Air Pollution, Labor Productivity and Individual

ed
Air Pollution, Labor Productivity and Individual

Consumption
iew
[Preliminary version: 12-28-2021]
Shuai Chen Min Wang Dandan Zhang
v
Abstract
re
This paper combines a tracking record of prison inmates’ monthly consumption and
labor productivity with daily air pollution data to examine the effects of air pollution
on consumption decisions through the substitution and income channels. A 10-unit
er
increase in the average daily air pollution index is found to reduce inmates’ monthly
consumption expenditures by 3.6% and the income channel explains 78% of the effects.
Such effects are heterogeneous across demographic groups, consumption goods and
pe
different levels of air pollution. The paper extends the growing literature on
environmental effects on productivity and consumption by integrating the two in a
quasi-experimental setting.
ot
Key word: Air pollution; labor productivity; consumption; inmates
JEL Codes: Q51, Q53, I15, O15.

tn
rin
 Shuai Chen, China Academy for Rural Development (CARD) and School of Public Affairs,
Zhejiang University, 310058, China; Email: shuaichen@zju.edu.cn. Min Wang, China Center for
ep
Economic Research, National School of Development, Peking University, Beijing, 100871, China;
Email: wangmin@nsd.pku.edu.cn. Dandan Zhang, China Center for Economic Research, National
School of Development, Peking University, Beijing, 100871, China; Email:
ddzhang@nsd.pku.edu.cn. Acknowledgement: Generous financial support was given by the
Pr
National Natural Science Foundation of China (71973003; 71973004; 20AZD086) and the National
Key R&D Program of China (2020YFA0608600); Ethics approval was obtained from the Australian
National University (2013/494).
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4014350
ed
Air Pollution, Labor Productivity and Individual
Consumption
iew
[Preliminary version: 12-28-2021]
Abstract
v
re
This paper combines a tracking record of prison inmates’ monthly consumption and
labor productivity with daily air pollution data to examine the effects of air pollution
on consumption decisions through the substitution and income channels. A 10-unit
increase in the average daily air pollution index is found to reduce inmates’ monthly
consumption expenditures by 3.6% and the income channel explains 78% of the effects.
er
Such effects are heterogeneous across demographic groups, consumption goods and
different levels of air pollution. The paper extends the growing literature on
environmental effects on productivity and consumption by integrating the two in a
pe
quasi-experimental setting.
Key word: Air pollution; labor productivity; consumption; inmates

ot
JEL Codes: Q51, Q53, I15, O15.

tn
rin
ep
Pr
1. Introduction
ed
That a country’s environment – the air its people breathe, the water they drink,
and so on – will have a profound effect on its overall economic performance, hardly
iew
seems surprising. A growing body of literature has discovered consistent evidence of
pollution’s damaging effects on human health.1 This direct health effect is undoubtedly
v
of primary importance in the overall impacts of pollution. Moreover, it is normal to
re
expect that this pollution-associated impairment in human health, by negatively
impacting the lives of people and their daily production and consumption activities,
er
prevents them from being at their productive best.
Recent studies have found that pollution detrimentally affects the production of
pe
individuals both by reducing their labor supply at the extensive margin (Carson et al.,
2011; Hanna and Oliva, 2015; Viard and Fu, 2015) and their labor productivity at the
ot
intensive margin (Graff Zivin and Neidell, 2012; Fu et al., 2021; Chang et al., 2018;
2019; Archsmith et al., 2018; He et al., 2019). However, due to data limitations, the
tn
impacts of pollution on individual consumption have received limited attention.
Pollution’s effects on consumption also warrant attention. Consumption is both

rin
the major determinant of individual welfare and, at a macro level, is the largest
component of GDP in most economies. In this paper, we use high-frequency data with
ep
1 The economics literature on environmental health effects has blossomed in the last two decades.
See Chay and Greenstone (2003), Currie and Neidell (2005), Ebenstein (2012), Ebenstein et al.
(2012), Anderson (2020), Chen et al. (2013), Greenstone and Hanna (2014), Schlenker and Walker
(2016), Arceo et al. (2016), Knittel et al. (2016), Deschenes et al. (2017), Deryugina et al. (2019),
Pr
and He et al. (2020), for example.
fully recorded information on individual consumption and production in a prison
ed
environment to examine whether air pollution alters individuals’ consumption
decisions, and the channels through which air pollution may induce such changes in
iew
consumption behavior.
The impacts of air pollution on an individual’s consumption decisions can be split
into two broad divisions. It has been widely shown that air pollution can increase
v
defensive expenditures, i.e., compensatory investments in health.2 Less well known are
re
air pollution’s effects on the consumption of non-defensive and utility-generating goods.
Since an individual could have preferences related to both consumption goods and
er
health status, two environmental consumption effects could arise through the health
pe
channel.
First, since both health status and consumption of goods feature in an individual’s
utility function, a consumer good can be a substitute or complement for health status.
ot
That is, conditional on his income level, an improvement in health status will reduce or
increase a person’s demand for a particular good. We call this effect the substitution
tn
effect of air pollution on individual consumption.
Second, since air pollution affects labor supply, labor productivity and defensive
rin
expenditures, it may also alter an individual’s budget constraint. We call the impact of
such a shift on non-defensive consumption, the income effect of air pollution on

ep
2Recent studies have documented that air pollution increases medical (Deschênes et al., 2017;
Barwick et al., 2018; Deryugina et al., 2019) and preventive expenditures, such as health insurance
(Chang et al., 2018), particulate-filtering facemasks (Zhang and Mu, 2018) and home-use air
Pr
purifiers (Ito and Zhang, 2020).
individual consumption.
ed
Since air pollution can impact an individual’s daily health status in subtle and
pervasive ways (e.g., irritations of eye, ear, nose and throat, and mild headaches)
iew
without generating any defensive expenditures (Graff Zivin and Neidell, 2013), one can
expect air pollution’s impacts on individual daily consumption to be more prominent
than those on defensive expenditures. Moreover, since the scale of most individuals’
v
daily consumption is incomparably greater than their defensive expenditures3, a small
re
impact on individuals’ daily consumption can translate into a large welfare change.
However, the existing literature provides little information on the impacts of air
er
pollution on individuals’ daily consumption.4 This paper develops a theoretical model
pe
relating air pollution and individual consumption, and uses a unique, individual level
data set to identify the income, substitution and overall effects of air pollution on
individual consumption.
ot
Two clear challenges are faced by a researcher trying to quantitatively estimate
the relationship between air pollution and individual consumption. First, appropriate
tn
data is difficult to access. Such a study requires frequent and complete tracking of an
individual’s expenditures over a long period of time. In addition, to decompose the

rin
3 Using the transaction data of credit and debit cards in China, Barwick et al. (2018) shows that,
ep
relative to healthcare spending, spending on the daily necessities could be three times as large and
transactions three times as frequent.
4 Barwick et al. (2018) uses city-level transaction data of credit and debit cards in China to study
the impact of air pollution on healthcare spending. In the paper, they also use spending on daily
necessities as the comparison group and find that a temporary increase in PM2.5 significantly reduces
Pr
card transactions in daily necessities.
overall effect of air pollution on these expenditures into substitution and income effects,
ed
the individual’s labor supply and productivity need to be precisely measured.
Second, identifying the causal effect of air pollution on changes in consumption
iew
is challenging. Air pollution exposure can be an endogenous outcome of an individuals’
optimization behaviors, such as residential sorting (Chay and Greenstone, 2005; Chen
et al., 2017; Freeman et al., 2019; Khanna et al., 2021), avoidance behavior (Neidell,
v
2009; Moretti and Neidell, 2011; Graff Zivin and Neidell, 2009), and even decision
re
making (Chew et al., 2021). Traffic congestion can also be a potential confounder, as it
generates both air pollution and emotional stress for individuals (Chang et al., 2019).
er
Finally, since an individual may move around in different locations on a daily basis
pe
(Chen et al., 2020; 2021), assigning air pollution exposure to the individual based on
his general location may lead to measurement error. A final concern is the sometimes
suspicious manipulation of daily air pollution reports (Ghanem and Zhang, 2014); these
ot
may also generate measurement error.
To overcome these challenges, we use a unique dataset that contains the monthly
tn
consumption expenditure and piece-rate wage information of all male inmates
imprisoned in Shenzhen City, China, from July 2004 to August 2015. In the prison,
rin
thousands of inmates are required to engage in manual work on a daily basis, often
manufacturing parts which are then sold to private companies through prison/company
ep
agreements.5 Inmate laborers receive piece-rate wages and purchase consumption goods
Pr
5 Many countries throughout the world have adopted a policy of putting inmates to work. In the
at a grocery store inside the jail. Detailed individual monthly production and consumption
ed
information has been recorded in the administrative system since the establishment of the
prison in July 2004. In the prison, all inmates have to work on the same fixed schedule
iew
and receive a piece-rate wage once a month. The inelastic labor supply eliminates the
possible impact of air pollution on labor supply choice, a challenge complicating previous
studies. With the piece-rate pay system, the monthly wage each inmate receives also
v
perfectly measures his labor productivity. Since both production and consumption are
re
conducted in an isolated and confined space, the prison provides an ideal environment,
absent from residential sorting, avoidance behavior and daily location changes, to
er
examine the impacts of outside shocks (fluctuations of air pollution) on production and
pe
consumption decisions of individuals.
Moreover, since the prison provides free medication for the inmates who are sick,
the potential substitutions between defensive expenditures against air pollution and
ot
non-defensive consumption expenditures are also absent in our study. Hence, the
income effect of air pollution on the inmate’s daily consumption is only driven by the
tn
channel of labor productivity. Meals are free to inmates, but they have to use their own
income to buy other consumption goods, such as snacks, personal care, and cigarettes,
rin
U.S.A. in 2017, approximately a third of the prison population of 2.3 million worked (The
Economist, 2018, see https://www.economist.com/united-states/2018/04/12/how-convict-labour-
increased-inequality in details), and prisoners employed in manufacturing provided 4.2% of total
ep
U.S. manufacturing employment in 2005 (Poyker, 2019). Since most prison workers are paid at
rates below the federal minimum wage (or even unpaid in some states), the policy of forcing the
inmates to work could generate non-negligible impacts on local labor markets. Poyker (2019) shows
that “competition from cheap prison-made goods led to higher unemployment, lower labor-force
participation, and reduced wages (particularly for women) in counties that housed competing
Pr
manufacturing industries.”
from the grocery store inside the prison. They can only do so once a month, with the
ed
date fixed for each inmate. The dataset provides detailed records of quantity and price
of all goods each inmate purchased each time. The inability to choose the day on which
iew
to buy goods eliminates the endogenous exposure to air pollution that could be
potentially generated by the purchase date decision. Since the prison grocery store is
the only option for inmates, external selection is not a concern.
v
Most of the endogeneity issues mentioned in the literature are precluded as
re
concerns since inmates’ production and consumption are physically distant from the
outside world. However, one endogeneity issue may still remain: air pollution could be
er
correlated with an inmate’s wage income, the most important determinant of individual
pe
consumption. To address this potential endogeneity issue, we regress each inmate’s
total consumption expenditure on their air pollution exposure, which is instrumented
by thermal inversions. A thermal (i.e., temperature) inversion occurs when a mass of

ot
hot air sits above a mass of cold air near the ground, thereby trapping air pollutants.
After controlling for seasonal effects and related weather variables, including
tn
temperature, humidity, wind speed, sunshine duration and precipitation, which are also
likely to influence both productivity and consumption decisions, the occurrence of

rin
thermal inversions appears random and independent of economic activities; it has
therefore been widely used as a valid instrumental variable (IV) for air pollution (Jans
ep
et al., 2014; Hicks et al., 2015; Arceo et al., 2016; Fu et al., 2021; Chen et al., 2017;
2020).
Pr
Our IV estimates show that the negative effects of air pollution on the total
ed
consumption expenditure of inmates are both economically and statistically significant.
Specifically, a 10-unit increase in the average daily Air Pollution Index (API), in the
iew
entire month prior to a grocery shopping day, reduces inmates’ monthly consumption
expenditure by 3.6%. The results are robust across a variety of model specifications.
Consistent with the literature on the environmental productivity effect, we also find that,
v
conditional on the weather variables, air pollution significantly reduced inmates’
re
productivity, measured by their earnings paid under the piece rate system.
Using the consumption expenditures and earnings of inmates, a mediation

er
analysis was used to decompose air pollution’s effect on consumption expenditure into
pe
substitution and income effects. The results show that the income effect is dominant,
accounting for 78% of the overall causal effect of air pollution on individual
consumption.
ot
The impacts of changes in the API were heterogeneous, varying across the
different kinds of consumer goods, between inmates’ demographic groups, and with
tn
the base level of air pollution. More specifically, the individual consumption decision
may rationally respond to the rising air pollution by increasing spending on healthy
rin
and stress relief goods (e.g., sweets, fruits and personal care), and reducing that on
unhealthy goods (e.g., cigarettes); the consumption expenditures of younger, more

ep
educated and more productive inmates are more likely to be negatively affected by the
increased air pollution; and inmates’ consumption expenditure remains unresponsive

Pr
to air pollution until API exceeds 40, after which it declines sharply as API rises.
ed
The article’s contributions are threefold. First, it examines a novel channel
whereby air pollution alters an individual’s consumption decisions on non-defensive
iew
goods. Since the consumption effect estimated by the study is subtler and more
pervasive than standard air pollution effects, the study provides an additional rationale
for more stringent environmental policies in heavily polluted countries, as these may
v
boost aggregate consumption. Second, although previous studies have documented
re
different effects of air pollution on individual consumption, most have estimated a
partial effect, i.e., the effect on one specific defensive good. By using a complete record
er
of all goods consumed by an individual, the study can estimate not only the API’s effect
pe
on the individual’s overall consumption, but also its effect on the individual’s
consumption pattern. Finally, and most importantly, the characteristics of this data set
allow the study to estimate both the consumption and productivity effects of air
ot
pollution, and to cleanly decompose its effects on individual consumption expenditure
into substitution and income effects.

tn
The paper proceeds as follows. Section 2 presents a theoretical model that
illustrates possible channels through which air pollution can affect an individual’s
rin
consumption decisions. Section 3 and section 4 introduce the data and the empirical
strategy. Section 5 presents the regression results. Section 6 concludes the paper.
ep
Pr
2. Conceptual Framework
ed
This section uses a simple conceptual model to illustrate pollution’s effects on
inmates’ consumption decisions. Since each inmate’s labor supply is fixed, we assume
iew
a representative individual has preferences over health 𝐻 and consumption goods 𝑐,
𝑢(𝑐,𝐻), (1)
v
re
where 𝑐 = (𝑐1, 𝑐2,…,𝑐𝑛) is a vector of the consumption goods. The utility function is
assumed to be strictly concave and to satisfy the usual differentiability requirements
with the following sign assumptions for partial derivatives: 𝑢𝑖≡∂𝑢/∂𝑐𝑖 > 0, 𝑢𝑖𝑖≡∂2𝑢/∂
er
𝑐𝑖2 ≤ 0, (𝑖 = 1, 2, …, 𝑛); 𝑢𝑛 + 1≡∂𝑢/∂𝐻 > 0, 𝑢𝑛 + 1,𝑛 + 1≡∂2𝑢/∂𝐻2 ≤ 0. Without loss
pe
of generality, we assume independent demands so that 𝑢𝑖𝑗 = 0 for ≠𝑗
(𝑖,𝑗 = 1, 2, …, 𝑛), but consumption of good 𝑖 (𝑖 = 1, 2, …, 𝑛) may depend on the
consumer’s health status 𝑢𝑖,𝑛 + 1 ≠ 0. Consumption good 𝑖 is a complement to health

ot
if 𝑢𝑖,𝑛 + 1 > 0 and is a substitute for health if 𝑢𝑖,𝑛 + 1 < 0.

tn
To clearly show the channels through which pollution affects consumption, we
firstly assume away defensive expenditure and medical treatment that could improve
rin
health, and only consider a health function dependent on air pollution exposure 𝐴.
That is 𝐻 = 𝐻(𝐴). The individual faces a budget constraint of the form:

ep
𝑝𝑐 = 𝑦(𝐻) + 𝐼, (2)
where 𝑝 is the 𝑛 × 1 price vector. 𝑦(𝐻) is the individual’s income from piece work.
Pr
10
It therefore also measures labor productivity and is presumed to increase with health at
ed
a decreasing rate. 𝐼 is non-wage income, such as family transfers.
The individual’s problem is to maximize (1) through his choices of 𝑐 subject to
iew
(2). The first order conditions are:
𝑢𝑖 ‒ 𝑝𝑖 = 0 (𝑖 = 1, 2, …, 𝑛), (3)
v
𝑦 + 𝐼 ‒ 𝑝𝑐 = 0, (4)
re
where  is the Lagrange multiplier. Optimally, consumption and the Lagrange
multiplier are functions of prices, non-wage income and air pollution, 𝑐𝑖(𝑝,𝐼,𝐴) and 
(𝑝,𝐼,𝐴) (𝑖 = 1, 2, …, 𝑛).
er
Now, consider the total derivative of equation (3) with respect to 𝐴:
pe
𝑝𝑖𝑑/𝑑𝐴 ‒ 𝑢𝑖,𝑛 + 1𝐻'
𝑑𝑐𝑖 𝑢𝑖𝑖 𝑢𝑖𝑖
𝑑𝐴
= ⏟ + ⏟ (𝑖 = 1, 2, …, 𝑛) (5)
𝑖𝑛𝑐𝑜𝑚𝑒 𝑒𝑓𝑓𝑒𝑐𝑡 𝑠𝑢𝑏𝑠𝑡𝑖𝑡𝑢𝑡𝑖𝑜𝑛 𝑒𝑓𝑓𝑒𝑐𝑡
ot
Note that 𝑢𝑖𝑖 < 0 and 𝐻' < 0. Since, by the envelope theorem, = 𝑑𝑢/𝑑𝐼 measures
tn
the marginal utility of income, the first term in equation (5) exactly reflects the income
effect of air pollution on consumption good 𝑖. As will be shown below, 𝑑/𝑑𝐴 > 0,
and therefore the income effect is negative. That is, the loss in labor productivity
rin
resulting from air pollution exposure would reduce the individual’s consumption of any
good 𝑖. The second term reflects the substitution effect of air pollution. As air pollution
ep
reduces the health level, consumption on good 𝑖 would be further depressed if it is a
complement of health, i.e., 𝑢𝑖,𝑛 + 1 > 0. This substitution effect could be positive if
Pr
11
good 𝑖 is a substitute for health, i.e., 𝑢𝑖,𝑛 + 1 < 0. Notice that the substitution effect is
ed
mainly driven by the individual’s instinctual preference which is mainly biologically
determined.
iew
Totally differentiating equation (4) with respect to 𝐴 reads
∑𝑛 𝑝 ∂𝑐 /∂𝐴 = 𝑤'𝐻' (6)

𝑗=1 𝑗 𝑗
v
By substituting (6) into (5), we have
re
𝑛
𝑑 𝑤'𝐻' 𝐻'∑𝑗 = 1𝑢𝑗,𝑛 + 1/𝑢𝑗𝑗
𝑑𝐴
= + (7)
∑𝑛 𝑝2𝑢 ∑𝑛 𝑝2𝑢
𝑗 = 1 𝑗 𝑗𝑗 𝑗 = 1 𝑗 𝑗𝑗
er
Evidently the income effect presented in equation (5) has two components. The first
pe
term in (7) is positive and represents the income effect that is generated by air pollution-
caused loss of wage income. The second term in (7) is the aggregation of the
substitution between health and all other consumption goods, which measures the
ot
change of total expenditure that is purely caused by the health-loss induced substitution
tn
effects. Giving the binding budget constraint, 𝑑/𝑑𝐴 has to be positive, i.e., the
marginal utility of income increases in air pollution. Intuitively, individual utility is
generally concave in income, so when air pollution levels are high, an individual has a
rin
lower level of utility, and a marginal increase in income could greatly increase the
individual’s welfare.
ep
The model can be further generalized by allowing health to be dependent on the
consumption of goods 𝑐, i.e., 𝐻 = 𝐻(𝑐,𝐴). Say 𝑐𝑖 is the consumption of a pure utility-

Pr
12
generating good, where ∂𝑢/∂𝑐𝑖 > 0 and ∂𝐻/∂𝑐𝑖 = 0. If 𝑐𝑖 were a pure defensive
ed
expenditure or medical treatment, ∂𝑢/∂𝑐𝑖 = 0 and ∂𝐻/∂𝑐𝑖 > 0. Since air pollution
reduces the health level, 𝑑𝑐𝑖/𝑑𝐴 > 0 for the defensive expenditure or medical
iew
treatment. Finally, 𝑐𝑖 can also be modelled as a hybrid good that not only generates
utility but also has health impacts if ∂𝑢/∂𝑐𝑖 > 0 and ∂𝐻/∂𝑐𝑖 ≠ 0. For example,
cigarette smoking, which generates utility for the consumer, also hurts his health.
v
Pollution’s impacts on the consumption of hybrid goods are evidently the summation
re
of the effects on consumption of pure utility-generating goods and pure defensive
expenditure. er
3. Data
pe
This data contains prison inmates’ consumption and piece-rate wage incomes,
ot
and the data on air pollution and other weather variables.
3.1. Tracking Records of Inmates

tn
The inmates’ data are collected in the city of Shenzhen, China. Shenzhen lies in
the Pearl River Delta, in the southern province of Guangdong, in which much of
rin
China’s manufacturing occurs, and which borders Hongkong from north. It has long
been a magnet for migrant workers and its present population of 20 million makes it
ep
the fourth largest city in China. The study can thus provide some insights into the
impacts of air pollution on the productivity and daily consumption behavior of low-
Pr
13
wage manufacturing workers in a metropolitan city.
ed
We have access to all of the administrative records of 20,334 inmates who were
imprisoned in the male prison between July 2004 and August 2015 – a total of 433,369
iew
inmate-month observations. The prison warden records each inmate’s monthly
spending in the prison store and wage income. The prison takes inmates arrested in both
Shenzhen and the neighboring city of Dongguan. The average age of inmates was 34
v
years old, amongst whom 86% held rural household registration (Hukou) before
re
incarceration. On average, they had received 9 years of education and their average
sentence length was 4.6 years. er

The inmates’ incomes include wage income from the prison factory and family
pe
transfers. Inmates work for the prison factory 6 days a week doing manual assembly
work. The final products are mainly electronic products, such as USB ports and circuit
boards. They are paid at a fixed piece rate for each unit of product produced, and receive
ot
income monthly as a deposit into a personal prison trust account. Given the piece-rate
pay system, each inmate’s monthly wage is a perfect measure of his productivity in the
tn
corresponding working period.
Inmates were allowed to purchase consumption goods using their account balance
rin
once a month. The purchase dates differ across prison districts and are exogenously
given for inmates. On the purchase day, an inmate places an order by submitting a paper
ep
listing the quantities and prices of all consumption goods. The insider grocery store
sells more than 200 consumption goods, which can be categorized into foods, personal
Pr
14
care, personal stationery, phone cards and cigarettes.6 The dataset includes information
ed
specifying who placed the order and when, as well as the quantities and prices of
ordered commodities. The data allows us to calculate each inmate’s monthly
iew
expenditure and to categorize it (see Table 2 for summary statistics).
The data also includes detailed demographic information for each inmate. This
includes age, education level, crime type, sentence length and occupation before being
v
incarcerated. The information is summarized in Figure A1 and Table A1 in the
re
appendix.
3.2. Air Pollution

er
Air pollution data is from the China National Environmental Monitoring Center
pe
(CNEMC). Before 2013, CNEMC published a daily Air Pollution Index (API) for 120
cities in China. The API ranges from zero to 500 and is divided into six levels, namely
ot
Excellent (0 – 50), Good (51 – 100), Lightly Polluted (101-150), Moderately Polluted
(151-200), Heavily Polluted (201-300) and Severely Polluted (above 300).

tn
Since January 2013, CNEMC has published a real-time hourly Air Quality Index
(AQI) and the concentrations of air pollutants including PM2.5, PM10, O3, SO2, NO2,
rin
and CO, in around 1,400 monitoring stations, 11 of which are located in Shenzhen
ep
6 Foods and personal care are the two largest choice sets for inmates. There are around 130 food
items including drinks, cookies, noodles, canned food, sweets, snacks and fruits. Personal care has
around 40 products including soap, shampoo, body wash, facial cleanser, toothpaste, toilet paper,
Pr
face towels and other personal care items.
15
City.7 To obtain an air pollution measure for the period 2013 to 2015, we use the
ed
inverse-distance weighting (IDW) method to match air pollution data from monitoring
stations to Shenzhen Men’s Prison. The IDW method is widely used in the literature
iew
when calculating pollution measures (Currie and Neidell, 2005; Deschênes and
Greenstone, 2007; Schlenker and Walker, 2016).8 The basic algorithm is to take the
weighted average of all monitoring stations within the circle with a radius for the
v
centroid of each county. We choose 100 kilometers (km) as our radius and the results
re
are robust to different radii. The computation of AQI differs from that of API. Since
Shenzhen’s daily AQI is not available before 2013, we compute daily APIs during
er
2013-2015 so that we can have a balanced and uniform measure for air pollution across
pe
the whole sample period.9
3.3. Thermal Inversions

ot
We obtained thermal inversion data from the Modern-Era Retrospective analysis
for Research and Applications version 2 (MERRA-2) released by the National

tn
Aeronautics and Space Administration (NASA) in the U.S.10 MERRA-2 divides the
7 The data can be viewed at http://106.37.208.233:20035/. One may need to install Microsoft
rin
Silverlight.
8 This method has been recently criticized by Sullivan (2017) in the context of point pollution
sources. In the context of a difference in difference design that uses opening and closing of point
sources as the source of random variation in air pollution, the interpolation created by the IDW may
smooth out sharp spatial differences in exposure, creating bias in the estimates in either direction.
However, when using thermal inversions as the source of variation for air pollution, there are no
ep
sharp spatial differences and IDW will not distort the estimates.
9 See http://kjs.mep.gov.cn/hjbhbz/bzwb/dqhjbh/jcgfffbz/201203/W020120410332725219541.pdf
for detailed AQI calculation formula and explanation.

10 The data we use is M2I6NPANA (version 5.12.4) and can be downloaded at
https://disc.sci.gsfc.nasa.gov/uui/datasets/M2I6NPANA\_V5.12.4/summary?keywords=\%22ME
Pr
RRA-2\%22\%20M2I6NPANA\&start=1920-01-01\&end=2017-01-16.
16
earth using a 0.5 × 0.625-degree grid (approximately 50 × 60-km grid), and reports the
ed
air temperature for each 42 sea-level pressure layers every six hours starting from 1980.
We aggregate every six-hour temperature for each layer in the grids located in Shenzhen
iew
City, and define a thermal inversion as occurring in a six-hour period if the temperature
in the first layer (110 meters) is lower than that in the second layer (320 meters). We
also conduct a robustness check by coding inversions using differences in temperature
v
between the first and third layers (540 meters).
re
3.4. Weather
er
The weather data are taken from the China Meteorological Data Service Center
(CMDC).11 The CMDC records daily maximum, minimum, and average temperatures,
pe
precipitation, relative humidity, wind speed, and sunshine duration for 820 weather
stations in China, one of which is located in Shenzhen City. We match Shenzhen’s daily
ot
temperature, humidity, wind speed, sunshine duration and aggregate precipitation in the
whole month prior to each inmate’s order placement date. To capture the flexible non-
tn
linearity of temperature (Deschênes and Greenstone, 2011), we discretize the daily
temperature distribution using 5°C temperature bins.

rin
3.5. Summary Statistics
Table 1 presents summary statistics of key variables. There are 433,369

ep
inmate/month observations in the sample period July 2004 to August 2015. On average,
Pr
11 The data can be seen at http://data.cma.cn/.
17
each inmate placed 16 monthly purchase orders during incarceration. The average
ed
monthly expenditure was 205 Chinese Yuan (CNY), while the average monthly piece
rate wage was 66 CNY. That means that income transfers from family provided strong
iew
support for spending by inmates inside the prison. Figure 1 plots the time trends of
nominal monthly expenditure and earnings. It is evident that both monthly expenditure
and wage income monotonically increase over time.
v
re
[Insert Figure 1 here]
Air pollution, thermal inversions and other weather variables are individual-
er
specific monthly measures constructed by averaging the daily data in the entire month
prior to the purchase date of each inmate. The mean of the individual-specific monthly
pe
average APIs was 52.8 on average, with a maximum monthly average of 106. Recall
that the occurrence of thermal inversion is calculated within every six-hour period. In
ot
an average month, there were 5.7 six-hour periods in which a thermal inversion was
recorded. Summary weather statistics for the area are presented in Table A2 of the
tn
appendix.
rin
[Insert Table 1 here]
Finally, we present the summary statistics of consumption by inmates in Table 2.

ep
The two most frequently purchased consumption goods were cigarettes and snacks, the
expenditures on which averaged 93.7 CNY and 81 CNY per month respectively. The
Pr
inmates’ expenditures on cigarettes was close to their mean expenditure on all food
18
categories, which was 135 CNY.
ed
iew
4. Empirical Strategy
The following equation is used to estimate the effect of air pollution on the
v
consumption expenditures of inmates:
re
𝐿𝑜𝑔(𝑐𝑖,𝑡(𝑖)) = 𝛽0 + 𝛽1𝑃𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜀𝑖,𝑡(𝑖), (8)
er
where 𝑐𝑖,𝑡(𝑖) is inmate 𝑖’s consumption expenditure at the order placement date 𝑡(𝑖),
𝑃𝑖,𝑡(𝑖) and 𝑊𝑖,𝑡(𝑖) are respectively the average daily API and other weather variables
pe
to which inmate 𝑖 is exposed in the entire month before date 𝑡(𝑖). Note that 𝑡(𝑖)
occurs only once a month when the inmate is allowed to purchase consumption goods.
ot
The function 𝑓(𝑊𝑖,𝑡(𝑖)) resets the daily weather measures to monthly figures, which
include temperature-day bins and second order polynomials of all other weather
tn
variables, including average daily relative humidity, average daily wind speed, average
daily sunshine duration and cumulative precipitation in the entire month before date 𝑡
rin
(𝑖). Temperature-day bins are constructed by calculating the number of days when the
daily mean temperature is in the jth of 5°C bins in the month before date 𝑡(𝑖). We
ep
include a range of weather variables that might be correlated with thermal inversions
or influence prisoner health (Arceo et al., 2016; Adhvaryu et al., 2015) and labor
Pr
productivity (Graff Zivin and Neidell, 2012; Chang et al., 2016). Our results are robust
19
to alternative weather constructs such as finer temperature bins and higher order
ed
polynomials. 𝜃𝑖 is an individual fixed effect that captures all time-invariant
characteristics of an inmate, such as preference and physique. 𝑔(𝑡(𝑖)) denotes time
iew
fixed effects of inmate i’s order placement date t, including year fixed effects, month
fixed effects and weekday fixed effects, controlling for time trends and any seasonal
patterns or weekday effects.
v
All production and consumption are taking place in an isolated and closed
re
environment; our study, therefore, excludes most endogeneity problems relating to air
pollution. However, one endogeneity issue may remain: air pollution could be
er
correlated with the inmate’s wage income. First, as suggested by the literature and our
pe
model, air pollution may depress worker productivity and therefore the inmate’s piece-
rate wage income. Second, prison production is contingent on external orders which
may be positively linked to the intensity of local economic activity, which is a source
ot
of air pollution. Taking into account the positive relationship between income and
consumption, the linear regression estimates of pollution’s impact on consumption

tn
could be upward biased by the first of these explanations and downward biased by the
second.
rin
Following the literature (Jans et al., 2014; Hicks et al., 2015; Arceo et al., 2016;
Fu et al., 2017; Chen et al., 2017; 2020), we use thermal inversion as an instrument for
ep
air pollution to solve this potential endogeneity issue. Thermal inversion can be treated
as an exogenous shock because its formation only depends on meteorological factors

Pr
20
and is independent of local economic activities and human health (Arceo et al., 2016).
ed
However, a thermal inversion phenomenon is normally influenced by the daily weather
variation (Chen et al., 2017), which should be controlled in the IV analysis.
iew
Figure 2 shows the trend of monthly API and thermal inversions in Shenzhen
between July 2004 and August 2015. The thermal inversions variable is constructed as
the number of thermal inversions to which inmate 𝑖 is exposed in the entire month
v
before date 𝑡(𝑖). Note that, since thermal inversions are recorded in six-hourly intervals,
re
the number of inversions can range from zero to four in a 24 hour day. A strong positive
correlation between thermal inversion and API can be observed in the figure. Using
er
thermal inversions as the IV for air pollution, we estimate the following generalized
pe
2SLS model:
𝑙𝑛(𝑐𝑖,𝑡(𝑖)) = 𝛽0 + 𝛽1𝑃𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜀𝑖,𝑡(𝑖), (9)
𝑃𝑖,𝑡(𝑖) = 𝛼0 + 𝛼1𝐼𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜑𝑖,𝑡(𝑖), (10)

ot
where 𝐼𝑖,𝑡(𝑖) is the number of thermal inversions to which inmate 𝑖 is exposed in the
tn
entire month before date 𝑡(𝑖).

rin
All the regressions conducted were weighted by the number of inmates in each
ep
month; however, the results were robust to omission of regression weights. Standard
errors were clustered at the date on which orders were put in. This allowed the error
Pr
terms of inmates shopping on the same day to be correlated.

21
5. Results
ed
Since we rely on 2SLS to estimate the effects of air pollution on inmates’
consumption decisions, we begin by examining the first-stage estimation of 2SLS.
iew
Panel A in Table 3 reports the regression results of Equation (10). The coefficient of
thermal inversions is always positive and statistically significant under different model
v
settings. In Columns (1)-(3), we add weather controls, inmate fixed effects, and
re
year/month/weekday fixed effects in sequence. After controlling for the group of time
fixed effects, the coefficient of thermal inversions decreases in magnitude slightly, but
er
still remains significant at the 1% level. In Column (4), we consider the weighted
regression using the total number of purchase orders placed by inmates on each order
pe
placement day; the result remains unchanged. The estimated coefficient in Column (4)
suggests that one more occurrence of thermal inversion would increase the average
ot
daily API by 0.4 unit in the entire month prior to the order placement day. That is, a
one standard deviation (8.0) increase in thermal inversion increases the average daily
tn
API by 6.1 percentage points, or 0.2 standard deviations. In Columns (1)-(4), the
Kleibergen-Paap (KP) Wald F-statistics (Kleibergen and Paap, 2006) are all larger than
rin
the Stock-Yogo weak identification test critical values at 10% maximal IV size of 16.4
(Stock and Yogo, 2005), suggesting that the IV is strong and valid. We will now focus
ep
on 2SLS results in all that follows.

Pr
22
5.1. Results of the total causal effects
ed
Panel B in Table 3 presents our main results, i.e., the total causal effect of air
pollution on inmates’ consumption expenditure. It is evident from Panel B that the
iew
coefficient of air pollution estimated by 2SLS not only consistently keeps the same
negative sign, but also varies little in magnitude across different model settings. Indeed,
the stability of 2SLS results suggests that the IV is exogenous. Column (4) shows that,
v
after controlling for weather conditions, inmate individual effects,
re
year/month/weekdays fixed effects, and using weighted regression, a rise in API in the
entire month prior to the order placement date significantly reduces the inmates’ total
er
consumption expenditure. This suggests that every 10 unit increase in monthly average
pe
API reduces a typical inmate’s consumption expenditure by 3.6%. In contrast, the OLS
results presented in Table A3 in the appendix are very sensitive to the model settings
and the air pollution effect becomes insignificant after controlling for the time fixed
ot
effects. Moreover, the OLS systematically under-estimates the air pollution impact on
inmates’ consumption.
tn

rin
We conducted a sensitivity analysis to check the robustness of our analysis to
variations in the measurement used when specifying the regression variables. As shown
ep
in Table 4 and Figure 3, the estimated coefficients of API are robust to variations in the
measurement of inversions/expenditure, weather and time window exposure.

Pr
In our baseline, we treat a thermal inversion as existing whenever the temperature

23
in the second layer (320 meters) is higher than the temperature in the ground layer (110
ed
meters), and use 5°C wide temperature-day bins ranging from ≤-10°C to ≥35°C.
Column (2) in Table 4 presents results obtained when the definition of a thermal
iew
inversion uses the 540 meter layer rather than the layer at 320 meters. Column (3) in
Table 4 reports the results using finer temperature-day bins, i.e., 3°C-wide bins. The
results hardly change. The temperature-day bins clearly reflect the nonlinear impacts of
v
temperature. To highlight their importance, Column (4) of Table 4 presents the
re
regression results obtained when the bins are replaced by a quadratic term reflecting
average daily temperature in the month prior to the order placement date. The
er
coefficient for air pollution is still significantly negative under the simpler temperature
pe
control, but is slightly reduced. In Column (5), we use the ratio (%) of total consumption
expenditure relative to the piece-rate wage as an alternative measure of consumption.
The ratio reflects the relative changes between the total consumption expenditure and
ot
wage income, and is negative at the 1% significance level.

tn
Figure 3 presents the 2SLS results obtained when using different time windows
rin
to indicate inmates’ exposure to air pollution. Since inmates are only allowed to order
consumption goods once a month, it is natural to use average daily API in the entire
ep
month prior to the order placement date to measure inmates’ exposure to air pollution.
To test sensitivity to the duration as well as the intensity of pollution, we used the mean
Pr
API in a range of time windows. These included contemporaneous daily API; average
24
daily API in the past six months, three months, two months, three weeks, two weeks
ed
and one week before the day of placing the order; and future average daily API in the
one week, one month, three months and six months after the day of placing the order.
iew
Future air pollution should not affect current consumption decisions and is used as a
placebo test. Whilst doing these tests of sensitivity to the API time period, the window
used for inversion and weather was matched with that used for the API.
v
There are several findings worth noting. First, the contemporaneous API has no
re
significant impact on inmates’ consumption expenditure. Since inmates make the
consumption decision after a month’s exposure to air pollution and need to consume
er
the ordered goods over the coming month, the current air pollution should have little
pe
impact on their consumption plan. Second, all past average daily APIs had a significant
inverse relationship with inmates’ total consumption expenditure, but the magnitude
and significance of the effect was at its greatest when using a one month time window,
ot
i.e., average daily API in the month preceding order placement. Although average daily
APIs in longer periods were still significant, both their significance and their magnitude
tn
fell sharply, suggesting that the additional time was decision-irrelevant. The result is
consistent with He et al. (2019), who show that worker outcomes are responsive to day-
rin
to-day variation in air pollution with up to 30 days of delay. Finally, Figure 3 also shows
that future exposure to air pollution has no impacts on today’s consumption decision.
ep
The regression using future average daily API is actually a placebo test, and our study
passes all placebo tests, validating the results of the study.

Pr
25
ed
5.2. Channel decomposition: substitution effect and income effect
Our theoretical model predicts that air pollution may affect an individual’s
iew
consumption through two channels: a direct substitution effect driven by the
substitution between damaged health and utility-generating consumption goods, and an
v
indirect income effect driven by the depressed productivity and wage income. We now
re
explore these two channels. Our major results are presented in Table 5.
We firstly check whether air pollution reduces inmates’ piece-rate wage income
by estimating
er
𝑙𝑛(𝑤𝑖,𝑡(𝑖)) = 𝛽𝑤0 + 𝛽𝑤1 𝑃𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜀𝑖,𝑡(𝑖), (11)
pe
where 𝑤𝑖,𝑡(𝑖) is the wage income inmate i receives in the month of 𝑡(𝑖) and 𝑃𝑖,𝑡(𝑖) is
estimated from Equation (9). The inmates are paid monthly, and the pay period starts
ot
on the 1st day of the month and ends on the last day of the month. All paydays are at
tn
the beginning of the following month. Our purpose is to decompose the substitution
and income effects of air pollution on individual consumption, with the piece-work
rin
income used as a control in the consumption regression (i.e., Column (4) in Table 5).
Accordingly, in Column (2) we present the regression results using the same air
ep
pollution variable as was used in the consumption regression, i.e., average daily API in
Pr
26
the entire month prior to the order placement day12. These show that a 10-unit increase
ed
in the average daily API reduces prisoners’ monthly earnings by 4.2%. The result is
directly comparable to previous studies – Graff and Neidell (2012) and Chang et al.
iew
(2016) – that used U.S. data to study the effect of concurrent air pollution on work
productivity.
An inmate receives income from two sources: wage income and family transfers.
v
Both are directly deposited into his personal trust account at the prison. When buying
re
grocery items on the scheduled date, the budget available to an inmate will be his total
account balance. If air pollution were related to inmates’ non-wage income, the
er
complexity of the above-discussed income effect of air pollution could be compounded
pe
by non-productivity factors. To test whether there is any correlation between air
pollution and outside transfer, we subtract the inmate’s last month’s wage income from
his account balance on the order placement date and regress the residual on the average
ot
daily API in the month prior to the order placement date. The results (see Column (3)
in Table 5) show that income transfers from outside the prison are independent of
tn
inmates’ air pollution exposure. This verifies the use of prisoners’ piece-work based
rin
12Note that in Column (2) of Table 5, the time window of the air pollution measure differs from the
time window of the working time, which is the inmate’s pay period, but they coincide if the allowed
shopping date is on the 1st day of the month. The data show that the inmates’ order placement days
tend to be clustered at the beginning of each month (the mean and median of the date is the 7th of a
ep
month). Hence the measurement error of the air pollution measure used in Column (2) is limited.
Moreover, since both the payday and shopping day in each month are fixed for all inmates, the
measurement error is completely exogenous and should not bias the regression results. We also
estimate the impact of average daily API in the month of the pay period on the inmate’s pay received
at the start of the following month. The 2SLS estimator is -0.0040, almost the same as that in
Pr
Column (2), and also significant at the 1% level.
27
earnings as the sole mediator in the following mediation analysis.
ed
We next decompose the total causal effect of air pollution on inmates’
consumption expenditure into the direct substitution and indirect income effects. If air
iew
pollution is randomly assigned, we can easily decompose the total effect into two causal
ones by using traditional mediation analysis.13 However, once air pollution is
endogenous, the two regressors in the mediation model – wage income and air pollution
v
– are both endogenous, and a complex causal chain arises. That is, the endogenous
re
treatment (air pollution) and its outcome (inmates’ piece-rate wage income) jointly
cause a second outcome of interest, inmates’ consumption. Conventionally, a separate

er
dedicated instrument for the wage income is required to achieve identification, but we
pe
only have one instrumental variable, which is unable to unpack the causal chain with
standard IV estimation. However, Deppel et al. (2020) show that, if air pollution is
endogenous in the regression of consumption on air pollution only because of omitted

ot
variables that affect wage income, and through wage income also affect consumption,
a single instrumental variable is enough to identify the total effect, direct effect and
tn
indirect effect. That is to say, in the following first-stage regression, we can use the
temperature inversion to instrument the wage income conditional on the air pollution,
rin
𝑙𝑛(𝑤𝑖,𝑡(𝑖)) = 𝛼𝑤0 + 𝛼𝑤1 𝐼𝑖,𝑡(𝑖) + 𝛼𝑤2 𝑃𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜑𝑖,𝑡(𝑖) (12)
ep
13That is, estimating the direct causal effect from the mediation model, i.e., regressing consumption
on both the mediator, i.e., piece-rate wage income, and the air pollution, and calculating the indirect
Pr
causal effect by differencing the total causal effect and the direct causal effect.
28
and then estimate the second-stage mediation model using the following equation,
ed
𝑙𝑛(𝐶𝑖,𝑡(𝑖)) = 𝛽𝐶0 + 𝛽𝐶1𝑃𝑖,𝑡(𝑖) + 𝛽𝐶2𝑙𝑛𝑤𝑖,𝑡(𝑖) + 𝑓(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜀𝑖,𝑡(𝑖) (13)
iew
where 𝑤𝑖,𝑡(𝑖) are the estimated values of 𝑤𝑖,𝑡(𝑖) in the first stage - Equation (12).
𝐶
Following Deppel et al. (2020), 𝛽𝐶1, 𝛽𝐶2 ∙ 𝛽𝑤1 and 𝛽𝐶1 + 𝛽2 ∙ 𝛽𝑤1 would be,
respectively, the direct, indirect and total causal effects of air pollution on the inmates’
v
consumption expenditure. Intuitively, the identification approach implies that, if the
re
following two conditions are satisfied, the air pollution would be exogenous in Equation
(13) and therefore the parameters of the mediation model can be well estimated. First,
er
the correlation between wage income and air pollution is the only channel for the air
pollution to be endogenous in regression (9). Second, temperature inversion is a valid

pe
instrument to identify the causal effect of wage income on consumption when
conditional on air pollution. As already discussed, in the isolated and closed prison
ot
environment, air pollution can be correlated with wage income both through a direct
environmental productivity effect and through outside-world economic activities.

tn
Hence, in Equation (13), when conditional on air pollution, the endogeneity can only
arise from the omitted outside-world economic activities, which affect consumption
rin
through the wage income. This is exactly the setting required by Deppel et al. (2020).
Column (4) in Table 5 presents the regression results of equation (13). Based on
ep
the estimators from Equations (11) and (13), we can calculate the direct, indirect and
total causal effects of air pollution on the inmates’ consumption expenditure, 𝛽𝐶1
Pr
29
𝐶
=‒ 0.0008, 𝛽𝐶2 ∙ 𝛽𝑤1 =‒ 0.0028 and 𝛽𝐶1 + 𝛽2 ∙ 𝛽𝑤1 =‒ 0.0036. The total causal effect
ed
𝐶
𝛽𝐶1 + 𝛽2 ∙ 𝛽𝑤1 estimated by Equation (13) can be shown to equal 𝛽1, which is estimated
by Equation (9) and presented in Column (1) in Table 5. The comparison shows that
iew
the total causal effects estimated by the two different approaches are exactly the same,
confirming that the identification approach proposed by Deppel et al. (2020) is valid
for our study. Evidently, the direct substitution effect and indirect income effect
v
respectively account for 78% and 22% of the total effect of air pollution on inmates’
re
consumption expenditure, and the income effect is the main force driving the impacts
of air pollution on individual daily consumption. To the best of our knowledge, the
er
above analysis could be the first in the literature that decomposes the total causal effect
pe
of air pollution on individual consumption into the two different channels.

ot
5.3. Heterogeneity of consumption results
To assess how inmates and inmates’ consumption choices are differentially

tn
affected by air pollution, we explore the potential heterogeneity by running multiple
subsample regressions. We first separate inmates’ aggregate consumption expenditure

rin
into five different consumption expenditure components – food, personal care, personal
stationery, phone cards and cigarettes – to test the heterogeneity effects of air pollution
ep
on the expenditure on different consumption goods. We then divide the population of
inmates into different groups by their characteristics to test the heterogeneous

Pr
30
consumption effects of air pollution on different individuals.
ed
5.3.1. Heterogeneity of consumption goods
Table 6 presents the 2SLS estimates of air pollution’s effects on each
iew
consumption expenditure component. As shown in the table, when inmates are exposed
to more polluted air during the month before they place their purchase orders, they
v
reduce their spending on food and cigarettes, but increase their expenditure on personal
re
care. Expenditure on personal stationery, however, is unresponsive to air pollution.
Within the category of food, more heterogeneity is observed. Specifically, air pollution
er
induces inmates to consume more sweets and fruits, and less beverages, cookies,
noodles and canned food, and has little effect on pre-packaged snacks such as nuts.
pe
Shenzhen’s air pollution level is relatively low when compared to that in most
Chinese cities; during the most polluted month in our sample period, the API was only
106, and its day-to-day changes go largely unnoticed by prison inmates. This suggests
ot
that air pollution’s effects on consumption expenditure by inmates are mainly driven
tn
by its impacts on prisoner health.
Our results suggest several patterns whereby air pollution has impacts on
rin
consumption. First, the discomfort caused by greater air pollution exposure may lead
individuals to be depressed and lose appetite, which in turn reduces their consumption
ep
of carbohydrate foods and meats, such as cookies, noodles and canned food, but may
increase their consumption of stress-relieving foods such as sweets. Second, air

Pr
pollution based discomfort may incline people to consume fewer unhealthy goods, such
31
as cigarettes, and more healthy goods, such as fruits. Thirdly, air pollution may make
ed
daily physical activity less attractive, and so reduce the demand for beverages. Fourthly,
people may increase their spending on personal toiletries in response to the dirty air.
iew
Finally, as Table 6 shows, cigarettes are the most sensitive items to air pollution – a 10-
unit increase in average daily API leads to a nearly 7.9% reduction in expenditure on
them. Since cigarettes are the prisoners’ second largest expenditure item (behind
v
snacks), accounting for 22.7% of inmates’ total consumption expenditure, this response
re
to air pollution has much larger impacts than other consumption goods.
er
5.3.2. Heterogeneity amongst inmates

pe
The impacts of air pollution on the amount prisoners spend each month varies
across the prison population. We now explore these variations in terms of prisoner age,
ot
education and productivity level. As shown in Table 7, air pollution significantly
reduces total spending by inmates younger than 40, but has no impact on inmates aged
tn
above 40. Moreover, although air pollution has no impact on spending by illiterate
inmates (see Column (1) in Panel B of Table 7), it reduces the total consumption
rin
expenditure of all other education levels, with the magnitude of the contraction
increasing with prisoner education. The results in Table 8 further suggest that the total
ep
consumption expenditure of more productive inmates is more sensitive to air pollution.
In sum, these results show that the consumption expenditures of younger, more
Pr
educated and more productive inmates are more likely to be negatively affected by
32
increases in air pollution.
ed
iew
Inmates who are younger, more educated and more productive are also more
likely to be high earners and therefore to spend more on consumption. To understand
v
the source of the heterogeneity, we explore whether differences in inmates’
re
consumption levels can explain these patterns. To assess this, we run a subsample
regression, splitting the whole sample into 10 decile consumption groups based on the
er
consumption expenditure of each inmate. Figure 4 presents the results of the subsample
pe
regression for each decile. It shows that air pollution significantly reduces the total
consumption expenditure of the three largest spending decile groups, i.e., the 8th, 9th
and 10th, but has no significant effect on spending by the other seven deciles. The
ot
finding suggests that the variations in the pattern of spending by prisoners depends on
tn
the variations in the amounts they spend.

rin
5.4. Nonlinearity of consumption results
Throughout our analysis, we have assumed air pollution to have a linear effect on
ep
inmates’ consumption. We now move on to examine whether the relationship is non-
linear. To test the nonlinear effect, we cannot simply add higher-order forms of API
Pr
33
into the regression, as higher-order forms of API could also be endogenous and we do
ed
not have enough valid IVs to instrument them. The approach we adopt to test for
nonlinearity is to introduce a spline term 𝛽2[𝑃𝑖,𝑡(𝑖) ‒ 𝜅] ∙ 𝐼(𝑃𝑖,𝑡(𝑖) ≥ 𝜅) in the regression
iew
equation,
𝐿𝑜𝑔(𝐸𝑥𝑝𝑒𝑛𝑑𝑖𝑡𝑢𝑟𝑒𝑖,𝑡(𝑖)) = 𝛽0 + 𝛽1𝑃𝑖,𝑡(𝑖) + 𝛽2[𝑃𝑖,𝑡(𝑖) ‒ 𝜅] ∙ 𝐼(𝑃𝑖,𝑡(𝑖) ≥ 𝜅) + 𝑓
v
(𝑊𝑖,𝑡(𝑖)) + 𝜃𝑖 + 𝑔(𝑡(𝑖)) + 𝜀𝑖,𝑡(𝑖), (14)
re
where 𝜅 represents the threshold value at which non-linearity is assumed to arise,
where 𝐼(𝑃𝑖,𝑡(𝑖) ≥ 𝜅) equals one if 𝑃𝑖,𝑡(𝑖) ≥ 𝜅, and zero otherwise. With the presence of
er
the spline term, 𝛽1 becomes the marginal effect of API on consumption, provided that
API is less than 𝜅, and 𝛽1 + 𝛽2 is the marginal effect of API on consumption when
pe
API is larger than or equal to 𝜅. Hence, a non-zero 𝛽2 is evidence of nonlinearity in
the relationship between air quality and monthly spending by prisoners.

ot
Two difficulties remain in estimating equation (11): the threshold value 𝜅 is
unknown and the spline term [𝑃𝑖,𝑡(𝑖) ‒ 𝜅] ∙ 𝐼(𝑃𝑖,𝑡(𝑖) ≥ 𝜅) could be endogenous. To

tn
address the first issue, we try all possible candidate values of 𝜅, one by one, to estimate
equation (11). More specifically, we start at 𝜅 = 30 and increase the value by 10 at

rin
each try until 𝜅 = 80. If a threshold value of 𝜅 does exist, the coefficient of 𝛽2
should significantly change its sign or magnitude when regressing on true 𝜅. To solve
ep
the endogeneity problem of the newly added spline term, we construct another thermal
inversion, i.e., temperature difference between the 1st layer and 3rd layer, to instrument
Pr
the variable, and continue using thermal inversions as defined in the benchmark
34
regression to instrument 𝑃𝑖,𝑡(𝑖).
ed
Figure 5 and Table A4 present the results of our test for nonlinearity. As shown
in Figure 5, the estimate of 𝛽1 is statistically significant if and only if 𝜅 is set above
iew
or equal to 50; the estimate of 𝛽2 is always statistically significant; and the coefficient
of 𝛽1 + 𝛽2 is statistically significant if and only if 𝜅 is set above or equal to 40.
Moreover, the magnitudes and significances of these estimates are rather stable once
v
they become statistically significant. Since 𝛽1 and 𝛽1 + 𝛽2 respectively measure the
re
marginal effect of our API measure below and above 𝜅, the results imply that air
pollution negatively affect inmates’ consumption if and only if average daily API in the
er
month prior to the order placement day exceeds 40. Moreover, the spending by
pe
prisoners drops at an increasing rate as levels of air pollution rise. Recall that, by the
officially announced air quality standards in China, daily air quality with API less than
50 is considered as excellent, with air pollution posing little or no risk. Our result
ot
implies that the official standard of air quality could be higher.

tn

rin
6. Conclusion
In this paper, we merged a unique data set containing records of the individual-
ep
level piece-work wage income and monthly consumption expenditures of prison
inmates, with data on air pollution and weather conditions. The combined data was then
Pr
35
used to study the effect of air pollution on individual prisoners’ consumption behavior.
ed
We found that a 10 unit increase in average API exposure resulted a significant and
robust 3.6 percentage drop in inmates’ total monthly consumption expenditure. The
iew
impact was primarily driven by the income effect, i.e., the air pollution depressed
worker productivity and wage income. Our results also indicated that individuals may
have changed their consumption patterns in response to the air pollution, and that ‘high’
v
spenders are more likely to be affected by the air pollution. Moreover, individual
re
decisions on consumption responded to air pollution if and only if the average daily
API was high enough, i.e., exceeding 40.

er
The consumption impacts of air pollution estimated in the paper offer a novel
pe
angle on the intertwining of air pollution with the economy, i.e., identifying a new
morbidity impact of air pollution that is more subtle, but more pervasive, than standard
health encounters. More specifically, our study shows that clean air, by stimulating
ot
people to consume more, can boost both the individual’s welfare and the nation’s GDP.
We concede that our data came from a rather unique population group, limiting
tn
the generalizability of the results across the whole population in the country.
Nevertheless, the size of the productivity effect of air pollution obtained was consistent
rin
with previous studies and should have general implications for labor-intensive
manufacturing in China. Our estimate of air pollution’s effects on consumption should

ep
be considered as a lower bound indicator of its effects on average households in the
economy since some consumption expenditures that are necessary to prisoners, e.g.,
Pr
36
travel expenditure, could also be depressed by air pollution (Chen et al., 2021).
ed
Another of the paper’s major contributions is to decompose air pollution’s
impacts on individual spending into income and substitution effects, and to show that
iew
the income effect is dominant. The qualitative results of our study should be valid for
most population groups although the quantitative results may differ. Finally, although
our study may sacrifice some representativeness, we gained greatly from the clean
v
identification. The prison naturally provides an ideal experimental environment for
re
studying consumption behaviors of individuals; the detailed consumption expenditure
of each inmate is perfectly documented and inmates have no options regarding when
er
and where to buy consumption goods. Many common and inevitable endogeneity
pe
problems in previous studies, such as self-selection and avoidance behavior, become
irrelevant in our study. Moreover, the inelastic labor supply, and the piece-work pay
system inside the prison factory, allowed us to clearly identify how air pollution
ot
differentially affected consumption through the income and substitution effects.

tn
rin
ep
Pr
37
References
ed
Adhvaryu, Achyuta, James Fenske, Namrata Kala, and Anant Nyshadham. "Fetal origins of mental
health: Evidence from Africa. " Centre for the Study of African Economies, University of
Oxford, No. 2015-15 (2015).
iew
Anderson, Michael L. "As the wind blows: The effects of long-term exposure to air pollution on
mortality." Journal of the European Economic Association 18.4 (2020): 1886-1927.
Arceo, Eva, Rema Hanna, and Paulina Oliva. "Does the effect of pollution on infant mortality differ
between developing and developed countries? Evidence from Mexico City." The Economic
v
Journal 126.591 (2016): 257-280.
re
Archsmith, James, Anthony Heyes, and Soodeh Saberian. "Air quality and error quantity: Pollution
and performance in a high-skilled, quality-focused occupation." Journal of the Association
of Environmental and Resource Economists 5.4 (2018): 827-863.
er
Barwick, Panle Jia, Shanjun Li, Deyu Rao, and Nahim Zahur. "The Morbidity Cost of Air Pollution:
Evidence from Consumer Spending in China." NBER Working Paper No. w24688 (2018).
Carson, Richard T., Phoebe Koundouri, and Céline Nauges. "Arsenic mitigation in Bangladesh: a
pe
household labor market approach." American Journal of Agricultural Economics 93.2
(2011): 407-414.
Chang, Tom, Joshua GraffZivin, Tal Gross, and Matthew Neidell. "Particulate pollution and the
productivity of pear packers." American Economic Journal: Economic Policy 8.3 (2016):
ot
141-69.
Chang, Tom, Joshua GraffZivin, Tal Gross, and Matthew Neidell. "The effect of pollution on worker
productivity: evidence from call center workers in China." American Economic Journal:
tn
Applied Economics 11.1 (2019): 151-72.
Chang, Tom Y., Wei Huang, and Yongxiang Wang. "Something in the air: Pollution and the demand
for health insurance." The Review of Economic Studies 85.3 (2018): 1609-1634.
rin
Chay, Kenneth Y., and Michael Greenstone. "The impact of air pollution on infant mortality:
Evidence from geographic variation in pollution shocks induced by a recession." Quarterly
Journal of Economics 118.3 (2003): 1121-1121.
ep
Chay, Kenneth Y., and Michael Greenstone. "Does air quality matter? Evidence from the housing
market." Journal of political Economy 113.2 (2005): 376-424.
Chen, Shuai, Paulina Oliva, and Peng Zhang. "The effect of air pollution on migration: evidence
Pr
from China." NBER Working Paper No. w24036 (2017).
38
Chen, Shuai, Yuyu Chen, Ziteng Lei, and Jie-Sheng Tan-Soo. "Impact of air pollution on short-term
movements: evidence from air travels in China." Journal of Economic Geography 20.4
ed
(2020): 939-968.
Chen, Shuai, Yuyu Chen, Ziteng Lei, and Jie-Sheng Tan-Soo. "Chasing clean air: Pollution-induced
travels in China." Journal of the Association of Environmental and Resource Economists
iew
8.1 (2021): 59-89.
Chen, Yuyu, Avraham Ebenstein, Michael Greenstone, and Hongbin Li. "Evidence on the impact
of sustained exposure to air pollution on life expectancy from China’s Huai River policy."
Proceedings of the National Academy of Sciences 110.32 (2013): 12936-12941.
v
Currie, Janet, and Matthew Neidell. "Air pollution and infant health: what can we learn from
California's recent experience?." The Quarterly Journal of Economics 120.3 (2005): 1003-
re
1030.
Chew, Soo Hong, Wei Huang, and Xun Li. "Does haze cloud decision making? A natural laboratory
experiment." Journal of Economic Behavior & Organization 182 (2021): 132-161.
er
Dippel, Christian, Robert Gold, Stephan Heblich, and Rodrigo Pinto. "Mediation Analysis in IV
Settings With a Single Instrument." (2020), https://christiandippel.com/IVmediate_.pdf.
pe
Deryugina, Tatyana, Garth Heutel, Nolan H.Miller, David Molitor and Julian Reif. "The mortality
and medical costs of air pollution: Evidence from changes in wind direction." American
Economic Review 109.12 (2019): 4178-4219.
Deschênes, Olivier, and Michael Greenstone. "The economic impacts of climate change: evidence
from agricultural output and random fluctuations in weather." American Economic Review
ot
97.1 (2007): 354-385.
Deschenes, Olivier, Michael Greenstone, and Joseph S. Shapiro. "Defensive investments and the
tn
demand for air quality: Evidence from the NOx budget program." American Economic
Review 107.10 (2017): 2958-89.
Ebenstein, Avraham. "The consequences of industrialization: evidence from water pollution and
rin
digestive cancers in China." Review of Economics and Statistics 94.1 (2012): 186-201.
Ebenstein, Avraham, Maoyong Fan, Michael Greenstone, Guojun He, and Maigeng Zhou. "New
evidence on the impact of sustained exposure to air pollution on life expectancy from
ep
China’s Huai River Policy." Proceedings of the National Academy of Sciences 114.39
(2017): 10384-10389.
Freeman, Richard, Wenquan Liang, Ran Song, and Christopher Timmins. "Willingness to pay for
clean air in China." Journal of Environmental Economics and Management 94 (2019): 188-
Pr
216.
39
Fu, Shihe, V. Brian Viard, and Peng Zhang. "Air Pollution and Manufacturing Firm Productivity:
Nationwide Estimates for China." The Economic Journal, ueab033 (2021),
ed
https://doi.org/10.1093/ej/ueab033.
Graff Zivin, Joshua, and Matthew Neidell. "The impact of pollution on worker productivity."
American Economic Review 102.7 (2012): 3652-73.
iew
Ghanem, Dalia, and Junjie Zhang. "‘Effortless Perfection:’ Do Chinese cities manipulate air
pollution data?." Journal of Environmental Economics and Management 68.2 (2014): 203-
225.
Greenstone, Michael, and Rema Hanna. "Environmental regulations, air and water pollution, and
v
infant mortality in India." American Economic Review 104.10 (2014): 3038-72.
re
Graff Zivin, Joshua, and Matthew Neidell. "Days of haze: Environmental information disclosure
and intertemporal avoidance behavior." Journal of Environmental Economics and
Management 58.2 (2009): 119-128.
Graff Zivin, Joshua, and Matthew Neidell. "The impact of pollution on worker productivity."
er
American Economic Review 102.7 (2012): 3652-73.
Hanna, Rema, and Paulina Oliva. "The effect of pollution on labor supply: Evidence from a natural
pe
experiment in Mexico City." Journal of Public Economics 122 (2015): 68-79.
He, Guojun, Tong Liu, and Maigeng Zhou. "Straw burning, PM2.5, and death: evidence from
China." Journal of Development Economics 145 (2020): 102468.
He, Jiaxiu, Haoming Liu, and Alberto Salvo. "Severe air pollution and labor productivity: Evidence
ot
from industrial towns in China." American Economic Journal: Applied Economics 11.1
(2019): 173-201.
tn
Hicks, Daniel, Patrick Marsh, and Paulina Oliva. "Air pollution and procyclical mortality: Causal
evidence from thermal inversions." NBER working paper (2016).
Ito, Koichiro, and Shuang Zhang. "Willingness to pay for clean air: Evidence from air purifier
rin
markets in China." Journal of Political Economy 128.5 (2020): 1627-1672.
Jans, Jenny, Per Johansson, and J. Peter Nilsson. "Economic status, air quality, and child health:
Evidence from inversion episodes." Journal of Health Economics 61 (2018): 220-232.
ep
Kleibergen, Frank, and Richard Paap. "Generalized Reduced Rank Tests using the Singular Value
Decomposition." Journal of Econometrics 133.1 (2006): 97-126.
Khanna, Gaurav, Wenquan Liang, Ahmed Mushfiq Mobarak, and Ran Song. "The productivity
Pr
consequences of pollution-induced migration in China." NBER Working Paper No.

w28401 (2021).
40
Knittel, Christopher R., Douglas L. Miller, and Nicholas J. Sanders. "Caution, drivers! Children
present: Traffic, pollution, and infant health." Review of Economics and Statistics 98.2
ed
(2016): 350-366.
Moretti, Enrico, and Matthew Neidell. "Pollution, health, and avoidance behavior evidence from
the ports of Los Angeles." Journal of Human Resources 46.1 (2011): 154-175.
iew
Neidell, Matthew. "Information, avoidance behavior, and health the effect of ozone on asthma
hospitalizations." Journal of Human Resources 44.2 (2009): 450-478.
Poyker, Michael. "Economic consequences of the US convict labor system." Institute for New
Economic Thinking Working Paper Series 91 (2019).
v
Schlenker, Wolfram, and W. Reed Walker. "Airports, air pollution, and contemporaneous health."
re
The Review of Economic Studies 83.2 (2016): 768-809.
Stock, James, and Motohiro Yogo. "Asymptotic distributions of instrumental variables statistics
with many instruments. " Vol. 6. Chapter, (2005).
er
Sullivan, Daniel M. "The True Cost of Air Pollution: Evidence from the Housing Market." Working
Paper, (2017).
pe
Viard, V. Brian, and Shihe Fu. "The effect of Beijing's driving restrictions on pollution and
economic activity." Journal of Public Economics 125 (2015): 98-115.
Zhang, Junjie, and Quan Mu. "Air pollution and defensive expenditures: Evidence from particulate-
filtering facemasks." Journal of Environmental Economics and Management 92 (2018):
517-536.
ot
tn
rin
ep
Pr
41
Figures and graphs
ed
v iew
re
er
Figure 1. Time trend of inmates’ monthly expenditure and piece-rate wage
pe
Notes: The light-blue bar denotes the total monthly consumption expenditure of inmates. We have
missing records of expenditure in the period from Jul. 2004 to Sep. 2004 and the whole year of 2011.
The red dropline depicts inmates’ average monthly piece-rate wage. The black line presents the ratio of
expenditure/wage.
ot
tn
rin
ep
Pr
42 / 59
ed
v iew
re
er
Figure 2. Positive correlation between thermal inversion and API
Notes: Light-blue bar denotes the monthly average API that inmates were exposed to. The dark-green
pe
dropline depicts the corresponding monthly number of thermal inversions.
ot
tn
rin
ep
Pr
43 / 59
ed
v iew
re
er
Figure 3. The effect of API on the total consumption expenditure by different
time windows of air pollution exposure
pe
Notes: This figure presents the effects of air pollution on inmates’ total monthly consumption expenditure
by different time windows of air pollution exposure. API is calculated by averaging the daily APIs in
these different time windows. The circle denotes the point estimate of the total causal effect and the
whisker denotes the 95% confidence intervals. Our baseline (average daily API in the entire month prior
to the shopping day) is highlighted in red horizontal line.
ot
tn
rin
ep
Pr
44 / 59
ed
v iew
re
er
Figure 4. The effect of API on the total expenditure by expenditure quantiles
Notes: This figure presents the effects of air pollution on inmates’ consumption expenditure by 10
pe
quantile consumption subsamples (Q1-Q10). The quantile is divided by the average monthly expenditure
of each inmate during our research period. The circle denotes the point estimate of the effect and the
whisker denotes the 95% confidence intervals.
ot
tn
rin
ep
Pr
45 / 59
ed
v iew
re
er
Figure 5. Nonlinear impacts using spline regression
Notes: This figure shows the marginal impact of API on inmates’ consumption expenditure and its
corresponding 95% confidence interval at ascending knots of API on a spline regression. The circle dots
pe
and square dots respectively represent the marginal impact for observations where API is smaller and
larger than the knot.
ot
tn
rin
ep
Pr
46 / 59
Table 1. Summary statistics
ed
Variable Def./Unit Mean SD Min Max
Total commodity expenditure
Total expenditure CNY (yuan) 205.43 144.34 0.102 1586.0
Piece-rate wage CNY (yuan/month) 65.87 76.92 0 3882.9
iew
Total account balance CNY (yuan/month) 163.68 813.43 0 73143.2
Air pollution
API Air pollution index (0-500) 52.82 15.07 27.07 105.95
Thermal inversions
Inversions # of inversions (1st layer) 5.687 7.958 0 38
Notes: Number of observations=433,369; number of prisoners=20,334. Total account balance is the
v
balance of inmates’ trust account on the day placing the purchase order. Summary statistics for
weather controls and inmates’ characteristics information see Table A1-A2 in the Appendix.
re
er
pe
ot
tn
rin
ep
Pr
47 / 59
Table 2. Summary statistics of expenditure structure
ed
Expenditure (Yuan) Spending share (%)
Expenditure by categories
Mean SD Mean SD
Food 135.03 124.01 55.59 31.22
Beverage 55.54 47.73 8.87 14.19
iew
Cookies, noodles, and canned food 29.01 38.83 12.34 16.03
Sweets 23.79 22.79 2.41 6.06
Snacks 81.00 66.53 27.32 23.07
Fruits 48.45 42.24 4.64 11.17
Office suppliers 13.51 18.80 0.43 4.82
Personal care 25.71 24.95 20.15 24.36
v
Phone cards 8.22 8.07 1.62 5.62
Cigarettes 93.67 69.61 22.22 26.65
re
er
pe
ot
tn
rin
ep
Pr
48 / 59
Table 3. 2SLS estimates of the total effect of API on Log(Expenditure)
ed
(1) (2) (3) (4)
Panel A: 1st-stage estimates
Dep. Var.: API
Thermal inversions 0.4547*** 0.4721*** 0.3864*** 0.4197***
iew
(0.0052) (0.0060) (0.0045) (0.0049)
1st-stage F-statistics 495.5 123.5 154.3 38.72
KP F-statistics 77.76 60.94 74.19 74.63
Panel B: 2nd-stage estimates
Dep. Var.: Log(Expenditure)
API -0.0053*** -0.0036*** -0.0034*** -0.0036***
v
(0.0014) (0.0005) (0.0008) (0.0007)
Model specifications
re
Weather controls Yes Yes Yes Yes
Inmate FE No Yes Yes Yes
Year-month-weekdays FE No No Yes Yes
Sample weight er No No No Yes
Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects and
weather controls, and are weighted by the number of inmates in the jail across different days.
Weather controls include temperatures bins and second order polynomials in average humidity,
pe
wind speed, total sunshine duration, and cumulative precipitation. Standard errors are clustered
by the date and are listed in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
ot
tn
rin
ep
Pr
49 / 59
e d
w
Table 4. Robustness checks
e
Dep. Var. Log(Expenditure) Expenditure/Wage
Scenarios
Alternative
standard errors
(1)
Alternative
layer of inversions
(2)
Flexible
temperature controls
(3) (4)
v i
Measurement
of consumption
(5)
API -0.0036***
(0.0007)
-0.0038***
(0.0007)
-0.0032***
(0.0008)
r e
-0.0036***
(0.0010)
-0.5201***
(0.1173)
Temperature
IV- inversions
5°C bins
Layer 1
5°C bins
Layer 2
er 3°C bins
Layer 1
Quadratic
Layer 1
5°C bins
Layer 1
KP F-statistics
S.E. clustered by
API composition
153.05
Robust
Composite
153.33
Date
Composite
p e 76.08
Date
Composite
93.47
Date
Composite
74.63
Date
Composite
o t
Notes: Less than 0.1% observations are missing in Column (5) regression as Expenditure/Wage (%) cannot be defined when
piece rate wage income equals to zero. All regressions control for inmate fixed effects, year/month/weekday fixed effects
n
and weather controls, and are weighted by the number of inmates in the jail across different days. Standard errors are clustered
by date and are listed in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
ir n t
e p
P r 50 / 59
Table 5. Mechanisms and mediating effect
e d
Dep. Var. Log(Expenditure)
(1) Baseline
Mechanisms
Log(Piece-rate wage)
(2)
Log(Balance - Wage)
(3)
i ew
Mediating effect
Log(Expenditure)
(4)
API -0.0036***
(0.0007)
-0.0042***
(0.0010)
-0.0002
(0.0029)
e v -0.0008***
(0.0003)
Log(Piece-rate wage)
r r 0.6622***
(0.1742)
KP F-statistics
Mediating effect (%)
74.63
ee
74.63 74.63 29.84
78
Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects and weather controls, and are
parentheses. *** p<0.01, ** p<0.05, * p<0.1.
t p
weighted by the number of inmates in the jail across different days. Standard errors are clustered by date and are listed in
n o
ir n t
e p
P r 51 / 59
Table 6. Heterogeneous analyses: by commodity category
e d
Category Total food
(1)
Beverage
(2)
Cookies, noodles,
and canned food
(3)
Sweets
(4)
Snacks
(5)
Fruits
(6)
Office supplies
(7)
i
(8)
ew
Personal care Phone cards
(9)
Cigarettes
(10)
API -0.0059** -0.0130*** -0.0148*** 0.0226***

0.0033 0.0128*** -0.0007
e v
0.0056*** -0.0072** -0.0081***
Observations
(0.0024)
264,971
(0.0027)
133,356
(0.0028)
168,317
(0.0029)
74,596
(0.0028)
237,222
r
(0.0026)
78,059
(0.0027)
305,003 r (0.0013)
305,003
(0.0028)
265,019
(0.0028)
158,651
KP F-statistics 40.88 28.95 28.75 18.66 34.51
ee 28.83 39.37 39.37 28.68 19.10

Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects and weather controls, and are weighted by the number of inmates in the jail
across different days. Standard errors are clustered by date and are listed in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
t p
n o
ir n t
e p
P r 52 / 59
Table 7. Heterogeneous analyses: by age and education
ed
(1) (2) (3) (4) (5)
Age cohort <30 30-40 40-50 50-60 >=60
Panel A: by age cohort
iew
API -0.0042*** -0.0037*** -0.0021 0.0053 0.0123
(0.0010) (0.0013) (0.0019) (0.0035) (0.0111)
KP F-statistics 254.3 397.0 209.7 42.85 44.52
Observations 199,946 157,781 62,755 10,989 1,522
Below Primary Junior Senior College
Education
v
primary school high school high school or above
Panel B: by education level
re
API -0.0010 -0.0016 -0.0037*** -0.0078*** -0.0059**
(0.0072) (0.0014) (0.0010) (0.0022) (0.0030)
KP F-statistics 227.7 350.2 444.4 155.2 67.04
Observations 7,319 er 118,525 245,003 47,132 14,525
Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects and
weather controls, and are weighted by the number of inmates in the jail across different days.
Standard errors are clustered by date and are listed in parentheses. *** p<0.01, ** p<0.05, *
pe
p<0.1.
ot
tn
rin
ep
Pr
53 / 59
Table 8. Heterogeneous analyses: by work type and wage level
ed
Wage level Work type in Jail
Management
Subsample Above average Below average Production work
and logistics
iew
(1) (2) (3) (4)
API -0.0016** -0.0062*** -0.0005 -0.0042***
(0.0007) (0.0016) (0.0017) (0.0008)
KP F-statistics 178.95 34.53 43.54 69.33

Observations 171,118 261,875 24,532 408,461
v
Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects
re
and weather controls, and are weighted by the number of inmates in the jail across different
days. Standard errors are clustered by date and are listed in parentheses. *** p<0.01, **
p<0.05, * p<0.1.
er
pe
ot
tn
rin
ep
Pr
54 / 59
Appendix
ed
v iew
re
(a) Number of inmates (b) Duration of stay
er
pe
(c) Age of prisoners (d) Education distribution

ot
Figure A1. Inmates in Shenzhen Men’s Prison

Notes: This figure presents the basic information of inmates in Shenzhen Men’s Jail, including (a)
number of inmates (monthly total numbers), (b) distribution of inmates’ stay (counting measure by
tn
months), (c) age distribution, and (d) education distribution.

rin
ep
Pr
55 / 59
Table A1. Summary statistics for inmates’ characteristics
ed
Variable Definition Obs. Mean SD Min Max
Age Year 20,334 33.77 8.174 18 93
Hukou 1-urban; 0-rural 20,334 0.137 0.344 0 1
Marry status 1-marry; 0-otherwise 20,334 0.340 0.474 0 1
iew
1-management and logistics;
Work type in Jail 20,334 0.028 0.166 0 1
0-production work
Education levels (1-5) 20,334 2.863 0.767 1 5
Education-1 Management and logistics 636 3.048 0.783 1 5
Education-0 Production work 20,036 2.857 0.766 1 5
Sentence length # of month 20,672 51.40 32.18 6 240
v
Notes: Definitions of education levels: 1-below primary school; 2- primary school; 3- junior high
school; 4-senior high school; 5-colledge and above.
re
er
pe
ot
tn
rin
ep
Pr
56 / 59
Table A2. Summary statistics for weather variables
ed
Variable Mean SD Min Max
Temperature bins
Tbin <=10°C Days (total) 0.437 1.275 0 10
Tbin 10-15°C Days (total) 2.244 4.035 0 20
iew
Tbin 15-20°C Days (total) 5.444 6.546 0 22
Tbin 25-30°C Days (total) 6.743 7.080 0 25
11.20
Tbin 30-35°C Days (total) 12.09 7 0 28
Tbin >=35°C Days (total) 1.037 2.273 0 16
Other weather vairables
v
Humidity % 72.39 8.427 47.107 90
Rainfall mm (total) 4.905 5.457 0 46.93
re
Solar duration hours (daily average) 5.174 1.649 1.250 10.06
Wind force m/s (daily average) 2.257 0.270 1.521 3.657
Alternative measurement of thermal inversions
Inversions-L2 # of inversions (2nd layer) 4.487
er 5.859 0 24
Notes: All weather variables are constructed from daily measures in the entire month
prior to the shopping day of each inmate.
pe
ot
tn
rin
ep
Pr
57 / 59
Table A3. Reduced effects and OLS comparison
ed
(1) (2) (3) (4)
Panel A: Reduced estimates
Inversions -0.0239*** -0.0015*** -0.0014*** -0.0017***
iew
(0.0005) (0.0003) (0.0003) (0.0003)
R-squared 0.0859 0.6214 0.6325 0.6277
Panel B: OLS comparison
API 0.0051*** 0.0005*** 0.0001 -0.0001
(0.0003) (0.0001) (0.0002) (0.0002)
R-squared 0.0779 0.6214 0.6324 0.6276
v
Weather controls Yes Yes Yes Yes
Inmate FE No Yes Yes Yes
re
Year-month-weekdays FE No No Yes Yes
Sample weight No No No Yes
Notes: Panel A is the reduced estimates, in which the IV (thermal inversions) is directly
linked to the expenditure, while Panel B presents the OLS estimation results of air
er
pollution on expenditure as a comparison for the 2SLS estimates reported in Table 3.
All regressions control for inmate fixed effects, year/month/weekday fixed effects and
weather controls, and are weighted by the number of inmates in the jail across different
pe
days. Standard errors are clustered by date and are listed in parentheses. *** p<0.01, **
p<0.05, * p<0.1.
ot
tn
rin
ep
Pr
58 / 59
e d
w
Table A4. Nonlinear evidence: spline regression
e
Cutoff: API=
API spline β1
(1)
30
-0.0050
(2)
40
-0.0069*
(3)
50
-0.0053***
(4)
60
-0.0049***
(5)
70
-0.0044***
v i(6)
80
-0.0042***
API spline β2
(0.0056)
-0.0038***
(0.0039)
-0.0037***
(0.0020)
-0.0035***
(0.0015)
-0.0035***
r e
(0.0010)
-0.0035***
(0.0008)
-0.0032***
Lincom (β1+β2)
(0.0007)
-0.0054
(0.0007)
-0.0106**
er
(0.0007)
-0.0089***
(0.0007)
-0.0084***
(0.0008)
-0.0079***
(0.0010)
-0.0074***
Lincom-se
KP F-statistics
(0.0056)
38.39
(0.0039)
45.73
p e
(0.0021)
38.08
(0.0017)
36.93
(0.0014)
342.3
(0.0014)
40.37
o t
Notes: All regressions control for inmate fixed effects, year/month/weekday fixed effects and weather
controls, and are weighted by the number of inmates in the jail across different days. Standard errors are
n
clustered by date and are listed in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
ir n t
e p
P r 59 / 59

Air Pollution, Labor Productivity and Individual

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Air Pollution, Labor Productivity and Individual

Uploaded by

Copyright:

Available Formats

ed

Air Pollution, Labor Productivity and Individual

Shuai Chen Min Wang Dandan Zhang

Key word: Air pollution; labor productivity; consumption; inmates

JEL Codes: Q51, Q53, I15, O15.

Key word: Air pollution; labor productivity; consumption; inmates

JEL Codes: Q51, Q53, I15, O15.

impacts of pollution on individual consumption have received limited attention.

Pollution’s effects on consumption also warrant attention. Consumption is both

and He et al. (2020), for example.

The impacts of air pollution on an individual’s consumption decisions can be split

effect of air pollution on individual consumption.

such a shift on non-defensive consumption, the income effect of air pollution on

purifiers (Ito and Zhang, 2020).

expect air pollution’s impacts on individual daily consumption to be more prominent

Two clear challenges are faced by a researcher trying to quantitatively estimate

individual’s expenditures over a long period of time. In addition, to decompose the

card transactions in daily necessities.

Second, identifying the causal effect of air pollution on changes in consumption

may also generate measurement error.

consumption expenditure and piece-rate wage information of all male inmates

the only option for inmates, external selection is not a concern.

total consumption expenditure on their air pollution exposure, which is instrumented

by thermal inversions. A thermal (i.e., temperature) inversion occurs when a mass of

likely to influence both productivity and consumption decisions, the occurrence of

thermal inversions appears random and independent of economic activities; it has

Using the consumption expenditures and earnings of inmates, a mediation

unhealthy goods (e.g., cigarettes); the consumption expenditures of younger, more

increased air pollution; and inmates’ consumption expenditure remains unresponsive

whereby air pollution alters an individual’s consumption decisions on non-defensive

pollution, and to cleanly decompose its effects on individual consumption expenditure

into substitution and income effects.

The paper proceeds as follows. Section 2 presents a theoretical model that

assumed to be strictly concave and to satisfy the usual differentiability requirements

(𝑖,𝑗 = 1, 2, …, 𝑛), but consumption of good 𝑖 (𝑖 = 1, 2, …, 𝑛) may depend on the

consumer’s health status 𝑢𝑖,𝑛 + 1 ≠ 0. Consumption good 𝑖 is a complement to health

if 𝑢𝑖,𝑛 + 1 > 0 and is a substitute for health if 𝑢𝑖,𝑛 + 1 < 0.

To clearly show the channels through which pollution affects consumption, we

That is 𝐻 = 𝐻(𝐴). The individual faces a budget constraint of the form:

The individual’s problem is to maximize (1) through his choices of 𝑐 subject to

𝑢𝑖 ‒ 𝑝𝑖 = 0 (𝑖 = 1, 2, …, 𝑛), (3)

reduces the health level, consumption on good 𝑖 would be further depressed if it is a

∑𝑛 𝑝 ∂𝑐 /∂𝐴 = 𝑤'𝐻' (6)

marginal utility of income increases in air pollution. Intuitively, individual utility is

The model can be further generalized by allowing health to be dependent on the

consumption of goods 𝑐, i.e., 𝐻 = 𝐻(𝑐,𝐴). Say 𝑐𝑖 is the consumption of a pure utility-

and the data on air pollution and other weather variables.

3.1. Tracking Records of Inmates

sentence length was 4.6 years. er

corresponding working period.

ordered commodities. The data allows us to calculate each inmate’s monthly

3.2. Air Pollution

(151-200), Heavily Polluted (201-300) and Severely Polluted (above 300).

face towels and other personal care items.

3.3. Thermal Inversions

We obtained thermal inversion data from the Modern-Era Retrospective analysis

for Research and Applications version 2 (MERRA-2) released by the National

for detailed AQI calculation formula and explanation.

also conduct a robustness check by coding inversions using differences in temperature

linearity of temperature (Deschênes and Greenstone, 2011), we discretize the daily

temperature distribution using 5°C temperature bins.

3.5. Summary Statistics

Table 1 presents summary statistics of key variables. There are 433,369

11 The data can be seen at http://data.cma.cn/.