Satellite Detection of Air Pollution: Air Quality Impacts of Shale Gas Development in Pennsylvania

Abstract

This paper estimates the impact of shale gas development on local particulate matter pollution by exploiting a quasi-experimental setting in Pennsylvania where some wells experienced pre-production and/or production activities whereas some others were permitted but not spud between 2000-2018. We measure local PM pollution using daily aerosol optical depth (AOD) over a 3 kilometers circular area around every shale gas well. Using a spatial difference-indifferences model, we find that both shale gas pre-production and production activities increase daily AOD significantly, by 1.35%-2.19% relative to the baseline. The effect of pre-production is slightly larger than production activities, but both effects attenuate with distance from the centroid well. Accounting for airborne spillovers, fracking increases AOD by 1.27%-5.67%, which translates to 0.017ug/m 3-0.062ug/m 3 increase in PM 2.5 concentration. This increase in PM 2.5 is associated with 20.11 additional deaths.

Full text

Satellite Detection of Air Pollution: Air Quality Impacts of Shale
Gas Development in Pennsylvania
Ruohao Zhanga,∗, Huan Lib, Neha Khannaa,
Daniel M. Sullivanc, Alan J. Krupnickd, and Elaine L. Hill e
aDepartment of Economics, Binghamton University, Binghamton, NY
bDepartment of Economics, North Carolina A&T State University, Greensboro, NC
cJ.P. Morgan Chase, Washington DC.
dResource for Future, Washington DC.
eDepartment of Public Health Sciences, University of Rochester, Rochester, NY
This version: May 21, 2020
Abstract
This paper estimates the impact of shale gas development on local particulate matter pollu-
tion by exploiting a quasi-experimental setting in Pennsylvania where some wells experienced
pre-production and/or production activities whereas some others were permitted but not
spud between 2000 – 2018. We measure local PM pollution using daily aerosol optical
depth (AOD) over a 3 kilometers circular area around every shale gas well. Using a spatial
difference-in-differences model, we find that both shale gas pre-production and production
activities increase daily AOD significantly, by 1.35% – 2.19% relative to the baseline. The ef-
fect of pre-production is slightly larger than production activities, but both effects attenuate
with distance from the centroid well. Accounting for airborne spillovers, fracking increases
AOD by 1.27% – 5.67%, which translates to 0.017µg/m3–0.062µg/m3increase in PM 2.5
concentration. This increase in PM 2.5 is associated with 20.11 additional deaths.
Key words: shale gas, particulate matter, aerosol optical depth, spatial DID, mortality
impact
JEL codes: I15 I18 Q52 Q53 R11 R12
∗Corresponding author. Tel. +1 (). E-mail: rzhang47@binghamton.edu.
1

1 Introduction
The United State’s shale gas industry has developed rapidly in the past decades, growing from
1.6% of total natural gas production in 2000 to 69% in 2018 (Sieminski,2014). The boom in
shale gas production is largely due to the application of a new technology known as hydraulic
fracturing. Recent work has shown that the hydraulic fracturing technology and its massive
deployment impacts not only local economic but also local environmental conditions, including
ground water contamination (Hill and Ma,2017;Osborn et al.,2011;Jackson et al.,2013) and
chemical exposures to surface water (Olmstead et al.,2013). In addition, various air pollutants
including CO, NOx, SOx, particulate matter (PM), and volatile organic compounds (VOC) are
released to the air from unconventional wells’ preparation and fracking operations (Allen et al.,
2014;Litovitz et al.,2013). However, the effects of shale gas development on local air pollution
have not been evaluated.
We examine whether shale gas development in Pennsylvania has led to detectable changes in
local PM pollution between 2000 and 2018 and the magnitude of those changes. Pennsylvania is
the largest producer of shale gas among all states, and produces almost 30% of total shale gas in
the whole country (EIA Website,2019). Our particular focus on PM rather than the other types
of air pollutants is important for two reasons: First, there is documented public concern that
shale gas drilling activities contribute to local PM pollution (Litovitz et al.,2013), yet there is
little causal evidence linking the two. Furthermore, while the literature has documented health
effects on populations living close to unconventional wells (Currie et al.,2017;Hill,2018), the
channels explaining these effects are uncertain. PM pollution has known adverse health impacts
(Atkinson et al.,2014), so understanding the causal effects of shale gas development on local
PM pollution is relevant to policy.
Since the shale gas wells usually locate in the rural areas which are mostly not covered
by the ground-based air quality monitors, we use a satellite based remote sensing data, daily
Aerosol Optical Depth (AOD) data, as an indicator of PM concentration. AOD is a unit-less
high-frequency and high-resolution measurement of PM concentration provided by specialized
instruments on NASA’s satellites. AOD measures the degree to which aerosols prevent the
transmission of light by absorption or scattering of light through the entire vertical column of
the atmosphere from the ground to the satellite sensors; therefore a higher value of AOD implies
2

higher concentration of PM pollution (Liu et al.,2004;Donkelaar et al.,2016). While the daily
AOD captures the short-lived nature of the pollution (Sarigiannis et al.,2017), the measurment
is sensitive to weather (Kumar et al.,2007;Foster et al.,2009). To solve this problem, we
include as regressors a set of daily weather variables including precipitation, temperature, and
dew point. We use AOD to estimate the pollution of shale gas industry, and covert the pollution
in AOD to PM 2.5 using Lee et al. (2011)’s method for the purpose of estimating the mortality
impact.
We define our study unit, named “Pollution area” (P-area), as a circular area of 3 kilometers
radius around each unconventional well. The effect of wind on PM is captured econometrically,
as detailed below. We expect that during the well pre-production phase, PM pollution will
increase because well preparation activities, such as drilling and the associated commercial
vehicle traffic, bring dust and diesel combustion to well sites and nearby roads. Similarly, we
expect elevated PM pollution during gas production from on-site diesel combustion (Litovitz
et al.,2013) and fugitive emissions.1Since the two periods may have different impacts on
local PM pollution, we divide the life cycle of unconventional wells into three phases: inactive,
pre-production, and production, and focus on the effects of the pre-production and production
treatments on local PM pollution.
To estimate the causal effects of the two treatments on local PM pollution in the vicin-
ity of hydraulically fractured wells, we need to construct the appropriate counterfactual. The
Pennsylvania Department of Environmental Protection reports that not every well that is per-
mitted eventually gets spudded and drilled (PA DEP Web,2012). We use the PM pollution in
the vicinity of these wells to construct the counterfactual because they are not associated with
pre-production or production activities but are likely located in areas with similar geology and
social-economic conditions as the wells that are spudded or producing. On the basis of this,
we assign P-areas to the treatment group if their centroid wells have ever been in either the
pre-production or production treatment, while P-areas are included in the control group if their
centroid wells were permitted but never started the pre-production phase (i.e. not spudded).
Such a quasi-experiment setting allows us identify the causal difference-in-differences estimates
of the effect of the two treatments on local air quality in the vicinity of hydraulically fractured
1Fugitive emissions include benzene, toluene, ethyl-benzene, xylem other toxic hydrocarbons (Srebotnjak and
Rotkin-Ellman, 2014). These aerosols can interact with sunlight and water vapor to form liquid particles,
and are considered the secondary source of PM pollution.
3

wells.
It is important, however, to note that the estimation of the treatment effects is complicated
by the potential “spillover” of one well’s activities on its neighboring wells’ P-areas. The spillover
may arise from two channels. First, because PM pollution is air borne and travels with wind,
a well’s pre-production and production operations may increase PM pollution in downwind P-
areas. Second, a cluster of wells may share infrastructure, such as road access to the fracking
site, pipelines, waste pits, and other facilities needed for the fracking operation, thereby lowering
the marginal change in PM pollution at a new well’s P-area. To deal with the spatial spillover
effects, we implement a spatial difference-in-differences model (Delgado and Florax,2015), that
allows the potential outcome of spatial units to be affected by not only their own treatment
status, but also their neighboring units’ treatment through spillovers. The model uses daily
information on wind speed and direction to model the potential airborne spillovers from one
well to another. We account for the second channel by including the number of wells in the
pre-production phase, in production, or plugged in the 20 km radius around every well area as
control variables in our regression.
Our data set includes all 20,677 unconventional wells in PA between February 24, 2000
and September 20, 2018. We have an unbalanced panel with 11,470 areas (17,506,147 area-
day observations) in the treatment group and 9,207 areas (14,025,840 area-day observations)
in the control group. Among the treatment group observations, we have 1,004,184 area-day
observations in the pre-production phase and 3,607,236 area-day observations in the production
period.
We find statistically detectable changes in daily AOD during both the pre-production and
production phases of a marginal unconventional well. Not surprisingly, the marginal increase
in AOD is higher during the pre-production phase (2.19% relative to the baseline AOD) than
during production (1.35% of baseline). Furthermore, while the airborne spillover effects decline
with distance from a centroid well, they can be felt not only within the P-area but also as far
as 10 km downwind. Accounting for airborne spillovers, fracking increases AOD by 1.27% for
the whole sample, and by 5.67% for the subsample of P-areas with a treated well. Based on
Lee et al. (2011), these overall increases in AOD imply that daily PM concentrations increased
by 0.017µg/m3and 0.062µg/m3, respectively, in the average P-area. Using the concentration
4

response functions in Lepeule et al. (2012) and Fowlie et al. (2019), we estimate that this
resulted in an additional 20 deaths between 2010 and 2017 in 671 census block groups (across
40 counties) which contains at least one shale gas well, with a total population of about 840,000
and an annual average death rate of 12 per 1000.
Our results are relevant for policy makers who seek to understand the welfare effect of shale
gas development. They are also relevant for the communities located close to shale gas wells in
terms of understanding the local air quality impact.
2 Link between Shale Gas Development and Local Air Pollution
2.1 The Link
The innovation of hydraulic fracturing and horizontal drilling technology decreased the produc-
tion cost of shale gas significantly, making unconventional production economically feasible and
boosting the size of the shale gas industry. The rapid expansion of natural gas development
increased the supply of natural gas and lowered the prices relative to the scenario without hy-
draulic fracturing (Newell and Raimi,2014). Abundant natural gas with relatively lower prices
has facilitated the displacement of coal to natural gas in power plants, leading to air quality
improvement. At the same time, relatively lower prices encourage more consumption of energy.
Newell and Raimi (2014) show that the boom of shale gas development has reduced green house
gas emissions in the US, which is driven by the fact that the retirement of coal-fired power
plants dominates the effects of increased energy consumption.
Despite the positive global and regional environmental externality, the shale gas boom has
raised concerns regarding local air quality because of the extensive activities associated with
well preparation and gas production. Most fracking activities come with diesel combustion and
dust, increasing emissions of ambient pollutants like CO, NO, hydrocarbons, PM, etc. It usually
takes several months to complete well preparation (Hill,2018). Activities include building roads,
clearing sites, and transporting heavy equipment. Fracking requires a large amount of heavy
equipment, such as drilling rigs, high-volume fracking pumps, and large size storage tanks,
and this equipment typically arrives at the site on heavy trucks. According to Graham et al.
(2015), it takes roughly 1,500 heavy-truck trips to deliver equipment and materials to a site and
to remove the construction and drilling wastes from the site. In addition, off-road heavy-duty
5

engines are used to construct drill rigs and hydraulic fracturing pumps (Roy et al.,2014). When
the construction is completed, a completion venting is performed for cleaning and bringing the
well to production (Roy et al.,2014). The venting occurs multiple times during the whole
life-cycle of an unconventional well as a maintenance procedure, and is known to be a major
source of of volatile organic compounds (VOCs) emissions from unconventional wells.
During the production period, on-site equipment, including compressors to maintain the
pressure of produced natural gas and other diesel machinery for well maintenance, result in
diesel and natural gas combustion and air pollution. Additionally, when gas is flared, vented, or
accidentally leaked during production, it also releases toxic air pollutants. Toxic air pollutants
and VOCs including benzene, toluene, ethyl-benzene, xylem other toxic hydrocarbons come from
direct and fugitive emissions of hydrocarbons at the well and from associated infrastructure such
as condensate tanks (store liquid separated from produced natural gases), dehydrators (remove
water from the produced natural gas), waste water impoundment pits, and pipelines (Srebotnjak
and Rotkin-Ellman,2014). In addition, many of these aerosols can interact with sunlight and
water vapor to form liquid particles, and are considered a secondary source of PM pollution. In
fact, secondary aerosols contribute a large portion of total PM, and are the dominating source
of PM in many cases (Larsen et al.,2012;Heo et al.,2009;Lewandowski et al.,2008;Huang
et al.,2014).
2.2 Local Air Pollution Measurements
Since many of the pollutants emitted during the well’s pre-production and production phase
are either primary or secondary sources of PM, we use PM to indicate the impact of shale
gas development on local air quality. PM measurements are available from multiple sources,
including the EPA’s ground-based mobile monitoring, network stations, aircraft measurements,
and a satellite platform as summarized in Field et al. (2014). Only the satellite platform provides
a hyperlocal, daily measurement of PM. We therefore take advantage of NASA’s satellite based
measurements of aerosal optical depth (AOD), which is a high-frequency and high resolution (3
km ×3 km) measure, and is known to be one of the most robust aerosol parameters retrieved
by the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA’s satellites (Streets
6

et al.,2013).2 3
More importantly, the literature has shown that AOD is a good predictor of PM of different
sizes: PM2.5(size <2.5µm) and PM10 (size <10µm) (Liu et al.,2004;Donkelaar et al.,2016).
Higher AOD indicates worse air quality, and therefore higher PM pollution. An advantage of
using AOD is that it offers daily air quality observations with high geographical resolution. The
spatial scale of the air pollution from shale gas wells is small. Companies conduct the drilling
process on about a 3-acre pad of land, with a number of trucks that become part of an oil
and gas drilling process. Given the sources of air pollution are from truck traffic and on-site
construction and production process, we focus on a 3 km circular area around each single well.
Our use of AOD as a measure of air quality is not unique in the economics literature. Zou
(2019) studies the current EPA policy of intermittent monitoring of environmental standards,
and uses AOD to measure air quality when ground monitoring is off. His study finds air quality
significantly worse on unmonitored days. Foster et al. (2009) studies the air pollution impact of
a voluntary pollution reduction program in Mexico, and its consequence on infant health. They
find a significant drop in AOD (increase in air quality) associated with the program, along with
a significant drop in infant mortality due to respiratory illness associated with the decreasing in
AOD. Sullivan and Krupnick (2019) uses AOD to measure the air quality in individual counties
across the US, and argue that due to the limited number of ground monitors, many counties
are mistakenly assigned as being in “attainment” with the 2015 National Ambient Air Quality
Standards for PM. They estimate that 24.4 million people live in attainment areas that AOD
data suggests should be in nonattainment. A similar result is also found by Fowlie et al. (2019).
3 Empirical Model
3.1 Estimating the Average Treatment Effects
Our baseline model follows a difference-in-differences framework:
qid =ηcTc
id +ηpTp
id +Aid
0Λ + Zid
0ζ+µi+σd+uid (1)
2There are two NASA satellites with MODIS instruments: Aqua and Terra. We use Terra because it has a
longer observation period (start from Feb. 2000, whereas Aqua starts from May 2002).
3AOD is generated by the following method: the remote sensors record the interaction between electromagnetic
radiation and aerosols including solid and liquid particles in the atmosphere, then convert the recorded results
to AOD by applying the radioactive transfer models (Remer et al.,2005).
7

Here, qid represents the AOD of P-area ion date d.Tc
id and Tp
id are treatment dummy variables,
with Tc
id = 1 and Tp
id = 1 indicating the centroid well of P-area iis in the pre-production or
production period respectively on date d, and zero otherwise. The two way fixed effects are
P-area fixed effect µiand date fixed effect σd; thus in the difference-in-differences framework, ηc
and ηcestimate the average treatment effects of the pre-production and production treatments,
respectively.
Z0
id and A0
id are vectors of covariates. Z0
id contains four weather variables: precipitation,
dew point, temperature, and wind speed. A0
id includes three additional variables to address the
density of wells in the surrounding 20 km radius around every centroid well of P-area i. They
are the daily counts of wells with pre-production, production, and inactive status.4Together,
these variables account for the potential sharing of infrastructure, such as main roads, rigs and
pipelines in the fracking area across wells. At the same time, the density of inactive wells –
wells that used to be in the pre-production or production phase before date, but do not have
on site activities anymore, reflects temporal variation in local geological conditions of shale gas
stock that we expect is correlated with the productivity of wells in the area.
3.2 Estimating the Average Treatment Effects with Spatial Spillovers
Motivation: spillover path and quantification
PM travels with wind, bringing a well’s pre-production and production treatment effects to
downwind areas. This spillover effect would not be a concern for us if it affected the neighboring
P-areas’ air quality randomly. Non-randomness in wind blown spillover may arise from spatial
segregation between treatment and control groups or temporal variation of wells’ pre-production
and production treatments.
Figure 1shows the location of permitted wells in PA. While there is negligible regional
segregation between treatment and control groups (see Figure 2and 3), the temporal variation
in the pre-production and production phases among wells (see Figure 4) suggests that the
spillover effects might be nonrandom in our study setting.5
4Inactive wells are those wells that are without any operations.
5The pattern of nonrandom spatial and temporal spillovers is more clear in Table 1,Panel II. The summary
statistics show that the densities of wells in the pre-production and production phases are much higher
around treated group than control group.
8

Figure 1: Permitted Wells in 2000 – 2018
Figure 2: Wells in Treatment Group
9

Figure 3: Wells in Control Group
Figure 4: Well Location by Permit Years
We follow Delgado and Florax (2015)’s model to handle the nonrandom spatial spillover.
Delgado and Florax (2015) account for a spatial spillover in potential outcome through a binary
time-invariant spatial weight matrix, which equals 1 if the neighboring area is within a threshold
10

distance, and zero if not. In our setting, we expect the spillover effects through wind to attenuate
by distance. Therefore, we distinguish 3 circular rings around each P-area’s centroid well: 0–2
km, 2–5 km, and 5–10 km (hereafter referred to as bins), and allow the spillover effects from
the upwind wells’ pre-production and production activities to differ by bins. In addition to
distance, wind direction matters too. The closer the wind direction to the geographic direction
of the two wells, the stronger the spillover effects would be, conditional on the distance between
the two wells. This can be addressed by the angle between any two wells’ geographic direction
and the wind direction. Therefore, our spatial weight matrix is continuous and various by date,
because the wind direction is different every day.
Figure 5: Pollution Transportation by Wind
Define wbin
id as a 1 ×Nmatrix (spatial weight matrix) for the P-area with centroid well ion
day d, where Nis the total number of centroid wells (of both treatment group and control group
P-areas) in our sample. The jth element in wbin
id , indicated by wbin
ijd , measures the magnitude
11

of the spillover effects that the P-area with centroid well iwould receive from the well jin the
bin bin on day d. We allow wbin
ijd to vary over time depending on jth well’s treatment status
on day dand whether it is located upwind of well i. To quantify wbin
ijd , we use the two wells’
geographical locations and wind direction on the day d. As shown in Figure 5, suppose θij d is
the angle between the wind direction and the geographical direction between the wells jand i
on day d, and xij is the perpendiculars distance between the two wells. Then wbin
ijd is defined as:
w0−2
ijd =cos(θij d) if θij d ≤πand 0 < xij ≤2, w0−2
ijd = 0 otherwise
w2−5
ijd =cos(θij d) if θij d ≤πand 2 < xij ≤5, w2−5
ijd = 0 otherwise
w5−10
ijd =cos(θij d) if θij d ≤πand 5 < xij ≤10, w5−10
ijd = 0 otherwise
The definition of wbin
ijd follows the spillover path: it implies that the weight is zero if the
the well jis downwind of the P-area i, and the weight is positive if the the well jis located in
the upwind of the P-area i. A smaller angle between the wind direction and the geographical
direction from well jto P-area imeans the treatments from well jhas larger effect on P-
area i. The design of spatial weight matrix follows the spirit of the Gaussian point source
dispersion model6, in which the aerosol travels along the downwind direction, and diffuses along
the crosswind direction.
Model specification
In the light of Delgado and Florax (2015), we use the following spatial difference-in-differences
model as our preferred benchmark model:
qid =ηcTc
id +ηpTp
id +X
3bins
βbin,cwbin
id Hbin,c
d+X
3bins
βbin,pwbin
id Hbin,p
d
+Aid
0Λ + Zid
0ζ+µi+σd+uid
(2)
Where qid,Tc
id,Tp
id,A0
id, and Z0
id have the same definitions as in Equation (1).
wbin
id is the spatial weight matrix, Hbin,c
dis a vector of dummies for all wells, in which an
element is equal to 1 if and only if the well is in P-area i’s bin and this well was in the pre-
production phase on day d. In matrix notion, wbin
id Hbin,c
dis a weighted sum of all pre-production
treatments belonging to a given bin bin for P-area ion day d, indicating the spillover effects of
6Gaussian point source dispersion model is a fundamental model in atmospheric science. See Wikipedia page.
12

pre-production treatment that P-area ireceived from the bin on day d. The interpretation can
be generalized to the other 5 variables.
4 Data
We create a comprehensive data set that includes every well that was permitted in Pennsylvania
between February 24, 2000 and September 20, 2018. Our data set includes detailed information
on each well and daily information on air pollution and weather from multiple data sources.
4.1 Pennsylvania Shale Gas Data
Our shale gas well data is compiled from two sources published by Pennsylvania Department
of Environmental Protection: “Oil Gas Locations – Unconventional” and “Oil & Gas Well
Production Report”. These sources include information submitted by well operators, as required
by PA DEP (Regulation Code Section 78a.121). Both data sources contain a unique well
identifier and well coordinates, allowing us to merge information across these sources.
For each well we have information on its geographic coordinates and permit date, regard-
less of whether the well is spudded or not. For wells that were drilled, we have current well
status, along with spud date, production period, and completion date if already plugged. This
information allows us to determine each well’s activities on any given day. Given our focus on
air quality, we consider three different phase in the life-cycle of a well: not yet spudded (or
inactive), pre-production phase, and production phase.
In total, we obtained 20,677 unconventional wells that were permitted over our study period.
As shown in Figure 1, most of these permitted wells are located in the northeastern corner
PA (Susquehanna and Bradford counties) and southwestern corner (Washington and Greene
counties). This is because the depth of Marcellus Shale base in these regions ranges from 5,000
to 8,500 feet, higher than other parts of Pennsylvania, suggesting that these areas are especially
productive (PSU Web,2009).
4.2 Local Air Quality: MODIS AOD
There are two NASA satellites with MODIS instruments: Aqua and Terra. We use the data
provided by Terra because it has a longer observation period (starting from Feb. 2000, whereas
13

Aqua starts from May 2002). There are 4 levels of MODIS data available: L0 to L3. The higher
level means the data is more pre-processed, but has lower spatial resolution. We use level 2
data7because it provides daily AOD observations at 3km ×3km pixel resolution8. L2 data not
only satisfies the requirements of our research design, but is also properly processed and can be
directly used in our analysis.9
With the prior that the pre-production and production periods primarily affect local air
quality, we define the P-area, our study unit, as an area surrounding a well with the 3km radius.
There are two difficulties in determining a P-area’s AOD (i) L2 AOD data is for 3km ×3km
square pixels, whereas P-areas are 3km radius circles. (ii) The pixels in the L2 data change
every day according to the orbit of the satellite. We overcome these difficulties as follows: First
we overlap all pixels with P-areas to find the square-circle intersections between pixels and every
particular P-area for each day separately. Second, We assign every pixel a daily weight based
on the daily intersection area to calculate the daily weighted average of AOD for each P-area.
For example, suppose on a day d, a P-area overlaps with pixel 1 and pixel 2 only. Pixel 1
has x1dkm2intersection area with P-area on day dand has AOD equals to q1d. Pixel 2 has
x2dkm2intersection area with P-area and AOD of q2d. Then on day dthe weight of pixel 1
is x1d
x1d+x2d, and the weight of pixel 2 is x2d
x1d+x2d. The weighted average AOD of P-area becomes
q1dx1d
x1d+x2d+q2dx2d
x1d+x2d. In general, where a P-area overlaps with J pixels on day d, the weighted
average AOD for P-area iis:
AODid =q1dx1d
x1d+x2d+· · · +xJ d +· · · +qjd xjd
x1d+x2d+· · · +xJ d 
+· · · +qJ d xJ d
x1d+x2d+· · · +xJ d 
4.3 Weather
Our key control variables include daily information on local weather. We do so for three reasons:
First, the weather variables account for the possible correlation between weather conditions and
the choice of spud date. Second, the weather variables are important confounders of the strong
7We use version 6.1 MOD04 3K HDF data file. In the data file, we choose the layer “Cor-
rrected Optical Depth Land”.
8The level 2 data also provide 5km ×5km resolution.
9L0 is raw spectral channel, and L1 is calibrated and geolocated radiance. Neither of them can be directly
used. L3 also provide AOD , but the resolution becomes 1◦×1◦global grid, and the data is either 8 days
or 1 month.
14

association between AOD and PM2.5(Kumar et al.,2007;Foster et al.,2009), and both pre-
production and production activities affect AOD through PM2.5. Third, wind information helps
us address the spatial spillover, which occurs due to the spatial clustering of wells in PA.
The weather data is from the Parameter-elevation Regressions on Independent Slopes Model
(PRISM), a spatial climate database. PRISM provides weather data including precipitation,
mean temperature and mean dew point temperature at a daily frequency. One advantage of
PRISM is that it is based on a spatial resolution of 4 km2pixels, which is comparable with the
size of a well’s P-area and the spatial resolution of AOD data. We process the daily precipitation,
temperature and dew point data in a similar manner to the AOD data. That is, we overlay the
4km2grid of weather data with P-areas and calculate the weighted average for each P-area
using the (time-invariant) interaction areas as weights.
Daily information on wind speed and direction is from the National Centers for Environ-
mental Prediction (NCEP)-U.S. Department of Energy Reanalysis II (NCEPRII). These daily
data sets are available at resolution of 2.5 degree in latitude and longitude. We assign wind
information to each P-area based on the 2.5 degree square that the centroid well is located in.
Wind speed serves as an additional control variable, and wind direction is used to address the
spatial spillover effects.10
5 Results
5.1 Model Estimation
Defining the Pre-production and Production Treatments
The duration of the two phases is unique to each well. Hence, we use each well’s timeline: permit
date, spud date, drilling and production phase, and plug date, to define its two treatments.
Specifically, the first day of the pre-production phase is set either on the spud date or on the
permit date, depending on which date is later.11 Since the data sources do not report the
10The wind is decomposed into two component: U wind speed and V wind speed. U wind is the east-west
component of wind. Positive U wind means the wind is from west to east, and negative U wind is from east
to west. V wind is the north-south component of wind. Positive V wind means the wind is from south to
north, and negative V wind is from north to south. The combination of U wind and V wind provides the
wind direction and wind speed.
11In general, a well should be permitted before being spud, but there are few cases in which permit date was
later than the spud date.
15

end of the pre-production phase, we set it as the day before the first day of production, under
the assumption that wells start to produce immediately after the pre-production phase ends.
See Appendix A.1 for additional details on how we use the production report to assign the
pre-production and production treatment.
Clustered the Standard Error
When estimating the model, we cluster the standard errors by well pads. A well pad may
consist of multiple wells, where P-areas frequently overlap. Yet, in our sample, 7,884 out of
20,677 unconventional wells do not have well pad information. We therefore assume that wells
close to each other are located on the same well pad, and artificially assign the well pad ID
to every individual well. In particular, any well that is closer than 63m to any other well is
designated to be in the same well pad (Muehlenbachs et al.,2015). By comparing artificially
designated well pad ID and original well pad ID if any, only 1.9% of wells with original well pad
ID is mistakenly assigned into different artificial well pad.
5.2 Descriptive Statistics
The final sample is an unbalanced panel data consisting of 22,067 wells on 4,691 days, from
February 24, 2000 – September 20, 2018. Out of the 22,067 permitted wells, 11,470 wells
experienced pre-production or/and production phases and are included in the treatment group,
the remaining 9,207 wells were not spudded and are included in the control group.12 The
majority of the wells in our sample were permitted after 2007 and are located in southwestern
and northeaster corner of PA, as seen previously in Figure 4.
For the wells that have been spudded (i.e., wells in the treatment group), the pre-production
phase lasts on average 406 days. Once the pre-production phase ends, wells produce for as long
as 1,805 days (5 years), on average, until the production is completed. Table 1presents summary
statistics for our main analysis sample. We compare the mean statistics from the full sample
and subsamples with treatment group observations in the pre-treatment period, pre-production
period, production period, and control group observations separately. The mean daily AOD is
0.23 per P-area, with a standard deviation of 0.26. As shown in Table 1, through the whole
12In our data set, out of the 11,470 wells in the treatment group, 2,184 wells experienced pre-production phase
only, 94 wells had information on production phase only, and 9,192 wells experienced both phases.
16

Table 1: Descriptive Statistics
Panel I: Key Variables
Treated
Full Sample Pre-Treatment Construction Production Control
AOD 0.2275 0.2397 0.2005 0.1946 0.2267
(0.2608) (0.2726) (0.2215) (0.2160) (0.2619)
Precipitation (mm) 1.2192 1.2141 1.1974 1.2287 1.2230
(4.4304) (4.4876) (4.0946) (4.2256) (4.4521)
Temperature (Celsius) 12.3195 12.3527 12.0260 12.4069 12.2875
(7.9863) (7.8705) (8.3065) (8.3144) (7.9817)
Dew Point (Celsius) 5.4960 5.4368 5.2807 5.7628 5.4972
(8.5333) (8.4274) (8.7536) (8.8462) (8.5307)
Wind Speed (m/s) 3.9783 3.9655 3.9954 3.9791 3.9886
(2.1202) (2.1165) (2.1345) (2.1190) (2.1230)
Panel II: Spillover Variables
Treated
Full Sample Pre-Treatment Construction Production Control
Pre-production treatment 0.2227 0.1115 1.8842 0.3392 0.1760
in 0-2 km ring (0.9799) (0.6676) (2.5011) (1.2325) (0.8339)
Pre-production treatment 0.7874 0.5145 2.7146 1.6458 0.6796
in 2-5 km ring (2.3467) (1.9266) (4.1321) (3.2900) (2.1039)
Pre-production treatment 2.2152 1.4814 6.0632 4.6138 1.9974
in 5-10 km ring (5.2077) (4.3891) (7.8726) (6.8318) (4.8720)
Production treatment 0.8618 0.2343 1.1749 3.8689 0.6429
in 0-2 km ring (2.2909) (1.1308) (2.3953) (3.9253) (1.8729)
Production treatment 3.0912 1.1458 5.8850 11.5251 2.5108
in 2-5 km ring (7.2320) (4.2817) (8.8516) (11.4143) (6.2403)
Production treatment 8.2211 3.3991 16.1812 27.5720 7.1076
in 5-10 km ring (18.5921) (11.5792) (23.0690) (28.6077) (16.7869)
Density Pre-production 30.8528 20.2306 87.1349 62.1165 28.5484
(Number of wells, 20 km) (50.6045) (42.4766) (63.8144) 56.6762 48.9225
Density Production 113.8012 47.4958 223.1352 363.2004 103.0471
(Number of wells, 20 km) (203.3654) (121.7000) (226.0386) (262.6684) (192.5515)
Density Inactive 17.5413 7.4077 30.1685 54.2171 16.5211
(Number of wells, 20 km) (35.1427) (24.4105) (36.2823) (49.1666) (32.5558)
Number of observations 31,531,987 12,894,727 1,004,184 3,607,236 14,025,840
Number of wells 20,677 11,470 11,376 9,286 9,207
AOD is missing for all observations receiving pre-production treatments with centroid well OBJECTID 383270.
107 wells are transformed from conventional to unconventional, their pre-production phases are not considered.
17

study period (Feb. 2000 – Sep. 2018), the treated areas’ pre-treatment P-areas’ average AOD
is higher than control, but treated P-areas’ average AOD is smaller. This might be explained
by the combination of (i) the majority of wells were constructed and producing in later years
and (ii) the AOD is declining over time in our sample, as shown in the first panel of Figure A1.
The mean statistics of weather variables are relatively stable across full sample and subsamples.
The rest of Figure A1 describes the yearly trend of weather variables.
In Panel II of Table 1, we summarizes the mean statistics of the spatial spillover variables.
The non-random spatial clustering is clearly shown: subsample of pre-production observations
has the largest mean values for the spatial spillover treatments of pre-production across all 3
distance bins as well as the highest density of wells in the pre-production phase in a 20km
radius around centroid well. Likewise, the subsample of production observations has the largest
mean value of spatial spillover treatments of production across all 3 distance bins as well as
the highest density of wells under production in a 20km radius around centroid well. These
statistics suggest the existence of the nonrandom spatial and temporal spillovers. We also find
the subsample of production observations has a higher average density of inactive wells than
the other samples.
5.3 Testing the Common Trend Assumption
The validity of our difference-in-differences estimation depends on the credibility of our control
group P-areas to provide a reliable counterfactual. One way to assess the credibility is to test if
the treatment and control groups share a common trend prior to the occurrence of the treatment.
To test common trend assumption, we compare the daily differences between pre-treatment
observations from the treated group and their corresponding control group observations on the
same dates. Because we have staggered treatments on a large number of treated P-areas, the
pre-treatment observations from the treated group covers all dates of our sample period, so all
control observations are used in the test. We take a subset of data that consists of pre-treatment
observations for treated P-areas and all observations for control group P-areas, and estimate
the following model:
qid =Aid
0Λ + Zid
0ζ+µi+σd+σd×Ti+uid,(3)
18

where Tiis the group fixed effect, with Ti= 1 indicating P-area iis in the treated group (i.e.,
pre-production or production), and Ti= 0 indicating P-area iis in the control group. σdis
the day dummies. The coefficients on the interaction term σd×Ticapture the difference in the
time trend in AOD between the treated and control groups. Standard errors are clustered by
pad. Under the null hypothesis of a common trend, the coefficients on σd×Tiare insignificant.
Rejecting the null hypothesis implies the absence of the common trend of AOD between the
two groups.
Figure A2 shows the estimated coefficients of σd×Tialong with 99% confidence intervals.
Because 87% of the coefficients are not significantly different from zero, we conclude that the
pre-treatment trends in AOD in both groups are similar, and thus the P-areas with permitted
but not spudded centroid wells serve as a suitable control group, providing us an appropriate
counterfactual for pollution in the absence of shale gas development.
5.4 Main Results
Table 2summarizes the average treatment effect of each P-area’s own centroid well. Our baseline
DID model shows that, on average, a P-area’s AOD increases significantly when its centroid well
is in pre-production and production. However, the estimates confound the effect due to wind
blown pollution from upwind wells. Hence, our preferred benchmark estimates is based on the
spatial DID model. Similar to the baseline DID model, the spatial DID model also shows that
a well’s pre-production and production activities significantly increase the daily AOD in its P-
area. In terms of the magnitudes, the baseline model overestimates the treatment effect, possibly
because there is temporally non-random variation of pre-production and production activities
(see Figure 4), and the coefficients pick up the non-random airborne spillovers. According to
the results from the spatial DID model, pre-production activity significantly increases a well’s
P-area AOD by 0.00429, which is 2.19% relative to the average AOD in our sample. Similarly,
a well’s production activities significantly increases AOD in its P-area by 1.35%. At the same
time, we find that a well’s pre-production activities increase its P-area’s AOD more than the
production activities. Our results are similar in magnitude to Zou (2019), who found a 1.6%-
1.8% increasing in AOD when the ground PM monitor is off.
The results from the spatial DID model show that the spillover effects from wells in upwind
19

areas increase local AOD. Furthermore, while this effects attenuates by distance, both for pre-
production and production treatments, it is statistically detectable at least as far as 10km from
its emission source. Additionally, pre-production activities from upwind wells have a larger
spillover effect than production activities in all the three distance rings.
5.5 Overall AOD Impact of Shale Gas Industry
We use the estimates reported in Table 2to estimate overall increasing in AOD in each P-area
due to unconventional shale gas development. Let the overall AOD increases in P-area ion
date dbe sid, then
ˆsid = ˆηcTc
id + ˆηpTp
id +X
3bins
ˆ
βbin,cwbin
id Hbin,c
d+X
3bins
ˆ
βbin,pwbin
id Hbin,p
d.(4)
ˆηc, ˆηp,ˆ
βbin,c and ˆ
βbin,p are estimated coefficients from our main result shown in the last column
of Table 2. We find that the overall effect of fracking increases AOD by 0.00276 for the whole
sample, and 0.01031 for the subsample of observations under treatment. This represents a 1.27%
and 5.67% increase in AOD above the background level which we estimate as ˆ
bid =qid −ˆsid,
where qid is the observed AOD level.13
6 Robustness Checks
To evaluate the validity and robustness of the causal effects found in our main analysis, we
conduct several additional analyse. First, we consider the fact that unobserved confounders
such as the local communities’ preferences and bargaining power in leasing land to the shale
gas industry that correlate with both well location and social-economic conditions, so local air
quality may lead to non-random sorting of wells into treatment and control groups. To address
this, we consider three different subsamples: one includes wells from the treatment group only,
and the other two use two different matching criteria. Second, considering the uncertainty
regarding the start and end of the pre-production phase due to data limitations, we use two
alternative pre-production phase definitions. Finally, we conduct two falsification tests. All of
these sensitivity analyses support the average treatment affects in our main analysis.
13The percentage estimation is calculated by ˆsid/ˆ
bid.
20

Table 2: Summary results of average treatment effects: main analysis
DID Spatial DID
Treatments Coefficient Std. Err. Coefficient Std. Err.
Own Effects
Pre-production treatment 0.00624∗∗∗ (0.00040) 0.00429∗∗∗ (0.00039)
Production treatment 0.00463∗∗∗ (0.00040) 0.00258∗∗∗ (0.00036)
Spillover Effects
Pre-production treatments in 0–2 km ring 0.00101∗∗∗ (0.00010)
Pre-production treatments in 2–5 km ring 0.00050∗∗∗ (0.00004)
Pre-production treatments in 5–10 km ring 0.00008∗∗∗ (0.00002)
Production treatments in 0–2 km ring 0.00066∗∗∗ (0.00008)
Production treatments in 2–5 km ring 0.00026∗∗∗ (0.00003)
Production treatments in 5–10 km ring 0.00002∗∗∗ (0.00001)
Control Variables
Weather, well densities in 20km
circular area, Two-way FEs Y Y
Adjusted R20.76 0.76
Sample Size 31,531,987 31,531,987
Note: The full results are reported in Appendix Table A3. The daily weather controls are mean
precipitation, mean dew point, mean temperate and wind speed. The 20 km circular background
condition include numbers of wells of three different operation status: pre-production, production,
and inactive. Standard errors are clustered by well pads.
6.1 Using Treatment Group Only
Each well in the treatment group began pre-production or production phases at a different point
in time, staggering the treatments in our sample. As Goodman-Bacon (2018) and Athey and
Imbens (2018) have discussed, not only observations in the control group, but treatment group
observations during the pre-treatment period and post-treatment period can also provide a valid
counterfactual. For the sake of eliminating any confounding unobserved factors, we re-estimate
the standard and spatial difference-in-differences models using a subsample of P-areas from the
treatment group only.
The average treatment effects are reported in Table 3are highly consistent with the main
results: pre-production treatment and production treatment increase the AOD significantly; pre-
production increases AOD more than production activities; and spillover effects from upwind
wells’ pre-production and production operations are significant and attenuate with distance.
The magnitudes of the estimated coefficients are comparable to these in Table 2.
21

Table 3: Robustness check: using treatment group P-areas only
DID Spatial DID
Treatments Coefficient Std. Err. Coefficient Std. Err.
Own Effects
Pre-production treatment 0.00651∗∗∗ (0.00040) 0.00475∗∗∗ (0.00040)
Production treatment 0.00488∗∗∗ (0.00060) 0.00281∗∗∗ (0.00059)
Spillover Effects
Pre-production treatments in 0–2 km ring 0.00086∗∗∗ (0.00010)
Pre-production treatments in 2–5 km ring 0.00048∗∗∗ (0.00004)
Pre-production treatments in 5–10 km ring 0.00008∗∗∗ (0.00002)
Production treatments in 0–2 km ring 0.00060∗∗∗ (0.00009)
Production treatments in 2–5 km ring 0.00025∗∗∗ (0.00003)
Production treatments in 5–10 km ring 0.00003∗∗∗ (0.00001)
Control Variables
Weather, well densities in 20km
circular area, Two-way FEs Y Y
Adjusted R20.76 0.76
Sample Size 17,506,147 17,506,147
Note: The full results are reported in Appendix Table A4. The weather controls are daily mean pre-
cipitation, daily mean dew point, daily mean temperate, wind speed. The 20 km circular background
condition include numbers of wells of three different operation status: pre-production, production,
and inactive. Standard errors are clustered by well pads.
6.2 Using Matched Samples
Figure 4suggests non-randomness of the treatment across locations and permit issue dates.
For example, there is a cluster of wells in the northeastern part of PA receiving permit during
2012–2015, whereas there is another cluster of wells in the southwestern part receiving permit
after 2015. We trim this sample using two matching strategies to eliminate possible systematic
differences driven by such non-randomness. In both matching strategies, wells in the control
group are used repeatedly. The first strategy uses one-to-one matching by distance, permit year,
and permit month. That is, each P-area in the treatment group is matched with the closest
P-area in the control group, such that the centroid wells from the two P-areas are permitted in
the same month and same year. The second strategy is one-to-one matching by distance, permit
year, and county. In this case, each P-area in the treatment group is matched with its closest
P-area in the control group from the same county, and the centroid wells of the two P-areas were
permitted in the same year. These two matching strategies allow us to exclude control group
observations from P-areas that are temporally and geographically distant from the regional
clusters in the treatment group, so as to obtain a potentially more reliable counterfactual.
22

The first matching strategy gives us 11,368 wells in the treatment group and 3,000 wells in
the control group. The average distance between a matched treatment P-area and a control P-
area is 16.43 KM. The second matching strategy gives us 11,293 wells in the treatment group and
2,833 wells in the control group. The average distance between the matched treatment P-area
and control P-area is 3.45 KM. For both subsamples, we find that the AOD balance is slightly
improved, with the difference in the average AOD between the pre-treatment observations and
control observations reduced from 0.013 to about 0.011. Figure A3 and Figure A4 demonstrate
the locations of the wells in two matched subsamples. Using the two matched subsamples, we
re-run the DID and spatial DID model, with the results reported in Table 4. The results are
remarkably consistent with the main findings reported in Table 2.
Table 4: Robustness check: using matched samples
Matched sample 1 Matched sample 2
DID Spatial DID DID Spatial DID
Own Effects
Pre-production treatment 0.00609∗∗∗ 0.00436∗∗∗ 0.00624∗∗∗ 0.00452∗∗∗
(0.00039) (0.00038) (0.00039) (0.00038)
Production treatment 0.00449∗∗∗ 0.00262∗∗∗ 0.00475∗∗∗ 0.000291∗∗∗
(0.00051) (0.00048) (0.00051) (0.00049)
Spillover Effects
Pre-production treatments in 0–2 km ring 0.00093∗∗∗ 0.00093∗∗∗
(0.00010) (0.00010)
Pre-production treatments in 2–5 km ring 0.00043∗∗∗ 0.00048∗∗∗
(0.00004) (0.00004)
Pre-production treatments in 5–10 km ring 0.00008∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002)
Production treatments in 0–2 km ring 0.00061∗∗∗ 0.00060∗∗∗
(0.00008) (0.00009)
Production treatments in 2–5 km ring 0.00026∗∗∗ 0.00027∗∗∗
(0.00003) (0.00003)
Production treatments in 5–10 km ring 0.00002∗∗∗ 0.00002∗∗∗
(0.00001) (0.00001)
Control Variables
Weather, well densities in 20km
circular area, Two-way FEs Y Y
Adjusted R20.76 0.76 0.76 0.76
Sample Size 21,949,389 21,949,389 21,583,079 21,583,079
Note: The full results are reported in Appendix Table A5 and Table A6. The weather controls are daily
mean precipitation, daily mean dew point, daily mean temperate, wind speed. The 20 km circular back-
ground condition include numbers of wells of three different operation status: pre-production, production,
and inactive. Standard errors are clustered by well pads.
23

6.3 Using Extended Pre-production Phase
In the main analysis, we assume that pre-production operations start on the spud date. But, the
activities associated with pre-production phase might begin before the spud date. For example,
building roads for accessing the well site and delivering equipment to the well site takes several
weeks (Hill,2018), both of which usually take place before spudding a well. Taking this into
consideration, we extend the pre-production phase by one week and one month before the spud
date, respectively, and re-estimate our model under these new definitions of the pre-production
treatment.
Table 5reports the results. The results of both baseline DID and spatial DID are not
sensitive to the change of pre-production phase length. These results suggest the possibility
that some pre-production activities may have commenced months before spud date, but the
accuracy of our results stands.
Table 5: Robustness check: spatial DID using extended pre-production phase
Extended by 7 days Extended by 30 days
DID Spatial DID DID Spatial DID
Own Effects
Extended pre-production treatment 0.00626∗∗∗ 0.00434∗∗∗ 0.00620∗∗∗ 0.00433∗∗∗
(0.00040) (0.00039) (0.00039) (0.00038)
Production treatment 0.00466∗∗∗ 0.00260∗∗∗ 0.00470∗∗∗ 0.00264∗∗∗
(0.00049) (0.00046) (0.00049) (0.00046)
Spillover Effects
Pre-production treatments in 0–2 km ring 0.00101∗∗∗ 0.00101∗∗∗
(0.00010) (0.00010)
Pre-production treatments in 2–5 km ring 0.00050∗∗∗ 0.00050∗∗∗
(0.00004) (0.00004)
Pre-production treatments in 5–10 km ring 0.00008∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002)
Production treatments in 0–2 km ring 0.00066∗∗∗ 0.00066∗∗∗
(0.00008) (0.00008)
Production treatments in 2–5 km ring 0.00026∗∗∗ 0.00026∗∗∗
(0.00003) (0.00003)
Production treatments in 5–10 km ring 0.00002∗∗ 0.00002∗∗
(0.00001) (0.00001)
Control Variables
Weather, well densities in 20km
circular area, Two-way FEs Y Y Y Y
Adjusted R20.76 0.76 0.76 0.76
Sample Size 31,531,987 31,531,987 31,531,987 31,531,987
Note: The full results are reported in Appendix Table A7.
24

6.4 Falsification Test
To validate the causal nature of our estimates, we implement falsification tests by excluding
observations with true treatments and assigning placebo treatments on alternative days. We
consider 2 placebo tests with arbitrary placebo treatments: in the first placebo test, we assign a
placebo treatment from 1,080 days (about 3 years) before the pre-production phase to 720 days
(about 2 years) before pre-production phase. In the second placebo test, we assign a placebo
treatment from 1,440 days (about 4 years) before the pre-production phase to 1,080 days (about
3 years) before the pre-production phase. We assign a single placebo treatment for each test
and do not distinguish between placebo pre-production and placebo production treatments.
The results are reported in Table 6: the coefficients on the placebo treatments are statis-
tically insignificant, while the true spillover effects remains positive and significant. Thus, our
results survive the two falsification tests, showing little evidence that the true treatments effect
are distorted by AOD trends in the pre-treatment period.
Table 6: Robustness check: falsification tests
Placebo Test 1 Placebo Test 2
DID Spatial DID DID Spatial DID
Placebo Treatment 0.00019 0.00036 -0.00015 0.00008
(0.00034) (0.00033) (0.00034) (0.00034)
True Spillover Effects
Pre-production treatments in 0–2 km ring 0.00120∗∗∗ 0.00119∗∗∗
(0.00012) (0.00012)
Pre-production treatments in 2–5 km ring 0.00055∗∗∗ 0.00055∗∗∗
(0.00005) (0.00005)
Pre-production treatments in 5–10 km ring 0.00013∗∗∗ 0.00013∗∗∗
(0.00002) (0.00002)
Production treatments in 0–2 km ring 0.00070∗∗∗ 0.00070∗∗∗
(0.00010) (0.00010)
Production treatments in 2–5 km ring 0.00029∗∗∗ 0.00029∗∗∗
(0.00004) (0.00004)
Production treatments in 5–10 km ring 0.00002∗0.00002∗
(0.00001) (0.00001)
Control Variables
Weather, well densities in 20km
circular area, Two-way FEs Y Y Y Y
Adjusted R20.76 0.77 0.76 0.77
Sample Size 26,920,567 26,920,567 26,920,567 26,920,567
Note: The full results are reported in Appendix Table A8. Standard errors are clustered by well pads.
25

7 Welfare Analysis
Is the estimated increase in AOD due to shale gas development economically meaningful? To
answer this question, we use a concentration response function to estimate the increase in
mortality due to the increase in local PM pollution associated with the overall increase in AOD
due to the shale gas development.
7.1 Convert AOD to PM 2.5
Epidemiological concentration response functions describe the magnitude of a population level
health response from exposure to pollution. A first step in applying a concentration response
function is to translate the change in AOD to a change in PM 2.5 concentration. To do this, we
utilize the random coefficient model proposed by Lee et al. (2011). To predict the daily PM 2.5
concentrations for each P-area14, we use daily PM 2.5 concentration data from all 41 monitors
located in Pennsylvania to estimate the random coefficient model, and then use the estimated
coefficients to predict the daily change in PM 2.5 concentration for each P-area due to fracking
activities. Specifically, we estimate the following regression:
P Mjd = (α+ud)+(β+vd)qjd +mj+jt ,(5)
where jdenotes the monitor site, and ddenotes date. qjd is the average AOD value in a 3KM
radius area surrounding each PM 2.5 monitor. mjis the monitor specific random intercept,
udis the date specific random intercept, and vdis the daily random component in the slope of
AOD. We assume mj∼N(0, σ2
m), (ud, vd)∼N((0,0),Σ), and
Σ = 

σ2
uσuσv
σuσvσ2
v


,
and estimate equation 5using maximum likelihood. Figure A5 shows that the predicted PM
2.5 concentration fits the true PM 2.5 concentration in the vicinity of ground monitors with
R2= 0.78. Let ˆαand ˆ
βbe the estimated coefficients, let ˆudand ˆvdbe the daily value of random
components. Recall that the background AOD impact is ˆ
bit =qid −ˆsid, where qid is the observed
14We do not observe daily PM 2.5 concentration on a fine geographic scale. This is our original motivation
for using satellite-based AOD data rather than ground level measures of PM 2.5 concentration.
26

AOD, and ˆsid is the estimated AOD impact due to the shale gas development. The estimated
PM 2.5 concentration and ambient/background PM 2.5 concentration can be represented by
d
P M 2.5id = ˆα+ ˆud+ ( ˆ
β+ ˆvd)qid + ˆmi,(6)
g
P M 2.5id = ˆα+ ˆud+ ( ˆ
β+ ˆvd)ˆ
bid + ˆmi.(7)
where iis the P-area index and dis the date index. qid is the observed AOD, and ˆmiis the
estimated random intercept of each P-area. Then, the daily change in PM2.5 concentration due
to shale gas development ∆ d
P M 2.5id is estimated as
∆d
P M 2.5id =d
P M 2.5id −g
P M 2.5id = ( ˆ
β+ ˆvd)ˆsid .(8)
We find that on average, shale gas development increases the daily PM concentration by
0.017mg/m3for the whole sample, and 0.062mg/m3for the P-areas in the treatment group.
7.2 Mortality Impact
We estimate the change in mortality rates due to the estimated change in PM 2.5 concentration
at the census block group level by year. We obtain the mortality data from CDC WONDER at
the county level, which is the finest spatial resolution at which mortality data are available to the
public. We obtain population data from the American Community Survey at the census block
group level. Let the mortality rate for census block group kin year tbe λkt. We assume that
mortality is uniformly distributed across census block groups within a county, so λk0t=λk00tif
census block group k0and k00 are in the same county. Let the impact in PM 2.5 concentration of
shale gas development on census block group kbe ∆P M 2.5kt. We approximate ∆P M 2.5kt by
averaging the yearly average overall impact in PM 2.5 concentration of P-areas whose centroid
wells are located in census block group k. That is, ∆P M 2.5kt =1
Nkt Pi∈Ik,d∈t∆d
P M 2.5id, where
Nkt is the total number of centroid wells in census block kand date in year t(i∈Ikand d∈t).
(Krewski et al.,2009;Lepeule et al.,2012) estimate mortality concentration-response func-
tions for PM 2.5 concentration. They utilize a Cox proportional-hazard model with log-linear
functional form, which is also used by the EPA for Regulatory Impact Analysis (Fowlie et al.,
27

Table 7: Mortality Impact, 671 census block group containing P-areas
Year Cardio COPD All Death Population
2010 0.69
(3,104.83) 0.09
(555.43) 1.22
(9,522.19) 841,848
2011 1.16
(3,084.53) 0.15
(568.97) 2.10
(9,765.50) 845,607
2012 1.32
(3,022.89) 0.17
(543.39) 2.42
(9,670.84) 843,169
2013 1.11
(3,090.76) 0.15
(579.48) 2.04
(9,912.32) 845,133
2014 1.35
(3,053.58) 0.18
(555.91) 2.53
(9,866.78) 843,801
2015 1.96
(3,122.42) 0.28
(589.06) 3.70
(10,159.49) 838,444
2016 1.48
(3,111.75) 0.19
(553.01) 2.73
(10,074.72) 833,749
2017 1.78
(3,133.03) 0.23
(592.83) 3.36
(10,488.52) 828,150
Total 10.85
(24,723.80) 1.44
(4,538.08) 20.11
(79,420)
Numbers in parentheses describe the death count.
2019). We use the same method to estimate the impact in mortality rate:
∆Deathskt =P opkt λkt 1−exp(−ˆγ∆P M 2.5kt),(9)
where ˆγis the proportional hazard coefficient estimated by Lepeule et al. (2012), and P opkt
is the block group k’s population in year t. Appendix A.8 shows the detailed derivation of
Equation 9.
Following previous studies, we estimate mortality due to all causes, cardiovascular disease,
and chronic obstructive pulmonary disease (COPD). Let ˆγALL, ˆγCARD, and ˆγC OP D be the esti-
mated coefficient of the proportional hazard coefficients. Lepeule et al. (2012)’s study suggests
that for every 1 ug/m3increment in PM 2.5 concentration, ˆγALL = 0.0131, ˆγCARD = 0.0231,
and ˆγCO P D = 0.0157. Table 7reports the estimated impact on mortality of the shale gas indus-
try through PM 2.5 pollution for 671 census block groups in Pennsylvania where unconventional
wells are located. Table 8reports the estimated impact on mortality of all causes in top 4 coun-
ties in terms of the numbers of active shale gas wells. Our results suggest that from 2010 to 2017,
shale gas development caused an additional 20.11 deaths in a population of 840,000 through PM
28

Table 8: Mortality Impact of All Death, counties with most P-areas with active wells
Year Washington Susquehanna Bradford Greene
2010 0.28
(1,043.55) 0.05
(343.67) 0.11
(440.22) 0.10
(377.18)
2011 0.40
(1,085.06) 0.12
(354.03) 0.22
(436.16) 0.16
(351.16)
2012 0.43
(1,085.50) 0.14
(316.28) 0.20
(416.10) 0.18
(367.02)
2013 0.39
(1,101.69) 0.13
(347.16) 0.14
(453.45) 0.17
(354.29)
2014 0.50
(1,096.74) 0.18
(377.50) 0.15
(467.68) 0.22
(319.16)
2015 0.78
(1,133.99) 0.32
(383.46) 0.21
(484.69) 0.36
(385.46)
2016 0.65
(1,112.75) 0.19
(350.74) 0.14
(442.50) 0.28
(371.86)
2017 0.83
(1,161.92) 0.24
(373.37) 0.16
(469.44) 0.34
(366.65)
Total 4.26
(8,821.21) 1.34
(2.846.20) 1.33
(3,610.24) 1.82
(2,892.78)
P-areas (Active Wells) 1,785 1,553 1,465 1,336
All P-areas 2,776 2,519 3,338 2,100
Numbers in parentheses describe the death count.
29

2.5 emissions. Among all counties, Washington County is mostly affected, with an additional
4.26 deaths between 2010 and 2017. Using a value of statistical life of $7.4 million (2006 dollars)
(https://www.epa.gov/environmental-economics/mortality-risk-valuation), the total mortality
effect of additional 20.11 deaths can be translated to an economic loss of $148.814 million. This
is a lower bound of the total economic loss which is expected to be larger when considering all
types of pollution associated with shale gas development.
8 Conclusion and Discussion
The paper estimates use satellite based AOD data to detect the PM pollution from shale gas
wells at hyperlocal area. We find significant impact in PM concentration due to the shale gas
wells pre-production and production activities in the vicinity of well, with the marginal increase
in AOD is higher during pre-production (2.19% relative to the baseline AOD) than during
production (1.35% of baseline). Our results suggest that the PM pollution from shale gas wells
can travel through wind for up to 10 kilometers, but the pollution disperses and the impact
decreases in distance. Accounting for airborne spillovers, fracking increases AOD by 1.27%
for the whole sample, and by 5.67% for the subsample of P-areas with a treated well. These
overall increases in AOD imply that daily PM concentrations increased by 0.017mg/m3and
0.062mg/m3, respectively, in the average P-area. Besides, we estimate the impact in mortality
caused by the PM emissions from shale gas wells. We find that from 2010 to 2017, there are 671
census block groups across 40 counties that have shale gas wells located in, and there are about
840,000 populations living around shale gas wells. The estimated PM emissions from shale gas
wells causes additional 20.11 deaths among these communities.
Currie et al. (2017) and Hill (2018) find that the shale gas development has negative impact
in local health outcomes, but there is limited knowledge about through what channels shale
gas development affecting health. While there are several studies showing how the shale gas
development generates air and water pollution, there is no previous research directly link these
pollution to local health outcomes. Our study contribute to the literature by filling this gap.
This paper not only provides understandings of how the shale gas wells affect the air quality
through PM emissions, but also shows empirical evidence that shale gas wells cause extra
mortality in the local communities through generating PM pollution.
30

Since we only focus on the PM pollution fro shale gas industry, we are not able to estimate
the overall externalities. Future researches are needed for investigating the welfare impact of
shale gas industry through other channels, such as other air pollutants and water contamination.
31

A Appendix
A.1 Production Data and Assignment of Treatments
We utilize the information on spud date, plug date, and production period to determine the
on-site activities and sources of air pollution on a specific day. We assume a well is subject
to pre-production activities from the spud date till the start of production. Assuming the
production period is continuous, we define the production period from the first production date
to either the last production date or the plug date, depending on which date is later.15 We use
the well production reports to determine the start and end of production for each well. The
production reports are available annually before July 2010 , then biannual between July 2010
and December 2014, and eventually every month.
The production reports do not have the specific starting and completion dates. We use the
report date and duration to derive the first and last date of the production phase: for each
well, we find the earliest production report cycle with positive production, and then define the
first production date as the last date in that production report cycle minus production days
plus one. The last production date is defined as the first date in the last production report
cycle plus production days minus one. The estimated last production date is used to verify the
reliability of and to fix the measurement error in the plug date provided in “Oil Gas Locations
– Unconventional” data set.
For some wells with limited and incomplete information, we make additional assumptions
to define the pre-production phase. For instance, some wells were spudded but not producing
so only the start date of the pre-production phase is available, but the end date is unknown.
Others were spudded and producing but spud date is missing so only the pre-production end
date is available, but the pre-production start date remains unknown. In these cases we use the
average pre-production phase length of wells within a 50 km radius to define the pre-production
phase. There are 1,698 wells with estimated pre-production phase length.
We do not consider the post-production period in this paper, so all observations after pro-
duction period end date are removed. In the whole sample, we have 10,479 P-areas under
15There are 50 wells with plug date before the last date of production. 47 of them have plug date more
than 500 days before production end date. All 50 wells have positive production after their listed plug date.
Therefore, we assume these plug dates are incorrectly measured.
32

treatments. There are 1,846 P-areas with only pre-production treatment, 15 P-areas with only
production treatment, and 8,618 P-areas with both treatments. The remaining 8,391 P-areas do
not experience any treatment. Our definition of treatments is very consistent with well status
as reported in “Oil Gas Locations – Unconventional”. Among 10,479 wells with treatments,
only 5 of them are labelled as “Proposed But Never Materialized” or “Operator Reported Not
Drilled”. Based on their production report, they are mistakenly labelled.
A.2 Additional Summary Description
Table A1: Bivariate Correlations Between Variables
AOD Pre-production treatment Production treatment Pre-pro duction treatment Pre-production treatment Pre-pro duction treatment Production treatment Pro duction treatment
in 0-2 km ring in 2-5 km ring in 5-10 km ring in 0-2 km ring in 2-5 km ring
AOD 1 - - - - - - -
Pre-production treatment -0.0188 1 - - - - - -
Production treatment -0.0454 -0.0652 1 - - - - -
Pre-production treatment -0.0254 0.3075 0.0427 1 - - - -
in 0-2 km ring
Pre-production treatment -0.0485 0.1489 0.1315 0.2564 1 - - -
in 2-5 km ring
Pre-production treatment -0.0752 0.1340 0.1655 0.2262 0.4439 1 - -
in 5-10 km ring
Production treatment -0.0513 0.0248 0.4718 0.0949 0.1987 0.2546 1 -
in 0-2 km ring
Production treatmen -0.0710 0.0701 0.4191 0.1369 0.2608 0.3606 0.5865 1
in 2-5 km ring
Production treatmen -0.0900 0.0777 0.3741 0.1427 0.2801 0.4274 0.4922 0.7381
in 5-10 km ring
Density Pre-production -0.0928 0.2409 0.2037 0.3730 0.6088 0.7789 0.3059 0.4044
(Number of wells, 20 km)
Density Production -0.0928 0.0873 0.4789 0.1630 0.3008 0.4229 0.6113 0.7995
(Number of wells, 20 km)
Density Inactive -0.0592 0.0479 0.3648 0.0684 0.1182 0.1652 0.4376 0.5254
(Number of wells, 20 km)
Precipitation (mm) 0.0276 -0.0009 0.0008 -0.0014 -0.0057 -0.0128 -0.0007 -0.0055
Dew Point (Celsius) 0.3704 -0.0046 0.0112 -0.0065 -0.0137 -0.0247 0.0081 0.0012
Temperature (Celsius) 0.3852 -0.0067 0.0039 -0.0082 -0.0147 -0.0239 0.0012 -0.0044
Wind Speed (m/s) -0.1207 0.0015 0.0001 0.0049 0.0044 0.0046 0.0070 0.0042
prod 10 Density Pre-production Density Production Density Inactive Precipitation (mm) Dew Point (Celsius) Temperature (Celsius) Wind Speed (m/s)
in 5-10 km ring (Number of wells, 20 km) (Number of wells, 20 km) (Number of wells, 20 km)
AOD - - - - - - - -
Pre-production treatment - - - - - - - -
Production treatment - - - - - - - -
Pre-production treatment - - - - - - - -
in 0-2 km ring Pre-production treatment - - - - - - - -
in 2-5 km ring Pre-production treatment - - - - - - - -
in 5-10 km ring Production treatment - - - - - - - -
in 0-2 km ring Production treatment - - - - - - - -
in 2-5 km ring Production treatment 1 - - - - - - -
in 5-10 km ring Density Pre-production 0.4475 1 - - - - - - -
(Number of wells, 20 km)
Density Production 0.8542 0.5218 1 - - - - -
(Number of wells, 20 km)
Density Inactive 0.5024 0.2064 0.6089 1 - - - -
(Number of wells, 20 km)
Precipitation (mm) 0.0110 -0.0141 -0.0114 -0.0041 1 - - -
Dew Point (Celsius) -0.0059 -0.0303 -0.0043 0.0095 0.1916 1 - -
Temperature (Celsius) -0.0094 -0.0308 -0.0095 -0.0096 0.1091 0.9490 1 -
Wind Speed (m/s) 0.0025 0.0050 0.0030 0.0143 0.1030 -0.2275 -0.2536 1
33

Figure A1: Time Trend – AOD and weather variables
A.3 Common Trend Test
Figure A2: Common Trend Test
Table A2: Common Trend Significance Dates
Significance Level 10% 5% 1% 0.1% Overall
Number of Significant Dates 1,337 1,028 598 315 4,609
Significant Dates Ratio 29.01% 22.30% 12.97% 6.83%
34

A.4 Additional Discussion of Main Results
Figure 2,3and 4show the geographical location of all unconventional wells in PA. We find
clusters of unconventional wells in northeast and southwest PA, with a spatial segregation
between the treatment and control groups as shown in figure 2and 3. This can be explained by
figure 4, as the distribution of treatment group with both pre-production and production are
quite similar to wells with permits issued in 2008-2011, and the distribution of control group
is quite similar to wells with permits issued in 2012-2015. These wells belong to the control
group because of the late development of the local shale gas industry, and it is possible they
will operate in the future.
The coefficients of ”Density Pre-production” and ”Density Production” are significantly
negative, but the coefficient of ”Density Inactive” is significantly positive. That is because the
density variables have ambiguous effect in the baseline model: First, larger densities means less
infrastructure construction and lower AOD. Second, larger densities means more pre-production,
production and inactive wells nearby, which increases AOD. Inactive wells may also contribute
to AOD because of the cleanup process after the well no longer operated.
Table A3 reports the estimation results of the baseline model and spatial difference-in-
differences model. Column (1) and (2) report the results from the baseline model, and column
(3) and (4) report the results from spatial difference-in-differences model. Column (1) gives
very counter-intuitive results, as pre-production phase does not affect air quality, and produc-
tion period makes air quality even better. Without controlling density variables, the estimation
is heavily biased, because the segregation of treatment group and control group causes the treat-
ments to be non-random in location and time, and makes tow groups incomparable. Column (3)
does not include density variables, but by doing spatial difference-in-differences, the coefficients
of ”Pre-production” and ”Production” become significantly positive. However, some coefficients
of spatial spillover variables are still counter-intuitive. That is because spatial spillover vari-
ables are correlated with number of wells nearby, and partially account for the densities. It also
means that density variables are confounders of both treatment dummy variables and spatial
spillover variables, and excluding density variables in the regression may downward bias the
estimation results. Column (2) and column (4) include density variables in the regression, and
the treatment effects of pre-production phase and production period are significantly positive.
35

Table A3: Main Analysis
AOD DID Spatial DID
(1) (2) (3) (4)
Own Effects
Pre-production treatment 0.00028 0.00624∗∗∗ 0.00191∗∗∗ 0.00429∗∗∗
(0.00033) (0.00040) (0.00033) (0.00039)
Production treatment −0.00078∗∗∗ 0.00463∗∗∗ 0.00102∗∗∗ 0.00258∗∗∗
(0.00030) (0.00049) (0.00032) (0.00036)
Spillover Effects
Pre-production treatment from 0–2 km ring 0.00029∗∗∗ 0.00101∗∗∗
(0.00008) (0.00010)
Pre-production treatment from 2-5 km ring −0.00015∗∗∗ 0.00050∗∗∗
(0.00004) (0.00004)
Pre-production treatment from 5–10 km ring −0.00066∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002)
Production treatments from 0–2 km ring 0.00037∗∗∗ 0.00066∗∗∗
(0.00006) (0.00008)
Production treatments from 2–5 km ring 0.00009∗∗∗ 0.00026∗∗∗
(0.00002) (0.00003)
Production treatments from 5–10 km ring −0.00035∗∗∗ 0.00002∗∗
(0.00001) (0.00001)
Covariates
Number of Pre-production treatment wells in 0-20km −0.00013∗∗∗ −0.00016∗∗∗
(0.00000) (0.00000)
Number of production wells in 0-20km −0.00008∗∗∗ −0.00009∗∗∗
(0.00000) (0.00000)
Number of inactive wells in 0-20km 0.00023∗∗∗ 0.00022∗∗∗
(0.00001) (0.00001)
Daily precipitation 0.00018∗∗∗ 0.00020∗∗∗ 0.00018∗∗∗ 0.00020∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
Daily mean dew point 0.01319∗∗∗ 0.01247∗∗∗ 0.01300∗∗∗ 0.01245∗∗∗
(0.00012) (0.00011) (0.00012) (0.00011)
Daily mean temperature −0.00339∗∗∗ −0.00286∗∗∗ −0.00315∗∗∗ −0.00287∗∗∗
(0.00009) (0.00009) (0.00009) (0.00009)
Wind Speed −0.00155∗∗∗ −0.00157∗∗∗ −0.00154∗∗∗ −0.00158∗∗∗
(0.00008) (0.00008) (0.00008) (0.00008)
Well FE Y Y Y Y
Date FE Y Y Y Y
Cluster PAD PAD PAD PAD
Observations 31,531,987 31,531,987 31,531,987 31,531,987
R20.76073 0.76259 0.76122 0.76268
Adjusted R20.76054 0.76240 0.76103 0.76248
Residual Std. Error 0.12761 (df = 31506614) 0.12712 (df = 31506611) 0.12748 (df = 31506608) 0.12709 (df = 31506605)
Wind speed is the daily wind speed at the location of centroid well of each P-area;
Standard errors are clustered by well pad;
Standard errors: ∗∗∗p<.01,∗∗ p<.05,∗p<.1;
36

A.5 Full Results of Robustness Check
Table A4: Sensitivity Analysis 1 – Using Treatment Group Only
AOD Baseline Model Spatial Diff-in-diff Model
(1) (2) (3) (4)
Pre-production treatmen 0.00058∗0.00651∗∗∗ 0.00200∗∗∗ 0.00475∗∗∗
(0.00033) (0.00040) (0.00035) (0.00040)
Production −0.00069∗0.00488∗∗∗ 0.00070∗0.00281∗∗∗
(0.00038) (0.00060) (0.00042) (0.00059)
Pre-production 2 0.00019∗∗ 0.00086∗∗∗
(0.00009) (0.00010)
Pre-production 5−0.00015∗∗∗ 0.00048∗∗∗
(0.00004) (0.00004)
Pre-production 10 −0.00064∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002)
Production 2 0.00032∗∗∗ 0.00060∗∗∗
(0.00007) (0.00009)
Production 5 0.00008∗∗∗ 0.00025∗∗∗
(0.00003) (0.00003)
Production 10 −0.00034∗∗∗ 0.00003∗∗
(0.00001) (0.00001)
Density Pre-production −0.00013∗∗∗ −0.00016∗∗∗
(0.00000) (0.00000)
Density Production −0.00009∗∗∗ −0.00010∗∗∗
(0.00000) (0.00000)
Density Inactive 0.00021∗∗∗ 0.00021∗∗∗
(0.00001) (0.00001)
Daily Precipitation 0.00016∗∗∗ 0.00019∗∗∗ 0.00016∗∗∗ 0.00019∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
Daily mean dew point 0.01374∗∗∗ 0.01296∗∗∗ 0.01355∗∗∗ 0.01294∗∗∗
(0.00014) (0.00014) (0.00014) (0.00014)
Daily mean temperature −0.00342∗∗∗ −0.00284∗∗∗ −0.00318∗∗∗ −0.00285∗∗∗
(0.00011) (0.00011) (0.00011) (0.00011)
Wind Speed −0.00181∗∗∗ −0.00185∗∗∗ −0.00180∗∗∗ −0.00186∗∗∗
(0.00008) (0.00008) (0.00009) (0.00008)
Well FE Y Y Y Y
Date FE Y Y Y Y
Cluster PAD PAD PAD PAD
Observations 17,506,147 17,506,147 17,506,147 17,506,147
R20.76242 0.76444 0.76294 0.76454
Adjusted R20.76221 0.76423 0.76272 0.76432
Residual Std. Error 0.12672 (df = 17490026) 0.12618 (df = 17490023) 0.12658 (df = 17490020) 0.12615 (df = 17490017)
Standard errors: ∗p<0.1; ∗∗ p<0.05; ∗∗∗p<0.01
37

Table A5: Sensitivity Analysis 2 – Standard Diff-in-Diff Using Matched Samples
AOD Matched sample 1 Matched sample 2 Matched sample 1 Matched sample 2
(1) (2) (3) (4)
Pre-production 0.00050 0.00047 0.00609∗∗∗ 0.00624∗∗∗
(0.00032) (0.00032) (0.00039) (0.00039)
Production −0.00051 −0.00057∗0.00449∗∗∗ 0.00475∗∗∗
(0.00031) (0.00032) (0.00051) (0.00051)
Density Pre-production −0.00013∗∗∗ −0.00013∗∗∗
(0.00000) (0.00000)
Density Production −0.00009∗∗∗ −0.00009∗∗∗
(0.00000) (0.00000)
Density Inactive 0.00021∗∗∗ 0.00021∗∗∗
(0.00001) (0.00001)
Daily precipitation 0.00018∗∗∗ 0.00017∗∗∗ 0.00020∗∗∗ 0.00019∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
Daily mean dew point 0.01347∗∗∗ 0.01347∗∗∗ 0.01271∗∗∗ 0.01270∗∗∗
(0.00013) (0.00014) (0.00012) (0.00013)
Daily temperature −0.00341∗∗∗ −0.00344∗∗∗ −0.00284∗∗∗ −0.00286∗∗∗
(0.00010) (0.00010) (0.00010) (0.00010)
Wind Speed −0.00173∗∗∗ −0.00174∗∗∗ −0.00176∗∗∗ −0.00177∗∗∗
(0.00008) (0.00008) (0.00008) (0.00008)
Well FE Y Y Y Y
Date FE Y Y Y Y
Cluster PAD PAD PAD PAD
Observations 21,949,389 21,583,079 21,949,389 21,583,079
R20.76163 0.76201 0.76361 0.76400
Adjusted R20.76142 0.76180 0.76340 0.76379
Residual Std. Error 0.12710 (df = 21930365) 0.12700 (df = 21564321) 0.12657 (df = 21930362) 0.12647 (df = 21564318)
The matched subsample 1 is matched by distance, permit year and permit month.
The matched subsample 2 is matched by distance, permit year and county.
Standard errors: ∗p<0.1; ∗∗ p<0.05; ∗∗∗p<0.01
38

Table A6: Sensitivity Analysis 3 – Spatial Diff-in-Diff Using Matched Samples
AOD Matched sample 1 Matched sample 2 Matched sample 1 Matched sample 2
(1) (2) (3) (4)
Pre-production 0.00191∗∗∗ 0.00192∗∗∗ 0.00436∗∗∗ 0.00452∗∗∗
(0.00033) (0.00033) (0.00038) (0.00038)
Production 0.00088∗∗∗ 0.00092∗∗∗ 0.00262∗∗∗ 0.00291∗∗∗
(0.00034) (0.00034) (0.00048) (0.00049)
Pre-production 2 0.00022∗∗ 0.00021∗∗ 0.00093∗∗∗ 0.00093∗∗∗
(0.00009) (0.00009) (0.00010) (0.00010)
Pre-production 5 −0.00015∗∗∗ −0.00015∗∗∗ 0.00043∗∗∗ 0.00048∗∗∗
(0.00004) (0.00004) (0.00004) (0.00004)
Pre-production 10 −0.00065∗∗∗ −0.00065∗∗∗ 0.00008∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002) (0.00002) (0.00002)
Production 2 0.00033∗∗∗ 0.00031∗∗∗ 0.00061∗∗∗ 0.00060∗∗∗
(0.00007) (0.00006) (0.00008) (0.00009)
Production 5 0.00008∗∗∗ 0.00009∗∗∗ 0.00026∗∗∗ 0.00027∗∗∗
(0.00002) (0.00002) (0.00003) (0.00003)
Production 10 −0.00035∗∗∗ −0.00035∗∗∗ 0.00002∗∗ 0.00002∗∗
(0.00001) (0.00001) (0.00001) (0.00001)
Density Pre-production −0.00016∗∗∗ −0.00016∗∗∗
(0.00000) (0.00000)
Density Production −0.00009∗∗∗ −0.00009∗∗∗
(0.00000) (0.00000)
Density Inactive 0.00021∗∗∗ 0.00020∗∗∗
(0.00001) (0.00001)
Daily precipitation 0.00018∗∗∗ 0.00017∗∗∗ 0.00020∗∗∗ 0.00019∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
Daily mean dew point 0.01328∗∗∗ 0.01328∗∗∗ 0.01269∗∗∗ 0.01268∗∗∗
(0.00013) (0.00013) (0.00013) (0.00013)
Daily mean temperature −0.00316∗∗∗ −0.00320∗∗∗ −0.00285∗∗∗ −0.00288∗∗∗
(0.00010) (0.00010) (0.00010) (0.00010)
Wind Speed −0.00171∗∗∗ −0.00173∗∗∗ −0.00177∗∗∗ −0.00178∗∗∗
(0.00008) (0.00008) (0.00008) (0.00008)
Well FE Y Y Y Y
Date FE Y Y Y Y
Cluster PAD PAD PAD PAD
Observations 21,949,389 21,583,079 21,949,389 21,583,079
R20.76214 0.76252 0.76370 0.76409
Adjusted R20.76193 0.76231 0.76350 0.76389
Residual Std. Error 0.12697 (df = 21930359) 0.12686 (df = 21564315) 0.12655 (df = 21930356) 0.12644 (df = 21564312)
The matched subsample 1 is matched by distance, permit year and permit month.
The matched subsample 2 is matched by distance, permit year and county.
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
39

Table A7: Robustness check: spatial DID using longer pre-production windows
Window 1 Window 2 Window 1 Window 2
window 0.00626∗∗∗ 0.00620∗∗∗ 0.00434∗∗∗ 0.00433∗∗∗
(0.00040) (0.00039) (0.00039) (0.00038)
prod 0.00466∗∗∗ 0.00470∗∗∗ 0.00260∗∗∗ 0.00264∗∗∗
(0.00049) (0.00049) (0.00046) (0.00046)
window 2 0.00101∗∗∗ 0.00101∗∗∗
(0.00010) (0.00010)
window 5 0.00050∗∗∗ 0.00050∗∗∗
(0.00004) (0.00004)
window 10 0.00008∗∗∗ 0.00008∗∗∗
(0.00002) (0.00002)
prod 2 0.00066∗∗∗ 0.00066∗∗∗
(0.00008) (0.00008)
prod 5 0.00026∗∗∗ 0.00026∗∗∗
(0.00003) (0.00003)
prod 10 0.00002∗∗ 0.00002∗∗
(0.00001) (0.00001)
Treatment window Count −0.00013∗∗∗ −0.00013∗∗∗ −0.00016∗∗∗ −0.00016∗∗∗
(0.00000) (0.00000) (0.00000) (0.00000)
Treatment prod count −0.00008∗∗∗ −0.00008∗∗∗ −0.00009∗∗∗ −0.00009∗∗∗
(0.00000) (0.00000) (0.00000) (0.00000)
Plug Count 0.00023∗∗∗ 0.00023∗∗∗ 0.00022∗∗∗ 0.00022∗∗∗
(0.00001) (0.00001) (0.00001) (0.00001)
PPT 0.00020∗∗∗ 0.00020∗∗∗ 0.00020∗∗∗ 0.00020∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
TDMEAN 0.01247∗∗∗ 0.01247∗∗∗ 0.01245∗∗∗ 0.01245∗∗∗
(0.00011) (0.00011) (0.00011) (0.00011)
TMEAN −0.00286∗∗∗ −0.00286∗∗∗ −0.00287∗∗∗ −0.00287∗∗∗
(0.00009) (0.00009) (0.00009) (0.00009)
Wind Speed −0.00157∗∗∗ −0.00157∗∗∗ −0.00158∗∗∗ −0.00158∗∗∗
(0.00008) (0.00008) (0.00008) (0.00008)
Well and Date FEs Y Y Y Y
Observations 31,531,987 31,531,987 31,531,987 31,531,987
R20.76259 0.76259 0.76268 0.76268
Adjusted R20.76240 0.76240 0.76248 0.76248
Residual Std. Error 0.12712 0.12712 0.12709 0.12709
Residual Std. Error (df = 31506611) (df = 31506611) (df = 31506605) (df = 31506605)
Note: Window 1 start from 7 days before spud date, Window 2 starts from 30 days before spud date.
40

Table A8: Robustness check: falsification tests
AOD Baseline Model Spatial Diff-in-diff Model
(1) (2) (3) (4)
Placebo Treatment 0.00019 −-0.00015 0.00036 0.00008
(0.00034) (0.00034) (0.00033) (0.00034)
Pre-production 2 0.00120∗∗∗ 0.00119∗∗∗
(0.00012) (0.00012)
Pre-production 5 0.00055∗∗∗ 0.00055∗∗∗
(0.00005) (0.00005)
Pre-production 10 0.00013∗∗∗ 0.00013∗∗∗
(0.00002) (0.00002)
Production 2 0.00070∗∗∗ 0.00070∗∗∗
(0.00010) (0.00010)
Production 5 0.00029∗∗∗ 0.00029∗∗∗
(0.00004) (0.00004)
Production 10 0.00002∗0.00002∗
(0.00001) (0.00001)
Density Pre-production −0.00015∗∗∗ −0.00015∗∗∗ −0.00017∗∗∗ −0.00017∗∗∗
(0.00000) (0.00000) (0.00000) (0.00000)
Density Production −0.00008∗∗∗ −0.00008∗∗∗ −0.00008∗∗∗ −0.00009∗∗∗
(0.00000) (0.00001) (0.00000) (0.00000)
Density Inactive 0.00023∗∗∗ 0.00023∗∗∗ 0.00022∗∗∗ 0.00022∗∗∗
(0.00001) (0.00001) (0.00001) (0.00001)
Daily precipitation 0.00029∗∗∗ 0.00029∗∗∗ 0.00029∗∗∗ 0.00029∗∗∗
(0.00003) (0.00003) (0.00003) (0.00003)
Daily mean dew point 0.01272∗∗∗ 0.01272∗∗∗ 0.01270∗∗∗ 0.01270∗∗∗
(0.00012) (0.00012) (0.00012) (0.00012)
Daily mean temperature −0.00321∗∗∗ −0.00321∗∗∗ −0.00321∗∗∗ −0.00321∗∗∗
(0.00010) (0.00010) (0.00010) (0.00010)
Wind Speed −0.00155∗∗∗ −0.00155∗∗∗ −0.00157∗∗∗ −0.00157∗∗∗
(0.00008) (0.00008) (0.00008) (0.00008)
Well FE Y Y Y Y
Date FE Y Y Y Y
Cluster PAD PAD PAD PAD
Observations 26,920,567 26,920,567 26,920,567 26,920,567
R20.76519 0.76519 0.76526 0.76526
Adjusted R20.76497 0.76497 0.76504 0.76504
Residual Std. Error 0.12952 0.12952 0.12950 0.12950
Residual Std. Error (df = 26895198) (df = 26895198) (df = 26895192) (df = 26895192)
Note: Standard errors are clustered by wells pad.
41

A.6 Additional maps of wells
Figure A3: Matched sample 1 – wells matched by distance, permit year, and permit month
Figure A4: Matched sample 2 – wells matched by distance, permit year, and county
42

A.7 Convert AOD to PM 2.5
Figure A5: PM 2.5 Prediction, PM 2.5 Monitors in Pennsylvania
A.8 Cox proportional-hazard model
Estimating the Cox proportional-hazard model yields the relative mortality rate (RR), which
is the ratio of mortality rates under different pollution concentration.
RR =λ(X, P M 0
2.5)
λ(X, P M 00
2.5)=exp(ˆγ(P M 0
2.5−P M 00
2.5)).(10)
Xis a vector of covariates related with mortality, λ(X, P M 0
2.5) and λ(X, P M 00
2.5) are mortality
rates under two different PM 2.5 concentrations P M 0
2.5and P M 00
2.5. ˆγis the estimated coefficient
of RR, which indicates the mortality impact by changing PM 2.5 concentration from P M 0
2.5to
P M 00
2.5.
43

References
Allen, D. T., Torres, V. M., Thomas, J., Sullivan, D. W., Harrison, M., Hendler, A., Herndon, S. C.,
Kolb, C. E., Fraser, M. P., Hill, A. D., et al. (2014). Measurements of methane emissions at natural gas
production sites in the united states. Proceedings of the National Academy of Sciences, 110(44):17768–
17773.
Athey, S. and Imbens, G. W. (2018). Design-based analysis in difference-in-differences settings with
staggered adoption. NBER, working paper.
Atkinson, R. W., Kang, S., Anderson, H. R., Mills, I. C., and Walton, H. A. (2014). Epidemiological
time series studies of pm2.5 and daily mortality and hospital admissions: a systematic review and
meta-analysis. Respiratory Epidemiology, 69:650–662.
Currie, J., Greenstone, M., and Meckel, K. (2017). Hydraulic fracturing and infant health: New evidence
from pennsylvania. Science Advances, 3(12).
Delgado, M. S. and Florax, R. J. (2015). Difference-in-differences techniques for spatial data: Local
autocorrelation and spatial interaction. Economics Letters, 137:123–126.
Donkelaar, A. V., Martin, R. V., Brauer, M., Hsu, N. C., Kahn, R. A., Levy, R. C., Lyapustin, A.,
Sayer, A. M., and Winker, D. M. (2016). Global estimates of fine particulate matter using a combined
geophysical-statistical method with information from satellites, models, and monitors. Environmental
Science and Technology, 50:3762–3772.
EIA Website (2019). Shale gas production. https://www.eia.gov/dnav/ng/ng prod shalegas s1 a.htm,
Accessed in: March 2020.
Field, R. A., Soltis, J., and Murphy, S. (2014). Air quality concerns of unconventional oil and natural
gas production. Environmental Science Processes and Impact, 16:954–969.
Foster, A., Gutierrez, E., and Kumar, N. (2009). Voluntary compliance, pollution levels, and infant
mortality in mexico. American Economic Review, 99(2):191–197.
Fowlie, M., Rubin, E. A., and Walker, R. (2019). Bringing satellite-based air quality estimates down to
earth. AEA Papers and Proceedings, 109(1):283–288.
Goodman-Bacon, A. (2018). Difference-in-differences with variation in treatment timing. NBER, working
paper.
Graham, J., Irving, J., Tang, X., Sellers, S., Crisp, J., Horwitz, D., Muehlenbachs, L., Krupnick, A., and
Carey, D. (2015). Increased traffic accident rates associated with shale gas drilling in pennsylvania.
Accident Analysis and Prevention, 74:203–209.
Heo, J. B., Hopke, P. K., and Yi, S. (2009). Source apportionment of pm2.5 in seoul, korea. Atmospheric
Chemistry and Physics, 9:4957–4971.
Hill, E. and Ma, L. (2017). Shale gas development and drinking water quality. American Economic
Review, 107(5):522–525.
Hill, E. L. (2018). Shale gas development and infant health: Evidence from pennsylvania. Journal of
Health Economics, 61:134–150.
Huang, R.-J., Zhang, Y., Bozzetti, C., Ho, K.-F., Cao, J.-J., Han, Y., Daellenbach, K. R., Slowik, J. G.,
Platt, S. M., Canonaco, F., Zotter, P., Wolf, R., Pieber, S. M., Bruns, E. A., Crippa, M., Ciarelli, G.,
Piazzalunga, A., Schwikowski, M., Abbaszade, G., Schnelle-Kreis, J., Zimmermann, R., An, Z., Szidat,
S., Baltensperger, U., Haddad, I. E., and 1, A. S. H. P. (2014). High secondary aerosol contribution
to particulate pollution during haze events in china. Nature, 514:218–222.
44

Jackson, R. B., Vengosh, A., Darrah, T. H., Warner, N., Down, A., Poreda, R. J., Osborn, S. G., Zhao,
K., and Karr, J. D. (2013). Increased stray gas abundance in a subset of drinking water wells near
marcellus shale gas extraction. Proceedings of the National Academy of Sciences, 110(28):11250–11255.
Krewski, D., Jerrett, M., Burnett, R. T., Ma, R., Hughes, E., Shi, Y., Turner, M. C., Pope III, C. A.,
Thurston, G., Calle, E. E., and Thun et al, M. J. (2009). Extended follow-up and spatial analysis
of the american cancer society study linking particulate air pollution and mortality. Health Effects
Institute Boston, MA.
Kumar, N., Chu, A., and Foster, A. (2007). An empirical relationship between pm2.5 and aerosol optical
depth in delhi metropolitan. Atmospheric Environment, 41:4492–4503.
Larsen, B., Gilardoni, S., Stenstr¨om, K., Niedzialek, J., Jimenez, J., and Belis, C. (2012). Sources for
pm air pollution in the po plain, italy: Ii. probabilistic uncertainty characterization and sensitivity
analysis of secondary and primary sources. Atmospheric Environment, 50:203–213.
Lee, H. J., Liu, Y., Coull, B. A., Schwartz, J., and Koutrakis, P. (2011). A novel calibration approach of
modis aod data to predict pm2.5concentrations. Atmospheric Chemistry and Physics, 11(15):7991–
8002.
Lepeule, J., Laden, F., Dockery, D., and Schwartz, J. (2012). Chronic exposure to fine particles and
mortality: an extended follow-up of the harvard six cities study from 1974 to 2009. Environmental
health perspectives, 120(7):965.
Lewandowski, M., Jaoui, M., Offenberg, J. H., Kleindienst, T. E., Edney, E. O., Sheesley, R. J., and
Schauer, J. J. (2008). Primary and secondary contributions to ambient pm in the midwestern united
states. Environmental Science Technology, 42(9):3303–3309.
Litovitz, A., Curtright, A., Abramzon, S., Burger, N., and Samaras, C. (2013). Estimation of regional air-
quality damages from marcellus shale natural gas extraction in pennsylvania. Environmental Research
Letters, 8(1):014017.
Liu, Y., Park, R. J., Jacob, D. J., Li, Q., Kilaru, V., and Sarnat, J. A. (2004). Mapping annual
mean ground-level pm 2.5 concentrations using multiangle imaging spectroradiometer aerosol optical
thickness over the contiguous united states. Journal of Geophysical Research, 109(D12):D22206.
Muehlenbachs, L., Spiller, E., and Timmins, C. (2015). The housing market impacts of shale gas devel-
opment. American Economic Review, 105(12):3633–3659.
Newell, R. G. and Raimi, D. (2014). Implications of shale gas development for climate change. Environ-
mental Science & Technology, 48(15):8360–8368.
Olmstead, S. M., Muehlenbachs, L. A., Shih, J.-S., Chu, Z., and Krupnick, A. J. (2013). Shale gas
development impacts on surface water quality in pennsylvania. Proceedings of the National Academy
of Sciences, 110(13):4962.
Osborn, S., Vengosh, A., Warner, N., and Jackson, R. (2011). Methane contamination of drinking water
accompanying gas well drilling and hydraulic fracturing. Proceedings of the National Academy of
Sciences, 108(20):8172–8176.
PA DEP Web (2012). All permits issued & wells drilled. http://files.dep.state.pa.us/OilGas/BOGM/BOG
MPortalFiles/OilGasReports/2012/2009Wellspermitte-drilled.pdf, Accessed in: March 2020.
PSU Web (2009). Depth of marcellus shale base. http://www.marcellus.psu.edu/resources/images/marce
llus-depth.gif, Accessed in: Mar. 2020.
Remer, L. A., Mattoo, S., Levy, R. C., and Munchak, L. A. (2005). Modis 3km aerosol product: algorithm
and global perspective. Journal of the Atmospheric Sciences, 62(4):947–973.
45

Roy, A. A., Adams, P. J., and Robinson, A. L. (2014). Air pollutant emissions from the development,
production, and processing of marcellus shale natural gas. Journal of the Air and Waste Management
Association, 64(1):19–37.
Sarigiannis, D., Handakas, E., Kermenidou, M., Zarkadas, I., Gotti, A., Charisiadis, P., Makris, K.,
Manousakas, M., Eleftheriadis, K., and Karakitsios, S. (2017). Monitoring of air pollution levels
related to charilaos trikoupis bridge. Science of The Total Environment, 609:1451 – 1463.
Sieminski, A. (2014). Implications of the u.s. shale revolution. EIA US-Canada Energy Summit, Chicago.
Srebotnjak, T. and Rotkin-Ellman, M. (2014). Fracking fumes: Air pollution from hydraulic fracturing
threatens public health and communities. NRDC Issue BRIEF, 14:10–a.
Streets, D. G., Canty, T., R., G., Carmichael, de Foy, B., Dickerson, R. R., Duncan, B. N., Edwards,
D. P., Haynes, J. A., Henze, D. K., Houyoux, M. R., Jacob, D. J., Krotkov, N. A., Lamsal, L. N.,
Liu, Y., Lu, Z., Martin, R. V., Pfister, G. G., Pinderm, R. W., Salawitch, R. J., and Wecht, K. J.
(2013). Emissions estimation from satellite retrievals: A review of current capability. Atmospheric
Environment, 77:1011–1042.
Sullivan, D. M. and Krupnick, A. (2019). Using satellite data to fill the gaps in the us air pollution
monitoring network. RFF, working paper.
Zou, E. (2019). Unwatched pollution: The effect of intermittent monitoring on air quality. American
Economic Review, forthcoming.
46