NCEE 0
NATIONAL CENTER FOR
ENVIRONMENTAL ECONOMICS
What Do Property Values Really Tell Us?
A Hedonic Study of Underground Storage Tanks
Dennis Guignet
Working Paper Series
Working Paper# 12-01
March, 2012
^e.0 sr^ U.S. Environmental Protection Agency
g ra National Center for Environmental Economics
s z 1200 Pennsylvania Avenue, NW (MC 1809)
^ Washington, DC 20460
sP http://www.epa.gov/economics
ProT^-
-------
What Do Property Values Really Tell Us?
A Hedonic Study of Underground Storage Tanks
Dennis Guignet
NCEE Working Paper Series
Working Paper # 12-01
March, 2012
DISCLAIMER
The views expressed in this paper are those of the author(s) and do not necessarily represent those
of the U.S. Environmental Protection Agency. In addition, although the research described in this
paper may have been funded entirely or in part by the U.S. Environmental Protection Agency, it
has not been subjected to the Agency's required peer and policy review. No official Agency
endorsement should be inferred.
-------
What Do Property Values Really Tell Us?
A Hedonic Study of Underground Storage Tanks
By: Dennis Guignet1
Last Revised: March 1st, 2012
Abstract:
Hedonic property value models are widely used, but are susceptible to omitted variable
bias and potentially invalid conjectures regarding the assumed measure of environmental
quality. This paper focuses on an application where both are of particular concern:
leaking underground storage tanks. I estimate a hedonic model using quasi-experimental
and spatial econometric techniques. Similar to previous studies, I examine how house
prices vary with distance to the disamenity. This approach yields little evidence that
prices are adversely impacted. However, to better measure risks, I utilize home-specific
data on correspondence from environmental regulators, and find a 9-12% depreciation
when households are well-informed.
JEL Classification: D63 (Externalities); 118 (Government Policy; Regulations; Public
Health); Q51 (Valuation of Environmental Effects);
Keywords: hedonic analysis, housing prices, leaking underground storage tanks, LUST,
contaminated sites, groundwater contamination, remediation benefits.
This research was supported by funding from the US Environmental Protection Agency (contract
EH08H000849 via Industrial Economics). I thank Anna Alberini for her support and guidance. This paper
extends on earlier work co-authored with Jeff Zabel, who I thank for his valuable comments. I am also
grateful to Maureen Cropper, Erik Lichtenberg, Doug Lipton, Ted McConnell, Charles Towe, and Patrick
Walsh for comments on earlier drafts. I thank the Maryland Department of Environment and the National
Center for Smart Growth Research and Education for data support. Any views and conclusions expressed
are my own, and do not necessarily reflect the views and policies of the US EPA.
1 National Center for Environmental Economics; US Environmental Protection Agency; Mail Code 1809 T
1200 Pennsylvania Avenue, N.W.; Washington, DC 20460; guignet.dennis@epa.gov
-------
1
I. Introduction
In the absence of markets for environmental quality, researchers rely on non-
market valuation techniques to estimate the value of environmental amenities and
disamenities. One of the most widely used revealed preference approaches is hedonic
property value models, where the value of an environmental commodity is inferred from
its impact on home prices. Hedonics is an attractive technique because housing
transaction data is readily available, the method relies on actual market behavior, and it
presumably captures all aspects of a change in welfare.
There are, however, two practical issues in obtaining valid welfare estimates. The
first is omitted variable bias. If there are unobserved influences on house prices that are
correlated with the assumed measure of environmental quality, then the implicit price
estimates of interest may be biased. The second is whether the measure of environmental
quality assumed in the hedonic model reflects what buyers and sellers in the market are
actually aware of, and care about.1 To address, or at the very least illustrate, these issues,
I focus on an application where both are of particular concern: groundwater pollution
from leaking underground storage tanks (LUSTs), such as those found at gas stations.
I estimate a hedonic property value model using a unique and comprehensive
dataset of house transactions, neighborhood characteristics, underground storage tank
(UST) facilities, leak investigations, and groundwater well contamination tests, for three
Maryland counties (Baltimore, Frederick, and Baltimore City) from 1996-2007.
Disentangling the implicit price of LUSTs and cleanups is challenging because the
placement of UST facilities (and hence potential leaks) may be correlated with the spatial
-------
2
distribution of other amenities and disamenities. Moreover, the UST facilities themselves
pose both desirable and undesirable characteristics besides contamination.
I take several steps to reduce potentially confounding effects on home values. I
include (i) extensive controls in the hedonic regressions (home and neighborhood
attributes), (ii) neighborhood fixed effects, and (iii) comparable non-leaking USTs. The
latter, along with temporal variation in the discovery of leaks, allow for a spatial
difference-in-difference regression framework (see Horsch and Lewis, 2009). I also
implement a "propensity score" type of hedonic model, where in the first stage I estimate
the probability that a leak is discovered at the individual UST facilities. This is then
included in the right-hand side of the hedonic model, which essentially allows for a
comparison of home values around leaking and non-leaking tanks that have a similar
propensity for a leak to be discovered. Repeat sales and spatial econometric models are
estimated to check the robustness of the results.
Previous hedonic property value studies often rely on proximity of a home to the
disamenity and discrete informational events to proxy environmental quality (Boyle and
Kiel, 2001; Farber, 1998). However, it remains unclear whether this is always a valid
measure for environmental and health risks, especially in the context of LUSTs. A unique
contribution of this paper is that in addition to measuring risk solely by proximity to the
disamenity, I also account for home-specific variation in information and pollution,
which I measure with domestic groundwater well tests and correspondence from the
Maryland Department of Environment (MDE).
-------
3
I ask four main research questions. First, analogous to the conventional
identification strategy, is the value of houses in close proximity to a UST adversely
impacted when a leak is discovered, and how does this change when cleanup is
undertaken and completed? Second, does this effect differ depending on whether the
primary exposure pathway (private groundwater wells) is present? Third, how are prices
impacted at a subset of homes where households have additional information regarding
the disamenity? Fourth, what are the implications for using hedonics in benefit-cost
analysis?
Despite extensive efforts to control for potential omitted variable bias, in general,
I find that the typical LUST has little effect on the price of surrounding homes (e.g.,
within 500 meters). This holds even if a home relies on a private well. Based solely on
this conventional identification strategy of relying on distance, a researcher may conclude
that home values are not impacted by LUST sites, and therefore that cleaning up or
preventing these leaks yields no benefits (at least as capitalized in property values).
However, I do find a significant 9-12% depreciation among homes where the private well
was tested for contamination. These households face actual (or suspected) risks and are
relatively well-informed since they receive correspondence from MDE.
This illustrates how hedonic analyses can lead to vastly different conclusions
depending on the assumed measure of environmental quality, and brings into question
findings from previous hedonic studies where more refined environmental measures were
not available. More care is needed towards how we measure an environmental good in
the right-hand side of our hedonic models. Particular attention should be given to what
information buyers and sellers in the market have regarding the amenity or disamenity of
-------
4
interest, how they perceive this information, and how well our assumed environmental
measure reflects these perceptions.
The rest of this paper is organized as follows. Section II reviews the hedonic
literature relevant to this analysis. Section III provides background information on
LUSTs. The empirical framework is outlined in section IV. I describe the data in section
V, and present the results in section VI. Section VII concludes.
II. Previous Literature
II.A. The Hedonic Price Method
In a differentiated good market (e.g. housing) the matching of buyers and sellers
forms a hedonic price schedule, where the price of the differentiated good is a function of
the attributes composing that good. The marginal contribution of each attribute to the
overall value of the good is the marginal implicit price. Rosen (1974) demonstrated that
in equilibrium the marginal implicit price equates to the buyer's marginal willingness to
pay, implying that marginal welfare estimates can be obtained with no information on
people's underlying preferences, and that only the hedonic equation needs to be
estimated.
To estimate non-marginal welfare changes one may have to pursue the second
stage of Rosen's procedure, and estimate the underlying demand functions. However,
under certain assumptions non-marginal welfare impacts from sufficiently local
disamenities (i.e. only affect a few homes so as not to shift the hedonic price function) are
simply windfall gains or losses to the property owners, and therefore can be estimated
-------
5
solely from the hedonic price function (Palmquist, 2005). I make this assumption here
since pollution from a LUST usually only migrates a few hundred meters, at most.
Hedonic models have been used extensively to value air quality and visibility (e.g.
Chattopadhyay, 1999; Kim and Goldsmith, 2009), water quality (e.g. Leggett and
Bockstael, 2000; Walsh et al., 2011), noise (e.g. Pope, 2008, Day et al., 2007) and health
risks (e.g., Gayer et al., 2002; Davis, 2004), among other things. There are several areas
of research that are particularly important to understanding the effects of LUSTs on home
values.
II. B. Groundwater Quality and Residential Property Values
To my knowledge the few studies investigating the effects of groundwater
contamination on residential property prices generally find little or no effect. For
example, Malone and Barrows (1990) found that nitrates in the groundwater did not
affect home prices. Examining seven different towns in Wisconsin, Page and Rabinowitz
(1993) report that assessed values are not affected by well contamination from landfills,
industrial sites, or pesticide run-off Dotzour (1997) finds that groundwater pollution
does not lead to a significant difference in the average house price in Wichita, Kansas,
which is not surprising since most (if not all) of the homes were connected to the public
water system. These studies provided valuable contributions, but are well over a decade
old, and econometric techniques and data quality have improved greatly since that time.
More recently, Boyle et al. (2010) find that home prices decline by 0.5-1% for
each 0.01 mg/1 of arsenic above the 0.05 mg/1 standard set by the EPA. This depreciation
appears to be temporary since prices rebound within a few years. Boyle et al. speculate
-------
6
this may be due to the availability of in-home water treatment systems or dissipation of
perceived risks once media attention stops.
II. C. Contaminated Sites and Residential Property Values
There is a significant literature on the effects of larger contaminated sites (e.g.
Superfund sites) on home values. Often the identification strategy in these studies is to
account for proximity to the site, and allow the implicit price of proximity to differ before
and after a contamination-related event (e.g. discovery of contamination, listing on the
"3
National Priorities List, cleanup being undertaken, and cleanup completion). Each event
represents new information that may change public perceptions of environmental and
health risks, and in turn, affect property values. The change in the premium for distance
from a site reflects a change in residents' welfare.4
Kohlhase (1991) and Michaels and Smith (1990) are among the earliest to study
the effects of contaminated sites on property values. Farber (1998) reviews these and
subsequent hedonic studies and finds that property values increase, on average, by $3,500
for each additional mile from a contaminated site. However, Boyle and Kiel (2001) find
significant variation across studies, ranging from $190 to $11,450 per mile. Most find
that home prices decrease when a site is placed on the NPL (Kiel, 1995; Farber, 1998;
Boyle and Kiel, 2001; Jackson, 2001), but Kiel and Williams (2007) find that this may
not be the case at all sites.
Evidence that home values rebound during and after cleanup is mixed (Kiel and
Zabel, 2001; Dale et al., 1999; McCluskey and Rausser, 2003; Kiel and Williams, 2007).
Even though cleanup reduces objective risks, property values may not rebound because
-------
7
of a lingering social stigma (e.g., Messer et al, 2006; Gregory and Scatterfield, 2002).
The site may still be perceived as a threat, and the surrounding community publicly
shunned.
Overall, the mixed results from this literature suggests that there is significant
heterogeneity across sites, and more refined measures of environmental quality are
needed to control for such heterogeneity.
II. I). LUSTs and Residential Property Values
While there is a vast literature on how larger types of contaminated sites affect
property values, comparability of these studies to LUSTs is unclear. LUSTs are more
numerous, less publicized, relatively smaller in size, and pollution is more local. LUSTs
are comparatively homogeneous in that contamination mainly consists of petroleum
products, and the sites are generally gas stations, or similar types of commercial and
industrial facilities. In contrast, Superfund and other contaminated sites are comprised of
a wide assortment of prior land uses and pollutants. Most hedonic studies focusing on
larger contaminated sites are concerned with just a single site or assume that only the
nearest site affects property values, but there are numerous USTs and LUSTs within a
single housing market.
There are few studies on LUSTs and residential property values. Simons et al.
(1997) estimate a hedonic model using a cross-section of home sales in Cuyahoga
County, Ohio, and find a 17% depreciation associated with homes within 300 feet of a
leak at a registered UST facility. They find no effect associated with proximity to
registered non-leaking tanks or non-registered LUSTs. In a later analysis, Simons et al.
-------
8
(1999) find that "contamination" from nearby gas stations reduces home values by 14-
16%.5 Isakson and Ecker (2010) focus on 50 USTs in Cedar Falls, Iowa, which
environmental regulators categorized as "no risk," "low risk," and "high risk." They find
that the prices of homes adjacent to a high risk LUST are about 11% lower.
Due to the small sample size and cross-sectional nature of these studies, one
should use caution when interpreting the results as causal. In contrast, I utilize a large
panel of home sales over 11 years, which allows me to better identify the causal impact
of LUSTs on property values. Using the same dataset, my study extends on an earlier
analysis by Zabel and Guignet (2011), who emphasized the need to exploit both spatial
and temporal variation in identifying the causal effects of LUSTs on home values.
Zabel and Guignet (2011) include neighborhood fixed effects and spatial
econometric techniques to minimize potential biases from unobserved spatially correlated
influences on house prices. Even more importantly, they observed home sales before and
after the leak, allowing them to establish a pre-leak baseline, and analyze how prices
change upon the discovery of a LUST, and completion of a leak investigation. They
examined home prices within 100, 200, 500, 1000, and 2000 meters of a LUST, and
checked whether the impact of a leak varied depending on the severity of contamination,
the presence of an exposure pathway, and publicity surrounding the site.
In general, Zabel and Guignet (2011) found that the typical LUST had little effect
on home values, but more publicized (and more contaminated) sites can cause more than
a 10%) depreciation at homes up to 1,000 meters away. Focusing on these sites
individually, however, revealed substantial heterogeneity in the price impact, ranging
-------
9
from a 12% depreciation to a 14% appreciation. This again suggests the need for a more
refined measure of environmental quality.
In this paper, I devise a quasi experiment focusing only on leaks at UST facilities
registered with Maryland's Oil Control Program. In contrast, Zabel and Guignet (2011)
focus on all LUSTs, including historical sites where regulators were previously unaware
that an old inactive UST was (or had been) present. When focusing only on registered
USTs a clear counterfactual exists, homes near non-leaking registered UST facilities.
These can be compared to homes near leaking USTs, both before and after the leak. This
framework allows me to estimate difference-in-difference and "propensity score" types of
hedonic models. Additionally, in this paper a repeat sales model is estimated to check the
robustness of the results.
A key contribution of this research is that, in contrast to previous hedonic
analyses, I utilize home-specific variation in information and environmental quality,
namely well tests and correspondence from environmental regulators.
II. E. How to Measure Risk
Despite its widespread use, it remains unclear whether distance to the source of
pollution is always an acceptable proxy for environmental and health risks, especially in
the context of LUSTs. First of all, if the general public is unaware of the pollution
problem, then no threat is perceived and distance is unrelated to perceived risk. In this
case, there would be no premium for distance of a home from the disamenity.
While people can see gas stations, and other UST facilities, it is unclear whether
they are always aware of a LUST near their home. USTs are underground and there may
-------
10
be no obvious visual cues of contamination. When a leak does occur, there is little media
attention, if any, and if there is, it is restricted to only the most severe cases.6
The Maryland Department of Environment (MDE) requires a responsible party
(usually the UST owner) to notify the public only in the most severe cases, where a
n
corrective action plan is necessary. Notification is only required for "members of the
public directly affected by the release and planned corrective action" (COMAR,
26.10.09.08). Under Maryland real estate disclosure laws, sellers are not required to
disclose information about any nearby pollution unless the for-sale property is actually
contaminated.
Additionally, simply looking at proximity to a LUST assumes that the spatial
extent of the effect on property values is the same across all sites, and homogeneous in all
directions, but this may not be true. The spread of contamination plumes are complicated
by unobserved groundwater flows (Page and Rabinowitz, 1993). Cameron (2006) shows
the importance of accounting for directional heterogeneity around a contaminated site
and presents a method for doing so, but her approach is not applicable here because the
effect of LUSTs on home values is too local and there are too few sales to statistically
analyze individual sites.
Following the traditional approach in the hedonic literature, I examine the impact
of proximity to a LUST on house prices, and how this varies across informational events
(e.g. leak discovery, cleanup, and cleanup completion). In addition, I have compiled a
unique dataset of private well contamination tests and correspondence from MDE, which
allows me to identify households who are relatively well-informed and face actual (or
-------
11
suspected) risks. Well contamination levels are observed at the end of the complicated
hydrogeological processes, and thus provide a measure of risk that already accounts for
spatial heterogeneity of contamination around an individual LUST site, and across sites.
III. Background on Leaking Underground Storage Tanks
There are about 595,000 industrial and commercial facilities that store petroleum
or other hazardous substances in underground tanks (US EPA, 201 Id). Tanks could
eventually leak as a result of corrosion, cracks, defective piping, or spills during refilling
and maintenance. Leaking contaminants can seep into the soil and local groundwater.
These pollutants may migrate to surrounding water bodies and ecological systems via
surface run-off or groundwater flows.
Human health can be adversely affected by the consumption of contaminated
groundwater, inhalation of vapors, and dermal contact with contaminants. Those most at
risk are among the 15% of Americans who rely on private groundwater wells, which are
not regulated by the Safe Drinking Water Act, and for which there are no testing,
monitoring, and treatment requirements (US EPA, 2011c).The majority of the regulated
USTs contain petroleum substances, the by-products of which include harmful
compounds, such as benzene (a proven carcinogen), and toluene, ethyl benzene, and
xylenes (commonly abbreviated as BTEX), which affect the kidneys, liver, and nervous
system (US EPA, 2011b). Furthermore, motor fuel can contain harmful additives, such
as Methyl tertiary butyl ether (MTBE), a former gasoline additive and suspected
carcinogen (Toccalino, 2005; US EPA, 2011b).
-------
12
Congress first mandated the US Environmental Protection Agency (EPA)
establish a comprehensive program regulating USTs in 1984, by adding Subtitle I to the
Resource Conservation and Recovery Act. Since then, the EPA has encouraged the States
to develop their own UST programs, for which they can seek formal approval. In total,
about 495,000 leaking underground storage tanks (LUSTs) have been identified
throughout the United States. Cleanups have been initiated at 470,460 LUST sites, and
completed at 401,874 sites (US EPA, 201 Id), making the UST program perhaps the
largest remediation program for which the EPA is responsible. For comparison, there are
currently a total of 1,298 sites on the Federal National Priorities List (NPL) and 354 sites
have been deleted (US EPA, 201 la).
With the cost of cleanup ranging from a few thousand to millions of dollars at
each LUST site (US EPA, 2004; Khan et al., 2004), and given large number of sites,
potential health risks, and extensive government involvement, it is useful and important
to find out what the benefits of cleanup and prevention are.
IV. The Empirical Model
IV. A. A Difference-in-Difference Approach
Consider a single housing market. I posit that the price of home i in neighborhood
j at period t is a function of structural characteristics of the home (e.g. interior
square footage) and its location, where the latter includes UST facilities and perceived
environmental and health risks associated with that location. Formally,
-------
13
Pyt = USTijt,ni}-t) , where ;c!Jt denotes home structure and neighborhood
characteristics, USTljt is the presence (or number of) UST facilities in close proximity,
and Mijt denotes perceived environmental and health risks.
Risk perceptions are formed from a given information set about the disamenity
and location: = ¦E(USTijt,LUSTijtfTestit ) . The vector LUST^ denotes the
presence of a leaking UST within a given distance of home i in each of the three stages of
the contamination/cleanup process.
Briefly, based on MDE practice, if a leak is (i) discovered then an investigation is
undertaken by the environmental regulators to assess the situation and determine the
appropriate actions. MDE may require that (ii) cleanup be undertaken, which could
include removal of the tank, excavation of contaminated soil, and the extraction and
treatment of groundwater, among other things. Not all LUSTs undergo active cleanup
efforts. Petroleum products naturally degrade over time, so if there is no public or
environmental threat then ongoing monitoring and natural attenuation are sometimes
deemed the best course of action (US EPA, 2004; Khan et al., 2004). If cleanup is
undertaken, it is usually complete by the time the leak investigation enters the third and
final stage, (iii) closure of the case, which is reached when the regulatory agency no
longer considers the LUST a threat.
If buyers and sellers in the housing market are aware of a LUST and perceive it as
a disamenity or risk to human health, then one would expect home prices to decrease
upon the discovery of a leak, and to rebound back to pre-leak levels after cleanup. If these
conditions do not hold, then property values may be unaffected. It is also possible that
-------
14
prices may not rebound after cleanup because the site may still be perceived as a threat,
or because of a residual social stigma (Messer et al, 2006; Gregory and Scatterfield,
2002).
A unique aspect of this study is the inclusion of home-specific information
regarding leaks and pollution in private groundwater wells, denoted Testit. If the
Maryland Department of Environment (MDE) suspects that contamination has migrated
into a private groundwater well, they will notify the residents, usually with a personal
letter informing them about the LUST and requesting to test their well. After testing,
MDE sends a follow-up letter with the test results and regulatory standards. If
contamination is found, additional tests and notification letters may occur. This is not
common to all homes near a LUST, thus the households whose wells are tested are
relatively well informed about actual or potential risks.
I do not observe perceived risks directly and must therefore estimate a reduced-
form hedonic model. Assuming a log-linear functional form, the model is
lnpi/t = + USTiJt y + LUSTijt0 + Testita + Vj + Mt + siJt (1)
where Vj is a neighborhood specific fixed effect to control for all unobserved time-
invariant neighborhood influences, Mt denotes quarterly and annual fixed effects to
o
capture overall market trends, and s^t is a disturbance term. The coefficients to be
estimated are fi, v, 0, and a.
USTl]t includes all UST facilities, whether leaking or not. The coefficient y
captures the baseline effect of desirable and undesirable characteristics associated with
-------
15
these facilities and the surrounding area. This yields a clean quasi-experimental
framework where the "treatment" is the discovery of a leak (denoted by LUSTlJt). Home
sales around registered USTs that never leak serve as a control, sales around LUSTs
before the leak is discovered are the treated group before the treatment, and sales after a
leak is discovered are the treated group after the treatment. This is similar to what Horsch
and Lewis (2009) refer to as a spatial difference-in-difference approach.
Assuming that the unobserved characteristics captured by y do not change over
time in a manner correlated with LUST events, the elements of vector & are the causal
effects of (i) a leak being discovered, (ii) contamination being cleaned up, and (iii) a leak
case being closed, on the value of homes in close proximity. The coefficient a is the
additional impact of groundwater well testing and contamination on the value of a home.
IV.B. A Propensity Score Approach
In addition to OLS estimation of equation (1), I also pursue several variants of it.
The first entails a two-stage control function or propensity score type of approach
(Wooldridge, 2002, Chp. 18). This framework essentially compares the price of homes
around leaking UST facilities to a control group of homes around non-leaking facilities
that have similar propensities for a leak to be discovered. This approach may better
control for confounding influences associated with a particular facility that are correlated
with the discovery of a leak. For example, if larger UST facilities are more likely to leak,
and larger facilities are more of a nuisance to residents, then we must properly control for
such heterogeneity in the baseline in order to accurately estimate 0.
-------
16
The "propensity score" approach is done in two steps. In the first step I estimate a
probit model of the probability that a leak is discovered at each individual UST facility.
Formally:
where leakk is a dummy variable equal to one if a leak is discovered at UST facility k,
and tankk is a vector of characteristics of the facility (e.g., age, number of tanks, site
use) and its location (e.g., hydrogeology, exposure pathway and receptors, neighborhood
socio-demographics). # is a vector of unknown coefficients. I posit that the propensity
that a leak is discovered follows a normal distribution, so #{¦) is a standard normal
cumulative density function. A probit model is then estimated via the method of
maximum likelihood.
The predicted propensity of a leak at each UST facility [PSk = is
then used to derive the expected number of leaks around each home, formally
where denotes the set of all UST facilities in close proximity to home i (e.g., 500
meters). In the second step, I add leaksi to equation (1), yielding:
Inpyj. = + USTiJtrf + leaks ^ + LUSTijt0
PSk — Prob(leakk) = Q>(tankkilr)
(2)
leakSi = Zk&iPSk
(3)
+Testita + Vj + Mt + . (4)
-------
17
Similar to a propensity score model, in (4) I estimate the average treatment effect 0 (i.e.,
the effect of a LUST on home values) conditional on the propensity for "treatment." In
theory, the estimated coefficient
-------
18
Let n denote the number of observed home sales, then P is a n*l vector containing the
natural log of the price for all sales, Wj is a row-standardized n*n spatial weight matrix
defining the neighbor relations, and p is the spatial lag coefficient which is meant to
absorb unobserved and potentially confounding characteristics associated with the
location of a home. The model later estimated also accounts for spatial autocorrelation. In
other words, in equation (6), s — W2e + u, where IV: is a row-standardized nxn spatial
weight matrix defining the relationship between neighboring disturbance terms, and u is a
nx 1 vector of iid error terms.
V. The Data
The empirical analysis focuses on single-family home sales from 1996-2007 in
three Maryland Counties: Baltimore, Frederick, and Baltimore City (see figure 1). I focus
on Maryland because a comprehensive dataset of home transactions was available, and I
could physically access the leak investigation files at the Maryland Department of
Environment (MDE). I selected these counties because they have a good mix of urban
and rural areas, and homes served by public water versus private wells.9 The four main
components of the dataset are described below.
V.A. Underground Storage Tanks
The State of Maryland requires all tanks meeting certain criteria be registered
with its Department of Environment (COMAR 26.10.02). MDE's Oil Control Program
provided data on all registered USTs in Maryland. Attention is restricted to the 3,516
registered UST facilities in Baltimore (1495), Baltimore City (1562), and Frederick (459)
-------
19
Counties.10 Table 1 shows that the majority of UST facilities are in areas served by
public water, but there are 426 USTs in areas where households rely on private
groundwater wells. Among the 1,300 UST facilities where the use is listed, 574 (44.2%)
are gas stations, 305 (23.5%) are classified as commercial, and 421 (32.4%) as
industrial.11 The average UST facility has three tanks and a total capacity of 17,363
gallons. Just over half (53.9%) of the facilities had no active tanks at the beginning of the
study period. From 1996 to 2007 leaks were discovered at 138 (3.92%) of these
registered UST facilities (see figure 2).
V.B. Leaking Underground Storage Tanks
MDE's Oil Control Program provided data on 42,100 petroleum "cases" in
Maryland, which includes routine compliance checks, the opening and closing of USTs,
and leak investigation and remediation cases. Out of these cases, 284 pertain to leak
investigations for vapor intrusion, or soil and groundwater contamination, in the study
area, and that were first opened between 1996 and 2007. Lesser cases where
contamination was not found or was minimal and could not conceivably affect property
prices were disregarded, leaving 255 cases. I disregard investigations that were not linked
to a UST at a valid address, leaving 219 cases. To ensure a relatively homogeneous set of
LUSTs and better control for pre-leak conditions I focus only on the 138 leak
investigations that were undertaken at a registered UST facility.
The State of Maryland does maintain an electronic database of LUST cases, but
much of the key information remains in hard copy files, which are available only at the
MDE office in Baltimore. I spent over 200 hours reviewing and retrieving records from
individual case files at MDE.
-------
20
Gathering site-specific details turned out to be crucial to the analysis because
despite the fact that LUSTs involve a relatively homogeneous class of pollutants
(petroleum by-products), I find substantial heterogeneity in pollution severity, public
knowledge, and investigation activities. Among the 138 LUST sites, 34 (24.6%) were in
a private well area (see table 2). There was evidence confirming that contamination
migrated to neighboring properties in just 27 cases (19.6%). As of the end of 2007, active
cleanup efforts had been undertaken at 61 LUSTs (44.2%). Remediation technologies
included soil excavation, pump-and-treat, vacuum extraction, soil vapor extraction,
recovery sumps, containment walls, concrete caps, and bioremediation (e.g. oxygen and
enzyme injections). Considering the 84 leak cases that were resolved by the end of
2008, the average was open for 1.79 years (median 1.24 years), the shortest was a day,
and the longest was 10.48 years.
V.C. Home Sales
Data on single family homes come from Maryland Property View (MDPV) 1996-
2007, which compiles the tax assessment databases maintained by the tax assessor's
office in each county of the state. There are a total of 244,169 single-family homes with
valid geographic coordinates: 59,671 in Frederick County, 152,488 in Baltimore County,
12
and 32,015 in Baltimore City. The hedonic analysis focuses on the 132,840 sales from
13
1996-2007 for this set of homes. The average transaction rate per year is thus 4.53%.
The median price over that period is $215,063 in Baltimore County, $279,627 in
Frederick County, and $125,931 in Baltimore City (2007$).14
Descriptive statistics of the home characteristics are shown in table 3. MDPV
contains geographic coordinates and several structural characteristics for each home (e.g.,
-------
21
interior square footage, lot size, the number of bathrooms, etc.). I derived several
locational variables using a Geographic Information System (GIS) and data from various
sources.15 I define neighborhoods according to the 2000 Census block groups for
Baltimore and Frederick counties, and by census tract for Baltimore City.16 This
produces 498, 127, and 200 "neighborhoods," respectively. Other spatial attributes are
included to control for local variation within a neighborhood (e.g., distances to major
roads and public open space).
Distance from a home to surrounding USTs was calculated using GIS. The
average sale is 718 meters from the nearest registered UST, and 2.2 km from the nearest
LUST. There are 17,963 sales (13.5%) within 200m of a UST and 65,367 (49.2%) within
0-500m. Considering only these sales, there were 3.58 USTs within 500m of the average
home sold. Almost half (48.2%) of all single-family homes (not just sales) are within
500m of a UST, confirming that USTs truly are ubiquitous.
Identification of the effect of LUSTs on property values requires that transactions
occur during the various stages of the leak investigation/cleanup process. Table 4 shows
the number of sales during each of these stages that are within 0-200 and 200-500 meters
of the LUST. Notice there are relatively few sales, and thus few observations for
statistical identification, in the more rural private well areas, which is where households
face potentially higher risks.
V. I). Potable Well Contamination Tests
If MDE suspects a household's well has been contaminated by a leak, a letter is
sent notifying the residents of the situation and requesting to test their well. MDE then
-------
22
sends the test results to the residents. If warranted, regularly scheduled testing and
correspondence will continue.
During 1996-2007 there were over 7,700 potable well tests conducted at 670
different homes and businesses, 633 of which were single-family homes. Only 50 single-
family home transactions took place after the well had been tested (18 in Baltimore and
32 in Frederick counties), corresponding to 11 different LUST cases. Often MDE found
minimal contamination, if any, and it was therefore not necessary to continue testing.
However, in some cases several testing events were warranted. At 16 homes only one
well test was undertaken prior to the sale, but the well water at 34 homes was tested
17
multiple times.
If contamination is found at a residence to be sold, the prospective seller is
required by law to disclose such information. Contamination was found at 23 of the 50
18
sales where testing occurred: BTEX in 11 domestic wells and MTBE in 19 (see table 5).
Ten home sales took place where pollution levels in the potable well exceeded regulatory
standards. Granulated active carbon (GAC) filters, which essentially eliminate all
pollutants, were installed and maintained by MDE for nine of these 10 home sales.
On average, the most recent MDE-conducted well test relative to the sale date was
1.55 years prior to the sale (the median is 124 days). At 32 (64%) of the tested homes,
testing occurred both before and after the transaction, suggesting that sellers and buyers
were aware of the LUST and groundwater contamination.
-------
23
VI. Hedonic Regression Results
VIA. Base Hedonic Model Results
In the first set of hedonic price regressions I estimate several variants of equation
(1). The dependent variable is the natural log of the sale price (2007$). In table 6, I
estimate a single hedonic price function for all three counties.19 Since these counties
could be considered separate housing markets, the estimated coefficients associated with
attributes of the home and its location, as well as the year and quarter time effects, are
allowed to vary across the counties. For now, however, I constrain the estimated price
effects of a UST and leak and cleanup events to be the same across all counties.
Perceived pollution risks are measured by three dummy variables denoting that a
LUST is within 500 meters, and is in one of the three stages: i) leak discovered, ii)
cleanup, iii) post-closure.20 The corresponding regression coefficients can be interpreted
as a percent change in price relative to the pre-leak values. To absorb any unobserved
confounding influences on prices, I include a dummy variable denoting whether a non-
leaking UST, or a LUST that has not yet leaked, are within 500 meters. In a difference-
in-difference framework, these dummies denote the "control" group of homes, and the
"treated" group prior to treatment, respectively.
In model 6.A I do not include neighborhood fixed effects or other variables to
control for unobserved effects on prices. This model serves as a baseline to compare with
models where I better control for confounding influences and heterogeneity. Only the
coefficient estimates of interest are shown, but the sign and significance of those not
21
presented are as expected. The -0.0201 coefficient on non-leaking UST within 0-500m
-------
24
suggests that homes near a UST sell for about 2% less. Homes near a UST that will
eventually leak tend to sell for 7% less (as seen by LUST within 500m). Neither result
should necessarily be interpreted as causal.
The 0.1126 coefficient on leak discovered suggests an 11% increase in property
values when a leak is found and an investigation opened, which is against initial
expectations. It is possible that the public is unaware of the discovery of a LUST, or does
not perceive it as a threat, but this would imply no change in prices. This counterintuitive
appreciation may be due to omitted variables associated with a home and its location,
which I better account for in later regressions. As seen by cleanup and post-closure,
model 6.A suggests a small and statistically insignificant price effect when cleanup is
undertaken, and the investigation subsequently completed.
Model 6.B includes neighborhood fixed effects. Accounting for unobserved
heterogeneity in this fashion does reduce the counterintuitive appreciation upon leak
discovery by more than 50% (relative to model 6.A), suggesting that the neighborhood
fixed effects may be absorbing some of the omitted variable bias. However, the discovery
of a leak still seems to lead to an unexpected 4.11% increase in property values.
Leak discovery is well distributed both spatially and temporally, so the coefficient
on leak discovered is unlikely to pick up unobserved effects due to a particular location
or time period. This coefficient could, however, absorb unobserved effects that are
systematically occurring at LUSTs in different locations and time periods in a manner
correlated with leak discovery.
-------
25
Anecdotally, at least eight of the leak investigations were opened because
contamination was found during redevelopment. This is one potential explanation as to
why leak discovery is associated with an increase in home values. Unfortunately, I could
22
not identify a clear measure of redevelopment to include in the hedonic regressions.
To better control for unobserved heterogeneity that may vary over time or within
a neighborhood, in model 6.C I add the natural log of the median price of all single-
family home sales within 500 meters of a home, and that took place within 3 years prior
to the sale. Clearly this variable is endogenous, but I do not wish to make any inference
about the resulting coefficient. It is included solely to absorb potentially confounding
local (and possibly time varying) influences on home prices. This adds explanatory
power, but does not change the LUST related coefficients.
Model 6.D is a repeat sales model. Following equation (5), the home and location
specific time invariant characteristics are differenced out. The median neighbor price is
included to help control for unobserved local trends. The estimated coefficient on leak
discovered is much smaller, suggesting only a 1.6% appreciation, which is not
statistically different from zero. Cleanup seems to have a small and insignificant effect
on prices, but closure of a leak investigation leads to a marginally significant 4.4%
depreciation, suggesting a residual perception of risk or public stigma after the
environmental threat is eliminated. However, unless the residential housing market is
simply slow to capitalize this stigma, one would also expect lower home values during
the discovery and cleanup of a leak.
-------
26
VI.B. Hedonic Results by County
In tables 7 through 9, I repeat the above regressions for each county individually.
There are too few sales in close proximity to a LUST to estimate a separate repeat sales
model for each county. Instead, spatial autoregressive and autocorrelation models are
23
estimated following equation (6). Again, only the coefficients of interest are displayed.
Focusing first on Baltimore City (table 7), in the specifications controlling for
unobserved neighborhood effects (models 7.B and 7.D) we see no significant effect of
LUSTs on home prices.
The Baltimore County results (table 8) suggest that although the prices of homes
within 500 meters of a UST and future LUST site tend to be lower, the discovery of a
LUST, cleanup, and closure of a leak investigation have no adverse effect on prices. In
fact, models 8.B through 8.D suggest a counterintuitive 5% appreciation in home prices
upon the discovery of a leak. There is some evidence that home values appreciate 2.68%
to 5.41% upon cleanup and closure of a leak investigation, relative to before the leak was
discovered (see models 8.C and 8.D).
In Frederick County (table 9), the results suggest that the discovery of a LUST
leads to a 2.53% to 4.68% increase in home prices, although this is only marginally
significant in some specifications. There is some weak evidence that prices are slightly
lower during the cleanup of a LUST site (but this is only marginally significant at best).
I next examine whether the baseline effects of USTs on home values vary
according to the type of UST facility, and whether this affects the estimated LUST
impacts. In the state of Maryland's database of registered USTs, facilities are classified as
-------
27
gas stations, commercial, industrial, or as "unknown" (which means that the use was not
specified on the UST inspection report).
In table 10, for Baltimore City (model 10.A) we see some evidence that the
baseline effect of proximity to a UST on home prices depends on the type of facility. For
example, homes within 500 meters of a gas station are associated with 6% lower values,
all else constant (although this effect is not statistically significant at the conventional
levels, p-value = 0.11). On the other hand, among UST facilities where the use is not
specified, prices tend to be 4% higher. Based on a Wald test, for Baltimore City I reject
the null that the baseline price effects are equal across different types of UST facilities (p-
value=0.0316). For Baltimore (model 10.B) and Frederick (model 10.C) counties,
however, I fail to reject this null hypothesis, suggesting that different types of UST
facilities are associated with similar effects on the price of homes within 500 meters.
Comparing the results in table 10 to those in tables 7 through 9, it appears that
controlling for the baseline effects across different types of UST facilities does not affect
the estimated impacts of the discovery of a leak, cleanup, and closure of a leak
investigation. One exception is that model 10.A suggests that in Baltimore City, home
prices decrease by 9.97% upon closure of a leak investigation, as opposed to the
statistically insignificant 6.6% decline found when we do not distinguish between the
type of UST facility (see model 7.C).
Focusing just on Baltimore and Frederick counties, in table 111 examine whether
the presence of an exposure pathway matters. Separate regressions are estimated for
homes in private well areas versus those in areas connected to the public water system.
-------
28
Again we see that leaks and related activities generally have a small and often
insignificant effect on surrounding home prices.
Previous specifications sometimes revealed an unexpected appreciation in home
values when a leak was discovered. The results here suggest that this primarily occurs to
homes connected to the public water system, and hence where the primary exposure
pathway to contamination is not present. As seen in model 1 l.A, the discovery of a LUST
is associated with a 4.21% increase in price for homes connected to the public water
system. Among homes that rely on private wells (model ll.B), however, leak discovery
is associated with a statistically insignificant 2.95% appreciation in value.
In short, the analysis thus far suggests that, if anything, leak discovery is
associated with an unexpected small, albeit statistically significant, increase in
surrounding home values. However, this result does not hold across all specifications,
and is most notably absent in the results focusing on homes that rely on private
groundwater wells (which consists of households who are most at risk), and the repeat
sales model (which one could argue best controls for omitted variables). In the next
section I further examine this result using a more refined quasi-experimental framework.
VI. C. Hedonic Results from a Refined Quasi Experiment
In table 12 I focus only on homes within 500 meters of a registered UST, and as
such, this hedonic model only compares homes near USTs where a leak did and did not
occur. Model 12. A focuses on all three counties. This is a spatial difference-in-difference
model comparing home sales around registered non-leaking USTs (the control group), to
sales near LUST sites (treated group), both before and after the leak. The LUST within
-------
29
500m dummy is included to account for any "pre-treatment" price differences between
the control and treated groups. The fact that this coefficient estimate is statistically
insignificant, lends support to the selection of the control group. Even in this refined
model, the only significant result is an unexpected 4.64% appreciation upon the discovery
of a leak. A similar result holds even when focusing on homes with private wells (model
12.C).
An alternative two-step "propensity score" type of model, as shown in equations
(2)-(4), is estimated for models 12.B and 12.D. Since the treatment (i.e., the discovery of
a leak) is "assigned" to USTs, and not homes per se, I estimate the discovery of a leak at
a UST facility (see equation 2). The binary dependent variable equals one if a leak is
discovered at a UST facility from 1996-2007 (n=138), and zero otherwise (n=3,378).
Several probit models of leak discovery are estimated and discussed in the appendix.
I use the estimated coefficients from the first stage probit model to estimate the
probability (or propensity) of leak discovery at each UST facility. The mean predicted
probability of a leak among non-leaking USTs is 3.49% (median=1.28%), compared to
14.39%) (median=12.9%>) among the 138 LUSTs.24 Figure 3 displays the distribution of
predicted propensities, and shows that the common support is fairly wide (0.0037 to
0.7198).
I next calculate the predicted number of leaks near each home by summing up the
predicted leak probability for all USTs within 500 meters (see equation 3). The hedonic
model (equation 4) is then estimated. In table 12, model 12.B focuses on all homes
within 500 meters of a UST. As seen by the coefficient on the Propensity Score variable,
-------
30
each additional predicted LUST site is associated with a 5.3% decline in home prices.
Caution is warranted in interpreting this as a causal effect, but this does suggest that even
conditional on a nearby UST facility, "treatment" is not necessarily random. Nonetheless,
the inclusion of the "propensity score" does not significantly change the estimated effects
of a LUST on home prices. We still see small and statistically insignificant impacts,
except for the 4.03% appreciation upon the discovery of a leak. The results are similar
25
when focusing just on homes that rely on private wells (model 12.D).
Despite extensive efforts to reduce potentially confounding influences, I find no
robust evidence that prices are adversely impacted by a nearby LUST. Perhaps potential
buyers and sellers of homes that are merely in close proximity to the disamenity do not
perceive it as a threat, or are completely unaware of it. In the last set of regressions I
examine an alternative measure of the disamenity, where I know households are relatively
well-informed and face actual or suspected risks.
VI. I). Hedonic Results with Well Tests
The regressions in table 13 focus only on homes in private well areas, and include
a dummy denoting whether the well water at individual homes was tested prior to a sale
{Well Tested). Model 13.A considers all homes that rely on private wells. The significant
coefficient on Well Tested suggests that the price of tested homes decreases by 11.36%.
Models 13.B and 13.C focuses only on the sale of homes that are most at risk (those that
rely on private wells and are within 500 meters of a UST); both models suggest a 10.85%
and a 11.37% depreciation, respectively.
-------
31
There are only 50 sales where MDE tested the well prior to the transaction.
Despite this small number of observations the results suggest a fairly large and
statistically significant depreciation. To make sure these dummies are not just picking up
unobserved heterogeneity associated with this subset of homes, in table 14 I add a
dummy denoting observations where a transaction took place before the well was tested
{Sold before Well Test). All else constant, if sales where well testing occurred prior to the
sale are similar to those where testing occurred later, then this dummy controls for any
unobserved heterogeneity associated with this subset of homes. There is little change in
the Well Tested coefficient, and in fact, the inclusion of this control bolsters the result. I
calculate the impact as:
Well Test Impact = fexp(aWell Tested - aSold brfore Well Test) - 1} (7)
and find an 11% to 12% depreciation in these cases where the households are relatively
well-informed about the LUST and groundwater pollution, and face actual (or potential)
risks.
Although the results are not reported here, in subsequent analysis I found that if a
test shows pollution levels above the regulatory standard, then prices decrease about
14%), but this is not statistically different from the 10%> depreciation among homes where
the tests revealed no contamination, or levels below the standard.
-------
32
VII. Conclusion
The goal of this paper was to illustrate two practical issues with hedonic property
value models using an empirical application to groundwater pollution from leaking
underground storage tanks (LUSTs). Omitted variable bias and the validity of the
assumed environmental measure are of particular concern in this context. Disentangling
the implicit price of LUSTs, and hence inferring welfare impacts, is challenging because
LUSTs are relatively unpublicized pollution events, and the spatial distribution of UST
facilities and leaks, may be correlated with other confounding influences on property
values.
Focusing on three Maryland counties (Baltimore, Frederick, and Baltimore City)
from 1996-2007, I conducted a detailed hedonic study on house prices. I control for a
large set of home and neighborhood attributes in the hedonic regressions, including
neighborhood fixed effects. To further reduce omitted variable bias, I implement
difference-in-difference and "propensity score" approaches by accounting for leaking and
non-leaking tanks, and exploiting the temporal and spatial variation in the discovery of
leaks. As a robustness check, repeat sales and spatial autoregressive models are
estimated.
Hedonic models examining home values around Superfund sites and other
undesirable land uses generally rely on distance to the site and contamination related
events to measure the magnitude of the disamenity (see Boyle and Kiel, 2001; Farber,
1998). Following this approach, I find that homes simply near a LUST (e.g., 500 meters)
do not typically decline in value upon the discovery of a leak, even when an obvious
exposure pathway is present (private groundwater wells). Similarly, there is no clear
-------
33
evidence that prices respond to cleanup and closure of a leak investigation. Based solely
on this typical identification strategy, a researcher may conclude that LUSTs do not
impact home values, and therefore that there is little benefit to preventing and cleaning up
these leaks (at least as reflected in home values).
However, in this application, and perhaps others, it remains unclear whether
distance is always the best proxy measure for environmental quality. Residents who are
merely living near a LUST may not always perceive it as a threat, or may not even be
aware of it.
A unique aspect of this paper is that I incorporate an alternative environmental
measure that captures home-specific variation in information and pollution, namely
domestic groundwater well test results from the Maryland Department of Environment
(MDE). Households whose wells were tested are relatively well-informed because they
receive correspondence from MDE. The mere testing of a private well by MDE signals to
a household that there is suspected contamination, and perhaps even health risks.
Furthermore, the test results may reveal that the private well is in fact contaminated.
Among these tested homes I find a 9-12% decline in value, which reflects a real
welfare loss to these well-informed households. This result may also be partially
capturing heterogeneity in pollution severity across LUST sites, because testing is more
likely to take place at more severe sites. In either case, this illustrates the importance of
properly selecting a valid measure of environmental quality for hedonic property value
models, and brings into question the findings of past hedonic studies where more
attention to the assumed environmental measure may be warranted.
-------
34
Hedonic property value models are a useful non-market valuation tool that can be
used to value a variety of environmental goods. However, in future applications we must
be more cognizant of the assumed measure of these goods, and pay particular attention
towards what information buyers and sellers in the real estate market possess, how they
perceive this information, and how well our assumed environmental measure reflects
these perceptions.
-------
35
Appendix: Pr obit Models of the Discovery of a Leak at a UST Facility
This appendix presents the estimation results from the first step of a two-step
"propensity score" type of hedonic framework discussed in section IV.B. Following
equation (2), a probit model of the probability a leak is discovered at an individual UST
facility is estimated. The binary dependent variable equals one if a leak is discovered at a
UST facility between 1996 and 2007 (n=138), and zero otherwise (n=3,378).
The parameters in the probit model are estimated via the method of maximum
likelihood. The estimated average marginal effects are displayed in table Al. Model
Al.A includes characteristics of the UST system and facility, county dummies, and a
dummy denoting the presence of the primary exposure pathway (Private Well Area). The
results suggest that a leak is 7.9% more likely to occur, and be discovered, at a gas
station. Leaks are discovered more often among larger facilities with more USTs (as seen
by # tanks at facility).
A leak is 2.86% more likely to be discovered among UST facilities in the private
well area, where the primary exposure pathway to the surrounding population is present
and where USTs are more extensively regulated and monitored. In model Al.B, the
positive coefficient on # homes in 500m w/ Pvt Well suggests that in the presence of an
exposure pathway, the larger the potentially exposed population, the more likely a leak
will be discovered (although this effect is not statistically significant at conventional
levels). Finally, I find that leaks are less likely to be discovered when the groundwater
aquifer is relatively deep below the surface. In models Al.C and Al.D I include census
block group characteristics, which statistically speaking are not associated with the
probability that a leak is discovered.
-------
36
Works Cited
Anselin, Luc (1988), Spatial Econometrics: Methods and Models, Dorddrecht: Kluwer
Academic Publishers.
Bivand, Roger, with contributions by Micah Altman, Luc Anselin, Renato Assuntjao,
Olaf Berke, Andrew Bernat, Eric Blankmeyer, Marilia Carvalho, Yongwan Chun,
Bjarke Christensen, Carsten Dormann, Stephane Dray, Rein Halbersma, Elias
Krainski, Nicholas Lewin-Koh, Hongfei Li, Jielai Ma, Giovanni Millo, Werner
Mueller, Hisaji Ono, Pedro Peres-Neto, Gianfranco Piras, Markus Reder, Michael
Tiefelsdorf and Danlin Yu (2010), "spdep: Spatial dependence: weighting
schemes, statistics and models. R package version 0.5-21. http://CRAN.R-
proi ect. org/package=spdep.
Bockstael, Nancy and Kenneth McConnell (2006), Environmental and Resource
Valuation with Revealed Preferences. Springer Publishing. Dordrecht, Holland.
Boyle, Kevin J., Nicolai V. Kuminoff, Congwen Zhang, Michael Devanney, Kathleen P.
Bell (2010), "Does a property-specific environmental health risk create a
"neighborhood" housing price stigma? Arsenic in private well water," Water
Resources Research, 46, (W03507), 10.
Boyle, Melissa A. and Katherine A. Kiel (2001), "A Survey of House Price Hedonic
Studies of the Impact of Environmental Externalities," Journal of Real Estate
Literature, 9(2), 117-144.
Cameron, Trudy Ann (2006), "Directional heterogeneity in distance profiles in hedonic
property value models," Journal of Environmental Economics and Management,
51, 26-45.
Chattopadhyay, Sudip (1999), "Estimating the Demand for Air Quality: new Evidence
Based on the Chicago Housing Market," Land Economics, 25(1), 22-38.
COMAR (Code of Maryland Regulations), http://www.dsd.state.md.us/comarA Accessed
July 17, 2009.
Dale, Larry, James C. Murdoch, Mark A. Thayer, & Paul A. Waddell (1999), "Do
Property Values Rebound from Environmental Stigmas? Evidence from Dallas,"
Land Economics, 75(2), 311-326.
Davis, Lucas W. (2004), "The Effect of Health Risk on Housing Values: Evidence from a
Cancer Cluster," The American Economic Review, 94(5), 1693-1704.
Day, Brett, Ian Bateman, and Iain Lake (2007), "Beyond implicit prices: recovering
theoretically consistent and transferable values for noise avoidance from a
hedonic property price model," Environmental Resource Economics, 37, 211-232.
-------
37
Dotzour, M. (1997), "Groundwater Contamination and Residential Property Values," The
Appraisal Journal, 65(3), 279-285.
Ekeland, Ivar, James J. Heckman, and Lars Nesheim (2004), "Identification and
Estimation of Hedonic Models," Journal of Political Economy, 112(1), S60-
S109.
Farber, Stephen (1998), "Undesirable facilities and property values: a summary of
empirical studies," Ecological Economics, 24, 1-14.
Gayer, Ted, James T. Hamilton, and W. Kip Viscusi (2000), "Private Values of Risk
Tradeoffs at Superfund Sites: Housing Markets Evidence on Learning about
Risk," The Review of Economics and Statistics, 82(3), 439-451.
Gayer, Ted, James T. Hamilton, and W. Kip Viscusi (2002), "The Market Value of
Reducing Cancer Risk: Hedonic Housing Prices with Changing Information,"
Southern Economic Journal, 69(2), 266-289.
Gregory, Robin S., and Theresa A. Scatterfield (2002), "Beyond Perception: The
Experience of Risk and Stigma in Community Contexts," Risk Analysis, 22(2),
347-977.
Guignet, Dennis (2011), What do Property Values Really Tell Us? Evidence from
Revealed and Stated Preference Studies. Diss. University of Maryland, College
Park, Maryland; available at http://drum.lib.umd.edu/handle/1903/11994.
accessed Nov. 15, 2011.
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.
Jackson, Thomas O. (2001), "The Effects of Environmental Contamination on Real
Estate: A Literature Review," Journal of Real Estate Literature, 9(2), 93-116.
Khan, Faisal I., Tahir Husain, Ramzi Hejazi (2004), "An overview and analysis of site
remediation technologies," Journal of Environmental Management, 71, 95-122.
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 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.
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.
-------
38
Kim, Jungik and Peter Goldsmith (2009), "A Spatial Hedonic Approach to Assess the
Impact of Swine Production on Residential Property Values," Environmental
Resource Economics, 42, 509-534.
Kohlhase, Janet (1991), "The Impact of Toxic Waste Sites on Housing Values," Journal
of Urban Economics, 30, 1-26.
Leggett, Christopher G. and Nancy E. Bockstael (2000), "Evidence of the effects of
Water quality on Residential Land Prices," Journal of Environmental Economics
and Management, 39, 121-144.
LeSage, James and R. Kelley Pace (2009), Introduction to Spatial Econometrics.
Chapman & Hall/CRC Press. Boca Raton, FL.
Malone, P. and R.L. Barrows (1990), "Groundwater Pollution's Effects on Residential
Property Values, Portage County, Wisconsin," Journal of Soil and Water
Conservation, 45(2), 346-348.
McCluskey, Jill J. & Gordon C. Rausser (2003), "Stigmatized Asset Value: Is It
Temporary or Long-Term?," The Review of Economics and Statistics, 85(2), 276-
285.
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.
Messer, Kent D., William D. Schulze, Katherine F. Hackett, Trudy A. Cameron, & Gary
H. McClelland (2006), "Can Stigma Explain Large Property Value Losses? The
Psychology and Economics of Superfund," Environmental and Resource
Economics, 33, 299-324.
Page, G. W. and H. Rabinowitz (1993), "Groundwater Contamination: Its Effects on
Property Values and Cities," Journal of the American Planning Association,
59(4), 473-481.
Palmquist, Raymond B. (1982), "Measuring Environmental Effects on Property Values
Without Hedonic Regressions," Journal of Urban Economics, 11(3), 333-347.
Palmquist, Raymond B. (2005), "Property value models," in: G. Maler, J.R. Vincent
(Eds.), Handbook of Environmental Economics, vol. 2, North-Holland,
Amsterdam, pp. 763-819.
Petrin, Amil, and Kenneth Train (2003), "Omitted Product Attributes in Discrete Choice
Models," National Bureau of Economic Research Working Paper 9452,
Cambridge, MA, January.
-------
39
Pope, Jaren C. (2008), "Buyer information and the hedonic: The impact of seller
disclosure on the implicit price for airport noise," Journal of Urban Economics,
63, 498-516.
R Development Core Team (2010). R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-
900051-07-0, URL http://www.R-proiect.org.
Rosen, Sherwin (1974), "Hedonic Prices and Implicit Markets: Product Differentiation in
Pure Competition," The Journal of Political Economy, 82(1), 34-55.
Rosenbaum, Paul R., and Donald B. Rubin (1983), "The central role of the propensity
score in observational studies for causal effects," Biometrika, 70(1), 41-55.
Simons, Robert A., William Bowen, Arthur Sementelli (1997), "The Effect of
Underground Storage Tanks on Residential Property Values in Cuyahoga County,
Ohio," Journal of Real Estate Research, 14(1), 29-42.
Simons, Robert A., William Bowen, Arthur Sementelli (1999), "The Price and Liquidity
Effects of UST Leaks from Gas Stations on Adjacent Contaminated Property,"
The Appraisal Journal, 67, 186-194.
Toccalino, Patty (2005), "Human-Health Effects of MTBE: A Literature Summary,"
United States Geological Survey,
http://sd.water.usgs.gov/nawqa/vocns/mtbe hh summarv.html. accessed
September 8, 2009.
US EPA (2004), United States Environmental Protection Agency, "How to Evaluate
Alternative Cleanup technologies for Underground Storage Tank Sites: A Guide
for Corrective Action Plan Reviewers," EPA 510-R-04-002,
www.epa.gov/oust/pubs/tums.htm. accessed July 23, 2008.
US EPA (201 la), United States Environmental Protection Agency, "Cleaning up the
Nation's Hazardous Waste Sites," Oct. 3, 2011, Web,
http://www.epa.gov/superfund/index.htm. accessed Nov. 16, 2011.
US EPA (201 lb), United States Environmental Protection Agency, "Drinking Water
Contaminants," Sept. 30, 2011, Web,
http://www.epa.gOv/safewater/contaminants/index.html#listmcl. accessed Nov.
16, 2011.
US EPA (201 lc), United States Environmental Protection Agency, "Private Drinking
Water Wells," Oct. 4, 2011, Web, http://water.epa.gov/drink/info/well/. accessed
Nov. 16, 2011.
-------
40
US EPA (201 Id), United States Environmental Protection Agency, "Underground
Storage Tanks," Nov. 16, 2011, Web, http://www.epa.gov/oust/. accessed Nov.
16, 2011.
US EPA (201 le), United States Environmental Protection Agency, "Handbook on the
Benefits, Costs, and Impacts of Land Cleanup and Reuse," Web,
http://vosemite.epa.gov/ee/epa/eed.nsf/webpages/LandHandbook.html. accessed
Feb. 29, 2012.
Walsh, Patrick J., J. Walter Milon, and Davis O. Scrogin (2011), "The Spatial Extent of
Water Quality Benefits in Urban Housing Markets," Land Economics, 87 (4),
628-644.
Wooldridge, Jeffrey M. (2002), Econometric Analysis of Cross Section and Panel Data.
Cambridge, MA: The MIT Press.
Zabel, Jeffrey and Dennis Guignet (2011), "A Hedonic Analysis of the Impact of LUST
Sites on House Prices," paper presented at the Inaugural Association of
Environmental and Resource Economists Summer Conference, Seattle, WA, June
(www.webmeets.com/aere/2011/Prog/viewpaper.asp?pid=535. accessed June 17,
2011).
-------
41
Figures and Tables
Figure 2. Registered Underground Storage Tanks and Leaks.
o Registered USTs
A Leaking Registered USTs
5 10 20 30 40
Kilometers
-------
42
Figure 3. Kernel Density Estimate of the Probability that a Leak is Discovered at a UST
Facility (from Model Al.B in Table A1 of the appendix).
o
U~f '
Probability of a Leak: Kernel density estimate
i. i
i i
i i
Density
10 20
L 1
i
i
'1
I
i
l
i
l
\
U ^ ^ _
l
l
I 1
1 1 1 I I
0 .2 .4 6 .8
Probability that LUST Discovered
nonleaking UST
leaking UST
kernel = epanechnikov, bandwidth = 0.0031
Note: Gray dashed lines denote common support.
Table 1. Number of Registered Underground Storage Tank Facilities by County and
Water Source.
County: Public Water Area Private Well Area Total
Baltimore City 1,562 - 1,562
Baltimore 1,228 267 1,495
Frederick 300 159 459
Total 3,090 426 3,516
Table 2. Number of Leak Cases at Registered UST Facilities by Water Area.
Public Water Area Private Well Area Total
Baltimore City 32 - 32
Baltimore 58 18 76
Frederick 14 16 30
Total 104 34 138
-------
43
Table 3. Attributes of Single Family Home Sales in Baltimore City, Frederick, and
Baltimore Counties.
Variable
Obs
Mean
Std. Dev.
Min
Max
price of home (2007$)
132840
263877
174084
15000
1979828
interior square footage
132840
1816.56
811.37
104
7976
lot size (acres)
132840
0.4424
0.6260
0.002
5
number full baths
132840
1.7445
0.7319
1
10
number half baths
132840
0.5279
0.5473
0
10
porch size (sqft)
123402
256.11
226.88
0
4260
number of fireplaces
103165
0.7489
0.6377
0
40
basement (dummy)
132840
0.8168
0.3868
0
1
number of stories
132840
1.6479
0.4630
1
4
attached garage (dummy)
132840
0.3602
0.4801
0
1
low quality construction3
132840
0.0043
0.0654
0
1
average quality construction3
132840
0.8016
0.3988
0
1
good quality construction3
132840
0.1871
0.3900
0
1
high quality construction3
132840
0.0042
0.0647
0
1
age of home (years)
132840
39.5311
30.1462
1
207
in private groundwater well area (dummy)
132840
0.1799
0.3841
0
1
distance to central business district (kilometers)13
132840
13.37
6.88
0.178
49.79
median home price in neighborhood (2007$)°
129688
235472
141029
1
6401731
meters to nearest public open space (meters)
132840
817
1086
0
10744
distance to nearest commercial zone (meters)
132840
671
723
0
9697
distance to nearest major road (meters)
132840
989
1089
0
10496
a. Dummy variables based on classification by tax assessors.
b. Central business district defined as Baltimore's inner harbor for Baltimore County and City, and
the City
of Frederick for Frederick County,
c. Median sales price over last 3 years, for all single-family homes within 500 meters of sale.
-------
44
Table 4. Number of Sales During LUST Investigation and Cleanup Events.
Entire Area Baltimore Pvt Well
LUST Stage Entire Area (repeat sales) Baltimore Frederick City Area
0 - 200 meters
Leak Discovery
216
63
111
59
46
33
Cleanup
98
18
53
41
4
7
Post-Closure
381
63
226
21
134
11
200 - 500 meters
Leak Discovery
1097
327
567
260
270
103
Cleanup
518
94
326
175
17
19
Post-Closure
2241
360
1166
110
965
74
Table 5. Private Well Contamination Levels Prior to Sale (n=23 sales with
contamination).
# Wells
Last Test Mean Level (ppb)
Max Mean Level (ppb)
Contaminated
Mean
Median
Mean
Median
BTEX
11
22.53
0
748
45
MTBE
19
213.70
0.49
252.92
18.90
-------
45
Table 6. Base Hedonic Price Regression Results for Entire Study Area (Baltimore City,
Frederick, and Baltimore County).
Model 6.A Model 6.B Model 6.C Model 6.Da
VARIABLES ln(price) ln(price) ln(price) A ln(price)
Non-leaking UST within 500m
(dummy)
-0.0201**
-0.0067*
-0.0033
(0.008)
(0.004)
(0.004)
LUST within 500m (dummy)
-0.0724***
-0.0234**
-0.0186*
(0.026)
(0.011)
(0.010)
x leak discovered
(dummy)
0.1126***
0.0477***
0.0488***
0.0162
(0.020)
(0.012)
(0.012)
(0.019)
x cleanup (dummy)
-0.0061
-0.0051
-0.0035
-0.0140
(0.030)
(0.017)
(0.016)
(0.029)
x post-closure (dummy)
0.0136
0.0046
0.0063
-0.0439*
(0.038)
(0.014)
(0.013)
(0.025)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.1762***
0.1475***
(0.013)
(0.017)
Neighborhood Fixed Effects
No
Yes
Yes
No
(Number of Fixed Effects)
(729)
(729)
Repeat Sales Model
No
No
No
Yes
Home Characteristics:
Home Structure x County
Yes
Yes
Yes
Yes
Home Location x County
Yes
Yes
Yes
No
Year and Quarter Dummies
x County
Yes
Yes
Yes
Yes
Observations
132,831
132,831
132,831
27,128
R-squared
0.770
0.628
0.635
0.224
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by neighborhood group. (Neighborhoods are defined
by census block groups for Baltimore and Frederick Counties, and census tract for Baltimore City),
a. Repeat Sales Model.
-------
46
Table 7. Base Hedonic Price Regression Results for Baltimore City.
Model 7.A Model 7.B Model 7.C Model 7.D
VARIABLES ln(price) ln(price) ln(price) ln(price)
Non-leaking UST within 500m
(dummy)
0.0337
0.0168
0.0091
0.0374**
(0.029)
(0.017)
(0.014)
(0.019)
LUST within 500m (dummy)
0.4682***
0.0738*
0.0726*
0.0397
(0.097)
(0.041)
(0.040)
(0.053)
x leak discovered
(dummy)
-0.2110*
0.0169
0.0120
0.0270
(0.119)
(0.045)
(0.047)
(0.056)
x cleanup (dummy)
0.0076
-0.0075
-0.0012
0.0407
(0.159)
(0.094)
(0.096)
(0.182)
x post-closure (dummy)
-0.5053***
-0.0641
-0.0666
-0.0585
(0.105)
(0.044)
(0.043)
(0.061)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.1780***
(0.032)
Spatial Lag
0.0198***
(0.003)
Spatial Autocorrelation
0.7915***
(0.008)
Census Tract Fixed Effects
No
Yes
Yes
No
(Number of Tracts Effects)
(127)
(127)
Home Characteristics:
Home Structure
Yes
Yes
Yes
Yes
Home Location
Yes
Yes
Yes
Yes
Year and Quarter Dummies
Yes
Yes
Yes
Yes
Observations
22,508
22,508
22,508
22,508
R-squared
0.539
0.338
0.347
Log Likelihood
-13,788.01
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by census tract, except in model 7.D, where a
nonzero correlation is allowed for the 15 nearest neighbors.
-------
47
Table 8. Base Hedonic Price Regression Results for Baltimore County.
Model 8.A Model 8.B Model 8.C Model 8.D
VARIABLES ln(price) ln(price) ln(price) ln(price)
Non-leaking UST within 500m
(dummy)
-0.0269***
-0.0125***
-0.0076*
-0.0152***
(0.009)
(0.004)
(0.004)
(0.0003)
LUST within 500m (dummy)
-0.1413***
-0.0552***
-0.0495***
-0.0911***
(0.029)
(0.017)
(0.015)
(0.011)
x leak discovered
(dummy)
0.1095***
0.0565***
0.0577***
0.0546***
(0.024)
(0.016)
(0.014)
(0.013)
x cleanup (dummy)
0.0419
0.0092
0.0082
0.0334**
(0.043)
(0.024)
(0.022)
(0.015)
x post-closure (dummy)
0.0650*
0.0248
0.0268*
0.0541***
(0.037)
(0.017)
(0.015)
(0.011)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.2017***
(0.012)
Spatial Lag
0.0025***
(0.000)
Spatial Autocorrelation
0.6821***
(0.004)
Block Group Fixed Effects
No
Yes
Yes
No
(Number of Fixed Effects)
(479)
(479)
Home Characteristics:
Home Structure
Yes
Yes
Yes
Yes
Home Location
Yes
Yes
Yes
Yes
Year and Quarter Dummies
Yes
Yes
Yes
Yes
Observations
75,881
75,881
75,881
75,881
R-squared
0.792
0.693
0.701
Log Likelihood
-6,686.333
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by census block group, except in model 8.D,
where a nonzero correlation is allowed for the 7 nearest neighbors.
-------
48
Table 9. Base Hedonic Price Regression Results for Frederick County.
Model 9.A
Model 9.B
Model 9.C
Model 9.D
VARIABLES
ln(price)
ln(price)
ln(price)
ln(price)
Non-leaking UST within 500m
(dummy)
-0.0050
-0.0033
0.0023
-0.0013***
(0.010)
(0.005)
(0.005)
(0.000)
LUST within 500m (dummy)
0.0011
-0.0021
0.0031
0.0019
(0.024)
(0.010)
(0.010)
(0.002)
x leak discovered
(dummy)
0.0545**
0.0253*
0.0274*
0.0468***
(0.027)
(0.015)
(0.016)
(0.011)
x cleanup (dummy)
-0.0686**
-0.0140
-0.0101
-0.0323*
(0.031)
(0.017)
(0.019)
(0.018)
x post-closure (dummy)
0.0372
-0.0134
-0.0076
-0.0048
(0.045)
(0.025)
(0.022)
(0.019)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.1362***
(0.013)
Spatial Lag
0.0013***
(0.000)
Spatial Autocorrelation
0.7604***
(0.007)
Block Group Fixed Effects
No
Yes
Yes
No
(Number of Fixed Effects)
(123)
(123)
Home Characteristics:
Home Structure
Yes
Yes
Yes
Yes
Home Location
Yes
Yes
Yes
Yes
Year and Quarter Dummies
Yes
Yes
Yes
Yes
Observations
34,442
34,442
34,442
34,451
R-squared
0.878
0.863
0.867
Log Likelihood
-18,568.87
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by census block group, except in model 9.D,
where a nonzero correlation is allowed for the 15 nearest neighbors.
-------
49
Table 10. Hedonic Price Results: Varying Baseline Effects by Type of UST Facility.
Baltimore City Baltimore County Frederick County
VARIABLES Model 10.A Model 10.B Model 10.C
UST within 500m (dummy)
x gas station
x commercial
x industrial
x unknown
LUST within 500m (dummy)
x leak discovered (dummy)
x cleanup (dummy)
x post-closure (dummy)
Neighborhood Characteristics:
ln(Median Neighbor Price)
Neighborhood Fixed Effects
(Number of Fixed Effects)
Home Characteristics:
Home Structure
Home Location
Year and Quarter Dummies
-0.0626
(0.039)
0.0328
(0.025)
-0.0221
(0.081)
0.0434**
(0.019)
0.1111***
(0.038)
-0.0158
(0.054)
-0.0045
(0.102)
-0.0997**
(0.040)
0.1766***
(0.032)
Yes
(127)
Yes
Yes
Yes
-0.0141
(0.009)
-0.0183
(0.014)
-0.0212**
(0.011)
-0.0001
(0.006)
-0.0353**
(0.015)
0.0549***
(0.014)
0.0077
(0.021)
0.0199
(0.016)
0.2023***
(0.012)
Yes
(479)
Yes
Yes
Yes
-0.0026
(0.008)
0.0029
(0.013)
-0.0088
(0.009)
0.0042
(0.005)
0.0018
(0.012)
0.0295*
(0.016)
-0.0055
(0.020)
-0.0088
(0.022)
0.1363***
(0.013)
Yes
(123)
Yes
Yes
Yes
Observations
R-squared
22,508
0.348
75,881
0.701
34,442
0.867
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by neighborhood group. (Neighborhoods are
defined by census block groups for Baltimore and Frederick counties, and census tract for Baltimore
City).
-------
50
Table 11. Hedonic Price Results for Private Well v. Public Water Areas (Baltimore and
Frederick Counties).
VARIABLES
Public Water Area
Model 11.A
Private Well Area
Model ll.B
Non-leaking UST within 500m
(dummy)
LUST within 500m (dummy)
x leak discovered (dummy)
x cleanup (dummy)
x post-closure (dummy)
Neighborhood Characteristics:
ln(Median Neighbor Price)
Neighborhood Fixed Effects
(Number of Fixed Effects)
Home Characteristics:
Home Structure x County
Home Location x County
Year and Quarter Dummies x County
-0.0040
(0.004)
-0.0317***
(0.012)
0.0421***
(0.012)
0.0096
(0.016)
0.0157
(0.012)
0.2450***
(0.014)
Yes
(527)
Yes
Yes
Yes
0.0048
(0.006)
0.0102
(0.016)
0.0295
(0.020)
0.0271
(0.038)
-0.0228
(0.023)
0.1180***
(0.009)
Yes
(227)
Yes
Yes
Yes
Observations
R-squared
86,433
0.722
23,890
0.786
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by census block group.
-------
51
Table 12. Quasi-Experimental Hedonic Results for Homes within 500 meters of UST.
All Counties Private Well Area
VARIABLES Model 12.A Model 12.B Model 12.C Model 12.D
# of USTs within 0-500 m
-0.0010
-0.0005
0.0070
0.0079
(0.001)
(0.001)
(0.005)
(0.005)
"Propensity Score"
(pred. # leaks O-SOOm)^
-0.0530**
-0.0051
(0.027)
(0.056)
LUST within 500m (dummy)
-0.0118
0.0320
(0.011)
(0.023)
x leak discovered
(dummy)
0.0464***
0.0403***
0.0398**
0.0549***
(0.011)
(0.010)
(0.019)
(0.020)
x cleanup (dummy)
0.0001
-0.0074
-0.0238
-0.0043
(0.016)
(0.014)
(0.054)
(0.056)
x post-closure
(dummy)
-0.0005
-0.0105
-0.0422
-0.0154
(0.013)
(0.009)
(0.027)
(0.019)
Neighborhood
Characteristics:
ln(Median Neighbor Price)
0.1972***
0.1967***
0.1003***
0.1009***
(0.022)
(0.022)
(0.022)
(0.022)
Neighborhood Fixed Effects
Yes
Yes
Yes
Yes
(Number of Fixed Effects)
(670)
(670)
(176)
(176)
Home Characteristics:
Home Structure
Yes
Yes
Yes
Yes
x County
Yes
Yes
No
No
Home Location
Yes
Yes
Yes
Yes
x County
Yes
Yes
No
No
Year and Quarter Dummies
x County
Yes
Yes
Yes
Yes
Observations
65,367
65,367
5,252
5,252
R-squared
0.547
0.547
0.786
0.786
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by neighborhood. (Neighborhoods are defined
by census block groups for Baltimore and Frederick Counties, and census tract for Baltimore City),
f Standard Errors for predicted number of leaks are not adjusted to account for two-step procedure (see
section VLB).
-------
52
Table 13. Hedonic Price Results with Private Well Testing (only homes in private well
areas).
All Homes Within 500m of UST
VARIABLES (dep variable = ln(price)) Model 13.A Model 13.B Model 13.C
# of USTs within 0-500 m
0.0088**
0.0074
0.0082*
"Propensity Score"
(pred. # leaks 0-500m)
(0.004)
(0.005)
(0.005)
-0.0049
Non-leaking UST within 500m
(dummy)
-0.0063
(0.008)
(0.055)
LUST within 500m (dummy)
-0.0078
(0.015)
0.0269
(0.023)
x leak discovered (dummy)
0.0449**
0.0617***
0.0754***
(0.019)
(0.020)
(0.021)
x cleanup (dummy)
0.0389
0.0018
0.0193
(0.041)
(0.054)
(0.053)
x post-closure (dummy)
-0.0191
-0.0280
-0.0050
(0.023)
(0.027)
(0.017)
Well Tested
-0.1136***
-0.1085**
-0.1137***
(0.031)
(0.042)
(0.041)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.1180***
0.0998***
0.1003***
(0.009)
(0.022)
(0.022)
Neighborhood Fixed Effects
Yes
Yes
Yes
(Number of Fixed Effects)
(227)
(176)
(176)
Home Characteristics:
Home Structure
Yes
Yes
Yes
x County
Yes
No
No
Home Location
Yes
Yes
Yes
x County
Year and Quarter Dummies
x County
Yes
Yes
No
Yes
No
Yes
Observations
23,890
5,252
5,252
R-squared
0.786
0.787
0.787
*** p<0.01, ** p<0.05, * p<0.1
Note: Std errors are in parentheses. Errors are clustered by census block group,
f Standard Errors for predicted number of leaks are not adjusted to account for two-step procedure
(see section VI.C).
-------
53
Table 14. Hedonic Price Results with Private Well Tests: A Robustness Check (only
homes in private well areas).
All Homes
Within 500m of UST
VARIABLES
Model 14.A
Model 14.B
Model 14.C
# of USTs within 0-500 m
0.0084*
0.0064
0.0072
(0.005)
(0.005)
(0.005)
"Propensity Score"
-0.0177
(pred. # leaks 0-500m)
(0.058)
Non-leaking UST within 500m
-0.0059
(dummy)
(0.008)
LUST within 500m (dummy)
-0.0099
0.0222
(0.016)
(0.025)
x leak discovered (dummy)
0.0453**
0.0624***
0.0753***
(0.019)
(0.020)
(0.021)
x cleanup (dummy)
0.0410
0.0052
0.0189
(0.041)
(0.053)
(0.052)
x post-closure (dummy)
-0.0163
-0.0231
-0.0034
(0.025)
(0.029)
(0.017)
Sold before Well Test
0.0112
0.0287
0.0354
(0.019)
(0.025)
(0.025)
Well Tested
-0.1112***
-0.0963**
-0.0974**
(0.031)
(0.041)
(0.041)
Neighborhood Characteristics:
ln(Median Neighbor Price)
0.1179***
0.0998***
0.1003***
(0.009)
(0.022)
(0.022)
Neighborhood Fixed Effects
Yes
Yes
Yes
(Number of Fixed Effects)
(227)
(176)
(176)
Home Characteristics:
Home Structure
Yes
Yes
Yes
x County
Yes
No
No
Home Location
Yes
Yes
Yes
x County
Yes
No
No
Year and Quarter Dummies
x County
Yes
Yes
Yes
Well Testing Impact
-0.1152***
-0.1175***
-0.1244***
(0.033)
(0.043)
(0.041)
Observations
23,890
5,252
5,252
R-squared
0.786
0.787
0.787
*** p<0.01, ** p<0.05, * p<0.1
-------
54
Appendix Tables
Table Al. Probit of Leak Discovery at UST Facility, Estimated Average Marginal
Effects (all 3,516 registered facilities in study area).
VARIABLES Model Al.A Model Al.B Model Al.C Model Al.D
Facility Characteristics:
Industrial Facility (dummy)
-0.003288
-0.003745
-0.003613
-0.004006
(0.0080)
(0.0075)
(0.0076)
(0.0074)
Gas Station (dummy)
0.079332***
0.079571***
0.079326***
0.078071***
(0.0160)
(0.0160)
(0.0159)
(0.0156)
Facility Age
-0.000188
-0.000211
-0.000217
-0.000218
(0.0002)
(0.0002)
(0.0002)
(0.0002)
age missing (dummy)
-0.009141
-0.009566
-0.009705
-0.009827
(0.0066)
(0.0062)
(0.0062)
(0.0061)
Facility built after 1996 (dummy)
0.001162
-0.000250
-0.000281
-0.000061
(0.0132)
(0.0121)
(0.0120)
(0.0120)
Active USTs (dummy)
-0.001801
-0.001738
-0.001765
-0.001777
(0.0043)
(0.0041)
(0.0041)
(0.0041)
# tanks at facility
0.003096***
0.002949***
0.002960***
0.002945***
(0.0007)
(0.0007)
(0.0007)
(0.0007)
# previous leaks w/in 500m
0.002641
0.002228
0.002208
0.002349
(0.0028)
(0.0027)
(0.0027)
(0.0026)
Location Characteristics:
Baltimore County (dummy)
0.017067***
0.015266***
0.014088**
0.014514**
(0.0057)
(0.0056)
(0.0063)
(0.0077)
Frederick County (dummy)
0.024674***
0.024656***
0.022970**
0.023189**
-0.0111
(0.0112)
(0.0115)
(0.0135)
Private Well Area (dummy)
0.028553***
0.018831**
0.016669**
0.013952*
(0.0101)
(0.0098)
(0.0100)
(0.0092)
# homes in 500m w/ Pvt Well
0.000097
0.000100*
0.000106*
(0.0001)
(0.0001)
(0.0001)
Depth to groundwater (meters)
-0.000364**
-0.000372**
-0.000335*
(0.0002)
(0.0002)
(0.0002)
Block Group Characteristics:
median home value ($ 1000s)
0.000019
(0.0000)
% pop in poverty
-0.006715
(0.0324)
% housing vacant
0.026298
(0.0378)
% housing own occupied
0.013834
(0.0136)
Log Likelihood
-463.366110
-459.292861
-459.196168
-458.358468
Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1
-------
55
1 In the context of contaminated sites, these and other issues with hedonic property value methods are
reviewed by US EPA (201 le, chapters 6 and 9).
2 As first presented by Brown and Rosen (1982) and further discussed by Epple (1987), among others, this
second stage procedure generally lacks proper identification because buyers simultaneously choose implicit
prices when choosing a housing bundle. Two approaches have arisen to circumvent this identification
problem (see Bockstael and McConnell, Ch. 6, pg 177). First, one can make specific functional form
assumptions that imply identification mathematically. Second, analyzing several markets at once
introduces proper instruments into the hedonic price function. In essence, analyzing several markets allows
us to observe the "same" households' choices when facing different price schedules, thus tracing out the
underlying bid functions. In contrast, Ekeland et al. (2004) argue that identification of the bid and offer
functions can be obtained by using data within a single market by relying on differences in the curvature of
the hedonic price, and bid and offer functions.
3 The National Priorities List (NPL) is the list of Superfund sites which have been assessed to be the most
harmful and are therefore inline for, or in the process of, remediation through CERCLA (US EPA, 201 la).
4 Some researchers have gone beyond this approach and explicitly model risk perceptions, and how
perceptions update with new information. Focusing on a Superfund site in Grand Rapids, MI, Gayer et al.
(2000, 2002) infer a value per statistical cancer case avoided of $3.9-8.3 million. Similarly, Davis (2004)
estimates the value to avoid a statistical case of pediatric leukemia to be $3-9.2 million.
5 Simons et al. (1999) define contamination based on a 3 point scale, where 1= well test confirmed
contamination at the home, 2= home is adjacent and down-gradient from a LUST, and 3= home is adjacent
to a '1' or '2', down-gradient, and within 50-100 ft of the contamination plume. Only 11 contaminated
homes were sold which is too few for a conventional hedonic analysis. Instead they compare the actual
sales price to the predicted price from a hedonic regression that did not explicitly account for LUSTs.
6 A Lexis Nexis and Google search for local news articles from 1997-2008 on LUSTs in Maryland
uncovered 19 articles covering only 10 LUST sites. For comparison there are 138 LUSTs just in the three
Maryland counties considered in this paper. Search keywords included combinations of "Maryland," "gas
station," "leaking," "underground," "tank," "UST," "oil," "leak," and "LUST."
7 At more severe LUST sites the responsible party may be required to submit a corrective action plan,
which must provide adequate protection for human health and the environment (COMAR 26.10.09.07).
81 later assume ffIJt is normally distributed, and allow for correlation at different levels of spatial
aggregation, including clustering at the census block group or tract level, and more formal spatial
autocorrelation models (see LeSage and Pace, 2009).
9 According to the 1990 census, virtually all homes in Baltimore City are served by the public water
system, and in Frederick and Baltimore Counties 43% and 8% of the homes use private wells, respectively,
and the rest are mainly connected to the public water system.
101 disregard UST facilities that are i) classified as farms, residences, or government facilities, ii) relatively
small tanks that are not regulated by MDE, or iii) missing a valid street address.
11 Facility uses are as of May, 2009, and come from the Maryland Department of Environment's inspection
reports.
12 By home I mean a unique tax identification number that existed at least one year during the study period
and corresponds to a single-family home.
131 restrict attention to arms' length sales, and exclude home sales with a price less than $15,000 (2007$) or
greater than $2 million, a lot size greater than 5 acres or listed as zero, more than 10 full baths or 10 half
baths, no full baths listed, and interior square footage listed as zero.
-------
56
14 Prices were converted to 2007 dollars using the National Consumer Price Index developed by the US
Dept of Labor's Bureau of Labor Statistics (ftp://ftp.bls.gov/pub/special.reauests/cpi/cpiai.txt. accessed
Nov. 12, 2010).
15 Data sources included the Baltimore County Department of Public Works, Frederick County Division of
Planning, the Maryland Department of Planning, Maryland Department of Natural Resources, Federal
Highway Administration, United States Geological Survey, and Maryland Geological Survey.
16 Block groups in Baltimore City are relatively small and there are not enough single family-home sales to
include block group fixed effects, therefore tract level fixed effects are used instead.
17 Six homes had 2 tests prior the sale, seven homes had 3-5 tests, 11 homes had 6-20 tests, and 10 homes
had 21-60 tests.
18 BTEX is the summation of four commonly cited petroleum contaminants, all of which are individually
regulated by the EPA under the Safe Drinking Water Act. The Maximum Contaminant Levels (MCL) are
5 parts per billion (ppb) for benzene, 100 ppb for toluene, 700 ppb for ethyl benzene, and 10,000 ppb for
xylenes. MTBE is a former gasoline additive and suspected carcinogen. The regulatory threshold for
MTBE in Maryland is 20 ppb, which is based on the EPA's taste and odor health advisory of 20-40 ppb in
drinking water.
19 Instead of a log-linear relationship, I enter the natural log of interior square footage and lot size as
explanatory variables in the hedonic regressions. A quadratic term for age is also included. Values for a
few attributes are missing from some observations, in which case these are coded to zero and a companion
missing value dummy is included. More specifically, 29,675 (22%) sales were missing the number of
fireplaces, 9,428 (7%) sales were missing porch square footage, and 376 (less than 0.5%) were missing a
construction quality classification.
20 The results are robust to the use of other distance buffers including 100, 200, 500, and 1,000 meters. The
estimates are not reported here. See Zabel and Guignet (2010) for a more thorough examination.
21 Full results are provided by Guignet (2011, pg 77-85).
22 In earlier drafts I attempted to instrument for leak discovery, which in theory would eliminate
confounding effects such as redevelopment. An instrument was constructed by estimating the probability
that a leak is discovered at a UST in a given year, as a function of characteristics of the facility, tank
system, geology, and the 2005 adoption of stricter UST regulations in groundwater sensitive areas in
Maryland (COMAR, 26.10.02.03). This constructed instrument was then used in a two stage least squares
procedure. Unfortunately the approach did not prove fruitful, possibly because the predicted probability
that a leak is discovered in a given time period is extremely low, and due to the lack of time-varying
instruments.
23 Spatial autoregressive models were estimated in R using the "spdep" package (Bivand, 2010; R
Development Core Team, 2010).
24 These estimates are from the preferred specification (model Al.B in the appendix). The predicted
propensities, and subsequent hedonic results, do not change substantially when the other specifications in
table Alare used.
25 Typically, in a propensity score regression framework the second stage standard errors are biased
downward because they do not account for the sampling variation in the first stage parameter estimates.
Although the standard errors can be adjusted via asymptotic formulas or by bootstrapping the first stage
(Wooldridge, 2002, pg 614; Petrin and Train, 2003), such adjustments in this application are complicated
by the fact that several UST facilities can be linked to a single home sale. Since the coefficient estimate
corresponding to the predicted number of leaks is generally statistically insignificant, and its inclusion does
-------
57
not significantly change the estimated implicit price of the discovery of a leak and cleanup activities, I do
not attempt such adjustments here.
------- |