Information, Awareness, and Menta Lhealth - Evidence From Air Pollution Disclosure in China - 2023

Journal of Environmental Economics and Management 120 (2023) 102827
Contents lists available at ScienceDirect
Journal of Environmental Economics and

Management
journal homepage: www.elsevier.com/locate/jeem
Information, awareness, and mental health: Evidence from air

pollution disclosure in China✩
Tingting Xie a , Ye Yuan b ,∗, Hui Zhang c
a School of Economics, Central University of Finance and Economics, China
b School of Economics, Institute for Global Health and Development, Peking University, China
c
School of Economics , Peking University, China
ARTICLE INFO ABSTRACT
JEL classification: This paper assesses mental health responses to information on environmental risks. We ex-
D80 ploit the progressive implementation of a national program in China that introduces more
I10 comprehensive air pollution monitoring and provides real-time air-pollution information to
Q53
the public. The program leads to a sharp increase in public awareness and attention to air
Q58
pollution issues and results in a large increase in the sensitivity of individual’s mental health to
Keywords: changes in air quality, especially among those with more exposure to pollution information
Environmental information and those more susceptible to mental illnesses. Information of worsening air quality has
Mental health
a direct effect on mental health as a source of stressors and an indirect behavioral effect
Environmental regulations
through reducing outdoor activities and social integration. Our findings shed light on the design
Air pollution
and delivery of environmental information disclosure programs, especially for countries with
pressing environmental threats.
1. Introduction
With rising attention to environmental safety, immense amounts of environmental information are now accessible to the public.
Information disclosure policies are increasingly being used as an effective tool to promote public supervision and engagement in
environmental protection in diverse areas, including water safety, greenhouse gas emissions, air pollution, radiation, and climate
change (see, for example, Bennear and Olmstead (2008), Neidell (2009), Deschenes et al. (2017); and Greenstone and Jack (2015)).1
An important research agenda is to understand the role of environmental information in shaping individual’s awareness, cognition,
decision-making, mental well-being, and overall welfare in response to environmental risks. Recent literature has established that
the disclosure of environmental information mitigates exposure to environmental hazards, increases willingness to pay (WTP) for
means of avoidance, and reduces the mortality and morbidity burden of environmental risks (Chay and Greenstone, 2005; Bennear
and Olmstead, 2008; Barwick et al., 2019; Freeman et al., 2019; Ito and Zhang, 2020; Tu et al., 2020; Deschenes et al., 2017). Yet
✩ We thank seminar participants at Peking University, Fudan University, Jinan University, and Renmin University of China for helpful discussions, suggestions,
and comments. Any remaining errors are our own. Yuan gratefully acknowledges the National Natural Science Foundation of China (No. 72203004) and the
PKU School of Economics Research Seed Fund for research funding. The authors declare no financial support from or held position at any organization that has
interests or is relevant to this article.
∗ Corresponding author.
E-mail addresses: xietingting19@pku.edu.cn (T. Xie), yuanye.econ@pku.edu.cn (Y. Yuan), nk94zhang@pku.edu.cn (H. Zhang).
1 The United States was among the earliest to disclose information on pollution through the 1986 Toxic Release Inventory (TRI) Program. Other recent
environmental information disclosure programs include the Performance Evaluation and Ratings Program (PERP) in Indonesia, Green Rating Project (GRP) in
India, and EcoWatch Program in the Philippines.
https://doi.org/10.1016/j.jeem.2023.102827
Received 5 March 2022
Available online 30 May 2023
0095-0696/© 2023 Elsevier Inc. All rights reserved.
T. Xie et al. Journal of Environmental Economics and Management 120 (2023) 102827
much less is known about the role of environmental information in shaping individual’s mental well-being, especially in countries
with deteriorating environmental quality.
We aim to fill this gap in knowledge by investigate whether the public disclosure of real-time information on air pollution changes
the way individuals respond mentally to air pollution. Specifically, we estimate the causal impact of air pollution on individual’s
mental health in periods before and after the information on local pollution levels is readily disclosed, and investigate whether the
information disclosure changes – reduces or magnifies – the pollution’s impact on mental health. We then study the underlying
mechanisms for this change.
China’s recent real-time monitoring and disclosure program for air pollution provides the empirical setting for this study. Air
pollution is a major environmental challenge for China, which is one of the most polluted countries in the world. In 2010, China’s
annual concentration of a primary air pollutant – fine particulate matters with a diameter of 2.5 μm or less (PM2.5) – was over 60
μg∕m3 , five times higher than the World Health Organization’s safety standard (World Health Organization, 2005). Air pollution is
the fifth leading risk factor for all-cause mortality in China and caused 1.1 million deaths in 2015 (Forouzanfar et al., 2016). Yet, a
national system for monitoring and reporting air pollution was nonexistent before the 2010s. Amid a rising outcry over the lack of air-
pollution information (Ghanem and Zhang, 2014; Greenstone et al., 2022), the government launched a nationwide program in 2013
to undertake real-time monitoring and automated reporting of air pollution (henceforth, the information program). The program was
implemented progressively across cities and covered 98% of China’s population by the end of 2016. The comprehensive population
coverage and the progressive roll-out provide an ideal empirical setting to study the impact of such a sharp and permanent increase
in pollution information on various behavioral and mental responses to air pollution.2
Our empirical strategy exploits two sources of exogenous variations. First, we take advantage of the progressive roll-out schedule
of the information program across different cities in a general difference-in-differences design. Second, we adopt thermal inversions
as the instrumental variable for air pollution to elicit the causal impact of air pollution on mental health both before and after the
information program. Essentially, we take advantage of the sharp and staggered roll-out of the information program to investigate
whether there exists a ‘‘break’’ in the marginal effect of air pollution on mental health after the information program. In the language
of Barwick et al. (2019), we refer to the marginal effect of air pollution on mental health as the gradient of mental health to air
pollution, and study whether and how such a gradient changes when information on air pollution becomes readily available.
We provide robust evidence that the disclosure of real-time information on air pollution has significantly increased the gradient of
mental health to air pollution. The introduction of the information program has more than doubled this gradient. This pattern is more
pronounced for individuals with better access to pollution information, for those with more susceptibility to mental illnesses, and
for those living in cities with little prior information on local air quality. While previous literature has documented that information
can reduce the impact of environmental threats on individual’s physical health through facilitating avoidance, our finding highlights
that such information can also magnify the impact of environmental threats on individual’s mental health.
We find a potential awareness-alertness-stress pathway through which the disclosure of pollution information may lead to a
greater gradient of mental health to air pollution. We explain in three steps. First, the provision of real-time pollution information
has increased public awareness of the severity of local pollution levels as well as the health damages of air pollution. As a direct
sign of awareness, the intensity of active online searches for air pollution, health impact of air pollution, and defensive measures
against air pollution, all significantly increased in response to an increase in pollution levels. Second, this increased awareness of
the severity and harms of local air pollution would naturally make individuals more alert to the fluctuations in air pollution (Chun
et al., 2011; Gottlieb, 2012). Third, with higher alertness and attention, information on a worsening air quality is more likely to
register in the mental system and trigger a series of stress responses, such as anxiety and sleep disturbance (Miller et al., 2016;
Thomson, 2019; Braithwaite et al., 2019; Li et al., 2017). The physiology and neuroscience literature shows that an accumulation
of such stress responses can cause constant wear and tear on the body and the mental system, and result in a higher risk of both
short- and long-term physical and mental disorders (Juster et al., 2010; McEwen, 2000, 2012). Indeed, individuals who lived in
regions that exhibited a larger increase in pollution-related search activities experienced a greater increase in stress and depressive
responses to air pollution. In addition, the increase in the gradient of mental health to air pollution became more pronounced when
an episode of bad air quality persisted for a longer time, suggesting an accumulation of stress responses over time.
The second driving factor of the greater gradient of mental health to air pollution is avoidance behaviors. Consistent with prior
studies, we find that individuals are more likely to reduce the frequency of physical exercises, social activities with relatives, friends,
and neighbors, and other outdoor activities after being more informed of the severity of air pollution. These avoidance behaviors,
however, may generate mental health burdens. Physical exercise and social activities have been shown to have substantial protective
effects on mental health (Kleppang et al., 2018; Chan et al., 2019; Doré et al., 2016; Seeman, 1996; Kaplan et al., 2012; Lecerof
et al., 2015). The greater gradient of avoidance behaviors to air pollution also leads to more disrupted daily routines and reduced
social integration, which may contribute to a greater gradient of mental health to air pollution. In addition, the information program
has led to a substantial increase in the gradient of defensive expenditure to air pollution, such as in purchasing air purifiers (Barwick
et al., 2019; Ito and Zhang, 2020).3 While such an increase in defensive expenditure reflects a greater WTP for air quality, it may
also lead to an increase in the financial burden and stress.
2 In a recent study closest to ours, Barwick et al. (2019) exploit the same empirical setting – i.e., the progressive rollout of the information program – to
estimate the effect of information disclosure on individual’s behavioral responses to air pollution. They document a substantial social benefit of the information
disclosure in terms of mitigating the detrimental health effects of air pollution through facilitating avoidance and raising WTP for defensive measures against
air pollution. Our work complements to this strand of research and provides a novel perspective on mental health.
3 As documented by Ito and Zhang (2020), average price of regular air purifiers was $455 in 2010s, or 25% of per-capita annual income in our sample. Air
purifiers with capacity to remove PM2.5 were $139 more expensive, and cost an additional $21 when replacing a filter.
2
Our findings are in line with the general hypothesis that information has a direct impact on individual’s mental health4 ; and
that environmental information played a role in shaping individual’s response to environmental threats. However, our results do
not imply that the overall welfare effects of the information program are negative. Such calculation would require estimating the
effects along various other dimensions. In this respect, our estimates complement recent findings of Barwick et al. (2019) and Ito
and Zhang (2020), who quantify the positive effects of the information program on individual’s physical health and WTP through
pollution avoidance behaviors and defensive expenditure during the same period. In addition, considering that air quality improved
substantially after the information program, our estimates implies that the information program led to additional welfare gains
through an improvement in mental health. A back-of-envelope calculation suggests that the information program led to additional
saving on direct medical spending for depression of $3.36 billion per year when national concentration of PM2.5 dropped from the
2013 level to the 2016 level.5 Should the level of PM2.5 drop to the WHO guideline limit (10 μg∕m3 ), the estimated additional
saving on depression treatment alone would amount to $12 billion per year.
Our findings are informative to policymakers, especially in countries with pressing environmental threats. Our findings show that
the provision of environmental information may be a double-edge sword: It raises the gradient of mental health to air pollution, and
the change in the level of individual’s mental health depends on the change in air pollution levels. In the case of deteriorating
environmental quality, public disclosure of such information may create unintended mental health burdens on the population;
on the other hand, when the environmental quality is improving, the public disclosure may lead to an improvement in mental
health. Therefore, the design and implementation of an information disclosure program may take into account the general trend of
environmental quality and should be implemented along with programs of pollution control and environmental protection.
Our study contributes to the literature on the impacts of air pollution on the risk of mental illness and subjective well-being.
Recent studies in psychology and neuroscience have shown that exposure to air pollution, especially the particulate matters (PMs),
may severely harm the central nervous system (Ross et al., 2018; Mehta et al., 2015; Thomson, 2019) and increase the risk of
anxiety and depression (Pun et al., 2017; Power et al., 2015). The economic literature shows that the mental health play a critical
role in explaining the large adverse effect of air pollution on individuals’ subjective well-being (Zhang et al., 2017; Levinson, 2012).
However, little research has considered the role of pollution information in shaping the relationship between air pollution and
mental health. Our study aims to fill this gap in knowledge.
Our study reconciles an important discrepancy in estimates of the gradient of mental health to air pollution. Such a discrepancy
exists between countries as well as across time periods within the same country (Greenstone and Jack, 2015; Greenstone et al.,
2021). For example, Chen et al. (2018b) estimate a much larger gradient of mental illness to air pollution than Xue et al. (2019),
whereas both studies use the same nationally representative data from China, but cover different sample periods. Our study shows
that the disclosure of accurate, easily accessible information on air pollution could explain an important part of this discrepancy. In
particular, the disclosure of air-pollution information since 2013 can reconcile the larger gradient of mental health to air pollution
estimated by Chen et al. (2018b) (based on post-information-program data) and the smaller gradient by Xue et al. (2019) (based
on pre-information-program data).
Our study also complements the literature that evaluates the impact of information provision on compensatory responses to
pollution exposure. Previous studies have substantiated the effects of environmental information disclosure in shaping decision-
making, including having higher WTP for defensive devices and housing with better air quality (Chay and Greenstone, 2005;
Freeman et al., 2019; Ito and Zhang, 2020; Deschenes et al., 2017; Tu et al., 2020); adjusting consumption and travel patterns
to reduce exposure (Sun et al., 2019; Neidell, 2009; Barwick et al., 2019); reducing labor supply (Graff Zivin and Neidell, 2012;
Hanna and Oliva, 2015; Borgschulte et al., 2018); and moving to locations with better air quality (Qin and Zhu, 2018). Our paper
adds to this burgeoning strand of research by delineating the important role of pollution information in shaping the mental health
response to environmental risks, and find that environmental information amplifies the impact of environmental amenity or adversity
on individual’s mental well-being. Moreover, many forms of avoidance responses to pollution exposure, such as reducing outdoor
leisure activities, changing consumption patterns, and reducing social activities, may generate unintended mental-health effects.
Finally, our study provides a new perspective to the emerging literature that quantifies the social welfare effects of pollution
information disclosure in developing countries, and China in particular. Several studies have evaluated the welfare effects of the 2013
information program in China and documented substantial social benefits in terms of improving regulation of local polluters (Deng
and Axbard, 2021), reducing strategic underreporting of pollution levels (Greenstone et al., 2022), raising public awareness on
pollution, mitigating harms of pollution exposure, and raising WTP for defensive measures (Barwick et al., 2019; Ito and Zhang,
2020). We document an under-studied effect of the information program through the lens of mental health. Our results suggest the
need to account for mental health effects of public disclosure of environmental risks and the need for a proper deliver of information
on environmental risks, especially for developing countries in which environmental threats are severe and rising and households
may lack effective defensive measures.
We proceed as follows. Section 2 describes the institutional background of the public disclosure of air-pollution information in
China. Section 3 describes the data sources and key variables. Section 4 presents our empirical methodology, baseline results, and
robustness checks. Section 5 discusses the mechanism. Section 6 considers the welfare implications and Section 7 concludes.
4 See recent evidence on the effect of information on mental health in the setting of social media (Braghieri et al., 2022; Twenge and Campbell, 2019),
internet access (Donati et al., 2022), and pandemic information dissemination (Golberstein et al., 2019; Ko et al., 2020).
5 This estimate is conservative because it does not take into account medical saving from the reduced likelihood of other mental illnesses or the productivity
gains from improved mental health.
3
Fig. 1. Annual trend for national concentrations of PM2.5 in China. Notes: This figure plots the annual level of national PM2.5 concentrations in China from 2000
to 2018. This series of ground-level PM2.5 concentrations is obtained, computed, and calibrated based on NASA’s MERRA-2 data collections and corresponding
algorithms. See Section 3.1 for further details.
2. Background
The air of major cities in China was among the dirtiest in the world in the 2010s (Zhang and Wang, 2011). The average
concentration of PM2.5 in 2010 was over 60 μg∕m3 —five times higher than the WHO’s Air Quality Guidelines (10 μg∕m3 , Guan
et al. (2014)). See Fig. 1 for the annual trend of PM2.5 concentrations in China from 2000 to 2018. However, much like during
episodes of the ‘‘killer’’ fog in London in 1952 (Wang et al., 2016), air quality in China was not regularly monitored and its health
effects were poorly understood prior to 2012. Severe smog episodes in major cities such as Beijing were frequent by early 2013
and angered the Chinese public.6 Public outcry arose from the lack of effective disclosure of pollution information and the lack
of effective government responses. To address the situation, the Chinese government has implemented a series of environmental
regulations at both local and national levels (see Appendix Section A.2). The national program of comprehensive monitoring and
real-time reporting of air pollution is a crucial policy initiative and is a milestone in China’s history of air-pollution monitoring and
disclosure. We next discuss the program’s background.
2.1. Information and awareness of air pollution prior to 2012
Limited information on air pollution The monitoring and public disclosure of air-pollution information was limited prior to 2012.
The Ministry of Environmental Protection (MEP) had built monitoring stations and mandated the collection of air pollution data in
some cities (mostly provincial capitals and regional economic centers like Beijing and Shanghai) since the early 2000s (Ghanem and
Zhang, 2014). This initial system of pollution monitoring collected information on only three air pollutants: sulfur dioxide (SO2),
nitrogen dioxide (NO2), and suspended particulates with a diameter of 10 μm or less (PM10), which were aggregated to a measure
of air quality called the air pollution index (API). API only partially reflected true air quality because it did not incorporate PM2.5,
which has already risen to be the most important air pollutant in many Chinese cities since the early 2000s. In addition, API was
not reported by mass media or news broadcasts prior to 2012. In cities with prior monitoring of API, daily API was reported in
weather reports of local newspapers and available in limited form in local government website (which often required registration
or approval to access). In contrast, for the rest of Chinese cities, there was no air pollution monitoring or reporting before 2012.
6 For example, in mid-January 2013, the air quality index (AQI) in Beijing soared as high as 993, far beyond levels WHO deems extremely dangerous. For
comparison, during the same month in New York, the AQI was below 20.
4
Misreporting on air pollution data Due to competing economic and political interests, API data were often strategically underreported
by local officials prior to 2012 (Andrews, 2008; Ghanem and Zhang, 2014; Ghanem et al., 2020). For cities with operating air
pollution monitoring stations since the 2000s, API data were collected, compiled by the local environmental authorities, and reported
to the central environmental ministry (then called the State Environmental Protection Administration, or SEPA) on a daily basis.
By the end of year, SEPA used API data to rank the performance of city officials, based on a key metric being the annual share of
‘‘blue sky days’’. A blue sky day was defined as a day with an API below 100 (API can range from 0 to 500). Cities with 85% blue
sky days or higher were praised for good air pollution control and granted the ‘‘National Model City for Environmental Protection’’.
However, this artificial blue sky day threshold created a strong incentive for local officials to underreport API data to meet
the blue sky day standard. Early studies of Andrews (2008) and Chen et al. (2012) have found evidence of API data tampering in
some cities; Ghanem and Zhang (2014) investigate API data from all cities with prior monitoring of API (113 in total) and have
documented sharp discontinuities of reported API data at the blue sky threshold (as a sign of data manipulation) for roughly half
of the cities. In contrast, there was no evidence of systematic underreporting of API at values other than the blue sky threshold.
Overall, residents in these cities were presented with a rosier picture on the threat of air pollution around the blue sky threshold.
Limited awareness of air pollution Due to limited access to air-pollution information before 2012, the public awareness of air pollution
and its severe harm on human health was extremely limited. Before 2012, government agencies and the media all referred to the
low visibility due to high concentrations of air pollutants as fog rather than smog.7 Since fog, or wu in Chinese, is understood to
be a natural phenomenon that is often associated with natural beauty in Chinese poems and songs, there was insufficient public
awareness of the substantial health risks from being exposed to such high levels of air pollution. Indeed, there were frequent reported
cases in which an elderly exercised outdoor in ‘‘foggy’’ days and was hospitalized due to acute respiratory conditions.8
In addition, the (effort) cost of obtaining information on air pollution was nontrivial before 2012. In cities with collection of API
data, daily API was available in newspapers as an less salient part of the weather report. However, smartphone and mass media had
largely replaced the newspaper as a dominant way of mass communication; but the mass media did not report API then. This further
restricted the dissemination of information on daily air pollution levels. The public could search online for daily API. However, we
show in Fig. 2 that such searching was very low before 2012. Overall, the public awareness on the severity of air pollution and its
daily fluctuations was extremely limited prior to the information program.
2.2. The information program
The national program of monitoring and reporting of air pollution in 2013 marked a turning point for the public disclosure of
air-pollution information in China. It brings two major changes in the collection and dissemination of air pollution information. First,
modern pollution monitoring stations were built across the country and more air pollutants, particularly the primary air pollutant,
PM2.5, were added to the list of real-time monitoring. Since 2013, more than 1400 U.S. EPA-grade monitoring stations, each
equipped with a real-time monitoring and automated reporting system, were installed across the country to track the concentrations
of six main pollutants, the newly added PM2.5, O3, and CO, the existing PM10, NO2, and SO2, as well as an Air Quality Index (AQI)
encompassing all six pollutants.
Second, a data-streaming system was established to stream hourly air-pollution information directly to the central government
and broadcast to the general public in real time on the MEP’s website. Third-party smartphone apps and websites were allowed to
stream and report real-time air-pollution data directly from the cloud. This spurred a surge in news media, websites, and smartphone
apps that tracked live air-pollution data (Barwick et al., 2019). As a result, the manipulation and misreporting of air-pollution
data were no longer feasible (Greenstone et al., 2022). Indeed, Liang et al. (2016) find that after the program, publicly disclosed
PM2.5 hourly concentrations were very close to those reported by US embassy/consulate based on US-controlled PM2.5 monitors.
Overall, the information program drastically increased the dissemination of real-time, high-quality, high-frequency information on
air pollution.
The 2013 information program marked a new era of environmental protection and regulation in China. It shifted government’s
strategy from prioritizing economic growth to a more balanced growth path with environmental protection. At the central level,
PM2.5 was officially recognized as a primary air pollutant and was listed as part of national objective of pollution control. The MEP
revised the national air quality standard in 2012 to set a national standard for PM2.5. For the first time in the history of national
five-year plans, the 13th Five-Year Plan required cities at prefecture level or higher to reduce PM2.5 concentrations by 18% from
2015 to 2020 (Wang, 2018). At the local level, the evaluation system for local officials was reformed accordingly: the ‘‘blue sky
day’’ metric was officially discontinued in 2013, and city officials were instead evaluated based on their performance in reducing
average annual PM2.5 concentrations (Greenstone et al., 2021).
7 For example, on November 27, 2011, Beijing newspaper headlines, as well as the China Meteorological Administration, attributed widespread flight delays
and cancellations in Beijing and northern cities to the dense fog. Local environmental bureaus, however, reported that air pollution was light during the
day. PM2.5 was not mentioned in the media or official announcements by the airport. NASA satellite showed an AOD reading of 4.5 or higher on that
day for Beijing and many other northern cities, compared with a standard range of 0.1 to 0.15 in the United States. See a Reuters report on the event at
https://cn.reuters.com/article/instant-article/idUSLNE7B401520111205.
8 For example, http://www.gov.cn/govweb/fwxx//jk/2007-12/25/content_842632.htm (in Chinese).
5
Fig. 2. Online search volume for pollution-related keywords. Notes: This figure plots the time trends of daily online search volume for pollution-related keywords
per 10,000 population, measured by the Baidu search intensity index. Panel A plots the total search index for all pollution-related keywords, which is the sum of
three categories of indices: the search index for major air pollutants (Panel B), the search index for the harms and health effects of air pollution (Panel C), and
the search index for defensive measures against air pollution (Panel D). See Appendix Table A2 for the list of keywords included in each category of indices.
Rollout schedule of the information program The information program was rolled out progressively across cities. Fig. 3 tracks the
program’s coverage across time and shows that in 3 years after 2013, the newly established monitoring network covered more than
98% of the entire population. The program had an initial pilot phase in 2012. The national roll-out started in 2013 by constructing
1436 new U.S. EPA-grade monitoring stations in three waves: (1) 74 provincial capitals and special administrative zones; (2) 116
cities from the list of Environmental Improvement Priority Cities and the list of National Environmental Protection Exemplary Cities
(both lists were compiled before 2007); and (3) the remaining 177 cities (Barwick et al., 2019; Deng and Axbard, 2021). Appendix
Figure B1 depicts the roll-out schedule on China’s map. After completion of the monitoring stations, the debut of real-time, publicly
disclosed pollution information generally occurred 6 to 10 months later; the exact timing varied due to manpower constraints on
local environmental bureaus.
Importantly, the roll-out schedule for monitoring stations and information disclosure is uncorrelated with day-to-day variation
in local pollution levels. The three-wave schedule was primarily based on predetermined hierarchical designations: administrative
centers in the first wave, environmental priority cities in the second wave (the list was compiled in 2007), and the remaining
cities in the last wave. For each city, the exact rollout date was driven mostly by the physical constraints of installing monitoring
stations and integrating the data-streaming system. Appendix Figure B2 shows that variation in the concentration of air pollution
was mostly random across roll-out dates in different cities. Moreover, no national policies coincided with the schedule or the spatial
coverage of the information program. Three regional environmental policies partially overlapped in timing: the CO2 cap-and-trade
pilot program, energy-reduction program, and fuel-switching program. We use dummy variables to control for these in all empirical
analyses.
2.3. Rising awareness of air pollution
The information program has substantially increased public awareness of the severity of local pollution level and its health
damages. This is reflected in the surge of pollution-related media reports and individual’s active search for pollution information.
6
Fig. 3. Population coverage of the information program across time. Notes: This figure plots the share of cities and population that have started monitoring and
reporting real-time air pollution across time. The solid line plots the share in cities and the dashed line plots the share in total population.
The reporting frequency of daily pollution levels and pollution-related news in government websites, news media, and mobile phone
apps increased substantially. For example, almost immediately after the program’s rollout, the frequency of ‘‘smog’’ being mentioned
in the People’s Daily – the official mouthpiece of the central government – jumped up 8 times from 2 days per month in 2012 to 16
days per month in 2013; the frequency of smog mentions in local news media increased by 60% immediately after the program’s
rollout in the city and rose up persistently afterwards (Barwick et al., 2019). There was also a surge in documentaries and special
media reports on air pollution that aimed to educate and raise public attention to air pollution and its health consequences. A
notable example is a prominent documentary titled ‘‘Under the Dome’’ by journalist-turned-environmentalist Chai Jing, which had a
long-term impact on the public awareness and WTP for cleaner air in China (Tu et al., 2020).
The information program also led to a drastic increase in individual’s active searching for pollution information. Fig. 2 plots the
trends in the daily city-level Baidu search volume of pollution-related keywords (normalized by city population) in three main
categories: air quality and air pollutants, health risks of pollution exposure, and defensive measures against air pollution. We
observe a clear surge in the search volume of all three categories after the information program. The search for each pollution-
related keyword shows a sharp increase as well (Appendix Figures B3 to B5). This shows an increase in public demand for pollution
information over a broad range of pollution issues. In contrast, search volumes for non-pollution-related keywords, such as culture,
art, entertainment, pension, and financial market, showed no sign of noticeable increases in search volume after the information
program. We present formal regression analyses of information-seeking behaviors in Section 5.
Overall, the information program led to a substantial increase in public information and public awareness of air pollution. This
facilitated avoidance behaviors against pollution exposure, such as through adjusting day-to-day transit modes and outdoor activities
and having higher WTP for air purifiers and housing in less polluted areas (Barwick et al., 2019; Neidell, 2009; Ito and Zhang, 2020).
However, the arrival of high-frequency information on polluted air and associated avoidance behaviors may create additional mental
health burdens due to additional stress, rearranged schedules, cancelled social activities, financial stress due to additional defensive
expenditure, etc. We now turn to the empirical analyses.
3. Data and methodology
3.1. Data and variables
Mental health We use the rich measure of mental health and individual characteristics from the China Family Panel Studies (CFPS),
a nationally representative survey, to track individual’s mental health responses to air pollution before and after the information
program. The CFPS covers 162 counties in 25 provinces of China, which represents 94.5% of China’s total population (excluding
7
Hong Kong, Macao, and Taiwan). Appendix Figure B6 demonstrates the geographic representativeness of CFPS counties on the map
of China. We use CFPS data from the 2010, 2012, 2014, and 2016 waves.9
Adult mental health is measured using the K6 scale, which is a six-item mental health screening instrument for depression and
distress developed by Kessler et al. (2003). The six items are summed to yield a score between zero and 24. A higher score indicates
more severe mental illness. The scale demonstrates excellent internal consistency and reliability in epidemiological studies (Kessler
et al., 2002).10 Validation studies of the K6 against diagnostic interviews have shown that the scale has a classification accuracy of
0.92 at a cutoff-point ≥13 to detect severe mental illness (Kessler et al., 2003). The brevity and accuracy of the K6 scale make it
an ideal screening instrument for serious mental illness in population-based health surveys (Kessler et al., 2003; Furukawa et al.,
2003).
The CFPS collects and reports the K6 scale across all survey waves.11 We standardize the original K6 scale to be between zero
and one to represent each respondent’s predicted probability of having mental illness (Kessler et al., 2010). Alternatively, we define
three severity indicators of mental illness (mild, moderate, and severe) based on dichotomous cutoff values. In addition, we define
three alternative indices of mental health: constant stress, confidence in the future, and life satisfaction, all standardized to be valued
between zero and one.12
We also obtain data from the China Health and Retirement Longitudinal Study (CHARLS) as supplementary data for empirical
analysis. The CHARLS surveys a nationally representative sample of Chinese residents aged 45 and above and measures survey
respondents’ mental health using the 10-item Center for Epidemiologic Studies Depression scale (CESD). We standardize the CESD
score to be between zero and one to ensure comparability with the standardized K6 scale.
Air pollution The level of air pollution is measured using the concentration of fine particulate matter with a diameter of 2.5 μm
or less (PM2.5). PM2.5 is a good proxy for air pollution and serves as an ideal subject of interest for this study for three reasons.
First, PM2.5 has already risen to be a primary air pollutant in many Chinese cities since early 2000s, and is the most important air
pollutant in China during 2010–2018 (Greenstone et al., 2021). As a result, there exists a strong correspondence between PM2.5
and the other measure of particulate matters, PM10, and between PM2.5 and the composite AQI (Appendix Figure B7).
Second, PM2.5 was the main policy target for nationwide pollution control policies since 2013. The 13th Five-Year Plan (2016–
2020) of the central government set clear, explicit objective on PM2.5 abatement for all prefecture-level cities, which was the first
time in the history of China’s five-year plans (Wang, 2018). The State Council then issued the Action Plan on Air Pollution Prevention
and Control in 2013 to outline concrete policies to reduce the PM2.5 concentrations. In other words, PM2.5 was at the epicenter of
a series of aggressive pollution abatement policies during the sample period.
Third and most importantly, PM2.5 was at the center of public attention and the most searched air pollutant of the time. Appendix
Figure B8 vividly depicts the predominance of PM2.5 among all other air pollutants in online search since 2013. Among the six
primary air pollutants (included in the composite AQI) in 2011–2019, the search for PM2.5 was low and ranked only the 4th in
early 2010s, but quickly climbed to the top of the list after 2012 and dominated any other air pollutants ever since.13
Overall, three features of PM2.5 – being a primary air pollutant, the center of pollution abatement policies, and the focus of
public attention on environmental issues – all make PM2.5 as an ideal subject of interest for our study to investigate how pollution
information changes the way individuals respond mentally to air pollution.
To overcome the difficulty whereby PM2.5 data are only available after the information program, we obtained PM2.5
concentrations from NASA’s MERRA-2 aerosol data.14 In addition to the extensive temporal coverage, the MERRA-2 PM2.5 records
have been validated to have strong correlations and good accuracy with ground-monitor-based records (Buchard et al., 2017, 2016).
The MERRA-2 PM2.5 has the advantage of covering all of China and having a geographic resolution of 1 km × 1 km per grid.
We validate that MERRA-2 PM2.5 strongly correlates with ground-monitored PM2.5 when the latter is available. Appendix
Figure B9 shows that annual MERRA-2 PM2.5 series closely tracks ground-monitored PM2.5 after the information program. A simple
regression analysis of monthly MERRA-2 PM2.5 and ground-monitored PM2.5 in 2014–2018 shows a correlation coefficient ranging
9 Some cities in 2016 wave data (survey conducted in 2015) had yet implemented the information program, and serve as the ‘‘never-treated’’ control group.
We exclude the 2018 wave from the baseline sample to mitigate the issue of forbidden comparisons for the staggered DID design (Jakiela, 2021; De Chaisemartin
and D’Haultfoeuille, 2022). See Section 4.2 for discussion. In addition, we conduct a robustness check by including the 2018 wave data.
10 Initially developed for use in the annual U.S. National Health Interview Survey, the K6 is now widely adopted in epidemiological studies (e.g., Furukawa
et al. (2003), Prochaska et al. (2012) and Weissman et al. (2015)) and has been validated across countries (Sakurai et al., 2011; Carrà et al., 2011; Lee et al.,
2012). In particular, Lee et al. (2012) and Kessler et al. (2010) have validated the consistency and reliability of the K6 scale in the Chinese population and the
cutoff of 13 to detect severe mental illness.
11 The CFPS reported the 20-item Center for Epidemiologic Studies Depression (CESD-20) scale in the 2012 survey wave. The CESD is one of the most widely
used instruments to detect clinical depression in the field of psychiatric epidemiology (Radloff, 1977; Gotlib and Hammen, 2008). We confirm in the 2012 data
that the K6 scale is strongly correlated with the CESD-20 scale, with a correlation coefficient of 0.85.
12 Across survey waves, the CFPS asked the following questions: ‘‘To what extent do you feel stressed in the past month?’’; ‘‘To what extent do you feel
confident about your future?’’; and ‘‘Overall, how satisfied are you with your life?’’ Each question is answered with zero to five. We standardize each variable
to be valued between zero and one.
13 For example, the annual search volume for PM2.5 was 4–6 times higher than all other pollutants during 2013–2017. In comparison, the search for other
air pollutants was stable and gradually increased during this period.
14 The PM2.5 data series are computed using the NASA Moderate Resolution Imaging Spectroradiometer algorithm, based on the Goddard Earth Observing
System Model (version 5.12.4). The data series of satellite-based measures of particulate matters, such as PM2.5, PM10, and AOD, are widely adopted in
economics and environmental science research when ground-monitored pollution data are not available (Greenstone et al., 2021; Barwick et al., 2019; Buchard
et al., 2017, 2016).
8
from 0.97 to 1.007, depending on whether month and year fixed effects and weather conditions are included as covariates in the
regression (Appendix Table A1).15 Overall, MERRA-2 PM2.5 series represents a good measure for ground-level PM2.5 concentrations
in our analytical sample.
Thermal inversions Thermal inversion is a frequently used instrumental variable for air pollution.16 A thermal inversion is the
reversion of the air temperature in Earth’s troposphere (the region of the atmosphere nearest Earth’s surface), which prevents the
ventilation of air pollutants to upper atmosphere and traps the air pollutants at ground level.17 Daily number of thermal inversions
measures the number of times the air temperature reverses in the troposphere over a day, and a higher number of thermal inversions
strongly predicts higher air pollution at ground level.
We obtain thermal inversion from NASA’s data collection and corresponding algorithms, and use it as an instrumental variable
for the level of air pollution in our empirical analysis. We match the monthly concentration of air pollution and the monthly number
of thermal inversions to the CFPS county and survey date. Appendix Figure B11 presents the monthly variation of thermal inversions
in our sample.18
Weather conditions We include various weather covariates. Controlling for these weather conditions helps to isolate the impact of
air pollution on mental health from weather-related factors, since weather conditions are found to affect individuals’ moods, social
behavior, and health (Berry et al., 2010; Cunningham, 1979). These variables include temperature, precipitation, cloud thickness,
relative humidity, wind speed, and air pressure, as well as an indicator for bad weather (heavy fog, rain, snow, hail, or thunder).
We obtain weather data from the China Meteorological Data Service Center and use the inverse-distance weighting algorithm to
convert weather variables from station to county level.19 Finally, all weather variables are matched to CFPS respondents by county
and month of interview.
Baidu search index We use the Baidu search index to represent awareness of and active search for pollution information. Baidu is
the most widely used search engine in China20 ; it has published daily search indices that summarize the total number of queries for
top keywords since 2011. The search index is generated using an algorithm similar to Google Trends (Vaughan and Chen, 2015).
We obtain the Baidu indices for three broad categories of top-searched pollution-related keywords: (1) keywords for air pollution
and main pollutants (such as ‘‘smog’’, ‘‘PM2.5’’, and ‘‘air pollutant’’); (2) those for the health consequences of air pollution (such as
‘‘the harm of smog’’ and ‘‘the harm of PM2.5’’); and (3) those for defensive measures against air pollution (such as ‘‘face masks’’ and
‘‘air purifier’’). We obtained 27 top keywords (Appendix Table A2). We aggregate the search indices to city-day level and normalize
them by the city population. We also obtain an alternative index normalized by the number of households with internet access as a
robustness check. In addition, we also collect search indices for top non-pollution keywords and migration-related keywords to be
used in the robustness analyses.
Regional socioeconomic covariates We collect city-level socioeconomic variables, such as GDP per capita (deflated to 2010) and city
population, from various statistical yearbooks. We also define a list of dummy variables for contemporaneous regional environmental
policies and include them as controls in our analyses. These policies include the CO2 cap-and-trade pilot program, energy-reduction
program, and fuel-switching program. See Appendix Section A.2 for further details.
3.2. Sample construction and summary statistics
We merge various datasets according to the county geocode and interview date. The initial sample contains 133,260 observations
from 34,172 respondents. We exclude individuals below 16 years old because the K6 scale is designed to measure mental health in
the adult population. We also drop observations with missing values for the K6 scale or key explanatory variables, and obtain an
analytical sample of 93,091 observations from 32,219 respondents.
Table 1 reports the summary statistics. The standardized K6 score has an average of 0.14. The score distribution is right-skewed
(Appendix Figure B12). About 37.3% of respondents reported having mild depressive symptoms (K6 ≥ 4); 12.6% moderate depressive
15 According to Buchard et al. (2016), MERRA-2 computed PM2.5 are closer to ground-monitored observation values during summer than winter. Most CFPS
surveys were conducted during the summer (Appendix Figure B10). We reestimate our baseline specification on the sample of summer-surveyed respondents and
obtain almost identical results. See Section 4.2 for discussion.
16 See, for example, Arceo et al. (2016), Hicks et al. (2016), Jans et al. (2018), Chen et al. (2018b), He et al. (2019), Sager (2019), Chen et al. (2020),
Deschenes et al. (2020), Fu et al. (2021), and Chen et al. (2022).

17 Normally, temperature decreases as altitude increases, and thus air pollutants can rise to upper atmospheric layers and disperse. Only under relatively rare
circumstances is the temperature in an upper atmospheric layer higher than the layers below. This constitutes a thermal inversion, under which the warm layer
of air traps ambient pollutants near ground level by reducing vertical circulation.
18 Thermal inversion may affect the concentrations of all air pollutants. We cannot and do not intend to isolate the effect of PM2.5 on mental health from
other air pollutants using thermal inversion as an IV; rather, PM2.5 serves as a proxy for air pollution. We cannot use composite measures of air pollution, such
as AQI and API, because data on these measures are not available for most Chinese cities before the information program.
19 We compute county-level weather variables by taking the weighted average of station-level variables within a 100 km radius around the county centroid.
Alternative radii are adopted and the results remain robust.

20 According to the China Internet Network Information Center, China had 0.56 billion internet users by the end of 2012, of which more than 99% had heard
of Baidu (seconded by Google) and more than 98% had used it in the past 6 months.
9
Table 1
Summary statistics.
Variable Obs Mean S.D. Min Max
Panel A: Mental health
K6 score (standardized)a 93,091 0.14 0.17 0 1
Mild depression (K6 ≥ 4) 93,091 0.37 0.48 0 1
Moderate depression (K6 ≥ 8) 93,091 0.13 0.33 0 1
Severe depression (K6 ≥ 13) 93,091 0.04 0.20 0 1
Constantly stressful 93,091 0.17 0.19 0 1
Panel B: Individual covariates
Female 93,091 0.51 0.50 0 1
Age 93,091 46.82 16.10 16 98
Years of schooling 93,091 7.03 4.82 0 22
Annual income per memberb 93,091 11,935 18,531 0 1,483,971
Agricultural hukou 93,091 0.72 0.45 0 1
Married 93,091 0.89 0.32 0 1
Household Expenditure 93,091 53,592 93,002 0 5,169,220
Panel B: Pollution and weather variables
Information program dummy 93,091 0.48 0.50 0 1
PM2.5 (μg/m3 ) 93,091 66.78 33.25 3.26 195.92
Thermal inversionsc 93,091 9.29 8.20 0 31
Cloud thickness (km)d 93,091 4.94 1.70 0 7.59
Temperature 93,091 14.10 5.12 0.45 24.52
Wind speed (m/s) 93,091 25.56 9.35 6.71 65.96
Precipitation (mm) 93,091 684.45 487.83 0 2517.58
Relative humidity (%) 93,091 65.00 11.20 32.48 89.03
Bad weather dummye 93,091 0.18 0.38 0 1
Panel C: Regional development and environmental policy variables
GDP per capita 93,091 55,431 34,400 5441 262,922
GDP growth rate (%) 93,091 9.95 6.52 −15.3 109
Sec indus ratio (%)f 93,091 46.81 11.38 13.57 85.45
Total population (10,000) 93,091 593.11 441.14 89.08 3391
Environmental policy dummyg 93,091 0.30 0.46 0 1
Notes: This table presents summary statistics for the main variables in the empirical analyses.
a K6 score is the sum of K6, normalized to be between zero and one.
b Household annual income per household member.
c Thermal inversions is the average daily number of thermal inversions in the month prior to the interview date.
d
Cloud thickness is related to various weather conditions, such as precipitation.
e
Bad weather dummy indicates that the weather is heavy fog, rain, snow, hail, or thunder.
f
Secondary industry ratio is the share of total value-added of secondary industries in city GDP.
g
Environmental policy dummy indicates whether the city implemented the CO2 cap-and-trade pilot program, energy-reduction program, or fuel-switching program.
symptoms (K6 ≥ 8); and 4.1% severe depressive symptoms (K6 ≥ 13),21 which is very close to the national all-age prevalence rate
of severe depression of 3.99% in 2017 (Ren et al., 2020), and lower than the U.S. all-age prevalence rate of 8.1% during 2013–
2016 (Brody et al., 2018). The gender ratio is balanced, the average age is 46.8 years old, the majority are married (89%), the
average years of schooling are 7.0, and the average annual household income per capita is about CNY 12,000.
Average monthly PM2.5 concentrations in our sample are 66.21 μg∕m3 , close to the national average in the early 2010s. Monthly
concentrations of PM2.5 vary from 3.22 to 211.04 μg∕m3 , with a standard deviation of 33.63 μg∕m3 . The average monthly number
of thermal inversions is 11.74. Statistics on weather conditions and the metrics of regional economic development are within a
reasonable range.
3.3. Empirical methodology
In this section we elaborate on our empirical methodology to assess to what extent the sharp increase in the accessibility of
pollution information may change the relationship between mental health and air pollution. A conventional staggered DID model
would be specified as follows:
𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝜃𝐷𝑗𝑡 + 𝑋𝑖𝑗𝑡 𝜖 + 𝑊𝑗𝑡 𝜙 + 𝐶𝑗𝑡 𝜇 + 𝜆𝑖 + 𝜂𝑗 + 𝛿𝑡 + 𝜋𝑝𝑡 + 𝜀𝑖𝑗𝑡 , (1)
21 To check the validity of these K6 cutoffs, we compare the K6 cutoff values against those of the CESD-20 scale in the 2012 CFPS sample. The CESD-20 scale
has a set of clinically validated cutoff values (Radloff, 1977): A cutoff score ≥16 reflects individuals at risk for clinical depression; ≥23 for moderate depression;
and ≥28 for severe depression. The K6 cutoff values of 4, 8, and 13 represent the same percentile ranks in the 2012 CFPS sample – 70th, 90th, and 96th
percentile rank, respectively – as the CESD-20 cutoff values of 16, 23, and 28.
10
where 𝐾6𝑖𝑗𝑡 is the standardized K6 scale for respondent 𝑖 in county 𝑗 at date 𝑡; 𝐷𝑗𝑡 is the program dummy that equals one after the
information program was implemented; 𝜆𝑖 , 𝜂𝑗 , and 𝛿𝑡 are individual, county, and time fixed effects, respectively; and other covariates
are discussed later.
The DID coefficient, 𝜃, measures the post-program change in the level of mental health, which consists of two potential effects:
the effect due to the disclosure of pollution information, and the effect caused by post-program level of air pollution. We expect the
former effect to be negative as the awareness of the severe air pollution and its health damages may act as an informational shock
that impairs individual’s mental health. On the other hand, we expect the latter effect to be positive as air quality improved after
the program (Greenstone et al., 2021).
We aim to disentangle these two effects and identify how the pollution information changes individual’s mental response to air
quality. Motivated by Barwick et al. (2019), we adopt the following specification:
𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝛽𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 + 𝛾𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 × 𝐷𝑗𝑡 + 𝛼1 𝐷𝑗𝑡 (2)

+ 𝑋𝑖𝑗𝑡 𝜖 + 𝑊𝑗𝑡 𝜙 + 𝐶𝑗𝑡 𝜇 + 𝜆𝑖 + 𝜂𝑗 + 𝛿𝑡 + 𝜋𝑝𝑡 + 𝜀𝑖𝑗𝑡 .
Eq. (2) is our baseline specification. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 is the log of the 30-day average concentration of PM2.5 prior to interview
date 𝑡 for county 𝑗. 𝑋𝑖𝑗𝑡 is a set of individual time-varying covariates—age, age squared, urban status (hukou), marital status,
household income, family size, and education years. 𝑊𝑗𝑡 and 𝐶𝑗𝑡 represent a vector of weather conditions and regional socioeconomic
factors, respectively. We control for individual fixed effects (𝜆𝑖 ), county fixed effects (𝜂𝑗 ), quadratic functions of year–month (𝛿𝑡 ),
and province-by-year fixed effects (𝜋𝑝𝑡 ). In particular, 𝛿𝑡 accounts for seasonal fluctuations in diseases (such as the flu) and mental
illnesses (e.g., some mental disorders are more likely to occur in winter), and 𝜋𝑝𝑡 accounts for time-varying province-specific factors,
such as provincial investment in environmental protection and economic growth, which may correlate with regional trend of air
pollution and the population’s mental health. Finally, standard errors are clustered at the household level.
The key difference to Eq. (1) is the inclusion of 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 and its interaction with the program dummy (𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 ×𝐷𝑗𝑡 )
in Eq. (2). The coefficient of 𝛽 measures the marginal effect of air pollution on mental health – termed as the gradient of mental
health to air pollution in the language of Barwick et al. (2019) – before the information program. Our primary coefficient of interest,
𝛾, measures the change in the gradient of mental health to air pollution after the information program. In other words, 𝛾 measures
how the disclosure of pollution information changes the way individuals respond mentally to air pollution.
Sources of exogenous variations Eq. (2) exploits two sources of exogenous variations. First, we adopt thermal inversion as an
instrumental variable (IV) for air pollution to identify the gradient of mental health to air pollution.22 Second, we exploit the
staggered rollout of the information program to identify the change of the gradient of mental health to air pollution caused by the
disclosure of pollution information.
Thermal inversion has been widely adopted as an instrument for air pollution in prior studies (see, for example, Arceo et al.
(2016), Hicks et al. (2016), Jans et al. (2018), Chen et al. (2018b), He et al. (2019), Sager (2019), Chen et al. (2020), Deschenes
et al. (2020), Fu et al. (2021), and Chen et al. (2022)).23 These studies show that thermal inversion is highly predictive of the
variation in air pollution at daily, weekly, monthly, and yearly levels. In our sample, Appendix Figure B13 presents a scatter plot
between the county-by-month average level of air pollution and days of thermal inversions and shows a strong correlation. We
verify in regression analysis the strong correlation between air pollution and thermal inversions (the pseudo first stage of Eq. (2))
at both the individual level (Appendix Table A3) and the county level (Appendix Table A4). We also verify the strong correlation
from short- to medium-run (from the baseline of one month to six months) in Appendix Table A5.
Thermal inversion is considered to be a valid instrument for two reasons. First, the formation of a thermal inversion depends
on the rare co-occurrence of multiple meteorological factors across layers of Earth’s atmosphere, and is independent of economic
activities at ground level.24 Second, a thermal inversion does not present a health risk or directly affect mental health (Arceo et al.,
2016). Therefore, thermal inversion has been widely adopted to instrument for air pollution.
We then exploit the variation in the roll-out schedule of the information program to identify the change of the gradient of
mental health to air pollution due to the disclosure of pollution information. As we explained in the background, the program’s
roll-out schedule is determined by the administrative hierarchy (e.g., provincial capitals) and preassigned designations (e.g., the
environmental protection priority list set in 2007). Moreover, the exact date of implementation in each city is mainly driven by
the physical and manpower constraints of MEP. Therefore, it is unlikely to correlate with the short-term variation of local pollution
22 We need to adopt an IV for the variation of air pollution because the variables in Eq. (2) are measured in biannual CFPS surveys. The temporal variation
of air pollution is very likely to correlate systematically with time-varying regional economic activities. Barwick et al. (2019) estimate a specification similar
to Eq. (2) without an IV and exploits the weekly variation in air pollution, which is more likely to be exogenous.
23 Most of aforementioned studies use the daily to monthly thermal inversions as an IV for air pollution to study the effect of short-term variation in air
pollution on various short-term outcomes of interest, including mental health. Our study follows the same methodology to identify the short-term causal effect
of air pollution by exploiting the exogenous variation in short-term thermal inversions. Two recent exceptions to this practice are Fu et al. (2021) and Chen
et al. (2022), both of which use longer-term average variation (annually or five-year average) in air pollution and thermal inversion. Fu et al. (2021) use the
annual cumulative thermal inversions as an IV for annual air pollution to study the effect of air pollution on annual measures of firm productivity. Chen et al.
(2022) use the five-year average thermal inversions as an IV for five-year average air pollution to study the effect of air pollution on long-term behaviors of
inflows and outflows of migration.
24 In Eq. (2), we control for meteorological factors, month fixed effects, and province-by-year fixed effects. The variation in thermal inversions we exploit is
therefore net of meteorological factors at ground level, seasonal effects, permanent differences across counties, county fixed effects, and year-varying differences
across provinces.
11
levels. If cities with a higher level or a rising trend in air pollution were more likely to implement the program sooner (or later),
we would observe differential pre-trends in air pollution before the start of the program. We present evidence on the lack of such
pre-trends in Appendix Figure B2. We also conduct a series of robustness tests to verify that the baseline results from Eq. (2) are
unlikely to be biased by unobservable confounders.
Drawing from latest staggered DID methodologies We draw from the recent literature on staggered DID methodology to reduce potential
biases in our estimation. This literature shows that the DID estimator from a staggered design equals a convex combination of all
possible 2-by-2 treatment effects, if the treatment effect does not change over time but may still vary across groups (Goodman-Bacon,
2021; De Chaisemartin and d’Haultfoeuille, 2020). The presence of time-varying treatment effects may bias DID estimates because
when an early-treated group serves as the ‘‘control’’ for a later-treated group, the 2-by-2 DID estimator considers the time-varying
treatment effect of the early-treated group as part of the counterfactual ‘‘untreated’’ outcome (Goodman-Bacon, 2021). Borusyak
and Jaravel (2017) refer to such case as the ‘‘forbidden comparison’’.25 See Appendix Figure B14, Panel (A) for the illustration
of a forbidden comparison when the treatment effect is a trend-break (Goodman-Bacon, 2021). The recent literature proposes
different staggered DID estimators, mostly by eliminating the forbidden comparisons and ensuring all comparisons are between
‘‘switchers’’ and ‘‘never-treated’’ or between ‘‘switchers’’ and ‘‘not-yet-treated’’ units (for example, Callaway and Sant’Anna (2021)
and De Chaisemartin and d’Haultfoeuille (2020); see De Chaisemartin and D’Haultfoeuille (2022) for an excellent review).
Motivated by this burgeoning literature, we first show in an event-study figure that time-varying treatment effect in our data
stabilizes to a level shift between survey waves, thus will not cause a severe bias in the DID coefficient estimate (see Appendix Figure
B14, Panel (B) for an illustration). Then, following the suggestion of Goodman-Bacon (2021), we correct for the potential bias by
discarding variations coming from forbidden comparisons, and alternatively, by fitting a trend break at the program’s introduction
(in the form of 𝐷 × (𝑡 − 𝑡∗ ), as in Goodman-Bacon and Schmidt (2020)) to account for the potential time-varying component of
treatment effect. Results from both analyses are similar to the baseline and discussed in Section 4.2.
4. Results
4.1. Baseline results
Table 2 reports the estimation results of Eq. (2). We include the set of control variables progressively across columns. The
progressive addition of individual, household, weather, regulatory, and socioeconomic factors causes little change in the estimates
of 𝛽 and 𝛾, which suggests the limited role of observed and unobserved confounding factors in our regression specification (Altonji
et al., 2015). Our most preferred specification is reported in Column 5, in which a full set of controls and fixed effects are included.26
Two findings are noteworthy. First, the estimate of 𝛽 validates a strong gradient of mental health to air pollution before the
information program: A 100% increase in the concentration of PM2.5 (66.7 μg∕m3 , or 2 standard deviations (SD)) in the month
prior to the interview date leads to a 0.047 unit increase in the standardized K6 (0.28 SD), which is estimated to be statistically
significant at the 5% level. This is consistent with findings in psychology and neuroscience whereby PM2.5 damages the central
nervous system (Ross et al., 2018; Mehta et al., 2015; Thomson, 2019) and increases the risk of anxiety and depression (Pun et al.,
2017; Power et al., 2015).
To build intuition about the magnitude of our baseline effects, we provide two benchmarks. We first compare to previous studies
on the causal effect of air pollution on mental health. Our estimate of 𝛽 is almost identical to that of 0.048 obtained by Chen
et al. (2018b), who estimate the causal effect of PM2.5 on standardized K6 in the pre-program period using the same dataset. The
estimate is also similar qualitatively to that of Zhang et al. (2017) based on standardized measures of subjective well-being. We
then benchmark the magnitude of our estimates against the effect of a sudden unemployment spell on mental health. Comparing
our estimates to the average effect of unemployment on mental health in a meta-analysis by Paul and Moser (2009), we find that
the impact of a 1-SD increase in PM2.5 concentrations on mental health before the information program (0.14 SD of K6 score) is
around 37 percent of the causal effect of job loss due to factory closure or mass layoffs (averaged 0.38 SD of six measures of mental
health).27
Second and more importantly, the estimate of 𝛾 shows that introduction of the information program has led to a statistically
significant increase in the gradient of mental health to air pollution. After the information program, a 100% increase in the
concentration of PM2.5 leads to an additional 0.083 unit increase in the standardized K6 (0.49 SD), which is estimated to be
statistically significant at the 1% level. The estimated change in the gradient (𝛾) is almost twice as large as the pre-information
gradient (𝛽). To put this into perspective, before the information program a one-SD increase in PM2.5 concentration increases the
standardized K6 by 0.014 SD; after the program, a one-SD increase in PM2.5 leads to a 0.038 SD increase in the standardized K6,
similar in magnitude to the estimate of Xue et al. (2019) based on post-information-program data. This result thus reconciles the
25 The issue of forbidden comparison is likely the most severe when there exists a group that is treated most of the time, and during a period when all
groups are treated (De Chaisemartin and D’Haultfoeuille, 2022). This has led Jakiela (2021) to propose dropping the ‘‘always-treated’’ groups and periods when
all groups are treated. There is no ‘‘always-treated’’ group in our data, and we have excluded from our baseline analysis the period when all groups have been
treated (the 2018 CFPS wave). We conduct a sensitivity check to include the 2018 wave later.
26 For space constraints, we only present the estimated coefficients for 𝛽 and 𝛾 in the main text. Complete estimation results are available upon request.
27 In the meta-analysis, Paul and Moser (2009) include studies estimating various measures of mental health, including indicators of distress, depression,
anxiety, psychosomatic symptoms, subjective well-being, and self-esteem. Paul and Moser (2009) estimate an average treatment effect of unemployment as 0.38
SD of mental health measures based on studies that exploit quasi-experimental variation in job loss due to factory closures and mass layoffs.
12
Table 2
Effects of information disclosure on the gradient of mental health to air pollution.
Variables (1) (2) (3) (4) (5)
K6 K6 K6 K6 K6
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) (𝛽) 0.049** 0.050** 0.046** 0.047** 0.047**
(0.022) (0.022) (0.022) (0.022) (0.022)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 (𝛾) 0.076** 0.076** 0.081*** 0.084*** 0.084***
(0.030) (0.030) (0.030) (0.030) (0.030)
Observations 93,091 93,091 93,091 93,091 93,091

Individual FE Yes Yes Yes Yes Yes
County FE Yes Yes Yes Yes Yes
Interviewer FE Yes Yes Yes Yes Yes
Prov × Year FE Yes Yes Yes Yes Yes
Cubic month controls Yes Yes Yes Yes Yes
Individual controls Yes Yes Yes Yes
Weather controls Yes Yes Yes
City economic controls Yes Yes
Env. policy controls Yes
Kleibergen–Paap F-stat 592 593 616 601 601
Mean Dep. Var. 0.141 0.141 0.141 0.141 0.141
Notes: This table presents the estimated effects of information disclosure on the gradient of mental health to air pollution.
The dependent variable is the standardized K6. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) is the log of the average concentration of PM2.5 in the month
prior to the respondent’s interview date in the respondent’s county of residence. 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 is the dummy variable that equals
one if real-time air-pollution information is publicly available in the county. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒
are instrumented by log of thermal inversions and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒. Regression models are specified in Eq. (2).
All regressions across columns control for individual fixed effects, county fixed effects, interviewer fixed effects, a quadratic
function in month–year, and province-by-year fixed effects. Individual controls include age, age squared, years of schooling,
hukou status, marriage status, and family size. Weather controls include temperature, wind speed, cloud thickness, precipitation,
relative humidity, and an indicator for extreme weather conditions (heavy fog, rain, snow, hail, or thunder). City economic
controls include GDP per capita (price deflated to 2010), GDP growth rate, and share of value-added from secondary industry
to total city GDP. Environmental policy dummy indicates whether the city implemented the CO2 cap-and-trade pilot program,
energy-reduction program, or fuel-switching program. Standard errors in parentheses are clustered at the household level. ***P
< 0.01, **P < 0.05, *P < 0.1.
discrepancy in estimated dose response of mental illness to air pollution between Chen et al. (2018b) and Xue et al. (2019), and
highlights the importance of air-pollution information in magnifying the mental health impact of air pollution.
We further explore the effect on the risk of depression and subjective well-being (SWB). We define a set of dummies to indicate
mild depression (K6 ≥ 4), moderate depression (K6 ≥ 8), and severe depression (K6 ≥ 13).28 Table 3, Columns 1–3 report the
results and show that the information disclosure has led to a large and statistically significant increase in the gradient of mild and
moderate depression to air pollution, and a positive and marginally significant effect on the gradient of severe depression. Column
4 shows a consistent pattern of an increased gradient of constant stress to air pollution. Columns 5 and 6 use two measures of SWB
as dependent variables – an indicator for confidence in the future and an index of life satisfaction – and show that introduction of
the information program has significantly reduced the (negative) gradient of SWB to air pollution.
It is worth noting that before the information program, an increase in air pollution increased the risk of depression and stress,
but had a muted effect on SWB (i.e., the small and statistically insignificant estimate of 𝛽 in Table 3, Columns 5–6). This pattern
is also documented by Zhang et al. (2017) based on pre-information-program data. Zhang et al. (2017) conclude that air pollution
reduces individuals’ short-term happiness and increases the rate of depressive symptoms, but has no immediate effect on long-term
measures of SWB such as life satisfaction. However, our results show that air pollution would have a significantly negative effect on
life satisfaction after pollution information becomes readily available. This suggests that information on air pollution has changed
individuals’ ‘‘mental model’’ of air pollution, potentially by inducing them to reassess the severity of local pollution levels and
updating their expectations regarding the harms of air pollution on their physical and mental health. We explore these factors in
later analyses.
Event study We validate that there was no noticeable change in the gradient of mental health to air pollution before the information
program. We estimate an event-study specification:
∑
𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝛼1 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 + 𝛾𝑞 ⋅ 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 × 𝑑(𝑡 = 𝑞)𝑗 (3)
𝑞≠−1
∑
+ 𝛽𝑞 ⋅ 𝑑(𝑡 = 𝑞)𝑗 + 𝑋𝑖𝑗𝑡 𝜖 + 𝑊𝑗𝑡 𝜙 + 𝐶𝑗𝑡 𝜇 + 𝜆𝑖 + 𝜋𝑝𝑡 + 𝜀𝑖𝑗𝑡 ,
𝑞≠−1
28 Lee et al. (2012) and Kessler et al. (2010) have validated the K6 cutoff of 13 to detect severe mental illness in the Chinese population. We cross-validate
the cutoff values in 2012 CFPS data against the 20-item CESD scale. The CESD-20 scale has a set of clinically validated cutoff values (Radloff, 1977). The K6
cutoff values we adopt correspond to the CESD-20 cutoff values at the same percentile rankings in the 2012 CFPS sample. See footnote 21 for more details.
13
Table 3
Effects of information disclosure on the gradient of mental health to air pollution.
Variables (1) (2) (3) (4) (5) (6)
Mild depression Moderate depression Severe depression Constant stress Confidence in future Life satisfaction
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) (𝛽) 0.209*** 0.007 −0.015 0.078*** 0.085 0.024
(0.066) (0.049) (0.031) (0.026) (0.145) (0.141)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 (𝛾) 0.325*** 0.192*** 0.078* 0.077** −0.558*** −0.603***
(0.096) (0.068) (0.043) (0.037) (0.205) (0.204)
Observations 93,091 93,091 93,091 93,301 92,695 92,695

Individual FE Yes Yes Yes Yes Yes Yes
County FE Yes Yes Yes Yes Yes Yes
Interviewer FE Yes Yes Yes Yes Yes Yes
Time FE Yes Yes Yes Yes Yes Yes
Control variables Yes Yes Yes Yes Yes Yes
Kleibergen–Paap F-stat 597 597 597 598 595 595
Mean Dep. Var. 0.373 0.126 0.041 0.166 0.64 0.70
Notes: This table presents the estimated effects of information disclosure on the gradient of mental health to air pollution. Dependent variables are the indicator
for mild depression (K6 ≥ 4), moderate depression (K6 ≥ 8), severe depression (K6 ≥ 13), the indicator for feeling stressful constantly, the indicator for confidence
in the future, and self-reported life satisfaction. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) is the log of the average concentration of PM2.5 in the month prior to the respondent’s interview
date in the respondent’s county of residence. 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 is the dummy variable that equals one if real-time air-pollution information is publicly available in the
county. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 are instrumented by log of thermal inversions and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒. Regression models
are specified in Eq. (2). All regressions control for individual fixed effects, county fixed effects, interviewer fixed effects, a quadratic function in month-year,
and province-by-year fixed effects. Control variables are specified the same as in Table 2. Standard errors in parentheses are clustered at the household level.
***P < 0.01, **P < 0.05, *P < 0.1.
where the quarter-to-program dummy 𝑑(𝑡 = 𝑞)𝑗 equals one if the respondent is interviewed 𝑞 quarters since the start of the
information program at county 𝑗, and 𝑑(𝑡 = −1) is the last quarter before the program and is omitted as the reference period.
We set the event study window from 5 quarters before the program to 6 quarters after the program to ensure sufficient sample size
for the estimation of each 𝛾𝑞 coefficient.29 We dummy out the remaining sample periods. PM2.5 and its interaction terms with the
quarter-to-program dummies are instrumented by corresponding terms of thermal inversions.
Estimates of 𝛾𝑞 ’s trace the evolution of the gradient of mental health to air pollution in quarters relative to the program’s rollout.
Fig. 4 plots the estimates of 𝛾𝑞 ’s with the 95% confidence interval. Estimation results are reported in Appendix Table A6. All pre-
program coefficient estimates, 𝛾−5 to 𝛾−2 , are small and statistically insignificant, suggesting a lack of differential change in the
gradient of mental health to air pollution across cities before the program’s rollout.30
The post-program coefficients are positive, grow in size in the first two quarters after the program (𝛾0 and 𝛾1 ), but stabilize
quickly afterwards (𝛾2 to 𝛾5 ). This suggests that the gradient of mental health to air pollution has stabilized between survey waves
(CFPS surveys are spaced two years apart). In the language of Goodman-Bacon (2021), the treatment effects experienced by the
‘‘early-treated’’ group had changed into a level shift before being included as part of the counterfactual outcome in a 2-by-2 DID
comparison, as illustrated in Appendix Figure B14, Panel (B). Therefore, such a mostly level-shifting treatment effect is unlikely to
cause severe biases in our estimation.
It is worth noting that Eq. (3) is different from an event study specification in standard DID (Eq. (1)). First, an instrumental
variable is used to elicit the exogenous variation of air pollution. Second, the coefficients of interest, 𝛾𝑞 ’s, measure the change in
the gradient of mental health to air pollution, rather than the change in the level of mental health as in standard DID. In other
word, 𝛾𝑞 ’s do not track the change in average mental health after the information program, but track the change in individuals’
mental sensitivity to a ‘‘shock’’ in air quality after the program. This helps explain the increase of 𝛾𝑞 ’s after the program: Individuals
obtained real-time, high-quality information on local pollution levels and became more aware of the substantial health damages of
pollution exposure after the information program; they therefore became more sensitive to changes in air pollution afterwards.
Caveats Some caveats of above baseline results are worth noting. First, we study the effect of information disclosure on individual’s
gradient of mental health to short-term fluctuation in air pollution. Our estimates cannot speak directly to a change in long-term
air quality.31 Second, our sample period is 2010–2016, when air pollution in China was severe. Air quality improved substantially
afterwards. The gradient of mental health to air pollution may be different in post-2016 period, such as due to reduced public
attention to air pollution issues. Third and relatedly, our estimates are muted to the possibility that years of experience with pollution
information might teach individuals to learn to adapt and mitigate the effects of such information shock on their mental health.
Lastly, the information program changed the dissemination of pollution information in various aspects, including extending the
29 We use a quarter-to-program specification in Eq. (3) rather than month-to-program because the sample size of some month-to-program bins is too small to
estimate the event-study coefficients precisely (see the number of observations in each month-to-program bin in Appendix Figure B15).
30 We also conduct a conventional event-study estimation (with only quarter-to-program dummies and no interactions with log air pollution) and find similarly
small and insignificant pre-program coefficients, confirming the lack of deviating pre-trends in the level of mental health before the program. See discussion in
the next section.
31 One of our robustness analysis looks at the change in air quality up to 6 months and finds that the effects, if anything, increase in the short to medium
term. Longer-term effects, however, could be quite different.
14
Fig. 4. Event study estimates of the effect of information program on the gradient of mental health to air pollution. Notes: This figure plots event study estimates
of the gradient of mental health to air pollution before and after the information program. The regression equation is specified in Eq. (3). The dependent variable
is standardized K6 scores. Coefficients of interaction terms between log PM2.5 and quarter-to-policy dummies, as well as 95% confidence intervals, are plotted
in the figure. Regression results are reported in Appendix Table A6, Column (1).
public access, improving the data quality, raising the reporting frequency, and shifting the main source of information from local
newspapers to mass media and smartphone apps. All of these changes occurred at the same time at the introduction of the program,
leaving little identifiable variation of different aspects of the information program. We therefore cannot disentangle the effect of
each aspect of the information program. With these caveats in mind, we move on to check the robustness of our baseline results.
4.2. Robustness analyses
In this section, we first follow the latest literature on staggered DID methodology to deal with the potential biases of time-varying
treatment effects. We then examine the robustness of the baseline results to alternative model specifications, sample restrictions,
and variable definitions, and show that the results are unlikely to be driven by unobserved confounding factors.
Dealing with potential time-varying treatment effects As discussed in Section 3.3, the recent literature on staggered DID design shows
that the presence of time-varying treatment effects may bias the DID estimator. We eliminate potential biases due to time-varying
treatment effects in two ways. First, we follow the decomposition method of Goodman-Bacon (2021) to directly discard variations
from ‘‘forbidden comparisons’’—the 2-by-2 DID comparisons between a late-treated group and an early-treated group. We do so by
dropping observations of early-treated units after their first post-program wave so that these early-treated units will not be used as a
‘‘control’’ group.32 Appendix Table A7 show that the estimated coefficient of interest, 𝛾, is 0.094, similar to the baseline estimate of
0.084. Alternatively, in a similar fashion as Goodman-Bacon and Schmidt (2020), we fit a trend break at the program’s introduction
(in the form of 𝐷 × (𝑡 − 𝑡∗ )) to separate out the potential time-varying component of treatment effects. Results are presented in
Appendix Table A8. The estimated 𝛾 is 0.086, very close to the baseline estimate.
Conventional staggered DID We estimate a conventional staggered DID specification, Eq. (1). Appendix Table A9 shows that the
estimated DID coefficient (𝜃) is close to zero and statistically insignificant. As discussed earlier, the coefficient of 𝜃 may consist of
two effects on mental health: the effect from the information disclosure and the effect from improved air quality. We estimate an
event-study specification of Eq. (1) and plot the coefficients of time-to-policy dummies in Appendix Figure B16. As we expected, the
post-program coefficients show that there was a deterioration of mental health immediately after the program’s introduction, likely
due to the awareness of the local pollution levels, but the effect gradually subdued and turned to the opposite direction after the
third quarter as air quality improved. Despite the different interpretation of event-study coefficients as in Fig. 4, the close-to-zero
pre-program coefficients in Figure B16 again validate that there existed no deviating pre-trends across cities in different waves of
the program’s rollout.
32 A potential drawback from such a practice is that we discard about 25% of sample and restrict to a shorter panel for the early-treated group.
15
Placebo test We conduct a placebo test by replacing the information program dummy (𝐷𝑗𝑡 ) in Eq. (2) with a placebo policy dummy
defined as occurring 1 year, 2 years, or 3 years prior to the actual start date of the information program. The rationale is that
respondents in cities with different program roll-out dates should not experience changes in the gradient of mental health to air
pollution before the information program. The results are presented in Appendix Table A10 and show no statistically significant
change in the gradient of mental health to air pollution following the placebo policy time.
Physical health and cognitive functioning We show that the estimated increase in the gradient of mental health to air pollution is
unlikely to be driven by physical health or cognitive functioning. We conduct two tests. First, in Appendix Tables A11 and A12,
we estimate Eq. (2) using various measures of cognitive functioning and physical health as the dependent variable.33 Results show
that the information program does not lead to any statistically significant changes in the gradients of cognitive functioning or
physical health outcomes to air pollution. This suggests that information program is unlikely to change the physical health effect
of air pollution. Second, in Appendix Tables A13 and A14, we re-estimate Eq. (2) after controlling for the measures of cognitive
functioning and physical health. Estimation results are almost identical to the baseline results in Table 2. Therefore, the estimated
increase in individual’s gradient of mental health to air pollution is unlikely to be driven by changes in cognitive functioning or
physical health.34
Labor supply and migration We check whether the estimated change in the gradient of mental health to air pollution is driven by
changes in labor supply or moving to cities with cleaner air. We use household income and migration status as the dependent variable
in Eq. (2) and find no statistically significant change in the gradients of household income or migration status to air pollution after the
information program (Appendix Table A15). This is expected in our setting as migration to a different city is a long-term decision
that involves substantial financial cost, and for the general public, such cost may outweigh the short-term health burden of air
pollution in their place of residence.35 We interpret this result cautiously, however, because we do not have information on changes
in occupation or within-city residential location, and our sample features a relatively low cross-city migration rate compared with
the national average.36 In addition, we assess the online search intensity for migration-related keywords – such as migration-required
permits and migration-related means – and find no statistically significant increase in the gradient of searching for these keywords
to air pollution after the program (Appendix Table A16).37
Different sample specifications Appendix Table A17 shows that our results are robust to excluding observations when the interviewer
rated the respondent as lacking comprehension, interest, or credibility or expressing excessive doubt, which may generate additional
measurement error on self-reported mental health. Appendix Table A18 shows that results are robust to the following sample
specifications: (1) excluding observations when the survey was conducted outside of the regular survey window (Column 1);38
(2) excluding observations when the respondent was interviewed during extreme weather (Column 2); (3) excluding cities that
implemented some form of pollution monitoring in the early 2000s (Columns 3 and 4);39 and (4) excluding individuals who ever
lived in a different city during the sample period (Column 5). Lastly, Appendix Table A19 shows that results are consistent when
we include CFPS 2018 data in the analysis.40
Clustering of standard errors and removal of ‘‘bad controls’’ Appendix Table A20, Columns 1 and 2 show that the estimated coefficients
of interest (𝛾) are statistically significant at the 5% level and 10% level, respectively, when we cluster standard errors at county
and city level. Appendix Table A20, Column 3 shows qualitatively similar results when we remove individual fixed effects. Column
4 shows that the results are almost unchanged when we remove labor market outcomes and educational attainment from the set of
covariates, which may introduce the ‘‘bad control’’ problem.
33 We adopt four measures of cognitive functioning: standardized test scores for math, memory, reading comprehension, and IQ. We use eight measures of
physical health: indicators for (1) acute disease, (2) chronic disease, (3) cardiovascular disease, (4) respiratory disease, (5) circulatory/digestive/urinary disease,
(6) cognitive disorders, (7) self-reported to be healthy, and (8) ever hospitalized in the past month.
34 If a decline of physical health would aggravate individual’s mental vulnerability to air pollution, the improvement in air quality after the information
program (see Fig. 1) would lead to an improvement in physical health and a reduction of individual’s gradient of mental health to air pollution, which is
opposite to our baseline finding. In this regard, our estimated increase in the gradient of mental health to air pollution after the information program would be
a lower bound to the true effect. We thank an anonymous referee for this suggestion.
35 We expect that such preference for pollution-induced migration may be stronger over the longer term and for the better educated and more financially
accomplished individuals. Such pattern of heterogeneous pollution avoidance behaviors have been observed over a 15-year time span in Chen et al. (2022).
36 We define migration status as ever having lived outside one’s home town for at least 6 months in a year. The rate of cross-city migration is 12.6% in our
sample, compared with 20% nationally in 2014 (based on the same definition of migration), according to China’s National Bureau of Statistics.
37 See detailed discussion on the construction of search intensity measures in Section 5.1.
38 Most CFPS interviews are conducted in the summer and winter, when the interviewers – most of whom are college students – are on school holiday. See
Appendix Figure B12 for the survey month distribution of our sample observations.
39 For example, cities with a U.S. embassy or consulate had installed on-site PM2.5 monitors before the information program, and so residents may have some
prior knowledge of local pollution levels.

40 We do not include CFPS 2018 data in the baseline sample because the information program has been implemented in all cities by the end of 2016, leaving
no variation in the dummy of information program in the 2018 wave. We exclude this sample period to mitigate the issue of forbidden comparisons as suggested
by the latest staggered DID methodology (Jakiela, 2021; De Chaisemartin and D’Haultfoeuille, 2022).
16
Alternative dataset We conduct a sensitivity analysis based on the CHARLS data, which measures respondent’s mental health using
a 10-item CESD and provides a nationally representative sample of China’s elderly population. We construct the CHARLS sample in
the same way. The CHARLS sample is similar in basic household characteristics to the CFPS sample, except that respondents were
on average 12 years older than those in the CFPS (59 vs. 47 years old). We include the CHARLS waves 2011, 2013, and 2015.
Appendix Table A21 shows that the results are consistent: introduction of the information program substantially increases the
gradient of mental health to air pollution across various mental health outcomes, including standardized CESD score and dummies
for depression. It is noteworthy that the estimated gradient of mental health to air pollution – both before and after the information
disclosure (𝛽 and 𝛽 + 𝛾) – are larger than those estimated in the CFPS data. A potential explanation is that the older cohorts in the
CHARLS are more vulnerable physiologically and mentally to air pollution, and therefore would be more sensitive to the information
of deteriorating air quality. We show, based on a heterogeneous effect analysis in the next section, that the information program has
a larger effect on the gradient of mental health to air pollution for older cohorts than younger ones. Finally, we show in Appendix
Table A22 that the pattern is consistent in the pooled sample of CFPS and CHARLS.41
4.3. Heterogeneous effects
Heterogeneity by age group We observe that older cohorts – aged 50 years and above – experience a larger sensitivity of mental
health to air pollution after the information program than younger cohorts (Appendix Table A23). This is likely due to the fact
that older cohorts are more vulnerable in physical health and mental health to air pollution (Zhang et al., 2018; Neidell, 2009),
and therefore would be more mentally sensitive to information on air pollution. This age-dependent effect is also reflected in the
estimation results based on the older CHARLS sample.
Heterogeneity by gender and socioeconomic status We do not detect a clear treatment effect heterogeneity by gender. Although males
are in general more physically susceptible to harm from air pollution than females (Zhang et al., 2018), Appendix Table A24,
Columns 1–2 show that males are not more responsive in mental health to pollution information than females.
We observe a notable heterogeneity with respect to socioeconomic status. Appendix Table A24, Columns 3–4 show that urban
residents are more responsive to pollution information than rural ones, and Columns 5–6 show that more educated individuals are
more responsive to pollution information than less educated. This is likely because urban residents and the better educated were
more exposed to pollution information when such information became available—such as being more likely to have internet access
and own a smartphone. We provide further analysis based on different exposure to pollution information later.
Heterogeneity by predicted susceptibility to mental illness We find heterogeneous effects across quartiles of predicted susceptibility to
mental illness. We follow the practice of Braghieri et al. (2022) and define individual’s susceptibility to mental illness in two steps.
First, we implement a least absolute shrinkage and selection operator (LASSO) in the pre-program period to predict an indicator
of moderate depression (𝐾6 ≥ 8) using pre-determined individual-level characteristics (age, gender, educational attainment, and
an indicator for rural hukou) and all two-way interactions between these characteristics. We then use the model selected by the
Extended Bayesian Information Criterion (EBIC) to generate an index of susceptibility to mental illness as the pre-program average
predicted likelihood of moderate depression. We separately estimate the baseline specification in subsamples of individuals defined
by quartiles of the index of susceptibility. Appendix Table A25 presents the estimation results and show that individuals with
a greater susceptibility to mental illness exhibited a stronger increase in the gradient of mental health to air pollution after the
information program.
Heterogeneity by city’s prior monitoring status As discussed in the background, there exists a group of cities with some limited forms
of air pollution monitoring and reporting prior to the information program. Residents in these cities had prior exposure to pollution
information (they could obtain daily API data from local newspapers), thus might respond differently than residents in cities with
no prior pollution monitoring or reporting.
The information program creates different change in the information disclosure relative to the pre-program period in these two
groups of cities. For cities with no prior monitoring, the program provides air pollution information for the first time and reports such
information at a high frequency; for cities with some prior monitoring, the program improves the quality of pollution information
(by eliminating potential data manipulation) and increases the reporting frequency (from daily to hourly) (Andrews, 2008; Ghanem
and Zhang, 2014; Greenstone et al., 2022, 2021). After the information program, both groups of cities have installed the same
high-standard monitoring stations and started reporting real-time pollution information at the same hourly frequency. Therefore,
residents in cities with no prior pollution information (the no-information group) experienced a greater relative increase in exposure
to air pollution information than residents in cities with some prior pollution information (the some-information group).42 We expect
41 CFPS and CHARLS use different scales to measure mental health—CFPS uses the K6 scale while CHARLS uses the 10-item CESD. We standardize both scales
before running the pooled regression.

42 Note that we cannot fully disentangle the effect of information access from the effect of information quality in such a subsample analysis, because the
information program introduced high-quality pollution information to both groups of cities. In addition, the reporting frequency increased by different extents
– from none to hourly versus from daily to hourly – in both groups of cities. Therefore, we interpret the difference treatment effects between these two groups
as driven by different change in the exposure to pollution information relative to their respective pre-program level of information exposure.
17
a larger increase in the gradient of mental health to air pollution in the no-information group. We divide our sample into two groups
accordingly and estimate the baseline specification on each.43
Appendix Table A26 presents the estimation results separately on the no-information and misreported-information groups. Two
findings are worth noting. First, before the information program, the gradient of mental health to air pollution (𝛾 in Eq. (2)) is
larger in the some-information group than that in the no-information group (0.103 vs. 0.059). This is expected because in the some-
information cities, although pollutant information was potentially underreported, residents had some information and awareness of
the severity of local air pollution.44 This helps to explain the larger gradient of mental health to air pollution in the some-information
group before the information program.
Second and more importantly, the post-program change in the gradient of mental health to air pollution (𝛽 in Eq. (2)) is large
and statistically significant in the no-information group, and larger than that in the some-information group (0.129 vs. 0.068).45
This is likely because the post-program change in the individual’s exposure to pollution information is larger in the no-information
cities than the some-information cities. Consistently, we find that the increase in the gradient of online searching for pollution
keywords to air pollution is greater in the no-information group relative to the some-information group (Appendix Figure B17),
supporting the notion that residents in no-information group experienced a greater increase in information exposure relative to
their pre-program level. On the other hand, the estimated positive 𝛽 in the some-information group suggests that the increase in
information quality and reporting frequency also contributes to a larger gradient of mental health to air pollution for residents with
some prior information exposure. Summing up, the post-program gradient of mental health to air pollution (𝛾 + 𝛽) is similar in level
between the two groups of cities (0.171 versus 0.184). Overall, Appendix Table A26 presents consistent results that both the pre-
and post-program gradients of mental health to air pollution depend on the level of exposure to available pollution information.
So far we have shown that the information program increased the gradient of mental health to air pollution. The pattern is
more pronounced among individuals who were more susceptible to mental illnesses and those experiencing a greater increase in
the exposure of pollution information. Our results also suggest that different aspects of the information dissemination, including the
ease of access, data quality, and reporting frequency, are all important drivers of our baseline results. Motivated by these findings,
we hypothesize that individuals who received real-time frequent pollution information would become more alert and stressful when
local air pollution worsens, especially when they were aware of the health consequences of air pollution, and therefore had a greater
mental sensitivity to changes in air quality. We explore this in the next section.
5. Channels
We proceed to investigate the channels through which the information program results in a greater gradient of mental health
to air pollution. We proceed in two steps. First, we show that after the information program, individuals became more aware and
attentive to air pollution issues, as reflected in a greater gradient of active information seeking to air pollution. We then show
that the air pollution information may induce two types of behavioral responses that affect mental health: a direct stress effect by
which news of severe air pollution is a source of stressors to individual’s mental system, and an indirect avoidance effect by which
information of severe air pollution induces more avoidance behaviors, such as reduced outdoor exercise, cancelled outings, and
reduced social activities, which in turn may adversely affect mental health.
5.1. Awareness and active search for pollution information
The information program prompted individuals to be more attentive to air pollution and more active in seeking pollution
information when the local air quality worsened. The gradient of online search intensity to air pollution increased substantially after
the information program. We measure search intensity using the Baidu search index of pollution-related keywords.46 We categorize
pollution-related keywords into three main groups: (1) keywords on air pollution and air pollutants, (2) the health consequences of
air pollution, and (3) means of avoidance and defensive equipment. Based on city-by-month Baidu search index data matched with
monthly air pollution levels, we estimate the specification below:
𝐼𝑛𝑑𝑒𝑥𝑐𝑡 = 𝛼0 + 𝛽𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑐𝑡 + 𝛾𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑐𝑡 × 𝐷𝑐𝑡 + 𝛼1 𝐷𝑐𝑡 (4)

+ 𝑊𝑐𝑡 𝜙 + 𝐶𝑐𝑡 𝜇 + 𝜂𝑐 + 𝛿𝑡 + 𝜋𝑝𝑡 + 𝜀𝑐𝑡 ,
43 Appendix Table A27 tabulates sampled CFPS cities included in the three rollout waves of the information program. The dark shaded cells indicate the
some-information cities that were documented by Ghanem and Zhang (2014) as being suspected of API data manipulation before the information program, the
rest of cells indicate the no-information cities. The some-information group accounts for only 30% of observations in our sample; in addition, the information
program had been rolled out in all some-information cities in two waves.
44 Ghanem and Zhang (2014) find that underreporting mainly occurred around a policy-induced cutoff of API = 100, below which a day would be counted as a
‘‘blue sky day’’ in the city’s environmental evaluation. Therefore, local officials had incentives to underreport near this cutoff to meet the blue sky day threshold.
There was no clear sign of underreporting at other pollution levels. Therefore, residents in the some-information group may have imprecise, underreported
information on API around the cutoff value of 100, but had relatively accurate information on API at other API levels.
45 We caution that the estimates of 𝛾 and 𝛽 are not statistically significant in the some-information group, likely due to the smaller sample size and that the
variation of the program’s rollout schedule is limited in the some-information group (Appendix Table A27).
46 Baidu is the most widely used internet search engine in China, with more than 98% of internet users having used its service in the past 6 months.
18
Table 4
Effects of information disclosure on the gradient of search for pollution information to air pollution.
Search category (1) (2) (3) (4)
All keywords Air pollutants Pollution’s harm Defensive devices
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) (𝛽) 0.107*** 0.341*** 0.009*** 0.005
(0.018) (0.051) (0.003) (0.012)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 (𝛾) 0.103*** 0.272*** 0.018*** 0.049**
(0.027) (0.068) (0.006) (0.024)
Observations 15,309 15,309 15,309 15,309

Month FE Yes Yes Yes Yes
City FE Yes Yes Yes Yes
Prov × Year FE Yes Yes Yes Yes
Control Variables Yes Yes Yes Yes
Kleibergen–Paap F-stat 24 24 24 24
Mean Dep. Var. 0.200 0.314 0.008 0.341
Notes: This table presents estimated effects of information disclosure on the pollution gradient of search volume for pollution-
related information. Dependent variable is log monthly search volume per 10,000 city population for specified keywords.
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) is the log of the average concentration of PM2.5. 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 equals one if real-time air-pollution information
is publicly available. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 are instrumented by log of thermal inversions and its
interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒. Regression models are specified in Eq. (4). All regressions control for city fixed effects, month fixed
effects, and province-by-year fixed effects. Weather controls include temperature, wind speed, cloud thickness, precipitation,
relative humidity, and an indicator for extreme weather conditions (heavy fog, rain, snow, hail, or thunder). Socioeconomic
controls include GDP per capita (price deflated to 2010), GDP growth rate, and share of value-added from secondary industry
to total city GDP. Environmental policy dummies include whether the city implemented the CO2 cap-and-trade pilot program,
energy-reduction program, or fuel-switching program. Standard errors in parentheses are clustered at the city level. ***P < 0.01,
**P < 0.05, *P < 0.1.
where the dependent variable is the total number of Baidu search queries for top pollution-related keywords per 10,000 population
in city 𝑐 in month 𝑡 (see Appendix Table A2 for the list of included keywords).47 Other variables are specified the same as in Eq. (2).
Standard errors are clustered at the city level. Our coefficient of interest, 𝛾, denotes the change in the gradient of search intensity
to air pollution after the information program, and the coefficient 𝛽 represents the pre-program gradient.
Table 4 reports the estimation results and shows that the introduction of the information program has doubled the gradient of
search intensity (𝛽 = 0.107 and 𝛾 = 0.103, both are estimated statistically significantly at the 1% level, Column 1). This pattern is
consistently observed for all three categories of pollution-related keywords, including searches for major air pollutants (Column 2),
the health effects of air pollution (Column 3), and defensive measures against air pollution (Column 4). Appendix Table A29 reports
similar results for searches for each included keyword. In addition, we estimate the search volumes from personal computers (PCs)
and mobile platforms (smartphones and tablets) separately. Appendix Table A30 shows that the pattern of an increased gradient of
search intensity is stronger on mobile platforms than on PCs, consistent with the trend of surging pollution-broadcasting apps after
the information program (Barwick et al., 2019).
In addition, we conduct an event-study analysis analogous to Eq. (3) (Appendix Table A32). Fig. 5 plots the coefficients of
quarter-to-program dummies and shows that all pre-program coefficients are small and statistically insignificant, which confirms
the absence of changes in the gradient of search intensity to air pollution before the information program. After the information
program, the gradient of search intensity rose immediately and stabilized after two quarters, similar to the dynamic pattern of
changes in the gradient of mental health to air pollution (Fig. 4). Such a matching pattern of event-study plots between online
searching and mental health suggests that the information seeking plays an important role in explaining our baseline results on
mental health, which we explore further in the next section. Overall, we have validated that the information program led to a
drastic increase in the gradient of active information seeking to air pollution.
As a robustness check, we show that the increase in information seeking is only observed in pollution-related topics and not
in other domains of commonly searched keywords, such as culture, art, entertainment, pension, and financial market (Appendix
Table A31). This rules out the case in which unobserved policies or social events may coincide with the roll-out schedule for the
information program and drive up overall search activities on the internet.
5.2. Stress response to pollution information
The increased information seeking shows that individuals have dedicated more attention to information related to air pollu-
tion (Chun et al., 2011; Gottlieb, 2012). This surge in searching activities may lead to an accumulation of information shocks. We
conjecture that under heightened attention, frequent news of severe air pollution may represent a source of constant stressors that
overload the stress response system, which eventually leads to a deterioration of mental health.
47 A potential concern is that not everyone has access to the internet, and therefore are not able to search online for pollution information. We conduct a
robustness check in Appendix Table A28 and show that the results are robust when we standardize the search volume by the number of households with internet
access instead of the city population.
19
Fig. 5. Event study estimates of the effect of information program on the gradient of search volume to air pollution. Notes: This figure plots event study estimates
of the change in the online pollution gradient of search volume before and after the information program. The dependent variable, the online search index, is
the daily online search volume for pollution-related keywords per 10,000 population. Panel A plots the total search index for all pollution-related keywords,
which is the sum of three categories of indices: the search index for major air pollutants (Panel B), the search index for the harms and health effects of air
pollution (Panel C), and the search index for defensive measures against air pollution (Panel D). Regression results are reported in Appendix Table A32. See
Appendix Table A2 for the list of keywords included in each category of indices.
The concept of allostatic load provides insight into our conjecture. Allostatic load refers to the cost of chronic exposure to elevated
or fluctuating stress responses resulting from repeated events that the individual experiences as stressful (McEwen, 2017). Recent
studies in psychology and neuroscience have found that exposure to air pollution triggers the human body to produce a series of
endocrine and neural responses – such as producing the stress hormone – to facilitate the fight-or-flight reaction (Miller et al., 2016;
Thomson, 2019; Braithwaite et al., 2019), even when individuals are not aware of their exposure to air pollution (Li et al., 2017).
Better informed, more attentive individuals, who are aware of the health harms of air pollution, are more likely to experience these
stress responses when air pollution worsens. Although the body will naturally ‘‘cool down’’ after a single episode of stress response,
frequent stress responses can cause constant wear and tear on the physical and mental system, and result in higher risk of physical
and mental illness, including chronic stress and depression (McEwen, 2012, 2000; Juster et al., 2010). Appendix Figure B18 depicts
such a scenario, in which constant stressors may overwhelm the stress response system, increase the allostatic load, and raise the
risk of mental illnesses (Juster et al., 2010).
In our setting, both air pollution and the information of air pollution represent potential stressors. Because the information
program has significantly increased the awareness of and attention to air pollution, individuals have become more sensitive to news
on air pollution and more vulnerable to these information stressors (Maier et al., 2015). Although we lack data on biomarker-based
measures of stress responses, we provide four pieces of evidence that pollution information may act as stressors that increases the
gradient of mental health to air pollution.
Evidence of stress responses First, we define an indicator of constant stress if the respondent reported feeling stressed constantly over
the past month and estimate the baseline specification. Table 3, Column (3) shows that the information program has statistically
significantly increased the gradient of constant stress to air pollution. Event study plot in Appendix Figure B19 suggests that such
increase is on impact and remains persistent.
20
Table 5
Role of pollution information in the relationship between air pollution and mental health.
(1) (2) (3) (4) (5) (6)
Panel A: Effect of information disclosure by different changes in search intensity
Variables K6 K6 K6 K6 K6 K6
Search category Air pollutants Harmful effects Defensive devices
Below mean Above mean Below mean Above mean Below mean Above mean
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) 0.172*** −0.007 0.060* 0.024 −0.022 0.161***
(0.053) (0.031) (0.031) (0.035) (0.034) (0.036)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 0.094 0.143*** 0.056 0.118** 0.096** 0.164***
(0.059) (0.049) (0.050) (0.046) (0.047) (0.046)
Observations 38,169 54,362 38,311 54,256 36,685 55,861
Panel B: Effect of information disclosure by different window of air pollution exposure
Variable K6 K6 K6 K6 K6 K6
Window of pollution 1-month 2-month 3-month 4-month 5-month 6-month
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) 0.047** 0.024* 0.029** 0.058*** 0.104*** 0.226***
(0.022) (0.014) (0.014) (0.017) (0.021) (0.043)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 0.083*** 0.150*** 0.222*** 0.275*** 0.227*** 0.320***
(0.030) (0.042) (0.050) (0.059) (0.057) (0.077)
Observations 93,091 93,091 93,091 93,091 91,769 87,951
Panel C: Correlation between search for pollution information and measures of mental health
Variables K6 Mild Moderate Severe Constant stress Confidence in
Depression Depression depression future
Search index 0.666*** 2.158*** 0.935** 0.374* 0.816*** −0.632**
(0.168) (0.559) (0.371) (0.220) (0.212) (0.282)
Observations 31,831 31,831 31,831 31,831 31,831 31,831
Notes: Panel A presents the heterogeneous effects of information disclosure based on the extent of change in search intensity. The sample is divided into two
groups by whether the 1-year change in search volume is above or below the mean. Columns 1–2 divide the sample by the change in search volume for major
air pollutants; Columns 3–4 by the change in search volume for health effects of air pollution; and Columns 5–6 by the change in search volume of defensive
measures against air pollution. Panel B presents estimated effects of information disclosure on the gradient of mental health to air pollution across different
windows of air pollution exposure. Columns 1–6 present the results when the concentration of air pollution is measured by 1-month, 2-month, . . . , 6-month
average prior to the interview date, respectively. Panel C presents the estimated correlation search for air-pollution information and various measures of mental
health in sample after the information program. Search index is the log monthly search volume per 10,000 city population for top pollution-related keywords.
Other regression covariates and fixed effects are specified the same as in Table 2. Standard errors in parentheses are clustered at the household level. ***p <
0.01, **p < 0.05, *p < 0.1.
Second, we expect that individuals living in cities that exhibit larger increases in searching activities for pollution information,
thus receiving more information stressors, would experience a greater increase in the gradient of mental health to air pollution. We
divide cities into two groups based on the change in search intensity for air pollution 1 year after the information program. Table 5,
Panel A, Columns 1–2 show that individuals living in cities with a larger increase in search intensity experienced a significantly
larger increase in the gradient of mental health. This pattern is consistent over different types of search indices (Panel A Columns
3–6).
Third, we estimate the cumulative effects of air pollution by measuring the average level of air pollution over longer term prior
to the survey date. We expect an increasing cumulative effect of air pollution on mental health before the program, and that the
information program would further amplify such cumulative effect. This is because awareness of being exposed to a longer period of
severe pollution would engender more frequent stress responses and further overload the mental system. Table 5, Panel B presents
the estimation results when we compute the average level of air pollution in 1-month to 6-month before the survey date. Results
show that the estimates of both 𝛽 and 𝛾 get larger if the average concentration of air pollution is measured over a longer time
window (from 1 month to 6 months). This shows that cumulative effects of air pollution are larger than contemporaneous effect
and the awareness of such long-term exposure to severe air pollution would further amplify the adverse effects of air pollution on
mental health.
Lastly, we verify that the amount of pollution information received, measured by the search intensity for pollution information,
strongly correlates with worse mental health. We restrict the sample to observations after the information program and estimate
𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝛽𝐼𝑛𝑑𝑒𝑥𝑗𝑡 + 𝑋𝑖𝑗𝑡 𝛾 + 𝑊𝑗𝑡 𝜙 + 𝐶𝑗𝑡 𝜇 + 𝜆𝑖 + 𝜂𝑗 + 𝛿𝑡 + 𝜋𝑝𝑡 + 𝜀𝑖𝑗𝑡 . All covariates are specified the same as in Eq. (2). Table 5, Panel
C shows that an increase in the search intensity for pollution-related information strongly predicts more stress and worse mental
status, measured by a string of mental health indices such as K6 and dummies for constant stress and depression. We conduct a
robustness check by replacing the aggregate search index with the search index for each keyword (Appendix Tables A33 and A34),
and the results are consistent.
Overall, these results suggest that information on severe air pollution can be an important source of stress and plays an important
role in explaining the greater gradient of mental health to air pollution after the information program.
21
Table 6
Effect of information on the gradient of avoidance behaviors to air pollution.
(1) (2) (3) (4) (5)
Panel A: Effect on the gradient of social activities and outdoor exercise
Variables CFPS CHARLS
Outdoor Visiting fam- Caring for Visiting Playing
exercise per ily/friends family/friends sport/dance cards/chess
week living away club with friends
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) 0.066 0.129 0.115 0.032 0.100
(0.112) (0.154) (0.114) (0.089) (0.098)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 −0.806*** −0.076** −0.059** −0.038** −0.059***
(0.171) (0.032) (0.024) (0.017) (0.019)
Observations 93,091 25,988 25,988 25,988 25,988
Mean Dep. Var. 2.231 0.339 0.113 0.087 0.185
Panel B: Effect on the gradient of expenditure on transportation and outdoor activities
Variables CFPS CHARLS
Automobile Automobile Eating out & Training & Furniture &
& Transport & Fuels & Clothes shopping Sports & Durables
Transport Fitness shopping
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) 0.117 0.459 −0.262 0.660 0.544
(0.192) (0.297) (0.454) (0.494) (0.474)
𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) × 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 −0.891*** −0.191*** −0.177** −0.341*** −0.553***
(0.285) (0.062) (0.087) (0.103) (0.099)
Observations 93,091 25,988 25,988 25,988 25,988
Mean Dep. Var. 435.577 442.658 182.449 128.972 217.547
Notes: This table presents the estimated effects of information disclosure on the gradient of outdoor activities to air pollution (Panel A) and gradient of expenditure
on outdoor activities to air pollution (Panel B). Column (1) is based on the baseline CFPS sample; Columns (2) to (5) are based on the CHARLS sample. Both
samples are restricted to observations with nonmissing values of the dependent variable. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) is the log of the average concentration of PM2.5 in the
month prior to the respondent’s interview date in the respondent’s county of residence. 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 is the dummy variable that equals one if real-time air-pollution
information is publicly available in the county. 𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛) and its interaction with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒 are instrumented by log of thermal inversions and its interaction
with 𝐷𝑖𝑠𝑐𝑙𝑜𝑠𝑢𝑟𝑒. Dependent variables are log frequency of outdoor exercises per week in Column (1) and indicators of social activities in Columns (2) to (5) in
Panel (A), and are in log terms in Panel (B). Regression models are specified in Eq. (2). All regressions control for individual fixed effects, county fixed effects,
interviewer fixed effects, a quadratic function in month-year, and province-by-year fixed effects. Control variables are specified the same as in Table 2. Standard
errors in parentheses are clustered at the household level. ***P < 0.01, **P < 0.05, *P < 0.1.
5.3. Avoidance behaviors to pollution information
We now show that individuals’ avoidance behaviors can be another potential channel that leads to the greater gradient of mental
health to air pollution.
An extensive literature shows that reliable and easily accessible pollution information may prompt individuals to adopt various
forms of avoidance and defensive behaviors to reduce exposure to pollution (see, for example, Barwick et al. (2019), Neidell (2009),
Qin and Zhu (2018) and Ito and Zhang (2020)). Low-cost and effective solutions include reducing outdoor activities. However, the
psychology literature has presented substantive evidence that outdoor exercise represents a key element in preserving good mental
health (Doré et al., 2016; Chan et al., 2019; Kleppang et al., 2018; Doré et al., 2018; Yasunaga et al., 2018); social ties and social
integration maintained through social activities also strongly predict a lower risk of stress and mental health disorders (Seeman,
1996; Lecerof et al., 2015; Hammer, 2000; Kaplan et al., 2012). Therefore, the information-induced pollution avoidance behaviors,
such as reducing social integration and outdoor activities, may serve as a contributing factor to explain why the disclosure of
pollution information would magnify individual’s gradient of mental health to air pollution.
Evidence of avoidance responses We show that the information program amplifies the gradient of outdoor activities to air pollution.
Table 6, Panel A shows that the information program significantly decreased the gradient of frequency of outdoor exercises and
social activities to air pollution,48 including outdoor exercise (Column 1); visiting family and friends in the past month (Column
2); caring for a family member or friend who was living far away (Column 3); going to a gym, dance club, or community center
(Column 4); and playing cards or chess with friends (Column 5). Event study plot in Appendix Figure B20 further shows that the
gradient of outdoor exercises to air pollution dropped immediately after the information program and remained low for several
quarters afterwards.
Results are consistent based on outdoor-activity-related expenditure. Table 6, Panel B shows that the information program
significantly reduced the gradient of outdoor-activity-related expenditure to air pollution, including the log expenditure on
automobiles and transportation (Columns 1 and 2); eating out and shopping for clothes (Column 3); training, sports, and fitness
48 Because the CFPS survey provides limited scope in measuring respondents’ outdoor activities and social activities, we turn to CHARLS data for additional
evidence (Table 6, Columns 2 to 5).
22
(Column 4); and durable goods such as furniture (Column 5). Event study plot in Appendix Figure B21 shows that the gradient of
outdoor expenditure to air pollution decreased immediately after the program and remained low quarters afterwards. This pattern
is consistent with Barwick et al. (2019), who use detailed credit card records and find a large reduction in the gradient of outdoor
shopping frequencies to air pollution after the information program and the reduced gradient remained low for several quarters
afterwards.
Overall, Table 6 shows that pollution information prompts individuals to reduce their daily routines of outdoor exercise and
social activities when air pollution worsens. This pattern of a larger gradient of avoidance behaviors to air pollution is consistent
with the larger gradient of online searching for avoidance and defensive measures in Fig. 5, Panel D.
The fact that pollution information induces more avoidance behaviors has been consistently documented in prior studies, such
as Barwick et al. (2019) and Ito and Zhang (2020). Yet, few studies have considered that these avoidance behaviors and defensive
spending are associated with potential trade-offs in mental health in terms of disrupted outdoor activities, reduced social integration,
and greater financial distress. Considering the substantial benefits of physical exercise and social ties on mental health (Seeman,
1996; Kleppang et al., 2018), the disruption of daily routines and social engagement may contribute to the larger gradient of mental
health to air pollution after the information program. In addition, many defensive equipment are costly, such as the purchase of
PM-proof face masks and air purifiers (Barwick et al., 2019; Ito and Zhang, 2020). For example, as noted by Ito and Zhang (2020),
average price of regular air purifiers was $455 in 2010s, which accounts for about 25% of per-capita annual income in our sample.
Air purifiers with capacity to remove particulate matters – those with high-efficiency particulate arrestance (HEPA) filters – are $139
more expensive than regular air purifiers, and cost $21 more when replacing a filter. The larger gradient of defensive spending to
air pollution may give rise to additional financial stress and contribute to the greater mental gradient to air pollution.
Nevertheless, it is noteworthy is that the greater gradient of avoidance behaviors to air pollution does not necessarily lead to a
reduction in outdoor activities and social ties. The change in overall frequency of outdoor activities depends on the change in levels
of air pollution. When air quality is good or improving, the greater gradient of avoidance behaviors implies a greater increase in
outdoor activities and social integration as people become better informed of the environmental amenity.
6. Welfare discussion
So far we have shown that the disclosure of air-pollution information magnifies the gradient of mental health to changes in air
pollution levels. However, this does not imply that the information disclosure always leads to welfare loss. The welfare effect of
the information disclosure depends on the direction of trends in air quality. In situations under which the environmental quality is
continuously deteriorating, a public disclosure of such information may create additional mental health burdens on the population;
in contrary, when the environmental quality is improving, the public disclosure may lead to an improvement in mental health.
Considering the substantial improvement of air quality in recent years (Fig. 1), we expect a notable welfare gain after the information
program. In this section, we conduct two analyses to quantify the welfare gain.
Back-of-envelope calculation We first adopt a simple back-of-envelope calculation to compute the economic burden saved if the
level of annual PM2.5 is reduced from the sample average (66.8 μg∕m3 ) to the national target level of 34.5 μg∕m3 in 2021 when
the population is fully aware of such reduction.
We have estimated that after the information program, a 100% reduction in PM2.5 concentrations results in an additional 0.078-
percentage-point decrease in the probability of severe depression. Bringing PM2.5 concentrations down from our sample average
to the 2021 target level would represent a 48% reduction, which translates into a 3.8-percentage-point reduction in the rate of
severe depression. Based on 1.16 billion adult population (aged 16 and above) from the 2010 Census, we estimate that reaching the
2021 target of pollution control would result in a 44.1 million reduction in the population with severe depressive symptoms. Hu
et al. (2007) estimate that the annual direct medical cost (formal hospital care including inpatient and outpatient care) of severe
depression is CNY 8090 in China in 2002 (11,488 in 2016 CPI-adjusted CNY, or $1641 in 2016 USD). Lu et al. (2021) find that only
9.5% of depression patients would seek medical treatment in China (the rate is higher for patients with more severe symptoms).
Combining these estimates, we compute the direct medical cost saved by reducing the level of PM2.5 from our sample average to
the 2021 national target level to be $6.82 billion per year (in 2002 dollars).49
We conduct three additional calculations: If the level of PM2.5 is reduced by 10 μg∕m3 from the sample average, the estimated
saving of direct medical cost would be $2.11 billion per year; if the level of PM2.5 is reduced from the sample average to the WHO
recommended level (at 10 μg∕m3 ), the estimated saving would amount to $12 billion per year; when the level of PM2.5 was reduced
from the 2013 level (62.9 μg∕m3 , the first year of the information program) to the 2016 level (47.9 μg∕m3 , the end of our sample),
the estimated saving amounted to $3.36 billion per year.
Two caveats are worth mentioning for this simple back-of-envelope calculation. We may overestimate the reduction in the
depression rate if the depression-reducing effect of pollution regulations diminishes as air quality continues to improve. On the other
hand, we have only considered the direct saving of formal hospital spending on depression, and have not taken into account savings
on medications and informal care, improved labor productivity from better mental health, and savings from reduced probability of
suicide due to depression.
49 Total adult population (1.16 billion) × effect of pollution reduction on depression rate (0.078) × reduction in pollution level ((66.8–34.5)/66.8) × share of
patients with depression who seek treatment (9.5%) × annual direct medical cost ($1641 per year) = $6.82 billion per year.
23
Stated-preference-based estimate of WTP We also conduct a calculation of WTP for clean air based on the stated-preference approach
of Levinson (2012) and Zhang et al. (2017). This approach regresses individual’s mental health on the measure of air pollution and
income, and computes WTP as the marginal rate of substitution between air quality and income to maintain the same level of mental
health. See Appendix Table A35 and Appendix Section A.3 for details of regression results and the calculation of WTP. However,
because of the potential endogeneity of income measures and missing information on defensive expenditure against air pollution,
we exercise caution in interpreting this estimate of WTP.50
With the caveat in mind, we find that individuals were willing to pay an additional CNY 530, or 4.44% of annual income,
to compensate for a 1 μg∕m3 increase in PM2.5 per year per person after the information program, compared to the baseline
pre-program WTP of CNY 300 (2.51% of annual income). This implies that individuals are willing to pay 1.8 times more for
cleaner air when they have accurate, real-time information on air quality. In comparison, Ito and Zhang (2020) find that the
information program has led to a 1.45- to 1.77-fold increase in WTP for clean air, based on a revealed-preference approach with air
purifier transaction data. Overall, our estimated increase in WTP for clean air suggests that Chinese households were willing to pay
substantially more for an improvement in air quality after they are fully informed of local pollution levels and the health effects of
pollution exposure. This finding adds to the recent literature on the value of pollution information in China (Ito and Zhang, 2020;
Barwick et al., 2019), and shows that further enforcement of air-pollution regulations in China would generate substantial welfare
gains after information on air quality becomes publicly available and easily accessible.
7. Conclusion
This paper is among the first to assess the mental health effects of information on environmental risks. This question is of
particular policy relevance as countries around the world are increasingly concerned with climate change and environmental
safety and are adopting public information disclosure programs to promote public supervision and engagement in environmental
protection. It is thus vital to evaluate the effects of the public information disclosure program on all aspects of social well-beings,
especially for developing countries with mounting environmental threats.
We focus on a pivotal policy change in China that resulted in the comprehensive monitoring and provision of previously
unavailable, real-time air quality information to the public. Based on matched data on respondents’ K6 from nationally representative
surveys and air quality measures from satellite imaging, we find robust evidence that access to real-time information on air pollution
may significantly increase the gradient of mental health to the change in air pollution, especially for the elderly population. This
may impose an unintended mental health burden and an additional loss of social welfare if the air quality continues to deteriorate.
On the other hand, if air quality improves, the information disclosure policy would generate welfare gains. A back-of-envelope
calculation suggests that if the level of PM2.5 in China was to reduce from the sample average to the WHO recommended level, the
estimated medical saving on depression alone would amount to $12 billion per year.
Although our empirical analysis focuses on a national information program in China, our findings have broader policy
implications, especially for developing countries with large and rising environmental threats. Information disclosure policies are
increasingly being proposed and implemented by governments in developing countries,51 yet the mental health effects of these
policies are overlooked and understudied. Evidence of a substantial mental health response to information on air pollution in China
raises questions regarding the impact of environmental information disclosure on overall welfare consequences in other developing
countries. While fully acknowledging the value of environmental information in facilitating avoidance and mitigating harms to
physical health, our paper highlights the important role of mental health responses to environmental information in the design and
delivery of environmental information policies in developing countries.
Our findings also have broader implications for the potential mental health effects of public information disclosure during public
crisis, such as during a pandemic. Recent studies show that exposure to frequent news on COVID-19 infection and death rates could
engender adverse mental health outcomes such as increasing the risk of anxiety, depression, and a host of behavioral problems,
especially among the elderly and adolescents (Su et al., 2021; Holman et al., 2020). The increased use of social media and exposure
to pandemic news during the peak of the pandemic have been found to worsen individuals’ anxiety and depression among college
students in China (Gao et al., 2020; Li et al., 2020). In a New England Journal of Medicine article, Pfefferbaum and North (2020) argue
that news of mounting death rates and looming severe shortages of medical resources contribute to widespread emotional distress
and increased risk for psychiatric illness associated with COVID-19. They suggest that contact with pandemic-related news should be
monitored and limited, especially for children. Our study provides supporting evidence that frequent, negative information during
a pandemic can lead to increased risk of mental illnesses when individuals are mentally sensitive to information shocks. Reduced
exposure to ‘‘negative’’ news and limited use of social media among students and adolescents may mitigate such mental health
effects of information during a public crisis like COVID-19.
Appendix A. Supplementary data
Supplementary material related to this article can be found online at https://doi.org/10.1016/j.jeem.2023.102827. In the sup-
plementary material, we collect the analyses, discussions, figures, and tables omitted from the main text of the paper ‘‘Information,
Awareness, and Mental Health: Evidence from Air Pollution Disclosure in China.’’
50 See Levinson (2012) for discussion of the advantages and disadvantages of the stated-preference approach compared with other methods for computing
WTP for air quality.

51 For example, the Performance Evaluation and Ratings Program (PERP) in Indonesia, Green Rating Project (GRP) in India, and EcoWatch Program in the
Philippines.
24
References
Altonji, J.G., Elder, T.E., Taber, C.R., 2015. Selection on observed and unobserved variables: Assessing the effectiveness of catholic schools. J. Polit. Econ. 113
(1), 115–184.
Andrews, S.Q., 2008. Inconsistencies in air quality metrics:‘blue sky’ days and PM10 concentrations in Beijing. Environ. Res. Lett. 3 (3), 034009.
Arceo, E., Hanna, R., Oliva, P., 2016. Does the effect of pollution on infant mortality differ between developing and developed countries? Evidence from Mexico
city. Econ. J. 126 (591), 257–280.
Barwick, P.J., Li, S., Lin, L., Zou, E., 2019. From Fog to Smog: The Value of Pollution Information. NBER Working Paper (w26541).
Bennear, L.S., Olmstead, S.M., 2008. The impacts of the ‘‘right to know’’: Information disclosure and the violation of drinking water standards. J. Environ. Econ.
Manag. 56 (2), 117–130.
Berry, H.L., Bowen, K., Kjellstrom, T., 2010. Climate change and mental health: a causal pathways framework. Int. J. Public Health 55 (2), 123–132.
Borgschulte, M., Molitor, D., Zou, E., 2018. Air Pollution and the Labor Market: Evidence from Wildfire Smoke. Technical report, Working Paper.
Borusyak, K., Jaravel, X., 2017. Revisiting event study designs. Available at SSRN 2826228.
Braghieri, L., Levy, R., Makarin, A., 2022. Social media and mental health. Amer. Econ. Rev. 112 (11), 3660–3693.
Braithwaite, I., Zhang, S., Kirkbride, J.B., Osborn, D.P., Hayes, J.F., 2019. Air pollution (particulate matter) exposure and associations with depression, anxiety,
bipolar, psychosis and suicide risk: a systematic review and meta-analysis. Environ. Health Perspect. 127 (12), 126002.
Brody, D.J., Pratt, L.A., Hughes, J.P., 2018. Prevalence of Depression Among Adults Aged 20 and Over: United States, 2013–2016. Working Paper, US Department
of Health and Human Services, Centers for Disease Control.
Buchard, V., Da Silva, A., Randles, C., Colarco, P., Ferrare, R., Hair, J., Hostetler, C., Tackett, J., Winker, D., 2016. Evaluation of the surface PM2.5 in version
1 of the NASA MERRA aerosol reanalysis over the United States. Atmos. Environ. 125, 100–111.
Buchard, V., Randles, C., Da Silva, A., Darmenov, A., Colarco, P., Govindaraju, R., Ferrare, R., Hair, J., Beyersdorf, A., Ziemba, L., et al., 2017. The MERRA-2
aerosol reanalysis, 1980 onward. Part II: Evaluation and case studies. J. Clim. 30 (17), 6851–6872.
Callaway, B., Sant’Anna, P.H., 2021. Difference-in-differences with multiple time periods. J. Econometrics 225 (2), 200–230.
Carrà, G., Sciarini, P., Segagni-Lusignani, G., Clerici, M., Montomoli, C., Kessler, R., 2011. Do they actually work across borders? Evaluation of two measures of
psychological distress as screening instruments in a non Anglo-Saxon country. Eur. Psychiatry 26 (2), 122–127.
Chan, J.S., Liu, G., Liang, D., Deng, K., Wu, J., Yan, J.H., 2019. Special issue–therapeutic benefits of physical activity for mood: a systematic review on the
effects of exercise intensity, duration, and modality. J. Psychol. 153 (1), 102–125.
Chay, K.Y., Greenstone, M., 2005. Does air quality matter? Evidence from the housing market. J. Polit. Econ. 113 (2), 376–424.
Chen, S., Chen, Y., Lei, Z., Tan-Soo, J.-S., 2020. Impact of air pollution on short-term movements: evidence from air travels in China. J. Econ. Geogr. 20 (4),
939–968.
Chen, Y., Jin, G.Z., Kumar, N., Shi, G., 2012. Gaming in air pollution data? Lessons from China. BE J. Econ. Anal. Policy 13 (3).
Chen, S., Oliva, P., Zhang, P., 2018b. Air Pollution and Mental Health: Evidence from China. Technical report, National Bureau of Economic Research.
Chen, S., Oliva, P., Zhang, P., 2022. The effect of air pollution on migration: evidence from China. J. Dev. Econ. 156, 102833.
Chun, M.M., Golomb, J.D., Turk-Browne, N.B., 2011. A taxonomy of external and internal attention. Annu. Rev. Psychol. 62, 73–101.
Cunningham, M.R., 1979. Weather, mood, and helping behavior: Quasi experiments with the sunshine samaritan. J. Personal. Soc. Psychol. 37 (11), 1947.
De Chaisemartin, C., d’Haultfoeuille, X., 2020. Two-way fixed effects estimators with heterogeneous treatment effects. Amer. Econ. Rev. 110 (9), 2964–2996.
De Chaisemartin, C., D’Haultfoeuille, X., 2022. Two-Way Fixed Effects and Differences-In-Differences with Heterogeneous Treatment Effects: a Survey. Technical
report, National Bureau of Economic Research.
Deng, Z., Axbard, S., 2021. Informed Enforcement: Lessons from Pollution Monitoring in China. NHH Dept. of Economics Discussion Paper (01).
Deschenes, O., Greenstone, M., Shapiro, J.S., 2017. Defensive investments and the demand for air quality: Evidence from the NOx budget program. Amer. Econ.
Rev. 107 (10), 2958–2989.
Deschenes, O., Wang, H., Wang, S., Zhang, P., 2020. The effect of air pollution on body weight and obesity: Evidence from China. J. Dev. Econ. 145, 102461.
Donati, D., Durante, R., Sobbrio, F., Zejcirovic, D., 2022. Lost in the net? Broadband internet and youth mental health.
Doré, I., O’Loughlin, J.L., Beauchamp, G., Martineau, M., Fournier, L., 2016. Volume and social context of physical activity in association with mental health,
anxiety and depression among youth. Prev. Med. 91, 344–350.
Doré, I., O’loughlin, J.L., Schnitzer, M.E., Datta, G.D., Fournier, L., 2018. The longitudinal association between the context of physical activity and mental health
in early adulthood. Ment. Health Phys. Act. 14, 121–130.
Forouzanfar, M.H., Afshin, A., Alexander, L.T., Anderson, H.R., Bhutta, Z.A., Biryukov, S., Brauer, M., Burnett, R., Cercy, K., Charlson, F.J., et al., 2016. Global,
regional, and national comparative risk assessment of 79 behavioural, environmental and occupational, and metabolic risks or clusters of risks, 1990–2015:
a systematic analysis for the global burden of disease study 2015. Lancet 388 (10053), 1659–1724.
Freeman, R., Liang, W., Song, R., Timmins, C., 2019. Willingness to pay for clean air in China. J. Environ. Econ. Manag. 94, 188–216.
Fu, S., Viard, V.B., Zhang, P., 2021. Air pollution and manufacturing firm productivity: Nationwide estimates for China. Econ. J. 131 (640), 3241–3273.
Furukawa, T.A., Kessler, R.C., Slade, T., Andrews, G., 2003. The performance of the K6 and K10 screening scales for psychological distress in the Australian
national survey of mental health and well-being. Psychol. Med. 33 (2), 357–362.
Gao, J., Zheng, P., Jia, Y., Chen, H., Mao, Y., Chen, S., Wang, Y., Fu, H., Dai, J., 2020. Mental health problems and social media exposure during COVID-19
outbreak. Plos One 15 (4), e0231924.
Ghanem, D., Shen, S., Zhang, J., 2020. A censored maximum likelihood approach to quantifying manipulation in China’s air pollution data. J. Assoc. Environ.
Resour. Econ. 7 (5), 965–1003.
Ghanem, D., Zhang, J., 2014. ‘Effortless perfection:’ do Chinese cities manipulate air pollution data? J. Environ. Econ. Manag. 68 (2), 203–225.
Golberstein, E., Gonzales, G., Meara, E., 2019. How do economic downturns affect the mental health of children? evidence from the national health interview
survey. Health Econ. 28 (8), 955–970.
Goodman-Bacon, A., 2021. Difference-in-differences with variation in treatment timing. J. Econometrics 225 (2), 254–277.
Goodman-Bacon, A., Schmidt, L., 2020. Federalizing benefits: the introduction of supplemental security income and the size of the safety net. J. Public Econ.
185, 104174.
Gotlib, I.H., Hammen, C.L., 2008. Handbook of Depression. Guilford Press.
Gottlieb, J., 2012. Attention, learning, and the value of information. Neuron 76 (2), 281–295.
Graff Zivin, J., Neidell, M., 2012. The impact of pollution on worker productivity. Amer. Econ. Rev. 102 (7), 3652–3673.
Greenstone, M., He, G., Jia, R., Liu, T., 2022. Can technology solve the principal-agent problem? Evidence from China’s war on air pollution. Am. Econ. Rev.
Insights 4 (1), 54–70.
Greenstone, M., He, G., Li, S., Zou, E., 2021. China’s War on Pollution: Evidence from the First Five Years. Technical report, National Bureau of Economic
Research.
Greenstone, M., Jack, B.K., 2015. Envirodevonomics: A research agenda for an emerging field. J. Econ. Lit. 53 (1), 5–42.
Guan, D., Su, X., Zhang, Q., Peters, G.P., Liu, Z., Lei, Y., He, K., 2014. The socioeconomic drivers of China’s primary PM2. 5 emissions. Environ. Res. Lett. 9
(2), 024010.
25
Hammer, T., 2000. Mental health and social exclusion among unemployed youth in scandinavia. a comparative study. Int. J. Soc. Welf. 9 (1), 53–63.
Hanna, R., Oliva, P., 2015. The effect of pollution on labor supply: Evidence from a natural experiment in Mexico city. J. Public Econ. 122, 68–79.
He, J., Liu, H., Salvo, A., 2019. Severe air pollution and labor productivity: Evidence from industrial towns in China. Am. Econ. J. Appl. Econ. 11 (1), 173–201.
Hicks, D., Marsh, P., Oliva, P., 2016. Air Pollution and Procyclical Mortality: Causal Evidence from Thermal Inversions. NBER Working Paper.
Holman, E.A., Thompson, R.R., Garfin, D.R., Silver, R.C., 2020. The unfolding COVID-19 pandemic: A probability-based, nationally representative study of mental
health in the United States. Sci. Adv. 6 (42), eabd5390.
Hu, T.-w., He, Y., Zhang, M., Chen, N., 2007. Economic costs of depression in China. Soc. Psychiatry Psychiatr. Epidemiol. 42 (2), 110–116.
Ito, K., Zhang, S., 2020. Willingness to pay for clean air: Evidence from air purifier markets in China. J. Polit. Econ. 128 (5), 1627–1672.
Jakiela, P., 2021. Simple Diagnostics for Two-Way Fixed Effects. Technical report, arXiv preprint arXiv:2103.13229.
Jans, J., Johansson, P., Nilsson, J.P., 2018. Economic status, air quality, and child health: Evidence from inversion episodes. J. Health Econ. 61, 220–232.
Juster, R.-P., McEwen, B.S., Lupien, S.J., 2010. Allostatic load biomarkers of chronic stress and impact on health and cognition. Neurosci. Biobehav. Rev. 35
(1), 2–16.
Kaplan, K., Salzer, M.S., Brusilovskiy, E., 2012. Community participation as a predictor of recovery-oriented outcomes among emerging and mature adults with
mental illnesses. Psychiatr. Rehabil. J. 35 (3), 219.
Kessler, R.C., Andrews, G., Colpe, L.J., Hiripi, E., D K, M., Normand, S.T., Walters, E.E., Zaslavsky, A.M., 2002. Short screening scales to monitor population
prevalences and trends in non-specific psychological distress. Psychol. Med. 32 (6), 959.
Kessler, R.C., Barker, et al., 2003. Screening for serious mental illness in the general population. Arch. Gen. Psychiatry 60 (2), 184–189.
Kessler, R.C., Green, J.G., Gruber, M.J., Sampson, N.A., Bromet, E., Cuitan, M., Furukawa, T.A., Gureje, O., Hinkov, H., Hu, C.-Y., et al., 2010. Screening for
serious mental illness in the general population with the K6 screening scale: results from the WHO world mental health (WMH) survey initiative. Int. J.
Methods Psychiatr. Res. 19 (S1), 4–22.
Kleppang, A.L., Hartz, I., Thurston, M., Hagquist, C., 2018. The association between physical activity and symptoms of depression in different contexts–a
cross-sectional study of norwegian adolescents. BMC Public Health 18 (1), 1–12.
Ko, N.-Y., Lu, W.-H., Chen, Y.-L., Li, D.-J., Wang, P.-W., Hsu, S.-T., Chen, C.-C., Lin, Y.-H., Chang, Y.-P., Yen, C.-F., 2020. COVID-19-related information sources
and psychological well-being: An online survey study in Taiwan. Brain Behav. Immun. 87, 153.
Lecerof, S.S., Stafström, M., Westerling, R., Östergren, P.-O., 2015. Does social capital protect mental health among migrants in Sweden? Health Promot. Int. 31
(3), 644–652.
Lee, S., Tsang, A., Ng, K.L., Ma, Y.L., Guo, W., Mak, A., Kwok, K., 2012. Performance of the 6-item Kessler scale for measuring serious mental illness in Hong
Kong. Compr. Psychiatry 53 (5), 584–592.
Levinson, A., 2012. Valuing public goods using happiness data: The case of air quality. J. Public Econ. 96 (9–10), 869–880.
Li, H., Cai, J., Chen, R., Zhao, Z., Ying, Z., Wang, L., Chen, J., Hao, K., Kinney, P.L., Chen, H., et al., 2017. Particulate matter exposure and stress hormone
levels: a randomized, double-blind, crossover trial of air purification. Circulation 136 (7), 618–627.
Li, J., Yang, Z., Qiu, H., Wang, Y., Jian, L., Ji, J., Li, K., 2020. Anxiety and depression among general population in China at the peak of the COVID-19 epidemic.
World Psychiatry.
Liang, X., Li, S., Zhang, S., Huang, H., Chen, S.X., 2016. PM2. 5 data reliability, consistency, and air quality assessment in five Chinese cities. J. Geophys. Res.:
Atmos. 121 (17), 10–220.
Lu, J., Xu, X., Huang, Y., Li, T., Ma, C., Xu, G., Yin, H., Xu, X., Ma, Y., Wang, L., et al., 2021. Prevalence of depressive disorders and treatment in China: a
cross-sectional epidemiological study. Lancet Psychiatry 8 (11), 981–990.
Maier, C., Laumer, S., Eckhardt, A., 2015. Information technology as daily stressor: Pinning down the causes of burnout. J. Bus. Econ. 85, 349–387.
McEwen, B.S., 2000. Allostasis and allostatic load: implications for neuropsychopharmacology. Neuropsychopharmacology 22 (2), 108–124.
McEwen, B.S., 2012. Brain on stress: how the social environment gets under the skin. Proc. Natl. Acad. Sci. 109 (Supplement 2), 17180–17185.
McEwen, B., 2017. Stress: Homeostasis, rheostasis, reactive scope, allostasis and allostatic load. In: Reference Module in Neuroscience and Biobehavioral
Psychology. Elsevier.
Mehta, A.J., Kubzansky, L.D., Coull, B.A., Kloog, I., Koutrakis, P., Sparrow, D., Spiro, A., Vokonas, P., Schwartz, J., 2015. Associations between air pollution and
perceived stress: the veterans administration normative aging study. Environ. Health 14 (1), 1–10.
Miller, D.B., Ghio, A.J., Karoly, E.D., Bell, L.N., Snow, S.J., Madden, M.C., Soukup, J., Cascio, W.E., Gilmour, M.I., Kodavanti, U.P., 2016. Ozone exposure
increases circulating stress hormones and lipid metabolites in humans. Am. J. Respir. Crit. Care Med. 193 (12), 1382–1391.
Neidell, M., 2009. Information, avoidance behavior, and health the effect of ozone on asthma hospitalizations. J. Hum. Resour. 44 (2), 450–478.
Paul, K.I., Moser, K., 2009. Unemployment impairs mental health: Meta-analyses. J. Vocat. Behav. 74 (3), 264–282.
Pfefferbaum, B., North, C.S., 2020. Mental health and the Covid-19 pandemic. N. Engl. J. Med. 383 (6), 510–512.
Power, M.C., Kioumourtzoglou, M.-A., Hart, J.E., Okereke, O.I., Laden, F., Weisskopf, M.G., 2015. The relation between past exposure to fine particulate air
pollution and prevalent anxiety: observational cohort study. BMJ 350.
Prochaska, J.J., Sung, H.-Y., Max, W., Shi, Y., Ong, M., 2012. Validity study of the K6 scale as a measure of moderate mental distress based on mental health
treatment need and utilization. Int. J. Methods Psychiatr. Res. 21 (2), 88–97.
Pun, V.C., Manjourides, J., Suh, H., 2017. Association of ambient air pollution with depressive and anxiety symptoms in older adults: results from the NSHAP
study. Environ. Health Perspect. 125 (3), 342–348.
Qin, Y., Zhu, H., 2018. Run away? Air pollution and emigration interests in China. J. Popul. Econ. 31 (1), 235–266.
Radloff, L.S., 1977. The CES-D scale: A self-report depression scale for research in the general population. Appl. Psychol. Meas. 1 (3), 385–401.
Ren, X., Yu, S., Dong, W., Yin, P., Xu, X., Zhou, M., 2020. Burden of depression in China, 1990–2017: findings from the global burden of disease study 2017.
J. Affect. Disord. 268, 95–101.
Ross, J.A., Gliebus, G., Van Bockstaele, E.J., 2018. Stress induced neural reorganization: a conceptual framework linking depression and Alzheimer’s disease.
Prog. Neuro-Psychopharmacol. Biol. Psychiatry 85, 136–151.
Sager, L., 2019. Estimating the effect of air pollution on road safety using atmospheric temperature inversions. J. Environ. Econ. Manag. 98, 102250.
Sakurai, K., Nishi, A., Kondo, K., Yanagida, K., Kawakami, N., 2011. Screening performance of K6/K10 and other screening instruments for mood and anxiety
disorders in Japan. Psychiatry Clin. Neurosci. 65 (5), 434–441.
Seeman, T.E., 1996. Social ties and health: The benefits of social integration. Ann. Epidemiol. 6 (5), 442–451.
Su, Z., McDonnell, D., Wen, J., Kozak, M., Abbas, J., Šegalo, S., Li, X., Ahmad, J., Cheshmehzangi, A., Cai, Y., et al., 2021. Mental health consequences of
COVID-19 media coverage: the need for effective crisis communication practices. Glob. Health 17 (1), 1–8.
Sun, C., Zheng, S., Wang, J., Kahn, M.E., 2019. Does clean air increase the demand for the consumer city? Evidence from Beijing. J. Reg. Sci. 59 (3), 409–434.
Thomson, E.M., 2019. Air pollution, stress, and allostatic load: linking systemic and central nervous system impacts. J. Alzheimer’s Dis. 69 (3), 597–614.
Tu, M., Zhang, B., Xu, J., Lu, F., 2020. Mass media, information and demand for environmental quality: Evidence from the ‘‘under the dome’’. J. Dev. Econ.
143, 102402.
Twenge, J.M., Campbell, W.K., 2019. Media use is linked to lower psychological well-being: Evidence from three datasets. Psychiatr. Q. 90, 311–331.
Vaughan, L., Chen, Y., 2015. Data mining from web search queries: A comparison of google trends and baidu index. J. Assoc. Inf. Sci. Technol. 66 (1), 13–22.
Wang, A.L., 2018. Explaining environmental information disclosure in China. Ecol. Law Q. 44 (4), 865–924.
26
Wang, G., Zhang, R., Gomez, M.E., Yang, L., Zamora, M.L., Hu, M., Lin, Y., Peng, J., Guo, S., Meng, J., et al., 2016. Persistent sulfate formation from London
fog to Chinese haze. Proc. Natl. Acad. Sci. 113 (48), 13630–13635.
Weissman, J., Pratt, L.A., Miller, E.A., Parker, J.D., 2015. Serious Psychological Distress Among Adults, United States, 2009–2013, Vol. 203. US Department of
Health and Human Services.
World Health Organization, 2005. WHO air quality guidelines for particulate matter. In: Ozone, Nitrogen Dioxide and Sulfur Dioxide: Global Update.
Xue, T., Zhu, T., Zheng, Y., Zhang, Q., 2019. Declines in mental health associated with air pollution and temperature variability in China. Nature Commun. 10
(1), 2165.
Yasunaga, A., Shibata, A., Ishii, K., Koohsari, M.J., Oka, K., 2018. Cross-sectional associations of sedentary behaviour and physical activity on depression in
Japanese older adults: an isotemporal substitution approach. BMJ Open 8 (9), e022282.
Zhang, X., Chen, X., Zhang, X., 2018. The impact of exposure to air pollution on cognitive performance. Proc. Natl. Acad. Sci. 115 (37), 9193–9197.
Zhang, J., Wang, C., 2011. Co-benefits and additionality of the clean development mechanism: An empirical analysis. J. Environ. Econ. Manag. 62 (2), 140–154.
Zhang, X., Zhang, X., Chen, X., 2017. Happiness in the air: How does a dirty sky affect mental health and subjective well-being? J. Environ. Econ. Manag. 85,
81–94.
27

Information, Awareness, and Menta Lhealth - Evidence From Air Pollution Disclosure in China - 2023

Uploaded by

Copyright:

Available Formats

You might also like

Information, Awareness, and Menta Lhealth - Evidence From Air Pollution Disclosure in China - 2023

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Information, Awareness, and Menta Lhealth - Evidence From Air Pollution Disclosure in China - 2023

Uploaded by

Copyright:

Available Formats

Journal of Environmental Economics and Management 120 (2023) 102827

Contents lists available at ScienceDirect

Journal of Environmental Economics and

Information, awareness, and mental health: Evidence from air

ARTICLE INFO ABSTRACT

gains from improved mental health.

2.1. Information and awareness of air pollution prior to 2012

2.2. The information program

2.3. Rising awareness of air pollution

3. Data and methodology

3.1. Data and variables

3.2. Sample construction and summary statistics

Deschenes et al. (2020), Fu et al. (2021), and Chen et al. (2022).

Alternative radii are adopted and the results remain robust.

b Household annual income per household member.

3.3. Empirical methodology

𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝜃𝐷𝑗𝑡 + 𝑋𝑖𝑗𝑡 𝜖 + 𝑊𝑗𝑡 𝜙 + 𝐶𝑗𝑡 𝜇 + 𝜆𝑖 + 𝜂𝑗 + 𝛿𝑡 + 𝜋𝑝𝑡 + 𝜀𝑖𝑗𝑡 , (1)

𝐾6𝑖𝑗𝑡 = 𝛼0 + 𝛽𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 + 𝛾𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑗𝑡 × 𝐷𝑗𝑡 + 𝛼1 𝐷𝑗𝑡 (2)

4.1. Baseline results

Observations 93,091 93,091 93,091 93,091 93,091

Observations 93,091 93,091 93,091 93,301 92,695 92,695

term. Longer-term effects, however, could be quite different.

4.2. Robustness analyses

prior knowledge of local pollution levels.

4.3. Heterogeneous effects

before running the pooled regression.

5.1. Awareness and active search for pollution information

𝐼𝑛𝑑𝑒𝑥𝑐𝑡 = 𝛼0 + 𝛽𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑐𝑡 + 𝛾𝐿𝑜𝑔(𝑝𝑜𝑙𝑙𝑢𝑡𝑖𝑜𝑛)𝑐𝑡 × 𝐷𝑐𝑡 + 𝛼1 𝐷𝑐𝑡 (4)

Observations 15,309 15,309 15,309 15,309

5.2. Stress response to pollution information

5.3. Avoidance behaviors to pollution information

evidence (Table 6, Columns 2 to 5).

Appendix A. Supplementary data

WTP for air quality.

You might also like