NCEE Working Paper
The External Costs of Industrial
Chemical Accidents: A Nationwide
Property Value Study
Dennis Guignet, Robin R. Jenkins, Christoph
Nolte and James Belke
Working Paper 23-01
February, 2023
U.S. Environmental Protection Agency fif
National Center for Environmental Economics livtt flr
https://www.epa.gov/environmental-economics envIronmental^conomics
-------
The External Costs of Industrial
Chemical Accidents:
A Nationwide Property Value Study
Dennis Guignet1, Robin R. Jenkins2, Christoph Nolte3, and James Belke4
Last Revised: January 31, 2023
1. Corresponding Author: Department of Economics, Appalachian State University, 416 Howard
Street, ASUBox 32051, Boone, NC, 28608-2051. Ph: 828-363-2117. guignetdb@appstate.edu.
2. National Center for Environmental Economics, US Environmental Protection Agency.
3. Department of Earth and Environment, Boston University.
4. Office of Emergency Management, US Environmental Protection Agency.
-------
The External Costs of Industrial Chemical Accidents:
A Nationwide Property Value Study
Dennis Guignet, Robin R. Jenkins, Christoph Nolte, and James Belke
ABSTRACT:
Industrial chemical accidents involving fires, explosions, or toxic vapors impose external costs on
nearby communities. We examine changes in residential property values using nationwide data on
chemical facilities, accidents, and residential transactions within a spatial difference-in-differences
framework. We find that accidents with direct offsite impacts lower home values within 5.75 km
by 2-3%, an effect that remains for at least 15 years. We estimate an average loss of $5,350 per
home, which translates to a $39.5 billion loss to communities around the 661 facilities where an
offsite impact accident occurred. We assess the assumptions needed for a formal welfare
interpretation and conclude these results roughly approximate losses experienced by nearby
residents.
Keywords: chemical accident, hedonic, nonmarket valuation, property value, Risk Management
Plan, welfare effects
JEL Classification: D61, L50, Q51, Q53
Acknowledgements: 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.
Additional data support was provided by the Private-Land Conservation Evidence System (PLACES) at Boston
University. 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. We thank Patrick Walsh for helpful comments on
earlier versions of this research.
Disclaimer: The results and opinions are those of the authors and do not reflect the position of Zillow Group. Although
the research described in this paper may have been funded entirely or in part by the U.S Environmental Protection
Agency (EPA), it has not been subjected to the Agency's required peer and policy review. No official Agency
endorsement should be inferred and the results and opinions do not necessarily reflect the position of the EPA.
1
-------
I. INTRODUCTION
Accidents at industrial chemical facilities occur with a frequency and intensity that may impose
substantial social costs. These accidents involve fires, explosions, and drifting toxic vapors, all
of which can directly impact nearby populations. Impacts include injuries and deaths, damages
to nearby properties and the environment, and requirements that the surrounding community
evacuate or take shelter to avoid potential harm. In 2016, EPA estimated that at least 40 million
people (or about 12% of the U.S. population), and perhaps as many as 177 million (55%), were at
risk of experiencing impacts from an accident at these facilities (U.S. EPA, 2016).1 Evidence
suggests that environmental justice is a concern as communities located near industrial chemical
facilities have disproportionately larger income disparities, higher proportions of minority
households, and live in houses with already depressed values (Guignet et al. 2022; Elliot et al.
2004). The social costs imposed by accidental chemical releases can exacerbate these existing
inequalities. This paper provides the first ever national level estimates of the magnitude of such
social costs.
To reduce the impacts of chemical accidents experienced by nearby communities, the US
Environmental Protection Agency (EPA) administers the Risk Management Plan (RMP)
program.2 Section 112(r) of the 1990 Clean Air Act Amendments required EPA to publish
regulations and guidelines to prevent accidents at facilities using certain hazardous chemicals. The
Amendments followed public outrage at the mid-1980s catastrophe in Bhopal, India, where a
pesticide production facility accidentally released a toxic cloud that killed thousands of people.3
In 1996, EPA published a rule that established the RMP Program, requiring regulated facilities to
(1) undertake hazard assessment; (2) develop an accident prevention program; and (3) plan
emergency response activities in case of an accident. Facilities are covered by the program if they
hold above a threshold quantity of a regulated substance.4 At present, 140 toxic chemicals,
including ammonia, chlorine, hydrofluoric acid, and methane, are regulated under the RMP
Program.
1 In the Regulatory Impact Analysis for the 2017 "Accidental Release Prevention Requirements" rule, EPA reported
that approximately 177 million people would be impacted if a hypothetical worst-case scenario accident occurred 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).
2The safety of workers is addressed by the U.S. Occupational Safety and Health Administration.
3US EPA, "Emergency Planning and Community Right-to-Know Act (EPCRA) Milestones Through The Years,"
Accessed at: https://www.epa.gov/epcra/epcra~iiiilestones~through~
vears#:~:text=The%20Bhopal%20disaster%20was%20oire.storage%2C%20releases%20and%20emergencv%20resp
onse. 8 Aug 2022.
4 More specifically, a facility is regulated under the RMP program only if it holds above a threshold quantity in a
"process", as opposed to consideration of sitewide quantities. Under the RMP rule, a "process" means any activity
involving a regulated substance, including any use, storage, manufacturing, handling, or on-site movement of such
substances, or combination of these activities. A single process includes any group of "vessels" that are
interconnected, or separate vessels that are located such that a regulated substance could be involved in a potential
release (40 CFR part 68.3). For example, the quantities of separate containers of the same regulated substance that are
located such that they could be involved in a single accidental release event are aggregated into a single "process" for
purposes of determining whether a threshold quantity is exceeded.)
2
-------
As of 2020, the RMP program regulated close to 12,000 facilities that processed or stored certain
high-risk chemicals. These facilities include a wide range of industrial categories ranging from
complex petroleum refineries, chemical manufacturers, and paper producers, to less complex and
more numerous food and beverage manufacturers, water and wastewater utilities, and agricultural
chemical distributors and wholesalers. From 2004 to 2019, reports to EPA by RMP facilities
show an average of 202 accidents per year. About a quarter of these accidents caused measurable
impacts to offsite communities, including hospitalizations, other medical treatments,
evacuations, shelter-in-place events, or property and environmental damage.
Despite the almost 25-year age of the RMP program, there are no estimates of the value of its
social benefits. EPA updated the requirements for program facilities in 2017 and 2019, and most
recently, proposed amendments in August 2022 (US EPA 2022a, 2022b). Among other
provisions, the 2022 proposal would require root cause analysis of most accidents, third party
compliance audits for certain facilities with multiple accidents, and enhanced worker authority
to "stop work" in situations with a potential for a catastrophic release. Analyses accompanying
these final and proposed rule updates included a comparison of regulatory costs to baseline
accident damages, but lacked estimates of the social benefits from reducing the probability of
accidents (US EPA 2016, 2019, 2022).
It is well-established that hedonic property value results generally lack a formal welfare
interpretation in cases of non-marginal changes and when the hedonic price surface is changing
over time (Klaiber and Smith, 2013; Kuminoff and Pope, 2014). Banzhaf (2021) recently
proposed an approach that allows for inference of a formally valid, bounding welfare measure
based directly on first-stage hedonic property value models that use a difference-in-differences
(DID) design. Under this approach there is no need to estimate Rosen's (1974) second-stage bid
and offer functions. To our knowledge, our study is the first to explore this approach in the
context of chemical accident prevention, thereby facilitating a formal welfare interpretation of
the results; and thus helping inform benefit-cost analyses of RMP and other chemical security
policies that protect surrounding "fence-line" communities.
We start with nationwide data for 2004 to 2019 on facilities regulated by EPA's RMP program,
including information on the number of accidents and their impacts. Regulated facilities are
required to report such information, including details of onsite and offsite impacts. We combine
this with Zillow's nationwide ZTRAX data of residential parcels and transactions. Within a
hedonic regression framework, we use a DID design to examine differences in property prices
before and after an accident. We compare those differences between homes that are near versus
far from an accident. Our study is the first ever to assess the nationwide property value impacts of
chemical facility accidents. Prior research on similar disamenties has focused on only one or a few
accident cases (e.g., Carroll et al. 1996; Hansen et al. 2006; Grislain-Letremy and Katossky 2014;
and Herrnstadt and Sweeney 2019), or on a sub-national region within the US (Guignet, et al.
2022).
Our paper contributes three additional unique analyses. The effects on home prices of accidents of
different severity are estimated, as are the different effects of single versus multiple accidents. We
also examine the persistence of any adverse price impacts over time.
3
-------
The results suggest that homes as far as 5.75 km away are impacted by a chemical accident, but
the adverse price effects are limited to the most severe cases; i.e., accidents resulting in deaths or
injuries to people in the surrounding community, damage to offsite properties and environmental
systems, and/or the evacuation or sheltering-in-place of offsite populations. Among those homes,
an average price decline of 2% to 3% is experienced. We do not find evidence of systematically
different price declines among homes that experience multiple chemical accidents, but do find that
home values remain depressed for at least 15 years after an offsite impact accident occurred.
The average loss in a home's value is about $5,350 (2021$ USD), and this translates to a $39.5
billion loss due to the offsite impact accidents that occurred at 661 different facilities from 2004
to 2019. The loss to the community where an offsite impact accident occurs was, on average, $59.8
million (the median loss was $24.9 million). We provide evidence that the depreciation in prices
due to an offsite impact accident may be constant over our study period, which is a necessary
assumption to support interpretation of these results as a theoretical upper bound of the ex post
welfare loss to nearby residents (Banzhaf, 2021). For smaller shocks, such as may be the case for
our estimated 2% to 3% loss, the bounding estimate better approximates the true ex post welfare
loss to nearby residents (Banzhaf, 2021).
The remainder of this paper includes a brief description of EPA's RMP program, a literature
review, details on our data and methods, and a summary of our results. We conclude with a
discussion of the necessary assumptions to interpret the estimates as national-level ex post welfare
impacts on nearby communities, and the analytical advantages afforded by detailed, broad-
coverage datasets like that provided by Zillow's ZTRAX program.
II. LITERATURE
Soon after the first chemical facility accident data became public following establishment of the
RMP Program, several publications explored factors that correlated with accidents occurring
between 1994 and 2000. These studies examined how facility characteristics, applicable federal
regulations, and firm financial variables related to accidents; and reported on the correspondence
between accident risk and socioeconomic status of the surrounding communities (Kleindorfer, et
al., 2003, Elliot, et al., 2003, Kleindorfer et al., 2004, Elliot et al., 2004). For example, Kleindorfer
et al. (2004) identified a positive relationship between a facility's debt-to-equity ratio and accident
propensity. Elliot et al. (2004) concluded that larger RMP facilities and those using a larger number
of chemicals are disproportionately located in counties with higher median incomes, but also
greater levels of income inequality and a higher proportion of African Americans.
Multiple hedonic case studies have examined home values near petroleum refineries, chemical
plants, and natural gas pipelines, and find that prices decline following a chemical explosion
(Flower and Ragas, 1994; Carroll et al., 1996; Hansen et al., 2006; Liao et al., 2022). Such adverse
effects may vary from case to case. Focusing on homes near pipelines in San Bruno, California,
Herrnstadt and Sweeney (2019) find that a 2010 pipeline explosion and subsequent mail
4
-------
notifications to all households living within 2,000 feet of a natural gas pipeline resulted in no
impact on prices.
In a nationwide analysis of properties near natural gas distribution pipelines, Cheng et al. (2021)
find that home values within 1 km decline by 7.4% after an explosion, compared to a control group
of homes l-2km away. Cheng et al.'s spatial difference-in-differences approach is similar to the
identification strategy implemented in our analysis, as well as to hedonic studies of similar types
of disamenities and releases of hazardous chemicals. Guignet et al. (2018) examine property value
changes around high-profile releases from underground storage tanks (USTs) at retail gas stations
and find that homes within 3 km depreciate an average of 6%. Guignet and Nolte (2021) conduct
a nationwide study of hazardous waste treatment, storage, and disposal facilities (TSDFs). They
caution against a causal interpretation but do find evidence that home values within 750 meters
may decrease up to 5% after the discovery of contamination.
Our study also relates to a branch of literature on air pollution and home values. In general air
quality improvements increase home values (e.g., Chay and Greenstone 2005; Bayer, et al. 2009;
Grainger 2012; Bento et al. 2015; Lang 2015; and Amini et al. 2021). Much of this literature
focuses on "criteria air pollutants" as designated under the Clean Air Act, rather than toxic
pollutants. However, there are several notable exceptions focusing on EPA's Toxic Release
Inventory (TRI) Program. The TRI Program requires firms with threshold quantities of reportable
chemicals to disclose fugitive emissions. Currie et al. (2015) examine residential transactions near
1,600 industrial facilities in five U.S. states and conclude that the opening of a facility that reports
toxic air emissions to the TRI led to an 11% price decrease of homes within 0.5 miles.
Mastromonaco (2015) examines house price changes in several California counties with existing
firms newly required to report to the TRI in 2001 and finds up to 11% lower prices within one
mile. Mastromonaco interprets the house price impacts in this context as responses to firms
maintaining threshold chemical quantities, rather than to changing emissions levels. A working
paper by Moulton et al. (2018) examines nationwide home values around facilities newly required
in March 2000 to report pollution releases to the TRI. They find that home prices within a half
mile of firms emitting at high levels decrease by approximately 8%, with smaller price declines
experienced by homes up to 5 miles away. Banzhaf (2021) examines the effect of changing the
number of plants required to report to the TRI on house values in the Los Angeles area between
1995 and 2000. He finds that homes located within a mile of reporting plants experience negative
and significant price effects relative to homes located 1 to 2 miles away.
Most reported emissions to the TRI entail routine and intentionally emitted pollutants. In contrast,
pollution incidents reported under the RMP Program are exclusively the result of infrequent,
accidental air emissions and the resulting explosions, fires, and clouds of toxic vapors. The results
of hedonic studies on criteria air pollutants and TRI facilities are not necessarily transferable when
examining the impacts of the RMP program.
5
-------
To our knowledge, there is only one nonmarket valuation study specifically on the RMP Program.
In an earlier study, Guignet et al. (2022) use DID and triple difference approaches, along with
coarsened exact matching techniques, to examine the impact of chemical accidents on home prices
in a tri-state region of the US (Michigan, Ohio, and Pennsylvania). They find that the typical
accident yields no effect on surrounding home prices, but homes within 5 km of an accident that
impacts surrounding populations (i.e., leads to offsite injuries, property damage, shelter-in-place,
or evacuations of people in the surrounding community) experience a 5% to 7% decline in price.
We expand on Guignet et al.'s (2022) regional case study in multiple ways. First, we include a
more current study period, and examine national scale price impacts. This is important as it will
enable assessment of national-level policies and programs. By expanding our study area and time
period, we include more facilities and home sales in our analysis. The larger sample size allows us
to determine the appropriate spatial extent of any price impacts using higher resolution 250-meter
bins, rather than the one-kilometer bins used by Guignet et al. (2022). Second, we examine how
price impacts vary not just by whether the accident resulted in offsite impacts or not, but also if an
accident included any onsite impacts that were required by EPA to be reported. Third, we examine
the role of multiple accidents in updating residents' perceived risks, and the subsequent price
effects.
An additional important contribution enabled by the larger sample size, is that we can look more
in-depth at how price impacts evolve over time. After an accident, property values in nearby
communities could experience a brief period of decline or might suffer a persistent negative
impact. Stigma is a phenomenon explored in the economics literature, although mostly in relation
to cleanups of contaminated sites such as those on the National Priorities List (NPL). Messer, et
al. (2006) examine up to 30 years of house price fluctuations in metropolitan areas neighboring
prominent NPL sites. The study concludes that cleanups occurring over lengthy 10+ year periods
do cause stigma, and that neighboring property values do not rebound enough to compensate for
losses from the original contamination. One mechanism through which prices may remain
depressed is the re-sorting of households following a contamination event (or chemical accident),
whereby higher income families move away from contamination and, because they cannot afford
otherwise, lower income families move toward it. Cleanup would lead to the opposite effect - i.e.,
gentrification. Banzhaf (2012) provides a discussion of gentrification that clearly identifies
potential distributional concerns. A persistent price decline from chemical accidents would affect
house values that are already lower than average due to proximity to an industrial facility.
Studies in the context of toxic air emissions and pipeline explosions have found some evidence of
stigma leading to a persistent discount in house prices. In their study of toxic air emissions, Currie
et al. (2015) find that the initial decline in house prices due to the opening of a facility remains,
even after the plant is closed. Hansen et al. (2006) study home sales located within a mile of two
pipelines to estimate the impacts of a 1999 fuel pipeline explosion in Bellingham, Washington.
6
-------
Following the accident, property prices were significantly lower, with the effect diminishing with
distance, from a 4.6% decline for a property 50 feet from the pipeline to 0.2% at 1,000 feet away.
They find that prior to the explosion neither of two regional pipelines affected nearby property
values, but for the five-year-time period following the event there was a significant negative effect
of proximity to the pipeline that experienced the explosion (though it diminished in magnitude
with each passing year). The researchers attribute the persistent five-year effects to households
receiving new information about the location and risks of the pipeline, but also to attention-
grabbing media coverage that may have led people to overestimate risk.
In contrast, other studies of site contamination find little evidence of persistent negative effects.
Taylor, et al. (2016) compare the impact on home prices of commercial properties with no known
contamination to commercial properties with remediated contamination and find that any
differences in price are largely indistinguishable. The paper concludes that stigma does not persist
once a contaminated site is remediated. Guignet, et al. (2018) similarly find no evidence of stigma
in highly publicized cases of leaking and remediated underground storage tanks, nor do Guignet
and Nolte (2021) at remediated hazardous waste sites. Both studies find surrounding home values
rebound after cleanup is complete. We contribute to this persistence and stigma literature by
examining how initial price declines due to a clearly noticeable chemical accident evolve over
time.
One of our most important contributions is to assess a formally valid, welfare interpretation of our
hedonic results. DID is an increasingly popular estimation approach in the hedonic property value
literature (Parmeter and Pope, 2013; Guignet and Lee, 2021), and for causal inference in general.
However, the results from hedonic property value studies generally lack a formal welfare
interpretation in cases of non-marginal changes and when the hedonic price surface is changing
over time (Klaiber and Smith, 2013; Kuminoff and Pope, 2014). Though many approaches have
been suggested, there is no commonly agreed upon best approach to improve the interpretation of
hedonic estimates as welfare changes (Bishop et al., 2020).
With a specific focus on the DID design, Banzhaf (2021) demonstrates that a change in price along
the same ex post price gradient is a lower bound of the Hicksian equivalent surplus for an
improvement in quality. Conversely, it can be interpreted as an upper welfare bound to the nearby
community for a decrease in quality, as is the case for chemical accidents at industrial facilities. In
later models we allow the entire hedonic price surface to vary over time, which allows for a formal
ex post (i.e., post-accident) welfare interpretation of the results.
To our knowledge, our study is the first comprehensive, nationwide non-market valuation study
of accidents at chemical facilities. Our analyses better characterize price impact heterogeneity and
the potential persistence of price effects over time, thus providing more detailed insights about the
impacts of industrial accidents and similar disamenities. Furthermore, our estimated capitalization
effects and assessment of a formal welfare interpretation can inform benefit-cost analyses of
7
-------
federal regulations, as well as state and local decision-making. This is particularly relevant in light
of climate change and the vulnerability of industrial operations to increasingly severe storms,
floods, and wildfires (Flores, et al. 2021; US Chemical Safety Board, 2017; Chemical Industries
Association, 2015). Our results can inform more socially efficient decisions regarding the RMP
program and other federal chemical security policies, as well as local land use decisions and
climate adaptation strategies.
III. DATA
The empirical analysis focuses on home transactions from 2004 to 2019 in the contiguous U.S.
that occurred within 10 km of an RMP facility where a chemical accident was reported. Data
describing all RMP facilities and reported accidents were provided by EPA's Office of Emergency
Management. We spatially and temporally link facilities with reported accidents to transaction
data of single-family homes from Zillow's ZTRAX database.
III. A. RMP Facilities and Accidents
The EPA maintains a nationwide database of all facilities regulated under the RMP program.
Every five years, regulated facilities must identify and describe any accidents with reportable
impacts that occurred over the prior five years. Reportable impacts include onsite fatalities,
injuries, and property damage, as well as offsite fatalities, hospitalizations, people in need of
medical treatment, number of people evacuated, number of people who were sheltered-in-place,
and finally offsite property and environmental damage.5 As a result, the RMP national database
contains a continuous record of accidents from regulated facilities, beginning five years prior to
facilities' initial submissions at the program's inception in 1999. Our analysis focuses on the
1,822 facilities where at least one chemical accident was reported to have occurred from 2004 to
2019. Figure 1 shows that these facilities and accidents are quite dispersed across the contiguous
U.S., but with a higher spatial concentration in the rust belt around the Great Lakes, along the east
coast and Gulf of Mexico, and in portions of California and the Northwest. Facilities participate in
a variety of industrial activities, with the six most common being farm supplies wholesalers
(13.1%), organic and inorganic chemical manufacturing (7.7%), refrigerated warehousing and
storage (6.0%), poultry processing (5.0%), water supply and irrigation systems (4.9%), and
petroleum refineries (4.7%).
These facilities all reported at least one accident, with a mean of 1.8 accidents, and a median of
one. There is, however, noticeable variation in the number of accidents reported. Most facilities
report just one accident (68.5%), but 31.5% report multiple accidents. The 90th and 95th percentiles
are three and five accidents, and a maximum of 30 accidents is reported by one facility.
5 See 40 CFR 68.42.
8
-------
Figure 1. Study area and RMP accident sites.
0 Offsite Impact Accidents Monterr#>+
• Reportable Onsite Impact Accidents
+ Nonreportable Accidenfeiada1aJara
+ V
i #mi
%
Havana
Esri, HERE, Garmin, (c) OpenStreetMap contributors, and the GIS
user community
Port-au-
A total of 3,236 chemical accidents are reported by the 1,822 RMP facilities from 2004 to 2019,
an average of 202 accidents per year. As shown in Table 1, most of the accidents were at least
partly due to equipment failure (62.5%), followed by human error, issues due to maintenance
activities or a lack thereof, and unexpected weather conditions. Most accidents involved the release
of a hazardous gas into the air (66.8%), but liquid spills and subsequent evaporation, chemical
fires, and even explosions are somewhat common. Only 34 accidents (1.1%) resulted in
uncontrolled chemical reactions. Of the 140 hazardous substances regulated under the RMP
program, the chemicals most often reported as released include ammonia, chlorine, hydrofluoric
acid, propane, sulfur dioxide, hydrogen sulfide, methane, butane, and a general "flammable
mixture" category.
Most of the accidents (2,275 or 70.3%) were required to be reported under the RMP regulations,
but 961 (29.7%) were not required to be reported, meaning that the accident did not result in RMP
reportable impacts (i.e., deaths, injuries, significant property damage, environmental damage, or
the shelter-in-place or evacuation of people in the surrounding community). In Figure 1, the
nonreportable and reportable onsite impact accidents are denoted by the black crosses and dots,
respectively. We include all reported accidents and examine for potential heterogeneity in the
housing price impacts. Almost a fourth of the accidents (789) resulted in impacts that were not
limited to the facility itself, directly affecting the surrounding environment and community. As
shown by the diamonds in Figure 1, we label such accidents as having offsite impacts. These
accidents resulted in offsite environmental damage (e.g., defoliation to trees, surface water
contamination, dead or injured animals), damage to properties located offsite, injuries or deaths to
offsite populations, and/or the evacuation or shelter-in-place of people in the surrounding
community.
9
-------
Details of both onsite and offsite impacts can be found in the lower portion of Table 1. About 1.5%
(48) accidents resulted in one or more deaths to people onsite, including facility workers and first
responders. Among those 48 accidents, there was an average of two deaths. One accident resulted
in a maximum of 15 deaths. About 42% of accidents resulted in injuries onsite, and among those
accidents the number of people injured ranged from one to 250. Onsite property damage was
reported in 780 cases, with onsite damages assessed at $8.01 million (2021$ USD), on average.6
Damage to the surrounding environment was reported among 5.9% of the accidents. Fortunately,
only one accident resulted in the death of an individual located offsite, but there were 249 accidents
(8.0%)) that resulted in injuries to the surrounding population. The number of people injured offsite
among those accidents ranged from one to over 14,000, with an average of 61 people injured per
accident. Offsite property damage occurred in 89 incidents, with an average assessed damage of
$2.06 million. The evacuation and shelter-in-place of people in the surrounding community
occurred in 312 (9.6%) and 220 (6.9%) cases, respectively; and impacted 307 and 2,369 people,
on average. Thirty-nine accidents resulted in more than 1,000 individuals being evacuated or
sheltered-in-place.
Table 1. Chemical accident descriptive statistics.
Variable3
Obs
Mean
Std. dev.
Min
Max
Causes of accidentb
Equipment failure
3,236
0.6252
0.4842
0
1
Human error
3,236
0.4796
0.4997
0
1
Maintenance activity/inactivity
3,236
0.1792
0.3836
0
1
Weather
3,236
0.0374
0.1897
0
1
Type of accidentb
Gas release
Liquid spill and evaporation
Fire
Explosion
Chemical Reaction
Impacts of accident
Onsite deaths
# onsite deaths (people)
Onsite injuries
# onsite injuries (people)
Onsite property damage
Total onsite property damage (2021$ USD)
Environmental damage
Offsite deaths
3,236
0.6681
0.4710
0
3,236
0.3446
0.4753
0
3,236
0.1165
0.3209
0
3,236
0.0405
0.1971
0
3,236
0.0105
0.1020
0
3,236
0.0148
0.1209
0
48
2.04
2.63
1
3,236
0.4203
0.4937
0
1,356
2.34
9.01
1
3,236
0.2420
0.4283
0
780
8,019,896
4.21E+07
1
3,236
0.0590
0.2357
0
3,236
0.0003
0.0176
0
1
1
1
1
1
1
15
1
250
1
5.94E+08
1
1
6 All nominal dollar values converted to 2021$ USD based on the Bureau of Labor Statistics annual US city average
"All Urban Consumers" consumer price index (CPI), available at: https://www.bts.gov/epi/tabtes/siipptemental~
fites/historicat~epi~n~202206.pdf. accessed 31 July 2022.
10
-------
# offsite deaths (people) 1 1.00 - 1 1
Offsite injuries 3,236 0.0803 0.2719 0 1
# offsite injuries (people) 249 61.19 887.22 1 14,003
Offsite property damage 3,236 0.0275 0.1636 0 1
Total offsite property damage (USDS) 89 2,062,809 1.72E+07 58 1.62E+08
Offsite evacuations ordered 3,236 0.0964 0.2952 0 1
# offsite people evacuated 312 307.05 2,861.53 1 50,000
Offsite shelter-in-place ordered 3,236 0.0686 0.2528 0 1
# offsite people sheltered-in-place 220 2,368.65 8,342.48 1 55,000
Note: The total number of observations is 3,236 chemical accidents.
(a) Variables are binary indicator variables unless otherwise noted.
(b) Cause and type of accident categorical variables are not necessarily mutually exclusive.
III. B. Residential Property Transactions
Residential parcels are individually linked to any of the 1,822 RMP accident sites that are within
10 km. We account for the timing of accidents relative to transactions of those parcels between
2004 and 2019. A total of 10,428,442 arms-length transactions of single-family homes are
observed within 10 km of one or more of the RMP accident sites.7 The data includes residential
transactions in 47 of the 48 states across the contiguous U.S.8
Distance of a home from an RMP facility is accounted for using 250-meter incremental distance
bins. A continuous distance measure to the nearest accident site was not an appropriate measure
in our context because a home can potentially be near multiple RMP facilities that have
experienced accidents. Accounting for proximity using bins allows us to track the number of RMP
facilities and accidents in each incremental distance zone. We wanted to maintain a high-spatial
resolution while also ensuring a sufficient number of transactions within each bin for initial
empirical diagnostics. We judged 250-meter incremental bins as an appropriate size considering
these tradeoffs. The number of sales observed in each 250-meter bin from a facility both before
and after an accident are displayed in Figure A.l of Appendix A.
Key variables and descriptive statistics are provided in Table 2. Several variables describing the
home are derived from ZTRAX's transaction and assessment databases. Attributes of the location
of a home were provided by the Private-Land Conservation Evidence System (PLACES) at Boston
University. PLACES uses assessor parcel numbers to link ZTRAX data to parcel boundaries based
on county and town-specific deductive string pattern matching and geographic quality controls
(Nolte, 2020). For each parcel, we identified the census tract using spatial joins, computed
Euclidean distances to the nearest highway (using TIGER road data), lake (>4ha) and river (using
7 We focus solely on full, arms-length transactions of single-family homes. Data cleaning and formatting details are
provided in Appendix A.
8 We exclude Washington, D.C. because it is a unique housing market, and Wyoming is excluded because it is a non-
disclosure state and no transactions from the available data were of homes within 10 km of an RMP accident site.
Limited available sales data for the other states often cited as non-disclosure states (Wentland et al., 2020) are
maintained in our analysis, including Idaho, Kansas, Louisiana, Mississippi, Montana, New Mexico, North Dakota,
Texas, and Utah.
11
-------
the waterbody polygons from the National Hydrography Database), as well as the proportion of
developed land cover within a 500-meter circular buffer around each home (using the 2011
National Land Cover Database).9
The average home sells for just under $260,000 (2021$ USD), is on a 0.28-acre lot, and is 1.3
stories high. The average number of bathrooms is 1.9, and the interior square footage and age of a
home, on average, are about 3,300 sq ft and 40 years. As can be seen by the companion missing
variable indicators for the house structure and acreage variables, the percent of observations
missing are most noticeable for the number of bathrooms (30%) and stories (16%). Missing values
are coded as zero, and are included in the later hedonic regression models, along with the
corresponding missing value indicators.
Table 2 reports location attributes showing that, on average, 52% of the land within 500 m of a
home is developed. About 36% of sales are of homes within 500 m of a highway, and 4.6% and
2.2% are within 500 m of a lake and 250 m of a river, respectively. The subsequent hedonic
regression models include spatial fixed effects at the census tract level, but location attributes are
included to capture local, within tract variation of amenities and disamenities near each home.
Table 2. Residential transactions descriptive statistics.
Variable
Obs
Mean
Std. dev.
Min
Max
Price (2021$)
10,428,442
259,182
190,929
15,000
999,972
Transaction year
10,428,442
2011.13
4.85
2004
2019
Quarter
10,428,442
2.52
1.06
1
4
Acres
10,272,126
0.2834
0.2709
0.05
2
Missing: Acres'
10,428,442
0.0150
0.1215
0
1
Stories
8,802,167
1.34
0.48
1
3
Missing: Stories'
10,428,442
0.1559
0.3628
0
1
Bathrooms
7,310,276
1.94
0.76
1
4.5
Missing: Bathrooms^
10,428,442
0.2990
0.4578
0
1
Interior square footage
9,845,806
3,312.90
2,345.69
750
15,000
Missing: Interior square
footage'
10,428,442
0.0559
0.2297
0
1
Age (years)
9,702,115
39.70
29.07
0
120
Missing: Age1,
10,428,442
0.0696
0.2546
0
1
% Land Developed w/in 0-
500m
10,428,442
52.44
23.36
0
100
Highway w/in 500m '
10,428,442
0.3568
0.4791
0
1
9 Census tract data comes from the National Historical Geographic Information System (Manson et al., 2018), and are
based on the tract boundaries from the 2016 American Communities Survey. Land cover data comes from the 2011
National Land Cover Database (Dewitz, 2019). The highways data is from the US Census Bureau's TIGER/Line
shapefiles (2019).
12
-------
Lake w/in 500mt 10,428,442 0.0456 0.2087 0 1
River w/in 250m* 10,428,442 0.0216 0.1452 0 1
Note: The final sample includes n= 10,428,442 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.
IV. METHODS
This section explains our empirical design and presents models of the price impacts of chemical
accidents, cumulative price effects and their attenuation over time, and the welfare effects
experienced by nearby residents.
IVA. Stacked spatial difference-in-differences design
Spatial difference-in-differences (DID) is a popular approach to infer causal impacts in hedonic
pricing models (Parmeter and Pope 2013; Guignet and Lee 2021). Davis's (2004) application in
valuing the implicit price of pediatric cancer risks, and Linden and Rockoff s (2008) analysis of
how proximity to registered sex offenders impacts home values, are among the first to demonstrate
the appeal of the DID strategy in a property value setting. The approach has since been used in
numerous environmental hedonic applications (e.g., Horsch and Lewis 2009; Atreya et al. 2013;
Bin and Landry 2013; Muehlenbachs et al. 2015; Haninger et al. 2017; Guignet et al. 2017).
The DID strategy closely resembles a classical experimental design. Figure 2 depicts homes near
an RMP facility, both before and after a chemical accident. "Treatment" in this quasi-experimental
setting is defined as being near an industrial facility where a chemical accident occurs. Homes
denoted by group A are the treated group, pre-treatment; and those in group B are the treated group,
post-treatment. The impacts of the chemical accident on the price of homes in group B are of
primary interest, but identifying the appropriate counterfactual is critical.
A simple before and after (or first differences) estimate would entail the difference in price
between groups B and A (i.e., PB — PA). Such a comparison, however, is susceptible to temporally
varying confounders. If the price of homes in a neighborhood are changing due to other unobserved
factors, then a first differences estimate would suffer from an omitted variable bias. Such concerns
are partially alleviated because our analysis is a stacked treatment design (Cengiz et al. 2019;
Deshpande and Li 2019; Fadlon and Nielsen 2021), where RMP accidents occur at different
locations and at different times. Confounding factors would have to be temporally correlated
across many RMP accidents and nearby neighborhoods to bias the first differences results, but it
is still possible. For example, a plausible situation might be that an industrial facility is no longer
as profitable as it used to be, and cost-savings measures may imply an increased risk of an accident.
Such facilities could tend to be in neighborhoods that are experiencing ongoing economic decline.
The DID strategy can further alleviate omitted variable bias concerns, and bolster causal inference,
by using homes in the broader neighborhood as a counterfactual. The intuition and key assumption
are that the price of homes in the broader neighborhood are affected by the same unobserved trends
13
-------
as the treated group, but are too far away to be impacted by the chemical accident. The homes in
these farther distance bins (denoted by groups C and D) serve as the control group, and so the
second difference in the DID strategy allows one to difference out the broader neighborhood trends
that could otherwise bias the price effects of interest. The DID estimate is (PB — PA) — (PD — Pc).
Figure 2. Illustration of the Difference-in-differences approach.
Pre-accident Post-accident
Site with
Accident
This DID strategy is applied within a traditional hedonic price regression model, where the
dependent variable is the natural log of the price of home in neighborhood j (i.e., census tract),
in housing market m (i.e., county), at time t (Pijmt)• The independent variables include a vector
of house and location characteristics {xtjmt), county-by-year and county-by-quarter fixed effects
(rmt), and neighborhood fixed effects (Vjm)•
We set out to answer five main research questions. First, does the typical accident tend to impact
the value of nearby homes? Second, how do the price impacts vary based on severity of an
accident? Third, what are the cumulative effects on home prices due to multiple accidents? Fourth,
do the price impacts tend to attenuate over time? And fifth, what are the formal welfare
implications?
IV.B. Price impacts of chemical accidents
Our empirical analysis largely follows equation (2) below, but for purposes of exposition we start
with the simplest model:
(1) lYl(Pijmt) jmtPmt
+ pRMPt + 8postit + x postit) +
T-mt Vjm ^ijmt
where £jymt is a normally distributed disturbance term (which we allow to be correlated for all
transactions within the same county). The subscripts on the parameter Pmt reflect that the slope
coefficients of the house and location attributes are allowed in some models to vary over time and
by market. Although not explicitly represented for notational ease, in our most comprehensive
models these attributes are interacted with market and year indicators, or even with market-by-
14
-------
year indicators. We are estimating a nationwide hedonic price model, but the hedonic price surface
is an equilibrium result. The "law of one price" states that identical houses should sell for the same
price throughout the entire assumed market (Bishop et al. 2020). Assuming the entire nation is a
single housing market would surely violate this principle. The inclusion of these interaction terms
allows the equilibrium price surface to vary across space and time with respect to the house and
location attribute dimensions.
RMPt is an indicator denoting that an RMP facility is in close proximity to the home (i.e., 0 to
5,750 meters). Section V.A describes how the distance used to define being in close proximity to
an RMP facility (i.e., within the treated zone distance) is established. For the "treated" group of
homes, no matter whether the sale occurs before or after an accident, RMP£ = 1, and so the
coefficient p captures the baseline price differences associated with an RMP facility being nearby.
In some later models we allow for heterogeneity of the price effects within the assumed treated
zone, and in such cases RMP£ is a vector of indicator variables corresponding to 250-meter
incremental bins.
The indicator denoting that an RMP facility had an accident nearby, or in the broader vicinity of,
a home (postit) reflects the post-treatment period, irrespective of being in the treated or control
group. In other words, the parameter 5 captures temporally varying, and otherwise potentially
confounding, factors affecting prices. Given the inclusion of year-by-county and quarter-by-
county fixed effects, 5 may be redundant, capturing only average within year and within quarter
variation associated with the pre- and post-accident periods. Nonetheless, we include postit to be
as thorough as possible in controlling for temporally correlated confounders.
The variable of primary interest is the interaction term RMPt X postit, which equals one when a
chemical accident occurred at an RMP facility near the home, as of the time of sale. The key
parameter y thus captures the average effect of the treatment on the treated (ATT).
We estimate variants of equation (1) to examine heterogeneity in any price effects based on
severity of the accident. In an earlier case study of just three states (MI, OH, and PA), Guignet et
al. (2022) found that the typical RMP accident did not affect home values on average, but
significant price declines were found among homes near chemical accidents that impacted offsite
populations. We explore such heterogeneity here with respect to offsite versus onsite impacts, as
well as whether the accident resulted in any reportable impacts in general. (Recall that the data
contain reported accidents that were not required to be reported under the RMP program.)
To examine heterogeneity in the housing price impacts with respect to accident severity, we
include an interaction term with a vector of accident characteristics accit, as follows:
(2) ln(Pijmt) — X-ijmtPmt
+ pRMPi + 8postit + r(RMPi x postit)
+ X postit ^ T-mt ^jm ^ijmt
The coefficient vector 0 captures incremental differences in the price effects of an accident,
depending on whether the accident was reportable, or resulted in offsite impacts.
15
-------
The percentage change in price for a non-reportable accident, a reportable accident resulting in
only onsite impacts, and an accident exhibiting more severe offsite impacts are calculated,
respectively, as:
(3a) %Apnr = {exp(y) — 1} x 100
(3b) %Aprep = {exp(y + 6lrep]) - 1} x 100
(3c) %Ap°ff = {exp(y + 0^rep^ + — l) x 100
where 0'rep' is the first element of the coefficient vector 0, and captures the incremental effect of
an accident with reportable onsite or offsite impacts, relative to a non-reportable accident. 6
is the second element of 0, and reflects the incremental effect of an accident yielding offsite
impacts, relative to an accident yielding only reportable onsite impacts. The percent change in
price estimates described by equations (3 a) through (3 c) are the DID estimates of primary interest
and represent the weighted average of the ATT.10
IV.C. Cumulative price effects and attenuation over time
In subsequent models we investigate the potential cumulative effects of multiple chemical
accidents occurring near a home. We include additional interaction terms with the number of RMP
sites (RMP_cnti) and accidents that occurred (post_cntit), and the number of more severe
accidents (acc_cntit), as shown:
(4) ln(pijmt) = xijmtpmt + pRMPi + padd{RMP_cnti - RMPt)
+8postit + 8add(post_cntit — postit)
+y(RMPi x postit) + 7add(/?MPj x (post_cntit — postjt))
+(/?MPj x postit x accit)0
+{RMPt x postit x (acc_cntit - accit))Gudd + rmt + vjm + eijmt
The corresponding dummy variables are subtracted from the RMP and accident count variables,
so that, for example, post_cntit — postit = 0 if there was just one accident (i.e., post_cntit = 1
10 Recent literature has cautioned against this average ATT interpretation in settings where the treatment events are
staggered over time (Goodman-Bacon, 2021; Marcus and Sant'Anna, 2021; Roth et al., 2022; Sun and Abraham,
2021). The primary criticism is that in some settings a subset of treated observations can receive a negative weight if
they serve as a control observation for subsequent treatment event comparisons. The spatial DID design implemented
here and by numerous other hedonic property value applications (e.g., Linden and Rockoff, 2008; Haninger et al.,
2017; Muehlenbachs et al., 2015; Guignet et al., 2018) utilize a clearly separate control group (i.e., farther away homes
located around the same disamenity). This setup is essentially a stacked DID design (Cengiz et al., 2019; Deshpande
and Li, 2019; Fadlon and Nielsen, 2021), which is one suggested approach to address concerns with staggered
treatment events (Goodman-Bacon, 2021, Roth et al., 2022).
16
-------
and postit = 1); and postjcntit — postit = 1 if there were two accidents (i.e., post_cntit =
2 and postit = 1), and so on. In other words, the differenced variable (post_cntit — postit) is
the number of additional accidents that occurred after the first. And so yadd will capture the price
impacts of each additional accident after the first, in a linear fashion. Each additional accident
could lead residents to perceive the risks posed by the site as greater (yadd < 0). On the other
hand, additional accidents may not yield any new information towards the surrounding
community's perceived risk, in which case additional price impacts may diminish and be
negligible when multiple accidents occur (yadd = 0). A similar interpretation follows for the
differenced (acc_cntit — accit) variable and 0add coefficient - the coefficient captures the
incremental price effect for each additional more severe accident, relative to the least impactful
(non-reportable) accident category.
A variant of equation (2) is also estimated to examine whether any adverse price effects attenuate
over time, remain constant, or intensify. The coefficients corresponding to a chemical accident
(ys) and characteristics of that accident (0S) are allowed to vary for each year 5 after the accident.
We can then flexibly examine whether any adverse price effects attenuate over time (i.e., become
less negative (ys < ys+1, V s = 1,.remain constant (ys = ys+1), or intensify (i.e., become
more negative (ys > rs+1)).
(5) ^{PijmtsJ ~ X-ijmtPmt pRMPi + 5pOStjt
+ T,Ui{Ys(RMPt x postis) + (RMPt x postis x accis)0s}
Vjm + ymt
IV D. Inferring welfare impacts to nearby residents
As discussed in section II, non-marginal results from first-stage hedonic price regressions
generally lack a formal welfare interpretation. Although hedonic property value studies compose
an increasingly large portion of the nonmarket valuation literature, they are often not used in
benefit-cost analyses of environmental policy (Petrolia et al., 2021). The lack of a formal, non-
marginal welfare interpretation is one reason why. Although several studies have provided
guidance on ways to derive welfare estimates and bounds (see Bishop et al. 2020 for a review),
only recently has progress been made to infer a formal, non-marginal welfare estimate in a DID
setting.
Banzhaf (2021) demonstrates that a change in price along the same ex post price gradient is a lower
bound of the Hicksian equivalent surplus for an improvement in quality. Conversely, a decrease
in quality, like that from a nearby chemical accident, would suggest a theoretical upper bound of
the Hicksian equivalent surplus to affected residents. Banzhaf shows that this bounding estimate
more closely approximates the true loss for smaller shocks, with the two equating as the shock
approaches a marginal change.
17
-------
To implement Banzhaf s approach, the hedonic price surface must be allowed to vary over time
(ideally) with respect to all dimensions. For the current study, this includes not just fimt (which
is already allowed to vary over time in most of our model results), but also p, 5, y, and 0. Building
off equation (2), the model to be estimated is:
jmtPmt
+ ptRMPi + 8tpostit + Yt(RMPi x postit)
+ X postit ^ T-mt ^jm ^ijmt
Additional interaction terms with transaction year indicators are added to allow pt, 5t, yt, and 0t
to vary freely by year. The coefficient subscripts denote this increased flexibility, but the
interactions with year indicators are not explicitly represented for notational ease.
Similar to equations (3a) through (3c) the percent change in price in a specific ex post year t can
be calculated. Banzhaf (2021) describes these estimates as a direct unmediated effect (DUE). It is
unmediated because all other attributes are held constant, and it is direct because it only considers
movement on the same ex post price surface.
The resulting estimates represent a theoretical upper bound of the monetized welfare loss to nearby
residents from an accident, in a given ex post year t. If these estimates are constant over time, then
more conventional calculations of the capitalization effects have the same welfare interpretation.
V. RESULTS
Results are presented regarding the spatial extent of the treatment effect on house prices, the
estimated price impacts of chemical accidents of different severity and at different distances, an
examination of potential cumulative price impacts from multiple accidents, and the attenuation
of price effects over time. The section ends with an assessment of the parallel trends assumption.
V.A. Determining the Treated and Control Groups
When implementing a spatial DID approach where "treatment" assignment is based on proximity
to an environmental amenity or disamenity, researchers often rely on a strategy first proposed by
Linden and Rockoff (2008), and later adapted by Haninger et al. (2017), Muehlenbachs et al.
(2015), and others. The basic idea is that the pre- and post-treatment event price gradients are first
estimated with respect to distance from the environmental commodity. If the treatment of interest
is believed to have a negative effect (as is the case for a chemical accident), then for homes nearest
the site one would expect the post-treatment gradient to fall below the pre-treatment gradient. As
distance from the disamenity increases, we would expect the post-treatment gradient to gradually
increase, moving towards the pre-treatment gradient. The distance where the two lines converge
marks the average spatial extent of the treatment effect on house prices, and informs the assumed
cutoff point between the treated and control groups.
18
-------
We first estimate a regression based on equation (2), but where separate interaction term vectors
for proximity to a facility pre- and post-accident are included. The uninteracted RMPt and postit
variables are now excluded because they become perfectly collinear with the pre- and post-
accident interaction terms.
(7) ln(pijmt) = xijrntp + y°(RMPi x preit) + y1(RMPi x postit)
+ (RMPi X postit ^ "I" tmt ^/m ^ijmt
Estimating the hedonic model in equation (7) allows us to use the estimates of y° and y1 to graph
the pre- and post-accident price gradients. Furthermore, by summing y1 and 0 we can graph the
post-accident price gradients for different types of accidents (i.e., non-reportable accidents,
reportable accidents resulting in only onsite impacts, and offsite impact accidents). Distance from
the accident site is measured using indicators denoting 250-meter incremental bins, going from 0-
250 m through 9,500-9,750 m. The farthest 9,750-10,000 m bin is the omitted category. Finally,
we note that in this initial diagnostic exercise we do not include interaction terms to allow (3 to
vary by county and year, but the final regression models do allow for such flexibility.
The results are shown in Figure 3. The pre-accident price gradient suggests that in general,
irrespective of an accident occurring, the prices of homes nearest an RMP facility are already
significantly depressed; a finding that is in line with Guignet et al.'s (2022) case study of Michigan,
Ohio, and Pennsylvania. Although this is not necessarily a causal effect, house prices nearest RMP
facilities tend to be lower in value, even when no accident has occurred. For example, homes
within one kilometer are associated with a 3% to 6% decline in price compared to homes in the
farthest distance bin, all else constant. This negative association remains statistically significant (p
< 0.10) out to 2 to 2.5 kilometers from the site.
The top panel (Panel (A)) of Figure 3 shows the price gradient for a nonreportable accident.
Although generally lower, comparison of the post-nonreportable accident price gradient to the pre-
accident gradient suggests little statistically significant effect from nonreportable accidents. This
is not surprising given that no onsite or offsite damages, injuries, etc. resulted from these accidents.
Nearby residents may not generally be aware that such nonreportable accidents even occurred.
Panel (B) of Figure 3 shows how prices are impacted by proximity to an accident that resulted only
in reportable onsite impacts (e.g., injuries or deaths to workers or first responders, or onsite
property damage). The differences between the pre- and post-reportable accident gradients are not
always statistically significant, but we do generally see that prices nearest the site significantly
declined after an accident. As distance increases, the post-accident gradient gradually converges
to the pre-accident gradient.
The price gradient with respect to accidents that resulted in impacts to offsite populations,
properties, and/or the environment provides the clearest evidence of the extent of impacts on house
prices. As shown in Panel (C) of Figure 3, there is a stark decrease in house prices after an offsite
impact accident. This negative effect diminishes with distance, becoming negligible around 5,750
meters from the site.
19
-------
Based on this diagnostic exercise we assume a treated group of homes within 0 to 5,750 meters of
an RMP facility, and a control group of homes 5,750 to 10,000 meters from the same set of RMP
facilities. We discuss the validity of the assumed treated and control groups in section V.D.
VB. Estimated price impacts of chemical accidents
We next estimate a series of hedonic price regressions following equation (2). For the first set of
models (Model 1) we assume that the price effects of interest are homogenous within the 0 to 5,750
meter treatment zone. A binary scalar denoting homes within 0 to 5,750 m of an RMP facility is
used for RMPt. The full hedonic regression results are displayed in Table B.l of Appendix B.
Although not of primary interest, the coefficient estimates corresponding to the house structure
and location characteristics are all significant and of the expected sign and magnitude, lending
credibility to our results (see Model 1A in Table B.l). House prices increase with lot acres, interior
square footage, and the number of stories and bathrooms. House prices decrease with age,
following a quadratic relationship. All else constant, prices are higher in areas where the immediate
vicinity is more developed, and when the home is near a lake or river. Being located within 500 m
of a highway is associated with lower home values.
The ATT estimates following equations (3a) through (3c) are calculated and displayed in Table 3.
Model 1A includes year-by-county and quarter-by-county fixed effects, as well as time-invariant
census tract fixed effects, but constrains the slope coefficients corresponding to the house and
location attributes to be the same over time and across counties. In this initial model we see a
negative and marginally significant 1.26% price decline corresponding to the occurrence of a
nonreportable chemical accident, although this effect is not robust in subsequent models. The
occurrence of a reportable accident that resulted in onsite fatalities, injuries, and/or property
damage leads to an average 2.15% decrease to the price of homes within 5.75 km. An even larger
decline of 3.27% is experienced by homes within 5.75 km of a chemical accident that impacts
offsite populations, property, and/or the environment.
Model IB includes separate county and year interaction terms with the house and location
characteristics, thus allowing the hedonic equilibrium to vary over space and time with respect to
these dimensions. The results suggest no statistically significant effects to surrounding home
values, on average, due to a nonreportable accident or a reportable accident that only led to onsite
impacts. However, Model IB does suggest a significant 2.25% average decrease in the value of
homes within 0 to 5,750 meters of an accident with offsite impacts. This finding is robust to the
inclusion of year-by-county interactions with all home and location characteristics in Model 1C,
suggesting a 1.91% decrease in home values.
20
-------
Figure 3. Pre- and post-accident price gradients.
0.05
Panel (A): Nonreportable accident.
£
—
£
ij
c
U
0
-0.05
-0.1
-0.15
-0.2
-0.25
% % /oo *°o \ % % % % ^% W\ \
U U V G>0 U0 V0 00 UQ Oq U0 Op uQ V0 UQ 00 VQ 00 UQ 00 Gfo 00 Uq 00 U0 00 CJ0 00 U0 VQ O0~V0 VQ U0
Panel (B): Reportable Accident, only onsite impacts.
£
w
c
u
I
o
U
-0.25
0.05
0
-0.05
-0.1
-0.15
-0.2
^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^^ ^ ^
V U U Vq °o uo K*o °o uo °o uo °o uo uo °o uo uo uo °o °o uo uo °o uo °o uo °o0
Panel (C): Reportable Accident with offsite impacts.
-0.25
/>s^°o^s^so^>s3°o^^sd'?;>s /qo
U U U Vq Oq uQ oQ uQ oQ uQ oQ uQ oQ uQ oQ uQ oQ uQ vQ oQ o0 oQ uQ oQ U0 Oq~V0 oQ uQ oQ uQ oQ vQ oQ~u0 oQ vQ oQ uQ^
Meters
Pre-Accident Post-Nonreportable Accident Post-Reportable Accident Post-Offsite Impact Accident
Note: Dotted lines display the 95% confidence intervals.
21
-------
Table 3. Base model results: Percent change in price due to an accident.
Model 1A
0-5,750 meters
Model IB
Model 1C
Nonreportable
Reportable
Offsite Impacts
-1.2604*
(0.7450)
-2.1457***
(0.7966)
-3.2726***
(0.6487)
-0.2150
(0.6320)
-0.8284
(0.6871)
-2.2541***
(0.5692)
0.1695
(0.5537)
-0.7530
(0.5910)
-1.9130***
(0.5982)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Yes
Year x County
Quarter x County
Yes
House x County
House x Year
Year x County
Quarter x County
Yes
House x County x Year
Year x County
Quarter x County
Yes
Observations
Adjusted R2
10,426,638
0.737
10,426,638
0.759
10,426,638
0.767
Note: Average percent change in price to homes within 0-5,750 meters of an accident. * p<0.10, ** p<0.05, *** p<0.01.
Standard errors in parentheses, clustered at the county level. Estimates calculated following equations (3a) through (3c)
using the "nlcom" command in Stata 17/MP, and are based on the coefficient estimates from the hedonic regression
models 1A, IB, and 1C. Note that 1,804 singleton observations were dropped from the regression model. Full regression
results are presented in Table B. 1 in Appendix B.
The magnitude of these average price effects could be considered small, but they are averaged over
a rather large 5.75 km spatial extent. To investigate how these average price effects vary with
distance from the RMP facility, we estimate Model 2, where RMPt is a vector of indicators
denoting whether a facility is within a series of 250-meter incremental bins from a home, starting
with 0 to 250 m, and extending out to 5,550 to 5,750 m. The corresponding percent change in
price estimates are again calculated following equations 3a through 3c.
The results in Figure 4 are based on Model 2B (a variant of Model IB).11 Separate year and county
terms are interacted with the house and location attributes to allow the hedonic equilibrium price
surface to vary across markets, both temporally and spatially. One could consider pursuing a more
flexible model where year-by-county terms are interacted with the house and location
characteristics (as we did in Model 1C). However, this was not our preferred model for two
reasons. First, the key results are similar across both specifications, but the former models with
separate year and county interactions terms are less computationally burdensome to estimate.
Second, given the far-extending 5.75 km price effects from offsite impact chemical accidents, it is
possible that allowing the slope coefficients to vary flexibly over time for each specific county
11 See Appendix B Table B.2 for full regression results of Model 2B as well as Models 2A and 2C (the corresponding
variants based on Models 1A and 1C).
22
-------
may absorb some of the price effects of interest. Nonetheless, the offsite impact accident results
are robust to these alternative specifications (see Figure B. 1 and Figure B. 2 in Appendix B).
Figure 4 shows noticeable spatial heterogeneity, with the patterns over space meeting two key
expectations. First, any negative price effects are only experienced in response to the most severe
accidents - i.e., the offsite impact accidents (see Panel (C)). Second, the point estimates suggest
that the negative price effects from offsite impact accidents are strongest among the nearest homes,
and gradually diminish with distance. Panel (C) in Figure 4 shows that homes nearest an accident
with offsite impacts experience a 4.37% decrease in price, on average. These negative price effects
generally diminish with distance, but remain significant out to 5,750 meters, where homes
experience an average decline of 0.92% after an offsite impact accident.
V C. Cumulative price effects and attenuation over time
To assess how home prices respond to the occurrence of multiple accidents, we estimate Model
3B (a variant of Model IB) following equation (4). The regression model results are presented in
Table B.3 of Appendix B. The slope coefficients corresponding to the number of additional non-
reportable, reportable, and offsite impact accidents are all statistically insignificant. A Wald test
confirms that these three coefficients are also jointly insignificant (p=0.5063), as is the sum of the
three accident count coefficients (p=0.3549). Overall, the results suggest that accidents subsequent
to the first, even if they yield offsite impacts, do not on average have statistically significant
impacts on surrounding home values. A possible explanation is that a first accident involving
offsite impacts such as, for example, an evacuation event, leads to price declines and a partial
turnover in the neighborhood to households with less risk aversion. New residents may choose to
accept the risk in exchange for a discounted house price because they cannot afford otherwise.
Thus, the risks become capitalized in home prices after the first offsite impact accident, and
subsequent accidents do not further depress home values. A different explanation is that household
transactions occurring near multiple accident sites may be in relatively heavily industrialized areas
in which house prices are already discounted to reflect perceived heightened risks of an industrial
accident.
Overall, the inferred percentage change in prices is similar to those estimated from previous
models (see Table B.4 in Appendix B). Nonreportable accidents and those yielding only onsite
reportable impacts result in no significant effect on surrounding home prices, on average, but
homes within 5.75 km of one and even two offsite impact accidents see statistically significant
declines in price - between a 1.5% and 2.0% decrease.
23
-------
Figure 4. Model 2B results: Percent change in price due to an accident, by 250-meter bins.
Panel (A): Nonreportable accident.
10
-10
^°o >so /oon /jo, y> ^
^ J? °o °n vr Xfc y->r <3? 4b >?
o o o o o o o o o o o o o °ooouoouo o
Panel (B): Reportable Accident, only onsite impacts.
oi) 0
K*? '£r>
C(0
% % %a
%
%> "
% 4
0>
'
'O "o
Panel (C): Reportable Accident with oflsite impacts.
°o uo °o °o °o uo °o
'
-------
To examine whether the decline in home values due to an offsite impact accident are long-lasting
or attenuate overtime, we estimate Model 4B (for full results, see Table B.5 in Appendix B), which
is a variant of Model IB based on equation (5). The estimated percent change in price due to an
offsite impact accident is allowed to vary for each year after the accident. As shown in in Figure
B.3, the estimates are generally negative, ranging from -2.06% to a statistical zero. A Wald test
fails to reject the null hypothesis that the negative effects from an offsite impact accident are equal
for the first 15 years after an accident (p=0.2896), suggesting that the negative price effects persist
for at least this 15-year duration. A marginally significant 1.78% appreciation corresponding to 15
to 16 years after an accident is then estimated. This is the last year we are able to observe in our
16-year study period, and so more data are needed to confirm whether the adverse price effects
attenuate at this time, on average, or if this appreciation is just an artifact of the available data.
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 can observe the pretreatment trends and compare the
treated and control groups. If house prices for the two 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 the treatment event.
We conduct an event study by estimating a variant of Model IB and equation (2), but where
interaction terms are included to allow the RMP and accident coefficients of interest to vary by
year 5 relative to the date of the accident (s = 0).
(8) ln(jPijmts) ~ %ijmtPmt Yis=-16{.^sP^^is Ps(.P^^is ^ RMPi)}
+ Is=o{Sspostis + ps(.postis x RMPi) + (postis x RMPt x accit)Os}
Vjm + ymt
Throughout our analysis we find robust evidence of adverse effects on home values due to offsite
impact accidents. Thus, these accidents are the focus of the event study graph in Figure 5. To allow
for a percent change in price interpretation, estimates of ps and 0S from equation (8) are
transformed following equations similar to (3a) and (3c).
The ps coefficients reflect the incremental price difference between the treated group (homes
within 0 to 5,750 meters of an RMP accident site) and the control group (homes located 5,750 to
10,000 meters from the site). Ideally, ps would be equal for all s < 0. In other words, given the
ideal counterfactual group the event study graph would visually show that any differences in the
pre-treatment periods are constant (i.e., the trends are parallel). As can be seen on the left side of
Figure 5, the pre-treatment differences in a given year prior to an accident are generally similar,
25
-------
are not statistically different from zero, and do not demonstrate any clear pattern of differences
that would violate the parallel trends assumption. A Wald test fails to reject the null hypothesis
that the percent change in price estimates in the pre-treatment periods are equal (p = 0.2562),
suggesting that the pre-treatment trends between the treated and control groups are parallel.
Acknowledging that with any non-classical experimental framework, caution is warranted when
making causal inference, this event study supports the parallel trends assumption, and bolsters our
interpretation that offsite impact accidents caused an average 2% to 3% decline in the value of
homes within 5.75 km.
VI. WELFARE IMPLICATIONS
Based on the different versions of Model 1, we estimate that homes within 5.75 km of an offsite
impact accident experience a decline in value ranging from 1.91% to 3.27%. Our middle estimate
(from Model IB) suggests a 2.25% decline. We can calculate the average capitalization effect after
an offsite impact accident as Ap°ff = ^ 0//), where = $232,187 is the average
transaction price after an offsite impact accident and among homes within 0 to 5.75 km, and
%Ap°ff = —2.25% is estimated from Model IB following equation (3c). This suggests an
average loss in value of $5,354 per home. Multiplying this by the 7,383,200 single-family
residences within 5.75 km of at least one of the 661 RMP sites where an offsite impact accident
takes place during our 2004 to 2019 study period suggests a total loss in housing stock value of
over $39.5 billion.
As described in section IV.D, Banzhaf (2021) proposes an adjustment to the more conventional
DID hedonic price regression model, where the hedonic price surface is allowed to temporally
vary with respect to all dimensions. We carry out such an adjustment here in order to facilitate a
more formal welfare comparison. A variant of Model IB is estimated following equation (6). Then,
similar to equation (3c), the percent change in prices for an offsite impact accident is calculated
for each year during our study period. The estimated percent change in prices by year are shown
in black in Figure 6. In this model the price effects of an offsite impact accident are allowed to
vary freely from year to year. The results suggest that a welfare calculation is highly sensitive to
the assumed ex post year. If we choose 2019 (the last year of our study period) as the ex post year,
for example, the welfare loss to residents from an offsite impact accident is not statistically
significant. In contrast, the results would be quite different in a different ex post year, such as 2016
or 2017, for example. A Wald test rejects the null hypothesis that these estimates are statistically
equal for each year from 2004 to 2019 (p=0.0053), but at the same time there is no clear monotonic
trend over time.
26
-------
Figure 5. Event study of the percent difference in price from an offsite impact accident: Based on variant of Model IB.
4
9
* ^
z,
Q.
3
« 2
^ n
//////, T 1 •
>
i_ u
Q_ O
c £:
— c
(D O i
u u _9 <
l
11 11II111 ::"1
C ^
o> o
S_ -M
£ ~o
3= (D
¦— s_
~ ro a
<
>
H
• 1
• 1
1 1
—1
1
\
H
vo Q-
5s E
o
u
c
_Q
o
-17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1
Years relative to
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
date of accident
Note: Graph shows the percent difference in price estimates among the 0 to 5.75 km treated group relative to the 5.75 to 10 km control group, by year relative to
the date of accident.
-------
The key question is whether these year-to-year fluctuations are just noise, or do they reflect true
changes in the hedonic equilibrium with respect to these RMP accident dimensions. We estimate
a variant of Model IB where the RMP and accident coefficients are constrained to only vary
linearly over time. Such a model allows for temporal variation in the hedonic price surface, as
required for Banzhaf s (2021) welfare bounding interpretation, while at the same time minimizing
noise leading to year-to-year fluctuations. More specifically, the estimated model is:
(9) ln(piJmt) = xijmtpmt + pRMPi + pt(RMPi x trendt) + 6postit
+8t(postit x trend^ + Y^RMPi x postit) + x postit x trendt)
+(/?MPj x postit x accit)0 + (RMPt x postit x accit x trendt)0t
Vjm jmt
where trendt is a time trend variable with 0=2004, 1=2005, ..., 15=2019. The blue line fitted in
Figure 6 shows the linear trend of the percent change in price resulting from an offsite impact
accident for each ex post year t, which are calculated as:
(10) %Ap°^ = jexp (y + (.Yt x trendf) + 6^re^ + (oj:rep^ x trend^j +
+ x trendf^jj — lj x 100
where trendy is the corresponding value of the time trend variable for year t, and the superscripts
in brackets denote the elements corresponding to reportable and offsite impact accidents in the
respective coefficient vectors 0 and 0t.
A Wald test fails to reject the null hypothesis that the sum of the trend slope coefficients is equal
to zero - i.e., HO: yt + 0^rep' + = 0 (p=0.3528). The trend slope coefficients are also jointly
insignificant (p=0.1417). Together, these results suggest that the equilibrium hedonic price surface
with respect to RMP accidents is constant over time. Under that assumption we can interpret the
aforementioned capitalization effects from Model IB as an upper bound of the ex post loss in
welfare to residents living within 5.75 km of an RMP accident that resulted in offsite deaths,
injuries, property or environmental damage, and/or the evacuation and sheltering-in-place of
surrounding populations. The average facility where an offsite impact accident occurred has
11,170 single-family homes within 5.75 km. Multiplying this by the $5,354 loss per household
suggests an average welfare loss to surrounding communities of $59,809 million. The median
number of single-family homes within 5.75 km of an offsite impact accident site is 4,646,
suggesting a median loss of $24,877 million. Considering all 7,383,200 single-family homes
around the 661 RMP sites that experienced at least one offsite impact accident, the ex post social
cost to the surrounding communities is substantial, suggesting a total loss of $39,533 billion.
Following Banzhaf s (2021) theoretical framework, and our assumption that the hedonic price
surface is constant over time with respect chemical accidents, then the formal interpretation is that
these results represent an upper bound of the ex post welfare loss to nearby residents. However,
our estimated per home price impacts of 2% to 3% are fairly small, and Banzhaf established that
-------
the bounding estimates approach the true value for small changes, with the upper bound
approximating the true value as the environmental shock approaches a marginal change.
Furthermore, our estimates only account for residents living in single-family homes, and therefore
disregard impacts to residents living in other types of housing (e.g., multi-family apartment
buildings and condos, townhomes, etc.) and impacts to businesses and others in the community.12
Empirically speaking, it is ambiguous whether our estimates are a lower or upper bound of the ex
post external costs imposed on surrounding communities due to offsite impact accidents. Finally,
additional work is needed to formalize an ex ante welfare interpretation of the results from DID
hedonic applications. Banzhaf (2021) suggests that movement along the ex ante price surface may,
in our context, provide a lower bound of the loss in welfare to nearby residents.
Figure 6. Percent change in price estimates from offsite impact accidents by year: Based on
variants of Model IB.
6
Year
Note: Black dots are the percent change in price estimates from a variant of Model IB following equation (6). The error bars
represent the 95% confidence intervals. The solid blue line represents the percent change in price estimates from a more
restrictive linear version of the model following equation (9), and are then calculated as per equation (10). The dotted blue line
represents the 95% confidence interval. All estimates are derived using the "nlcom" command in Stata 17/MP.
VII. CONCLUSION
Our analysis of industrial chemical accidents across the contiguous U.S. reveals mixed evidence
as to whether accidents resulting in minimal impacts, or where the impacts (e.g., injuries, deaths,
and property damage) were confined to the industrial property itself, affect home values, with
estimated losses ranging from 0% to 2%. However, we find robust, causal evidence that accidents
yielding direct impacts to the surrounding community significantly affect home prices. Such
accidents resulted in health impacts to nearby residents, offsite property damage and
12 Welfare losses may also be experienced by individuals who do not reside in the nearby community; for example,
by the employees at the facility, the facility owners, the emergency responders, people who visit the community, and
so on.
29
-------
environmental degradation, and/or people being evacuated or sheltered-in-place to avoid harm.
Although the average capitalization effect of 2% to 3% is somewhat small in magnitude, this effect
extends 5.75 km from the industrial facility, which is quite far compared to studies of similar
disamenities. Past literature generally found stronger local impacts, extending only a few hundred
meters and up to 3 km (e.g., Gamper-Rabindran and Timmins, 2013; Guignet et al., 2018; Guignet
and Nolte, 2021; Haninger et al., 2017; Muehlenbachs et al., 2015). We find that the adverse price
effects from these most severe chemical accidents persist for at least 15 years on average, and
possibly longer. Further analysis and data covering a longer study period are needed.
We do not find evidence that losses in value are systematically greater among homes that
experience multiple offsite impact accidents. Perhaps the first accident led home values to fully
capitalize risks, and additional accidents did not yield new information to update residents'
perceptions. It is also possible that less risk averse residents, who might face pressing priorities
and financial constraints, moved in after the first accident; or that multiple accidents occur in
communities hosting multiple industrial facilities and perceived baseline risk is already built in to
home values. Further research should explore the socioeconomic characteristics of communities
experiencing multiple accidents.
We adapt the procedure proposed by Banzhaf (2021) to estimate a formal upper bound of the ex-
post loss in welfare to residents living within 5.75 km of an offsite impact accident. We find that
such welfare calculations are extremely sensitive to the assumed ex post year. A model that
restricts the accident impacts to vary linearly over time suggests that the price effects of accidents
are constant. Assuming that the price effects of accidents are constant over our study period allows
for a welfare interpretation of the estimated capitalization effects. The formal interpretation is that
the estimated effects represent a theoretical upper bound of the ex post welfare loss, yet the smaller
the incremental impact, the closer the estimates are to a true loss. Our estimate of an average 2%
to 3% price change is fairly small, suggesting the estimates are close to the true losses to residents
living near an offsite impact accident. Additionally, our estimates only account for residents living
in single-family homes, and therefore disregard impacts to residents living in other types of
housing, as well as impacts to businesses and others in the community.
Our preferred model specification suggests an average loss of about $5,350 per household.
Considering the 7,383,200 single-family homes within 5.75 km of one of the 661 offsite impact
accident sites across the contiguous U.S., this implies a social cost to these nearby residents of
$39.5 billion. This translates to an average loss of $59.8 million for each site where an offsite
impact accident occurs. It is clear that the external costs to fence line communities of these most
severe industrial chemical accidents are substantial and critical to account for in benefit-cost
analyses used to inform policy and management decisions. This includes the recently proposed
"Safer Communities by Chemical Accident Prevention" rule (US EPA 2022a), which is intended
to further protect communities and the local environment. Among the criticisms of this proposal
was the lack of estimates of social benefits alongside significant estimates of costs (InsideEPA
2022).
This study demonstrates that large-scale nationwide benefits analyses are critical to inform equally
as extensive federal policy. Such analyses would be difficult without a widely available, and fairly
30
-------
accurate and consistent dataset like Zillow's ZTRAX database. These data allowed us to analyze
home prices around almost the entire population of chemical accidents reported to the RMP
program. Prior to the availability of the ZTRAX data, studies on similar EPA programs and
disamenities were limited mainly to local-scale case studies (e.g., Michaels and Smith, 1990;
Kohlhase, 1991; Flower and Ragas, 1994; Kiel, 1995; Carroll et al., 1996; Kiel and Zabel, 2001;
Hansen et al., 2006; Zabel and Guignet, 2012; Guignet, 2013; Liao et al., 2022). Any studies
attempting nationwide coverage were often spotty in nature (e.g, Kiel and Williams, 2007; Guignet
et al., 2018), or were forced to use spatially and temporally coarse data (e.g., Gamper-Rabindran
and Timmins, 2013; Greenstone and Gallagher, 2008).
There are three additional important benefits of using large-scale datasets like that provided by
Zillow's ZTRAX program. First, identifying the impacts of environmental commodities on house
prices is of primary interest to environmental economists, but such attributes often yield a
relatively small contribution to the overall price of a home. Such is the case here, where we
estimate a 2% to 3% average decline in value for each home near an offsite impact accident. While
the estimated per home price impacts are close to zero, in aggregate, the effects are huge. There is
substantial difficulty in statistically distinguishing these price impacts from zero, especially
considering the numerous other, often spatially correlated, location attributes that affect house
prices. Although the overall sample size may be reasonable, smaller case studies focused on a
municipality, county, or even a state or multi-state region, may not have a large enough number of
identifying observations to precisely estimate such effects. Our results showing a 2% to 3%
decrease in nearby home prices is estimated with remarkable precision. A smaller case study
resulting in similar point estimates could well dismiss the findings as null because they would be
less precise and potentially statistically insignificant. In addition, there is likely heterogeneity in
the price impacts across markets, and estimates from smaller case studies may not be representative
of the nationwide effects.
Second, stacked spatial DID study designs like ours and countless others in the literature rely on
spatially and temporally dispersed sub-experiments for statistical identification.13 In our context,
a causal interpretation hinges on the assumption that any unobserved influences on house prices
are not correlated with the location and timing of an accident. The plausibility of such an
assumption increases as we observe higher numbers of accidents at different locations and periods
in time. Spatially and temporally extensive datasets like ZTRAX facilitate the inclusion of high
numbers of treatment events and locations, and therefore reduce endogeneity concerns related to
spatially correlated confounders.
A final advantage of large-scale datasets like ZTRAX is that the large number of identifying
observations allows researchers to examine treatment heterogeneity in more detail. In our context,
this enables several useful directions, including heterogeneity with respect to accident severity,
distance, and time. Although we find statistically significant price effects extending out to 5.75
km, we are able to examine heterogeneity in those price impacts with respect to distance from the
disamenity at a fine 250-meter bin resolution. We also examine how price impacts evolve over
13 See Parmeter and Pope (2013) and Guignet and Lee (2021) for reviews.
31
-------
time, based on one-year bins denoting time since the accident. "Slicing" the data into finer groups
based on space and time like this would not be possible without the large number of identifying
observations due to the extensive study area and time period afforded by the ZTRAX data.
In that same vein, our welfare analysis following Banzhaf (2021) relies squarely on our ability to
model separate treatment effects by year. Identifying year-by-year treatment effects and inferring
bounding welfare estimates from first-stage hedonic models is novel, and to our knowledge has
only been investigated in two other studies - the original proposal by Banzhaf (2021) and an
application to hazardous waste site cleanups by Guignet and Nolte (2021). Estimating formal
welfare effects would be difficult in many applications without a large-scale dataset containing a
high number of identifying observations.
A formal welfare interpretation is necessary for including estimates from hedonic property value
studies in benefit-cost analyses and to inform efficient policy decisions. The estimates from our
analysis will help inform decision makers regarding future policies directed at reducing the
probability of accidents at chemical facilities, and at RMP facilities in particular. To fully, and
quantitatively, incorporate these estimates into regulatory analysis, however, further research is
needed on how different regulatory requirements (e.g., third party audits, or employee "stop work"
provisions) impact the probability of accidents.
32
-------
WORKS CITED
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.
Banzhaf, H. Spencer and Eleanor McCormick. 2012. Moving beyond Cleanup: Identifying the
Crucibles of Environmental Gentrification. In S. H. Banzhaf (Ed.) The Political Economy
of Environmental Justice. Palo Alto, CA. Stanford University Press.
Banzhaf, H. Spencer. 2021. Difference-in-Differences Hedonics. Journal of Political Economy
129(8): 2385-2414.
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.
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.
Bishop, Kelly C., Nicolai V. Kuminoff, Spencer Banzhaf, Kevin J. Boyle, Kathrine von
Gravenitz, Jaren C. Pope, V. Kerry Smith, Christopher D. Timmins. 2020. Best practices
in using hedonic property value models for welfare measurement. Review of
Environmental Economics and Policy 14(2): 260-281.
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.
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.
33
-------
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.
Dewitz, J., 2019. National Land Cover Database (NLCD) 2016 Products. U.S. Geological
Survey data release,, https://doi.ore/10.5066/P96HHBIE.
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.
34
-------
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.
Grainger, Corbett A. 2012. The distributional effects of pollution regulations: Do renters fully
pay for cleaner air? Journal of Public Economics. 96: 840-852.
Greenstone, Michael, and Justin Gallagher. 2008. Does Hazardous Waste Matter? Evidence from
the Housing Market and the Superfund Program. The Quarterly Journal of Economics,
123(3): 951-1003.
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. 2013. What Do Property Values Really Tell Us? A Hedonic Study of Pollution
from Underground Storage Tanks. Land Economics 89(2): 211-226.
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 R. Jenkins, James Belke, and Henry Mason. 2022. The Property Value
Impacts of Industrial Chemical Accidents. US EPA National Center for Environmental
Economics Working Paper, 22-04, Washington, DC, September.
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.
35
-------
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.
InsideEPA, 2022. Proposed RMP Rule's Costs, Lack of Benefits Likely to Spark Litigation."
August 25. Accessed at https://insideepa.com/daily-news/proposed-rmp-rule-s-costs-lack-
benefits-likely-spark-litigation.
Kiel, Katherine. 1995. Measuring the Impact of the Discovery and Cleaning of Identified
Hazardous Waste Sites on House Values. Land Economics 71(4): 428-435.
Kiel, Katherine A. and Michael Williams. 2007. The impact of Superfund sites on local property
values: Are all sites the same? Journal of Urban Economics 61: 170-192.
Kiel, Katherine and Jeffrey Zabel. 2001. Estimating the Economic Benefits of Cleaning Up
Superfund Sites: The Case of Woburn, Massachusetts. Journal of Real Estate Finance
Economics 22(2): 163-184.
Klaiber, H. Allen, and V. Kerry Smith. 2013. Quasi-Experiments, Hedonic Models, and Estimating
Trade-offs for Local Amenities. Land Economics 89(3): 413-431.
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.
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 of Hazardous Materials.
115: 9-16.
Kohlhase, Janet. 1991. The Impact of Toxic Waste Sites on Housing Values. Journal of Urban
Economics 30: 1-26.
Kuminoff, Nicolai V., and Jaren C. Pope. 2014. Do 'Capitalization Effects' for Public Goods
Reveal the Public's Willingness to Pay. International Economic Review 55(4): 1227-1250.
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.
36
-------
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.
Mastromonaco, Ralph. 2015. Do Environmental Right-to-Know Laws Affect Markets?
Capitalization of Information in the Toxic Release Inventory. Journal of Environmental
Economics and Management 71: 54-70.
Messer, Kent D.„ William D. Schulze, Katherine F. Hackett, Trudy A. Cameron, Gary H.
McClelland, Can stigma explain large property value losses? The psychology and
economics of Superfund, Environmental and Resource Economics 33 (2006) 299-324.
Michaels, R. Gregory and V. Kerry Smith. 1990. Market Segmentation and Valuing Amenities
with Hedonic Models: The Case of Hazardous Waste Sites. Journal of Urban Economics
28: 223-242.
Moulton, Jeremy G., Nicholas J. Sanders, and Scott Wentland. 2018. Toxic Assets: How the
Housing Market Responds to Environmental Information Shocks. Unpublished
manuscript, accessed 26 Aug 2022
at https://www.aeaweb.ore/conference/2019/preliminarv/paper/r99KZF6H..
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. 2020. High-resolution land value maps reveal underestination of conservation
costs in the United States. Proceedings of the National Academny of Sciences 117: 29577-
29583.
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.
Parmeter, Christopher F. and Jaren C. Pope. 2013. Quasi-experiments and Hedonic Property Value
Methods. In Handbook on Experimental Economics and the Environment, eds. John A.
List and Michael K. Price. Cheltenham, United Kingdom: Edward Elgar Publishing.
37
-------
Petrolia, Daniel R., Dennis Guignet, John Whitehead, Cannon Kent, Clay Caulder, and Kelvin
Amon. 2021. Nonmarket Valuation in the Environmental Protection Agency's Regulatory
Process. Applied Economic Perspectives and Policy 43(3): 952-969.
Rosen, Sherwin. 1974. Hedonic Prices and Implicit Markets: Product Differentiation in Pure
Competition. The Journal of Political Economy 82(1): 34-55.
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.eithub.io/assets/files/DiD Review Paper.pdf.
Sun, Liyang, and Sarah Abraham. 2021. Estimating Dynamic Treatment Effects in Event Studies
with Heterogeneous Treatment Effects. Journal of Econometrics 225(2): 175-199.
Taylor, Laura O., Daniel J. Phaneuf, and Xiangping Liu. 2016. Disentangling property value
impacts of environmental contamination from locally undesirable land uses: Implications
for measuring post-cleanup stigma." Journal of Urban Economics 93: 85-98.
US Census Bureau, 2019. 2019 TIGER/Line Shapefiles (machine-readable data files) [WWW
Document], URL https://www.census.gov/cgi-bin/geo/shapefiles/index.php (accessed
5.19.20).
U.S. Census Bureau. 2016. Census Bureau Projects U.S. and World Populations on New Year's
Day. Accessed at: https://www.censiis.eov/newsroom/press-releases/2016/cbl6-
tj tin 1. 9 Sept. 2021.
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-I-TX, May.
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.
U.S. EPA (Environmental Protection Agency). 2022a. Safer Communities by Chemical Accident
Prevention Proposed Rule. Office of Land and Emergency Management. Washington,
D.C. April 19. Accessed at: https://www.regulations.gov/document/EPA-HQ-OLEM-
2022-0174-0093
U.S. EPA (Environmental Protection Agency). 2022b. Regulatory Impact Analysis: Safer
Communities by Chemical Accident Prevention Proposed Rule. Office of Land and
38
-------
Emergency Management. Washington, DC. April 19. Accessed at:
https://www.reeiilations.eov/dociiment/EPA-HQ-OLEM-2Q22 3.
Wentland, Scott A., Zachary H. Ancona, Kenneth J. Bagstad, James Boye, Julie L. Hass, Marina
Gindelsky, Jeremy G. Moulton. 2020. Accounting for land in the United States:
Integrating physical land cover, land use, and monetary valuation. Ecosystem Services,
46: 101178.
Zabel, Jeffrey and Dennis Guignet. 2012. A Hedonic Analysis of the Impact of LUST Sites on
House Prices. Resource and Energy Economics 34(4): 549-564.
39
-------
ONLINE APPENDIX
Two online appendices are included. Appendix A presents the data cleaning and formatting
procedures, followed by supplemental descriptive statistics. Appendix B provides full hedonic
property value regression results for models reported in the main paper, as well as supplemental
results as a sensitivity analysis.
Appendix A. Data formatting details and supplemental data statistics
The hedonic analysis focuses on a total of 10,428,442 full property14, arms-length transactions of
single-family homes from 2004 through 2019, where the home is within 10 km of one or more of
the RMP facility accidents. When estimating the hedonic price regression models, 1,804 singleton
observations were dropped due to the inclusion of numerous spatiotemporal fixed effects.
Therefore, a sample size of 10,426,638 sales is reported in the regression results tables.
The transaction and assessor data for all available states in the contiguous U.S. were obtained
through Zillow's ZTRAX program. For this analysis we use the October 2021 release of the data
(downloaded on 19 May 2022). We do not rely on the geographic coordinates provided by
ZTRAX. Price impacts associated with RMP facility accidents can be local in nature, and so the
highest level of spatial precision possible is desired. At the same time, there are documented
concerns regarding the accuracy of the geographic coordinates provided directly in ZTRAX,
including missing data, mislocated data, and undocumented spatial variation in the geographic
coordinate datums (Nolte et al., 2021). We therefore relied on geo-located parcel boundary
polygons, which we obtained partially from open-access data sources and partially from Regrid
(www.reerid.com) through their "Data with Purpose" program. We used unique parcel identifiers
(assessor parcel numbers) to link these parcel boundary data to parcel records in the ZTRAX tax
assessor database using county and town-specific deductive string pattern matching and
geographic quality controls (Nolte, 2020). The Euclidean distance from the centroid of each
residential parcel in ZTRAX to each RMP accident site is calculated.
Our data cleaning and formatting starts with the 15,184,233 full property transactions of single-
family homes, where the homes are located within 10 km of at least one RMP accident. Single-
family homes were identified as those for which the land use code in the ZTRAX assessor database
was RR000, RR101, RR102, or RR999. Transactions with missing nominal sales price or with
token values of $1, $100, or $1,000 are eliminated, leaving a sample of 15,153,958. Transaction
prices are converted to 2021$ USD based on the Bureau of Labor Statistics annual US city average
"All Urban Consumers" consumer price index (CPI).15
Outlier observations with a real price less than $15,000 or greater than $1,000,000 are eliminated,
as are transactions of homes with lot sizes less than 0.05 acres or greater than 2 acres, leaving a
sample of 13,567,656 home transactions. These outlier cutoffs fall squarely between the lowest
and highest 1st and 5th percentiles; for example, the $15,000 value is between the lowest 1st and 5th
percentiles of the price distribution. Homes with less than one story or greater than three stories,
14 Partial sales (i.e., transactions where just a portion of a parcel is sold) are disregarded.
15 Bureau of Labor Statistics (BLS). https://www.bls.gov/cpi/tables/siipplemental-files/liistorical-cpi-ii-202206.txlf.
accessed 31 July 2022.
40
-------
less than 750 square feet or greater than 15,000 square feet, and/or less than one full bathroom or
greater than 4.5 bathrooms, were also eliminated, leaving a sample of 13,066,840 home sales.
These outlier cutoffs correspond approximately to the highest and lowest percentiles.
The age of the home at the time of a transaction was calculated as the difference between the year
of transaction and the effective year built variable in the ZTRAX assessor database. Sales where
the effective year built variable was missing (566,334) or where the calculated age was negative
(357,559) were recoded with an age of zero, and a companion missing age dummy was included.
Among transactions where the age of the home is not missing, we drop 87,953 sales where the
home age was greater than 120 years (which closely corresponds to the 99th percentile of 118
years).
The sample at this point entailed 12,860,064 unique single-family home transactions. It is with
this "cleaned" sample that the usual hedonic analysis might proceed. However, Nolte et al. (2021)
took great care in going above and beyond the usual hedonic property value study data protocols
to more confidently identify arms-length transactions in the ZTRAX data. Following the criteria
developed by Nolte et al. (2021), we drop an additional 2,431,622 sales (18.91% of the sample)
where there is low-confidence that the observations reflect an arms-length transaction. To
accomplish this we created an aggregated transaction filter that represents the lowest level of
confidence determined across all individual filter items, shown in Table A.l. Sales were dropped
when the aggregated filter had a value of zero, meaning at least one of the individual filters was
designated as "low confidence" following Nolte et al. (2021). As shown in Table A.l, Nolte et
al.'s (2021) efforts to identify how transaction contract types are used differently across states
contributed the most to this additional data cleaning step, which ultimately provided a final dataset
that is more confidently focused on arms-length transactions (see Table A.l).
The hedonic analysis in the main paper focuses on the final sample of 10,428,442 full, arms-length
transactions of single-family homes. Descriptive statistics and additional details of this sample are
presented in section III.B of the main text.
Table A.l. Comparison of sample to data filters developed by Nolte et al. (2021).
Confidence that arms-length sale
Transaction Filter High (=2) Medium (=1) Low (=0)
Aggregated filter: Min of individual filters
62.34%
18.75%
18.91%
Similarity between buyer and seller names
96.85%
1.23%
1.92%
Intra-family transaction flag in ZTRAX
98.38%
0%
1.62%
Public buyer and/or seller
99.42%
0.09%
0.49%
Type of transaction contract
84.04%
2.82%
13.14%
Type of mortgage loan
98.80%
0.46%
0.74%
Source of sales price value
81.72%
16.07%
2.21%
Note: Percentage of n= 12,860,064 home transactions categorized as high, medium, or low confidence that they reflect an
arms-length transaction. Transactions designated as "low confidence" under the aggregated filter are dropped from the
final sample. Additional details can be found at https://placeslab.org/ztrax (accessed 15 September 2022), and are further
described by Nolte et al. (2021).
41
-------
Figure A.l. Number of Pre- and Post-Accident Transactions.
I Pre-accident
I Post-accident
500,000
450,000
400,000
350,000
a; '
™ 300,000
<4—
° 250,000
-------
Appendix B. Supplemental analysis results.
Table B.l. Base models: Full hedonic property value regression results.
RMP 0-5750 m*
Post-accident*
Post-accident x 0-5750 m '
Post-Reportable
Accident x 0-5750 m*
Post-Offsite Impact
Accident * 0-5750 m*
ln(acres)
Missing: Acres'
Stories
Missing: Stories*
Bathrooms
Missing: Bathrooms*
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age*
Model 1A
0.0018
(0.0044)
0.0110*
(0.0062)
-0.0127*
(0.0075)
-0.0090
(0.0102)
-0.0116
(0.0088)
0.1210***
(0.0048)
-0.3316***
(0.0286)
Model IB
-0.0023
(0.0035)
0.0102*
(0.0055)
-0.0022
(0.0063)
-0.0062
(0.0087)
-0.0145*
(0.0080)
Model 1C
-0.0036
(0.0033)
0.0093*
(0.0050)
0.0017
(0.0055)
-0.0093
(0.0077)
-0.0118*
(0.0071)
0.0482***
(0.0061)
0.0654***
(0.0171)
0.0754***
(0.0067)
0.1966***
(0.0244)
0.4200***
(0.0153)
3.4506***
(0.1370)
-0.0066***
(0.0004)
0.0000***
(0.0000)
-0.8707***
(0.0344)
-------
% Land Developed w/in 0-5 00m
Highway w/in 500m '
Lake w/in 500m '
River w/in 250m'
Constant
0.0009***
(0.0001)
-0.0244***
(0.0018)
0.0456***
(0.0047)
0.0313***
(0.0087)
9.0290***
(0.1156)
12.1677* *:
(0.0039)
12.1681**=
(0.0037)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Yes
Year x County
Quarter x County
Yes
House x County
House x Year
Year x County
Quarter x County
Yes
House x County x Year
Year x County
Quarter x County
Yes
Observations
10,426,638
10,426,638
10,426,638
Adjusted R-squared 0.737 0.759 0.767
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
Table B.2. Hedonic property value regression results with incremental 250-meter bins.
Model 2A Model 2B Model 2C
RMPf
0-250m
-0.0908***
-0.0867***
-0.0943**
(0.0187)
(0.0185)
(0.0187)
250-500m
-0.0759***
-0.0693***
-0.0706**
(0.0144)
(0.0130)
(0.0139)
500-750m
-0.0683***
-0.0617***
-0.0625**
(0.0117)
(0.0105)
(0.0104)
750-1000m
-0.0534***
-0.0506***
-0.0498**
(0.0110)
(0.0095)
(0.0092)
1000-1250m
-0.0390***
-0.0384***
-0.0381**
(0.0094)
(0.0077)
(0.0076)
1250-1500m
-0.0347***
-0.0323***
-0.0343**
(0.0084)
(0.0073)
(0.0071)
1500-1750m
-0.0417***
-0.0347***
-0.0374**
(0.0077)
(0.0068)
(0.0067)
44
-------
1750-2000m
-0.0364***
-0.0329***
-0.0347**
(0.0079)
(0.0066)
(0.0064)
2000-2250m
-0.0357***
-0.0329***
-0.0346**
(0.0072)
(0.0063)
(0.0061)
2250-2500m
-0.0324***
-0.0308***
-0.0317**
(0.0067)
(0.0056)
(0.0055)
2500-2750m
-0.0271***
-0.0287***
-0.0297**
(0.0060)
(0.0051)
(0.0049)
2750-3000m
-0.0208***
-0.0235***
-0.0252**
(0.0062)
(0.0055)
(0.0054)
3000-3250m
-0.0117*
-0.0153***
-0.0178**
(0.0060)
(0.0050)
(0.0048)
3250-3500m
-0.0082
-0.0115**
-0.0140**
(0.0056)
(0.0046)
(0.0044)
3500-3750m
-0.0057
-0.0090**
-0.0122**
(0.0055)
(0.0045)
(0.0043)
3750-4000m
-0.0030
-0.0063
-0.0087*H
(0.0053)
(0.0046)
(0.0044)
4000-4250m
-0.0032
-0.0046
-0.0063
(0.0052)
(0.0042)
(0.0040)
4250-4500m
-0.0058
-0.0064
-0.0070*
(0.0048)
(0.0039)
(0.0037)
4500-4750m
-0.0047
-0.0059
-0.0072*
(0.0047)
(0.0039)
(0.0037)
4750-5000m
0.0034
0.0009
-0.0002
(0.0047)
(0.0038)
(0.0036)
5000-5250m
0.0025
-0.0003
-0.0016
(0.0042)
(0.0035)
(0.0032)
5250-5500m
0.0034
0.0019
0.0005
(0.0040)
(0.0034)
(0.0030)
5500-5750m
0.0007
-0.0013
-0.0026
(0.0035)
(0.0028)
(0.0027)
Post-accidcnt'
0.0093
0.0095*
0.0087*
(0.0062)
(0.0055)
(0.0050)
0-250m
-0.0129
0.0071
0.0166
45
-------
250-500m
500-750m
750-1000m
1000-1250m
1250-1500m
1500-1750m
1750-2000m
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
(0.0411)
0.0087
(0.0219)
0.0087
(0.0180)
-0.0107
(0.0158)
¦0.0295**
(0.0143)
-0.0296*
(0.0154)
-0.0195*
(0.0116)
-0.0212*
(0.0116)
-0.0141
(0.0105)
-0.0135
(0.0093)
-0.0104
(0.0088)
-0.0053
(0.0093)
-0.0087
(0.0090)
¦0.0179**
(0.0089)
-0.0140
(0.0088)
¦0.0170**
(0.0082)
¦0.0204**
(0.0082)
¦0.0192**
(0.0090)
-0.0119
(0.0083)
-0.0098
(0.0073)
(0.0372)
0.0262
(0.0178)
0.0268*
(0.0152)
0.0081
(0.0139)
-0.0102
(0.0120)
-0.0106
(0.0137)
-0.0058
(0.0103)
-0.0030
(0.0100)
0.0010
(0.0093)
-0.0000
(0.0085)
0.0002
(0.0084)
0.0045
(0.0085)
0.0054
(0.0077)
-0.0052
(0.0076)
-0.0038
(0.0075)
-0.0068
(0.0070)
-0.0107
(0.0070)
-0.0080
(0.0074)
-0.0031
(0.0071)
-0.0023
(0.0063)
(0.0373)
0.0342**
(0.0172)
0.0309**
(0.0144)
0.0081
(0.0130)
-0.0084
(0.0114)
-0.0054
(0.0129)
-0.0004
(0.0099)
0.0008
(0.0098)
0.0036
(0.0090)
0.0022
(0.0078)
0.0030
(0.0077)
0.0078
(0.0079)
0.0081
(0.0075)
-0.0012
(0.0070)
0.0020
(0.0070)
-0.0020
(0.0061)
-0.0069
(0.0061)
-0.0057
(0.0068)
0.0003
(0.0064)
0.0008
(0.0055)
-------
5000-5250m
-0.0076
-0.0000
0.0032
(0.0067)
(0.0059)
(0.0053)
5250-5500m
-0.0111*
-0.0051
-0.0023
(0.0063)
(0.0056)
(0.0051)
5500-5750m
-0.0091
-0.0044
-0.0025
(0.0058)
(0.0051)
(0.0046)
ost-Reportable Accident
0-250m
-0.0654
-0.0421
-0.0498
(0.0509)
(0.0462)
(0.0450)
250-500m
-0.0738**
-0.0577*
-0.0661*^
(0.0366)
(0.0313)
(0.0287)
500-750m
-0.0505*
-0.0415*
-0.0468*^
(0.0286)
(0.0238)
(0.0230)
750-1000m
-0.0140
-0.0077
-0.0100
(0.0228)
(0.0192)
(0.0176)
1000-1250m
0.0020
0.0023
-0.0023
(0.0204)
(0.0174)
(0.0156)
1250-1500m
-0.0038
-0.0004
-0.0026
(0.0197)
(0.0171)
(0.0163)
1500-1750m
-0.0044
-0.0008
-0.0023
(0.0171)
(0.0151)
(0.0149)
1750-2000m
-0.0046
-0.0031
-0.0037
(0.0156)
(0.0138)
(0.0135)
2000-2250m
-0.0062
-0.0076
-0.0085
(0.0165)
(0.0144)
(0.0139)
2250-2500m
0.0017
0.0010
0.0001
(0.0146)
(0.0131)
(0.0122)
2500-2750m
-0.0032
-0.0007
-0.0030
(0.0142)
(0.0130)
(0.0125)
2750-3000m
-0.0122
-0.0074
-0.0083
(0.0141)
(0.0127)
(0.0118)
3000-3250m
-0.0057
-0.0074
-0.0064
(0.0140)
(0.0124)
(0.0120)
3250-3500m
-0.0025
-0.0030
-0.0036
(0.0140)
(0.0123)
(0.0116)
3500-3750m
-0.0056
-0.0028
-0.0042
(0.0131)
(0.0112)
(0.0106)
3750-4000m
-0.0016
0.0004
-0.0002
(0.0121)
(0.0102)
(0.0093)
4000-4250m
0.0033
0.0040
0.0029
(0.0114)
(0.0093)
(0.0086)
4250-4500m
0.0077
0.0064
0.0057
(0.0123)
(0.0102)
(0.0099)
4500-4750m
-0.0050
-0.0033
-0.0047
-------
4750-5000m
5000-5250m
5250-5500m
5500-5750m
Post-Offsite Impact Accident^
(0.0111)
-0.0125
(0.0097)
-0.0143
(0.0092)
-0.0061
(0.0085)
-0.0018
(0.0074)
(0.0093)
-0.0109
(0.0080)
-0.0116
(0.0078)
-0.0053
(0.0068)
0.0014
(0.0062)
(0.0088)
-0.0113
(0.0075)
-0.0123*
(0.0073)
-0.0055
(0.0062)
0.0009
(0.0057)
0-250m
0.0113
-0.0097
0.0033
(0.0434)
(0.0380)
(0.0380)
250-500m
0.0108
-0.0063
0.0006
(0.0354)
(0.0304)
(0.0278)
500-750m
-0.0116
-0.0232
-0.0155
(0.0277)
(0.0245)
(0.0231)
750-1000m
-0.0372*
-0.0463**
-0.0397*^
(0.0212)
(0.0181)
(0.0171)
1000-1250m
-0.0244
-0.0241
-0.0173
(0.0186)
(0.0165)
(0.0151)
1250-1500m
-0.0179
-0.0224
-0.0183
(0.0168)
(0.0153)
(0.0145)
1500-1750m
-0.0135
-0.0200
-0.0172
(0.0162)
(0.0148)
(0.0142)
1750-2000m
-0.0069
-0.0141
-0.0125
(0.0151)
(0.0139)
(0.0132)
2000-2250m
-0.0176
-0.0182
-0.0143
(0.0145)
(0.0130)
(0.0124)
2250-2500m
-0.0218
-0.0234*
-0.0205*
(0.0138)
(0.0127)
(0.0122)
2500-2750m
-0.0103
-0.0152
-0.0105
(0.0136)
(0.0121)
(0.0117)
2750-3000m
-0.0186
-0.0257**
-0.0237*^
(0.0132)
(0.0121)
(0.0116)
3000-3250m
-0.0251*
-0.0300**
-0.0272*^
(0.0134)
(0.0125)
(0.0121)
3250-3500m
-0.0106
-0.0143
-0.0127
(0.0125)
(0.0116)
(0.0109)
3500-3750m
-0.0118
-0.0188*
-0.0167
(0.0122)
(0.0113)
(0.0111)
3750-4000m
-0.0096
-0.0146
-0.0143
(0.0105)
(0.0096)
(0.0093)
4000-4250m
-0.0077
-0.0131
-0.0124
(0.0111)
(0.0104)
(0.0104)
48
-------
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
ln(acres)
Missing: Acres'
Stories
Missing: Stories'
Bathrooms
Missing: Bathrooms^
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age'
% Land Developed w/in 0-5 00m
Highway w/in 500m '
Lake w/in 500m '
River w/in 250m'
Constant
-0.0023
-0.0081
-0.0087
(0.0104)
(0.0095)
(0.0090)
0.0015
-0.0045
-0.0052
(0.0095)
(0.0084)
(0.0081)
-0.0031
-0.0079
-0.0091
(0.0092)
(0.0075)
(0.0073)
0.0022
-0.0041
-0.0046
(0.0082)
(0.0067)
(0.0064)
-0.0001
-0.0053
-0.0058
(0.0070)
(0.0060)
(0.0055)
0.0011
-0.0062
-0.0054
(0.0063)
(0.0053)
(0.0048)
0.1203***
(0.0048)
-0.3302***
(0.0286)
0.0482***
(0.0061)
0.0653***
(0.0171)
0.0753***
(0.0067)
0.1962***
(0.0244)
0.4198***
(0.0153)
3 4492***
(0.1369)
-0.0065***
(0.0004)
0.0000***
(0.0000)
-0.8704***
(0.0344)
0.0009***
(0.0001)
-0.0238***
(0.0018)
0.0458***
(0.0047)
0.0317***
(0.0087)
9.0375***
12.1750***
12.1753**
-------
(0.1155)
(0.0041)
(0.0039)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Yes
Year x County
Quarter x County
Yes
House x County
House x Year
Year x County
Quarter x County
Yes
House x County x Year
Year x County
Quarter x County
Yes
Observations
Adjusted R-squared
10,426,638
0.737
10,426,638
0.759
10,426,638
0.767
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
50
-------
Figure B.l. Model 2A results: Percent change in price due to an accident, by 250-meter bins.
Panel (A): Nonreportable accident.
: - -
% \ \ \ \ % \ \ \ \
"% %
un °n un or
'O uO
Panel (C): Reportable Accident with offsite impacts.
** ** * «£ JcP n> of O 6>
$ ^ ^ .f s- $ ..?•
-10
%> "%> /eb *%¦ /jb ^ J(b "% % %¦ Vjb %l. "*V *%. '*%¦
^ f C£, 6£, C£, Og, U0 Jq Oq Oq Uq Oq Uq Op CJQ Oq
Meters
Post-Nonreportable Accident Post-Reportable Accident Post-Offsite Impact Accident
Note: Estimates calculated following equations (3a) through (3c) using the "nlcom" command in Stata 17/MP, and are based on
the coefficient estimates from the hedonic regression model 2A, which constrains the slope coefficients for the housing and
location attributes to be common across counties and years. Full regression results for model 2A are presented in Appendix B.
Supplemental analysis results.
Table B.l. Base models: Full hedonic property value regression results.
Model 1A
Model IB
Model 1C
51
-------
RMP 0-5750 mt
Post-accident'
Post-accident x 0-5750 m'
Post-Reportable
Accident x 0-5750 m1^
Post-Offsite Impact
Accident * 0-5750 m1^
ln(acres)
Missing: Acres'
Stories
Missing: Stories'
Bathrooms
Missing: Bathrooms^
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age'
% Land Developed w/in 0-5 00m
Highway w/in 500m '
Lake w/in 500m '
0.0018
(0.0044)
0.0110*
(0.0062)
-0.0127*
(0.0075)
-0.0090
(0.0102)
-0.0116
(0.0088)
0.1210***
(0.0048)
-0.3316***
(0.0286)
-0.0023
(0.0035)
0.0102*
(0.0055)
-0.0022
(0.0063)
-0.0062
(0.0087)
-0.0145*
(0.0080)
-0.0036
(0.0033)
0.0093*
(0.0050)
0.0017
(0.0055)
-0.0093
(0.0077)
-0.0118*
(0.0071)
0.0482***
(0.0061)
0.0654***
(0.0171)
0.0754***
(0.0067)
0.1966***
(0.0244)
0.4200***
(0.0153)
3.4506***
(0.1370)
-0.0066***
(0.0004)
0.0000***
(0.0000)
-0.8707***
(0.0344)
0.0009***
(0.0001)
-0.0244***
(0.0018)
0.0456***
-------
River w/in 250m'
Constant
(0.0047)
0.0313***
(0.0087)
9.0290***
(0.1156)
12.1677* *:
(0.0039)
12.1681**=
(0.0037)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Yes
Year x County
Quarter x County
Yes
House x County
House x Year
Year x County
Quarter x County
Yes
House x County x Year
Year x County
Quarter x County
Yes
Observations
10,426,638
10,426,638
10,426,638
Adjusted R-squared
0.737
0.759
0.767
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
Table B.2. Hedonic property value regression results with incremental 250-meter bins.
Model 2A
Model 2B
Model 2C
RMPt
0-250m
250-500m
500-750m
750-1000m
1000-1250m
1250-1500m
1500-1750m
-0.0908***
(0.0187)
-0.0759***
(0.0144)
-0.0683***
(0.0117)
-0.0534***
(0.0110)
-0.0390***
(0.0094)
-0.0347***
(0.0084)
-0.0417***
(0.0077)
-0.0867***
(0.0185)
-0.0693***
(0.0130)
-0.0617***
(0.0105)
-0.0506***
(0.0095)
-0.0384***
(0.0077)
-0.0323***
(0.0073)
-0.0347***
(0.0068)
-0.0943***
(0.0187)
-0.0706***
(0.0139)
-0.0625***
(0.0104)
-0.0498***
(0.0092)
-0.0381***
(0.0076)
-0.0343***
(0.0071)
-0.0374***
(0.0067)
1750-2000m
-0.0364***
(0.0079)
-0.0329**H
(0.0066)
-0.0347**^
(0.0064)
53
-------
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
Post-accident^
0-250m
250-500m
500-750m
750-1000m
¦0.0357***
(0.0072)
¦0.0324***
(0.0067)
¦0.0271***
(0.0060)
¦0.0208***
(0.0062)
-0.0117*
(0.0060)
-0.0082
(0.0056)
-0.0057
(0.0055)
-0.0030
(0.0053)
-0.0032
(0.0052)
-0.0058
(0.0048)
-0.0047
(0.0047)
0.0034
(0.0047)
0.0025
(0.0042)
0.0034
(0.0040)
0.0007
(0.0035)
0.0093
(0.0062)
-0.0129
(0.0411)
0.0087
(0.0219)
0.0087
(0.0180)
-0.0107
-0.0329***
(0.0063)
-0.0308***
(0.0056)
-0.0287***
(0.0051)
-0.0235***
(0.0055)
-0.0153***
(0.0050)
-0.0115**
(0.0046)
-0.0090**
(0.0045)
-0.0063
(0.0046)
-0.0046
(0.0042)
-0.0064
(0.0039)
-0.0059
(0.0039)
0.0009
(0.0038)
-0.0003
(0.0035)
0.0019
(0.0034)
-0.0013
(0.0028)
0.0095*
(0.0055)
0.0071
(0.0372)
0.0262
(0.0178)
0.0268*
(0.0152)
0.0081
-0.0346***
(0.0061)
-0.0317***
(0.0055)
-0.0297***
(0.0049)
-0.0252***
(0.0054)
-0.0178***
(0.0048)
-0.0140***
(0.0044)
-0.0122***
(0.0043)
-0.0087**
(0.0044)
-0.0063
(0.0040)
-0.0070*
(0.0037)
-0.0072*
(0.0037)
-0.0002
(0.0036)
-0.0016
(0.0032)
0.0005
(0.0030)
-0.0026
(0.0027)
0.0087*
(0.0050)
0.0166
(0.0373)
0.0342**
(0.0172)
0.0309**
(0.0144)
0.0081
-------
1000-1250m
1250-1500m
1500-1750m
1750-2000m
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
Post-Reportable Accident
0-250m
250-500m
(0.0158)
-0.0295**
(0.0143)
-0.0296*
(0.0154)
-0.0195*
(0.0116)
-0.0212*
(0.0116)
-0.0141
(0.0105)
-0.0135
(0.0093)
-0.0104
(0.0088)
-0.0053
(0.0093)
-0.0087
(0.0090)
-0.0179**
(0.0089)
-0.0140
(0.0088)
-0.0170**
(0.0082)
-0.0204**
(0.0082)
-0.0192**
(0.0090)
-0.0119
(0.0083)
-0.0098
(0.0073)
-0.0076
(0.0067)
-0.0111*
(0.0063)
-0.0091
(0.0058)
-0.0654
(0.0509)
-0.0738**
(0.0366)
(0.0139)
-0.0102
(0.0120)
-0.0106
(0.0137)
-0.0058
(0.0103)
-0.0030
(0.0100)
0.0010
(0.0093)
-0.0000
(0.0085)
0.0002
(0.0084)
0.0045
(0.0085)
0.0054
(0.0077)
-0.0052
(0.0076)
-0.0038
(0.0075)
-0.0068
(0.0070)
-0.0107
(0.0070)
-0.0080
(0.0074)
-0.0031
(0.0071)
-0.0023
(0.0063)
-0.0000
(0.0059)
-0.0051
(0.0056)
-0.0044
(0.0051)
-0.0421
(0.0462)
-0.0577*
(0.0313)
(0.0130)
-0.0084
(0.0114)
-0.0054
(0.0129)
-0.0004
(0.0099)
0.0008
(0.0098)
0.0036
(0.0090)
0.0022
(0.0078)
0.0030
(0.0077)
0.0078
(0.0079)
0.0081
(0.0075)
-0.0012
(0.0070)
0.0020
(0.0070)
-0.0020
(0.0061)
-0.0069
(0.0061)
-0.0057
(0.0068)
0.0003
(0.0064)
0.0008
(0.0055)
0.0032
(0.0053)
-0.0023
(0.0051)
-0.0025
(0.0046)
-0.0498
(0.0450)
-0.0661**
(0.0287)
55
-------
500-750m
-0.0505*
-0.0415*
-0.0468*^
(0.0286)
(0.0238)
(0.0230)
750-1000m
-0.0140
-0.0077
-0.0100
(0.0228)
(0.0192)
(0.0176)
1000-1250m
0.0020
0.0023
-0.0023
(0.0204)
(0.0174)
(0.0156)
1250-1500m
-0.0038
-0.0004
-0.0026
(0.0197)
(0.0171)
(0.0163)
1500-1750m
-0.0044
-0.0008
-0.0023
(0.0171)
(0.0151)
(0.0149)
1750-2000m
-0.0046
-0.0031
-0.0037
(0.0156)
(0.0138)
(0.0135)
2000-2250m
-0.0062
-0.0076
-0.0085
(0.0165)
(0.0144)
(0.0139)
2250-2500m
0.0017
0.0010
0.0001
(0.0146)
(0.0131)
(0.0122)
2500-2750m
-0.0032
-0.0007
-0.0030
(0.0142)
(0.0130)
(0.0125)
2750-3000m
-0.0122
-0.0074
-0.0083
(0.0141)
(0.0127)
(0.0118)
3000-3250m
-0.0057
-0.0074
-0.0064
(0.0140)
(0.0124)
(0.0120)
3250-3500m
-0.0025
-0.0030
-0.0036
(0.0140)
(0.0123)
(0.0116)
3500-3750m
-0.0056
-0.0028
-0.0042
(0.0131)
(0.0112)
(0.0106)
3750-4000m
-0.0016
0.0004
-0.0002
(0.0121)
(0.0102)
(0.0093)
4000-4250m
0.0033
0.0040
0.0029
(0.0114)
(0.0093)
(0.0086)
4250-4500m
0.0077
0.0064
0.0057
(0.0123)
(0.0102)
(0.0099)
4500-4750m
-0.0050
-0.0033
-0.0047
(0.0111)
(0.0093)
(0.0088)
4750-5000m
-0.0125
-0.0109
-0.0113
(0.0097)
(0.0080)
(0.0075)
5000-5250m
-0.0143
-0.0116
-0.0123*
(0.0092)
(0.0078)
(0.0073)
5250-5500m
-0.0061
-0.0053
-0.0055
(0.0085)
(0.0068)
(0.0062)
5500-5750m
-0.0018
0.0014
0.0009
(0.0074)
(0.0062)
(0.0057)
ost-Offsite Impact Accident
0-250m
0.0113
-0.0097
0.0033
-------
250-500m
500-750m
750-1000m
1000-1250m
1250-1500m
1500-1750m
1750-2000m
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
(0.0434)
0.0108
(0.0354)
-0.0116
(0.0277)
-0.0372*
(0.0212)
-0.0244
(0.0186)
-0.0179
(0.0168)
-0.0135
(0.0162)
-0.0069
(0.0151)
-0.0176
(0.0145)
-0.0218
(0.0138)
-0.0103
(0.0136)
-0.0186
(0.0132)
-0.0251*
(0.0134)
-0.0106
(0.0125)
-0.0118
(0.0122)
-0.0096
(0.0105)
-0.0077
(0.0111)
-0.0023
(0.0104)
0.0015
(0.0095)
-0.0031
(0.0092)
0.0022
(0.0082)
-0.0001
(0.0070)
0.0011
(0.0380)
-0.0063
(0.0304)
-0.0232
(0.0245)
-0.0463**
(0.0181)
-0.0241
(0.0165)
-0.0224
(0.0153)
-0.0200
(0.0148)
-0.0141
(0.0139)
-0.0182
(0.0130)
-0.0234*
(0.0127)
-0.0152
(0.0121)
-0.0257**
(0.0121)
-0.0300**
(0.0125)
-0.0143
(0.0116)
-0.0188*
(0.0113)
-0.0146
(0.0096)
-0.0131
(0.0104)
-0.0081
(0.0095)
-0.0045
(0.0084)
-0.0079
(0.0075)
-0.0041
(0.0067)
-0.0053
(0.0060)
-0.0062
(0.0380)
0.0006
(0.0278)
-0.0155
(0.0231)
-0.0397**
(0.0171)
-0.0173
(0.0151)
-0.0183
(0.0145)
-0.0172
(0.0142)
-0.0125
(0.0132)
-0.0143
(0.0124)
-0.0205*
(0.0122)
-0.0105
(0.0117)
-0.0237**
(0.0116)
-0.0272**
(0.0121)
-0.0127
(0.0109)
-0.0167
(0.0111)
-0.0143
(0.0093)
-0.0124
(0.0104)
-0.0087
(0.0090)
-0.0052
(0.0081)
-0.0091
(0.0073)
-0.0046
(0.0064)
-0.0058
(0.0055)
-0.0054
-------
ln(acres)
Missing: Acres'
Stories
Missing: Stories'
Bathrooms
Missing: Bathrooms^
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age'
% Land Developed w/in 0-5 00m
Highway w/in 500m '
Lake w/in 500m '
River w/in 250m'
Constant
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
(0.0063)
0.1203***
(0.0048)
-0.3302***
(0.0286)
0.0482***
(0.0061)
0.0653***
(0.0171)
0.0753***
(0.0067)
0.1962***
(0.0244)
0.4198***
(0.0153)
3 4492***
(0.1369)
-0.0065***
(0.0004)
0.0000***
(0.0000)
-0.8704***
(0.0344)
0.0009***
(0.0001)
-0.0238***
(0.0018)
0.0458***
(0.0047)
0.0317***
(0.0087)
9.0375***
(0.1155)
Yes
Year x County
Quarter x County
Yes
(0.0053)
(0.0048)
12.1750***
(0.0041)
House x County
House x Year
Year x County
Quarter x County
Yes
12.1753***
(0.0039)
House x County x Year
Year x County
Quarter x County
Yes
Observations
Adjusted R-squared
10,426,638
0.737
10,426,638
0.759
10,426,638
0.767
58
-------
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
in Appendix B.
59
-------
Figure B. 2. Model 2C results: Percent change in price due to an accident, by 250-meter bins.
Panel (A): Nonreportable accident.
10
Panel (B): Reportable Accident, only onsite impacts.
10
5
Panel (C): Reportable Accident with offsite impacts.
10
Meters
Post-Nonreportable Accident Post-Reportable Accident Post-Offsite Impact Accident
Note: Estimates calculated following equations (3a) through (3c) using the "nlcom" command in Stata 17/MP, and are based on
the coefficient estimates from the hedonic regression model 2C, which includes county-by-year interactions with the housing and
location attributes. Full regression results for model 2C are presented in Appendix B. Supplemental analysis results.
Table B.l. Base models: Full hedonic property value regression results.
Model 1A Model IB Model 1C
60
-------
RMP 0-5750 mt
Post-accident'
Post-accident x 0-5750 m'
Post-Reportable
Accident x 0-5750 m1^
Post-Offsite Impact
Accident * 0-5750 m1^
ln(acres)
Missing: Acres'
Stories
Missing: Stories'
Bathrooms
Missing: Bathrooms^
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age'
% Land Developed w/in 0-5 00m
Highway w/in 500m '
0.0018
(0.0044)
0.0110*
(0.0062)
-0.0127*
(0.0075)
-0.0090
(0.0102)
-0.0116
(0.0088)
0.1210***
(0.0048)
-0.3316***
(0.0286)
-0.0023
(0.0035)
0.0102*
(0.0055)
-0.0022
(0.0063)
-0.0062
(0.0087)
-0.0145*
(0.0080)
-0.0036
(0.0033)
0.0093*
(0.0050)
0.0017
(0.0055)
-0.0093
(0.0077)
-0.0118*
(0.0071)
0.0482***
(0.0061)
0.0654***
(0.0171)
0.0754***
(0.0067)
0.1966***
(0.0244)
0.4200***
(0.0153)
3.4506***
(0.1370)
-0.0066***
(0.0004)
0.0000***
(0.0000)
-0.8707***
(0.0344)
0.0009***
(0.0001)
-0.0244***
(0.0018)
-------
Lake w/in 500m '
River w/in 250m'
Constant
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
0.0456***
(0.0047)
0.0313***
(0.0087)
9.0290***
(0.1156)
Yes
Year x County
Quarter x County
Yes
12.1677***
(0.0039)
House x County
House x Year
Year x County
Quarter x County
Yes
12.1681***
(0.0037)
House x County x Year
Year x County
Quarter x County
Yes
Observations
10,426,638
10,426,638
10,426,638
Adjusted R-squared 0.737 0.759 0.767
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
Table B.2. Hedonic property value regression results with incremental 250-meter bins.
Model 2A
Model 2B
Model 2C
RMPf
0-250m
250-500m
500-750m
750-1000m
1000-1250m
1250-1500m
1500-1750m
-0.0908***
(0.0187)
-0.0759***
(0.0144)
-0.0683***
(0.0117)
-0.0534***
(0.0110)
-0.0390***
(0.0094)
-0.0347***
(0.0084)
-0.0417***
(0.0077)
-0.0867***
(0.0185)
-0.0693***
(0.0130)
-0.0617***
(0.0105)
-0.0506***
(0.0095)
-0.0384***
(0.0077)
-0.0323***
(0.0073)
-0.0347***
(0.0068)
-0.0943***
(0.0187)
-0.0706***
(0.0139)
-0.0625***
(0.0104)
-0.0498***
(0.0092)
-0.0381***
(0.0076)
-0.0343***
(0.0071)
-0.0374***
(0.0067)
1750-2000m
-0.0364***
-0.0329***
-0.0347***
62
-------
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
Post-accident^
0-250m
250-500m
500-750m
(0.0079)
¦0.0357***
(0.0072)
¦0.0324***
(0.0067)
¦0.0271***
(0.0060)
¦0.0208***
(0.0062)
-0.0117*
(0.0060)
-0.0082
(0.0056)
-0.0057
(0.0055)
-0.0030
(0.0053)
-0.0032
(0.0052)
-0.0058
(0.0048)
-0.0047
(0.0047)
0.0034
(0.0047)
0.0025
(0.0042)
0.0034
(0.0040)
0.0007
(0.0035)
0.0093
(0.0062)
-0.0129
(0.0411)
0.0087
(0.0219)
0.0087
(0.0180)
(0.0066)
-0.0329***
(0.0063)
-0.0308***
(0.0056)
-0.0287***
(0.0051)
-0.0235***
(0.0055)
-0.0153***
(0.0050)
-0.0115**
(0.0046)
-0.0090**
(0.0045)
-0.0063
(0.0046)
-0.0046
(0.0042)
-0.0064
(0.0039)
-0.0059
(0.0039)
0.0009
(0.0038)
-0.0003
(0.0035)
0.0019
(0.0034)
-0.0013
(0.0028)
0.0095*
(0.0055)
0.0071
(0.0372)
0.0262
(0.0178)
0.0268*
(0.0152)
(0.0064)
-0.0346***
(0.0061)
-0.0317***
(0.0055)
-0.0297***
(0.0049)
-0.0252***
(0.0054)
-0.0178***
(0.0048)
-0.0140***
(0.0044)
-0.0122***
(0.0043)
-0.0087**
(0.0044)
-0.0063
(0.0040)
-0.0070*
(0.0037)
-0.0072*
(0.0037)
-0.0002
(0.0036)
-0.0016
(0.0032)
0.0005
(0.0030)
-0.0026
(0.0027)
0.0087*
(0.0050)
0.0166
(0.0373)
0.0342**
(0.0172)
0.0309**
(0.0144)
-------
750-1000m
1000-1250m
1250-1500m
1500-1750m
1750-2000m
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
5500-5750m
Post-Reportable Accident
0-250m
250-500m
-0.0107
(0.0158)
-0.0295**
(0.0143)
-0.0296*
(0.0154)
-0.0195*
(0.0116)
-0.0212*
(0.0116)
-0.0141
(0.0105)
-0.0135
(0.0093)
-0.0104
(0.0088)
-0.0053
(0.0093)
-0.0087
(0.0090)
-0.0179**
(0.0089)
-0.0140
(0.0088)
-0.0170**
(0.0082)
-0.0204**
(0.0082)
-0.0192**
(0.0090)
-0.0119
(0.0083)
-0.0098
(0.0073)
-0.0076
(0.0067)
-0.0111*
(0.0063)
-0.0091
(0.0058)
-0.0654
(0.0509)
-0.0738**
0.0081
(0.0139)
-0.0102
(0.0120)
-0.0106
(0.0137)
-0.0058
(0.0103)
-0.0030
(0.0100)
0.0010
(0.0093)
-0.0000
(0.0085)
0.0002
(0.0084)
0.0045
(0.0085)
0.0054
(0.0077)
-0.0052
(0.0076)
-0.0038
(0.0075)
-0.0068
(0.0070)
-0.0107
(0.0070)
-0.0080
(0.0074)
-0.0031
(0.0071)
-0.0023
(0.0063)
-0.0000
(0.0059)
-0.0051
(0.0056)
-0.0044
(0.0051)
-0.0421
(0.0462)
-0.0577*
0.0081
(0.0130)
-0.0084
(0.0114)
-0.0054
(0.0129)
-0.0004
(0.0099)
0.0008
(0.0098)
0.0036
(0.0090)
0.0022
(0.0078)
0.0030
(0.0077)
0.0078
(0.0079)
0.0081
(0.0075)
-0.0012
(0.0070)
0.0020
(0.0070)
-0.0020
(0.0061)
-0.0069
(0.0061)
-0.0057
(0.0068)
0.0003
(0.0064)
0.0008
(0.0055)
0.0032
(0.0053)
-0.0023
(0.0051)
-0.0025
(0.0046)
-0.0498
(0.0450)
-0.0661**
64
-------
(0.0366)
(0.0313)
(0.0287)
500-750m
-0.0505*
-0.0415*
-0.0468*^
(0.0286)
(0.0238)
(0.0230)
750-1000m
-0.0140
-0.0077
-0.0100
(0.0228)
(0.0192)
(0.0176)
1000-1250m
0.0020
0.0023
-0.0023
(0.0204)
(0.0174)
(0.0156)
1250-1500m
-0.0038
-0.0004
-0.0026
(0.0197)
(0.0171)
(0.0163)
1500-1750m
-0.0044
-0.0008
-0.0023
(0.0171)
(0.0151)
(0.0149)
1750-2000m
-0.0046
-0.0031
-0.0037
(0.0156)
(0.0138)
(0.0135)
2000-2250m
-0.0062
-0.0076
-0.0085
(0.0165)
(0.0144)
(0.0139)
2250-2500m
0.0017
0.0010
0.0001
(0.0146)
(0.0131)
(0.0122)
2500-2750m
-0.0032
-0.0007
-0.0030
(0.0142)
(0.0130)
(0.0125)
2750-3000m
-0.0122
-0.0074
-0.0083
(0.0141)
(0.0127)
(0.0118)
3000-3250m
-0.0057
-0.0074
-0.0064
(0.0140)
(0.0124)
(0.0120)
3250-3500m
-0.0025
-0.0030
-0.0036
(0.0140)
(0.0123)
(0.0116)
3500-3750m
-0.0056
-0.0028
-0.0042
(0.0131)
(0.0112)
(0.0106)
3750-4000m
-0.0016
0.0004
-0.0002
(0.0121)
(0.0102)
(0.0093)
4000-4250m
0.0033
0.0040
0.0029
(0.0114)
(0.0093)
(0.0086)
4250-4500m
0.0077
0.0064
0.0057
(0.0123)
(0.0102)
(0.0099)
4500-4750m
-0.0050
-0.0033
-0.0047
(0.0111)
(0.0093)
(0.0088)
4750-5000m
-0.0125
-0.0109
-0.0113
(0.0097)
(0.0080)
(0.0075)
5000-5250m
-0.0143
-0.0116
-0.0123*
(0.0092)
(0.0078)
(0.0073)
5250-5500m
-0.0061
-0.0053
-0.0055
(0.0085)
(0.0068)
(0.0062)
5500-5750m
-0.0018
0.0014
0.0009
(0.0074)
(0.0062)
(0.0057)
Post-Offsite Impact Accident^
-------
0-250m
250-500m
500-750m
750-1000m
1000-1250m
1250-1500m
1500-1750m
1750-2000m
2000-2250m
2250-2500m
2500-2750m
2750-3000m
3000-3250m
3250-3500m
3500-3750m
3750-4000m
4000-4250m
4250-4500m
4500-4750m
4750-5000m
5000-5250m
5250-5500m
0.0113
(0.0434)
0.0108
(0.0354)
-0.0116
(0.0277)
-0.0372*
(0.0212)
-0.0244
(0.0186)
-0.0179
(0.0168)
-0.0135
(0.0162)
-0.0069
(0.0151)
-0.0176
(0.0145)
-0.0218
(0.0138)
-0.0103
(0.0136)
-0.0186
(0.0132)
-0.0251*
(0.0134)
-0.0106
(0.0125)
-0.0118
(0.0122)
-0.0096
(0.0105)
-0.0077
(0.0111)
-0.0023
(0.0104)
0.0015
(0.0095)
-0.0031
(0.0092)
0.0022
(0.0082)
-0.0001
(0.0070)
-0.0097
(0.0380)
-0.0063
(0.0304)
-0.0232
(0.0245)
-0.0463**
(0.0181)
-0.0241
(0.0165)
-0.0224
(0.0153)
-0.0200
(0.0148)
-0.0141
(0.0139)
-0.0182
(0.0130)
-0.0234*
(0.0127)
-0.0152
(0.0121)
-0.0257**
(0.0121)
-0.0300**
(0.0125)
-0.0143
(0.0116)
-0.0188*
(0.0113)
-0.0146
(0.0096)
-0.0131
(0.0104)
-0.0081
(0.0095)
-0.0045
(0.0084)
-0.0079
(0.0075)
-0.0041
(0.0067)
-0.0053
(0.0060)
0.0033
(0.0380)
0.0006
(0.0278)
-0.0155
(0.0231)
-0.0397**
(0.0171)
-0.0173
(0.0151)
-0.0183
(0.0145)
-0.0172
(0.0142)
-0.0125
(0.0132)
-0.0143
(0.0124)
-0.0205*
(0.0122)
-0.0105
(0.0117)
-0.0237**
(0.0116)
-0.0272**
(0.0121)
-0.0127
(0.0109)
-0.0167
(0.0111)
-0.0143
(0.0093)
-0.0124
(0.0104)
-0.0087
(0.0090)
-0.0052
(0.0081)
-0.0091
(0.0073)
-0.0046
(0.0064)
-0.0058
(0.0055)
-------
5500-5750m
ln(acres)
Missing: Acres'
Stories
Missing: Stories'
Bathrooms
Missing: Bathrooms^
ln(interior sqft)
Missing: Interior sqft
Age (years)
AgeA2
Missing: Age'
% Land Developed w/in 0-5 00m
Highway w/in 500m '
Lake w/in 500m '
River w/in 250m*
Constant
0.0011
(0.0063)
0.1203***
(0.0048)
-0.3302***
(0.0286)
0.0482***
(0.0061)
0.0653***
(0.0171)
0.0753***
(0.0067)
0.1962***
(0.0244)
0.4198***
(0.0153)
3 4492***
(0.1369)
-0.0065***
(0.0004)
0.0000***
(0.0000)
-0.8704***
(0.0344)
0.0009***
(0.0001)
-0.0238***
(0.0018)
0.0458***
(0.0047)
0.0317***
(0.0087)
9.0375***
(0.1155)
-0.0062
(0.0053)
-0.0054
(0.0048)
12.1750* *:
(0.0041)
12.1753**=
(0.0039)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
Yes
Year x County
Quarter x County
Yes
House x County
House x Year
Year x County
Quarter x County
Yes
House x County x Year
Year x County
Quarter x County
Yes
Observations
Adjusted R-squared
10,426,638
0.737
10,426,638
0.759
10,426,638
0.767
67
-------
Note: 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. Note that 1,804 singleton observations were
dropped from the regression model. Variables denoted with f are binary indicators.
in Appendix B.
Table B.3. Model 3B: Hedonic Regression Examining Price Effects of Multiple Accidents.
Model 3B
RMP 0-5750 nr
# RMP sites 0-5750 m
Post-accident^
# subsequent accidents
Post-accident x 0-5750 m
# subsequent accidents x 0-5750 m
Post-Reportable Accident x 0-5750 m '
# subsequent post-reportable accidents x 0-5750 m
Post-Offsite Impact Accident x 0-5750 m '
# subsequent post-offsite impact accidents x 0-5750 m
Constant
-0.0026
(0.0034)
-0.0006
(0.0018)
0.0086*
(0.0052)
-0.0025
(0.0016)
-0.0014
(0.0063)
-0.0034
(0.0028)
-0.0047
(0.0087)
0.0034
(0.0053)
-0.0139*
(0.0080)
0.0045
(0.0075)
12.1711**=
(0.0038)
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
House x County
House x Year
Year x County
Quarter x
County
Yes
68
-------
Observations 10,426,638
Adjusted R-squared 0.759
Note: Dependent variable is ln(price). * p<0.10, ** p<0.05, *** p<0.01. Standard errors
in parentheses, clustered at the county level. Regression model results of equation (4), and
based on a variant of Model IB. Regression model estimated using "reghdfe" command
in Stata 17/MP. Note that 1,804 singleton observations were dropped from the regression
model. Variables denoted with f are binary indicators.
Table B.4. Cumulative price impacts of multiple accidents based on Model 3B.
Number of Accidents
1 2 3 4 5
Nonreportable
-0.1395
-0.4768
-0.8129
-1.1479
-1.4817
(0.6270)
(0.6687)
(0.8075)
(1.0023)
(1.2256)
Reportable
-0.6033
-0.5986
-0.5938
-0.5891
-0.5844
(0.7027)
(0.8092)
(1.0677)
(1.3960)
(1.7556)
Offsite Impacts
-1.9758***
-1.5339**
-1.0901
-0.6442
-0.1963
(0.5729)
(0.7410)
(1.1117)
(1.5496)
(2.0145)
Note: * p<0.10, ** p<0.05, *** p<0.01. Estimates calculated following equations similar to (3a)
through (3c) using the "nlcom" command in Stata 17/MP, and are based on the coefficient
estimates from Model 3B regression results in Table B.3.
69
-------
Table B.5. Hedonic regression results examining evolution of accident price effects over time.
RMP 0-5750 mt
Post-accident^
Post-accident x 0-5750 m '
0-1 year
1-2 year
2-3 year
3-4 year
4-5 year
5-6 year
6-7 year
7-8 year
8-9 year
9-10 year
10-11 year
11-12 year
Model 4B
-0.0035
(0.0032)
0.0087
(0.0054)
-0.0073
(0.0058)
0.0005
(0.0049)
-0.0022
(0.0049)
-0.0068
(0.0065)
-0.0059
(0.0056)
-0.0042
(0.0051)
-0.0048
(0.0057)
0.0007
(0.0062)
-0.0021
(0.0061)
-0.0032
(0.0063)
-0.0018
(0.0058)
-0.0041
(0.0066)
70
-------
12-13 year
13-14 year
14-15 year
15-16 year
Post-Reportable Accident x 0-5750 m '
0-1 year
1-2 year
2-3 year
3-4 year
4-5 year
5-6 year
6-7 year
7-8 year
8-9 year
9-10 year
10-11 year
11-12 year
12-13 year
13-14 year
14-15 year
15-16 year
Post-Offsite Impact Accident x 0-5750 m '
0-1 year
-0.0028
(0.0064)
-0.0134
(0.0086)
-0.0123
(0.0104)
-0.0075
(0.0157)
0.0052
(0.0080)
-0.0078
(0.0080)
-0.0024
(0.0081)
-0.0086
(0.0091)
-0.0099
(0.0091)
0.0001
(0.0076)
-0.0023
(0.0081)
-0.0044
(0.0082)
0.0061
(0.0088)
0.0044
(0.0084)
0.0042
(0.0091)
-0.0005
(0.0100)
-0.0019
(0.0088)
0.0113
(0.0113)
0.0251
(0.0171)
0.0267
(0.0217)
-0.0104*
(0.0062)
71
-------
1-2 year
2-3 year
3-4 year
4-5 year
5-6 year
6-7 year
7-8 year
8-9 year
9-10 year
10-11 year
11-12 year
12-13 year
13-14 year
14-15 year
15-16 year
Constant
House attributes
Year Fixed Effects
Quarter Fixed Effects
Tract Fixed Effects
-0.0037
(0.0085)
-0.0042
(0.0089)
0.0036
(0.0086)
0.0010
(0.0096)
-0.0075
(0.0086)
-0.0126
(0.0085)
-0.0134
(0.0083)
-0.0189**
(0.0087)
-0.0207**
(0.0096)
-0.0233**
(0.0098)
-0.0109
(0.0102)
-0.0112
(0.0098)
-0.0058
(0.0103)
-0.0176
(0.0147)
-0.0016
(0.0189)
12.1690***
(0.0039)
House x County
House x Year
Year x County
Quarter * County
Yes
Observations
Adjusted R-squared
10,426,638
0.759
72
-------
Note: Dependent variable is ln(price). * p<0.10, ** p<0.05, *** p<0.01.
Standard errors in parentheses, clustered at the comity level. Regression model
results of equation (5), and based on a variant of Model IB. Regression model
estimated using "reghdfe" command in Stata 17/MP. Note that 1,804 singleton
observations were dropped from the regression model. Variables denoted with f
are binary indicators.
Figure B.3. Percent change in price by year after offsite impact accident: Based on Model 4B.
Q_
_C
QJ 0
Ctf)
c
CD
-C
u
£
** „**
."V >'
** **
>• >'
y* *
& t* **
* & i
<
>
<
;v
• <
~ «
»
<
>
<
<
»
i
(
' 1
>
> «
r
>
(
r
i (
i
> <
<
»
*
»
>
' (
>
4 5 6 7 8 9 10 11 12 13 14 15 16
Years after offsite impact accident
Note: Estimates calculated following an equation similar to equation (3c) using the "nlcom" command in Stata 17/MP, and are
based on the coefficient estimates from the hedonic regression model 4B in Table B.5 in Appendix B.
73
------- |