Housing Market Segmentation around Mild Disamenities
By
Hans R. Isaksona*
And
Mark D. Eckerb
March 8, 2013
a
Department of Economics
University of Northern Iowa
b
Department of Mathematics
University of Northern Iowa
*Contact Author:
Hans R. Isakson
Professor
Department of Economics
University of Northern Iowa
Cedar Falls, Iowa 50614-0129
Hans.isakson@uni.edu
Voice: 319-273-2950
Fax: 319-273-2922
Keywords: sorting, housing markets, disamenities, underground storage tanks, spatial
correlation, maximum likelihood
JEL Codes: Q51, Q53, R21
Running Title: Segmentation around Mild Disamenities
1
Housing Market Segmentation around Mild Disamenities
Abstract
This study examines housing sales for market segmentation around mild disamenity sites.
We segment households into groups by the number of disamenity sites very close to them
to more accurately estimate the parameters of a hedonic equation using a Simultaneouslyfitted Spatial Linear (SFSL) model. The results show that households living more than ¼
mile from at most one disamenity site are willing to pay a premium to live even farther
away. Households living within ¼ mile of multiple disamenity sites have no aversion to
living near the disamenities.
2
Housing Market Segmentation around Mild Disamenities
1.
INTRODUCTION
The Rosen (1974) hedonic model has become the standard model used to estimate the
implicit prices for a wide variety of environmental features (see Nelson, 2004, Simons
and Saginor, 2006, and Boyle and Kiel, 2001). However, the possibility of market
segmentation can create problems for hedonic models. If market segmentation exists,
Freeman (1979, p. 164) points out that a single hedonic model fitted using aggregated
sales data will product “faulty estimates of the implicit prices.” Straszheim (1975)
observes that allowing for multiple housing submarkets can significantly reduce the sum
of squared errors in hedonic regressions. He also observes that the coefficients of some
independent variables, such as proximity to an amenity/disamenity, can vary substantially
between submarkets. If market segmentation exists, Freeman (p. 164) recommends using
a separate hedonic price function for each segment of the market.
Housing submarkets can be segmented or stratified by a variety of factors, such as
structural features (size, age, etc.), neighborhood features (school district, median
income, etc.), or household characteristics (family size, race, etc.). F-tests will reveal if
disaggregating the housing sales data into separate submarkets lowers the sum of squared
errors. Previous studies of market segmentation include Goodman and Thibodeau (1998,
2003) in which housing markets are stratified by census track, zip code, and
combinations of adjacent elementary school districts; Jones, Leishman, and Watkins
(2003) in which housing markets are stratified by geographical areas; Kiel (1995) in
which housing markes are stratified by time of sale (before/during/after cleanup of two
superfund sites); Palm (1978) in which housing sales are stratified by MLS areas; Kuethe
3
and Keeney (2012) in which housing markets are stratified by the selling price of the
house. In all of these studies, segmentation of the housing market significantly improved
the hedonic model.
Segmentation by household aversion
Nearly all households will have a strong aversion to living near a severe
disamenity site, such as a hazardous waste site, and will pay a premium to living farther
away from it. We can also reasonably expect that household aversion to living near a
mild disamenity, such as a nearby, older, gasoline station, will be more variable. In other
words, household tastes or preferences for living near a mild disamenity can vary, and
thus households will sort themselves into submarkets around the disamenity sites. So,
segmenting a housing market into submarkets according to the degree of household
aversion to living near a mild disamenity has the potential to improve the accuracy of our
estimates of the implicit prices in a hedonic model.
When multiple, yet similar, disamenities exist within an urban area, we can group
houses by how many of these disamenities are nearby. We expect households very close
to multiple disamenities to be less averse to living near these disamenities compared to
households farther away. Therefore, we can disaggregate housing sales data into groups
based upon how many disamenities are very close to the house. This approach to
segmenting the housing market around disamenities is new to the literature.
The purpose of this study is to explore the implicit prices of a hedonic model
when housing sales data are divided into groups based upon how many disamenities are
located near each house. We choose sites with an aging underground storage tank that is
being monitored by the state department of natural resources as mild disamenity sites,
4
because sites of this type are plentiful in most urban areas and houses are often found
them. Their presence in a neighborhood would be obvious to potential buyers, and
proximity to these types of sites has been shown to have a negative impact on house
prices (Simons, Bowen and Sementelli 1997a, Zabel and Guignet, 2010).
2. PREVIOUS LITERATURE
Simons, Bowen and Sementelli (1997a and 1999) provide insight into the impact of USTs
on property values in Cuyahoga County, Ohio (Cleveland). Although they do not look at
the impact on nearby property values, Simons and Sementelli (1997b) report that
commercial sites with an aging UST suffer in terms of salability and financing, making
them less than one-half as liquid (ability to convert into cash) than clean, comparable
commercial sites in Cleveland, Ohio. Simons, et al (1997a) report about a 17% decline in
the selling prices of 83 houses within 300 feet of a site with an aging UST in Cleveland,
Ohio. Using the same data as their earlier study, Simons, et al (1999) report a 14% to
16% reduction in houses values within 300 feet of a gas station with a UST.
In a National Center for Environmental Economics (NCEE) study, Zabel and
Guignet (2010) test for a disamenity effect in houses that use well water near USTs and
report that proximity to a more publicized and more contaminated site can negatively
impact surrounding home values by more than 10 percent. But, they also find that many
aging USTs, especially the less contaminated sites, have little (negative) impact on
surrounding house values. So, it would appear that publicity about the presence of an
aging UST can amplify its influence on the prices of nearby houses.
5
These studies demonstrate that commercial sites with an aging UST can have a
negative impact on the selling prices of nearby houses. In particular, the findings from
Zabel and Guignet, suggest that proximity to a mild disamenity may not be capitalized
into the values of house located near it, while the Simon, et. al. studies suggest otherwise.
Interestingly, the mechanism through which a disamenity effect is transmitted
around aging USTs may not be the obvious one, namely contamination of underground
water. Instead, the publicity surrounding the site or perhaps the land use above the UST
may be much more important than what actually lies beneath it. Not surprisingly, aging,
poorly maintain USTs are often found beneath aging, poorly maintained commercial
buildings. So, if the above and below ground conditions are highly correlated, it does not
matter which one we observe as long as one is detectable by potential buyers. In
addition, the effect could be a stigma associated with these types of sites and have very
little to do with the presence of a UST. As a result, the specific mechanism transmitting
the disamenity influence is often confounded.
Varying degrees of household aversion to living near a disamenity can be
important, especially around mild disamenities. For example, if there were a sufficiently
large number of households with little or no aversion to living near mild disamenities,
then these households would be willing to pay a higher price than more averse
households for houses located near the disamenities, other things held constant. If no
allowances are made for the presence of housing segmentation by household aversion,
then the influence of one group of households can weaken or even wash-out the influence
of another group of households.
6
3. THE STUDY FRAMEWORK
To test for varying degrees of aversion to a mild disamenity, we need (1) a sufficient
number of house sales surrounding the disamenities, (2) a means of disaggregating
housing sales data into segments based upon aversion to living near multiple disamenity
sites, and (3) a model that allows for varying coefficients within each segment. We
achieve the first by pooling together the house sales surrounding commercial sites with
an aging UST beneath them. This pooling of the data around multiple sites is common in
the literature. Simons, et al (1997a) and Zabel and Guignet (2010) pool house sales data
surrounding multiple sites. Isakson and Ecker (2008) pool housing sales data around
multiple hog lots. As long as the sites are similar, as they are in this study, this sort of
pooling of the data is acceptable.
Dividing housing sales data into submarkets based upon aversion to living near
multiple disamenity sites can be more challenging, because there are no examples to draw
from in the literature. In this study, we use the number of disamenity sites within a very
close distance to the house (0.25 mile) to classify households according to their degree of
aversion to living near the mild disamenities. Thus, we create three groups as follows:
Low Aversion Households (LAHs) consists of houses located very near two or more
disamenity sites; Moderate Aversion Households (MAHs) consists of houses located very
near exactly one disamenity site; and High Aversion Households (HAHs) consists of
houses that are not very close to any disamenity site. Furthermore, an Analysis of
Covariance series of F-tests in an OLS context (see Ott and Longnecker, 2010)
statistically justified the creation of the segmentation of the data into these three
7
submarkets. 1 We check the sensitivity of our results to the ¼ mile definition of very
close by using definitions a bit less (0.2 mile) and a bit more (0.3 mile) than ¼ mile.
4. THE STUDY AREA AND DATA
This study uses data from Cedar Falls, Iowa, a medium sized, Midwestern college town.
The college (University of Northern Iowa) has about 12,000 students, over 95% of which
live on or very near campus. The city has two junior high schools and several elementary
schools of nearly equivalent quality that feed students into the city’s only public high
school. The city is highly decentralized with the average commute time only 14 minutes.
There is no dominant destination point, such as a central business district. Instead, jobs
and retail shops are disbursed throughout the city. Preliminary analysis of the data
reveals that premiums are not paid to gain proximity to any particular site in the city
(including the college), largely due, we suspect, to the short commute times. Several
small retail strip malls exist as well as a downtown area containing several restaurants,
retail shops, banks, a small hotel, and a small theater. The downtown area is also within
a mile of the city’s sewage treatment plant (STP), a source of occasional noxious odors
depending on weather conditions. The city has a newer (1995) industrial park containing
primarily warehouses on its far south side that is home to various, very low polluting
businesses and warehouses.
The city of Cedar Falls provides all utilities to its residents, including electricity,
gas, refuse, cable TV, internet, and water. All houses are connected to the city’s water
supply. Potable water comes from eight ground-water wells drawing from the SilurianDevonian limestone aquifer. The wells range in depth from 147 to 275 feet. The water
8
from this source is of such high quality that little treatment is needed; only chlorine and
fluoride are added at the well sites. Contamination of potable water supplies by the
disamenities in this study has not been reported by the IDNR.
Housing Sales and Census Data
Initially, the housing sales dataset contained every sale of a single-family house in the
city of Cedar Falls from January 2000 to November 20042; these sales were parsed by
using only those identified as “arm’s length transactions” by the county tax assessor’s
office. This time period was selected in order to avoid the contamination effects of the
recent housing bubble and subsequent crash. Additional refinement consists of choosing
only those sales with a selling price greater than $32,000 or less than $400,000, houses
with at least three but less than 12 rooms, at least 500 square feet of living area, and a lot
size less than 3 acres.3 Several homes were repeatedly sold (but not enough to perform a
repeat sales analysis) in this time frame; we use only the most recent sale.
Disamenity Data
We define a mild disamenity site as a commercial site with an aging UST beneath it. The
50 mild disamenities in the region are mostly neighborhood gasoline stations. We note
that none of them are located near heavy industrial sites (see Deaton and Hoehn (2004).
Information for the 50 mild disamenity sites was obtained from the Iowa Department of
Natural Resources (http://www.iowadnr.gov/land/ust/ustdbindex.html).
Registration of USTs is required by federal law (40 CFR parts 280 and 281). In
Iowa, all USTs, except certain small, residential USTs, must be registered with the Iowa
DNR (Iowa Code 455.B.474). These tanks must also be inspected for their likelihood of
contamination of the surrounding land. If the site containing the UST has evidence of
9
contaminants that do not exceed certain delimited levels, then it is classified as requiring
no further action. If the onsite contaminants exceed the delimited levels, but no receptors
(aquifer, plastic water lines, basement, sewers, or surface water) are within the transport
plum of the contaminants, then it is rated as a low risk and must be tested annually. If
receptors exist within the transport plum of the contaminants, then it is classified as high
risk and faces more intensive monitoring, testing, and possible remediation. Due to the
extreme depth of the aquifer source of city water, none of the USTs in this study show
signs of contaminating the city’s potable water supply. The IDNR data includes the exact
state-plane coordinates of each UST site.
Table 1 contains the summary statistics of the data, while Figure 1 shows the
location of all 2,078 sales together with the STP, the 50 disamenity sites, the Cedar River,
major highways, and Census block group boundaries. The classification of each
disamenity is indicated in Figure 1 using different sized boxes.
5. THE EMPIRICAL MODEL
The degree of aversion for each household to living near an aging UST is
explicitly measured using a count variable: the number of very close disamenity sites,
similar to Ihlanfeldt and Taylor (2004). A distance of ¼ mile is chosen to define “very
close.” That is, the count variable is equal to the number of disamenity sites within ¼
mile (about three city blocks) of the house.
After segmenting the data into submarkets based on the degree of household
aversion to living near disamenity sites (LAHs, MAHs and HAHs), we use the previously
cited Rosen hedonic approach to model the relationship between the price of a house and
10
its characteristics. Key aspects of the hedonic model are discussed below, namely, (1)
the set of variables (dependent and independent) used in the model, (2) the functional
relationship among these variables, and (3) the structure of the error term of the model.
The dependent variable
We select the selling price of a house as the dependent variable in our empirical
model. Selling price is a popular choice in hedonic models, but selling price per unit area
(dollars per square foot) is also common. The only difference between these two choices
is the role that area plays in the set of independent variables. Using selling price per
square foot of living area is equivalent to using selling price as the dependent variable
and then multiplying every independent variable by the square footage of living area.
Since we cannot justify having independent variables weighted by living area, we choose
to use the selling price of the house as our dependent variable.
The independent Variables
We select independent variables that reflect the (1) structural characteristics of the
property, (2) proximity of the property to points of influence (other than the mild
disamenity sites), (3) characteristics of the neighborhood in which the property is located,
(4) date at which the sale occurred and (5) disamenity specific data.
The structural characteristic data include all of the publically available, house
specific information: its living area, size of the lot, number of rooms and date built.
Although more detailed housing data would be desirable, the above variables capture the
most important features of a house relevant to buyers. Yet, there will be some, hopefully
small, amount of variance in housing prices due to missing characteristics, such as
number of bathrooms, style of the house, etc. Our model developed in the next section
11
includes a spatial correlation term that helps to mitigate the impact of missing variables
as well as any incorrectly specified variables (Brasington and Hite, 2005).
We also include variables that measure the Euclidian distance to various points of
influence (other than the 50 disamenity sites), including a sewage treatment plant (STP),
a nearby river, the nearest public school, and the nearest highway. The STP is located
just south of a small downtown retail area and, occasionally, emits offensive odors that
dissipate with increasing distance. The river is a major feature of the community, and
may positively influence the selling prices of houses located near it. Proximity to a
highway can save commuters travel expenses and time, and therefore, should positively
influence selling prices.
Characteristics of the neighborhood in which a house is located (census block
groups) can also affect a house’s selling price. We include median rent, median year that
houses were built, percent of housing units that are owner occupied, the number of
housing units, and median household income of the neighborhood in which the house is
located as additional independent variables. We expect that the selling prices of houses
will be positively related to the median rent, median year that houses were built in the
neighborhood, the percent of housing units that are owner occupied, and the median
household income of the neighborhood. We expect selling prices of houses to be
negatively related to the number of housing units in the neighborhood.
The date of sale is also included as an independent variable to control for the
effects of local market price trends. We expect to find gradually rising house prices
during the period of time included in this study.
Disamenity variables
12
We also include disamenity specific variables. The Euclidian distance for each house to
the nearest disamenity, the condition of the nearest disamenity, and, when appropriate,
the number of disamenities that are very close (within ¼ mile) to the house are included
to capture the influence of nearby disamenity sites.
For each sale, we include a disamenity proximity variable: the Euclidian distance
to the nearest disamenity. The average distance for a house to the nearest mild
disamenity is 0.48 miles with 27.6% of the sales (574/2078) having at least one aging
gasoline station within ¼ mile. The coefficient of this variable is the primary focus of
this study. A positive and statistically coefficient on this proximity variable is consistent
with the site being a disamenity. We are interested in any variations in this coefficient
across the three groups of households with different aversion: LAHs, MAHs and HAHs.
We expect that this coefficient will be very low, perhaps statistically insignificant, in
LAHs, slightly positive and statistically significant in MAHs, and strongly positive and
statistically significant in HAHs.
The number of disamenities very close to a house is relevant only for LAHs, since
the number of very close disamenities is defined to be one for MAHs and zero for HAHs.
However, for LAHs, we anticipate that house prices will not be affected by the number of
very close disamenities, because we sense that any household already living near two
aging gas stations, will not mind living near more than two.
We also include the IDNR classification of each disamenity in our analysis to
capture a potential information effect. Six of the disamenity sites are classified by IDNR
as High Risk (receptors exist within the transport plum of the contaminants), four are
classified as Low Risk (no receptors within the transport plum of the contaminants), and
13
the remaining 40 are classified as No Further Action Required (contaminants do not
exceed delimited levels). We create a binary variable to capture the classification of the
disamenities as follows: Condition equals one if the disamenity is classified as low risk
or high risk; zero if it is classified as no further action required. We combine the low and
high risk disamenity sites because there are so few (four) low risk disamenity sites. The
information about the condition of the disamenity sites is publically available at the
IDNR’s web site, but finding it takes considerable time and effort. Therefore, households
may not actually know the condition of a UST under a nearby gas station. If households
are sensitive to the condition of a disamenity (found at the IDNR website), then we
expect the coefficient of the condition variable to be negative and statistically significant.
A statistically insignificant coefficient for the condition variable is consistent with (1)
households not undertaking the search to find the condition information or (2) households
know the condition, but it does not impact their decision to live near a disamenity.
The Functional Form of the Empirical Model
Various functional forms for hedonic models are possible. The log-linear
specification is popular, in which the log of selling price is the dependent variable and the
independent variables are entered in a linear fashion. In the log-log form, all of the
variables are entered in the log form. We use a combination of the log-log form for some
independent variables and the log-linear form for the rest of the independent variables.
The manner in which the distance to the nearest disamenity is entered into the
model is not a trivial matter. The functional form of this important variable must be such
that it diminishes with increasing distance from a disamenity. Examining the coefficient
associated with distance to the disamenity will indicate the implicit price (marginal
14
willingness to pay) to live farther away from a disamenity. Therefore, we enter the log of
distance to the nearest disamenity to capture its influence. Because the dependent
variable is the log of price, the coefficient for the log of distance to the nearest disamenity
will represent an estimate of the elasticity of house price with respect to proximity to the
nearest disamenity.
Mean structure of the model
The mean structure of the empirical model contains several types of explanatory
variables typically thought to explain house prices. Most broadly, these variables include
(1) site level non-spatial characteristics of the home (house size characteristics including
living area, number of rooms, parcel size, year built, and time of sale), (2) spatial
characteristics (distances to the STP, a river, and the nearest highway), (3) neighborhood
data (median rent, median year that houses were built, percent of housing units that are
owner occupied, the number of housing units, and median household income), and (4)
disamenity level factors (distance to the nearest disamenity; classification of the nearest
disamenity, and number of disamenities within ¼ mile). Specifically, let
P = the selling price of the house,
S = lot size in acres,
H = the living area of the house,
d = distance to the nearest disamenity, and
C = condition of the nearest disamenity,
CT = number of disamenities within ¼ mile of the house.
t = the time of the sale,
R = number of rooms in the house,
D = a vector of variables measuring distance to points of influence,
N = a vector of neighborhood level variables,
15
We model the mean structure using the following non-linear functional form,
(1)
𝑃 = 𝛼𝑆𝛽1 𝐻𝛽2 𝑑 𝛽3 𝑒 𝛿𝐶+𝜆𝐶𝑇+𝜅𝑡+𝜔𝑅+𝜃𝐷
where the Greek letters represent parameters of the hedonic model. On the natural log
scale, the hedonic model (1) becomes:4
(2)
ln 𝑃 = ln 𝛼 + 𝛽1 ln 𝑆 + 𝛽2 ln 𝐻 + 𝛽3 ln 𝑑 + 𝛿𝐶 + 𝜆𝐶𝑇 + 𝜅𝑡 + 𝜔𝑅 + 𝜃𝐷
The lot size (S) and house size (H) parameters, 𝛽1 and 𝛽2, represents the (constant)
elasticity of price with respect to lot size and house size, respectively (see Isakson and
Ecker, 2001 and Ecker and Isakson, 2005 for discussion). The key parameter, 𝛽3,
represents elasticity of house price with respect to distance from the nearest disamenity.
The parameters associated with the remaining explanatory variables represent the
percentage change in price given a one unit change in the variable.
Structure of the error term and spatial correlation
To estimate the parameters of (2), we must include an error or disturbance term, ε,
in (2). Standard Ordinary Least Squares (OLS) regressions assume that  ~ N (0, 2 )
where normality and constant variance (  2 is a variability parameter, after accounting for
site-level covariates) is assumed. When dealing with spatial data, such as housing sales,
the potential exists for the OLS parameter estimates to be biased, especially any spatial
distance parameter estimates in the mean structure (such as distance to the STP and
distance to the nearest disamenity site). Visual examination of an empirical variogram
(Cressie, 1993) of the OLS residuals in (2) reveals the presence of spatial correlation.
Therefore, efforts to model this source of bias could be beneficial.
16
Spatial linear models assume that the errors are not independent, that is, two
comparable homes that are closer in space sell for a more similar price than two
comparable homes farther apart. For example, houses located near each other are also
near the same neighborhood amenities/disamenities. In this case, the selling prices of
nearby, comparable houses tend to be more highly correlated than comparable houses
farther apart. We build this spatial correlation into the model by assuming that
 ~ N (0, 2   2 )
(3)
where  2 is called the “nugget” , i.e. a micro-scale or measurement error variability, in
the geostatistical literature (Cressie, 1993). The sum  2   2 in equation (3) is termed
the spatial variability of the spatial process or “sill” (the variability of the home prices
after adjusting for individual home characteristics). Finally, for two home sales with
errors  i and  j , their spatial correlation is modeled as a function of their Euclidean
distance apart, d ij . Specifically, we adopt the spherical correlation structure, i.e.,
(4)
𝐶𝑜𝑟𝑟(ln(𝜖𝑖 ), ln⁡(𝜖𝑗 )) = 12(𝜙3𝑑𝑖𝑗3−3𝜙𝑑𝑖𝑗 +2)⁡⁡⁡⁡⁡⁡𝑖𝑓⁡𝑑𝑖𝑗⁡≤⁡⁡𝜙1 ⁡.
The parameter  directly controls the spatial correlation in the dataset and is termed the
“range” (technically,
1

is the exact value of the range in equation (4) (see Ecker, 2003)
for the spherical correlation structure). Thus, any two homes that are separated by a
distance of more than the range, have selling prices that are essentially uncorrelated.5
We make use of a Simultaneously-fitted Spatial Linear (SFSL) model by
combining the traditional hedonic model, (2), together with a spatial correlation
disturbance term, (3) and (4). We estimate the parameters of the SFSL model using SAS
17
for the LAHs, MAHs and HAHs separately. The spatial correlation terms (3) and (4) are
random effects models designed to capture extra spatial variability not explained in the
mean structure of the model.
6. MODEL FITTING AND RESULTS
To fit the SFSL model given by equations (3) and (4) simultaneously with the mean
structure coefficients in equation (2), seed values are required for the range, sill and
nugget spatial parameters. We use S-Plus to fit a theoretical variogram to the empirical
variogram from the residuals from the OLS regression to obtain these seed values. The
coefficients of the model are then simultaneously estimated using Proc Mixed (a
maximum likelihood technique) in SAS, and the results are reported in Table 2.
Structure and neighborhood variables
The size variables of Living Area, Number of Rooms and Parcel Size are all
positive and strongly significant in all three groups. As expected, bigger houses sell at
higher prices. Year built (minus the Median year built for the census block in which the
house presides) is positive and significant in all three groups; newer houses sell for higher
prices, compared to the average house in the region. Date sold is positive and significant,
indicating a 4.9% annual rate of appreciation over the 4 years of sales for HAHs, 5.4%
for MAHs and 5.3% for LAHs.
Distance to the STP is positive and strongly significant for HAHs. Due to the
smaller sample sizes associated with LAH and MAH sales, the distance to the STP
coefficient is actually larger for both the LAHs and MAHs, but only moderately
significant (P-value(HAH) = 0.0438 and P-value(MAH) = 0.0885). In all three groups,
18
no one wants to live near the sewage treatment plant. Distance to the nearest school is
significant and negative for the MAHs only, indicating that MAHs are willing to pay a
premium to live near a school. Distance to the nearest highway is marginally significant
for the MAHs (p-value=0.0848) and HAHs (p-value=0.0631). Neither group desires to
live near a highway. HAHs are willing to pay a premium to be close to the river.
For the census group block variables, household income is positive and
statistically significant for the MAHs and HAHs, but weak in magnitude. Prices and
incomes are, not surprisingly, highly correlated. People with higher incomes tend to buy
slightly more expensive houses. The number of occupied units is only statistically
significant, at the 0.05 level for the LAHs, while median rent is not significant for any
group. Percent owner occupied and median year built are only statistically significant in
HAHs, suggesting that HAHs prefer neighborhoods with newer and fewer, non-owneroccupied houses.
Disamenity Variables
Overall, the parameter estimates for the disamenity variables are consistent with
our expectations. The results indicate that the three groups of households, LAHs, MAHs
and HAHs, have significant differences in their tastes for some of the neighborhood
variables. The non-significant distance to the nearest disamenity variable for the LAHs
suggests that these buyers do not have an aversion to living close to multiple
disamenities. The positive and nearly significant coefficient for proximity to the nearest
disamenity for the MAHs (P-value = 0.1178, with only n=209) indicates that they might
be willing to pay a higher premium to be farther from the disamenity than the LAHs. The
positive and strongly significant distance to the nearest disamenity variable for the HAHs
19
indicates that these households are willing to pay a statistically significant premium to
live farther away from a disamenity. In particular, the coefficient for distance to the
nearest disamenity for the HAHs is actually smaller than the same coefficient for the
MAHs, but it is about five times as large as that for the LAHs. The elasticity of house
price with respect to proximity to a disamenity for the HAH is statistically significant and
indicates that these households are willing to pay a premium to live farther away from a
disamenity. In particular, HAHs are willing to pay 9.29% more for a house that is 10%
farther away from the nearest disamenity.
In addition, the count and condition disamenity variables reveal an interesting
pattern. The count variable is only relevant for the LAHs, because count is defined as
one for the MAHs and zero for the HAHs. The LAHs do not seem to mind if they live
near two or more disamenities, as indicated by the lack of statistical significance for the
proximity-to-disamenity parameter. The condition of a disamenity does not seem to
matter to any of the households. It would appear that the publicly available (at the IDNR
website) information does not influence household decisions regarding living near a
disamenity. Nothing about a disamenity influences the LAHs (as expected), while all
that matters for the MAHs and HAHs is their proximity to the nearest disamenity.
The spatial association parameters of range, sill and nugget for the SFSL model
are also reported at the bottom of Table 2. For all three groups, the nugget and sill
estimates are fairly close, while range parameter is 166 feet (0.0166 times ten thousand
feet) for the LAHs, 333 feet for the MAHs and 677 feet for the HAHs.
Sensitivity Analysis
20
Lastly, we use sensitivity analysis to test the choice of ¼ mile as the distance to divide
the sales data into three distinct groups, LAHs, MAHs and HAHs. We redefined these
three groups using two other distances: 0.2 and 0.3 miles. Then, we fit the SFSL model
for each group. Table 3 reports the results. The distance to the nearest disamenity
variable intensifies (increases) for HAHs, as the distance to the nearest disamenity
increases. This is consistent with a stronger aversion for the HAHs. The mild aversion
observed for MAHs at a distance of ¼ mile has disappeared by increasing the distance to
0.3 miles. No disamenity variable, including count itself, is significant for the LAHs,
regardless of the choice of distance (0.2, 0.25, or 0.3 miles). Lastly, condition is only
significant for the MAHs at distance of 0.2 and for the HAHs at distance of 0.3,
indicating that these groups might be aware of the condition of the disamenities.
However, this awareness effect is not consistent in terms of statistical significance and
changing signs across the groups and distance definitions.
7. SUMMARY AND CONCLUSIONS
This study develops a model of household aversion around mild disamenities: aging
gasoline stations. It estimates the parameters of a hedonic model modified to include
spatial correlation in the error term. Using housing sales data from a medium sized,
Midwestern city, the parameters (mean structure and error terms) are estimated
simultaneously using a maximum likelihood regression technique, the SFSL model.
We find evidence that households within ¼ of a mile of multiple disamenities do
not have a statistically significant aversion to living near disamenities. Neither the
number of disamenities nearby nor the condition of the nearest disamenity has an
21
influence on selling prices. Households living near (within ¼ mile of) exactly one
disamenity have a mildly significant willingness to pay a premium to live even farther
away from it. Lastly, households living more than ¼ mile from a disamenity are willing
to pay a highly statistically significant premium to live even farther away from it. The
magnitude of this willingness to pay premium is consistent with other studies of the
effects of disamenities on nearby house prices.
It is no surprise that some people prefer to live far away from mild disamenities,
such as aging gasoline stations. It should also not be a surprise that some people are not
willing to pay a premium to live farther away from a mild disamenity. This study not
only identifies each of these groups (and a third group that we categorize as households
with moderate aversion), it also quantifies for each group the implicit price (marginal
willingness-to-pay in a competitive market) as a function of proximity to the nearest mild
disamenity. Furthermore, if one were to simply pool together all three submarkets
identified in this study, then the lack of aversion of those living closest to the mild
disamenities would be overwhelmed by the bulk of those individuals living farther away.
Painting an accurate spatial picture of how and why certain factors affect housing prices
is a fundamental goal of all researchers.
22
REFERENCES
Boyle, M.A. and Kiel, K.A. (2001). A survey of house price hedonic studies of the
impact of environmental externalities. Journal of Real Estate Research, 9(2),
117-134.
Brasington, D.M. and D. Hite (2005). Demand for environmental quality: a spatial
hedonic analysis. Regional Science and Urban Economics, 35(1), 57-82.
Cressie, N (1993). Statistics for Spatial Data. New York: John Wiley and Sons.
Deaton, B.J. and J.P. Hoen (2004). Hedonic analysis of hazardous waste sites in the
presence of other urban disameties. Environmental Science and Policy, 7, 499-408.
Ecker, M.D. (2003). Geostatistics: Past, Present and Future. In Encyclopedia of Life
Support Systems (EOLSS) . Developed under the Auspices of the UNESCO,
EOLSS Publishers, Oxford, UK. [available at: www.eolss.net].
Ecker, M.D. and De Oliveira, V. (2008). “Bayesian Spatial Modeling of Housing Prices
Subject to a Localized Externality”. Communications in Statistics. 37, 2066-2078
Ecker, M.D. and Isakson, H. (2005). A unified convex-concave model of urban land
values. Regional Science and Urban Economics, 35, 265-277.
23
Goodman, A.C., Thibodeau, T.G. (1998). Housing market segmentation. Journal of
Housing Economics, 7, 121-143.
Goodman, A.C., Thibodeau, T.G. (2003). Housing market segmentation and hedonic
prediction accuracy. Journal of Housing Economics, 12, 181-201.
Ihlandfeldt, K.R. and L.O. Taylor (2004). Externality effects of small-scale hazardous
waste sites: evidence from urban commercial property markets. Journal of
Environmental Economics and Management, 47, 117-139.
Isakson, H. and Ecker, M.D. (2008). An Analysis of the Impact of Swine CAFOs on the
Value of Nearby Houses. Agricultural Economics, 39, 1-8.
Isakson, H. and Ecker, M.D. (2001). An Analysis of the Influence of Location in the
Market for Undeveloped Urban Fringe Land. Land Economics, 77(1): 30-41.
Kiel, K.A. (1995). Measuring the impact of the discovery and cleaning of identified
hazardous waste sites on house values. Land Economics, 71(4), 428-35.
Kuethe, T.H. and Keeney, R. (2012). Environmental externalities and residential
property values: Externalized costs along the house price distribution. Land
Economics, 88(2), 241-250.
24
Jones, C., Leishman, C., and Watkins, C. (2003). Structural change in a local urban
housing market. Environmental and Planing, 35, 1315-1326.
Nelson, J.P. (2004). Meta-analysis of airport noise and hedonic property values:
problems and prospects. Journal of Transport Economics and Policy, ve(1), 1-27.
Ott, R.L. and M. Longnecker, (2010). An Introduction to Statistical Methods and Data
Analysis, 6th edition. Brooks/Cole.
Palm, R. (1978). Spatial segmentation of the urban housing market. Economic
Geography, 54(3), 210-221.
Rosen, S. (1974). Hedonic prices and implicit markets: product differentiation in pure
competition. Journal of Political Economy. 82, 34-55.
Simons, R.A. and Saginor, J.D. (2006). A meta – analysis of the environmental
contamination and positive amenities on residential real estate values. Journal of
Real Estate Research, 28 (1/2): 71-104.
Simons, R.A., Bowen, W.M, and Sementelli, A.J. (1997a). The effect of underground
storage tanks on residential property value in Cuyahoga County, Ohio. Journal of
Real Estate Research, 14 (1/2): 29-42.
25
Simons, R.A. and Sementelli, A.J. (1997b). Liquidity loss and delayed transactions with
leaking underground storage tanks. Appraisal Journal, July, 255-260.
Simons, R.A., Bowen, W.M. and Sementelli, A.J. (1999). The price and liquidity effects
of UST leaks from gas stations on adjacent contaminated property. Appraisal
Journal, April, 186-194.
Straszheim, M.R. (1975). An econometric analysis of the urban housing market.
National Bureau of Economic Research, New York.
Zabel, J. and D. Guignet (2010). A hedonic analysis of the impact of disamenity sites on
house prices in Frederick, Baltimore, and Baltimore City counties. National Center
for Environmental Economics Working Paper #10-01, U.S. Environmental
Protection Agency. (available at
yosemite.epa.gov/ee/epa/eed.nsf/WPNumber/2010-01/$File/2010-01.PDF)
26
Figure 1 : Cedar Falls sales (o) and disamenitys (■) locations.
Note: Plot is in State Plane Coordinates
27
TABLE 1: Summary Statistics of Sales Data
ALL SALES
N=2,078
LAH
N=365
MAH
N=209
HAH
N=1504
Variable
Mean
Std
Dev
Mean
Std
Dev
Mean
Std
Dev
Mean
Std
Dev
Sale Price
$144,995
71,415
$108,350
39,838
$97,717
27,606
$160,415
75,398
1332.8
1960.6
5.9
489.5
28.5
1.4
1172.1
1941.2
5.7
412.5
26.1
1.3
1073.5
1938.9
5.5
309.7
23.6
1.2
1407.8
1968.3
5.9
505.4
25.7
1.5
0.27
0.16
0.23
0.14
0.21
0.10
0.28
0.16
2.56
1.39
2.55
1.40
2.54
1.46
2.57
1.37
3.43
1.17
3.07
0.78
3.12
0.78
3.56
1.27
1.20
0.72
0.82
0.37
0.85
0.36
1.35
0.77
0.44
0.27
0.39
0.16
0.33
0.16
0.47
0.30
0.73
0.46
0.63
0.53
0.68
0.44
0.76
0.43
0.48
0.31
0.15
0.05
0.20
0.04
0.60
0.28
0.64
0.11
1.29
0.31
3.05
0.15
1.40
0.36
1
0.19
NA
0.39
0
0.09
NA
0.28
Structural
Variables
Living Area
Year Built
Number of
Rooms
Lot Size
(acres)
Date of Sale
(0=1/1/2000)
Distances
(miles)
Distance to
STP
Distance to
River
Distance to
School
Distance to
Highway
disamenity
Variables
Distance to
disamenity(
miles)
Count
Condition
28
Variable
Census
Block
Variables
Median
Household
Income
Number
Occupied
Units
Percent
Owner
Occupied
Median Year
Build
Median Rent
Difference
Median Year
Mean
Std
Dev
Mean
Std
Dev
Mean
Std
Dev
Mean
Std
Dev
40,817
12,015
37,878
10,197
38,927
8,661
41,793
12,665
695.3
423.3
482.0
319.2
492.2
267.2
775.3
435.7
59.3
22.4
55.8
22.0
62.5
26.0
59.7
21.8
1962.3
12.8
1953.8
11.5
1954.0
8.4
1965.6
12.1
474.3
-1.7
189.9
20.9
472.7
-12.6
173.3
21.8
455.6
-15.1
154.8
21.5
477.2
2.7
197.9
18.7
29
TABLE 2: Results
Proximity to disamenity
Count
Condition
Distance to STP
Distance to School
Distance to Highway
Distance to River
Ln Living Area
Date Sold
Number of Rooms
Ln Parcel Size
Yr Blt – Median Yr Blt
Median HH Income
Num Occ Units
Pct Owner Occ
Median Yr Blt
Median Rent
Nugget
Sill
Range
Group A
n=365
0.0194
(0.5477)
-0.0020
(0.8794)
0.0415
(0.3049)
0.1149
(0.0438)
0.0013
(0.9888)
0.1126
(0.0848)
0.0209
(0.7768)
0.4575
(0.0001)
0.0526
(0.0001)
0.0576
(0.0001)
0.0933
(0.0001)
0.0045
(0.0001)
0.000002
0.2809
0.0001
(0.0366)
-0.0609
(0.5994)
0.0022
(0.3266)
-0.00007
(0.4883)
0.0119
0.0355
0.0166
Group B
n=209
0.0966
(0.1178)
N/A
Group C
n=1504
0.0929
(0.0068)
N/A
0.0397
(0.4795)
0.1102
(0.0885)
-0.4095
(0.0011)
0.1009
(0.3022)
0.0243
(0.7894)
0.3274
(0.0001)
0.0544
(0.0001)
0.0782
(0.0001)
0.2266
(0.0001)
0.0037
(0.0001)
0.000007
0.0224
0.00002
(0.8752)
-0.0415
(0.7733)
-0.0024
(0.4788)
-0.00007
(0.6065)
0.0126
0.0213
0.0333
-0.0462
(0.3362)
0.0814
(0.0001)
-0.0320
(0.5171)
0.0738
(0.0631)
-0.0796
(0.0082)
0.3030
(0.0001)
0.0490
(0.0001)
0.0648
(0.0001)
0.1200
(0.0001)
0.0059
(0.0001)
0.000008
(0.0001)
0.00003
(0.5209)
-0.2284
(0.0113)
0.0069
(0.0002)
-0.00002
(0.8088)
0.0095
0.0454
0.0677
Note: Numbers in parentheses are p-values.
30
TABLE 3: Sensitivity analysis to the choice of the distance that defines the count variable
Distance = 0.2 mile
Sample Size
Proximity to
disamenity
Count
Condition
LAH
202
MAH
189
HAH
1687
0.0846
(0.1632)
NA
0.0878
(0.0005)
NA
0.1521
(0.0030)
-0.0564
(0.1403)
LAH
365
MAH
209
HAH
1504
0.0194
(0.5477)
-0.0020
(0.8794)
0.0415
(0.3049)
0.0966
(0.1178)
N/A
0.0929
(0.0068)
N/A
0.0397
(0.4795)
-0.0462
(0.3362)
LAH
550
MAH
187
HAH
1341
0.0357
(0.2762)
0.0009
(0.9336)
0.0400
(0.2790)
0.0350
(0.5329)
NA
0.1088
(0.0095)
NA
-0.1120
(0.2226)
-0.1404
(0.0348)
-0.0167
(0.7531)
-0.0332
(0.2267)
-0.0354
(0.7903)
Distance = 0.25 mile
Sample Size
Proximity to
disamenity
Count
Condition
Distance = 0.3 mile
Sample Size
Proximity to
disamenity
Count
Condition
31
1
Details of this ANCOVA are available from the authors upon request.
2
The authors thank the Black Hawk County Board of Supervisors for providing the housing sales data
used in this study. Of course, any opinions expressed in this study are strictly those of the authors.
3
Because the original data set contained all transactions of houses, it included a few very small structures
(as small as 100 square feet) that sold for very low prices (as low as $7,500) as well as a few very large
houses with large amounts of land that sold for as much as $2.5 million. These outliers were excluded
because they were not representative of the typical house in Cedar Falls, Iowa.
4
Equations 1 and 2 do not yet contain an error term. The structure of the error term is discussed later in
this section.
5
Ihlanfeldt and Taylor (2004) use a SAR for the error term in their model.
32