1

The Effect of Air Pollution on Body Weight and Obesity:

Evidence from China

Olivier Deschenes

Department of Economics, University of California, Santa Barbara, IZA, and NBER

olivier@econ.ucsb.edu

Huixia Wang

School of Economics and Trade, Hunan University

huixia_wang@outlook.com

Si Wang

Center for Economics, Finance, and management Studies, Hunan University

siww26@gmail.com

Peng Zhang

School of Accounting and Finance, The Hong Kong Polytechnic University

peng.af.zhang@polyu.edu.hk

March 9, 2019

Abstract

We provide the first study on estimating the causal effect of air pollution on body weight. Using

the China Health and Nutrition Survey (CHNS), which provided detailed longitudinal health

and socioeconomic information for 13,226 adult individuals over 1989-2011, we find

significantly positive effect of air pollution, instrumented by thermal inversions, on body mass

index (BMI). Specifically, a 1 µg/m3 (1.59%) increase in average PM2.5 concentrations in the

past 12 months increases the BMI by 0.31%, and further increases the overweight and obesity

rates by 0.89 and 0.19 percentage points respectively. Our paper identifies a new cause of

obesity, and shed light on the morbidity cost of air pollution

JEL Codes: I12, I15, Q53

Keywords: Weight gain, obesity, air pollution, China.

Acknowledgements: We thank Mark Rosensweig and Ruixue Jia for helpful comments. Any remaining

errors are our own.

2

1 Introduction

Overweight and obesity have drawn much attention recently, as they are

important risk factors for a variety of chronic diseases, including diabetes,

cardiovascular and kidney diseases, musculoskeletal disorders, and colorectal,

pancreatic, gallbladder and kidney cancers (Prospective Studies Collaboration, 2009;

Berrington de Gonzales et al., 2010; Emerging Risk Factors Collaboration, 2011; Zheng

et al., 2011). In 2016, nearly 40% of adults were overweight (body-mass index

(BMI)>=25)1, and 11% of men and 15% of women were obese (BMI>=30) (WHO,

2018a). In 1975, the obesity rate was only 3.2% for men and 6.4% for women (NCD-

RisC, 2016). It is estimated that overweight and obesity lead to at least 2.8 million

deaths and 35.8 million disability-adjusted life years annually across the world (WHO,

2018b). Overweight and obesity have become a large economic burden, and costed

USD 2 trillion, or 2.8% of global GDP in 2014 (Dobbs et al., 2014).

It is therefore critical to understand the causes of obesity. In this paper, we

identify a new and important determinant of obesity: ambient air pollution, and

particularly, very fine particulate matter (PM2.5)2. Nowadays, more than 90% of global

population lives in places with poor air quality. Air pollution causes 4.2 million deaths

per year, and is responsible for one in nine deaths around the world (WHO, 2018c).

Understanding the link between air pollution and obesity is thus crucial for policy

1 Overweight sometimes is defined when BMI is between 25 and 30, which excludes obesity.

2 We measure ambient air pollution using PM2.5. Therefore, we use two terms interchangeably

throughout the paper.

3

makers. To our best knowledge, we are the first to estimate the causal effect of air

pollution on BMI and obesity3.

Our focus, China, provides a unique opportunity to study the relationship

between air pollution and obesity. Over the past decades, China’s GDP grew from USD

798.4 billion in 1989 to USD 6.68 trillion in 2011 (Panel A of Figure 1), with annual

growth rate of 8-10%. Meanwhile, the national average concentration of PM2.5

increased from 40.11 microgram per cubic meter (µg/m3) in 1989 to 68.38 µg/m3 in

2011, with a 70 percent increase. On the other hand, overweight and obesity have

become prevalent. The average BMI increased from 21.49 in 1989 to 23.76 in 2011,

with a 11 percent increment. In 1989, the overweight and obesity rate were only 8.57

and 0.48 percent respectively. However, after 22 years, these two numbers became

32.83 and 4.90 respectively. In 2014, China ranked the first in obese men (16.3% of

global obesity) and women (12.4% of global obesity) (NCD Risk Factor Collaboration,

2016).

Air pollution could affect body weight through several channels. First, air

pollution could cause metabolic dysfunction such as trigger oxidative stress and adipose

tissue inflammation (Xu et al., 2011) and slow down glucose metabolism (Toledo-

Corral et al., 2018). Second, air pollution could affect body weight indirectly through

elevating the risks of a number of chronic diseases, such as cardiovascular and

respiratory diseases, heart diseases, and some cancers (An et al., 2018a). Thirdly, air

3 Several studies in the health science literature find correlations between air pollution and obesity.

See the review in An et al. (2018).

4

pollution could cause sleep disruption (Zanobetti et al., 2009; Lawrence et al., 2018),

which in turn, could increase the BMI (Keith et al., 2006). Lastly, behavior responses

to air pollution such as staying indoors, reducing physical activities, and increasing

sedentary activities could reduce calories expended (Neidell, 2009; Jerrett et al., 2010;

McConnell et al., 2014; Li et al., 2015; An et al., 2018b). On the other hand, pollution-

induced depression and anxiety could increase stress-related intake from fats and sugars

and increase calories consumed (Laitinen et al., 2002; Ozier et al., 2008; Chen et al.,

2018).

Identifying the causal effect of air pollution on BMI and obesity is challenged

primarily because of the omitted-variable bias. Air pollution, as a byproduct of

economic activity, is typically correlated with economic confounders, such as income,

GDP, and macroeconomic conditions. Those economic confounders are also important

determinants of obesity (Cawley, 2015). The direction of the bias depends on the

correlation between air pollution and those confounders, as well as the effects of those

confounders on body weight.

To identify the causal effect of air pollution on BMI and obesity, we use thermal

inversions as an instrumental variable for air pollution. Thermal inversions occur when

temperature in the upper atmospheric layer is higher than that of the lower layer, and

thus trapped air pollution near the surface. As a complex meteorological phenomenon,

the formation of thermal inversions is typically independent of economic activities.

Importantly, we utilize the longitudinal structure of our health data and include

individual fixed effects. Therefore, our identification is from random fluctuation in air

5

pollution instrumented by thermal inversions across different years for the same

individual. In addition, we include weather controls in flexible specifications as well as

year-by-month fixed effects to control for seasonality. Thermal inversions have been

used as the IV for air pollution in several studies (e.g., see Arceo et al., 2016; Chen et

al., 2017; Fu et al., 2017; Chen et al., 2018).

Our data on BMI are from the China Health and Nutrition Survey (CHNS), the

longest and most comprehensive health survey in China. The CHNS provided detailed

information on health and nutrition along with socioeconomic and demographic data

for 13,226 adult (age>=18) individuals from eight provinces in China over the period

of 1989-2011. We use measured instead of self-reported body weight and height to

calculate the BMI and define overweight and obesity accordingly. We then match

health data with satellite-based pollution and thermal inversion data by county and

month for each interviewee.

We find significantly positive effect of PM2.5 on body weight. Specifically, a 1

µg/m3 (1.59%) increase in average PM2.5 concentrations in the past 12 months increases

the BMI by 0.31%, and further increases the overweight and obesity rates by 0.89 and

0.19 percentage points respectively. On the other hand, we do not find any short-run

effect such as past one to three months of PM2.5 on body weight.

We then conduct the heterogeneity analysis. We find similar effect between

male and female, as well as urban and rural residents. However, we find larger effect

on adults aged below 60 and the highly educated population, which may due to behavior

responses such as exercising. We then formally test the pollution effect on those

6

behavior responses including nutrition intake, physical and sedentary activities, and

sleeping. However, because the survey only recorded these behavior responses in a time

short window such as past three or seven days before the interview, we do not find any

meaningful effect on behavior responses.

This paper contributes to two strands of the literature. First, a growing body of

literature is seeking to understand the economic causes of obesity (Cawley, 2015). Most

of previous studies focused on economic factors, including fast food and groceries

prices (Courtemanche and Carden, 2011; Cotti and Tefft, 2013), proximity to fast food

outlets (Currie et al., 2010; Anderson and Matsa, 2011), food assistance programs

(Bhattacharya et al., 2006; Gundersen et al., 2012), income (Cawley et al., 2010; Akee

et al., 2013), education (Brunello et al., 2013; Clark and Royer, 2013), unemployment

rate (Ruhm, 2000; Charles and DeCicca, 2008), and peer and neighborhood effects

(Kling et al., 2007; Carrell et al., 2011). We show that the environment, in particular

the ambient air pollution, also plays an important role in causing obesity. A 1 µg/m3

increase in PM2.5 has the same effect with an additional fast food restaurant and a half

of an additional year of schooling on obesity.

Our paper also contributes to a growing body of literature on estimating the cost

of air pollution on a variety of economic outcomes, including mortality and morbidity

(Chay and Greenstone, 2003; Schlenker and Walker, 2015), labor productivity (Graff

Zivin and Neidell, 2012), labor supply (Hanna and Oliva, 2015), test scores (Ebenstein

et al., 2016), migration (Chen et al., 2017), mental health and dementia (Chen et al.,

2018; Bishop et al., 2018), and others. We estimate that 1 µg/m3 increase in average

7

PM2.5 concentrations induces a total of CNY 1.91 billion (USD 0.29 billion) health

expenditure on overweight and obesity.

The remainder of the paper is organized as follows. Section 2 describes the

mechanisms through which air pollution affects obesity. Section 3 describes the

empirical strategy. In Section 4, we report the data sources as well as the summary

statistics. We report the regression results, heterogeneity analysis, and mechanism tests

in Section 5. Section 6 discusses the implication of our results and concludes.

2 Mechanisms

Air pollution could affect body weight through several channels. First, air

pollution could lead to metabolic disorder, which is closely related to body weight (An

et al., 2018a). For example, Xu et al. (2011) find that PM2.5 exposure triggers oxidative

stress and adipose tissue inflammation, which further predispose to metabolic

dysfunction. Toledo-Corral et al. (2018) find that PM2.5 exposure has negative effect on

glucose metabolism.

Second, air pollution could affect body weight indirectly through elevating the

risks for a number of chronic diseases (An et al., 2018a). For example, air pollution

exposure could lead to cardiovascular and respiratory diseases, heart diseases, and some

cancers (WHO, 2018c). Those chronic diseases, in turn, could affect body weight (An

et al., 2018a).

Third, air pollution could affect body weight through sleep disorders.

Researchers have found that increased concentrations of PM significantly increase

8

sleep disorders for adults in the U.S. (Zanobetti et al., 2009) and children in China

(Lawrence et al., 2018). Sleep disorders, in turn, could increase BMI because of

decreased leptin, thyroid-stimulating hormone secretion, and glucose tolerance, and

increased ghrelin level (Keith et al., 2006).

Pollution could also affect body weight through behavior responses. Many

studies find that people are more likely to stay indoors (Neidell, 2009), reduce physical

activities, and increase sedentary behaviors such as sitting, reclining and lying (Jerrett

et al., 2010; McConnell et al., 2014; Li et al., 2015; An et al., 2018b). These behaviors

in turn will reduce calories expended and causes obesity (WHO, 2018a). On the other

hand, air pollution may increase calories consumed. For example, Chen et al. (2018)

find that air pollution is likely to induce a variety of mental illness, such as depression

and anxiety, which in turn, could induce stress-related intake from fats and sugars

(Laitinen et al., 2002; Ozier et al., 2008). Stress also helps release the hormone cortisol,

which can increase appetite for energy-intensive foods and lead to insulin resistance

and fat accumulation (Björntorp, 1997).

3 Empirical Strategy

Our goal is to estimate the causal effect of air pollution on body weight. The

primary challenge is the omitted-variable bias. As a byproduct of economic activity, air

pollution is typically correlated with economic confounders, such as income, GDP, and

macroeconomic conditions. Those confounders are also important determinants of

obesity, and their impacts could be either positive or negative. For example, additional

9

income could either increase or decrease body weight. If both food/sedentary pursuits

and good health/appearance are normal goods, additional income could increase

consumption on both sides. If more food/sedentary pursuits are consumed than good

health/appearance, additional income could increase body weight, and vice versa.

Researchers have found that people are more likely to spend the first few thousand

dollars on food, and invest additional money in weight loss for the sake of appearance

and health (Philipson and Posner, 2003; Lakdawalla et al., 2005). Therefore, there exists

an inverted U-shaped relationship between income and weight.

Because the ambiguous effect of those economic confounders on body weight

as well as the correlation between air pollution and economic confounders, the bias

direction of air pollution on body weight is unknown. To identify the causal effect, we

apply an instrumental variables (IV) approach. In particular, we use thermal inversions,

a meteorological phenomenon, to instrument air pollution.

Under normal conditions, temperature in the upper atmospheric layer is lower

than that of surface layer. Therefore, air pollutants could be transmitted from ground to

the upper layer and further be spread out. Under certain circumstances (see Arceo et al.

(2016)), temperature in the upper layer is higher than that of ground layer, which forms

a thermal inversion. Air pollutants are thus trapped near the ground leading to higher

air pollution concentrations.

Because thermal inversion is meteorological phenomenon, the formation is

typically independent of economic activity. Figure 2 plots the average monthly

cumulative thermal inversions over 1989-2011. Unlike PM2.5 which has a clear time

10

trend, the frequencies of thermal inversions are more or less random. To ensure our IV

meets the exclusion restriction criteria, we control for flexible weather variables, so that

thermal inversions affect body weight only through air pollution. Thermal inversions

have been used as IV for air pollution in several studies, including Hicks et al. (2015),

Arceo et al., (2016), Chen et al. (2017), Fu et al., (2017), Chen et al. (2018), and Jans

et al. (2018).

We propose the following 2SLS model to estimate the causal effect of air

pollution on body weight:

𝑌𝑖𝑡 = 𝛽0 + 𝛽1𝑃𝑖𝑡 + 𝑓(𝑊𝑖𝑡) + 𝛾𝑖 + 𝜎𝑡 + 𝜀𝑖𝑡 (1)

𝑃𝑖𝑡 = 𝛼0 + 𝛼1𝐼𝑖𝑡 + 𝑓(𝑊𝑖𝑡) + 𝛾𝑖 + 𝜎𝑡 + 𝑢𝑖𝑡 (2).

In the model, 𝑌𝑖𝑡 denotes the body weight measure, including BMI,

overweight, and obesity for individual 𝑖 at date 𝑡. We use 𝑃𝑖𝑡 to denote the average

concentration of PM2.5. Note that we do not have priori to specific the exposure

window. In other words, we do not know how long will it take for air pollution to

affect body weight. In our baseline, we choose the exposure window as 12 months.

For example, if an individual’s BMI was measured on June 15th 2000 in county 𝑐, we

use the average concentration of PM2.5 between July 1999 to June 2000 for that

county. Since our pollution data are only available at monthly level, we cannot

construct the exposure window based on specific dates, i.e., June 16 1999 to June 15

2000. We conduct two robustness checks on the exposure window. First, we vary the

exposure window from one month to 18 months. The effect is mostly precisely

estimated and large in magnitude for the exposure window of 12 months. Second, we

11

exclude the current month when we construct the 12-month exposure window. For

example, if an individual’s BMI was measured on June 15 2000, our baseline PM2.5

measure uses the average between July 1999 to June 2000, and our alternative

measure uses the average between June 1999 to May 2000. Our results are robust to

drop the month of the interview.

We instrument 𝑃𝑖𝑡 using number of thermal inversions, denoted by 𝐼𝑖𝑡, in the

same exposure window. We use 𝑓(𝑊𝑖𝑡) to denote weather variables, including

temperature, precipitation, relative humidity, sunshine duration, wind speed, and

pressure in flexible specifications in the same exposure window. We use individual

fixed effects, 𝛾𝑖, to control for any time-invariant individual-specific characteristics,

such as gender and geographic locations. We use 𝜎𝑡, year-by-month fixed effects, to

control for nation-wide seasonality. We also conduct a robustness check using date

fixed effects to replace year-by-month fixed effects. Although out estimates are still

statistically significant, the magnitude is larger and the first-stage Kleibergen-Paap

Wald rk (KP) F-statistic (Kleibergen and Paap, 2006) is weaker. Therefore, we use

year-by-month fixed effects as the baseline time fixed effects to be conservative.

We employ the two-way clustering (Cameron et al., 2011) and cluster the

standard errors at individual level and county-year-month level. This controls for the

autocorrelation for the same individual across different years as well as the

autocorrelation within each county-year-month cell. Our results are robust to

alternative clustering methods, which is discussed in the results section.

12

To sum up, our identification relies on comparing the BMI of the same

individual in a more inversion-intensive and thus more polluted year versus a less

inversion-intensive and polluted year, after we adjust the year-specific seasonality and

weather shocks, to identify the causal effect of air pollution on body weight.

4 Data

4.1 BMI and obesity

We obtain the BMI data from the China Health and Nutrition Survey (CHNS),

which is one of the longest and most comprehensive longitudinal health surveys in

China and is still ongoing. The CHNS is jointly conducted by the Carolina Population

Center at the University of North Carolina at Chapel Hill and the National Institute for

Nutrition and Health at the Chinese Center for Disease Control and Prevention. The

survey covered 15,000-19,000 individuals in 4400-7200 households from nine

provinces4 (two-digit code)5 over the period of 1989-20116. The sample was selected

using a multistage random cluster sampling method. Specifically, for each province,

two cities (four-digit code) and four counties (six-digit code) were randomly selected.

The survey then randomly selected urban districts (six-digit code) for cities and villages

4 These nine provinces include Liaoning, Heilongjiang, Jiangsu, Shandong, Henan, Hubei, Hunan,

Guangxi, and Guizhou.

5 In China, there are three administrative levels, provinces/municipal cities (two-digit code),

prefectures/cities (four-digit code), and counties/districts (six-digit code). See

http://www.stats.gov.cn/tjsj/tjbz/tjyqhdmhcxhfdm/.

6 These years include 1989, 1991, 1993, 1997, 2000, 2004, 2006, 2009, and 2011.

13

and towns for counties. Those were defined as communities. Finally, households were

randomly selected from those communities.

The CHNS provided rich information on health and nutrition as well as

socioeconomic and demographic characteristics for both rural and urban China. One

key advantage of the CHNS is that the body weight and height are measured by medical

staff, instead of self-reported by the interviewee. This is important because individuals

tend to underreport their weight, especially for heavier individuals (Cawley et al., 2015).

We calculate the BMI using the body weight measured by kilograms (kg) divided by

the square of the body height measured by square meters (m2). The unit of BMI is thus

kg/m2. Note that this formula only applies to adults aged>=18, and thus our main

sample only includes adults. We define a person is overweight if BMI>=25, and obese

if BMI>=30 (WHO, 2018a). Overweight sometimes is defined between 25 and 30. In

that case, obesity is excluded from the overweight category. Therefore, our measure on

overweight includes both overweight and obesity.

4.2 Air pollution

Our data on air pollution are from satellite-based Aerosol Optical Depth (AOD)

retrievals. In particular, we obtain the AOD data from the Modern-Era Retrospective

analysis for Research and Applications version 2 (MERRA-2) from the National

Aeronautics and Space Administration (NASA) of the U.S. The data are available at

50*60-km grid level for each month since 1980. We calculate the concentration of

PM2.5 following Buchard et al. (2016). We then aggregate from grid to county for each

14

month. This dataset has been used in previous studies (Chen et al., 2017; Fu et al., 2017),

and has been validated with ground-based pollution data (Chen et al., 2017). We do not

use ground-based pollution data mainly because they are only available after 2000 and

covered only a few cities.

4.3 Thermal inversions

We obtain the thermal inversions data also from MERRA-2. The data report the

air temperature for each 50*60-km grid for 42 atmospheric layers, ranging from 110

meters to 36,000 meters. The data are available at six-hour period from 1980. We

aggregate all data from grid to county. We determine the existence of a thermal

inversion if the temperature in the second layer (320 meters) is higher than that of the

first layer (110 meters) for each six-hour period, and then aggregate number of

inversions from each six-hour period to each month, and further average to the 12-

month exposure window.

4.4 Weather

The weather data are obtained from the National Meteorological Information

Center, which releases daily weather variables, including temperature, precipitation,

relative humidity, sunshine duration, wind speed, and pressure for more than 800

weather stations in China. We use the inverse-distance weighting (IDW) method to

convert weather data from station to county level and choose a radius of 200 km

(Deschenes and Greenstone, 2007). To account for non-linear effects of temperature,

15

we construct the number of days within each 5 °C bin and aggregate to the month and

year level (Deschenes and Greenstone, 2011). For other weather variables, we average

relative humidity, sunshine duration, wind speed, and pressure, and aggregate

precipitation. We also include the quadratic of each weather variables except for

temperature bins to account for the non-linear effects.

4.5 Summary statistics

Our final sample has 13,226 adult individuals from 71 counties/districts across

eight provinces over 1989-20117. We also plot the number of interviewees in each year

in Figure A1 in the Online Appendix. There were 3,569 interviewees in 1989, and the

number of interviewees increased to 7,729 to 1991 and was relatively stable afterwards.

Our sample is an unbalanced panel. In Figure A2 in the Online Appendix, we plot the

number of interviewees against the time frequency. There are 3,586 individuals (27%)

were interviewed only two times, followed by 2,127 individuals (16%) with three times.

Only 757 individuals (6%) are present throughout the whole sample period.

One concern is that pollution may induce people’s moving (Chen et al., 2017)

and thus bias our estimates. Note that the individual fixed effects absorb all initial

sorting into difference places. Therefore, only individual’s moving during our sample

period will bias the estimates. However, in our sample, all individuals remained in the

same county. This is because they will be excluded from the survey once they changed

the county of residence.

7 The county identifiers are not available for the province of Heilongjiang.

16

Interviews were only conducted from July to December, with most in the Fall

(Figure A3 in the Online Appendix). These three months (September to November)

account for 90% of total interviews. On the other hand, thermal inversions also have a

strong seasonality. Figure A4 in the Online Appendix plots average thermal inversions

across months. It is clear that most inversions occurred in the cold months. Therefore,

the first stage is weak when we use other exposure windows such as 4-6 months.

However, the seasonality will not bias our baseline estimates for two reasons. First, we

control for year-by-month fixed effects. Second, the interview time changes little across

years, as most of them are concentrated in the Fall. Thus, for our baseline exposure

window, i.e., 12 months, we mainly use year-to-year variation (e.g., October 1999-

September 2000 versus October 2003-Setember 2004), instead of season-to-season

variation across years (e.g., October 1999-September 2000 versus June 2003-Mary

2004).

Table 1 reports the summary statistics. We have three measures on body mass:

the BMI, a dummy variable for overweight and obesity. We also report the body weight

and height. In our sample, the average BMI is 22.62, with a standard deviation of 3.30.

The average BMI is 23.76 in 2011, the most recent period. Figure 3 plots the histogram

of BMI, and most observations centered between 18 to 25. There are some extreme

values, with the minimum of 4.83 and maximum of 63.78. Our results are robust if we

drop the top and bottom 0.5% of data.

The average overweight and obesity rates are 21.47% and 2.52% during our

sample period, and 32.84% and 4.90% in 2011. The average body weight is 58.03 kg,

17

and average height is 159.92 cm. Female accounts for 52% observations and has

slightly higher BMI and overweight and obesity rates than male. The same pattern has

been found in the U.S. (National Center for Health Statistics, 2014) and the world

(WHO, 2018a). On the other hand, female has lower body weight and height than male.

In our sample, the average concentrations of PM2.5 is 63 µg/m3, which is six

times higher than the WHO standard of ten µg/m3 (WHO, 2006). The concentration

varies from 23.75 to 141.32, with a standard deviation of 26.19. For our IV, the average

monthly cumulative inversion times are 22.46. Since the occurrence is determined at

each six-hour period, the probability of having an inversion is 22.46/(30*4)=18.72%.

5 Results

5.1 First-stage results

Table 2 reports the first-stage estimates. In column (1), we have individual

fixed effects and year fixed effects. In column (2), we replace year fixed effects with

year-by-month fixed effects, to control for year-specific seasonality. In the last column,

we further add weather controls.

Overall, we find a strong first-stage result. The estimated coefficients are all

statistically significant at the 1% level. However, the KP F-statistic is much higher and

well above the Stock-Yogo critical value of 16.38 (Stock and Yogo, 2005) when we

add weather controls, suggesting that it is preferred to control for weather variables.

The estimated coefficient in column (3) suggests that one more occurrence of thermal

18

inversion (4.45%) in the past 12 months increases the PM2.5 concentrations in the same

period by 0.34 µg/m3 (0.54%), with an elasticity of 0.12.

5.2 Second-stage results

Table 3 reports the main regression results. The dependent variables are BMI in

columns (1) and (2), a dummy variable for overweight in columns (3) and (4), and a

dummy variable for obesity in columns (5) and (6). Panel A is the IV estimates and

Panel B is the fixed effects estimates. All columns include individual fixed effects and

weather controls. Columns (1), (3), and (5) use year fixed effects while columns (2),

(4), and (6) uses year-by-month fixed effects.

Several observations can be drawn from this table. First, we find a statistically

significant and economically large effect of PM2.5 on BMI. Our preferred specification,

column (2), suggests that a 1 µg/m3 (1.59%) increase in average PM2.5 concentrations

in the past 12 months increases the BMI by 0.0691 units (0.31%). This indicates the

elasticity between PM2.5 and BMI is around 0.19. We can also convert the magnitude

using standard deviations. We conclude that a one standard deviation increase in PM2.5

concentrations increases the BMI by 0.55 standard deviations.

Second, air pollution increases the probability of being overweight, or with

BMI>=25. Column (4) suggests that a 1 µg/m3 increase in average PM2.5 concentrations

in the past 12 months increases the probability of being overweight by 0.89 percentage

points, or 4.15 percent of the mean (mean=21.47 percentage points). In other words, a

19

one standard deviation increase in PM2.5 in the past 12 months increases the probability

of being overweight by 0.57 standard deviations.

Third, the effect of PM2.5 on obesity rates is statistically weak. This may be

because only few individuals are obese. In our sample, the obesity rate was only 2.52

percent. Nevertheless, our estimate in column (6) suggests that a 1 µg/m3 increase in

average PM2.5 concentrations in the past 12 months increases the probability of being

overweight by 0.19 percentage points, or 7.54 percent of the mean. In other words, a

one standard deviation increase in PM2.5 in the past 12 months increases the probability

of being obese by 0.32 standard deviations.

Fourth, statistically speaking, the estimated coefficients differ little between the

model with year fixed effects and year-by-month fixed effects. This is because most

interviews were conducted in the Fall, and thus we mainly use year-to-year variation,

instead of season-to-season variation across years.

Lastly, the OLS estimates are much smaller in magnitude compared to the IV

model and are statistically insignificant. This suggests that the socioeconomic

confounders severely bias the estimates downward.

How large are these effects? During our sample period, the average PM2.5

concentrations increased from 40.11 µg/m3 in 1989 to 68.38 µg/m3 in 2011, with a 28.27

µg/m3 (70.48%) increment. Scale our estimates linearly, this increment increases the

average BMI by 1.95 units (9.07% evaluated at the mean BMI in 1989) and increases

the probability of being overweight and obese by 25.16 and 5.37 percentage points

20

respectively. Therefore, the increasing air pollution could be an important factor in

explaining the rising of obesity in China.

Another way to assess the magnitude of the coefficient is to convert the

probability of being overweight and obese to actual population when PM2.5 increases

by 1 µg/m3. The adult population in China in 2016 was 1.15 billion (China Statistical

Yearbook, 2017)8. Apply our estimates in columns (4) and (6), we conclude that a 1

µg/m3 increase in average PM2.5 concentrations is likely to induce additional 10.24

million overweight adult populations and 2.19 million adult obese populations in China.

In 2016, 1.9 billion adults were overweight and 650 million of them were obese

worldwide (WHO, 2018a). Therefore, a 1 µg/m3 increase in average PM2.5

concentrations in China contributes to 0.54% of global overweight rate and 0.34% of

global obesity rate.

5.3 Robustness checks

We conduct various robustness checks in Table 4. Column (1) is the baseline

model, in which we use year-by-month fixed effects to control for nation-wide year-

month shocks. In column (2), we replace year-by-month fixed effects using date fixed

effects, to control for nation-wide daily shocks, which is more demanding. We find

much larger effects for all three dependent variables. However, the KP F-statistic is

only 12.8, which is below the Stock-Yogo critical value of 16.38. Therefore, we use the

year-by-month fixed effects as the baseline time fixed effects.

8 In our sample, the adult population are referred to aged>=18. However, in the Chinese Statistical

Yearbook, the population data are only available for aged>=15.

21

Our baseline model includes weather variables in flexible specifications. This

is to meet the exclusion restriction and ensure that air pollution is the only channel

through which thermal inversions affect body weight. In column (3), we exclude

weather controls. Although the KP F-statistic is smaller, the magnitude and statistical

significance of the estimated coefficients change little.

Our baseline exposure window is 12 months, and we include the current month

of the interview. For example, if an individual was interviewed on June 15 2000, we

construct the exposure window from July 1999 to June 2000. In column (4), we drop

the current month, and therefore construct the exposure window from June 1999 to May

2000. Our results are almost identical, although with the statistical significance is lower.

Column (5) tests the robustness of our IV construction. In our baseline model,

we code thermal inversion using the temperature difference between the first (110

meters) and the second layers (320 layers). In column (5), we replace the second layer

with the third layer (540 meters). The results are very similar.

We then test the robustness of excluding extreme values of BMI in column (6).

The average BMI in our sample is 22.62, with a minimum of 4.83 and a maximum of

63.78. In column (6), we winsorize the top and bottom 0.5% observations, and the

results are very robust.

Our current definitions on overweight (BMI>=25) and obesity (BMI>=30) are

from WHO, which are mainly derived from Western populations. Zhou (2002)

proposed that the BMI cutoff of 24 for overweight and 28 for obesity is more

appropriate for the Chinese populations. Using the new cutoff, the average overweight

22

and obesity rates in our sample are 30.01% and 6.37% respectively, much higher than

the old cutoff (21.47% and 2.52% respectively). Column (7) reports the estimates using

the new standards. The effect on overweight is very similar to the baseline model.

However, we find a much larger (around 2.5 times) and more statistically significant

effect on obesity compared to the baseline model. This is because more individuals are

included in the new obese group. When we convert the estimates from percentage

points (0.50 percentage points) to percent of the mean (7.85%), it is comparable to the

old cutoff (7.54%).

Our BMI measure is derived from body weight and height. In columns (8) and

(9), we directly estimate the effect of air pollution on body weight and height separately.

As expected, we find significantly positive effect of PM2.5 on body weight. Specifically,

a 1 µg/m3 increase in PM2.5 concentrations increases the body weight by 0.1858 kg, or

0.32% (mean=58.03 kg). On the contrary, the effect on body height is statistically

insignificant and close to zero. This serves as a falsification test because our sample

only includes adults (age>=18) and their body heights are relatively constant across

years.

Table A1 in the Online Appendix reports the estimates under different

clustering of standard errors. Column (1) is the baseline model, which we employ the

two-way clustering and cluster the standard errors at individual level and county-year-

month level. In column (2), we keep the individual clustering and change the county-

year-month clustering to county-year clustering, which is more demanding. The KP F-

statistic is smaller but the statistical significance of the estimated coefficients changes

23

little. In column (3), we further aggregate the clustering level from county-year to

county. This controls for any autocorrelation within each county. The KP F-statistic is

much lower but the estimated coefficients for BMI and overweight are still significant

at the 5% level. In column (4), we still employ the two-way clustering but cluster at

county and year level. The estimated coefficient is still significant at the 5% level for

overweight, but only at 10% level for BMI and obesity. In the last two columns, we

employ the one-way clustering and cluster at county-year and county level respectively.

The results are mostly significant at the 5% level.

5.4 Alternative exposure windows

In this section, we explore the effects of different exposure windows. Our

baseline model uses 12-month exposure window, i.e., we use the average

concentrations of PM2.5 in the past 12 months including the month of the interview. In

Figure 4, we vary the exposure window from past one month to past 18 months.

Dependent variables are BMI in Panel A, and overweight and obesity in Panels B and

C. The point estimates are denoted by dots and the 95% confidence intervals are

denoted by whiskers.

The effects are precisely estimated for the one to three-month exposure

windows. However, the effects are close to zero and statistically insignificant at the 5%

level. When the exposure windows extend from four to nine months, the standard errors

are very large. This is mainly because the weak first stage.9 Most interviews are

9 Table A2 in the Online Appendix reports the first-stage estimates as well as the KP F-statistics for

these exposure windows.

24

conducted from September to November (Figure A3 in the Online Appendix), and the

occurrences of thermal inversions are less in the summer (Figure A4 in the Online

Appendix).

When we further extend the exposure windows to ten to 18 months, the standard

errors are smaller, and the coefficients are statistically significant for the exposure

window of 11-13 months at the 5% level, but not for 14-18 months. The magnitude is

the largest for the 11 and 12-month exposure windows. We thus choose the 12-month

exposure window as the baseline because it is mostly straightforward.

Since the first stage in several months are weak, we cannot determine the exact

time window that air pollution will have an impact on body weight. However, because

the estimates are precisely estimated in the first three months but the effect is close to

zero, we should be able to cautiously conclude that the pollution effect is not

contemporaneous, or at least not within the three months.

5.5 Heterogeneity analysis

Table 5 reports the heterogeneous estimates. We start with gender in columns

(1) and (2) first. In general, female has higher BMI, overweight and obesity rates than

male. The effects of PM2.5 on three body weight measures are higher than female.

However, given the standard errors, the difference is statistically insignificant.

Columns (3) and (4) report the subsample analysis by age. The elderly aged

above 60 accounts for 20% of total observations, and has higher BMI, overweight and

25

obesity rates than adults aged below 60. However, the effects of PM2.5 are statistically

insignificant for the elderly. The magnitudes are much smaller for overweight and

obesity compared with adults aged below 60. This suggests that behavior responses

such as exercising may play a role for young adults and are more likely to be affected

by air pollution. It could also be the case that the exposure is larger for young adults

since they need to work or commute outdoors.

We next explore the heterogeneity by education. Since the survey covered a

significant amount of rural residents, and the survey started in 1989, when China was

in the early stage of development, the average education level in the sample is low.

Table A3 in the Online Appendix reports the sample size and the average BMI,

overweight and obesity rates by seven educational groups. Nearly 30% of observations

received zero education. Around 20% and 30% of observations only graduated from

primary school and junior high school respectively. Less than 20% of observations

graduated from high school of above.

Researchers typically find that people with higher education tend to be in better

health, and obesity is no exception (Grossman, 2006; Cawley, 2015). However, this

seems not the case for China. Overall, the BMI as well as overweight and obesity rates

are higher in highly educated group. This may be because those highly educated

population are more likely to be rich, and the income-obesity gradient in China is still

in the increasing part of the inverted U-shaped relationship.

Because the small size of each seven educational group, we merge them into

two groups: low educated with zero or primary-school education, and high educated

26

with senior high-school education or above. Columns (5) and (6) report the effects of

PM2.5 on body weight for each group separately. The impacts are larger for highly

educated group, especially for the impact on overweight. This heterogeneity could be

caused by behavior responses to air pollution. For example, highly educated group may

spend more time on exercise. If this is the case, a positive pollution shock may reduce

the exercise time because they have more information on the harmful effect of air

pollution. Therefore, the effect of air pollution is larger on body weight for highly

educated group because they will exercise less.

The last two columns explore the heterogeneity by urban and rural residency.

Rural individuals account for 55% of observations and have lower BMI and overweight

and obesity rates. The effects of PM2.5 on body weight, however, do not differ much

between urban and rural residents.

5.6 Mechanism tests

Section 2 discussed several mechanisms through which air pollution affects

body weight. In this section, we test several behavior channels, including physical and

sedentary activity, food intake, and bed time. However, the survey only asked the

respondents to report their behaviors in a very short time window, such as past three

days or a week. On the other hand, we find that the effect of air pollution on body

weight is at least beyond the three-month exposure window, as discussed in Section 5.4.

Therefore, we fail to detect any meaningful short-run behavior responses. Nevertheless,

we report the results in Table 6.

27

We start with food intake. The CHNS recorded detailed information on total

calories consumed and by different categories during past three days. Because our

pollution data are at monthly level, we use the PM2.5 in the month of the interview. The

dependent variable is total calories consumed. Although the sign is positive, i.e., high

pollution increases calories intake, it is very imprecisely estimated. Through columns

(2)-(4), the dependent variables are nutrition intake for each category, including

carbohydrate, fat, and protein. We find suggestive evidence that air pollution increases

fat and protein intake but reduces carbohydrate intake.

We then test the effect of PM2.5 on physical and sedentary activities. The survey

asked how many minutes did a respondent spend in a typical week. Physical activity

includes martial arts such as Kung Fu, gymnastics/dancing/acrobatics,

running/swimming, soccer/basketball/tennis, badminton/volleyball, and others such as

table tennis and Tai Chi. Because more than a half of the individuals are rural residents,

there are a lot zero values for the physical activity. On the other hand, sedentary

activities include watching TV, watching videotapes/VCD/DVD, playing video games,

surfing the internet, online chatting, playing computer games, and

reading/writing/drawing. Columns (5) and (6) report the estimates. We do not find any

statistically significant impact.

Lastly, we estimate the effect of PM2.5 on bed time. The survey asked how much

time each day do you usually spend in bed sleeping or lying there, including nighttime.

We again fail to detect any significant effect. In the above estimates, we use the PM2.5

concentrations in the month of the interview. We also tried the average PM2.5

28

concentrations in the past 12 months, but do not find any significant effect. We also

tried the reduced-form estimate, i.e., thermal inversions on each behavior responses,

and use the corresponding exposure window, i.e., past three days for nutrition intake,

past seven days for physical and sedentary activity, and one day for bed time. Again,

we fail to find any meaningful effect.

6 Discussion

We find a significantly positive effect of air pollution on the BMI, overweight

and obesity rates. In this section, we compare our estimates with two strands of the

literature: those on estimating the causes of obesity and those on estimating the

economic cost of air pollution. In the last two subsections, we discuss policy

implications and research caveats as well as future direction.

6.1 Comparison with the literature on estimating the causes of

obesity

In the past several decades, the prevalence of overweight and obesity has

increased significantly in the U.S. and other developed countries (National Center for

Health Statistics, 2014; OECD, 2014). Therefore, economists have devoted significant

amount of research on understanding the economic causes of obesity (see Cawley (2015)

for a literature review).

First, we compare our estimates with Currie et al. (2010), who estimate the

effect of fast food restaurants on obesity in the U.S. They find that the presence of a

29

fast food restaurant within 0.1 miles of a school increases the obesity rates by 5.2

percent for the ninth graders. This effect is similar to the increase of average PM2.5

concentrations by 1 µg/m3, as we find that a 1 µg/m3 increases the obesity rates by 0.19

percentage points, or 7.54 percent in our study.

Second, we compare our estimates with Brunello et al. (203), who investigate

the effect of education on obesity. They find insignificant impact of schooling on

obesity for male. However, the effect is significant for female. Specifically, a 1

additional year of schooling reduces the prevalence of obesity by 14.83 percent for

women. As we do not find significantly gender difference in response to air pollution,

we use our estimate for the whole sample. Therefore, we can conclude that a 1

additional year of schooling has similar effect with a decrease of PM2.5 concentrations

by around 2 (14.83/7.54) µg/m3.

Third, we compare our estimates with Rhum (2005), who estimates the effect

of employment rate on obesity. The author finds that a 1% increase in employment rate

increases obesity rate by 40%, which is equivalent to 5.31 (40/7.54) µg/m3 increase in

PM2.5 concentrations in our study. In contrast to Rhum’s counter-cyclical findings on

obesity, Charles and DeCicca (2008) find a pro-cyclical effect. In other words, they

find that increased unemployment rate increases obesity rates. Specifically, they find

that a 1% increase in employment rate increases the obesity rate by 2.19 percentage

points for African-American men. This is equivalent to 11.53 µg/m3 (2.19/0.19)

increase in PM2.5 concentrations in our study. The effect, however, is statistically

insignificant for Hispanic and White men.

30

Lastly, we focus on peer and neighborhood effects. Using the Moving to

Opportunity program as an experiment, Kling et al. (2007) find that moving to a low-

poverty neighborhood reduces the probability of obesity by 4.8 percentage points

relative to the control group. This is equivalent to reduce PM2.5 concentrations by 25.26

µg/m3.

In summary, we show that the impact of air pollution on obesity is meaningful

and comparable to other economic causes. In particular, a 1 µg/m3 increase in average

PM2.5 concentrations in the past 12 months has the similar effect with an additional fast

food restaurant within 0.1 miles of a school, a half of an additional year of schooling,

and 0.19% increase in employment rate (or 0.09% increase in unemployment rate).

6.2 Comparison with the literature on estimating the economic cost

of air pollution

Overweight and obesity could lead to a variety of chronic diseases such as

diabetes, cardiovascular and kidney diseases, and some cancers, and therefore

significantly attribute to the medical cost. To shed light on the economic cost of air

pollution on overweight and obesity, we perform a back-of-the-envelope calculation

using the increase in overweight and obese population in response to a 1 µg/m3 increase

in PM2.5 concentrations multiply by the per-capita health expenditure attributable to

overweight and obesity.

Our estimate suggests that 1 µg/m3 increase in average PM2.5 concentrations

induces additional 10.24 million overweight and obese (BMI>=25) adult populations.

31

Qin and Pan (2015) estimate that overweight and obesity account for 5.29% of total

personal health expenditure in China during 2000-2009. In 2016, the per-capita health

expenditure in China was CNY 3,784 (China Statistical Yearbook, 2017), and thus the

overweight/obesity-related health expenditure per capita was CNY 200. Therefore, we

can conclude that 1 µg/m3 increase in average PM2.5 concentrations induces a total of

CNY 1.91 billion (USD 0.29 billion) health expenditure on overweight and obesity.

We compare our estimates with previous studies that estimate the effect of PM2.5

on other economic variables. Deryugina et al. (2016) find that a 1 µg/m3 decrease in

PM2.5 brings an annual benefit of USD 4.11 billion in terms of avoided mortality in the

U.S., which is 14 times larger than our estimate. Fu et al. (2017) and Chang et al. (2019)

find that 1 µg/m3 decrease in PM2.5 increases labor productivity in China by USD 2.99

billion and in the U.S. by USD 6.99 annually, which is 10 and 24 times larger than our

estimate respectively.

In terms of mental health, Chen et al. (2018) find that 1 µg/m3 decrease in PM2.5

reduces the medical expenditure on mental illness by USD 0.78 billion annually in

China, which is around 3 times larger than our estimate. On the other hand, Bishop et

al. (2018) focus on dementia and find that reducing PM2.5 by 1 µg/m3 saves medical

expenditure on dementia by USD 3.5-10.5 billion per year, which is 12-36 times larger

than our estimate.

To sum up, our study suggests that the cost of air pollution on overweight and

obesity are non-trivial. Although the magnitude is smaller than previous studies focused

on other economic variables, it is in the same order of scale. Note we may underestimate

32

the cost for two reasons. First, we only focus on the medical cost, but researchers find

that obesity has wide impacts on economic outcomes, including wages (Cawley, 2004)

and employment (Rooth, 2009). Second, the estimated percent of medical cost

attributable to overweight and obesity in Qin and Pan (2015) was calculated during the

period of 2000-2009. Since the prevalence of overweight and obesity is expected to

increase in China, the related medical cost will also be likely to increase in the future.

6.3 Policy implications

Many developing countries have remarkably poor air quality, which is often

considered as the first-order obstacles to economic development. In China, Premier Li

Keqiang has declared “The War against Air Pollution” and many acts and regulations

have been made to reduce air pollution. For example, the Air Pollution Prevention and

Control Action Plan made in 2012 intends to reduce the PM2.5 concentrations by 25%,

20%, and 15% for three economic centers: the Jing-Jin-Ji (Beijing, Tianji, and Hebei)

area, the Yangtze River Delta (Shanghai, Zhejiang, and Jiangsu), and the Pearl River

Delta (Guangdong) respectively by 2017. Based on our estimates, this policy will

generate large co-benefits of reducing overweight and obesity rates in China. We find

that a 1% (0.63 µg/m3) decrease in PM2.5 concentrations reduces the overweight and

obesity rates by 0.56 and 0.12 percentage points respectively. Using the mid-point of

these three goals (20%), this policy will reduces overweight and obesity rates by 11.2

and 2.4 percentage points respectively. This corresponds to a total benefit of CNY 24.07

billion (USD 3.65 billion).

33

The Chinese government began to realize the increasing prevalence of

overweight and obesity and the associated economic burden in China, and therefore

made several policies on obesity prevention and control. For example, in 2003, the

Bureau of Disease Control affiliated in the Ministry of Health made the Guidelines for

Prevention and Control of Overweight and Obesity of Chinese Adults. In 2007, the

Bureau of Disease Control made the guidelines for school-age children and teenagers.

In 2013, the nutrient information should be included on labels. Our study, on the other

hand, shows that reducing air pollution could be an important and effective strategy to

reduce overweight and obesity in China.

6.4 Caveats and future research direction

Two caveats exist in our study. First, due to the research design, i.e., using

thermal inversions as the IV for PM2.5, we cannot identify the effect of PM2.5 per se,

because air pollutants are highly correlated, and thermal inversions could also affect

other air pollutants, such as PM10, CO, and O3 (Arceo et al., 2016). Therefore, it is better

to illustrate our estimates as the effect of air pollution, instead of PM2.5 per se, on body

weight.

Second, although we believe our causal estimates of air pollution on body

weight are convincing, the mechanism tests are only suggestive. This is mainly because

the data on several behavior responses such as exercising, dieting, and sleeping were

only tracked in a short of time. Therefore, we cannot isolate the behavior channel from

34

the physiological channel. Future research should utilize better data on behavior

responses.

Although our focus is China, our methods are general and could be applied to

other countries. In fact, it is not clear whether air pollution will affect body weight in a

different context, e.g., for a developed countries. Even so, it remains unknown about

the magnitude, which may differ because the exposure to air pollution, the behavior and

biological responses are different across countries. We leave this for future research.

Figure 1. Time Trend of PM2.5, GDP, and Body Weight in China during 1989-2011

Notes: This figure shows the time trend of average PM2.5 concentrations and GDP (Panel A),

average BMI (Panel B), overweight (Panel C), and obesity rates (Panel D) in China during

1989-2011.

Figure 2. Time Trend of Thermal Inversions and PM2.5

Notes: This figure plots the average PM2.5 concentrations and average monthly cumulative

thermal inversion occurrences in China over 1989-2011.

Figure 3. Histogram of BMI

Notes: This figure plots the histogram of the BMI. The vertical red line indicates the cutoff of

25, which is used to define overweight. The vertical blue line indicates the cutoff of 30, which

is used to define obesity.

Figure 4. Impact of PM2.5 on Body Weight

Notes: This figure depicts the impacts of PM2.5 on the BMI (Panel A), overweight (Panel B),

and obesity (Panel C). PM2.5 is calculated using average concentrations from the past 18

months to the month of the interview. The circle denotes the point estimate, and the whisker

denotes the 95% confidence intervals.

Table 1. Summary Statistics

Variable Description N Mean SD Min Max

Body mass measure

BMI weight(kg)/height(m)^2 60,056 22.62 3.30 4.83 63.78

Overweight BMI>=25 60,056 21.47 41.06 0.00 100.00

Obesity BMI>=30 60,056 2.52 15.69 0.00 100.00

Weight kg 60,056 58.03 10.56 15.00 162.70

Height cm 60,056 159.92 8.42 105.00 188.00

Male

BMI weight(kg)/height(m)^2 28,548 22.45 3.16 4.83 57.01

Overweight BMI>=25 28,548 19.53 39.64 0.00 100.00

Obesity BMI>=30 28,548 1.93 13.76 0.00 100.00

Weight kg 28,548 61.83 10.48 15.00 162.70

Height cm 28,548 165.73 6.48 105.00 188.00

Female

BMI weight(kg)/height(m)^2 31,508 22.78 3.41 9.03 63.78

Overweight BMI>=25 31,508 23.23 42.23 0.00 100.00

Obesity BMI>=30 31,508 3.06 17.23 0.00 100.00

Weight kg 31,508 54.60 9.40 24.00 156.80

Height cm 31,508 154.66 6.23 120.60 179.00

Air pollution

PM2.5 µg/m3 60,056 63.00 26.19 23.75 141.32

Thermal inversions

inversions Number if inversion times 60,056 22.46 9.74 5.92 48.17

Notes: N=60,056. Unit of observation is individual-year. The survey covered 71 counties

during the period of August 31st, 1989 to December 11th, 2011. The interview surveyed the

measured weight (kg) and height (cm) on the interview day for the adult population (age ≥18).

The BMI is calculated as the weight (kg) divided by squared height (m). Overweight is a

dummy variable which equals to one if BMI is above or equal to 25. Obesity is a dummy

variable which equals to one if BMI is above or equal to 30. Average PM2.5 concentrations are

on the month level. The existence of a thermal inversion is determined within each six-hour

period, and then aggregated to the month level.

Table 2. First-stage Estimation: Effects of Thermal Inversions on PM2.5

PM2.5

(1) (2) (3)

Thermal inversions 0.3035*** 0.3010*** 0.3392***

(0.0767) (0.0769) (0.0742)

Individual FE Yes Yes Yes

Year FE Yes No No

Year-by-month FE No Yes Yes

Weather controls No No Yes

KP F-statistics 15.66 15.31 20.92

Notes: N=60,056. The dependent variable is PM2.5. Weather controls include 5 °C temperature

bins, second-order polynomials in average relative humidity, wind speed, sunshine duration,

and cumulative precipitation. Standard errors are listed in parentheses and clustered at

individual level and county-year-month level. *** p<0.01, ** p<0.05, * p<0.1.

Table 3. Second-stage Estimation: Effects of Air Pollution on Body Mass

BMI

Overweight

Obesity

(1) (2) (3) (4) (5) (6)

Panel A: IV

PM 2.5 0.0473** 0.0691***

0.0069** 0.0089***

0.0016 0.0019*

(0.0226) (0.0249)

(0.0030) (0.0034)

(0.0010) (0.0011)

KP F-statistics 21.69 20.92 21.69 20.92 21.69 20.92

Panel B: OLS

PM 2.5 0.0001 -0.0006

-0.0003 -0.0004

-0.0003 -0.0003

(0.0039) (0.0037)

(0.0005) (0.0005)

(0.0002) (0.0002)

Individual FE Yes Yes

Yes Yes

Yes Yes

Year FE Yes No

Yes No

Yes No

Year-by-month FE No Yes

No Yes

No Yes

Weather Controls Yes Yes Yes Yes Yes Yes

Notes: N=60,056. The dependent variables are the BMI in columns (1)-(2), overweight in

columns (3)-(4), and obesity in columns (5)-(6). Panel A is the IV estimate, in which we use

number of thermal inversions as an instrument for PM2.5. Panel B is the OLS estimate.

Weather controls include 5 °C temperature bins, second-order polynomials in average relative

humidity, wind speed, and sunshine duration, and cumulative precipitation. Standard errors

are listed in parentheses and clustered by individual and county-year-month. *** p<0.01, **

p<0.05, * p<0.1.

Table 4. Robustness Checks

Baseline

Date FE

No

Weather

Controls

Not

include

current

month

IV:

inversion

Times

(D2)

Drop

Lower

0.5% and

Upper

0.5%

Cutoff

change

Body

Weight

Body

Height

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Panel A: BMI

PM 2.5 0.0691***

0.1082**

0.0851***

0.0769**

0.0655***

0.0723***

-

0.1858***

0.0005

(0.0249)

(0.0439)

(0.0302)

(0.0312)

(0.0235)

(0.0248)

-

(0.0649)

(0.0318)

Panel B: Overweight

PM 2.5 0.0089***

0.0146**

0.0119***

0.0085**

0.0096***

0.0090***

0.0085**

-

-

(0.0034)

(0.0062)

(0.0043)

(0.0035)

(0.0031)

(0.0034)

-0.0035

-

-

Panel C: Obesity

PM 2.5 0.0019*

0.0031*

0.0023*

0.0025*

0.0013

0.0022**

0.0050**

-

-

(0.0011)

(0.0019)

(0.0012)

(0.0014)

(0.0010)

(0.0011)

(0.0021)

-

-

KP

F-statistics 20.92

12.8

15.31

13.33

25.51

21.04

20.92

20.92

20.92

Notes: N=60,056. The dependent variables are BMI in Panel A, a dummy variable for overweight in Panel B, and a dummy variable for obesity in Panel C.

Column (1) is the baseline model and uses year-by-month fixed effect. Column (2) replaces year-by-month fixed effects with date fixed effects. Column (3)

tests the robustness by excluding weather controls. Column (4) drops the current month to construct the exposure window. Column (5) changes the IV by

coding thermal inversions with the temperature difference between the first (110 meters) and the third layers (540 meters). Column (6) tests the robustness of

excluding extreme values of BMI by dropping the top and bottom 0.5% observations. Column (7) reports the estimates with new BMI cutoff of 24 for

overweight and 28 for obesity. Columns (8)-(9) estimate the effect of air pollution on body weight and height separately. *** p<0.01, ** p<0.05, * p<0.1.

Table 5. Effect of Air Pollution on Body Mass : By Gender, Age, Education, and Residence

(1) (2) (3) (4) (5) (6) (7) (8)

Male Female Below 60 Above 60 Low edu High edu Urban Rural

Panel A: BMI

PM 2.5

0.0744*** 0.0656**

0.0614** 0.0511

0.0647** 0.0822**

0.0534** 0.0581**

(0.0281) (0.0284)

(0.0239) (0.0372)

(0.0251) (0.0371)

(0.0271) (0.0249)

Mean of Dep.Var

22.4462 22.7788

22.5881 22.6792

22.5347 22.6911

23.0131 22.2985

S.D. of Dep. Var. 3.1566 3.4148 3.1592 3.7403 3.3482 3.2392 3.4121 3.1665

Panel B: Overweight

PM 2.5

0.0114** 0.0067*

0.0098*** 0.0017

0.0051 0.0171**

0.0080* 0.0063*

(0.0044) (0.0037)

(0.0036) (0.0046)

(0.0032) (0.0069)

(0.0042) (0.0034)

Mean of Dep.Var

0.1953 0.2323

0.2050 0.2434

0.2082 0.2189

0.2579 0.1792

S.D. of Dep. Var. 0.3964 0.4223 0.4037 0.4292 0.4060 0.4135 0.4375 0.3835

Panel C: Obesity

PM 2.5

0.0024* 0.0015

0.0022* 0.0001

0.0016 0.0032

0.0015 0.0020*

(0.0013) (0.0014)

(0.0012) (0.0024)

(0.0011) (0.0022)

(0.0016) (0.0010)

Mean of Dep.Var

0.0193 0.0306

0.0219 0.0367

0.0268 0.0231

0.0315 0.0201

S.D. of Dep. Var. 0.1376 0.1723 0.1463 0.1881 0.1614 0.1504 0.1748 0.1403

Observations

28,548 31,508

46,900 12,090

29,239 28,516

27,077 32,979

KP F-statistics 19.89 21.32 22.33 25.23 29.58 12.09 24.26 17.78

Notes: The dependent variables are BMI in Panel A, a dummy variable for overweight in Panel B, and a dummy variable for obesity in Panel C. Regression

models are estimated separately for each subsample. Weather controls include 5 °C temperature bins, second-order polynomials in average relative humidity,

wind speed, and sunshine duration, and cumulative precipitation. Standard errors are listed in parentheses and clustered by individual and county-year-month.

*** p<0.01, ** p<0.05, * p<0.1.

Table 6. Air Pollution Effects on Nutrition Intake and Activities

Energy

(past 3

days)

Carbo

(past 3

days)

Fat

(past 3

days)

Protein

(past 3

days)

Physical

Activity

(typical

week)

Sedentary

Activity

(typical

week)

Bed time

(typical

day)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

kcal g g g mins mins hours

PM2.5 Recent 1 mo. 2.7920

-0.0198

0.2133

0.1246

0.2579

-0.6880

-0.0031

(9.9628)

(1.6880)

(0.5161)

(0.3680)

(0.2537)

(1.9032)

(0.0114)

Mean of Dep.Var 2309.1414

345.6785

69.5265

69.4988

8.9424

287.8281

8.0289

S.D. of Dep. Var. 758.0885 132.8820 46.3892 25.6850 46.3247 220.7902 1.2891

Observations 54,929

54,929

54,929

54,929

36,693

24,437

25,460

Number of idind 12,633

12,633

12,633

12,633

9,801

8,119

8,351

KP F-statistics 12.07 12.07 12.07 12.07 16 15.13 14.59

Notes: The dependent variables are average energy (kcal) over past 3 days in column (1), average carbohydrate (g) intake over past 3 days in column (2),

average fat (g) intake over past 3 days in column (3), average protein (g) intake over past 3 days in column (4), total minutes of physical activities per week in

column (5), total minutes of sedentary activities per week in column (6), and total hours at bed per day in column (7). Physical activity includes martial arts

such as Kung Fu, gymnastics/dancing/acrobatics, running/swimming, soccer/basketball/tennis, badminton/volleyball, and others such as table tennis and Tai

Chi. sedentary activities include watching TV, watching videotapes/VCD/DVD, playing video games, surfing the internet, online chatting, playing computer

games, and reading/writing/drawing. PM2.5 is in the month of the interview and instrumented using thermal inversion in the same exposure window. Weather

controls include 5 °C temperature bins, second-order polynomials in average relative humidity, wind speed, and sunshine duration, and cumulative

precipitation. . Standard errors are listed in parentheses and clustered by individual and county-year-month. *** p<0.01, ** p<0.05, * p<0.1.

Online Appendix

Figure A1. Number of Interviewers by Interviewed Year 0

2,0

00

4,0

00

6,0

00

8,0

00

Num

ber

of

Inte

rvie

wers

1989 1991 1993 1997 2000 2004 2006 2009 2011

Notes: This figure plots the number of interviewers in each survey year (1989-2011).

Figure A2. Frequency of Observation

0

1,0

00

2,0

00

3,0

00

4,0

00

# o

f in

terv

iew

ers

2 3 4 5 6 7 8 9

Notes: This figure plots the number of interviewers with 2-9 observations over 1989-2011.

Figure A3. Number of Interviewers by Interviewed Month 0

5,0

00

10

,000

15

,000

20

,000

25

,000

Num

ber

of

Inte

rvie

wers

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Notes: This figure plots the number of interviewers in each month over 1989-2011.

Figure A4. Average of Thermal Inversion Times by Interviewed Month

0

10

20

30

The

rmal in

ve

rsio

ns a

ve

rag

e

1 2 3 4 5 6 7 8 9 10 11 12

Notes: This figure plots the average number of thermal inversions for all 71 counties for each

month from 1989 to 2011.

Table A1. Air Pollution Impacts on Body Mass by Different Cluster Ways

Two-way

County-year-mon

th Individual

Two-way

County-year

Individual

Two-way

County

Individual

Two-way

County

Year

One-way

Count-year

One-way

County

(1) (2) (3) (4) (5) (6)

Panel A: BMI

PM 2.5 0.0691*** 0.0691*** 0.0691** 0.0691* 0.0691** 0.0691**

(0.0249) (0.0254) (0.0329) (0.0319) (0.0276) (0.0329)

Panel B: Overweight

PM 2.5 0.0089*** 0.0089** 0.0089** 0.0089** 0.0089** 0.0089**

(0.0034) (0.0035) (0.0040) (0.0036) (0.0037) (0.0040)

Panel C: Obesity

PM 2.5 0.0019* 0.0019* 0.0019* 0.0019* 0.0019* 0.0019*

(0.0011) (0.0011) (0.0011) (0.0009) (0.0011) (0.0011)

KP F-statistics 20.92 18.50 9.400 7.896 14.53 9.400

Notes: N=60,056. The dependent variables are BMI in Panel A, a dummy variable for Overweight in Panel B, and a dummy variable for Obesity in Panel C.

Column (1) is the baseline model and uses the two-way clustering. The standard errors are clustered at individual and county-year-month level. Column (2)

also employs the two-way clustering, and standard errors are clustered at individual and county-year-month level. In Column (3), standard errors are clustered

at individual and county level. Column (5) and Column (6) use one-way clustering and standard errors are clustered at county-year and county level

respectively. All models include individual fixed effects, year-by-month fixed effects, and weather controls. Weather controls include 5 °C temperature bins,

second-order polynomials in average relative humidity, wind speed, and sunshine duration, and cumulative precipitation. *** p<0.01, ** p<0.05, * p<0.1.

Table A2: First-stage Estimation in alternative window: Effects of Thermal Inversions on PM 2.5

Current

month

Average

recent

2month

Average

recent

3month

Average

recent

4month

Average

recent

5month

Average

recent

6month

(1) (2) (3) (4) (5) (6)

Panel A: Recent 1-6 month

Thermal inversions 0.1093*** 0.1374*** 0.1130*** 0.0672* 0.0499 0.0639

(0.0344) (0.0459) (0.0421) (0.0391) (0.0423) (0.0466)

KP F-statistics 10.12 8.973 7.214 2.951 1.393 1.881

Average

recent

7month

Average

recent

8month

Average

recent

9month

Average

recent

10month

Average

recent

11month

Average

recent

12month

(7) (8) (9) (10) (11) (12)

Panel B: Recent 7-12 month

Thermal inversions 0.1258** 0.1146* 0.1138* 0.1761** 0.2335*** 0.3392***

(0.0538) (0.0602) (0.0674) (0.0703) (0.0690) (0.0742)

KP F-statistics 5.475 3.622 2.854 6.276 11.46 20.92

Average

recent

13month

Average

recent

14month

Average

recent

15month

Average

recent

16month

Average

recent

17month

Average

recent

18month

(13) (14) (15) (16) (17) (18)

Panel C: Recent 13-18 month

Thermal inversions 0.3222*** 0.2855*** 0.3022*** 0.3045*** 0.2942*** 0.3539***

(0.0730) (0.0703) (0.0706) (0.0738) (0.0800) (0.0810)

KP F-statistics 19.49 16.50 18.30 17.03 13.53 19.09

Notes: N=60,056. The dependent variable is PM2.5. PM2.5 is calculated using average

concentrations of PM2.5 from the recent 1-6 months (Panel A), 7-12 months (Panel B), and

13-18 months (Panel C). Thermal inversions are in the same exposure window. Weather

controls include 5 °C temperature bins, second-order polynomials in average relative

humidity, wind speed, sunshine duration, and cumulative precipitation in the same exposure

window. Standard errors are listed in parentheses and clustered by individual and

county-year-month. *** p<0.01, ** p<0.05, * p<0.1.

Table A3. Body Mass Mean by Education Groups

Level Category BMI Overweight Obesity N

0 No School 22.47 20.70% 2.66% 17242

1 Primary 22.64 21.12% 2.75% 12831

2 Lower Middle School 22.61 21.03% 2.33% 17998

3 Upper Middle School 22.78 22.51% 2.17% 7176

4 Technical or vocational

degree 22.94 26.07% 2.82% 2447

5 College 23.10 26.05% 2.50% 1762

6 Master or Higher 23.24 27.03% 2.70% 37

Notes: There are 7 levels of highest education attained in the survey: 0(none), 1(graduate from primary school), 2(lower middle school degree), 3(upper

middle school degree), 4(technical or vocational degree), 5(university or college degree), and 6(master’s degree or higher).