0.644
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: Other covariates include API on the interview date, 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age
band 10-24 is the reference category. All the regressions include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem
hour fixed effects; and a quadratic monthly time trend. Demographic controls include gender, household per capita income, and years of education. Weather
controls include 20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a
dummy for bad weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share.
Robust standard errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and
its nearest API reporting city boundary within a radius of 40km (i.e. 25miles). The significance of the male-female differences is derived from Wald tests.
API = air pollution index; SD = standard deviation. *10% significance level; **5% significance level; ***1% significance level.
41
Table S4a: Age-cohort effects of air pollution on verbal test scores, by gender and education level
Dependent variable
A. Primary school or below
B. Middle school or above
verbal test scores
All
Male
Female
Gender
differences
(chi-square)
All
Male
Female
Gender
differences
(chi-square)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
3-year-mean API
age25-34
-0.319**
-0.545***
-0.008
-0.537**
0.006
0.003
0.015
-0.012
(0.154)
(0.195)
(0.176)
(6.06)
(0.027)
(0.039)
(0.037)
(0.05)
3-year-mean API
age35-44
-0.248
-0.437**
0.037
-0.474**
0.006
-0.004
0.030
-0.034
(0.151)
(0.199)
(0.144)
(5.37)
(0.033)
(0.035)
(0.049)
(0.44)
3-year-mean API
age45-54
-0.336**
-0.649***
0.018
-0.667***
-0.070**
-0.058*
-0.071
0.013
(0.136)
(0.146)
(0.146)
(14.61)
(0.033)
(0.035)
(0.052)
(0.05)
3-year-mean API
age55-64
-0.272*
-0.650***
0.114
-0.764***
-0.083**
-0.063
-0.098
0.035
(0.161)
(0.178)
(0.158)
(19.61)
(0.039)
(0.042)
(0.064)
(0.27)
3-year-mean API
age65+
-0.320**
-0.694***
0.089
-0.783***
-0.126***
-0.142***
-0.038
-0.104
(0.145)
(0.163)
(0.154)
(19.29)
(0.043)
(0.050)
(0.082)
(1.36)
Observations
12,515
4,927
7,588
19,436
10,385
9,051
Overall R-squared
0.166
0.166
0.166
0.166
0.166
0.166
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: Other covariates include API on the interview date, 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age
band 10-24 is the reference category. All the regressions include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem
hour fixed effects; and a quadratic monthly time trend. Demographic controls include gender, household per capita income, and years of education. Weather
controls include 20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a
dummy for bad weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share.
Robust standard errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and
its nearest API reporting city boundary within a radius of 40km (i.e. 25miles). The significance of the male-female difference is derived from Wald tests.
API = air pollution index; SD = standard deviation. *10% significance level; **5% significance level; ***1% significance level.
42
Table S4b: Age-cohort effects of air pollution on math test scores, by gender and education level
Dependent variable
A. Primary school or below
B. Middle school or above
math test scores
All
Male
Female
Gender
differences
(chi-square)
All
Male
Female
Gender
differences
(chi-square)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
3-year-mean API
age25-34
-0.110
-0.111
-0.093
-0.018
-0.011
0.005
-0.027
0.032
(0.068)
(0.120)
(0.093)
(0.01)
(0.022)
(0.032)
(0.021)
(1.04)
3-year-mean API
age35-44
-0.107
-0.104
-0.092
-0.012
0.019
0.024
0.018
0.006
(0.074)
(0.128)
(0.071)
(0.01)
(0.025)
(0.032)
(0.024)
(0.05)
3-year-mean API
age45-54
-0.100
-0.120
-0.073
-0.047
0.015
0.037
-0.007
0.044
(0.073)
(0.113)
(0.074)
(0.12)
(0.023)
(0.029)
(0.025)
(2.22)
3-year-mean API
age55-64
-0.066
-0.096
-0.033
-0.063
0.046**
0.065**
0.027
0.038
(0.072)
(0.113)
(0.078)
(0.19)
(0.020)
(0.025)
(0.027)
(1.43)
3-year-mean API
age65+
-0.082
-0.114
-0.041
-0.073
0.005
0.018
-0.002
0.020
(0.069)
(0.110)
(0.079)
(0.26)
(0.024)
(0.026)
(0.043)
(0.22)
Observations
12,515
4,927
7,588
19,436
10,385
9,051
Overall R-squared
0.268
0.268
0.268
0.268
0.268
0.268
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: Other covariates include API on the interview date, 3-year-mean API and age cohort dummies 25-34, 35-44, 45-54, 55-64, and 65+ in 2014. The age
band 10-24 is the reference category. All the regressions include individual fixed effects; county fixed effects; year, month, day of week, and post meridiem
hour fixed effects; and a quadratic monthly time trend. Demographic controls include gender, household per capita income, and years of education. Weather
controls include 20°F indicators for temperature bins (that is, <25°F, 25–45°F, 45–65°F, 65–85°F, and >85°F), total precipitation, mean wind speed, and a
dummy for bad weather. County-level characteristics include gross domestic product (GDP) per capita, population density, and industrial value share.
Robust standard errors, clustered at the county level, are presented in parentheses. Air pollution data are matched between each CFPS county centroid and
its nearest API reporting city boundary within a radius of 40km (i.e. 25miles). The significance of the male-female difference is derived from Wald tests.
API = air pollution index; SD = standard deviation. *10% significance level; **5% significance level; ***1% significance level.
43
Table S5: Breakpoints for API value calculation
API index value
PM10 (μg/m3)
SO2 (μg/m3)
NO2 (μg/m3)
0
0
0
0
50
50
50
40
100
150
150
80
150
250
475
180
200
350
800
280
300
420
1600
565
400
500
2100
750
500
600
2620
940
Note: API = air pollution index; NO2 = nitrogen dioxide; PM10 = particulate matter 10
micrometers or less in diameter; SO2 = sulfur dioxide.
44
Table S6: Summary statistics
Variable
All
Male
Female
Mean
SD
Mean
SD
Mean
SD
verbal test scores
18.115
10.489
19.729
9.430
16.629
11.172
math test scores
10.438
6.403
11.496
5.924
9.464
6.667
API
73.516
32.684
73.197
31.714
73.810
33.549
7-day mean API
72.885
21.360
72.619
21.108
73.130
21.587
30-day mean API
72.992
17.118
72.801
17.078
73.168
17.153
90-day mean API
75.516
16.184
75.342
16.133
75.676
16.231
1-year mean API
84.002
20.806
83.822
20.863
84.167
20.753
2-year mean API
77.738
15.706
77.572
15.782
77.891
15.634
3-year mean API
74.882
13.166
74.705
13.227
75.044
13.108
log form of household per capita
income (Chinese yuan)
8.874
1.154
8.891
1.153
8.858
1.155
age
44.742
17.892
44.925
18.158
44.573
17.642
years of education
7.475
4.451
8.220
4.058
6.789
4.681
Source: Authorsestimations using CFPS survey 2010 and 2014.
Note: API = air pollution index; SD = standard deviation.
45
Table S7: Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
Dependent variable
1-day
7-day
30-day
90-day
1-year
2-year
3-year
verbal test scores
(1)
(2)
(3)
(4)
(5)
(6)
(7)
A. matching distance = 0 km
t
API
-0.009***
-0.007**
-0.006**
-0.006**
-0.007**
-0.007***
-0.008***
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
(0.003)
1
0
1k
ti
iAPI
k
=
-0.004
-0.021**
-0.040***
-0.031***
-0.042**
-0.060**
(0.006)
(0.009)
(0.013)
(0.011)
(0.016)
(0.024)
Observations
17,160
17,160
17,160
17,160
17,160
17,160
17,160
Overall R-squared
0.297
0.295
0.294
0.295
0.294
0.295
0.290
B. matching distance = 10 km
t
API
-0.008***
-0.005**
-0.005**
-0.006**
-0.007***
-0.007***
-0.007***
(0.002)
(0.003)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.010*
-0.030***
-0.045***
-0.027**
-0.036**
-0.060**
(0.005)
(0.008)
(0.013)
(0.010)
(0.016)
(0.024)
Observations
20,049
20,049
20,049
20,049
20,049
20,049
20,049
Overall R-squared
0.278
0.278
0.273
0.276
0.270
0.274
0.269
C. matching distance = 20 km
t
API
-0.007***
-0.004*
-0.003*
-0.005**
-0.005**
-0.006***
-0.006***
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.009*
-0.033***
-0.044***
-0.032**
-0.040**
-0.068***
(0.005)
(0.009)
(0.015)
(0.014)
(0.017)
(0.023)
Observations
23,423
23,423
23,423
23,423
23,423
23,423
23,423
Overall R-squared
0.278
0.277
0.275
0.275
0.274
0.277
0.276
46
Table S7 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
Dependent variable
1-day
7-day
30-day
90-day
1-year
2-year
3-year
verbal test scores
(1)
(2)
(3)
(4)
(5)
(6)
(7)
D. matching distance = 30 km
t
API
-0.006***
-0.003
-0.002
-0.004
-0.005**
-0.005**
-0.005**
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.011**
-0.035***
-0.043***
-0.033***
-0.041**
-0.062***
(0.005)
(0.008)
(0.012)
(0.012)
(0.016)
(0.022)
Observations
28,000
28,000
28,000
28,000
28,000
28,000
28,000
Overall R-squared
0.262
0.262
0.263
0.255
0.254
0.259
0.253
E. matching distance = 40 km
t
API
-0.004*
-0.001
0.000
-0.002
-0.002
-0.003
-0.003
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.013**
-0.035***
-0.044***
-0.043***
-0.060***
-0.086***
(0.005)
(0.008)
(0.011)
(0.012)
(0.016)
(0.021)
Observations
31,955
31,955
31,955
31,955
31,955
31,955
31,955
Overall R-squared
0.285
0.279
0.288
0.291
0.281
0.279
0.278
F. matching distance = 50 km
t
API
-0.004*
-0.000
0.001
-0.001
-0.002
-0.002
-0.002
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.015***
-0.038***
-0.047***
-0.039***
-0.055***
-0.082***
(0.005)
(0.008)
(0.010)
(0.010)
(0.014)
(0.019)
Observations
33,953
33,953
33,953
33,953
33,953
33,953
33,953
Overall R-squared
0.271
0.277
0.270
0.270
0.268
0.272
0.274
47
Table S7 (continued): Robustness checks — matching distance to the boundaries of API reporting cities (verbal test scores)
Dependent variable
1-day
7-day
30-day
90-day
1-year
2-year
3-year
verbal test scores
(1)
(2)
(3)
(4)
(5)
(6)
(7)
G. matching distance = 60 km
t
API
-0.004*
0.000
0.001
-0.001
-0.002
-0.002
-0.002
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.015***
-0.037***
-0.045***
-0.036***
-0.050***
-0.072***
(0.005)
(0.008)
(0.010)
(0.009)
(0.013)
(0.018)
Observations
35,265
35,265
35,265
35,265
35,265
35,265
35,265
Overall R-squared
0.278
0.269
0.272
0.270
0.269
0.273
0.269
H. matching distance = 70 km
t
API
-0.004*
0.000
0.001
-0.001
-0.002
-0.002
-0.002
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
(0.002)
1
0
1k
ti
iAPI
k
=
-0.015***
-0.035***
-0.042***
-0.034***
-0.045***
-0.066***
(0.005)
(0.008)
(0.010)
(0.009)
(0.012)
(0.017)
Observations
36,388
36,388
36,388
36,388
36,388
36,388
36,388
Overall R-squared
0.266
0.274
0.279
0.270
0.264
0.265
0.269
I. matching distance = 80 km
t
API