Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Willingness to Pay for Clean Air:
Evidence from Air Purifier Markets in China
Koichiro Ito
University of Chicago and NBER
Shuang Zhang
⇤
University of Colorado Boulder
This version: March 8, 2016
Preliminary draft – PLEASE DO NOT CITE OR CIRCULATE
Abstract
This paper provides among the first revealed preference estimate of willingness to pay (WTP)
for clean air in developing countries. We use product-by-store level transaction data on air
purifier sales in Chinese cities and city-level air pollution data. Our empirical strategy leverages
the Huai River heating policy, which created discontinuous quasi-experimental variation in air
pollution between the north and south of the river. Using a spatial regression discontinuity
design, we estimate the marginal willingness to pay for removing 1 ug/m3 PM10. Our findings
provide important policy implications for optimal environmental regulation.
⇤Ito: Harris School of Public Policy, University of Chicago, and NBER (e-mail: [email protected]). Zhang: De-partment of Economics, University of Colorado Boulder (e-mail: [email protected]). For helpful comments,we thank Douglas Almond, Richard Freeman, Rema Hanna, Shanjun Li, Matt Neidell, and seminar participants atHarvard, MIT, Cornell and Colorado Boulder.
1 Introduction
Air quality is remarkably worse in developing countries, and severe air pollution causes substantial
health and economic burdens for billions of people. For example, the annual average exposure to
fine particle pollution (PM2.5) in China is six times higher than that in the United States in 2010.1
Such high levels of air pollution cause large negative impacts on a variety of economic outcomes,
including infant mortality (Jayachandran, 2009; Arceo et al., 2012; Greenstone and Hanna, 2014),
life expectancy (Chen et al., 2013) and labor supply (Hanna and Oliva, 2015). Therefore, air
pollution is one of the first-order problems for economic development for many countries.
However, high health and economic burdens of air pollution do not necessarily imply that
existing environmental regulations are not optimal. The optimal environmental regulation depends
on the extent to which individuals value air quality improvements—that is, willingness to pay
(WTP) for clean air (Greenstone and Jack, 2013). If WTP for clean air is low, the current level
of air pollution can be optimal because the social planner would prioritize economic growth over
environmental regulation. On the other hand, if WTP is high, the current stringency of regulation
can be away from the optimum. Therefore, WTP for clean air is a key parameter for economists and
policymakers when considering tradeo↵s between economics growth and environmental regulation.
Despite the importance of this question, the economics literature provides limited empirical evidence
because obtaining a revealed preference estimate of WTP for clean air is particularly hard in
developing countries due to the limited availability of comprehensive data.
In this paper, we provide among the first revealed preference estimates of WTP for clean air in
developing countries. Our idea is that demand for home-use air purifiers, a main defensive invest-
ment for reducing indoor air pollution, provides valuable information for estimating a lower bound
of WTP for air quality improvements. We begin by developing a random utility model in which
consumers purchase air purifiers to reduce indoor air pollution. A key advantage of analyzing air
purifier markets is that one of the product attributes informs consumers and econometricians about
the purifier’s ability to reduce air pollution. We apply this framework to unique transaction data in
air purifier markets in 81 Chinese cities. For each retail store in the 81 cities, we observe product-
level information on monthly sales, monthly average price, and detailed product characteristics.
1http://data.worldbank.org/indicator/EN.ATM.PM25.MC.M3/countries?display=default
1
Our data cover January 2006 through December 2012.
Our empirical strategy is a spatial regression discontinuity (RD) design, which exploits discon-
tinuous valuation in air pollution created by a policy experiment at the Huai River boundary. The
so-called Huai River policy provided city-wide coal-based heating for cities north to the river, which
generated higher pollution levels in the north (Almond et al., 2009; Chen et al., 2013). The attrac-
tiveness of this spatial RD approach is twofolds. First, it allows us to exploit plausibly exogenous
variation in air pollution created by the policy. Second, the discontinuous di↵erence in air pollution
created by the policy has existed for a long time since the 1950s, which is particularly useful for
studying the demand for durable goods like air purifiers.
Furthermore, we use a measure of transportation costs as a supply shifter as an instrument for
price. For each product, we measure the distance from the city it sells to its manufacturing plant
if it is domestically produced, or to the port if it is imported.
We present visual and statistical evidence that, in winter months, 1) the PM10 level is signifi-
cantly higher in the north of the Huai River, and 2) there is a discontinuous and substantial increase
in the market share of air purifiers that remove particular matters just north to the river. Using
estimates from the RD design, we find that the local average of MWTP for removing 1 ug/m3 of
PM10 is 2 dollars. Our RD estimates are robust to using di↵erent functional forms of latitude and
di↵erent bandwidths from the river line.
Our study provides two primary contributions to the literature. The first contribution is that
we develop a framework to estimate WTP for improvements in environmental quality by estimating
demand for defensive investment. Earlier studies on avoidance behavior against pollution examine
whether individuals take avoidance behavior in response to pollution exposure.2 A key question
in recent studies is whether researchers can estimate WTP for improvements in environmental
quality from observing defensive investment in markets. To our knowledge, two recent papers ask
this question. Kremer et al. (2011) uses a randomized control trial (RCT) for water pollution in
Kenya. Our approach, a quasi-experimental experiment with non-experimental data, is closer to
the approach by Deschenes et al. (2012), in which they use medical expenditure data in the United
Sates at the individual level. There is no doubt that RCT would be the ideal empirical strategy
2For evidence in the United States, see Neidell (2009); Zivin and Neidell (2009); Zivin et al. (2011). For evidencein China, see Mu and Zhang (2014); Zheng et al. (2015). For evidence in other developing countries, see Madajewiczet al. (2007); Jalan and Somanathan (2008).
2
to answer the question. However, a large-scale RCT on pollution is not always feasible in many
countries. Therefore, quasi-experimental approaches are also important complements to address
this question. We believe that our quasi-experimental framework can be useful for other contexts
because our method relies on market-level sales and price data, which are more likely to exist in
most countries because manufacturers and retail stores usually collect scanner data on product
sales and price.3
The second contribution is that our analysis provides empirical evidence for an important“miss-
ing piece” in the literature on air pollution in developing countries, which o↵ers important policy
implications. Many recent studies show that severe air pollution in developing counties cause large
negative impacts on a variety of economic outcomes, including infant mortality (Jayachandran,
2009; Arceo et al., 2012; Greenstone and Hanna, 2014), life expectancy (Chen et al., 2013) and
labor supply (Hanna and Oliva, 2015). However, as emphasized by Greenstone and Jack (2013),
there is little evidence on revealed preference estimates of WTP for clean air. Our estimates provide
a lower bound of MWTP for improvements in air quality, which is a key parameter for policymakers
when considering tradeo↵s between economic growth and environmental regulation. Specifically, in
theory, more stringent environmental regulations can be justified if the marginal cost of regulation
is below our MWTP estimate.
2 Air pollution, Air Purifiers and the Huai River Policy in China
2.1 The main pollutant in Chinese cities
Northern and eastern China has perhaps the most polluted cities in the world according to NASA’s
global map on PM2.5 (particulate matter with diameter of 2.5 micrometers or less).4 Among
ambient pollution measures, PM2.5 has shown most consistently an adverse e↵ect on human health
(Dockery et al. (1993), Pope et al. (2009) and Correia et al. (2013)).
The main pollutant in Chinese cities is particulate matter. The Chinese Ministry of Environ-
3There are a few more related studies. Berry et al. (2012); Miller and Mobarak (2013) use randomized controlledtrials to estimate WTP for water filters and cook stoves per se instead of WTP for improvements in environmentalquality. Consumer behavior in housing markets is usually not considered to be “avoidance behavior”, but Chay andGreenstone (2005) is related to our study in the sense that they provide a quasi-experimental approach to estimateWTP for clean air.
4http://www.nasa.gov/topics/earth/features/health-sapping.html
3
mental Protection (MEP) releases an daily air pollution index (API) in 120 cities since 2000. In
each city, a number of monitors take hourly concentration measures of three air pollutants: PM10
(particulate matter with diameter of 10 micrometers or less), SO2 and NO2. Daily API is converted
from one of these pollutants that has the highest daily average value. During 2006-2012, among
the 78% of days when the MEP reported the specific type of pollutant from which API was taken
from, 91.2% were from PM10, 8.7% from SO2 and 0.1% from NO2.
The o�cial API, mostly based on ambient PM10, is the only accessible pollution information for
Chinese citizens during the time period of this study.5 Both daily API level and the main pollutant
type are reported to local residents by city weather channel, radio and newspapers.
2.2 Air purifiers
Among already known air purification technologies, the High E�ciency Particulate Arrestance
(HEPA) filter is most e↵ective against particulate matter. According to the US Department of
Energy, a HEPA filter must remove (from the air that passes through) at least 99.97% of particles
in 0.3 micrometer in diameter (DOE (2005)). It is more e↵ective for particles which are larger, for
example, PM10 and PM2.5. Recent clinical studies find that the uses of HEPA purifiers in various
settings are associated with improvements in health, including reduced asthma visits and asthma
symptoms among children, lower levels of markers for inflammation and heart disease and reduced
incidences of invasive aspergillosis among adults (Abdul Salam et al. (2010); Allen et al. (2011);
Lanphear et al. (2011)).
On the Chinese market, other Non-HEPA purification technologies are designed for removing
other target pollutants, not particulate matter. Activated carbon absorbs volatile organic com-
pounds (VOCs), but it does not remove particles. A catalytic converter is e↵ective in removing
VOCs and formaldehyde. An air ionizer generates electrically charged air or gas ions, which attach
to airborne particles that are then attracted to a charged collector plate. However, there are no
specific standards for air ionizers, and they also produce ozone and other oxidants as by-products.
A study by Health Canada finds that residential ionizer only removes 4% of indoor PM2.5 (Wallace,
2008).
5The Chinese government started to report PM2.5 in 2014. We focus on 2006-2012 because of the availability andrepresentativeness of our air purifier data in this time period.
4
An air purifier can use multiple air purification technologies, which must pass quality testing
by a certified testing center before entering the market. Air purifier companies are required to
make their product testing results available. We compiled testing certificates for most products
in our sample from company websites. Air purifiers with a HEPA filter report testing results on
the percentage of indoor PM2.5 it removes, and Non-HEPA air purifiers that use activated carbon
report testing results on the e↵ectiveness in removing VOCs. Testings on PM2.5 are not conducted
if a product does not have a HEPA filter. Air purifiers that combine a HEPA filter with activated
carbon to remove both particulate matter and VOC report testing results on both.
Market advertisements of air purifiers with a HEPA filter typically state that they can remove
more than 99% of PM2.5.
2.3 The Huai River policy and its recent reform
China has adopted the Soviet-era centralized heating system since 1958. Due to budget constraint,
the Chinese government decided to provide city-wide centralized heating to cities in North China
only (Almond et al. (2009)). North and South China are divided by the line formed by the Huai
River and Qinling Mountains. This line was used to divide the country for heating policy because
the average January temperature is roughly 0° Celsius along the line, and it is not a border used
for administrative purposes (Chen et al. (2013)). Cities north to the line have received unlimited
heating in winter every year. In contrast, cities south to the line have been denied centralized
heating supply from the government. The heating policy is therefore called the Huai River policy.
The heating supply to the north has been considered a public welfare entitlement, and it remains
the same today.
The centralized heating supply in the north relies on coal-fired heating systems. Two-thirds
of heat is generated by heat-only hot water boilers for one or several buildings in an apartment
complex, and the other one-thid combined heat and power generators for larger areas of a city. This
system is inflexible and energy ine�cient. Consumers have no means to control their heat supply,
and there is no measurement of heat consumption in apartments or even at the building level. The
incomplete combustion of coal in the heat generation process leads to the release of air pollutants,
especially particulate matter. Because most of heat is generated by boilers within an apartment
complex, the pollution from coal-based heating largely remains local. Almond et al. (2009) finds
5
that the Huai River policy led to higher total suspended particulates (TSP) levels in the north.
Chen et al. (2013) further finds that the higher pollution levels created by the policy led to a loss
of 5.5 years of life expectancies in the north.
The heating supply to the north has been consistent since the 1950s, while the payment system
under the policy had an important change in 2003. Prior to 2003, free heating was provided for
residents in the north, and employers or local governments were responsible to pay for household
heat bills (WorldBank (2005)). It was designed under the centrally planned economy, in which
public sector employment dominated the labor market. However, during China’s transition to a
market economy, heat billing became a practical problem. The size of private sector has increased
dramatically since the 1990s, and employers in private sector were not regulated to pay heat bills
for their employees. Further, many public sector employees have moved out of public housing and
purchased home in the private market, which made it di�cult for employers to pay their heat bills
in private homes.
In July 2003, the Chinese government issued a heat reform in northern cities.6 The reform
changed the payment system from free provision to consumer-based billing (WorldBank (2005)).
Individual households became responsible for paying their own heat bills each season, which is a
flat rate per square meter of floor area (without changes in the metering method).7 Whether heat
subsidy is provided by employers depends on the sector. In public sector, former in-kind transfer
were changed to a transparent payment for heat added to the wage. In contrast, private sector
employers were not explicitly required to provide heat subsidy to their employees. In the 2005 mini
census, 21% of labor force was in urban public sector in the 81 cities in our sample, suggesting
that only a small percentage of employees get heat subsidy after the reform. Anecdotally, from an
online survey on heat bills among more than 800 individuals in Qingdao (a northern city) in 2012,
78% do not receive any heat subsidy, 12% think that the amount of heat subsidy is very small, and
only 10% think that they get a reasonable subsidy for heat.8
6http://news.xinhuanet.com/zhengfu/2003-07/29/content_998737.htm
7The Chinese government has also been interested in introducing heat metering and consumption-based billingto improve the e�ciency of heat utilization. To learn more about the progress of the reform, we recently had aconversation with the energy team of the World Bank about their collaboration with the Chinese government onthe reform. According to their experiences in China, it has been di�cult to switch to energy-saving constructiontechnologies on a large scale. Little substantial progress along this dimension has been made so far. Thus, theconsumer-based billing at a flat rate is not equivalent to consumption-based billing.
8http://news.bandao.cn/news_html/201212/20121224/news_20121224_2048038.shtml
6
Our analysis focuses on 2006-2012, after the 2003 reform. We summarize our comparison on
winter heating between the north and the south since the reform. First, the way winter heating has
been provided remains the same after the reform. The centralized city-wide heating supply in the
north remains the same, where households have little option other than the coal-based heating that
generates higher pollution levels. In the south, households choose their own ways to stay warm in
winter, including using air conditioners, space heaters, heated blankets, and etc. Second, heating
cost in the north has changed since the 2003 reform. Northern households no longer enjoy free
heating and instead have to pay a substantial proportion of their heat bills from the centralized
heating, while households in the south remain to pay for heating methods of their choice. According
to our comprehensive search of heating costs in 20 cities within 3 degrees of latitude relative to the
Huai River boundary, household heating costs in the north are comparable to, or could even be
higher than, those in the south.9
3 Data and Descriptive Statistics
We compile comprehensive data from several sources. Our dataset is unique because it integrates air
pollution data with detailed market transaction data at the product-city-year-month level, which
are usually hard to obtain in developing countries.
Air purifier data
We obtain monthly market transaction data of air purifiers in 81 cities in 2006-2012 from a top
marketing consulting firm in China. In each city, the transaction data are collected every month
from a network of major department stores and electrical appliance stores, which take up on average
more than 80% of all in-store sales in a city. During 2006-2012, in-store sales consist of on average
over 95% of overall sales including in-store and online sales.10 The original sales and price data are
9For example, in Xi’an, a city within 1 degree of latitude north to the Huai River, the price of heating per squaremeter per winter is 3.9USD. For an apartment of 100 square meters, the household pays 390USD. The average subsidyin public sector is 177USD per employee, and the number of public employee per household is 0.32 in the 2005 minicensus. The average amount of subsidy per household is 57USD. Therefore, an average household’s out-of-pocketpayment is 333USD. In southern cities, space heater and heated blankets are the most common choices that couldcost 150-200USD including purchasing these devices and the electricity bill in winter for a similar size of home. If ahousehold choose a more expensive option, air conditioning, the electricity bill for three months in winter could bearound 240-280USD, and their entire cost depends on the price of the air conditioners that varies to a large extent.
10The marketing consulting firm provides detailed product-level data on in-store sales only. The share of onlinesales has increased dramatically since 2013, while the share of in-store sales has decreased. Data from the firm suggest
7
at the product-city-store-year-month level. Because our empirical strategy using the Huai River
policy relies on cross-sectional variation, we aggregate the transaction data to the product-city level,
for example, the total number of sales of a product during 2006-2012 by city, the average price in
2006-2012 by product-city, etc. There are 395 di↵erent products of 30 brands in our sample.
A unique feature of the data is that the type of filtration system that a product uses is included.
Importantly, we observe whether a product uses a High E�ciency Particulate Arrestance (HEPA)
filter, which enables us to quantify the amount of pollution reduction that a product brings. As
discussed in Section 2.2, the HEPA filter is most e↵ective against particulate matter. Based on the
US standard for HEPA filters, air purifier advertisements in Chinese cities highlight that the HEPA
filter can remove more than 99% of PM2.5. Combining the binary measure of having a HEPA filter
with pollution levels, we measure the amount of pollution reduction as follows. Assuming that
consumers are fully informed, the amount of pollution reduction that consumers think they will get
from a HEPA purifier is equal to the PM10 level, and 0 if they get a Non-HEPA purifier.
We also observe other useful product attributes, for example, the maximum coverage area for
two-thirds of all products in our sample, which is used to estimate the MWTP per square meter.
Pollution data
The o�cial air pollution index (API) is the only accessible pollution information for Chinese citizens
during the time period of this study. We obtain daily API data for 81 cities in our sample in 2006-
2012 from Chinese Ministry of Environmental Protection (MEP). In addition to the API level, the
specific type of pollutant from which API was taken from is also disclosed to the public. During
2006-2012, among the 78% of days when the MEP reported the type of pollutant on which API was
based, 91.2% were from PM10. The conversion from the concentration of each pollutant to API
is non-linear. For days that PM10 is reported as the main pollutant, we use the o�cial formula
used by Chinese MEP to convert daily API to daily PM10. To match with the product-city level
market transaction data, we aggregate the average PM10 in winter months and non-winter months
respectively, at the city level.
We are cautious in using the API data because recent studies find evidence on underreporting
that online sales increased to 36% in 2013 and to 47% in 2014. Therefore, to use the data most representative of allsales, we focus our analysis on 2006-2012.
8
of API at the margin of 100 (Chen et al. (2012), Ghanem and Zhang (2014)). The manipulation is
motivated by the blue-sky award, which defines a day with API below 100 as a blue sky day and link
the number of blue sky days in a year to the annual performance evaluation of city governments.
For our analysis using the city-level pollution measure, we are concerned about the extent to which
the manipulation a↵ects the average level of API.
To investigate potential manipulation in our sample, we perform McCrary density tests (Mc-
Crary (2008)) on daily API data for each of these 81 cities in 2006-2012. We find no statistically
significant discontinuity in the density of daily API at 100 in 75% of city-years in Figure A.2a,
while there is statistically significant discontinuity at 100 in 25% of city-years in Figure A.2b. To
examine to what extent the manipulation changes the average API in the 25% of city-years, we
use the distribution of API in the 75% non-manipulation city-years to plot a counterfactual dis-
tribution in the manipulation subsample. Figure A.2c shows the original distribution of API in
the 25% subsample, where the mean of API is 147.90. Figure A.2d shows the the counterfactual
distributions of API, where the mean of API is 147.95. The similarity of these two mean values of
API reduces the concern about the manipulation for our analysis.
Measure of transportation costs
We are concerned that the price of air purifiers is likely correlated with unobserved demand shocks,
particularly product-city level unobserved factors that generated cross-city variation within product.
We use a measure of transportation costs as a supply shifter as an instrument for price. For each
product, we geo-coded the location of its manufacturing plant if it is domestically produced, or the
location of the port if it is imported. Around 16% of products are imported. We then calculate
the distance from the city a product sells to its manufacturing plant for domestically produced
products, or to the port for imported products. We use this product-city level distance measure to
capture transportation costs and use it as the instrument for price.
Descriptive statistics
Table 1 reports summary statistics for key variables in our sample. In the air purifier data, 52%
of products have a HEPA filter. We report summary statistics for all products in column (1),
and for HEPA purifiers in column (2) and Non-HEPA purifiers in column (3). On average there
9
are 463 sales per product during 2006-2012, and HEPA purifiers have more sales than Non-HEPA
purifiers. HEPA purifiers are also more expensive than Non-HEPA purifiers. On average a HEPA
purifier costs $461 USD, while a Non-HEPA purifier costs $347. Becuase a purifier could have
multiple filters, we also report the other three main filteration systems: activated carbon, ionizer
and catalytic converter. These Non-HEPA filters are mostly used by Non-HEPA purifiers. On
the additional function as a humidifier, HEPA purifiers and Non-HEPA purifiers are very similar.
HEPA purifiers have slightly larger room coverage (44 square meters) than Non-HEPA purifiers (39
square meters). Finally, HEPA and Non-HEPA purifiers have similar transportation costs in terms
of the distance to factory or port.
Turning to city-level pollution measures, the average PM10 in winter months is 119 ug/m3,
while it is 95ug/m3 in non-winter months.
Figure 1 shows the location of the 81 cities on the China map in our analysis. The line of Huai
River/Qinling Mountains divides China into its North and South. Each dot represents a city in our
sample. All cities in our sample are located east to 100 degree of longitude. The river line east to
100 degree of longitude ranges between 32.6 and 34.2 degree of latitude. In our spatial RD approach
using the Huai River policy, we define a city’s relative latitude north to the river line. Because
the river line has several di↵erent curved segments, we divide the river line to five segments. In
each segment, we measure a city’s relative latitude to the middle point of the river latitude range.
For example, Beijing locates at 39.9 degree of latitude and 116.3 degree of longitude, and the
corresponding middle point of the river latitude range is 33.4 degree. Beijing’s relative latitude
north to the river line is 6.5 (39.9-33.4) degree. Cities in our sample locate between -12.9 and 14.8
degree north to the river line.
Figure A.1 adds locations of manufacturing plants of domestically produced products and ports
of imported products on the map. Most manufacturing plants and ports are located on the east
coast.
4 A Random Utility Model for Air Purifier Demand
Our goal is to obtain a revealed preference estimate of WTP for indoor air quality improvements
by analyzing demand for air purifiers. In a standard demand model for di↵erentiated products,
10
a consumer purchases an air purifier by considering utility from its product characteristics and
disutility from its price. An advantage of analyzing air purifier markets is that one of the product
characteristics called high-e�ciency particulate arrestance (HEPA) indicates the purifier’s ability
to reduce indoor particulate matter. The extent to which consumers value this characteristic, along
with the price elasticity of demand, provides useful information for their WTP for indoor air quality
improvements.
We begin with a standard random utility model similar to Berry (1994), Nevo (2001), and
Kremer et al. (2011) to model demand for air purifiers. Consider that consumer i in city c has
average ambient air pollution zc (particular matters). The consumer can purchase air purifier j at
price pcj to reduce indoor air pollution by �zcj = zc · aj . We describe purifier j’s ability to reduce
indoor particulate matters by aj 2 [0, 1]. We observe markets for c = 1, ..., C cities with i = 1, ..., Ic
consumers. In our estimation below, a market is defined as a city. The conditional indirect utility
of consumer i from purchasing air purifier j at market c is:
uijc = ��zjc + ↵pjc + ✓j + ⇠jc + ✏ijc, (1)
where �zcj is the pollution reduction, pjc is the price of product j in market c, ✓j is a utility gain
from unobserved and observed product characteristics for product j, ⇠jc is a product-city specific
demand shock, and ✏ijc is a mean-zero stochastic term. Assuming the error term ✏ijc has an extreme
value density function, the market share for product j in city c is:
sjc =exp(��zjc + ↵pjc + ✓j + ⇠jc)PJ
j0=0 exp(��zj0c + ↵pj0c + ✓j0 + ⇠j0c). (2)
For j = 1, ..., J , we observe product-level sales data in city c, from which we construct the
market share (sjc). The outside option (j = 0) is not to buy any purifier. We make a few assump-
tions to construct the market share for outside options (s0c). We define s0c following Berry (1994);
Nevo (2001). Assuming that the number of households in city c are potential buyers and that each
household purchases one or zero air purifier, s0c can be calculated by the di↵erence of the number
of households in city c and the total number of sales in city c . Then, we assume that �zc0 = 0 and
pc0 = 0. That is, the outside option is free and does not reduce indoor air pollution.11 We then con-
11This assumption could be violated if consumers can reduce indoor air pollution by purchasing some goods other
11
struct the market share for outside options by ln s0c = 0� ln�P
exp(��zj0c + ↵pj0c + ✓j0 + ⇠j0c)�.
The di↵erence between log market share for product j and log market share for outside options can
be described by,
lnsjc � lns0c = ��zjc + ↵pjc + ✓j + ⇠jc, (3)
where � is the marginal utility from having one unit of pollution reductions and ↵ is the marginal
disutility from paying one more dollar for the purchase. Therefore, marginal willingness to pay
(MWTP) for one unit of indoor air pollution reductions can be obtained by ��/↵.
Notably, when we include city fixed e↵ects in our prefered estimation below, the outside option
does not matter to the estimation because the term (lns0c) will be absorbed by city fixed e↵ects.
The estimation is, therefore, una↵ected by the assumption on outside options.
We interpret that our estimate of ��/↵ provides a lower bound of MWTP for one unit of
indoor pollution reductions. Our approach assumes that indoor air pollution levels equal to ambient
pollution levels (zc). A recent engineering study shows that indoor pollution levels are lower than
outdoor pollution levels in Beijing.12 If people understand that indoor air pollution level is lower
than outdoor pollution level, our approach underestimates �. An alternative approach would be to
rely on an engineering estimate of indoor-outdoor air pollution ratio, which would produce slightly
larger values for the MWTP. However, we want to be conservative about our estimate as much as
possible, and therefore, our estimation mostly focuses on estimating the lower bound of MWTP.
An advantage of studying air purifier markets is that aj (purifier j’s ability to reduce indoor
particulate matters) is straightforward, and consumers are well informed about it. As we explained
in Section 2.2, mechanically, if a purifier has a High E�ciency Particulate Arrestance (HEPA)
filter, it reduces 99% of indoor particular matters. On the other hand, if a purifier does not have a
HEPA filter, it is not e↵ective in reducing indoor particular matters. In advertisements and product
descriptions for both types of products on the Chinese market, consumers are well informed about
the di↵erence between HEPA purifiers and non-HEPA purifiers. Therefore, we define the pollution
than air purifiers. To our knowledge, we are not aware of other major indoor devices for indoor pollution reducationin China.
12A study from Tsinghua University finds that, in Beijing, the indoor concentration of PM2.5 is 67% of the outdoorconcentration of PM2.5. See http://news.tsinghua.edu.cn/publish/news/4204/2015/20150423100046963966000/20150423100046963966000_.html
12
reduction by:
�zcj = zc ·HEPAj =
8>><
>>:
zc if HEPAj = 1
0 if HEPAj = 0.
(4)
However, it could be possible that some consumers do not fully understand the di↵erence and
think that Non-HEPA purifiers could also reduce particular matters to some extent. We test this
possibility empirically by including zc in our baseline estimation below (without including city fixed
e↵ect). If consumers do not respond to the level of particular matters in purchasing Non-HEPA
purifiers, we would expect that the estimate of zc is 0. Again, in our preferred specification, zc is
absorbed by city fixed e↵ects.
Our measure of zc is PM10, which is converted from API using the o�cial formula. The estimate
��/↵ indicates the MWTP for one unit of reduction in PM10.
Furthermore, for two-thirds of air purifiers in our sample, we also observe each purifier’s max-
imum coverage of square meters. It is possible that customers value an air purifier that cov-
ers larger square meter. To test this possibility, our second approach defines �zcj by �zcj =
zct ·HEPAj ·Roomsizej . Using this definition, we estimate the MWTP for one unit of reduction
in PM10 per square meter.
5 Empirical Strategy and Results
The goal of our empirical analysis is to obtain consistent estimates of ��/↵ in equation (3), which
is MWTP for indoor air quality improvements. Our main empirical tasks include identifying both
� and ↵, both of which should not be correlated with omitted variables.
To address the first challenge, identifying �, we take advantage of a natural experiment at the
spatial border of the Huai river, which is described in section 2.3. This approach is appealing for
the following reasons. First, it allows us to exploit plausibly exogenous variation in air pollution
created by the natural experiment to estimate local average estimates of � in cities close to the
Huai River boundary. If people value air quality, our demand model in section 4 predicts that the
market shares for HEPA purifiers are higher in the north. Our RD design tests this hypothesis.
Second, the discontinuous di↵erence in air pollution created by the Huai River policy exists for
a long period of time (since the 1950s), which is particularly useful for studying the demand for
13
durable goods like air purifiers.
The second concern is on identifying ↵, where the price of air purifiers is likely correlated
with unobserved demand shocks. Our first step is including product fixed e↵ects, which absorbed
product-level omitted factors. However, we are still concernd about product-city level unobserved
factors that generated cross-city variation within product. We take a further step to use a measure
of transportation costs as a supply shifter as an instrument for price. For each product, we measure
the distance from the city it sells to its manufacturing plant if it is domestically produced, or to
the port if it is imported.
Baseline specification
We start with the baseline specification, where city fixed e↵ects are not included.
Baseline reduced form:
The reduced form regression estimates a discontinuous change in the market share of air purifiers
at the border of the Huai river. The baseline reduced-form regression is,
lnsjc�lns0c = �Northc⇤HEPAj+�Northc+f(Latc)+f(Latc)⇤HEPAj+↵pjc+Xc�+Longc+✓j+✏jc
(5)
where for city c, we define a latitude north to the Huai River boundary by Latc and a dummy
variable for cities north to the Huai River by Northc = 1 {Latc � 0} . The amount of pollution
reduction using HEPA air purifiers is measured by the interaction term Northc ⇤HEPAj . f(.) is
a smooth control function for latitude. We allow the function of latitude to di↵er between HEPA
purifiers and Non-HEPA purifiers by including f(Latc) ⇤ HEPAj . The price of air purifiers is
denoted by pjc. Our coe�cients of interests are � and ↵. The identify ↵, we instrument price pjc
with the distance from a product’s selling city to its manufacturing plant or port. The WTP to
remove the amount of pollution generated by the Huai River policy is measured by ��/↵. We
are also interested in �, which captures whether household response to higher pollution levels in
purchasing Non-HEPA purifiers.
14
We include product FE, ✓j , to absorb product characteristics. Furthermore, Lee and Lemieux
(2010) note that geographical discontinuity designs should be used with careful investigation of
omitted variables at the border. In our case, we are particularly concerned that the Huai river
border is long distanced, and therefore, observable and unobservable characteristics can be di↵erent
between cities in di↵erent parts of the Huai river. To address this concern, we compare observed
demographic characteristics of cities on either side of the Huai River and find little discrete changes
at the river boundary. We include these city demographics as covariates Xc. Finally, considering
that cities around the Huai river span from the west to the east, we narrow the scope of unobserved
di↵erences on either side of the river by including fixed e↵ects of longitude decile, Longc, where
Longc = 1, ..., 10.
Baseline second stage (2SLS)
For the second stage regression, we estimate two-stage least squares for,
lnsjc�lns0c = �PM10c⇤HEPAj+�PM10c+f(Latc)+f(Latc)⇤HEPAj+↵pjc+Xc�+Longc+✓j+✏jc
(6)
by using Northc as the instrument for PM10c, Northc⇤HEPAj as the instrument for PM10c⇤
HEPAj , and distance as the instrument for pjc. Therefore, � is identified by the discontinuous
cross-sectional variation in pollution between the north and the south of the Huai river. The
identification assumption is that the instrumentNorthc is uncorrelated with the error term given the
smooth control function f(Latc) and covariates. The MWTP to remove 1 unit of PM10 generated
by the Huai River policy is measured by ��/↵.
Final specification
We prefer a more restrictive specification by including city fixed e↵ects and use it in most of our
estimations below. Advantages of including city fixed e↵ects are twofolds. First, any unobserved
di↵erences in city characteristics along the Huai river are obsorbed by city fixed e↵ects. Second,
the estimation does not depend on assumptions on outside options because the term (lns0c) is also
absorbed by city fixed e↵ects. The two-stage least square is:
15
Second stage (2SLS) with city fixed e↵ects
lnsjc � lns0c = �PM10c ⇤HEPAj + f(Latc) ⇤HEPAj + ↵pjc + ✓j + �c + ✏jc (7)
where �c denotes city fixed e↵ects. Note that PM10c is no longer identifiable because it is
absorbed by �c. We are able to identify our coe�cients of interest, � and ↵.
Visual presentation
Because the Huai River policy provides heat in winter only, we begin by plotting the average PM10
in winter months (December-March) during 2006-2012 by a city’s latitude relative to the Huai River
boundary. Because very few cities locate in the farthest north and the farthest south, we focus on
cities located within 10 degrees of latitude to the river line.13 There are 70 cities in this range (out
of 81 cities in our sample). Figure 2a plots the average PM10 by 1.5 degree of latitude north to the
river boundary. The vertical line at 0 indicates the location of the river. Consistent with findings
on pollution in earlier studies (Almond et al. (2009), Chen et al. (2013)), in winter months, there
is a discontinuous increase in PM10 just north to the Huai River, suggesting that the coal-based
heating policy generates higher pollution levels in the north. In contrast, in Figure A.3, during
non-winter months (April-November), there is no discrete change in PM10 levels just north to the
Huai River. Therefore, we focus our analysis on winter months.
Turning to Figure 2b, it shows the market share of HEPA purifiers by 1.5 degree of latitude
north to the river line. In the south, the market share of HEPA purifiers are below 60%. A sharp
jump to over 70% appears just north to the river, and the higher market share persists in the north.
Moreover, the south and north trends are fairly smooth, suggesting that the choice of functional
form would have little impact on the estimated di↵erence of HEPA purifiers’ market share at the
river line.
These sharp increases of both PM10 and the market share of HEPA purifiers right north to the
river suggest that higher PM10 levels generated by the heating policy trigger higher demand for
HEPA purifiers.
The visual presentation also provides useful guidance in choosing the functional form of latitude
13All cities in our sample locate between -12.9 and 14.8 degree north to the river line.
16
in our estimation below. We use quadratic trend in the main specification. For comparison, we also
report results using linear trend and linear trend interacted with North, as well as cubic trend.
Estimation results
First stage:
Table 2 presents results of two first stage estimations: the first stage on PM10 in Table 2a and
the first stage on price in Table 2b.
In Table 2a, column (1) includes quadratic trend of latitude and interactions of quadratic
latitude and the HEPA dummy, and product FE, and column (2) adds demographic controls, and
column (3) further controls for longitude decile FE. Note that the preferred specification controlling
for city FE is not used because North is defined at the city-level. Column (1) and (2) show similar
estimates of a 15 units increase of PM10 just north to the Huai river, which is similar to the visual
size of the change in Figure 2a. When longitude FE are included in column (3), the magnitude
becomes smaller.
Turning to the first stage on price in Table 2b, in addition to the same specifications in Table
2a, we are also able to include city FE in column (4). We find similar estimates from all four
specifications that a 1 km increase in the distance to the manufacturing plant or the port increases
price by around $0.02 dollars. All estimates are statistically significant at the 1 percent level.
Reduced form and 2SLS:
In Table 3, Panel A presents the reduced form results. Consistent with Figure 2b, there is an
economically and statistically significant increase in HEPA purifiers’ market share in the north of
the river across all specifications. Interestingly, the North dummy is not statistically significant in
column (1)-(3), suggesting that households do not respond to higher pollution level in purchasing
Non-HEPA purifiers. Using estimates from our preferred specification (city FE controlled for) in
column (4), the reduced form estimate is 0.690/0.023=$30. The WTP to reduce the amount of air
pollution generated by the Huai river policy is 30 dollars.
Panel B reports 2SLS results. Estimates are robust to using di↵erent specifications. Again, we
find no evidence that households purchase Non-HEPA purifiers as response to higher PM10 levels.
17
Using our preferred specification in column (4), we find that the MWTP to reduce 1 unit of PM10
generated by the Huai river policy is 0.050/0.025=2 dollars.
Robustness tests
A possible concern on the preferred specification is that, high income or highly educated individuals
might prefer HEPA purifiers over Non-HEPA purifiers for non-pollution reasons, and therefore
interaction terms of income and HEPA or education and HEPA could be omitted variables. To
directly test these possibilities, in Table A.1, we include interactions of city-level GDP and the
HEPA dummy and interactions of average schooling and the HEPA dummy in all three columns.
Reduced form and 2SLS estimates are similar to our main estimates in Table 3, which reduces the
concern on di↵erential preferences on purifier types by income and education.
We also test whether our estimates in Table 3 are robust to using di↵erent functional forms of
latitude and di↵erent bandwidths from the river line. In Table 4, we use linear trend and linear
trend interacted with North in Panel A, quadratic trend in Panel B, and cubic trend in Panel
C. For Panel A and B, column (1) uses 81 cities in the full sample, column (2) 70 cities within
10 degrees to the river, column (3) 54 cities within 7 degrees, and column (4) 45 cities within 5
degrees. In narrower windows (within 7 degrees and within 5 degrees), Figure 2a and 2b do not
suggest cubic trend as appropriate. Therefore, using cubic trend in Panel C, we report estimates in
the full sample in column (1) and cities within 10 degrees in column (2). Estimates in Panel A and
B are robust to our main estimates in Table 3 that the MWTP is around 2 dollars in all windows.
In Panel C, the magnitude of the interaction of PM10 and HEPA becomes smaller. This is because
the first stage estimate on PM10 is much larger using cubic trend. In Table A.2, we report the first
stage estimates on PM10 using various functional forms and window lengths. Using cubic trend in
panel C, we find larger first stage estimates than those in Panel A and B.
Comparison of covariates and migration
A natural concern on the RD design is that if cities in the north are systematically di↵erent from
cities in the south, higher demand for HEPA purifiers could be explained by north-south di↵erences
other than the Huai River policy. Particularly, if households have higher income or education level
in cities just north to the river, these di↵erences could contribute to the observed jump in HEPA
18
purifiers’ market share. Therefore, we compare observed demographics of cities on either side of
the Huai River.
In Figure A.4, we report the mean of city-level demographic measures by 1.5 degree of latitude
relative to the Huai River (in the 10-degree latitude window). Little discrete change is observed
at the Huai River boundary on GDP per capita, population, share of industrial GDP and average
years of schooling. Some decrease of GDP per capita appear just north to the river, which are
in the opposite direction to the increase in air pollution and unlikely explain the increase in the
demand for HEPA purifiers. There is an increase in the share of GDP from manufacturing in the
first dot north to the river, but the increase does not persist in the north.
In Table A.3, we formally test whether these city-level demographics are di↵erent in the north
of the river compared to the south. For each of these covariates, we report an estimate of North
controlling for quadratic trend of latitude and an estimate also adding longitude decile FE as
controls. These di↵erences are not statistically significant, only except for GDP per capita when
longitude decile FE are not controlled for. In column (3), GDP per capita is lower in the north,
consistent with the observation from Figure A.4. However, in column (4) where longitude decile FE
are included, the estimate is statistically insignificant and the magnitude becomes much smaller.
None of these estimates controlling for longitude decile FE are statistically significant, supporting
the use of longitude decile FE as controls in one of our baseline specifications. Note that in our
preferred specification, city FE absorb any city-level di↵erences.
Another possible concern is about migration for cleaner air. If some households migrated from
the north to the south for cleaner air, our estimates on WTP from demand for air purifiers do
not take into account of these households. Although we can not completely rule out migration for
air quality, we do not find it very likely in China for the following reasons. First of all, internal
migration in China is strictly constrained by the Hukou system. The hukou, obtained at one’s
city of birth, is crucial for getting local social benefits and education opportunities, which makes
migration a more costly decision than that in countries without mobility restriction. Second, we
look into migration to the south using the 2005 census micro-data. Figure A.5 presents the number
of migrants currently residing south of the river by one’s city of origin. Migrants residing in the
south are mostly from the south, and very few of them are from the north. Particularly, residents
just north to the river are not more likely to move to the south than residents just south to the
19
river. Third, we find that northern migrants residing in the south are not among the richest in the
income distribution of their city of origin, and therefore the cost of migration for clean air is very
high relative to their income.
Interpretation
To interpret the higher demand for HEPA purifiers in the north as the consequence of the higher
pollution levels generated by the heating policy, one should exclude other potential confounders. We
find that observed demographics are not significantly di↵erent between the south and the north of
the Huai River. Another possible concern is that the Huai River boundary might a↵ect air quality
and the demand for HEPA purifiers in the north through other policies other than the heating
policy. For example, if the Huai River boundary is used for making economic policies, for example,
allocating heavy industries to the north which might also generate higher pollution level, we cannot
attribute higher pollution and higher market share of HEPA purifiers solely to the heating policy.
To incorporate such confounding policies, we have conducted a comprehensive search for policies
that are made di↵erently on either side of the Huai River. We fail to find other policies using the
Huai River to divide the country. This is consistent with the fact this line was used to divide the
country for heating policy because the average January temperature is roughly 0° Celsius along the
line, and it is not a border used for administrative purposes (Chen et al. (2013)).
Conceptually, one might still be concerned that the heating policy could lead to higher demand
for HEPA purifiers in the north through channels other than higher pollution levels. Particularly,
if the heating supply to the north has been a public welfare entitlement and subsidized heating
cost of northern households, while households in cities just south to the river (with similar winter
temperature) are responsible for paying their heating choice, northern households might have higher
disposable income because of the heating subsidy. If this is the case, the heating policy might
generate higher demand for HEPA purifiers through the subsidy channel instead of the pollution
channel. However, as we discussed in Section 2.3, the heat reform in 2003 changed the payment
system from free provision to consumer-based billing. Of critical importance is the change that
northern households have to pay a substantial proportion of their heat bills from the centralized
heating after 2003. From our comparison of heating costs in the north versus in the south in Section
2.3, household heating cost in the north could even be higher than that in the south. Therefore, in
20
our analysis during 2006-2012, heating subsidy has minimal e↵ect on household disposable income
in the north.
A final note is on the availability of HEPA purifier products between the north and south of
the river. If HEPA purifiers are more available in the north because appliance stores supply more
of them relative to Non-HEPA purifiers, what we observe in Figure 2b might reflect the di↵erence
in supply. To test this concern, in Figure A.6, we plot the fraction of HEPA purifier products (out
of all purifier products available) by 1.5 degree of latitude relative to the Huai River. We do not
observe a discontinuous jump on the supply side just north to the river.
MWTP for 1 unit pollution reduction per square meter
We propose an alternative way to define MWTP, which is the marginal willingness-to-pay for 1
unit of indoor pollution reduction per square meter, as discussed in Section 4. Table 5 reports
the 2SLS results, where the regressors of interests are Pollution*HEPA*Room size and Price. We
focus on column (4) using our preferred specification. The MWTP for 1 unit reduction in PM10
per square meter is 0.0011/0.0271=$0.04. The average maximum coverage area in our data is 41
square meters. Therefore, the MWTP for 1 unit reduction in PM10 in an average room of 41 square
meters is $1.64, close to our main findings in Table 3.
6 Policy implications
Our estimates provide a lower bound of MWTP for improvements in air quality in China. There
are two approaches to compare our estimate to estimates in the literature. One could combine
our estimate and a pollution-health estimate to infer health valuation, under assumptions that
households are aware of the relationship between PM10 and health outcomes and more precisely
the exact functional form, and compare a value of statistical life (VSL) estimate with those in
the literature. However, we prefer not to make unverifiable assumptions. We use an alternative
approach to directly compare our WTP estimate with WTP estimates in the literature. Readers
should keep the caveat in mind that, this is not an apple-to-apple comparison because existing
studies discuss WTP for di↵erent measures of improvements in environmental quality.
Among limited design-based evidence on revealed preference in developing countries, the WTP
21
for spring protection (improved water quality) is $2.96 per household in Kenya in Kremer et al.
(2011), which accounts for roughly 0.6% of annual household income.
In studies in developed countries, the magnitude of WTP for improvements in environmental
quality also varies depending on the measure of behavioral changes. On defensive expenditures,
Deschenes et al. (2012) finds that the reductions in NOx emissions decreased mean ozone con-
centrations by 6% and drug expenditures decreased by 1.9% annually. These estimates combined
with average levels of ozone and drug expenditure per capita suggest that the MWTP for one unit
reduction of ozone is $2.1 per capita per year. Using the hedonic approach, Chay and Greenstone
(2005) find that the Clean Air Act led to a decrease of TSPs by 10 ug/m3 and a rise in housing
prices by 2.5%. These estimates, combined with mean values of housing prices and TSPs, suggest
that the MWTP for one unit reduction in TSPs is $101.
We find that the lower bound of MWTP for one unit reduction in PM10 is $2 in China, suggesting
a considerable level of MWTP compared to findings above.
The estimate we provide is a key parameter for policymakers when considering tradeo↵s be-
tween economic growth and environmental regulation. China has recently declared “War Against
Pollution”,14 followed by a series of new policies. For example, $1.65 billion a year is o↵ered to
reward cities and regions that make “significant progress” in air pollution control,15 tougher fines
are enforced for polluters,16 and coal-fired power plants will be upgraded to cut pollution from
power plants by 60 percent by 2020.17 More prominently, China is going to start a national cap-
and-trade program in 2017 that will limit and put a price on greenhouse gas emissions.18 For each
environmental policy, policy-makers could compare the marginal cost to our MWTP estimate to
assess the policy. Specifically, more stringent regulations can be justified if the marginal cost of
regulation is below our MWTP estimate.
14http://news.xinhuanet.com/english/special/2014-03/05/c_133163557.htm
15http://www.nytimes.com/2014/02/14/world/asia/china-to-reward-localities-for-improving-air-quality.
html
16http://www.theguardian.com/environment/chinas-choice/2014/apr/25/china-environment-law-fines-for-pollution
17http://www.nytimes.com/2015/12/03/world/asia/china-plans-to-upgrade-coal-plants.html?_r=0
18http://www.nytimes.com/2015/09/25/world/asia/xi-jinping-china-president-obama-summit.html
22
7 Conclusion
Despite higher pollution levels and mortality rates from pollution, limited design-based evidence
shows very lowWTP for environmental quality in developing countries. For example, the VSL based
on WTP for environmental quality in Kenya is 10,000 times lower than that in US (Kremer et al.
(2011)). Does the low WTP for environmental quality imply that the current level of environmental
quality in developing countries is optimal? Or is WTP high, yet policy makers fail to express
the preference of citizens in policy making and implementation? These questions remain poorly
understood among economists and policy makers.
This paper provides new evidence on WTP for air quality from one of the most polluted coun-
tries in the world. Our estimates on the lower bound of WTP for improvements in air quality and
health valuation in China are substantially higher than previously understood for developing coun-
tries, which suggests that the current environmental regulations are not optimal and strengthening
environmental regulations will largely improve human welfare.
23
References
Abdul Salam, Z.-H., R. B. Karlin, M. L. Ling, and K. S. Yang (2010). The impact of portablehigh-e�ciency particulate air filters on the incidence of invasive aspergillosis in a large acutetertiary-care hospital. American Journal of Infection Control 38 (4), e1–e7.
Allen, R., C. Carlsten, B. Karlen, S. Leckie, S. van Eeden, S. Vedal, I. Wong, and M. Brauer (2011).An air filter intervention study of endothelial function among healthy adults in a woodsmoke-impacted community. American Journal of Respiratory and Critical Care Medicine 183, 1222–1230.
Almond, D., Y. Chen, M. Greenstone, and H. Li (2009). Winter heating or clean air? unintendedimpacts of china’s huai river policy. The American Economic Review: Papers and Proceed-ings 99 (2), 184–190.
Arceo, E., R. Hanna, and P. Oliva (2012). Does the e↵ect of pollution on infant mortality di↵erbetween developing and developed countries? evidence from mexico city. NBER Working PaperNo. 18349.
Berry, J., G. Fischer, and R. Guiteras (2012). Eliciting and utilizing willingness to pay: evidencefrom field trials in northern ghana. Unpublished manuscript .
Berry, S. T. (1994, July). Estimating discrete-choice models of product di↵erentiation. The RANDJournal of Economics 25 (2), 242–262.
Chay, K. Y. and M. Greenstone (2005, April). Does air quality matter? evidence from the housingmarket. Journal of Political Economy 113 (2), 376–424.
Chen, Y., A. Ebenstein, M. Greenstone, and H. Li (2013). Evidence on the impact of sustainedexposure to air pollution on life expectancy from china’s huai river policy. Proceedings of theNational Academy of Sciences 110 (32), 12936–12941.
Chen, Y., G. Z. Jin, N. Kumar, and G. Shi (2012). Gaming in air pollution data? lessons fromchina. The B.E. Journal of Economic Analysis & Policy 13 (3).
Correia, A., C. A. P. III, D. Dockery, Y. Wang, M. Ezzati, and F. Dominici (2013). The e↵ect ofair pollution control on life expectancy in the united states: An analysis of 545 us counties forthe period 2000 to 2007. Epidemiology 24 (1), 23–31.
Deschenes, O., M. Greenstone, and J. S. Shapiro (2012, July). Defensive investments and thedemand for air quality: Evidence from the NOx budget program and ozone reductions. NBERWorking Paper (ID 2109861).
Dockery, D., A. P. third, X. Xu, J. Spengler, J. Ware, M. Fay, B. Ferris, and F. Speizer (1993).An association between air pollution and mortality in six u.s. cities. New England Journal ofMedicine 329, 1753–1759.
DOE (2005). Specification for hepa filters used by doe contractors. DOE TECHNICAL STANDARDDOE-STD-3020-2005, U.S. Department of Energy, Washington, D.C.
Ghanem, D. and J. Zhang (2014). E↵ortless perfection: Do chinese cities manipulate air pollutiondata? Journal of Environmental Economics and Management 68, 203–225.
24
Greenstone, M. and R. Hanna (2014). Environmental regulations, air and water pollution, andinfant mortality in india. American Economic Review 104 (10), 3038–3072.
Greenstone, M. and B. K. Jack (2013, September). Envirodevonomics: A research agenda for ayoung field. NBER Working Paper (19426).
Hanna, R. and P. Oliva (2015). The e↵ect of pollution on labor supply: Evidence from a naturalexperiment in mexico city. 122, 68–79.
Jalan, J. and E. Somanathan (2008). The importance of being informed: Experimental evidenceon demand for environmental quality. Journal of Development Economics 87, 14–28.
Jayachandran, S. (2009). Air quality and early-life mortality: Evidence from indonesia’s wildfires.Journal of Human Resources 44 (4), 916–954.
Kremer, M., J. Leino, E. Miguel, and A. P. Zwane (2011). Spring cleaning: Rural water impacts,valuation, and property rights institutions. The Quarterly Journal of Economics 126 (1), 145–205.
Lanphear, B., R. Hornung, J. Khoury, K. Yolton, M. Lierl, , and A. Kalkbrenner (2011). E↵ectsof hepa air cleaners on unscheduled asthma visits and asthma symptoms for children exposed tosecondhand tobacco smoke. Pediatrics 127 (1), 93–101.
Lee, D. S. and T. Lemieux (2010). Regression discontinuity designs in economics. Journal ofEconomic Literature 48 (2), 281–355.
Madajewicz, M., A. Pfa↵, A. van Geen, J. Graziano, I. Hussein, H. Momotaj, R. Sylvi, and H. Ahsan(2007). Can information alone change behavior? response to arsenic contamination of ground-water in bangladesh. Journal of Development Economics 84, 731–754.
McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity design: Adensity test. Journal of Econometrics 142 (2), 698–714.
Miller, G. and M. Mobarak (2013). Gender di↵erences in preferences, intra-household externalities,and low demand for improved cookstoves. NBER Working Paper No.18964.
Mu, Q. and J. Zhang (2014). Air pollution and defensive expenditures: Evidence from particulate-filtering facemasks. E3 WP-078.
Neidell, M. (2009). Information, avoidance behavior, and health. Journal of Human Re-sources 44 (2), 450–478.
Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica 69 (2),307–342.
Pope, C. A. I., M. Ezzati, and D. Dockery (2009). Fine-particulate air pollution and life expectancyin the united states. New England Journal of Medicine 360, 376–386.
Wallace, L. (2008). E↵ectiveness of home air cleaners in reducing indoor levels of particles. Technicalreport, Health Canada.
WorldBank (2005). China - heat reform and building energy e�ciency project. Technical report,World Bank, Washington, DC.
25
Zheng, S., C. Sun, and M. Kahn (2015). Self-protection investment exacerbates air pollutionexposure inequality in urban china. NBER Working Paper No.21301.
Zivin, J. G. and M. Neidell (2009). Days of haze: Environmental information disclosure andintertemporal avoidance behavior. Journal of Environmental Economics and Management 58,119–128.
Zivin, J. G., M. Neidell, and W. Schlenker (2011). Water quality violations and avoidance be-havior: Evidence from bottled water consumption. American Economic Review: Papers andProceedings 101 (3), 448–453.
26
Figure 1: Huai River Boundary and City Locations
Notes: The line in the middle of the map is the Huai River-Qinling boundary. Each dot represents 1 city.There are 81 cities in our sample.
27
Figure 2: Regression Discontinuity Design at the Huai River Boundary
(a) PM10 in Winter
8010
012
014
016
0PM
10
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
(b) Market Share of HEPA Purifiers in Winter
.3.4
.5.6
.7.8
.9M
arke
t sha
re o
f HEP
A pu
rifier
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Notes: The sample includes 70 cities within 10-degree latitude relative to the Huai River. Figure 2a plotsthe average PM10 during winter (December-March) in 2006-2012 by 1.5 degree of latitude north to the HuaiRiver boundary. The vertical line at 0 indicates the location of the river. Each dot represents cities in 1.5degree of latitude and corresponds to the middle point of the range on the x-axis. For example, the dot at0.75 on the x-axis represents cities between 0 and 1.5 degree of latitude north to the river line. The y-axisindicates the average PM10 level of cities within 1.5 degree of latitude. Figure 2b shows the market share ofHEPA purifiers in winter by 1.5 degree of latitude north to the Huai River line.
28
Table 1: Summary Statistics
Whole sample HEPA purifiers Non-HEPA purifiers
(1) (2) (3)Panel A: Air purifier data (product level)
Percentage of HEPA purifiers 0.52(0.50)
Number of sales 463.07 535.52 383.18(1214.97) (1452.07) (879.83)
Price (USD) 407.21 461.47 347.38(343.94) (356.14) (320.40)
Humidifing 0.135 0.137 0.135(0.343) (0.345) (0.343)
Room coverage (square meter) 41.84 44.11 38.67(22.87) (24.25) (20.49)
Distance to factory or port (km) 885.58 884.54 886.73(344.26) (314.26) (375.44)
Panel B: Pollution data (city level)
PM10 in Winter (ug/m3) 118.53(22.97)
PM10 in Non-winter (ug/m3) 95.41(13.56)
Panel C: City demographics data (city level)
Population (1,000) 2498.69(419.11)
GDP per capita (USD) 8353.36(3371.61)
Share of inustrial GDP 0.50(0.11)
Years of schooling 8.29(0.04)
29
Table 2: First stage estimates on PM10 and Price
(a) First stage on PM10
PM10
(1) (2) (3)
North 14.36⇤⇤⇤ 15.40⇤⇤⇤ 7.85⇤⇤⇤
(0.76) (0.67) (0.83)Observations 4,686 4,686 4,686R2 0.43 0.55 0.75Quadratic latitude Y Y YQuadratic latitude*HEPA Y Y YModel FE Y Y YDemographic controls Y YLongitude decile FE Y
(b) First stage on Price of air purifiers
Price of Air Purifiers
(1) (2) (3) (4)
Distance to factory or port 0.020⇤⇤⇤ 0.018⇤⇤⇤ 0.020⇤⇤⇤ 0.018⇤⇤⇤
(0.003) (0.003) (0.003) (0.003)Observations 4,686 4,686 4,686 4,686R2 0.95 0.95 0.95 0.96Quadratic latitude Y Y Y YQuadratic latitude*HEPA Y Y Y YModel FE Y Y Y YDemographic controls Y YLongitude decile FE YCity FE Y
Notes: Each observation represents a product-city. City demographic controls include population, GDP percapita, share of industrial GDP and average years of schooling from City Statistical Yearbook (2006-2012)and the 2005 Census. Standard errors in brackets are clustered at the model level. * significant at 10% level;** significant at 5% level; *** significant at 1% level.
30
Table 3: Reduced form and 2SLS
Ln(market share)-Ln(outside option)
(1) (2) (3) (4)Panel A: Reduced form
North*HEPA 0.577⇤⇤ 0.573⇤⇤⇤ 0.584⇤⇤ 0.690⇤⇤⇤
(0.243) (0.206) (0.238) (0.244)
North 0.161 -0.197 -0.199(0.170) (0.149) (0.184)
Price -0.022⇤⇤⇤ -0.018⇤⇤⇤ -0.022⇤⇤⇤ -0.023⇤⇤⇤
(0.004) (0.004) (0.004) (0.004)Observations 4,686 4,686 4,686 4,686Panel B: 2SLS
PM10*HEPA 0.045⇤ 0.041⇤⇤ 0.041⇤⇤ 0.050⇤⇤⇤
(0.023) (0.017) (0.018) (0.019)
PM10 0.021 -0.013 -0.008(0.016) (0.012) (0.022)
Price -0.031⇤⇤⇤ -0.021⇤⇤⇤ -0.024⇤⇤⇤ -0.025⇤⇤⇤
(0.006) (0.005) (0.004) (0.005)Observations 4,686 4,686 4,686 4,686First-Stage F-Stat 10.27 10.51 12.59 16.35Quadratic latitude Y Y Y YQuadratic latitude*HEPA Y Y Y YModel FE Y Y Y YDemographic controls Y YLongitude decile FE YCity FE Y
Notes: Each observation represents a product-city. Panel A presents reduced-form estimates, where priceis instrumented by distance to factory/port. Panel B presents 2SLS results, where PM10*HEPA is instru-mented with North*HEPA, PM10 is instrumented with North, and price is instrumented with distance tofactory/port. City demographic controls include population, GDP per capita, share of industrial GDP andaverage years of schooling from City Statistical Yearbook (2006-2012) and the 2005 Census. Standard er-rors in brackets are clustered at the model level. * significant at 10% level; ** significant at 5% level; ***significant at 1% level. Stock-Yogo weak identification test critical value for two endogenous variables (10%maximal IV size): 7.03.
31
Table 4: Robustness
Ln(market share)-Ln(outside option)
(1) full sample (2) 10-degree (3) 7-degree (4) 5-degreePanel A: Linear and Linear*North
PM10*HEPA 0.045⇤⇤⇤ 0.050⇤⇤⇤ 0.051⇤⇤ 0.025⇤⇤⇤
(0.014) (0.019) (0.022) (0.009)
Price -0.028⇤⇤⇤ -0.025⇤⇤⇤ -0.029⇤⇤⇤ -0.014⇤⇤⇤
(0.005) (0.005) (0.008) (0.004)Observations 5,368 4,686 3,865 3,046First-Stage F-Stat 18.40 15.90 7.09 11.62Panel B: Quadratic
PM10*HEPA 0.054⇤⇤⇤ 0.050⇤⇤⇤ 0.049⇤⇤ 0.023⇤⇤⇤
(0.018) (0.019) (0.021) (0.009)
Price -0.028⇤⇤⇤ -0.025⇤⇤⇤ -0.029⇤⇤⇤ -0.014⇤⇤⇤
(0.005) (0.005) (0.008) (0.004)Observations 5,368 4,686 3,865 3,046First-Stage F-Stat 17.85 16.35 7.20 11.68Panel C: Cubic
PM10*HEPA 0.036⇤⇤ 0.025⇤⇤
(0.016) (0.013)
Price -0.027⇤⇤⇤ -0.024⇤⇤⇤
(0.005) (0.004)Observations 5,368 4,686First-Stage F-Stat 19.43 19.36Model FE Y Y Y YCity FE Y Y Y Y
Notes: Each observation represents a product-city. All results are from 2SLS regressions. Panel A controls forlinear latitude and linear latitude*north, and their interaction terms with HEPA. Panel B controls quadraticlatitude and their interaction terms with HEPA. Panel C controls for cubic latitude and their interactionterms with HEPA. Column 1 uses the full sample. Column 2 uses cities within 10 degrees relative to theHuai river. Column 3 uses cities within 7 degrees relative to the Huai river. Column 4 uses cities within 5degrees relative to the Huai river. In all regressions, PM10*HEPA is instrumented with North*HEPA, andprice is instrumented with distance to factory/port. Standard errors in brackets are clustered at the modellevel. * significant at 10% level; ** significant at 5% level; *** significant at 1% level. Stock-Yogo weakidentification test critical value for two endogenous variables (10% maximal IV size): 7.03.
32
Table 5: MWTP per Square Meter
Ln(market share)-Ln(outside option)
(1) (2) (3) (4)PM10*HEPA*Room size 0.0012*** 0.0009*** 0.0009*** 0.0011***
(0.0004) (0.0002) (0.0003) (0.0003)
PM10 0.0278 -0.0172* -0.0143(0.0213) (0.0100) (0.0208)
Price -0.0322*** -0.0182*** -0.0237*** -0.0271***(0.0099) (0.0065) (0.0067) (0.0087)
Observations 3,191 3,191 3,191 3,191Quadratic latitude Y Y Y YQuadratic latitude*HEPA Y Y Y YModel FE Y Y Y YDemographic controls Y YLongitude decile FE YCity FE Y
Notes: Each observation represents a product-city. This table presents 2SLS results, wherePM10*HEPA*Room size is instrumented with North*HEPA*Room size, PM10 is instrumented with North,and price is instrumented with distance to factory/port. City demographic controls include population,GDP per capita, share of industrial GDP and average years of schooling from City Statistical Yearbook(2006-2012) and the 2005 Census. Standard errors in brackets are clustered at the model level. * significantat 10% level; ** significant at 5% level; *** significant at 1% level.
33
Online Appendices Not For Publication
A Additional Figures
Figure A.1: Huai River Boundary and City Locations
Notes: The line in the middle of the map is the Huai River-Qinling boundary. Each dot represents 1 city.Each triangle represents a factory location or a port location.
34
Figure A.2: API distribution
(a) McCrary density test (75% of sample)
0.05
.1.15
0 200 400 600
(b) McCrary density test (25% of sample)
0.02
.04
.06
0 200 400 600
(c) Original distributions of the 25% sample
020
040
060
080
010
00nu
mbe
r of d
ays
in 7
yea
rs
0 100 200 300 400 500API
Original
(d) Counterfactual distribution of the 25% sample
020
040
060
080
010
00nu
mbe
r of d
ays
in 7
yea
rs
0 100 200 300 400 500API
Counterfactual
35
Figure A.3: Huai River: PM10 in non-winter months (April-November)
8010
012
014
016
0PM
10
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Notes: The sample includes 70 cities within 10-degree latitude relative to the Huai River. This figure plotsthe average PM10 during non-winter months (April-November) in 2006-2012 by 1.5 degree of latitude northto the Huai River boundary. The vertical line at 0 indicates the location of the river. Each dot representscities in 1.5 degree of latitude and corresponds to the middle point of the range on the x-axis. For example,the dot at 0.75 on the x-axis represents cities between 0 and 1.5 degree of latitude north to the river line.The y-axis indicates the average PM10 level of cities within 1.5 degree of latitude.
36
Figure A.4: Huai River and Demographics
040
0080
0012
000
1600
0G
DP
per c
apita
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
GDP per capita
010
0020
0030
0040
0050
00Po
pula
tion(
1000
)
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Population
.3.4
.5.6
.7Sh
are
of in
dust
rial G
DP
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Share of industrial GDP4
68
1012
1416
Year
s of
sch
oolin
g
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Years of schooling
Notes: The sample includes 70 cities within 10-degree latitude relative to the Huai River. These figuresplot the mean of each demographic variable by 1.5 degree of latitude north to the Huai River boundary.The vertical line at 0 indicates the location of the river. Each dot represents cities in 1.5 degree of latitudeand corresponds to the middle point of the range on the x-axis. For example, the dot at 0.75 on the x-axisrepresents cities between 0 and 1.5 degree of latitude north to the river line. The y-axis indicates the meanlevel of each variable in cities within 1.5 degree of latitude.
37
Figure A.5: Number of migrants to South of the river by city of origin
020
0040
0060
00N
umbe
r of m
igra
nts
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Number of migrants to South by city of origin
Notes: The sample includes 70 cities within 10-degree latitude relative to the Huai River. This figure plotthe number of migrations currently residing South of the Huai River by their city of origin. The verticalline at 0 indicates the location of the Huai river. Each dot represents cities in 1.5 degree of latitude andcorresponds to the middle point of the range on the x-axis. For example, the dot at 0.75 on the x-axisrepresents cities between 0 and 1.5 degree of latitude north to the river line. The y-axis indicates the meanlevel of each variable in cities within 1.5 degree of latitude.
38
Figure A.6: Fraction of available HEPA purifier products
.5.5
5.6
.65
Frac
tion
of a
vaila
ble
HEP
A pu
rifier
pro
duct
s
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10Degree North to the Huai River
Notes: The sample includes 70 cities within 10-degree latitude relative to the Huai River. This figure plotsthe fraction of available HEPA purifier products out of all purifier products by 1.5 degree of latitude northto the Huai River boundary. The vertical line at 0 indicates the location of the river. Each dot representscities in 1.5 degree of latitude and corresponds to the middle point of the range on the x-axis. For example,the dot at 0.75 on the x-axis represents cities between 0 and 1.5 degree of latitude north to the river line.The y-axis indicates the mean level of each variable in cities within 1.5 degree of latitude.
39
B Additional Tables
Table A.1: Controlling for GDP*HEPA and Schooling*HEPA
Ln(market share)-Ln(outside option)
(1) (2) (3)Panel A: Reduced form
North*HEPA 0.527** 0.562** 0.646***(0.208) (0.239) (0.250)
North -0.169 -0.185(0.151) (0.190)
Price -0.018*** -0.022*** -0.023***(0.004) (0.004) (0.004)
Observations 4,686 4,686 4,686Panel B: 2SLS
PM10*HEPA 0.044** 0.045** 0.053**(0.019) (0.021) (0.022)
PM10 -0.016 -0.012(0.013) (0.023)
Price -0.021*** -0.024*** -0.025***(0.005) (0.004) (0.005)
Observations 4,686 4,686 4,686First-Stage F-Stat 9.99 12.32 16.01Quadratic latitude Y Y YQuadratic latitude*HEPA Y Y YModel FE Y Y YDemographic controls Y YLongitude decile FE YCity FE YGDP*HEPA and Schooling*HEPA Y Y Y
Notes: Each observation represents a product-city. Panel A presents reduced-form estimates, where priceis instrumented by distance to factory/port. Panel B presents 2SLS results, where PM10*HEPA is instru-mented with North*HEPA, PM10 is instrumented with North, and price is instrumented with distance tofactory/port. GDP per capita*HEPA and schooling*HEPA are included in all regressions. Standard errors inbrackets are clustered at the model level. * significant at 10% level; ** significant at 5% level; *** significantat 1% level. Stock-Yogo weak identification test critical value for two endogenous variables (10% maximalIV size): 7.03.
40
Table A.2: First stage (robustness)
PM10
full sample 10-degree 7-degree 5-degreePanel A: Linear and Linear*North
North 19.44*** 16.28*** 21.63*** 18.23***(0.58) (0.68) (1.02) (1.54)
Observations 5,368 4,686 3,865 3,046R2 0.63 0.55 0.48 0.47Panel B: Quadratic
North 14.30*** 15.46*** 20.45*** 18.49***(0.54) (0.66) (0.95) (1.43)
Observations 5,368 4,686 3,865 3,046R2 0.64 0.55 0.49 0.47Panel C: Cubic
North 23.09*** 33.08***(0.70) (0.93)
Observations 5,368 4,686R2 0.65 0.57Model FE Y Y Y YDemographic controls Y Y Y Y
Notes: Each observation represents a product-city. Panel A controls for linear latitude and linear lati-tude*north, and their interaction terms with HEPA. Panel B controls quadratic latitude and their interactionterms with HEPA. Panel C controls for cubic latitude and their interaction terms with HEPA. Column 1uses the full sample. Column 2 uses cities within 10 degrees relative to the Huai river. Column 3 uses citieswithin 7 degrees relative to the Huai river. Column 4 uses cities within 5 degrees relative to the Huai river.Standard errors in brackets are clustered at the model level. * significant at 10% level; ** significant at 5%level; *** significant at 1% level.
41
Table A.3: Huai river and Demographics
Population GDP per capita Industrial GDP share Schooling(1) (2) (3) (4) (5) (6) (7) (8)
North -1725.701 -1837.974 -2800.805⇤⇤ 20.214 0.026 0.055 -0.089 -0.087(1170.940) (1152.711) (1396.275) (1572.412) (0.048) (0.050) (0.369) (0.401)
Observations 71 71 71 71 71 71 71 71R2 0.03 0.15 0.07 0.41 0.01 0.17 0.19 0.39Mean of Dependent Variable 2498.69 2498.69 8353.36 8353.36 0.50 0.50 8.29 8.29Quadratic latitude Y Y Y Y Y Y Y YLongitude decile FE Y Y Y Y
Notes: Each observation represents a city. The sample includes 70 cities within 10-degree latitude relative to the Huai River. Quadratic latitude isincluded in all regressions. Robust standard errors are in brackets.
42