Rural Light Pollution from Shale Gas Development and Associated Sleep and Subjective Well-Being

Abstract

The shale oil and gas boom has had large economic, environmental, and social impacts on rural communities in the United States. This study provides novel estimates of the impacts of shale oil and gas development on light pollution in rural areas of the United States. Using nationwide, time-calibrated DMSP-OLS database from 2000 to 2012, we find robust evidence that the shale oil and gas boom significantly increased light pollution in rural areas. We then assess associations between horizontal drilling and subjective self-rated health using nationwide data from the Behavioral Risk Factor Surveillance System (BRFSS) from 2000 to 2012. Our findings suggest that insufficient sleep and poor health (physical or mental) are associated with increased drilling in rural areas. These results provide support for drilling-related light pollution as an additional environmental pathway of concern for public health beyond the mechanisms of air or water pollution.

1. Introduction

Light pollution has significant implications for the environment and public health, and its effects have become more pronounced over time due to the increasing extent and radiance of artificially-lit areas (Kyba et al. 2017). Substantial economic values have been attached to affected outcomes, such as biodiversity, recreation, and public health. A recent paper values biodiversity in Oregon to be $2.5 million per year using a travel cost model (Kolstoe and Cameron, 2017). Stated preference methods have also found significant willingness to pay to preserve bird species: Loomis (2005) finds the average consumer surplus for birdwatching per person per day to be $30. With respect to human health, artificial lights at night (ALAN) are associated with sleep deprivation and mental health (Patel 2019; Xiao et al. 2020); sleep deprivation, in turn, has been shown to reduce cognition and labor market productivity, as well as elevate mortality risks associated with dementia, heart attacks, and vehicle accidents (Hafner et al. 2017; Paksarian et al. 2020; Ma et al. 2020; Jin and Ziebarth 2020, Prats-Uribe et al. 2018). A study in Australia quantified the financial and non-financial costs of inadequate sleep in 2016–2017 to be US$45 billion (Hillman et al. 2018) and another study estimates that $680 billion is lost due to sleep deprivation across 5 OECD countries (Hafner et al. 2017).

Despite efforts made by stakeholder groups (e.g., International Dark Sky Association) to both acknowledge and mitigate light pollution and its negative impacts, the primary causes of light pollution – spatial growth of cities, residential development and local economic activity (e.g., Chalkias et al. 2006; Radeloff et al. 2010; Olsen et al. 2014) – continue unabated. In the last two decades, the United States has also seen extraordinary growth in unconventional natural gas and oil development, which has been hypothesized to contribute to light pollution, especially in rural areas where drilling operations predominantly occur. That the shale boom has led to large increases in rural light pollution is visually clear (e.g., Figure 1; author generated). Yet, we are not aware of any studies that have quantitatively assessed the degree to which shale development in the U.S. impacts local light pollution.

Figure 1:

Bakken shale drilling and nighttime lights in North Dakota: 2000 vs. 2012

Notes: In this figure, we show a nighttime stable light maps from DMSP-OLS of western North Dakota. We include horizontal well data from Enverus and shale play boundaries from the U.S. Energy Information Administration.

This paper provides the first quantitative assessment of the impacts of shale drilling on rural light pollution in the United States. It verifies previous speculation that a relationship exists between shale development and light pollution in the Bakken (Barentine 2019). We capture light pollution at a spatial resolution of 25 km² grid cells using an intercalibrated version of the best available, long-term dataset on nighttime lights: the Defense Meteorological Satellite Program – Operational Linescan System (DMSP-OLS) nighttime lights data (Zhang et al., 2016). Relative to previous investigations, the high spatial frequency of our analysis is important to isolate drilling and related infrastructure-fueled light pollution apart from that related to general economic growth (Henderson, Storeygard, and Weil 2012) and commercial expansion, or “boomtown” effects (Smith and Wills 2018). Our main approach focuses on very-rural areas with minimal light pollution in 2000, and identifies how the growth of light pollution differs between rural areas with and without local, contemporaneous shale development. We find evidence that contemporaneous light pollution has grown significantly in areas with shale development – the presence of horizontal drilling increases the probability of exceeding an average digital number (ADN) of 5 by 28.2 percentage points, which is well over 100 percent compared to light pollution levels prior to the shale boom in 2001–2005. The magnitudes of these effects are consistent across U.S. states, and robust to subsample analyses based on different definitions of rurality. We also estimate spatial spillover models to examine the impact of drilling on light pollution in adjacent areas, and find fairly localized effects of up to 10 km. Compared to the large economic spillover effects documented in the literature (beyond 100 miles) (Feyer et al. 2017; Alcott and Keniston 2017), this supports our conclusion that our empirical strategy is able to control for a portion, though not all, of the positive local economic effects associated with light, and that our estimates of changes in light pollution are primarily driven by shale operations. We are, however, not able to account for every single confounding factor and discuss the limitations of our results in the Discussion section.

We also evaluate the potential health costs of increases in light pollution from drilling using repeated cross-sectional county-level self-reported measures of sleep deprivation, mental health, and physical health from the Behavioral Risk Factor Surveillance System (BRFSS). We provide indirect least squares (ILS) estimates of the impact of light pollution on health. Our models find evidence of a dose response relationship between the number of horizontal wells in a county and sleep deprivation, and we find some suggestive evidence supporting drilling impacts on poor physical and mental health.

Our work contributes to the literature that investigates the external costs of shale development in the U.S. There is growing evidence that the shale boom has negatively impacted local environmental quality, including water and air quality (e.g., Olmstead et al. 2013; Swarthout et al. 2015; Hill and Ma 2017). The extent to which the shale boom has impacted light pollution, however, has remained largely unstudied. Our analysis provides the first suggestive evidence that rural light pollution is associated with inadequate sleep and that locations with more drilling have a stronger association than places with no or less drilling. Our results bear direct implications for public health, as light pollution has been connected to disturbances to human health (e.g., Fonken et al. 2010; Schwimmer et al. 2014) and local biodiversity (e.g., Davies et al. 2013), and for weighing the economic costs of shale development against its economic and environmental benefits (Weber 2012; Brown et al. 2016; Newell and Raimi 2018). These findings will be increasingly important for policymaking in the future, as the U.S. EIA Annual Energy Outlook 2018 projects shale oil and gas production to continue its climb through the coming decades.

2. Background

U.S. natural gas and oil production have increased substantially between 2000 and 2017 (Figure 2). More than 200,000 horizontal wells have been drilled in the U.S. over this time period, which led to the proportion of natural gas from shale plays rising from 6.7 to 42 percent.¹

Figure 2:

Oil and natural gas production in the United States between 2000 and 2012

Note: Annual oil and natural gas production data is from the U.S. Energy Information Administration. We normalize annual values based on 2000 production levels.

The U.S. shale boom began around 2006, when domestic natural gas production rose dramatically (Figure 1).² In 2007, the proportion of natural gas production from shale resources was only ~6%; this proportion rose to ~43% in 2012.³ The shale boom has had large economic impacts (Weber 2012; Brown et al. 2016; Newell and Raimi 2018). However, there is concern that this growth has created significant environmental costs. Most research on environmental impacts related to shale development examines the effects on air and water quality (e.g., Swarthout et al. 2015; Hill and Ma 2017), yet there are also concerns that the shale boom has led to light pollution (Krulwich 2013; Schlanger 2017).

Light pollution has significant consequences for wildlife populations. It affects nighttime behavior and habits of terrestrial (Bennie et al. 2015) and marine (Davies et al. 2014) wildlife populations, particularly for species that use sun or moon light for guidance. It disrupts natural sleep and reproductive cycles, geographical orientation, and predator-prey relationships (Longcore and Rich 2004). Other effects of light pollution include changes in bird singing behavior (Miller 2006), estrus patterns in nocturnal primates (LeTallec et al. 2015), insect pollination (MacGregor et al. 2015), and fish biological rhythms (Brüning et al. 2015). These impacts have led to ecosystem-wide changes in biodiversity and growing disparities between entire taxonomic groups (Davies et al. 2013).

The impacts of light pollution also extend to human health and well-being. Artificial light disrupts melatonin secretion and circadian rhythm (Haim and Zubidat 2015) with corresponding changes on mood regulation, depression, and sleeping disorders (Cho et al. 2016). Light pollution-driven changes in circadian rhythms may also have contributed to recent growth in obesity and metabolic dysfunction (Fonken et al. 2010). Growing laboratory and epidemiological evidence also support the long-hypothesized relationship between nighttime light exposure and cancer rates (Kerenyi et al. 1990; Kloog et al. 2010; Schwimmer et al. 2014; Jones 2018).

While there is some work speculating that light pollution associated with shale development induces psychosocial stress (Fisher et al. 2017), sleep and mental health issues (Casey et al. 2018), and local ecosystems (Kiviat 2013), the literature directly connecting the recent resource boom to light pollution is extremely limited. Importantly, no work has documented the causal impact of U.S. shale development on light pollution. The most related work is Smith and Wills (2018), who also investigate the relationship between natural resource booms and illumination. These authors use nighttime lights data, aggregated to 100 km² grid cells, in a global analysis to estimate whether high oil prices and newly-discovered oil increased illumination, a proxy for economic growth. Several key differences make our investigation uniquely suited to better isolate the light pollution effects of shale development and their associated policy implications, including the sources of variation in nighttime light and the spatial scale.⁴

This lack of research is surprising given the potential extent of light pollution externalities from shale oil and gas development. First, the shale boom has occurred across the United States in areas that have renowned dark skies, such as International Dark Sky Places (Figure 3).⁵ Second, the 24/7 work cycle at many oil fields requires substantial lighting infrastructure to ensure safety for workers given the risky nature of occupations in the oil and gas industry (Witter et al. 2014). Natural gas compressor stations (Junkins 2017) and access roads (Jones et al. 2015) also require artificial lighting. Flaring, the prescribed burning of natural gas at well sites for the purposes of maintenance and production testing,⁶ is another source of light pollution that may last for multiple days or weeks (Ohio Environmental Protection Agency 2014). Flaring is done at different phases throughout the natural gas extraction process and is also used to get rid of natural gas that is a byproduct of oil wells in areas without nearby natural gas pipelines.⁷ Although light pollution stemming from active drilling and flaring may be short-lived, some well pads may experience extended light pollution due to a combination of factors, including horizontal drilling, re-fracturing, and high production decline rates, all of which increase drilling intensity at well pads. Regulations on flaring in the U.S. have limited the occurrence of such events, though flaring is a persistent issue in areas without natural gas pipelines. Systematic evidence that shale operations have increased light pollution would support the existence of a potential mechanism by which the shale boom impacts human health and the environment, such as mental health (Casey et al 2018) and biodiversity (Kiviat 2013).

Figure 3:

International Dark Sky Places and oil and gas development since 2000 in the U.S.

Note: Oil and gas well data is from Enverus and includes all vertical and horizontal oil and gas wells with post-2000 spud dates. Play coverage is from the United States Energy Information Administration’s Tight Oil and Shale Gas Plays layer. International Dark Sky Place locations are from the International Dark Sky Association.

3. Materials and Methods

3.1. Data

3.1.1. Light Pollution

The Defense Meteorological Satellite Program – Operational Linescan System (DMSP-OLS) collects visible and infrared imagery across a wide band of the earth’s surface multiple times a day. The satellites were designed by the U.S. military for nighttime cloud cover forecasting and other meteorological purposes (Mellander et al. 2015). One data byproduct of this system was recorded nighttime light data (Henderson et al. 2012). From these data, the National Oceanic and Atmospheric Administration (NOAA) create cloud-free, annualized images of stable night lights, which greatly reduced the potential measurement error associated with transient light sources.⁸ The resulting datasets have a resolution of one square kilometer (Doll 2008).

Each pixel is given what is known as an average digital number (ADN), which is a scaled value from 0 to 63, representing the frequency of nighttime light occurrence per pixel in a given year, normalized across multiple satellites. Higher ADNs indicate areas with more persistent nighttime light during the given year (Doll 2008). The ADN is not directly related to light radiance (Henderson et al. 2012) and is thus a within-year measure of relative brightness across the landscape. Other studies in the academic literature have used these data as direct proxies for urban expansion (Elvidge et al. 1999; Milesi et al. 2003),⁹ energy use and emissions (Doll et al. 2000), and sub-national economic growth (Doll et al. 2006). Because annual DMSP-OSP is not calibrated for intertemporal comparisons, this analysis uses time-calibrated DMSP-OLS stable lights data from 2000 to 2012 from Zhang et al. (2016) in order to facilitate comparisons of nighttime stable lights across time.¹⁰ A nighttime light grid cell of 25 km², hereinafter referred to as a “cell,” in a specific year represents the unit of observation.

We use two continuous measures of light pollution: (1) ADN; and (2) log(ADN+1). We also use three alternative DMSP-OLS average digital number (ADN) thresholds that constitute various levels light pollution: (1) ADN > 5; (2) ADN > 10; and (3) ADN > 20. The literature has used varied thresholds, but our thresholds follow most closely Gaston et al. (2015) which uses 0–5.5, 5.5–9.5 and 9.5–19.5, and Xiang and Tan (2017), which uses 0–5.5 (none), 5.5–10 (moderate), 10–30 (medium), 30–63 (strong).

3.1.2. Oil and gas development

Data on all oil and gas wells drilled in the U.S. from 2000 to 2012 are from Enverus, Inc. We designate all directional and horizontal wells that were drilled after 2000 as “unconventional” since those wells are most likely to involve high-volume hydraulic fracturing. To identify the level of oil and gas development, we calculate the number of horizontal and directional oil and gas wells in each 25 km² cell of the continental U.S. We also identify cells that are adjacent to those that experience drilling (used in the spatial spillover model), specifically those that are 0–5, 5–10, and 10–25km outside of cells with drilled wells. These data are annualized and based on the “spud” date, which is the date on which a well was drilled.¹¹

We collect U.S. shale play boundary data from the U.S. EIA.¹² To our knowledge, this is the most comprehensive database on shale play locations and boundaries in the U.S. We use these data to identify the percent shale coverage in the county of each cell. These measures provide a sense of whether shale development was locally possible. We supplement this geological database with county-level spatial data on shale play depth.¹³ Variation in shale depth allows for within-shale play variation in local resource quality and has been used elsewhere as an exogenous driver of where and when drilling takes place (Weber et al. 2016).

3.1.3. Covariates

We augment our main data set with a series of control variables. Using U.S. Census Bureau data, we identify the state, county, and zip code of each 25 km² cell, while also identifying whether each cell overlays an American Indian Area or Urban Area. We use data on boundaries of U.S. Census Bureau places, which represent locally-recognized incorporated areas where there are settled concentrations of people, to differentiate rural areas from even the smallest of towns. Specifically, we determine whether each cell is within a census place’s boundaries and, if not, whether any portion of the cell is within 10 and 20 km of a census place. Using interstate highway shapefile data from the National Highway Planning Network, we identify those cells that are within 10 km of an interstate highway. Lastly, we identify the land cover of each cell in the dataset prior to the shale boom using National Land Cover Dataset 2001.¹⁴ To control for fluctuations in local economic conditions and population levels, we use Bureau of Labor Statistics annual data on unemployment rates and labor force participation at the county level; the National Institutes of Health’s Surveillance, Epidemiology, and End Results (SEER) Program county-year population data; and median household income from the U.S. Census Bureau’s Small Area Income and Poverty Estimates (SAIPE) database.

3.1.4. Behavioral Risk Factor Surveillance System

The Behavioral Risk Factor Surveillance System (BRFSS) is a commonly used data source to evaluate community health and well-being. The BRFSS is implemented by the Centers for Disease Control (CDC) and is a repeated monthly survey at the county-level.¹⁵ We use data from 2000 to 2012, overlapping with the time period we study for the effect of drilling on light pollution. For our estimates of the effect of drilling on well-being, sleep and physical health, we use the following questions: 1) On average, how many hours of sleep do you get in a 24-hour period? Think about the time you actually spend sleeping or napping, not just the amount of sleep you think you should get; 2) During the past 30 days, for about how many days have you felt you did not get enough rest or sleep?; 3) Now thinking about your mental health, which includes stress, depression, and problems with emotions, for how many days during the past 30 days was your mental health not good?; 4) During the past 30 days, for about how many days did poor physical or mental health keep you from doing your usual activities, such as self-care, work, or recreation?; and 5) During the past 30 days, for about how many days have you felt very healthy and full of energy?. Not all questions are asked in each year. For example, Question 1 is asked for years 2009–2012 and does not provide a pre-drilling measure. We include a table in the Appendix that shows the years available for each question. We dichotomize these questions in the following way: 1) average sleep < 7 hours per night in last 30 days; 2) insufficient sleep > 7.8 days of last 30; 3) bad mental health more than 3.4 days; 4) poor health > 4.8 days of last 30; and 5) healthy > 18.5 days of last 30. We define measures 1–5 using the mean of the dependent variable such that each is equal to 1 if the county value is greater than the mean. Figure A5 in the appendix plots time series of average gas wells in counties with drilling and various measures of sleep and health. These plots provide stylized evidence of the negative correlation between wells drilled and measures of health.

3.2. Empirical Methodology

3.2.1. Light Pollution Model

We estimate the impact of various levels of drilling on nighttime stable lights. Our main model for analysis is shown in (1):

$${Light}_{ict}={\alpha }_0+\beta ·{Wells}_{ict}+\mu ·X_{ict}^{\mathrm{\prime }}+{\tau }_t+{\theta }_i+{ϵ}_{ict}$$ (1)

Our primary outcome of interest Light_ict is a binary or continuous variable indicating light pollution in cell i of county c in year t.¹⁶ For our binary outcomes, we use three alternative DMSP-OLS average digital number (ADN) thresholds that constitute levels of light pollution: (1) ADN > 5; (2) ADN > 10; and (3) ADN > 20. For the continuous variables, we report ADN and log(ADN+1). The main variable of interest is contemporaneous well development Wells_ict, which represents the number of wells drilled within cell i in year t; the parameter of interest, β, thus indicates the impact of each well drilled in cell i during year t on the probability that cell i will exceed a particular light pollution threshold. Alternatively, Wells_ict can be defined as a binary variable equal to 1 if there was a well drilled within the cell in year t, in which case β will be equal to the effect of any drilling during the year on a cell’s likelihood of experiencing light pollution above each threshold. In both cases, we hypothesize that β will be positive, i.e., contemporaneous oil and gas well drilling leads to increases in light pollution. In addition, the specification in Equation (1) controls for a vector of time-varying, county-level economic variables ($X_{ict}^{\mathrm{\prime }}$), including the unemployment rate, labor force participation rate, population level, and median household income level in county c of year t, year fixed effects (τ_t), and cell fixed effects (θ_i). Economic variables are included to capture general economic development associated with light pollution in addition to our sample restrictions. The residual (ϵ_ict) represents all other unobserved factors that contribute to light pollution. An unobserved variable of concern is the presence of astronomical observatories: both observatories and O&G companies may tend to locate in low-light pollution areas, and these facilities may respond to increased drilling to mitigate light pollution (e.g., by enacting agreements to limit light pollution around their facilities, including with the O&G industry). The unobserved, time-varying factor, in this case, would tend to attenuate our estimated impact of drilling on light pollution.

As an additional test of robustness, we estimate a model using data aggregated to the county-level:

$${\mathrm{\%}Light}_{ct}={\alpha }_0+\beta ·{Wells}_{ct}+\mu ·X_{ct}^{\mathrm{\prime }}+{\tau }_t+{\theta }_c+{ϵ}_{ct}$$ (2)

where light pollution in county c of year t is a function of contemporaneous well development, a vector of time-varying county-level variables, year fixed effects, county fixed effects and a residual. In this specification, %Light_ct is the percent of previously-unexposed cells (e.g., ADN=0 in 2000) within county c that has an ADN > 5 in year t. Thus, β represents the marginal impact of each well on the percentage of cells within the county that have an ADN > 5. We again hypothesize that contemporaneous oil and gas well drilling will increase light pollution, i.e., β will be positive.

3.2.2. Health Outcomes Model

To estimate the impact of light pollution on health outcomes at the county-level, we estimate the reduced form model described in model (2) above, except our outcomes are measures of health. The model takes the following form:

$${Health\_Outcome}_{ct}={\alpha }_0+\beta ·{Wells}_{ct}+\mu ·X_{ct}^{\mathrm{\prime }}+{\tau }_t+{\theta }_s+{ϵ}_{cst}$$ (3)

Where Health_Outcome_ct is one of the following outcomes (described above in Section 3.14): 1) average sleep < 7 hours per night in last 30 days per CDC; 2) insufficient sleep > 7.8 days of last 30; 3) bad mental health more than 3.4 days; 4) poor health > 4.8 days of last 30; and 5) healthy > 18.5 days of last 30. The rest of the model includes controls described above under model (2) and includes state and year fixed effects.

We also estimate models on numerous thresholds of drilling to better understand the dose response of more drilling on health outcomes. These models are the same as (3), except we replace Wells_ct with binary variables. The binary measures of drilling are the following: 1) any horizontal drilling and 2) # of horizontal wells greater than 5, 10, 20...,100 wells in the county.

4. Results

4.1. Summary statistics

Table 1 provides summary statistics of light pollution (as measured by DMSP-OLS ADN). Each observation is a 25 km² cell in a particular year. Where applicable, we measure the distance between an entity (e.g. census place) and each cell as the linear distance between (the boundary of) the entity and the boundary of that cell. Panel A shows average ADM in 2000 (the first year of our sample) by different “regions,” such as states with and without shale coverage (row 1) or counties with and without shale coverage (row 2). We check for systematic differences in light pollution before the onset of shale development in order to check for pre-existing regional differences in nighttime lights that are correlated with where shale development will eventually occur. The group averages reveal several differences in light pollution levels across regions. While states (and counties) with and without shale coverage have similar levels of nighttime lights, areas within 10 km of a census place, urban area, and an interstate highway all have higher light pollution levels in 2000. Pre-boom light pollution levels are also lower on Native American lands.¹⁷ In panel B, we also see evidence that areas with vertical oil and gas development before 2006 have higher mean ADN than areas without. In order to minimize differential light pollution trajectories due to these pre-shale boom differences, we restrict our estimation sample to 1) cells in counties with shale coverage, 2) cells that are outside of a 10 km buffer of either a census place or an interstate highway, as well as those outside of Native American reservations, and 3) cells without vertical well development in the period 2000–2005. To further focus on light pollution in rural areas, we make the final sample restriction of examining cells with ADM<5 (i.e. those that began with low levels of light pollution) in 2000. Most, if not all, industrial activity requires some form of lighting. We make these restrictions since we are interested in the effect of drilling and its directly-related activities (e.g., pipeline transmission; natural gas compressor stations) on rural light pollution in areas that generally do not experience oil and gas boomtown-related development (e.g., expanded commercial development). These restrictions help us isolate the strict light pollution effects of shale development from the general economic effects of the oil and gas boom. We are, however, not able to account for every single confounding factor. This could contaminate our interpretation of the effects being completely driven by oil and gas development-related infrastructure, a point to which we return in the Discussion section.

Table 1:

Summary Statistics

	Average digital number in 2000
Panel A	Within Region		Outside of Region
In state with shale acreage	6.26		6.28
In county with shale acreage	6.32		6.25
Census place buffer (10 km)	9.12		0.32
Urban area	33.15		3.06
Native American lands	3.77		6.46
Interstate highway buffer (10 km)	16.28		5.68
Panel B	No Vertical Wells		Vertical Wells > 0
Fraction with ADN > 5 (2001–2005)	0.0032		0.016
	Eventually Drilled		Never Drilled
Panel C	Mean	S.D.	Mean	S.D.
ADN	0.351	0.922	0.141	0.609

Notes: Panel A compares 2000 (i.e., initial year of analysis) levels of DMSP-OLS's average digital number (ADN) at 25 km² cells within different regions, denoted in each row: (1) states with and without shale, (2) counties with and without shale, (3) within and outside of a 10 km buffer of a U.S. Census Bureau-defined census place, (4) within and outside of urban areas, (5) within versus outside of Native American lands, and (6) within versus outside of a 10 km buffer of all Interstate Highways in the U.S. Panel B provides the fraction of cells exceeding a light pollution threshold of ADN=5 by whether cells have any vertical drilling in the years 2001–2005. Panel C provides the light pollution levels in 2000 for the main estimation sample by whether a grid cell is eventually drilled or not.

In Panel C of Table 1, we present summary statistics of ADN in 2000 (before the shale boom) by whether a grid cell is eventually drilled or not for our main estimation sample, i.e., cells with minimal light pollution (ADM<5) in 2000 in counties with shale but no vertical wells from 2000–2005, and outside of urban areas, Native American lands, and census place and interstate highway buffers. These averages provide baseline light pollution levels with which we can gauge the magnitudes of our estimated effects. All subsequent specifications use this restricted sample as our main estimation sample.

Table 2 presents summary statistics of self-reported sleep and health outcomes from the BRFSS (panel A) and sociodemographic characteristics (panel B) by intensity of horizontal drilling by county-year; the two well intensity measures (0<Wells<100 and Wells>100) are based on contemporaneous drilling and not cumulative drilling. Places that were never drilled are comparable to those that will eventually be drilled in terms of sleep and health outcomes. Presence of wells seems to decrease measures of sleep quality and health, although the relationship does not appear to be monotonic. In terms of sociodemographic characteristics, places that never experience drilling are more populated than those with eventual drilling. However, median income, poverty, and unemployment levels are similar across these areas.

Table 2:

Quality of Sleep and Health by Horizontal Well Exposure

	Panel A: Sleep and Health Outcomes
	Never Drilled		Eventually Drilled		0 < Wells < 100		Wells > 100
	Mean	S.D.	Mean	S.D.	Mean	S.D.	Mean	S.D.
Slept <7 in Last 24 Hours	0.315	0.089	0.307	0.082	0.326	0.089	0.371	0.100
Insufficient Sleep >7.8 of Last 30 Days	0.339	0.083	0.353	0.079	0.341	0.101	0.346	0.095
Bad Mental Health >3.4 in Last 30 Days	0.204	0.064	0.215	0.060	0.215	0.079	0.206	0.066
Poor Health >4.8 of Last 30 Days	0.264	0.106	0.260	0.090	0.287	0.114	0.291	0.110
Healthy >18.5 in Last 30 Days	0.572	0.071	0.564	0.072	0.548	0.123	0.551	0.162
Observations*	15,826		817		1,519		74
Unique Counties*	1,881		201		386		23
	Panel B. Demographic Characteristics
	Never Drilled		Eventually Drilled		0 < Wells < 100		Wells > 100
	Mean	S.D.	Mean	S.D.	Mean	S.D.	Mean	S.D.
Population	93,445	250,725	81,806	242,996	86,145	208,827	45,822	51,781
Poverty Rate	14.67	6.21	16.23	5.48	17.04	5.86	15.11	5.42
Unemp. Rate	6.42	2.90	5.71	1.95	6.35	2.64	6.11	2.38
Median HH Income	40,844	11,155	34,531	7,336	39,323	9,379	45,832	8,791
Observations	31,555		2,424		2,832		114
Unique Counties	2,442		634		549		270

Note: This table presents summary statistics of health measures relating to sleep, mental health, and overall health (panel A) and demographic characteristics (panel B) by the horizontal drilling intensity of a county. Specifically, we provide summary statistics for counties that are never drilled (column 1), counties that will eventually be drilled but have no horizontal wells (column 2), counties that have between 0 and 100 horizontal wells (column 3), and counties that have more than 100 wells (column 4); these well counts are for a given county-year and are not cumulative, which makes columns 3 and 4 not mutually exclusive.

^*Due to BRFSS data availability, the sample size for variables in panel A differs; we provide the maximum number of observations and number of unique counties available. See Appendix Table A2 for variable availability by year and Appendix Table A3 for specific samples for the regressions.

4.2. Drilling Impacts on Light Pollution

We start by exploring our results in graph form. In Figure 4, we highlight how the percent of cells with ADN > 5 evolves between 2000 and 2012 for two groups: (1) cells with a horizontal well within its boundary drilled at any time during the study period, and (2) cells without a horizontal well drilled at any time during the study period. Between 2000 and 2006, there is little evidence that the level of light pollution differs between cells with and without future drilling. After 2006, the level of light pollution starts to increase for the group with horizontal wells ever drilled. Starting in 2009, the difference in light pollution between the two groups becomes stark. In 2012, ~35% of cells with any horizontally drilled well exceed the light pollution threshold of ADN>5 relative to only ~3% for the group with no wells drilled at any point during the study period.¹⁸

Figure 4:

Growth in light pollution over time by treatment group status

Notes: In this figure, we highlight the percent of cells with a DMSP-OLS ADN > 5 between 2000 and 2012 for two groups of cells separated by whether they ever have a horizontal well drilled during the study period. In 2000, both groups have no cells with DMSP-OLS > 5 by design.

We present our baseline Equation (1) estimates in Table 3. Our outcome of interest is a binary variable equal to 1 if the cell’s average digital number (ADN) is greater than a certain light pollution threshold. In Columns 1 and 2, we apply a threshold of greater than 5, as used in Figure 4. In Columns 3 and 4, we use a threshold of greater than 10, which is similar to our estimate of mean ADN within 10 km of a U.S. Census Bureau-defined place (Table 1 Panel A; 9.12 ADN). In Columns 5 and 6, we use a threshold of greater than 20, which is similar to our estimate of mean ADN within 10 km of a U.S. interstate highway (Table 1 Panel A; 16.28 ADN). We estimate two models for each threshold: (1) where we only include the number of horizontal wells drilled within each cell in the current year; and (2) where we additionally include the numbers of vertical wells drilled during the year. By including both horizontal and vertical well counts in Columns 2, 4, and 6, we test whether there is a meaningful difference in the effect of well type on light pollution.

Table 3:

Impact of shale gas and oil development on rural light pollution

	(1)	(2)	(3)	(4)	(5)	(6)
Dep. Var.:	ADN > 5		ADN > 10		ADN > 20
# of horiz. wells	0.0601***	0.0602***	0.0409***	0.0409***	0.0127**	0.0127**
	(0.0124)	(0.0124)	(0.00989)	(0.00989)	(0.00509)	(0.00509)
# of vert. wells		−0.00200		−0.000685		−0.000175
		(0.00234)		(0.000619)		(0.000223)
Cell-level FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
Control variables	Y	Y	Y	Y	Y	Y
# of observations	260,556	260,556	260,556	260,556	260,556	260,556
R-squared	0.364	0.364	0.319	0.319	0.186	0.186

Notes: This table models 25 km² cell-level light pollution as a function of contemporaneous horizontal and vertical drilling using DMSP-OLS data from 2000–2012. The outcome variable is a binary variable equal to 1 if the cell has contemporaneous DMSP-OLS ADN > 5 (columns 1, 2), ADN > 10 (columns 3, 4), or AND > 20 (columns 5, 6). We additionally control for annual county-level population, poverty rate, unemployment rate, and median household income. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

In Table 3, we find clear evidence that horizontal oil and gas well development increases the likelihood of a cell experiencing light pollution above our various thresholds. Each well is associated with a 6.0 percentage point (pp) increase in the probability that the cell will have ADN > 5 (Column 1). This represents an increase in light pollution that is 20 times the baseline probability of exceeding the ADN-5 threshold from 2001–2005 (0.29 percent). Using a threshold of ADN > 10 in Column 3, we estimate that each well increased the probability of light pollution by 4.1 pp. With the most stringent threshold value of ADN > 20 (i.e., the level of light pollution similar to those near interstate highways) in Column 5, each well increased the probability of light pollution by 1.3 pp. The estimates in Columns 3 and 5 suggest that shale development increases the probability that a cell experiences light pollution similar to that experienced near interstate highways and U.S. Census Bureau-defined settled places. All estimated effects are statistically significant at the 5% level. Including the number of vertical wells drilled within 5 km of the cell (Columns 2, 4, and 6) does not alter the impact of horizontal wells.¹⁹

In Table 4, we instead estimate the impact of a cell experiencing any oil and gas well development in a given year. Our qualitative findings in Table 3 are supported using this alternative treatment definition. In Column 1, horizontal well development increases the likelihood of a cell experiencing light pollution of ADN>5 by ~28 pp. Using the more stringent light pollution thresholds of ADN > 10 and > 20, we estimate that horizontal well development increases the likelihood of light pollution by ~16 and ~5.1 pp, respectively. In Columns 2, 4, and 6, we find additional evidence that vertical oil and gas development has a much smaller effect than that associated with horizontal oil and gas development.

Table 4:

Impact of shale gas and oil development on rural light pollution, binary treatment

	(1)	(2)	(3)	(4)	(5)	(6)
Dep. Var.:	ADN > 5		ADN > 10		ADN > 20
Horiz. wells > 0	0.282***	0.283***	0.160***	0.160***	0.0505***	0.0505***
	(0.0375)	(0.0375)	(0.0352)	(0.0353)	(0.0164)	(0.0164)
Vert. wells > 0		−0.00843***		−0.00341***		−0.000821
		(0.00383)		(0.00173)		(0.000535)
Cell-level FE	Y	Y	Y	Y	Y	Y
Year FE	Y	Y	Y	Y	Y	Y
Control variables	Y	Y	Y	Y	Y	Y
# of observations	260,556	260,556	260,556	260,556	260,556	260,556
R-squared	0.365	0.365	0.289	0.289	0.176	0.176

Notes: This table models 25 km² cell-level light pollution as a function of contemporaneous horizontal and vertical drilling using DMSP-OLS data from 2000–2012. The outcome variable is a binary variable equal to 1 if the cell has contemporaneous DMSP-OLS ADN > 5 (columns 1, 2), ADN > 10 (columns 3, 4), or AND > 20 (columns 5, 6). We additionally control for annual county-level population, labor force size, unemployment rate, and median household income. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

We also estimate versions of Model 1 where Light_ict is measured as a continuous variable using either the mean DMSP-OLS within the 25km² cell or the natural log of this value + 1 to account for 0 values. We briefly describe these results, and point estimates are presented in Appendix Table A1. We find that each horizontal well is associated with an increase in the mean DMSP-OLS within the cell by 0.9 units (Columns 1–2, Panel A) or 17% (Columns 3–4, Panel A). When we use a binary treatment variable equal to 1 with the presence of horizontal drilling within a cell, we find that horizontal drilling is associated with a 4-unit or 87% increase in DMSP-OLS ADN.

Next, we explore whether the impact of drilling on rural light pollution is heterogeneous by states using a threshold of ADN > 5 (consistent with the specification in Table 3, Column 2). Estimated impacts are presented in Figure 5. All but one of our point estimates are positive and nine out of the twelve estimates are statistically significant at the 1% level. Our point estimate is negative in Montana, though there is more limited drilling here than in all other states shown. Ohio and Wyoming are the exceptions. There is significant variation in parameter estimates, which may reflect greater flaring and differences in topography that influence light dispersion across the landscape.

Figure 5:

Heterogeneity of effect across states with active shale development

Notes: In this figure, we present estimates of the specification in Column 2 of Table 2 on a state-by-state basis – the impact of an additional horizontal well on exceeding the light pollution threshold of DMSP-OLS ADN > 5. Oil and gas drilling data are from Enverus. All wells shown were drilled between 2000 and 2012. Shale/tight oil and gas play coverage is from the U.S. Energy Information Administration. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

4.3. Robustness checks

We evaluate the robustness of our results to alternative subsamples of the data based on different definitions of rurality in Table 5. Panel A re-estimates the specification from Table 3 (Column 2) using a continuous treatment variable of # of horizontal and vertical wells, and Panel B re-estimates the specification from Table 4 (Column 2) using the presence of any horizontal and vertical wells. All specifications use an outcome variable based on the ADN threshold of 5. In the main analyses, we deliberately dropped observations close to towns and cities – those within 10 km of an interstate highway or a U.S. Census Bureau place – as well as those with pre-existing vertical well development (in 2000–2005) in order to capture the effect of shale drilling and nearby related infrastructure on light pollution in rural areas that are separate from boomtown effects. Examples of potential boomtown effects include residential and commercial development near towns and expanded local infrastructure. These factors may influence local light pollution levels, but they are best viewed as being separate from light pollution effects that stem from drilling for both management and policy relevance. In Columns 1 and 2, we use alternative distance restrictions of 5 and 10 km buffers of U.S. Census places. In Column 3, we retain cells that are near interstate highways. Column 4 makes no sample restrictions based on pre-shale boom vertical well-development. Finally, Column 5 restricts the sample to cells with no light pollution (i.e., ADN = 0) in 2000, which represent the darkest areas in the country as of 2000, rather than to areas with minimal light pollution (i.e., ADN < 5) in 2000 in the main specification.

Table 5:

Robustness checks based on alternative subsets of the data

	(1)	(2)	(3)	(4)	(5)
	Dep. Var.: ADN > 5
Panel A	Census Place		No Interstate Highway drop	No Pre-2005 V.Well drop	Initial ADN = 0
	> 5 km	> 20 km
# of horiz. wells	0.0567***	0.0335***	0.0613***	0.0326***	0.0599***
	(0.0111)	(0.0104)	(0.0119)	(0.00725)	(0.0139)
# of vert. wells	−0.000400	−0.00375	−0.00248	0.00269**	−0.00180
	(0.00155)	(0.00240)	(0.00219)	(0.00119)	(0.00131)
# of observations	363,800	128,677	263,715	302,926	236,645
R-squared	0.373	0.333	0.370	0.364	0.359
	Dep. Var.: ADN > 5
Panel B	Census Place		No Interstate Highway drop	No Pre-2005 V.Well drop	Initial ADN = 0
	> 5 km	> 20 km
Horiz. wells > 0	0.290***	0.174***	0.288***	0.218***	0.276***
	(0.0328)	(0.0373)	(0.0363)	(0.0278)	(0.0410)
Vert. wells > 0	−0.00709**	−0.00625	−0.0111***	−0.00307	−0.00626**
	(0.00359)	(0.00383)	(0.00403)	(0.00397)	(0.00310)
# of observations	363,800	128,677	263,715	302,926	236,645
R-squared	0.375	0.334	0.370	0.367	0.352

Notes: Panels A and B of this table respectively re-estimates the specifications in Tables 2 and 3 using alternative subsamples. In Columns 1 and 2, we use alternative restrictions of > 5 and > 20 km from the closest U.S. Census Bureau-defined place (compared to the 10 km restriction used in Tables 2 and 3). In Column 3, we do not drop observations based on distance to the closet interstate highway. In Column 4, we make no restrictions based on pre-shale boom vertical well-development. Column 5 uses only those observations that had no light pollution in 2000 (i.e., DMSP-OLS ADN = 0). We control for annual county-level population, labor force size, unemployment rate, and median household income. All specifications include cell fixed effects and year fixed effects. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

The estimated effects of horizontal drilling on light pollution based on various subsamples are consistent with those in our main results and support the findings in earlier analyses. Parameter estimates suggest that each horizontal well increases the likelihood of light pollution by 3.3 to 6.1 pp, and any horizontal drilling increases the probability of exceeding the ADN threshold of 5 by 17.4 to 29 pp. Thus, our results are robust to alternative restrictions based on rurality, including the most restricted definition of areas with no light pollution in 2000. While we contextualize our effects as the impacts of shale development on rural light pollution, there is still a possibility that our effects are confounded by general commercial and residential development due to the shale boom.

We next estimate whether the effects of light pollution spillover into grid cells that do not themselves experience drilling, and present results in Table 6. We estimated models that allow for a spillover effect for drilling outside the cell in distance bins of 0–5, 5–10 and 10–25km. We find statistically precise impacts that extend to 10km outside of cells with drilling, but, as expected, the magnitudes of these effects attenuate greatly as the distance increases. These spillover effects have two important implications. First, that light pollution may spillover into the control group of grid cells suggests that our estimated light pollution effects are lower bounds of the true impact. Second, previous literature has found that economic spillovers persist to beyond 100 miles from counties with drilling (Feyer et al. 2017, Allcott and Keniston 2017). That our spatial spillover models find that light pollution impacts persist to only 10km and are attenuated 6-fold from the within-cell impacts suggests that our empirical strategy is able to control for a portion, though not all, of the positive economic effects associated with light.

Table 6:

Spillover Effects of Drilling on Light Pollution

	Panel A. Continuous Treatment
Dep. Var.:	ADN>5	ADN>10	ADN>20
# Horiz. wells in cell	0.0342***	0.0262***	0.00734***
	(0.00474)	(0.00431)	(0.00264)
# Horiz. Wells 0–5km outside cell	0.00699***	0.00383***	0.00140***
	(0.000854)	(0.000755)	(0.000497)
# Horiz. Wells 5–10km outside cell	0.00163***	0.000834**	0.000249
	(0.000583)	(0.000351)	(0.000173)
# Horiz. Wells 10–25km outside cell	0.000480*	0.000310	0.000130
	(0.000257)	(0.000201)	(9.03e-05)
Observations	265,577	265,577	265,577
R-squared	0.404	0.366	0.209
	Panel B. Binary Treatment
Dep. Var.:	ADN>5	ADN>10	ADN>20
=1 if # Horiz. wells in cell > 0	0.260***	0.151***	0.0488***
	(0.0389)	(0.0373)	(0.0168)
=1 if # Horiz. Wells 0–5km outside cell > 0	0.0398***	0.0177***	0.00441**
	(0.00888)	(0.00555)	(0.00173)
=1 if # Horiz. Wells 5–10km outside cell > 0	0.0160***	0.00606***	0.00159***
	(0.00328)	(0.00182)	(0.000608)
=1 if # Horiz. Wells 10–25km outside cell > 0	−0.00210	−0.00232**	−0.000735*
	(0.00133)	(0.000934)	(0.000386)
Observations	265,577	265,577	265,577
R-squared	0.360	0.276	0.169

Note: This table presents estimates of spillover effects of drilling on light pollution in areas that do not experience drilling (i.e., for grid cells outside of cells that contain positive number of horizontal wells). All regressions include controls as in Table 3, with additional controls for vertical drilling in adjacent cells. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

4.4. Falsification testing

We provide evidence that our estimated link between drilling and growth in rural light pollution is not driven by differential trends in light pollution prior to drilling. We examine this through an event study regression using only pre-shale boom data from 2000–2005 that have no horizontal well development:²⁰

$${Light}_{ict}={\alpha }_0+\beta \bullet {Ever\_Horizontal}_i\bullet {\tau }_t+\mu \bullet X_{ict}^{\mathrm{\text{'}}}+{\tau }_t+{\theta }_i+{\epsilon }_{ict}$$ (4)

In Equation (3), Ever_Horizonal_i is an indicator equal to 1 if cell i will ever host horizontal drilling in the future, and 0 otherwise. The parameters of interest are the vector of β that capture the year-by-year differential effect between areas with and without future horizontal drilling on the likelihood of experiencing light pollution across various thresholds, and indicates whether there were pre-boom differential trends in rural light pollution areas with and without future shale drilling.

Table 7 presents the estimates from model (4) using the three light pollution thresholds (ADN = 5, 10, and 20) as outcome measures. We find no evidence that areas with future drilling trended differently over time relative to areas without future drilling during the pre-boom period. This thus alleviates the concern that our estimated light pollution impacts are driven by pre-drilling differences in nighttime lights between areas with and without drilling.

Table 7:

Falsification testing on pre-boom changes in light pollution

	(1)	(2)	(3)
	ADN > 5	ADN > 10	ADN > 20
2001 * Ever Horizontal	0.00211	0.000943	0.00108
	(0.00430)	(0.00133)	(0.00126)
2002 * Ever Horizontal	−0.00407	−0.000592	−6.22e-05
	(0.00433)	(0.000394)	(0.000116)
2003 * Ever Horizontal	−0.000539	−0.000224	−0.000148
	(0.00471)	(0.000745)	(9.96e-05)
2004 * Ever Horizontal	0.000815	−0.000173	0.00103
	(0.00528)	(0.00127)	(0.00124)
2005 * Ever Horizontal	−0.000450	0.000672	0.000239
	(0.00485)	(0.000748)	(0.000221)
# of observations	119,640	119,640	119,640
R-squared	0.814	0.798	0.753

Notes: In this table, we test for whether there were differences in light pollution trajectories between areas with vs. without future shale development that preceded actual drilling. Based on Model (4), we use data from 2000–2005 and examine whether there was a differential trend in the likelihood of light pollution on an annual basis between areas that ever have a horizontal well after 2006 (Ever Horizontal == 1) versus those that never have a horizontal well (Ever Horizontal == 0). All specifications control for annual county-level population, labor force size, unemployment rate, and median household income, as well as cell fixed effects and year fixed effects. Standard errors are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

4.5. Light Pollution Impacts on Sleep and Mental Health

ADN is not easily translated to a radiance measure as previously described, and therefore, may not reflect economically meaningful changes in light pollution. We start by estimating OLS regressions of county-level ADN on measures of sleep and health in Figure A6. We stratify by urban, semi-urban and rural areas to see how these levels of ADN are associated with our sleep and health measures. These are simply the direct effect and do not consider endogeneity or places with and without drilling. We do not find strong associations between ADN threshold and the sleep measures, except for rural areas, with worse sleep (insufficient sleep above mean 7.8 days in last 30 days) associated with ADN>20.

To better assess the relationship between light pollution and sleep, we examine the sleep impacts associated with light pollution that are induced by horizontal drilling. Although omitted factors associated with fracking and light pollution (e.g. Cunningham, DeAngelo, and Smith (2020)) still prevent interpretation of these estimates as causal, the correlations using variation in drilling alleviate some of the endogeneity concerns associated with health outcomes. In Figure 6, we present model (2) regression results in graphical form such that each bin of number of wells (i.e., 5, 10, 20) is a separate regression with the binned number of wells as the independent variable of interest. We find a dose response for various measures of sleep and a somewhat weaker dose response for poor health (physical or mental). We find no clear relationship between drilling and reported healthy, energetic days (see Figure A7 in the appendix for estimates of healthy days). We present the associated regression results in Table A3 where for each outcome we present county and state fixed effects regressions separately.²¹ Counties with more than 100 wells drilled see increased incidences of sleeping less than 7 hours in the last 24 hours (6 percentage points) and insufficient sleep (2.5 percentage points). The incidence of poor health and bad mental health also increase by between 1 and 1.5 percentage points, depending on the number of wells drilled. In Table 8, we report the reduced form using a continuous value of horizontal drilling such that we can calculate ILS using our estimates of drilling on light pollution in Table 3. We find that each additional well increases the likelihood of sleep <7 hours and the likelihood of insufficient sleep respectively by 1.3 (not statistically significant) and 0.3 (p<0.05) percent, relative to the mean. There is also some evidence of adverse impacts on mental health, where an additional well is associated with an increase in the incidence of bad mental health of 0.2 percent (p<0.1). In the ILS, we do not find a statistically precise effect of light on number of healthy days in the last 30 days. Similar to Maguire and Winters (2017), who examine the impact drilling on life satisfaction and mental health using BRFSS data in Texas, we also find associations between drilling and self-reported mental health.

Figure 6:

Reduced Form Estimates of Wells Drilled on Health Outcomes

Note: This figure estimates the impacts of wells drilled on measures of sleep and health. Specifically, we regress an outcome of interest on a binary variable equal to 1 if the number wells exceed x (for x = 5, 10, 20, ..., 100), and plot the coefficient on the binary variable from each regression. Standard errors are clustered at the county level, where 90 and 95% confidence intervals are provided in figures. Hours of sleep (top left) is estimated off of 2009–2012 data as BRFSS did not ask this question prior to 2009.

Table 8:

Indirect Least Squares Impact of Light on Health Outcomes

Health Outcomes:	Sleep <7 Hours	Insufficient Sleep	Poor Health	Bad Mental Health	Healthy
A. Effect of Drilling on Health	0.000244	5.13e-05**	2.58e-05	2.20e-05*	−2.83e-05
	(0.000170)	(2.39e-05)	(2.02e-05)	(1.22e-05)	(9.14e-05)
B. Mean of Health Outcome	0.3197	0.3360	0.2683	0.2057	0.5704
C. Impact of Drilling on Light (ADM>5)	0.0602	0.0602	0.0602	0.0602	0.0602
Impact of Light on Health:
% Point Impact (A/C)	0.0041	0.0009	0.0004	0.0004	−0.0005
% Impact Relative to Mean (A/C/B)	0.0127	0.0025	0.0016	0.0018	−0.0008

Note: This table presents the impact of increasing the number of horizontal wells on sleep quality and overall health (A), the mean of the health measure (B), and the reduced form impact of increasing the number of horizontal wells on light pollution from Table 3, column 2 (C). Using these estimates, the last two rows respectively calculate the indirect least squares estimate of light pollution on health measures as a percentage point impact and the percent impact relative to the health outcome mean. Standard errors for the impact of drilling on health (row A) are clustered at the county level. Statistical significance at 1% (***), 5% (**), and 10% (*).

5. Discussion

In this study, we evaluate how the shale boom impacted rural light pollution. We find evidence that drilling increased the dispersion of nighttime stable lights in areas that, prior to the shale boom, had minimal light pollution by over 100 percent relative to pre-boom light pollution levels. We find no evidence that our results are driven by pre-existing, differential trends in nighttime lights between areas with and without drilling. Our findings are also robust to using subsamples based on geographic areas or alternative definitions of what constitutes a rural area. We find a dose response between number of horizontal wells and measures of insufficient sleep, where counties with more than 100 horizontal wells are 3 percentage points more likely to report insufficient sleep and 6 percentage points more likely to report sleep less than 7 hours per night (reduced panel 2009–2012). We find weaker associations between drilling and measures of poor health (physical or mental). Combining the reduced-form estimates of wells drilled on light pollution and sleep measures, we find that each additional well is correlated with an increased likelihood of insufficient sleep by 0.3 percent relative to the mean.

Several issues should be kept in mind when interpreting these results. First, the light impacts that we identify are potentially driven by both well operations and local economic development. In this case, the health impacts associated with changes in light can be either negative (from the undesirable effects of light pollution) or positive (from local economic development and growth), and attempts to isolate the negative health impacts of light pollution may result in attenuated estimates. On the other hand, the negative association between insufficient sleep and horizontal drilling may result from the combined effects of various nuisances that are correlated with drilling-related light pollution, such as air pollution or noise, and thus may be overstated. Additional work is therefore needed to isolate the negative sleep and health impacts through this particular mechanism; our results, however, provide suggestive evidence that is consistent with shale-related light pollution negatively impacting sleep (and, to a lesser extent, physical and mental health) using a nationwide sample. We caveat this with two additional limitations; (1) our measures of sleep, health and mental health are all self-reported, and (2) the BRFSS data do not survey every county in the US or ask the same questions every year, and so our estimates are based off of an unbalanced panel.

To our knowledge, this is the first study to demonstrate the effects of shale oil and gas development on rural light pollution and the associated correlations with sleep in the United States. We focus on very rural areas, far from interstate highways and U.S. Census Bureau-defined places, which allow us to recover changes in light pollution driven by shale operations apart from the boomtown effects of drilling. This is supported by models estimating spatial spillovers of light pollution being fairly local (up to 10km) vis a vis the large economic spillover effects documented in the literature (beyond 100 miles) (Feyer et al. 2017; Alcott and Keniston 2017). This provides support for an additional environmental pathway of concern beyond the commonly-suspected mechanisms of air or water pollution. The implications of our findings are crucial for the well-being and success of eco-systems, a matter of growing global concern especially in migratory corridors (Cabrera-Cruz et al. 2018), areas with exceptional biodiversity (Koen et al. 2018), and protected areas (Gaston et al. 2015), as well as for human health (Fisher et al. 2017; Casey et al. 2018).

Our analysis suggests several potential pathways for future investigation. First, future work could extend the time frame of analysis beyond 2012, still a time of rapid shale well drilling,²² to explore the permanency of light pollution in areas where the shale boom has slowed using later vintages of light pollution data as they become available.²³ If the impact is largely felt contemporaneous of development and disappears afterwards, then policy prescriptions should minimize light pollution during active well drilling and preparation. Second, while we are not aware of any state or county-based light pollution mandates, several groups have promoted best management practices to mitigate light pollution at drilling sites (e.g. the Apache Corporation and the Permian Basin Petroleum Association (Wicker et al. 2017) and the Nature Conservancy (The Nature Conservancy 2015)). Future work should assess the effectiveness of current approaches towards the mitigation of light pollution from shale development.

Our research is critical in light of growing evidence of the impacts of light pollution on human and non-human health and well-being, such as sleep as shown in this paper. Shale oil and gas development has certainly produced positive economic changes at both local and regional levels. Yet, our study provides support that these positive developments have come at a cost of greater light pollution in previously-low light pollution areas. As an externality of production, this impact should be considered in environmental and resource-related policymaking.

Supplementary Material

Online Appendix

Table A1: Impact of shale development on rural light pollution, continuous measures of light pollution

Table A2: BRFSS Availability by Year

Table A3: Impact of Horizontal Drilling on Sleep

Figure A1: Shale gas production as a proportion of total natural gas production in the United States, 2000-2012

Notes: Annual shale gas production and total natural gas production data are from the U.S. Energy Information Administration.

Figure A2: Growth in light pollution over time by treatment group status (DMSP-OLS ADN > 10)

Notes: In this figure, we highlight the percent of cells with a DMSP-OLS ADN > 10 between 2000 and 2012 for two groups of cells separated by whether they ever have a horizontal well drilled during the study period. We rely on an “ever drilled” measure of horizontal well development in this figure, though in our statistical models we rely on a contemporaneous measure of horizontal well development for our treatment variable. In 2000, both groups have no cells with DMSP-OLS > 10 by design.

Figure A3: Growth in light pollution over time by treatment group status (DMSP-OLS ADN > 20)

Notes: In this figure, we highlight the percent of cells with a DMSP-OLS ADN > 20 between 2000 and 2012 for two groups of cells separated by whether they ever have a horizontal well drilled during the study period. We rely on an “ever drilled” measure of horizontal well development in this figure, though in our statistical models we rely on a contemporaneous measure of horizontal well development for our treatment variable. In 2000, both groups have no cells with DMSP-OLS > 20 by design.

Figure A4: Drilling over Time by Well Type

Note: This figure presents the temporal change in the share of wells drilled by well type from 2000 to 2018.

Figure A5: Time Series of Horizontal Drilling, Sleep, and Health

Note: Figure plots times series of average new horizontal wells and measures of health outcomes within a county over time.

Figure A6: OLS Impact of Light Pollution on Health Outcomes

Note: Figure plots point estimates from a regression of binary measures of health outcomes on light pollution for the full sample and subsamples based on rurality. Light pollution is measured as the share of a county that exceeds 5, 10, or 20 ADM. Standard errors are clustered at the county level, where 90 and 95% confidence intervals are provided in figures.

Figure A7: Impacts on Healthy Days (OLS Light Estimates and Drilling Estimates)

Note: Figure on the left plots point estimates from a regression of binary measures of healthy days > 18.5 out of last 30 days on light pollution for the full sample and subsamples based on rurality. Light pollution is measured as the share of a county that exceeds 5, 10, or 20 ADM. Figure on the right plots point estimates from a regression of healthy days > 18.5 on a binary variable equal to 1 if the number wells exceed x (for x = 5, 10, 20, ..., 100). Standard errors are clustered at the county level, where 90 and 95% confidence intervals are provided in figures.