NCEE Working Paper
The Property Value Impacts of
Industrial Chemical Accidents
Dennis Guignet, Robin Jenkins, James Belke,
and Henry Mason
Working Paper 22-04
September, 2022
U.S. Environmental Protection Agency |k|f"CC jgf
National Center for Environmental Economics fw
https://www.epa.aov/environmental-economics environmental^conomics
-------
The Property Value Impacts of Industrial Chemical Accidents
Dennis Guignet, Robin Jenkins, James Belke, and Henry Mason
ABSTRACT: Using hedonic property value methods, we examine how chemical accidents at
industrial facilities impact home values in surrounding communities. The study focuses on
facilities regulated by the U.S. Environmental Protection Agency's (EPA) Risk Management Plan
(RMP) program, which is in place to reduce the risk of harm to offsite populations from accidental
chemical fires, explosions, and releases of toxic vapors. RMP facility and accident data were linked
to residential property transactions in Michigan, Ohio, and Pennsylvania occurring between 2004
and 2014. To facilitate causal inference, alternative quasi-experimental difference-in-differences
and triple differences models are estimated, where we compare homes near and far, and before and
after, an accident; as well as homes near RMP facilities where an accident did and did not occur.
We find that the typical accident does not affect home values, but accidents resulting in offsite
injuries, property damage, evacuations, or shelter-in-place orders lead to a 5% to 7% decrease in
the value of homes within five kilometers, this translates to an average loss of $12,000 to $17,000
per home. The benefits of the RMP program in avoiding these impacts are particularly relevant
from an environmental justice standpoint. We find that proximity to an RMP facility, irrespective
of any incidents, is already associated with significantly lower home values. These existing
inequities are exacerbated by chemical accidents that impact offsite populations.
KEYWORDS: chemical accident, environmental justice, hedonic, nonmarket valuation,
property value, Risk Management Plan
JEL CODES: D63; Q51; Q53
DISCLAIMER
The views expressed in this paper are those of the author(s) and do not necessarily represent those of
the U.S. Environmental Protection Agency (EPA). In addition, although the research described in this
paper may have been funded entirely or in part by the U.S EPA, it has not been subjected to the Agency's
required peer and policy review. No official Agency endorsement should be inferred.
-------
The Property Value Impacts of Industrial Chemical Accidents*
Dennis Guignet,a Robin Jenkins,b James Belke,c and Henry Masond
I. INTRODUCTION
To promote safe industrial activity, the U.S. Environmental Protection Agency (EPA) requires
that certain industrial facilities participate in the Risk Management Plan (RMP) program. The
purpose of the program is to protect the general public and the environment from accidental
chemical releases that can result in toxic clouds, fires, and explosions. While on-site worker
safety is regulated under the purview of the U.S. Occupational Safety and Health Administration
(OSHA), EPA's RMP program is specifically meant to reduce risks to surrounding populations,
living in nearby "fence-line" communities.
The RMP program covers a wide range of industry categories, from large multi-process
manufacturing plants such as petroleum refineries and chemical manufacturers, to smaller facilities
such as agricultural chemical distributors, water and wastewater treatment systems, and cold
storage facilities. In between there are a variety of other specialty chemical manufacturers, energy
production facilities, chemical storage terminals and warehouses, natural gas processing plants,
and food manufacturers, among others. In total, approximately 12,000 U.S. facilities are currently
regulated under the RMP program.1 EPA estimates that at least 40 million people (or about 12%
of the U.S. population), and perhaps as many as 177 million (55%), are at risk of being impacted
*The first author's work on this research was supported by the Dean's Club Summer Research Grant from the
Walker College of Business at Appalachian State University. The fourth author's work on this research was
supported by the U.S. Environmental Protection Agency (EPA) Research Participation Program administered by the
Oak Ridge Institute for Science and Education (ORISE) through an interagency agreement between the U.S.
Department of Energy (DOE) and the EPA. ORISE is managed by Oak Ridge Associated Universities (ORAU)
under DOE contract number DE-SC0014664. Property transaction data were provided by Zillow through the Zillow
Transaction and Assessment Dataset (ZTRAX). More information on accessing the data can be found at
http://www.zillow.com/ztrax. The results and opinions are those of the authors and do not reflect the position or
policies of the US EPA, DOE, ORAU/ORISE, or the Zillow Group. Any errors are those of the authors.
a Corresponding Author: Department of Economics, Appalachian State University, 416 Howard
Street, ASUBox 32051, Boone, NC, 28608-2051. Ph: 828-363-2117. guignetdb@appstate.edu.
b National Center for Environmental Economics, US Environmental Protection Agency.
c Office of Emergency Management, US Environmental Protection Agency.
d Calvert Research and Management.
1 According to EPA's RMP National Database, there were 11,759 active RMP facilities as of July 4, 2021 (Source:
EPA Office of Emergency Management).
1
-------
by an accident at these facilities (U.S. EPA, 2016).2 From 2004-2013, there was an average of
about 150 accidents with reportable impacts each year in the U.S.3 About 31% of those accidents
involved off-site deaths, injuries, property damage, evacuations, sheltering-in-place, or
environmental damage (U.S. EPA, 2016).
The RMP program has three primary objectives: accident prevention, damage mitigation, and
information provision. The program was authorized in the U.S. Clean Air Act (CAA) Amendments
of 1990 as a result of the catastrophic chemical accident that occurred in 1984 in Bhopal, India, as
well as other serious chemical accidents in the U.S. during the late 1980s.4 Section 112(r) of the
CAA Amendments created a "general duty" for owners and operators of "stationary sources"
holding extremely hazardous substances to identify potential hazards, take steps to prevent
releases, and to minimize the consequences of accidental releases that do occur.5 Section 112(r)
also required EPA to develop a list of at least 100 regulated substances for accident prevention6
and to develop "reasonable regulations" requiring facilities with over a threshold quantity of a
regulated substance to undertake accident prevention steps and submit a "risk management plan"
(RMP) to various local, state, and federal planning entities.7 The final RMP regulation
promulgated by EPA in 1996 requires that any stationary source that uses, stores, manufactures,
or handles, more than a threshold quantity of one or more of the 140 listed substances must submit
an RMP to prevent accidents and minimize any potential damage.8
The RMP regulations require regulated facilities to take a number of steps, such as: (i)
performing a hazard assessment to estimate the potential effects of an accidental release; (ii)
implementing a program to prevent accidental releases of regulated substances, including written
safe operating procedures, regular maintenance activities, compliance audits, employee training,
and other measures; and (iii) developing an emergency response program requiring specific
actions in response to an accidental release in order to minimize the impacts to human health
and the environment. The emergency response program includes procedures for informing the
public and local agencies responsible for responding to accidental releases, and annual
coordination with responders, as well as emergency health care and employee training measures.
Owners or operators of regulated facilities must develop and submit a report to EPA that
summarizes the facility's RMP. The report must be updated every five years or sooner, if certain
changes occur at the facility.
2 In the Regulatory Impact Analysis for the "Accidental Release Prevention Requirements" rule, EPA reported that
approximately 177 million people would be impacted by a hypothetical worst-case scenario accident at all RMP
facilities. However, under more likely alternative accident scenarios, EPA reports that approximately 40 million
people are potentially at risk (US EPA, 2016). The estimated percentages of the U.S. population are based on the
estimated total population of 324 million people at the end of 2016 (U.S. Census Bureau, 2016).
3 Some facilities reported accidental releases with no "reportable" impacts (i.e., no deaths, injuries, property damage,
etc.). Such accidents were not counted in determining the annual average number of accidents (US EPA, 2016).
4 For example, the Phillips Petroleum Company disaster of 1989 in Pasadena, Texas involved a series of explosions
and fire with over 20 worker deaths and over 150 injuries.
5 42 U.S.C. § 7412(r)(l)
6 CAA section 112(r)(3)-(5)
7CAA section 112(r)(7)(B)
8 61 FR 31668, June 20, 1996
2
-------
Economic analyses conducted by EPA for recent RMP program rulemakings have omitted
estimation of benefits and been limited to presentations of baseline damages (U.S. EPA 2016,
2019). This study fills an important gap by providing valuable information to decision makers
assessing the benefits of policies to prevent and mitigate industrial chemical accidents.
Focusing on over 3.6 million single-family home transactions in Michigan, Ohio, and
Pennsylvania from 2004-2014, we use hedonic property value methods to examine how chemical
accidents at RMP facilities impact nearby residents. To support causal inference, difference-in-
differences and triple differences quasi-experimental models are estimated, where we compare
homes near and far, and before and after, an accident; as well as homes near RMP facilities where
an accident did and did not occur. Although there are local case studies on the property value
impacts of similar disamenities (e.g., Carroll et al. 1996; Hansen et al. 2006; Grislain-Letremy and
Katossky 2014; and Herrnstadt and Sweeney 2019), to our knowledge this is the first
comprehensive broader-scale non-market valuation study on accidents at chemical facilities, and
the first specifically on the RMP program.
This study is also unique compared to quasi-experimental hedonic property value studies of other
EPA programs involving the cleanup of hazardous chemicals (e.g., Cassidy et al., 2021; Gamper-
Rabindran and Timmins, 2013; Guignet et al., 2018; Guignet and Nolte, 2021; Haninger et al.,
2017). Relative to lingering environmental contamination at Superfund sites, brownfields,
underground storage tank releases, and other types of contaminated industrial sites, the damages
addressed by the RMP program are immediate, and perhaps more evident in many cases. These
other categories of contamination often involve releases that are not immediately apparent, often
entailing, for example, slowly spreading releases into soil or groundwater. In contrast, RMP
accidents can involve fires, explosions, immediate shelter-in-place and evacuation orders, first-
responder activities, and other overt cues that clearly alert the surrounding community. Certain
information about RMP accidents may be more transparent; including the identity of the
responsible party, which may be reported by local press.
We find that the typical accident does not affect home values, on average, but accidents resulting
in offsite injuries, environmental damage9, evacuations, or shelter-in-place orders lead to a 5% to
7% decrease in the value of homes within five kilometers. Widespread awareness of accidents
resulting in offsite impacts among buyers and sellers in nearby housing markets is more likely
relative to accidents resulting only in onsite impacts. Accidents with offsite impacts are generally
more severe, and local communities are directly affected.
This finding is important to policy makers. Property value impacts from accidents affecting offsite
populations are particularly relevant for EPA's RMP program. EPA's responsibility is primarily
to reduce the risk of offsite harms, including to the general public and the environment (US EPA
1996).
RMP program benefits are also important from an environmental justice standpoint. RMP facilities
that are larger (indicated by number of workers), that use more regulated chemicals, and that have
9 Types of environmental damage include fish or animal kills; lawn, shrub, or crop damage; water contamination; etc.
3
-------
experienced more than one accident, are more likely to be in counties where a greater proportion
of the population is Black, and/or where (although the median income is higher) there are higher
levels of income inequality (Elliot, et al. 2004). Our data show homes within two kilometers of a
single RMP facility, irrespective of any accidents, tend to be 2% to 4% lower in value, all else
constant. We also find strong cumulative effects associated with proximity to multiple RMP
facilities. For example, being located within one kilometer of three RMP facilities is associated
with 13% lower home values. Additional property value declines due to chemical accidents at
these facilities further exacerbate existing inequities.
In the next section we review the relevant literature on the property value impacts of similar
disamenities involving releases of hazardous chemicals. The data and methods are described in
sections III and IV, respectively, followed by the results in section V. We provide concluding
remarks in section VI.
II. LITERATURE
Prior research involving RMP facilities includes descriptive analyses of data from the initial wave
of RMP submissions covering accident histories from 1994 to 2000. This work identified
directions for future research (Kleindorfer, et al., 2003), considered what aspects of an RMP
facility best characterized its "hazardousness" (Elliot, et al., 2003), examined the relationship
between facilities' propensity for accidents and financial condition, capital structure, regulatory
environment, and other characteristics (Kleindorfer et al., 2004), and explored the frequency of
RMP chemical accidents and riskiness of RMP facilities in relation to the socioeconomic status of
surrounding communities (Elliot et al., 2004). However, to date no study has explored the
relationship between RMP facility accidents and surrounding property values.
The economics literature has explored the effect on property values of a wide assortment of
environmental disamenities. Grislain-Letremy and Katossky (2014) provide a targeted review of
property price impacts of hazardous industrial facilities, reviewing studies that examined data for
one to a handful of affected regions. In general, these studies find that prices increase with distance
from the facility and that the magnitude of the price impact varies across regions.
The impact of ambient air pollution on property values has been examined with a variety of
property value data and assorted hedonic models.10 Recently the literature has focused on the
distributional impacts of property value appreciation, or on the attenuation of appreciation by labor
market impacts. For example, Bento, et al., (2015) examine the effect of the Clean Air Act
Amendments of 1990 on tract-level housing values near PM10 air monitors that are in a non-
attainment status and likely to be targeted for air quality improvements. They find significant
10 A broad review of the earlier hedonic property value literature examining ambient air quality is provided in meta-
analyses by Smith and Huang (1993; 1995).
4
-------
appreciation within 5 miles, and insignificant impacts on rents for the same distance. Agarwal, et
al. (2019) examine the effect of the NOx Budget Trading Program. House prices increased in areas
with low percentages of residents employed in manufacturing but decreased in areas with high
percentages, suggesting that negative impacts on labor markets counteract appreciation of housing
values. A variety of other studies find that air quality improvements increase housing values
(Amini, et al. 2021, Lang 2015, Grainger 2012, Bayer, et al. 2009, Chay and Greenstone 2005).
Many studies of the property value effects of ambient air quality have focused on criteria air
pollutants, but Currie, et al. (2015) explore the effect of toxic air pollutants. They examine housing
transactions near 1,600 industrial plant openings and closings for plants that report toxic air
emissions to the Toxic Release Inventory (TRI) in five U.S. states. The TRI requires firms to
report fugitive emissions including pollutant quantities that are accidentally emitted and those that
are routinely and intentionally emitted. They compare house prices before and after plant openings
and find that prices within 0.5 miles decrease by 11 percent after a plant opens. House prices do
not respond when a plant closes.
Research on how property values change in response to infrequent, accidental emissions events is
relatively uncommon, and mainly limited to a few geographically focused case studies. Flower
and Ragas (1994) studied house prices in Louisiana that were located near two petrochemical
refineries. They found that being close to one refinery with a large buffer zone was associated with
higher property values, except during a one-year period following a tank explosion. For the second
refinery, home prices declined with proximity. Carroll et al. (1996) studied property prices in
neighborhoods surrounding a chemical facility in Nevada before and after a 1988 explosion. They
found a 17.6% decline in property values in the two closest towns following the explosion. These
towns were located approximately 3.5 miles from the plant. An announcement regarding closure
of the facility was associated with a 38% rebound in home prices, suggesting that proximity to the
plant suppressed home values prior to the explosion.
Pipelines are not as visible to nearby residents as chemical facilities. For homeowners near
pipelines, post-accident property values may decline by more than they would if accident risks
were already capitalized into home values, as one would suspect for properties near more visible
RMP facilities. Cheng, et al. (2021) conduct a nationwide property value analysis of how accidents
at natural gas distribution pipelines impact nearby home values. Using a difference-in-difference
approach comparing pre- and post-incident transactions, their preliminary results suggest that the
value of homes within 1 km decline by 7.4% compared to a control group of houses l-2km from
an explosion. Liao, et al. (2022) examine properties near and distant from pipelines in a major city
in Taiwan before and after a catastrophic series of gas pipeline explosions, as well as soon after
the release of a map disclosing pipeline locations for the first-time. Following both treatments,
treated properties (within 1.3 km of any hazardous pipeline) were 2.6% lower in value than homes
farther away. Hansen et al. (2006) studied home sales located within a mile of two pipelines to
5
-------
estimate the impacts of a 1999 fuel pipeline explosion in Bellingham, Washington. Following that
event, property prices were significantly lower, with the effect diminishing over time and by
distance, from a 4.6% decline for a property 50 feet from the pipeline to 0.2% at 1,000 feet away.
Herrnstadt and Sweeney (2019) studied how property values responded to the combination of a
2010 pipeline explosion in San Bruno, California, and a subsequent mail notification to all
households living within 2000 feet of any natural gas pipeline. Using a difference-in-differences
approach comparing transactions within 2000 feet from a pipeline to those between 2000 and 4000
feet before and after the explosion and mailed notifications, they found no impact on prices. The
result is confirmed by a triple differences comparison against properties near pipelines managed
by other California utilities. Possibly, the information disclosure did not cause nearby homeowners
to revise their perceived risks; the observed risk of death from a pipeline explosion in California
is relatively low at only 0.053 deaths per million people annually (Hernstadt and Sweeney 2019).
Less dramatic releases at industrial and commercial facilities that result in lingering contamination
can also affect property values. Guignet et al. (2018) examine property value changes around high-
profile releases from underground storage tanks (USTs) at retail gas stations. Using difference-in-
differences hedonic equations and an internal meta-analysis, they find that homes within 3 km
depreciate an average of 6% after a release. Guignet and Nolte (2021) conduct a nationwide
property value study of hazardous waste treatment, storage, and disposal facilities (TSDFs) in need
of remediation. Although they do not necessarily interpret it as causal, they find evidence that
home values within 750 meters of a TSDF decrease up to 5% after the discovery of contamination
and initiation of a cleanup investigation. Both studies find causal evidence that home values
rebound by 6% to 7% after the contamination is cleaned up.
Chemical releases and accidents at RMP facilities differ from those studied by Guignet et al. (2018)
and Guignet and Nolte (2021) because RMP accidents involve fires, explosions, and releases of
toxic plumes into the air, which although sometimes severe, rarely result in lingering
environmental contamination. RMP accidents are distinct release events, which make them more
perceivable to surrounding residents. In cases involving an immediate threat, residents are notified
by first-responders, or indirectly informed by visual cues and local press. (Information about RMP
accidents is also made available to the public through reports to EPA.11) In contrast, releases of
hazardous substances from USTs and TSDFs often pollute the land, groundwater, or surface water.
The release events may happen years before they are discovered, and there is an element of
uncertainty to nearby property owners as to when and if a cleanup will occur. To our knowledge,
11 Accidents must be reported to EPA within 6 months of their occurrence. Once reported, the information is publicly
available though not easily accessible. Non-governmental organizations have occasionally obtained and republished
the accident information in a more publicly accessible format. For example, see OMB Watch, Press Release, October
29, 2009 at https://rirstameiMliirentcoalition.Org/2009/10/omb~watch~posts~details~o:f-risks~posed-bv-ii~s~chemieal~
facilities/ or Reynolds Journalism Institute. "Houston Chronicle. RJI partner to keep chemical company data site
available for journalists," June 12, 2019 at https://riionline.org/news/hoiiston~ehron.iete~rii~partiier~to~keep~ehemieat~
conipanv-data-site-available-for~ioiirnalists/.
6
-------
the current study is the first comprehensive, broader-scale non-market valuation study on accidents
at chemical facilities that result in fugitive air emissions, and the first specifically on the RMP
program.
III. DATA
The analysis focuses on home sales and RMP sites in Michigan, Ohio, and Pennsylvania from
2004 through 2014. Two datasets are combined to examine how RMP facilities and accidents
affect home values in the surrounding communities. The first dataset contains all RMP facilities
and chemical accidents covered by EPA's RMP program. The RMP facilities and accidents are
then spatially and temporally linked to transaction data of single-family homes obtained through
Zillow's ZTRAX program. Both datasets are discussed in detail below. Additional data were
obtained from a variety of sources to calculate distances between homes and local
amenities/dismenities, such as major roads, large lakes, and Superfund sites.
III. A. RMP Facilities and Accidents
Beginning with the initial RMP submissions in 1999, EPA maintains a national database of all
RMPs submitted to the Agency by the regulated facilities. All RMP submissions are required to
include a facility's five-year history of accidental releases from regulated processes that resulted
in deaths, injuries, significant property damage on site, or known offsite deaths, injuries,
evacuations, sheltering in place, property damage, or environmental damage.12 RMPs are
required to be updated and resubmitted at least every five years, so the RMP national database
contains a continuous record of accidents from regulated facilities spanning from mid-1994 (five
years prior to initial submissions) to the present.
Our analysis focuses on the 1,578 RMP facilities in Michigan, Ohio, and Pennsylvania (see Figure
1). These data were pulled from the national database by the EPA's Office of Emergency
Management (OEM) in June of 2018. The RMP facilities are involved in an array of different
industrial activities (as described in section I). Geographic coordinates are provided by the
facilities themselves, and most often refer to the facility centroid (42%) or plant entrance (16%).
In the remaining cases the coordinates correspond to some other point on the property (e.g., a
storage tank, administrative building, processing or treatment units, and so on.).
During our 2004-2014 study period, 209 accidents involving at least one RMP-regulated chemical
occurred at 124 of these facilities. Between 13 and 27 separate chemical accidents occurred in any
given year (see Figure A1 in the Appendix). Among the facilities where an accident occurred, most
had just a single accident (81), but some facilities had two (24), three (10), four (5), five (2), and
12 See 40 CFR 68.42.
7
-------
six accidents (1) during our study period. At the upper end, one facility in Schuylkill County, PA
reported 14 separate accidents.
Descriptive statistics for the 209 accidents are provided in Table 1. Most accidents (183 or 88%)
involved the release of just a single chemical, with the mean number of chemicals released being
1.2. However, at 26 (12.4%) of the accidents multiple chemicals were released. Sixty-four percent
of accidents involved a gas or vapor release, and 38% of the incidents involved a liquid spill and
evaporation of the chemical. Although not as common, chemical fires and explosions are observed
at 9% and 2% of the accidents, respectively. About 4% of the accidents involved an uncontrolled
chemical reaction. Environmental damage (e.g., fish or animal kills, defoliation of trees, and water
contamination) occurred in about 5% of the accidents. In about half the cases (48%) first
responders were notified and deployed.
Two accidents each resulted in one on-site mortality, and one accident led to the death of two
individuals. On-site injuries occurred at 79 (37.8%) of the accidents. Among the 40 accidents
where onsite property damage was reported, the average amount of damage was estimated at $3.4
million.
The typical accident did not directly impact offsite populations, but in a quarter of the accidents
(51 cases or 24.4%) there were direct impacts to individuals located offsite. No offsite mortalities
occurred, but offsite injuries requiring hospitalization or other medical treatments were
experienced at 19 of the accidents, with the number of people injured ranging from one to 25
individuals. Damage to offsite properties was reported in 8 cases, with the average amount of
damage valued at approximately $175,000. Among the 23 cases involving evacuations of nearby
communities, on average about 130 people were evacuated. Anywhere from 20 to over 4,500
people located offsite were ordered to shelter-in-place in 13 of the incidents.
III. 11 Residential Property Transactions
The initial analysis focuses on the 3,686,984 single-family home transactions that were within 25
km of at least one RMP site.13 The average home sold for a price of $159,217 (2019$), had a lot
size of 0.28 acres, was 1.5 stories, 52 years in age, and had 2.3 bedrooms, 1.7 bathrooms, and was
3,100 square feet in size. Values for lot acreage, number of stories, number of bathrooms, interior
square footage, and age was missing for about 12.9% to 16.7% of the sample, as indicated by the
companion missing value indicators. In our subsequent regression models the missing values are
coded as zero, and the corresponding missing indicator variables are included as independent
variables. Locational variables are included in the regression models to account for any within
13 Outlier transactions where the real sales price was less than $10,000 or greater than $1 million, where lot size was
greater than 1.3 acres (90th percentile), where the number of stories was greater than three (99th percentile), the interior
square footage was less than the lowest or greater than the highest percentiles (807 sqft. and 10,048 sqft, respectively),
the number of bedrooms was greater than seven, the number of bathrooms was greater than six, or when the census
tract identifiers were missing, are dropped. We also excluded likely non-arms-length sales where a token nominal
price of $1 was listed, or where different parcels had the exact same coordinates, as recommended by Nolte et al.
(2021).
8
-------
tract price variation due to proximity to local amenities and disamenities. Distance from each home
to the nearest primary road, Superfund site listed on the National Priorities List (NPL), and large
lake were all calculated within a Geographic Information System (GIS).
There are a substantial number of sales within each of the one-kilometer distance bins around an
RMP facility. This number steadily increases, starting with over 210k transactions within 0-1 km
of a site, all the way out to over 2.3 million transactions within 24-25 km of an RMP site (see
Figure A2 in the Appendix).
Narrowing the focus to transactions with an accident in close proximity, the numbers remain large.
Even within the nearest distance bin of 0-1 km, there are 9,657 and 9,289 home sales observed
before and after an accident respectively. Among the post-accident transactions, 2,944 are within
0-1 km of an accident that did involve offsite impacts, and 6,345 sales are within 0-1 km of an
accident that had no offsite impacts. The number of identifying transactions in each incremental
bin before and after an accident increases sharply as distance increases (see Figure A3 in the
Appendix). These transactions serve as the treatment and control groups in our quasi-experimental
framework.
IV. METHODS
IVA. The Spatial Difference-in-differences and Triple Differences Approach
The usefulness of the spatial DID approach for causal inference in hedonic pricing studies was
made clear through early applications by Davis (2004) and Linden and Rockoff (2008). The
approach remains at the forefront of hedonic pricing methods (Guignet and Lee, 2021), and has
been applied in numerous studies measuring environmental impacts on home values, including
analyses of contaminated site cleanups (Haninger et al., 2017; Guignet et al., 2018;), aquatic
vegetation and invasive species (Guignet et al., 2017; Horsch and Lewis, 2009), flood and
hurricane risks (Atreya et al., 2013; Bin and Landry, 2013), and unconventional shale gas
extraction (Muehlenbachs et al., 2015).
The DID strategy is intuitive in mimicking a classical experimental design. Consider the stylized
illustration in Figure 2. The top row represents homes in a community near an RMP facility where
a chemical accident occurs. The bottom row represents homes in a community near an RMP
facility where an accident does not occur. The left column denotes the time period before an
accident, and the right column denotes the time period after an accident. Each outer circle
represents homes farther from an accident than homes represented by each inner circle.
The price effect of interest is experienced by homes in group B. These are homes in close proximity
to an RMP facility after an accident occurs (i.e., the treated group, post-treatment). A simple cross-
sectional comparison of homes near an RMP site with an accident to those near a site without an
accident provides a first difference estimate of how the accident impacts home values, i.e., B-B'.
However, the RMP facilities and surrounding communities where accidents tend to occur or not
occur may be systematically different for unobserved reasons, which would confound such a first
difference estimate.
9
-------
Our cross-area DID estimate addresses this by exploiting temporal variation in when accidents
occur. This second differencing comes from comparing the change in home values before and after
the accident occurs (B-A), to the change in home values during this same time period around the
counterfactual RMP site (B'-A). Our cross-area DID estimates are represented as (B-A)-(B'-A').
Such estimates, however, could be biased if the trends in the neighborhoods where contamination
does and does not occur also differ systematically over time.
We can frame an alternative within-area DID estimate that uses homes in the same community,
but that are far enough away from the RMP site as to not be impacted by the accident, as the
counterfactual. In doing so, broader neighborhood-specific trends can be differenced out. Our
within-area DID estimate can be expressed as (B-A)-(D-C). This is the same as the conventional
spatial DID strategy implemented in preceding hedonic studies (e.g., Linden and Rockoff, 2008;
Haninger et al., 2017; Muehlenbachs et al., 2015). Nonetheless, such estimates could be biased if
the trends in homes value near versus far from an RMP facility are systematically different for
unobserved reasons.
Our triple differences estimate addresses that latter possibility by accounting for both layers of
controls and differencing out broader neighborhood trends and systematic differences between
homes that are near versus far from an RMP site. Our triple differences estimate in this stylized
illustration is represented as: {(B-A)-(B'-A)} - {(D-C)-(D'-C)}.
IV B. Hedonic Property Value Model
Our DID and triple differences approaches are couched within a hedonic regression framework,
where the dependent variable is the natural log of the price of home z, in neighborhood j, in housing
market m (i.e., county), at time t (Pijmt)• More formally:
(1) ln(Pijmt) — ijmtPmt
+ t{RMPijm > 0)p + RMPijm
-------
a specific location. £ijmt is a normally distributed disturbance term, which we allow to be
correlated for all transactions within the same county.
The parameters of direct interest are p, (p, ypre, yP"st and Qv"st, The vector RMPtjm denotes the
number of RMP sites within each one-kilometer incremental bin around a home (e.g., 0-lkm, 1-
2km, and so on), and > 0) is a vector of indicator variables denoting the presence of
at least one RMP facility in each one-kilometer bin. Estimates of p and
-------
direct offsite impacts and from an accident that does lead to offsite impacts are calculated,
respectively, as:
(which we later assume to be 7-10km, see section V.B for details).
In our more focused analysis of just homes within ten kilometers of an RMP accident, we drop all
home transactions that are only near an RMP site where no accident occurs. However, due to the
sometimes-clustered nature of RMP sites, we must still control for the presence and number of
nearby RMP facilities. Therefore, the within-area DID estimates are also calculated following
equations (4) and (5). Although the calculations are the same, dropping that initial control group
of homes near an RMP where no accident occurs degrades a triple differences interpretation.
IV C. Coarsened Exact Matching
To examine the robustness of our results, in later models we perform a pre-regression matching
exercise that prunes and re-weights observations so that the defined treatment and control groups
have more balanced distributions in terms of the observed attributes. More specifically, we use
coarsened exact matching (CEM) techniques (Blackwell et al., 2009; Iacus et al., 2012). The CEM
approach first divides the continuous attribute space for the relevant variables into discrete bins.
Then home transactions that fall within the same set of discrete attribute bins (i.e., have the same
value for all "coarsened" attributes) are matched. Homes are matched based on coarsened variables
of the number of bathrooms, age, interior square footage, lot acreage, and distance to a primary
road.14 In addition, we match homes in the treated and control groups only if they are in the same
county and sold in the same year. Transactions in the control group are dropped from the matched
sample if there are no transactions in the treated group that occurred in the same county and year,
and vice versa.
The CEM algorithm assigns weights to the treated and control group observations. A weight of
one is given to all treated home sales that could be matched to at least one control home sale. A
treated observation can potentially be matched to more than one control observation, and one
control observation can be matched to more than one treated observation, and so all maintained
control home sales are given a positive weight that may be less than, equal to, or greater than one.
Despite its advantages in flexibly matching observations to make the treated and control groups
14 Three coarsened attribute values are defined for each variable, where the cutoffs are roughly based on the 25th and
75th percentiles for each attribute. The coarsened acres variable is divided into four values, using the 25th, 50th, and
75th percentiles as cutoffs. An additional "missing" category is included for each of the coarsened variables, when
applicable.
12
-------
simultaneously more comparable across numerous dimensions, the CEM procedure has been used
in only a few other hedonic property value studies (Groves and Rogers, 2011; Guignet et al., 2018;
Guignet and Nolte, 2021; Qiu et al., 2017).
V. RESULTS
V.A. Proximity to RMP Sites
Before examining the impacts of a chemical accident on house prices, we first wish to understand
the baseline association between house prices and proximity to an RMP site, irrespective of an
accident. To do so, we estimate the percent change in price as a function of distance to the site
using the results from Model 1 (shown in Table Al in the Appendix). Following the estimates
from Equation (1), the percent change in price associated with one, two, or three RMP sites within
distance bin d of a home is calculated as:
(6) %A pfjf = {exp (p[d] + 0[d] x RMP[d]j) - l} x 100
Figure 4 illustrates the results and suggests that homes within 0-1 km of a single RMP site tend to
be 4.4% lower in value, all else constant. This negative association is relative to homes 24-25 km
from a single site and extends to two kilometers, with a 1.9% depreciation associated with being
within 1-2 km of a single site. Although we do not claim this as a causal estimate, especially since
RMP facilities may be located in areas that host multiple industrial uses, it does suggest that homes
nearest an RMP site tend to be lower in value due to disamenities associated with the site itself
and/or due to the surrounding neighborhood. After two kilometers we see statistically insignificant
negative, or even small positive price effects associated with proximity to an RMP site, ranging
from -0.4% to 1.1%.
Figure 3 also demonstrates strong cumulative effects associated with proximity to RMP sites.
Being within 0-1 km and 1-2 km of two RMP sites is associated with an 8.7% and 3.7% decline
in home values, all else constant. And homes within 0-1 km and 1-2 km of three RMP sites tend
to be 12.8%) and 5.4% lower in value, respectively. Again, we find no statistically significant
evidence of a negative association between home prices and proximity to RMP sites beyond two
kilometers. However, these negative associations should be considered conservative because the
estimates are conditional on census tract-level fixed effects.
All else constant, we find that homes nearest RMP facilities tend to be significantly lower in value,
especially in neighborhoods where multiple facilities are present. This is an important finding from
an environmental justice standpoint. Any adverse impacts from a chemical accident could further
depress the value of this critical financial asset; thus, further reducing the wealth of resident home
owners and diminishing landlords' incentives to maintain or improve the housing stock occupied
by renters.
13
-------
VB. Determining the Treated and Control Groups
As with any DID application, appropriately determining the treated and control groups is of the
utmost importance. Such is particularly true in the case of spatial DID applications, where there is
no clear assignment into the treated and control groups. Rather, researchers rely on data diagnostics
to determine what observations should fall into the treated or control group categories based on
spatial location (usually distance) to the commodity of interest. In such settings, researchers have
often relied on an approach by Linden and Rockoff (2008), and later adapted by Haninger et al.
(2017), Muehlenbachs et al. (2015), and others. We follow a similar approach by first estimating
a regression based on equation (1), where we define proximity to an RMP facility using one-
kilometer incremental bins, starting with 0-lkm, l-2km, and going all the way out to 24-25km.
The estimated coefficients are then used to flexibly graph the pre- and post-accident price
gradients. Expecting the price gradients to be most different nearest the site, we assume the
distance at which they converge to be the cutoff between the spatially defined treated and control
groups.
For typical accidents, estimates of equation (1) provided no clear evidence of where the spatially
defined thresholds between the control and treated group should be. However, accidents that result
in offsite impacts are where residents are better informed. For these accidents, the estimated pre-
and post-accident price gradients are displayed in Figure 4 (see Model 1 in Table Al of the
Appendix for the full results). All else constant, at closer distances we see that prices are
significantly lower after an accident that had offsite impacts. This decrement diminishes with
distance but based on the point estimates appears to possibly extend out to about seven kilometers.
Therefore, for the main analysis we assume the treated group of homes is within seven kilometers
or less of an RMP facility, and examine the price effects of an accident within this distance. As
can be seen in the second panel in Figure 3, for all but one of the bins within five kilometers a
Wald test rejects the null hypothesis that the pre-accident and post-offsite-accident distance
gradients are equal, at the p<0.10 level. Subsequent regression analyses focus on a treated group
within five kilometers, and disregard sales within 5-7 km due to concerns of potential treatment
effect spillover between the assumed treated and control group distance bins.
We limit the control group to homes within 7-10 km of an RMP facility. Figure 4 suggests no
systematic difference in prices before or after an accident at these farther distances and the p-values
for a Wald test of the null hypotheses that the pre- and post-accident gradients are equal in each
bin are all p>0.40. In addition, limiting the sample to transactions within 10km draws focus to the
neighborhoods nearest to industrial areas, where homes in the treated and control group distance
bins are likely more similar in terms of unobserved characteristics.
V C. Impacts of Chemical Accidents
We estimate a series of hedonic price regressions following equation (1). Each of five models are
explained in detail below with full regression results appearing in Table Al in the Appendix. As
per equations (2) through (5), we use the estimated coefficients to derive estimates of the percent
change in price due to an accident. All regressions include indicator and count variables denoting
14
-------
the presence of RMP facilities in general, at different distances. To account for differences in the
hedonic price surface across markets and time, the regressions include interaction terms between
each county-by-year combination and all house and location attributes. County-specific year and
quarter indicator variables, and census tract-level fixed effects are also included. Only estimates
of the percent change in home prices due to an accident are presented and discussed here.
The estimated percent changes in price due to an accident in Table 3 are based on Model 2, which
is estimated using the sample of almost 3.4 million transactions within 10km of an RMP site. This
model allows the price effects of an accident to differ for homes in each one-kilometer incremental
bin, from 0-1 km out through 6-7 km. The cross-area DID and triple difference estimates for a
chemical accident involving no offsite impacts are generally small and statistically insignificant.
In the 2-3km bin we even find a counterintuitive statistically significant 2.7% to 4.4% increase in
prices. More intuitive results are obtained when focusing on the price effects of accidents that
had direct offsite impacts to the surrounding community (i.e., injuries to offsite populations, offsite
property damage, or evacuation or shelter-in-place orders). The cross-area DID estimates suggest
statistically significant price impacts in each one-kilometer bin within five kilometers, ranging
from a 6.2% to 9.7% depreciation. The cross-area DID estimates even suggest a significant 4.3%
and 3.9% price decline in the 5-6 km and 6-7 km bins.
Those farther out price effects in the 5-6 km and 6-7 km bins are no longer significant for the triple
differences estimates, which better account for broader neighborhood trends. However, the
estimated price declines within five kilometers of an accident are fairly robust in the triple
differences estimates, ranging from a 4.3% to 7.8% deprecation and being statistically significant
in 3 of the 5 bins.
Given the results from our earlier distance gradient diagnostics (see Figure 4) and that we find
marginally significant results extending out to five kilometers here, in subsequent models we pool
the distance bins together and estimate a single average price effect for all homes within five
kilometers of a chemical accident. Model 3 is the same as Model 2, and is estimated using the
same sample of transactions, but pools the treatment bins into a single 0-5 km bin. The estimates
of the percent change in home prices due to an accident are presented in Table 4. Again, the cross-
area DID and triple difference estimates suggest that an accident involving no offsite impacts has
no effect on house prices in the surrounding community. The price effects from an accident that
resulted in offsite impacts remains robust, with the cross-area DID estimate suggesting a 6.7%
average decrease in price to homes within 5km. The more thorough triple differences estimate
suggests a marginally significant 4.8% price decline.15
The alternative within-area DID estimates are provided by Models 4 and 5 and are presented in
Table 5. Model 5 is estimated using the CEM weights and sample, which only includes transactions
matched by county, year, and observed (coarsened) house attributes (see section IV.C). These
models are estimated using a sample of home transactions that are within 0-5km (the treated group)
or 7-10km (the control group) of an RMP site where an accident occurs during the study period.
15 The percent change in price estimates discussed next are robust to an alternative 0-4km treatment bin, suggesting
an average price decrease of 5.3% to 7.4%.
15
-------
Sales that are only near RMP sites where no accidents occur, or are only within 5-7km of an RMP
accident site, are excluded.16 The CEM algorithm matches similar home sales in the treated (0-
5km) and control (7-10km) groups. Home transactions that are not in either group are given a zero
weight and excluded from the estimating sample in Model 5. We use the same, but unmatched,
sample in Model 4 to cleanly demonstrate the robustness of our results from this matching
procedure.
The results of models 4 and 5 again demonstrate that, on average, an accident involving no offsite
impacts does not affect surrounding home values, whereas an accident with offsite impacts does.
The estimated price effects from accidents with offsite impacts suggest an average 5.4% or 5.7%
depreciation in homes values within five kilometers of an accident.
VD. Assessing the Parallel Trend Assumption
A causal interpretation of our results hinges on the parallel trends assumption. In a well-defined
DID quasi-experiment the trajectory of the outcome variable experienced by the treated group in
the absence of treatment must be the same as that of the assumed control group in the post-
treatment period (Angrist and Pischke, 2009). We do not observe the true counterfactual (i.e., the
treated group absent the treatment), but we do observe the pretreatment trends. If house prices for
the treated and control groups follow similar trends before the occurrence of a chemical accident,
then it is more reasonable to assume those trajectories would have remained similar in the absence
of this treatment event.
To assess whether we can interpret the estimated 5% to 7% decline that is experienced among
homes within five kilometers of an "offsite" accident as causal, we conduct an event study (Hanna
and Olivia, 2010) by estimating a variant of equation (1) and Model 4 that includes annual lead
terms prejyms to denote transactions that occurred 5 years prior to an accident. Similarly, we
include the lag terms postijms and offijms to denote transactions that occur 5 years after any
accident, or after an accident yielding offsite impacts, respectively. The hedonic regression
estimated for this event study is:
(7) ln(Pijmt) — X-ijmtPmt ^ ^)P RMPijm(P + Hs=-14{P^^ijmsYs }
+ Zlio{postiJmsYPsOSt + (postijms x
Offtjms^s } 1-mt "I" Vjm + ^ijmt
The time component of the lead and lag variables is normalized so that 5=0 corresponds to the
exact day an accident occurred; denoted by the black vertical line in in Figure 5. The estimated
lead terms for the 0-5km treated group and 7-10km control group in yvsre are graphed on the left
side of Figure 5. Focusing on accidents resulting in offsite impacts, the lag terms yvs°st + Qvsost
are graphed to the right.
16 Because homes can be near multiple RMP sites, the corresponding RMP site and accident distance bins are still
included as right-hand side variables in the regression models.
16
-------
The pre-accident coefficients are generally very similar and follow a similar trend, suggesting that
the parallel trends assumption is likely valid. There is some divergence in the pre-trends very early
in the study period, but we attribute this largely to the small number of transactions observed more
than 10 years before an accident. It is mainly after an accident with offsite impacts occurs that we
see the price trends between the treated and control groups diverge. Home values within 0-5 km
begin to decline compared to homes 7-10 km away. This decrement seems to be largest around 5
to 6 years after an accident, and then prices may start to converge. Confirmation of the parallel
trends assumption supports a causal interpretation of the 5% to 7% depreciation among homes
within five kilometers of an accident that impacts offsite populations.
VI. CONCLUSION
Accidental chemical releases at industrial facilities can result in toxic clouds, fires, and explosions.
These accidents have negative impacts both onsite and off, faced by workers and nearby
communities. The U.S. EPA's RMP program is in place to reduce the harm from such events to
nearby communities. The program is meant to reduce the risk of chemical accidents and, should
one occur, ensure that protocols are in place to minimize damages. The RMP program is truly far-
reaching, with almost 12,000 U.S. facilities currently regulated under the program. At least 40
million people (or about 12% of the U.S. population), and perhaps as many as 177 million (55%),
are at risk of experiencing impacts from an accidental chemical release, should one occur (U.S.
EPA, 2016).
This study examines how chemical accidents at RMP facilities impact the welfare of residents in
surrounding communities using hedonic property value methods and an extensive dataset of 3.6
million single-family home transactions in Michigan, Ohio, and Pennsylvania. Our hedonic
regression models use a variety of difference-in-differences and triple differences identification
strategies, in conjunction with spatial and temporal fixed effects and coarsened exact matching
techniques, to estimate the causal impacts of chemical accidents on nearby home values. In
general, we have more confidence in the triple differences and within-area DID estimates,
compared to the cross-area DID estimates. First, the within-area DID and triple differences
estimates better account for local price trends that could otherwise confound the results. Second,
the cross-area DID approach is a staggered DID design, and recent literature has raised concerns
over the interpretation and appropriateness of DID settings where the treatment event is staggered
over time (Goodman-Bacon, 2021; Marcus and Sant'Anna, 2021; Sun and Abraham, 2021). Our
within-area DID and triple differences estimates are akin to a stacked DID design (Cengiz et al.,
2019; Deshpande and Li, 2019; Fadlon and Nielsen, 2021), which alleviates concerns regarding
staggered treatment events (Goodman-Bacon, 2021, Roth et al., 2022). Still, the cross-area DID
estimates instill confidence in our primary findings as they yield remarkably similar results, despite
relying on an entirely different set of counterfactuals.
The results indicate that the typical RMP accident, where adverse impacts are contained on-site,
does not impact surrounding home values. However, for the roughly 25% of incidents in our data
17
-------
in which the chemical accident resulted in injuries to offsite populations, offsite property damage,
or necessitated the evacuation or shelter-in-place of individuals located offsite, we find an average
5% to 7% decline in the value of homes within five kilometers of the accident.
This finding is not unexpected as households are likely to be well-informed about accidents with
offsite impacts. Such accidents overtly impose real risks and costs to nearby populations. Whereas
workers may be compensated thorough wage rates for facing onsite risks, causing chemical firms
to internalize at least some of the costs to employees of accidents, the costs of offsite damages are
external and borne by nearby homeowners and residents. Our findings illustrate that these external
costs are substantial, corresponding to an average loss in home value of $12,000 to $17,000 (see
Table 6 for details).
The environmental justice implications are particularly important. We find that proximity to an
RMP facility, even in the absence of any accidents, is associated with lower home values. Homes
within one kilometer of a single facility tend to be worth over 4% less, all else constant; and homes
within l-2km experience an almost 2% discount. This adverse relationship is particularly
noticeable in the presence of multiple facilities. For example, homes within a kilometer of three
RMP facilities are more than 12% lower in value, and those within l-2km are 5% lower. Additional
depreciation due to a chemical accident at a nearby facility will exacerbate these baseline
inequities. Elliot et al. (2004) find that risk factors associated with RMP accidents (e.g., number
of chemicals and employees), and facilities that have more than one accident, are more likely in
counties where a greater proportion of the population is Black, and where there are higher levels
of income inequality.
The RMP program serves a critical role in minimizing such inequities, and in reducing the risk to
populations living in the "fence-line" communities around industrial facilities more generally. The
increased occurrence of extreme storms, heat waves, floods, wildfires and other natural disasters
due to climate change, makes industrial facilities more vulnerable to experience accidental
chemical releases, which emphasizes the importance of the RMP program (Flores, et al. 2021; US
Chemical Safety Board, 2017; Chemical Industries Association, 2015). The U.S. has a long-
standing, proud industrial tradition, but at the same time it is important to consider the magnitude
of the negative externalities imposed by these facilities. This study provides significant insight into
the monetized external costs of these industrial chemical accidents.
18
-------
WORKS CITED
Agarwal, Sumit, Yongheng Deng, and Teng Li. 2019. Environmental regulation as a double-edged
sword for housing markets: Evidence from the NOx Budget Trading Program. Journal of
Environmental Economics and Management 96: 286-309.
Amini, Amirhossein, Kaveh Nafari, and Ruchi Singh. 2021. Effect of air pollution on house prices:
Evidence from sanctions on Iran. Regional Science & Urban Economics. Forthcoming.
Angrist, Joshua D., and Jorn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricist's Companion. Princeton: Princeton University Press.
Atreya, Ajita, Susana Ferreira, and Warren Kriesel. 2013. Forgetting the Flood? An Analysis of
the Flood Risk Discount over Time. Land Economics 89(4): 577-596.
Bayer, Patrick. Nathaniel Keohane, Christopher Timmins. 2009. Migration and hedonic valuation:
The case of air quality. Journal of Environmental Economics and Management. 58: 1-14.
Bento, Antonio, Matthew Freedman, and Corey Lang. 2015. Who Benefits from Environmental
Regulation? Evidence from the Clean Air Act Amendments. The Review of Economics and
Statistics 97(3): 610-622.
Blackwell, Matthew, Stefano Iacus, Gary King, and Guiseppe Porro. 2009. Cem: Coarsened exact
matching in Stata. The Stata Journal 9(4): 524-546.
Bin, Okmyung, and Craig Landry. 2013. Changes in Implicit Flood Risk Premiums: Empirical
Evidence from the Housing Market. Journal of Environmental Economics and Management 65:
361-376.
Boyle, M. A. and K. A. Kiel. 2001. A Survey of House Price Hedonic Studies of the Impact of
Environmental Externalities. Journal of Real Estate Literature 9(2): 117-144.
Carroll, T.M., Clauretie, T.M., Jensen, J. and Waddoups, M. September 1996. The Economic
Impact of a Transient Hazard on Property Values: The 1988 PEPCON Explosion in Henderson,
Nevada. The Journal of Real Estate Finance and Economics, Volume 13, Issue 2, pp 143-167.
Cassidy, Alecia, Elaine L. Hill, and Lala Ma. 2020. Housing Impacts of Hazardous Waste
Cleanup. SSRN Working Paper 3766930. Available at:
https://privpapers.ssm.com/sol3/papers.cfm7abstract id 530.
Chay, Kenneth Y., Michael Greenstone. 2005. Does Air Quality Matter? Evidence from the
Housing Market. Journal of Political Economy. 113(2): 376-424.
Chemical Industries Assocation, 2015. "Safeguarding chemical businesses in a changing climate:
How to prepare a Climate Change Adaptation Plan." March. Available at:
https://www.cia.org.uk/LinkClick.aspx?fileticket=KW8WF8CBZG0%3D&portalid=0
19
-------
Cheng, N., M. Li, P. Liu, Q. Luo, C. Tang, and W. Zhang. 2021. Pipeline and Property Values: A
Nationwide Hedonic Analysis of Pipeline Accidents Over the Past Three Decades. Poster
presentation at 2021 Annual AAEA meeting. Available at:
https://ideas.repec.Org/p/ags/aaea21/312676.html.
Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer. 2019. The Effect o Minimum
Wages on Low-Wage Jobs. The Quarterly Journal of Economics 134(3): 1405-1454.
Currie, Janet, Lucas Davie, Michael Greenstone, and Reed Walker. 2015. Environmental Health
Risks and Housing Values: Evidence from 1,600 Toxic Plant Openings and Closings. American
Economic Review. 105(2): 678-709.
Davis, Lucas W. 2004. The Effect of Health Risk on Housing Values: Evidence from a Cancer
Cluster. American Economic Review. 94(5): 1693-1704.
Deshpande, Manasi, and Yue Li. 2019. Who is Screened Out? Application Costs and the
Targeting of Disability Programs. American Economic Journal: Economic Policy 11(4): 213-
248.
Elliot, M.R., Kleindorfer, P.R., and Lowe, R.A. 2003. The Role of Hazardousness and
Regulatory Practice in the Accidental Release of Chemicals at U.S. Industrial Facilities. Risk
Analysis. 23(5): 883-896.
Elliot, M.R., Wang, Y., Lowe, R.A., and Kleindorfer, P.R. 2004. Environmental Justice:
Frequency and Severity of U.S. Chemical Industry Accidents and the Socioeconomic Status of
Surrounding Communities. Journal of Epidemiology and Community Health. 58: 24-30.
Fadlon, Itzik, and Torben H. Nielsen. 2021. Family Labor Supply Responses to Severe Health
Shocks: Evidence from Danish Administrative Records. American Economic Journal: Applied
Economics 13(3): 1-30.
Flores, David, Casey Kalman, Michelle Mabson, and Darya Minovi. 2021. Preventing "Double
Disasters": How the U.S. Environmental Protection Agency can protect the public from
hazardous chemical releases worsened by natural disasters. Center for Progressive Reform,
Earthjustice, and the Union of Concerned Scientists. July.
https://www.ucsusa.org/sites/default/files/2021-07/preventing-double-disasters%20FINAL.pdf
Flower, Patrick and Wade Ragas. 1994. The Effects of Refineries on Neighborhood Property
Values. Journal of Real Estate Research. 9: 319-338.
Gamper-Rabindran, Shanti, and Christopher Timmins. 2013. Does cleanup of hazardous waste
sites raise housing values? Evidence of spatially localized benefits. Journal of Environmental
Economics and Management, 65(3): 345-360.
Goodman-Bacon, Andrew. 2021. Difference-in-Differences with Variation in Treatment Timing.
Journal of Econometrics 225(2): 254-277.
20
-------
Grainger, Corbett A. 2012. The distributional effects of pollution regulations: Do renters fully
pay for cleaner air? Journal of Public Economics. 96: 840-852.
Grislain-Letremy, C. and A. Katossky 2014. The impact of hazardous industrial facilities on
housing prices: A comparison of parametric and semiparametric hedonic price models. Regional
Science and Urban Economics 49: 93-107
Guignet, Dennis, Charles Griffiths, Heather Klemick, and Patrick J. Walsh. 2017. The Implicit
Price of Aquatic Grasses. Marine Resource Economics 32(1): 21-41.
Guignet, Dennis, Robin Jenkins, Matthew Ranson, and Patrick J. Walsh. 2018. Contamination
and Incomplete Information: Bounding Implicit Prices using High-Profile Leaks. Journal of
Environmental Economics and Management 88: 259-282.
Guignet, Dennis, and Jonathan Lee. 2021. State of the Art of Hedonic Pricing. In Oxford Research
Encyclopedia of Environmental Science, ed. James R. Kahn. Oxford University Press,
doi: https://doi.org/10.1093/acrefore/978019938
Guignet, Dennis, and Christoph Nolte. 2021. Hazardous Waste and Home Values: An Analysis
of Treatment and Disposal Sites in the U.S. Appalachian State University Department of
Economics Working Paper, 21-07, Boone, NC, July.
Haninger, Kevin, Lala Ma, and Chrisopher Timmins. 2017. The Value of Brownfield Remediation.
Journal of the Association of Environmental and Resource Economists 4{ 1): 197-241.
Hanna, Rema Nadeem, and Paulina Olivia. 2010. The Impact of Inspections on Plant-Level Air
Emissions. The B.E. Journal of Economic Analysis and Policy 10(1): 1-29.
Hansen, J. L., E. D. Benson and D. A. Hagen. 2006. Environmental Hazards and Residential
Property Values: Evidence from a Major Pipeline Event. Land Economics 82(4): 529-541
Herrnstadt, Evan and Richard L. Sweeney. 2019. Housing Market Capitalization of Pipeline
Risk: Evidence from a Shock to Salience and Awareness. Unpublished Manuscript. Accessed at
http://www.richard-sweeney.com/pdfs/HS Pipelines.pdf. 4 Sept 2021.
Horsch, Eric J., and David J. Lewis. 2009. The Effects of Aquatic Invasive Species on Property
Values: Evidence from a Quasi-Experiment. Land Economics 85(3): 391-409.
Iacus, Stefano, Gary King, and Giuseppe Porro. 2012. Causal Inference without Balance Checking:
Coarsened Exact Matching. Political Analysis 20: 1-24.
Kleindorfer, P.R., Belke, J.C., Elliot, M.R., Lee, K., Lowe, R.A., and Feldman, H.I. 2003.
Accident Epidemiology and the U.S. Chemical Industry: Accident History and Worst-Case Data
from RMP*Info. Risk Analysis. 23(5): 865-881.
21
-------
Kleindorfer, P.R., Elliot, M.R., Wang, Y., and Lowe, R.A. 2004. Drivers of Accident
Preparedness and Safety: Evidence from the RMP Rule. Journal ofHazardous Materials. 115: 9-
16.
Lang, Corey. 2015. The dynamics of house price responsiveness and locational sorting: Evidence
from air quality changes. Regional Science and Urban Economics. 52: 71-82.
Liao, Wen-Chi, Yilan Luo, and Yajie Sun. 2022. Information shock of disaster and hazard:
Impact of Kaohsiung gas explosions and risk disclosure on the equalizing difference in the
housing market. Real Estate Economics. 1-40.
Linden, Leigh, and Jonah E. Rockoff 2008. Estimates of the Impact of Crime Risk on Property
Values from Megan's Laws. The American Economic Review 98(3): 1103-1127.
Marcus, Michelle, and Pedro H.C. Sant'Anna. 2021. The Role of Parallel Trends in Event Study
Settings: An Application to Environmental Economics. Journal of the Association of
Environmental and Resource Economists 8(2): 235-275.
Muehlenbachs, Lucija, Elisheba Spiller, and Christopher Timmins. 2015. The Housing Market
Impacts of Shale Gas Development. American Economic Review 705(12): 3633-3659.
Nolte, Christoph, Kevin J. Boyle, Anita M. Chaudhry, Christopher M. Clapp, Dennis Guignet,
Hannah Hennighausen, Ido Kushner, Yanjun Liao, Saleh Mamun, Adam Pollack, Jesse
Richardson, Shelby Sundquist, Kristen Swedberg, and Johannes H. Uhl. 2021. Studying the
Impacts of Environmental Amenities and Hazards with Nationwide Property Data: Best Data
Practices for Interpretable and Reproducible Analyses. SSRN Working Paper: Available at
SSRN: https://ssrn.com/abstract=3900806.
Roth, Jonathan, Pedro H.C. Sant'Anna, Alyssa Bilinski, and John Poe. 2022. What's Trending in
Difference-in-Differences? A Synthesis of the Recent Econometrics Literature. Unpublished
manuscript. Available at: https://ionathandroth.github.io/assets/files/DiD Review Paper.pdf.
Smith, V. Kerry, and Ju-Chin Huang. 1993. Hedonic Models and Air Pollution: Twenty-Five
Years and Counting. Environmental and Resource Economics 3: 209-227.
Smith, V. Kerry, and Ju-Chin Huang. 1995. Can Markets Value Air Quality? A Meta-Analysis of
Hedonic Property Value Models. Journal of Political Economy 103(1): 209-227.
Sun, Liyang, and Sarah Abraham. 2021. Estimating Dynamic Treatment Effects in Event Studies
with Heterogeneous Treatment Effects. Journal of Econometrics 225(2): 175-199.
U.S. Census Bureau. 2016. Census Bureau Projects U.S. and World Populations on New Year's
Day. Accessed at: https://www.census.gov/newsroom/press-releases/2016/cbl6-tpsl58.html, 9
Sept. 2021.
22
-------
U.S. Chemical Safety and Hazard Investigation Board. 2018. Organic Peroxide Decomposition,
Release, and Fire at Arkema Crosby Following Hurricane Harvey Flooding. Report Number:
2017-08-1-TX, May.
U.S. EPA (Environmental Protection Agency). 1996. Strategy for Coordinated EPA/OSHA
Implementation of the Chemical Accident Prevention Requirements of the Clean Air Act
Amendments of 1990). https://www.epa.gov/sites/default/files/2013-10/documents/oshaepa.pdf
U.S. EPA (Environmental Protection Agency). 2016. Regulatory Impact Analysis: Accidental
Release Prevention Requirements: Risk Management Programs Under the Clean Air Act, Section
112(r)(7). Office of Emergency Management, Washington, D.C. December 16.
U.S. EPA (Environmental Protection Agency). 2019. Regulatory Impact Analysis:
Reconsideration of the 2017 Amendments to the Accidental Release Prevention Requirements:
Risk Management Programs Under the Clean Air Act, Section 112(r)(7). Office of Emergency
Management, Washing, D.C. November 18.
23
-------
FIGURES AND TABLES
Figure 1. Tri-state Study Area, RMP Facilities, and Chemical Accidents.
si
sutors.
minify
Esri, HERE, Garmiri. (c) OpenStreetMap contributors, and the
GIS user community
RMP Facility (no accidents)
Accident (with no offsite impacts)
Accident (with offsite impacts)
2,500 5,000 10,000
Kilometers
Note: Map of tri-state study area includes Michigan, Ohio, and Pennsylvania, U.S.
24
-------
Figure 2. Conceptual Illustration of Difference-in-Differences and Triple Differences
Approaches.
Pre-accident Post-accident
Site with
Accident
Site with no
Accident
Note: As discussed in section IV A, based on the above illustration our cross-area DID estimates are represented as
U-'B'-. Ithe within-area DID estimates are represented as (B~A)-(D-Cp, and the triple differences estimates are
represented as: {(B-A)-(B'-A)} - {(D-C)-(D'-C')}.
IRMPSite 2RMPSites 3 RMP Sites
25
-------
Figure 4. Pre-Accident and Post-Offsite-Accident Price Gradients (top panel); and P-values
testing if pre- and post-offsite-accident coefficients are equal (bottom panel).
Note: Top panel shows the pre- and post-accident price gradient, where the latter pertains to accidents with offsite
impacts. Coefficient estimates are based on model 1 in Table A1 of the Appendix. The 95% confidence intervals are
shown by the dotted lines, with standard errors calculated via the delta method using the "nlcom" command in Stata
17/MP. The lower panel shows the p-values from a Wald test of the null hypothesis that the pre- and post-offsite
accident points are equal at distance d. For points below the p=0.10 line we reject the null that the gradients are
equal at distance d.
26
-------
Figure 5. Event Study Graph of Price Differentials by Year.
0.15
0.1
0.05
0
-0.05
-0.1
-0.15
-0.2
-0.25
i^ViV
I H (,
-15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11
Year (Date of Accident = 0)
• Treated Group (0-5km) ¦ Control Group (7-10km)
Note: Figure shows estimates based on a variant of Model 4 that includes interaction terms to allow ypre, yPosf, and 0Post
coefficients to vary by year, relative to the date of the accident. Pre-accident points show estimates of ypre, and post-accident
points pertain to accidents with offsite impacts, and thus show estimates of yPost + QV0St
27
-------
Table 1. Descriptive Statistics, RMP Accidents from 2004-2014 in MI, OH and PA
Variable
Obs Mean
Std. dev. Min
Max
# Chemicals Released
209
1.20
0.82
1
10
Gas release (dummy)
209
0.6411
0.4808
0
1
Spill/evaporation (dummy)
209
0.3828
0.4872
0
1
Fire (dummy)
209
0.0861
0.2812
0
1
Explosion (dummy)
209
0.0191
0.1373
0
1
Reactive Incident (dummy)
209
0.0431
0.2035
0
1
Environmental Damage (dummy)
209
0.0478
0.2140
0
1
First Responders Deployed (dummy)
209
0.4785
0.5007
0
1
Onsite Deaths (# of people)t
3
1.33
0.58
1
2
Onsite Injuries (# of people) t
79
1.76
1.59
1
9
Onsite Property Damage (nominal USD)
40
3,432,218
12,700,000
100
75,000,000
Offsitc Deaths (# of people)
0
-
-
-
-
Offsite Injuries (# of people)
19
3.68
6.09
1
25
Offsitc Environmental Damage (nominal
USD)t
8
175,441
258,824
250
750,000
Offsite Evacuated (# of people)t
23
130.13
231.92
1
735
Offsite Shelter-in-Place (# of people)
13
501.23
1281.36
20
4,677
Notes: There were a total of 209 chemical accidents from 2004-2014, at 124 different RMP facilities,
¦f Descriptive statistics here reported only for accidents where the respective variable has a nonzero value.
28
-------
Table 2. Single-Family Home Transactions Descriptive Statistics.
Variable
Obs
Mean
Std. Dev.
Min
Max
Price (2019$)
3,686,984
159,217
132,272
10,001
999,815
Acres
3,212,836
0.2779
0.2433
2.300E-05
1.3000
Missing: Acrest
3,686,984
0.1286
0.3348
0
1
Stories
3,071,319
1.53
0.51
0.75
3.00
Missing: Stories
3,686,984
0.1670
0.3730
0
1
Bedrooms
3,686,984
2.30
1.52
0
7
Bathrooms
3,117,285
1.67
0.75
0.5
6
Missing: Bathroomst
3,686,984
0.1545
0.3614
0
1
Interior square footage
3,164,277
3,107
1,701
807
10,048
Missing: Interior square footaget
3,686,984
0.1418
0.3488
0
1
Age (years)
3,191,921
52.23
31.30
0
232
Missing: Aget
3,686,984
0.1343
0.3409
0
1
Distance to primary road (km)
3,686,984
5.22
6.56
0.00
154.93
Distance to NPL site (km)
3,686,984
13.81
11.94
0.01
104.73
Distance to large lake (km)
3,686,984
11.46
7.89
0.00
59.00
Note: The final sample includes n=3,686,984 single-family home transactions. Descriptive statistics for some
variables are for a smaller sample due to missing values, as reflected by the corresponding missing value
indicators. Variables denoted with f are binary indicators.
-------
Table 3. Percent Change in Price from a Chemical Accident: One-kilometer Treatment Bins.
Model 2
All w/in 10km of RMP Site
Cross-area DID
Triple Diff
Cross-area DID
Triple Diff
No Offsite
No Offsite
Offsite
Offsite
0 to 1 km
0.8947
2.5222
-8.0546**
-6.1371
(1.7382)
(2.2555)
(3.8954)
(4.0178)
1 to 2 km
-0.3919
1.2149
-9 7147***
-7.8319**
(1.7256)
(2.3497)
(2.9725)
(3.1385)
2 to 3 km
2.7383**
4.3955**
-6.4483*
-4.4972
(1.2743)
(1.7760)
(3.7189)
(4.1000)
3 to 4 km
0.3710
1.9901
-8.0571***
-6.1396**
(1.5244)
(2.0302)
(2.7377)
(2.8822)
4 to 5 km
-1.1755
0.4186
-6.2371**
-4.2816*
(1.0383)
(1.2916)
(2.4991)
(2.4100)
5 to 6 km
-0.7101
0.8915
-4.3052**
-2.3095
(1.1612)
(1.4912)
(2.0302)
(1.9344)
6 to 7 km
-0.7859
0.8145
-3.9428*
-1.9395
(1.0568)
(0.7657)
(2.1289)
(1.6580)
House and Location Attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Observations
County
County
County x
x Year
x Year
Quarter
Yes
3,359,856
Note: Dependent variable is ln(price). *p<0.10, **p<0.05, ***p<0.01. Standard errors in parentheses, clustered at the county
level. Estimates calculated following equations (2)-(5) using the "nlcom" command in Stata 17/MP, and are based on the
coefficient estimates from the hedonic regression model 2. Full regression coefficient results presented in Table A1 in Appendix.
30
-------
Table 4. Percent Change in Price from a Chemical Accident: 0-5km Treatment Bin.
Model 3
All w/in 10km of RMP Site
0 to 5 km
House and Location Attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Observations
Cross-area DID
No Offsite
0.5616
(1.3883)
Triple Diff
No Offsite
2.0803
Cross-area DID
Offsite
-6.6905***
(1.7592) (2.3352)
County x Year
County x Year
County x Quarter
Yes
3,359,856
Triple Diff
Offsite
-4.8032*
(2.4678)
Note: Dependent variable is ln(price). *p<0.10, **p<0.05, ***p<0.01. Standard errors in parentheses, clustered at the county
level. Estimates calculated following equations (2)-(5) using the "nlcom" command in Stata 17/MP, and are based on the
coefficient estimates from the hedonic regression model 3. Full regression coefficient results presented in Table A1 in Appendix.
Table 5. Within-area DID Estimates of Percent Change in Price from a Chemical Accident: 0-
5km Treatment Bin.
Model 4
Model 5
0-5km or 7-10km of Accident CEM 0-5km or 7-10km of Accident
0 to 5 km
House and Location Attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Observations
Within-area
DID
No Offsite
1.8805
Within-area
DID
Offsite
-5.4222**
Within-area
DID
No Offsite
0.7805
Within-area
DID
Offsite
-5.7289*
(1.6825) (2.6284)
County x Year
County x Year
County x Quarter
Yes
1,037,013
(1.0844) (3.0592)
County x Year
County x Year
County x Quarter
Yes
893,115
Note: Dependent variable is ln(price). *p<0.10, **p<0.05, ***p<0.01. Standard errors in parentheses, clustered at the county
level. Estimates calculated following equations (4) and (5) using the "nlcom" command in Stata 17/MP, and are based on the
coefficient estimates from the hedonic regression models 4 and 5. Full regression coefficient results presented in Table A1 in
Appendix.
31
-------
Table 6. Average Loss in Home Values due to an Accident with Offsite Impacts (2019$).
Model 3
Cross-area DID Triple Diff
Model 4
Within-area DID
Model 5
Within-area DID
0 to 5 km
-16 949***
[-28,543 to -5,354]
-12,168*
[-24,421 to 85]
-13,736**
[-26,786 to -686]
-14,513*
[-29,702 to 676]
Note: * p<0.10, ** p<0.05, *** p<0.01. Average implicit price estimates based on percent change in price estimates from models
3 through 5 (in tables 4 and 5), multiplied by the pre-accident average price of $253,327. The 95% confidence intervals are
shown in brackets. Estimates and corresponding confidence intervals are calculated using the "nlcom" command in Stata 17/MP.
32
-------
APPENDIX
Figure Al. Number of RMP Accidents in Michigan, Ohio, and Pennsylvania by Year.
Figure A2. Number of Transactions within each One-Kilometer Distance from an RMP Site
2500
2000
T3
£
g 1500
1000
(O
00
4—
o
* 500
0
V" Tr V" & ^ Nvv ^;v ^
33
-------
Figure A3. Number of Pre- and Post-Accident Transactions in each One-Kilometer Bin.
300
34
-------
Table Al. Full Hedonic Regression Results.
Model 1
All w/in 25km
of RMP
Model 2
All w/in 10km
of RMP
Model 3
All w/in 10km
of RMP
Model 4
0-5km or
7-10km of
Accident
Model 5
CEM Sample
RMP Site3
0 to 1 km
0.0004
-0.0009
-0.0016
-0.0143
-0.0203
(0.0110)
(0.0108)
(0.0109)
(0.0174)
(0.0215)
1 to 2 km
-0.0008
-0.0015
-0.0019
0.0115
0.0156
(0.0072)
(0.0070)
(0.0072)
(0.0113)
(0.0126)
2 to 3 km
0.0052
0.0057
0.0056
-0.0007
0.0001
(0.0050)
(0.0050)
(0.0051)
(0.0083)
(0.0098)
3 to 4 km
0.0068*
0.0065*
0.0067**
0.0030
-0.0006
(0.0035)
(0.0034)
(0.0033)
(0.0040)
(0.0057)
4 to 5 km
0.0030
0.0031
0.0030
0.0077*
0.0132*
(0.0034)
(0.0033)
(0.0033)
(0.0045)
(0.0070)
5 to 6 km
0.0002
0.0002
0.0002
0.0021
0.0054
(0.0029)
(0.0030)
(0.0029)
(0.0056)
(0.0062)
6 to 7 km
0.0015
0.0009
0.0010
0.0001
-0.0015
(0.0033)
(0.0033)
(0.0033)
(0.0049)
(0.0062)
7 to 8 km
-0.0057*
-0.0064**
-0.0063**
-0.0089*
-0.0182**
(0.0031)
(0.0031)
(0.0031)
(0.0045)
(0.0077)
8 to 9 km
0.0009
-0.0005
-0.0004
-0.0020
0.0010
(0.0030)
(0.0026)
(0.0026)
(0.0047)
(0.0054)
9 to 10 km
0.0011
(0.0040)
10 to 11 km
0.0004
(0.0031)
11 to 12 km
-0.0031
(0.0027)
12 to 13 km
0.0023
(0.0027)
13 to 14 km
0.0005
(0.0027)
14 to 15 km
-0.0048
(0.0031)
15 to 16 km
0.0052
(0.0036)
16 to 17 km
0.0022
(0.0030)
17 to 18 km
-0.0074**
(0.0038)
18 to 19 km
0.0000
35
-------
19 to 20 km
20 to 21 km
21 to 22 km
22 to 23 km
23 to 24 km
RMP Site Count
-0.0018
(0.0027)
-0.0012
(0.0032)
0.0013
(0.0034)
0.0047
(0.0029)
0.0056*
(0.0030)
0 to 1 km
-0.0457***
-0.0526***
-0.0551***
-0.0524***
-0.0494***
(0.0098)
(0.0097)
(0.0102)
(0.0136)
(0.0177)
1 to 2 km
-0.0183***
-0.0260***
-0.0273***
-0.0400***
-0.0425***
(0.0062)
(0.0059)
(0.0061)
(0.0103)
(0.0135)
2 to 3 km
-0.0094*
-0.0180***
-0.0187***
-0.0174**
-0.0169
(0.0055)
(0.0056)
(0.0053)
(0.0086)
(0.0105)
3 to 4 km
-0.0020
-0.0108**
-0.0107***
-0.0126*
-0.0114
(0.0046)
(0.0042)
(0.0039)
(0.0065)
(0.0077)
4 to 5 km
0.0055
-0.0036
-0.0029
-0.0068
-0.0075
(0.0037)
(0.0028)
(0.0029)
(0.0049)
(0.0062)
5 to 6 km
0.0101***
0.0012
0.0015
-0.0035
-0.0041
(0.0039)
(0.0026)
(0.0028)
(0.0055)
(0.0067)
6 to 7 km
0.0096**
0.0007
0.0011
0.0019
0.0046
(0.0039)
(0.0029)
(0.0030)
(0.0055)
(0.0063)
7 to 8 km
0.0119***
0.0029
0.0030
0.0044
0.0049
(0.0039)
(0.0030)
(0.0030)
(0.0052)
(0.0069)
8 to 9 km
0.0078**
-0.0017
-0.0016
-0.0018
-0.0007
(0.0035)
(0.0026)
(0.0026)
(0.0043)
(0.0066)
9 to 10 km
0.0062*
-0.0028*
-0.0027
-0.0035
-0.0032
(0.0032)
(0.0016)
(0.0016)
(0.0029)
(0.0039)
10 to 11 km
0.0083***
(0.0029)
11 to 12 km
0.0109***
(0.0030)
12 to 13 km
0.0078***
(0.0029)
13 to 14 km
0.0098***
(0.0032)
14 to 15 km
0.0085***
(0.0031)
15 to 16 km
0.0055*
(0.0029)
16 to 17 km
0.0066*
36
-------
17 to 18 km
18 to 19 km
19 to 20 km
20 to 21 km
21 to 22 km
22 to 23 km
23 to 24 km
24 to 25 km
Pre-Accidenta
0 to 1 km
1 to 2 km
2 to 3 km
3 to 4 km
4 to 5 km
5 to 6 km
6 to 7 km
7 to 8 km
8 to 9 km
9 to 10 km
10 to 11 km
11 to 12 km
12 to 13 km
(0.0035)
0.0088***
(0.0033)
0.0041
(0.0025)
0.0042**
(0.0018)
0.0041*
(0.0023)
0.0036
(0.0024)
0.0019
(0.0019)
0.0003
(0.0015)
0.0008
(0.0015)
-0.0423**
(0.0210)
-0.0166
(0.0146)
-0.0251*
(0.0142)
-0.0047
(0.0113)
0.0039
(0.0092)
0.0044
(0.0098)
0.0045
(0.0078)
0.0107
(0.0080)
0.0054
(0.0073)
0.0162*
(0.0094)
0.0024
(0.0066)
-0.0022
(0.0056)
0.0121*
(0.0071)
-0.0479**
(0.0213)
-0.0204
(0.0148)
-0.0279*
(0.0148)
-0.0073
(0.0113)
0.0023
(0.0095)
0.0030
(0.0102)
0.0040
(0.0076)
37
-------
13 to 14 km
14 to 15 km
15 to 16 km
16 to 17 km
17 to 18 km
18 to 19 km
19 to 20 km
20 to 21 km
21 to 22 km
22 to 23 km
23 to 24 km
24 to 25 km
16 to 25 km
0 to 5 km
5 to 7 km
7 to 10 km
Post-Accident"
0 to 1 km
1 to 2 km
2 to 3 km
3 to 4 km
4 to 5 km
5 to 6 km
0.0117
(0.0076)
0.0147**
(0.0068)
0.0059
(0.0073)
0.0040
(0.0068)
0.0059
(0.0069)
0.0103
(0.0119)
0.0127
(0.0108)
0.0154*
(0.0088)
0.0105
(0.0073)
0.0108
(0.0097)
0.0067
(0.0090)
0.0053
(0.0065)
-0.0480*
(0.0258)
-0.0342
(0.0226)
-0.0103
(0.0153)
-0.0139
(0.0154)
-0.0197*
(0.0107)
-0.0153
0.0154*
(0.0079)
-0.0390
(0.0249)
-0.0244
(0.0220)
-0.0009
(0.0146)
-0.0036
(0.0144)
-0.0096
(0.0089)
-0.0041
-0.0138
(0.0108)
-0.0015
(0.0066)
0.0132*
(0.0076)
-0.0044
(0.0089)
0.0002
(0.0085)
0.0117*
(0.0067)
-0.0160**
(0.0067)
-0.0078
(0.0081)
0.0086
(0.0077)
38
-------
6 to 7 km
7 to 8 km
8 to 9 km
9 to 10 km
10 to 11 km
11 to 12 km
12 to 13 km
13 to 14 km
14 to 15 km
15 to 16 km
16 to 17 km
17 to 18 km
18 to 19 km
19 to 20 km
20 to 21 km
21 to 22 km
22 to 23 km
23 to 24 km
24 to 25 km
16 to 25 km
0 to 5 km
5 to 7 km
(0.0106) (0.0088)
-0.0166 -0.0039
(0.0125) (0.0089)
-0.0128
(0.0117)
¦0.0241**
(0.0105)
-0.0205*
(0.0105)
¦0.0225**
(0.0103)
¦0.0227**
(0.0101)
¦0.0288**
(0.0129)
¦0.0268**
(0.0111)
-0.0177*
(0.0100)
¦0.0232**
(0.0116)
-0.0164
(0.0111)
-0.0159
(0.0123)
-0.0209
(0.0130)
-0.0195
(0.0160)
-0.0186
(0.0146)
-0.0156
(0.0104)
-0.0147
(0.0129)
-0.0139
(0.0120)
-0.0021
(0.0077)
-0.0082
(0.0093)
-0.0078
0.0164*
(0.0088)
0.0086
0.0029
(0.0093)
0.0043
39
-------
7 to 10 km
Post-Accident
x Offsite"
0 to 1 km
1 to 2 km
2 to 3 km
3 to 4 km
4 to 5 km
5 to 6 km
6 to 7 km
7 to 8 km
8 to 9 km
9 to 10 km
10 to 11 km
11 to 12 km
12 to 13 km
13 to 14 km
14 to 15 km
15 to 16 km
16 to 17 km
17 to 18 km
18 to 19 km
19 to 20 km
-0.0639
(0.0470)
-0.0673*
(0.0354)
-0.0622
(0.0383)
-0.0563*
(0.0321)
-0.0238
(0.0346)
-0.0084
(0.0287)
-0.0023
(0.0299)
0.0309
(0.0258)
0.0218
(0.0241)
0.0247
(0.0207)
0.0573**
(0.0248)
0.0339*
(0.0202)
0.0375
(0.0295)
0.0197
(0.0281)
0.0348
(0.0240)
0.0337
(0.0220)
0.0267
(0.0210)
0.0368
(0.0234)
0.0373
(0.0251)
0.0403
(0.0260)
-0.0006
(0.0083)
-0.0929*
(0.0480)
-0.0983**
(0.0385)
-0.0937**
(0.0429)
-0.0877***
(0.0333)
-0.0526*
(0.0317)
-0.0369
(0.0240)
-0.0323
(0.0255)
(0.0075)
-0.0018
(0.0087)
(0.0089)
0.0139
(0.0095)
(0.0075)
0.0197*
(0.0115)
40
-------
20 to 21 km
0.0457**
21 to 22 km
(0.0232)
0.0573***
22 to 23 km
(0.0197)
0.0438**
23 to 24 km
(0.0210)
0.0367**
24 to 25 km
(0.0184)
0.0145
0 to 5 km
(0.0118)
-0.0748**
-0.0658*
-0.0651
5 to 7 km
(0.0303)
-0.0357
(0.0360)
-0.0264
(0.0464)
-0.0205
7 to 10 km
-0.0046
(0.0235)
-0.0050
(0.0214)
0.0086
(0.0251)
0.0017
Constant
11.5125***
(0.0141)
11.6612***
(0.0139)
11.6615***
(0.0168)
11.5509***
(0.0215)
11.5253***
(0.0544)
(0.0134)
(0.0135)
(0.0385)
(0.0498)
House and
Location Attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
County x Year
County x Year
County x Qtr
Yes
County x Year
County x Year
County x Qtr
Yes
County x Year
County x Year
County x Qtr
Yes
County x Year
County x Year
County x Qtr
Yes
County x Year
County x Year
County x Qtr
Yes
Observations
Adj. R-squared
3,686,898
0.664
3,359,856
0.675
3,359,856
0.675
1,037,013
0.689
893,115
0.686
Notes: Dependent variable is ln(price). *p<0.10, **p<0.05, ***p<0.01. Standard errors in parentheses, clustered at the county
level. Regression models estimated using "reghdfe" command in Stata 17/MP. (a) denotes binary indicator variables.
41
------- |