Content uploaded by Zack Hawley
Author content
All content in this area was uploaded by Zack Hawley on Nov 30, 2017
Content may be subject to copyright.
Where Does Racial Discrimination Occur? An Experimental Analysis across Neighborhood and
Housing Unit Characteristics
Published in Regional Science and Urban Economics, January 2014
Andrew Hansona Zackary Hawleyb
a Department of Economics, Marquette University, mailing address: P.O. Box 1881 Milwaukee, WI 5320.
e-mail: andrew.r.hanson@marquette.edu. phone: +1-414-288-5822.
b Corresponding author, Department of Economics, Texas Christian University, mailing address: P.O.
Box 298510 Fort Worth, TX 76129. e-mail: z.hawley@tcu.edu. phone: +1-817-257-6722.
This paper examines racial discrimination across several neighborhood and housing unit
characteristics including racial composition, rent, and distance from the urban core. We find that
African Americans face higher rates of discrimination than whites in a wide range of racially
mixed neighborhoods, in higher rent areas, closer to central cities, and in low vacancy areas.
These results are robust to various parameterizations of the local smoothing empirical
specification. The location of discrimination supports the current/future customer prejudice and
perceived preference hypotheses as a cause of discrimination in housing markets but not the
landlord taste-based hypothesis.
JEL: J15, C93
Keywords: Racial Discrimination, Rental Housing, Field Experiment
Acknowledgements: The authors thank Bentley Coffey, participants at the Discrimination and
Economic Outcomes session at the Western Economic Association International 85th annual
meeting, and participants at the Texas Christian University Economics Department seminar
series. We also thank Dan McMillen and two anonymous referees for helpful suggestions on
improving the original manuscript.
1
I. Introduction
The Department of Housing and Urban Development (HUD) annually receives between
700 and 1,000 complaints alleging discrimination on the basis of race or national origin,1 and
spends approximately $25 million on enforcing Fair Housing Laws. About 2 percent of all
complaints end with a charge of discrimination, and about 1 percent end in a referral to the
Department of Justice (DOJ) for enforcement. In fiscal year 2010, DOJ actions resulted in over
$1 million in penalties for Fair Housing Law violators, not including out of court settlements.2
Despite the dollars spent on enforcement and penalties landlords, real estate agents, and
mortgage brokers face for violating these laws, racial discrimination is still apparent in housing
markets across the U.S.3 The dollars spent on enforcement of Fair Housing Laws and the fact
that agents still violate them leads to at least two questions: What motivates agents to
discriminate? And, are there ways to reduce the cost of identifying violators?
This paper examines how discrimination changes with neighborhood and housing unit
characteristics, which offers insight into the causes of racial discrimination in housing markets
and identifies areas where agents are most likely to violate Fair Housing Laws. We measure
discrimination using data from a within-subjects field experiment, or matched-pair housing audit
of landlords advertising rental housing on-line. The experiment communicates with landlords via
e-mail, using names to signal race, and measures differential response rates between African
1 State-level Fair Housing Assistance Programs (FHAP) also handles discrimination complaints. These programs
receive an additional 3,000-4,000 complaints annually. More of these complaints (7-8 percent) end up in a charge of
discrimination, but fewer are referred to the Department of Justice for enforcement.
2 The DOJ recently settled with Bank of America for $335 million in response to allegations that it’s Countrywide
Financial subsidiary practiced discrimination against African American and Hispanic borrowers by charging them
higher fees and steering them into subprime mortgages (New York Times, “Country Wide Will Settle Bias
Lawsuit”, December 21, 2011)
3 Three decades of academic work confirms that racial discrimination exists in housing markets. See Yinger (1986)
for a classic example, and Hanson and Hawley (2011) for a recent example with citations to other newer studies.
2
Americans and whites. We match our measure of discrimination to data on housing unit and
2010 Census neighborhood characteristics.
We use a local polynomial smoothing estimation procedure to find where in the
distribution of characteristics discrimination is more likely to occur. This estimation method
provides a smooth prediction of discrimination outcomes across various characteristics of the
neighborhood or housing unit. Local polynomial smoothing is advantageous in this setting as it
does not assume a functional form for where discrimination may occur, it allows the data to
determine the relationship. We examine several different dimensions of neighborhood and
housing unit characteristics including, racial composition, distance to the city center, rental rates,
and vacancy rates.
The results suggest there are particular areas within cities that are more prone to higher
rates of discrimination against African American home-seekers. African Americans face higher
rates of discrimination than whites in a wide range of racially mixed neighborhoods, in higher
rent areas, closer in to the center city, and in low vacancy areas. The different locations of
increased discrimination provide evidence that supports several hypotheses on the cause of
discrimination. Generally, the results suggest current/future customer prejudice or perceived
preference as a cause of discrimination and do not support landlord taste as a cause. The results
also suggest that targeting enforcement efforts to particular areas of cities may help to reduce
costs.
The next section of the paper is a discussion of the previous research on housing market
discrimination, and describes how examining where discrimination occurs can inform the causes
of discrimination. Section 3 describes the experimental design and neighborhood data. Section 4
outlines the methodology for examining discrimination across neighborhood characteristics
3
using non-parametric estimation. Section 5 presents the results, and the final section of the paper
concludes.
II. Background and Causes of Discrimination
Discrimination against racial minorities in the housing market is well documented by
experimental studies. 4 These studies typically use data from pairs of actors that visit real estate
agent offices on separate occasions and record treatment to researchers. Matched pair, or within-
subjects experiments, often referred to as audits, began with Yinger (1986) and continue to be
used in the literature to study discrimination today. Other studies that use data from in-person,
matched pair experiments include Yinger (1991), Roychoudhury and Goodman (1992), Page
(1995), Ondrich et al. (1998), Ondrich et al. (1999), Ondrich et al. (2000), and Ondrich et al.
(2003), Choi et al. (2005), Zhao (2005), and Zhao et al. (2006). More recently, correspondence
studies that use names to identify race, and e-mail communication, have appeared in the
literature. These studies maintain the advantages of the experimental design, but avoid some of
the problems associated with using in-person actors (see Heckman (1998) and Heckman and
Siegelman (1993) for a detailed critique of in-person experiments). Studies that use e-mail based
communication to study discrimination in the housing market include Carpusor and Loges
(2006), Ahmed and Hammarstedt (2008), Ahmed et al. (2008), Ahmed and Hammarstedt (2009),
Ewens et al. (2012), Ahmed and Hammarstedt (2010), Bosch et al. (2010), Hanson and Hawley
(2011), and Hanson et al. (2011).
All of the literature that examines discrimination in the housing market finds that
discrimination occurs against minority clients to some degree. The literature examines a variety
4 Experimental studies of racial discrimination have become the standard in separately identifying racial
discrimination from other confounding factors. Other studies that use observational data on home prices (sales, self-
reported, or assessed) suffer from bias caused by correlation between unobserved factors at the unit, person, and
neighborhood level with outcomes of interest and race. While we see value in these studies for their ability to
examine important outcomes, we question the magnitudes associated with the level of discrimination they find.
4
of ways that housing agents may practice unequal treatment of minority clients, including
steering, providing information on or showing additional units, or asking for future visit
opportunities. The Hanson et al. (2011) study even examines the text of landlord replies to
inquires about rental housing to show that landlords use more positive language and are more
descriptive about units when replying to white customers.
We extend this literature along the lines of Yinger (1986) to examine the characteristics
of neighborhoods where discrimination happens, and use this to inform the likely cause of
discrimination. Yinger (1986) hypothesizes three causes of discrimination in housing markets,
and how they relate to the racial composition of neighborhoods.5
The first hypothesis Yinger describes is that agents (landlords) discriminate because of
their own tastes or prejudice in dealing with minority clients. Yinger points out that this cause of
discrimination may vary with landlord characteristics. For our purposes, because we do not have
data on landlord characteristics, this cause of discrimination should result in a constant
relationship between neighborhood characteristics and discrimination.
The second hypothesis is that landlords discriminate because they act on behalf of
prejudice customers (current or future tenants). This cause of discrimination is driven by
statistical discrimination as described by Phelps (1972). Statistical discrimination implies an
agent (landlord) uses past experiences to formulate the expected payoff for each potential home-
seeker and selects a lessee by profit maximization. This type of discrimination may also be
linked to landlords attempting to prevent the surrounding neighborhood from “tipping” beyond
an acceptable minority share for white residents, as described in Card et al. (2008). Once a
neighborhood reaches a tipping point share of minorities, it tends to become all minority
5 Yinger puts these in the context of discrimination by real estate agents, but the concepts apply equally well to
landlords of rental properties, which we examine here.
5
residents, as whites find the neighborhood unacceptable. Another form of perceived prejudice
that landlords may react to is from local public goods provision. Alesina et al. (1999) show that
areas with more ethnic fragmentation spend less on ‘productive’ public goods, such as education,
roads, sewers, and trash pickup. If landlords care about the level of local public goods provided
in their area, they may intentionally try to keep minorities from locating in their communities in
order to keep the level of local public goods high.
Lastly, the landlord may treat minority and white clients differently because of what they
perceive to be different preferences for neighborhoods or housing units among these groups. This
would again be considered statistical discrimination by the definition in Phelps (1972). Yinger
points out, and the Card et al. (2008) study confirms, that whites have a preference for
neighborhoods with a vast white majority of residents. African Americans, on the other hand,
have preference for integrated neighborhoods.
III. Experimental Design and Data
The key elements of studying where housing market discrimination occurs are an
unbiased measure of discrimination, and data on local neighborhood characteristics. Our
unbiased measure of discrimination comes from a field experiment conducted by Hanson and
Hawley (2011). For data on neighborhood characteristics, we match the address of housing units
in Hanson and Hawley (2011) data to 2010 census tract level neighborhood characteristics using
ArcGIS software.
The Hanson and Hawley (2011) field experiment is a within-subjects, or matched pair
audit correspondence study, of landlords across the United States.6 The data come from
landlords of rental properties advertised on Craigslist.org. Craigslist allows participants to place
and reply to on-line advertisements specific to local markets for jobs, housing, companionship,
6 For details of the original experiment see Hanson and Hawley (2011).
6
and other goods and services, although the experiment uses only listings pertaining to the rental
housing market. Each landlord is sent two emails, one email from an email address associated
with a white name and one email from an address with an African American name. The
experiment design relies on the names of the potential renters to signal race to the landlord.
The fictitious renters’ names come from Bertrand and Mullainathan (2004), who use
Massachusetts birth certificate data from 1974 to 1979 to identify names highly associated with
each race. The first names used to designate a white renter are Brad, Brendan, Brett, Matthew,
Neil, Geoffrey, Todd, Greg, and Jay. The first names used to identify an African American renter
are Darnell, Hakim, Jamal, Jermaine, Kareem, Leroy, Rasheed, Tremayne, and Tyrone. The last
names for white renters are Davis, Ryan, Murphy, O’Brien, Baker, McCarthy, Young, Jones, and
Wright. The last names used to represent African American renters are Johnson, Washington,
Robinson, Jackson, Hall, Parker, Williams, Jones, and Cooper.
The experiment consists of 4,728 audits, or 9,456 e-mail inquiries to advertisements for
rental housing from Atlanta, Boston, Chicago, Dallas, Washington, D.C., Houston, Los Angeles,
New York, Seattle, and San Francisco. All e-mail inquiries are sent between 9 am and 12 pm on
the day after a landlord posts an advertisement (always a Wednesday). Inquiries are sent from g-
mail account addresses in the following format: firstname.lastname.###@gmail.com, where ###
is a three-digit number unique to each name. The overall response rate to e-mail inquiries is
53.9%, with 63.6% of landlords responding to at least one e-mail inquiry from a pair of e-mails.
We use approximately 2,000 audits from the original experiment that include one African
American and one white name, the exact number depends on the neighborhood characteristic we
examine.7 We examine discrimination across neighborhood demographics by identifying the
7 The sample size also limited by ArcGIS software not recognizing some addresses information, and by some
landlords offering incomplete address information.
7
addresses of the advertised housing units, and matching them to census information. The
neighborhood information is determined at the census tract level. The addresses for each of the
rental properties are geo-coded and identified with a unique census tract. With the census tract id
information, we match the 2010 census data for each location and corresponding audit.
IV. Estimating Discrimination Rates across Neighborhood Characteristics
We use the within-subjects experimental design to determine if a given landlord treats
emails inquiries equally or discriminates by only replying to one e-mail. To measure
discrimination across characteristics, we calculate the discrimination rate against African
Americans and whites. The discrimination rate against African Americans (whites) is the ratio of
landlords that responded to the white (African American) home-seeker but not to the African
American (white) home-seeker divided by the number of audits.
##
AA
Rate
of repliesto only whiteauditor
Dof audits
=
We calculate a discrimination rate for each “bin” in our non-parametric specification, where the
size of the bin depends on the characteristic in question and the distribution of each variable.
This means that we are aggregating landlords that rent units in census tracts with similar
characteristics to create a discrimination rate. The original experiment measures discrimination
at the landlord level, using one unit per landlord so that we are not counting discriminating (or
non-discriminating) landlords more than once.
There are four different characteristics for which we evaluate the changes in the
discrimination rates. Table 1 provides a description and summary statistics for each of the
neighborhood characteristics. The most influential neighborhood characteristic for discrimination
studies is the population racial composition. We use the percentage of white residents within the
neighborhood as the characteristic of interest. The average percentage of white residents in the
8
sample is 61.8 percent. The second characteristic is how far the neighborhood is from the city
center. We calculate the distance measure as straight-line distance from the centroid of the
neighborhood, or census tract, to the tallest building for each respective city. The average
distance to the city center for units in our sample is 10.6 miles.
Most units advertised on craigslist include an advertised rent, and we use this
information, combined with data on city median rents to create a ratio of median rent for each
unit in our sample. 8 We use a ratio, rather than the nominal value of rent for each unit, to control
for the vast differences in nominal rent values across the areas studied. The rent ratio provides
insight to the level of discrimination in higher public good areas, as these qualities of the
neighborhood are usually reflected in higher rents. The average landlord advertised rental rate is
$1,488.04, and the average rental ratio is 1.08 or very close to the median rent for the city.
To understand how the supply of rental units in the neighborhood may affect
discrimination, we calculate the vacancy rate as the number of for rent vacant rental units divided
by the number of rental households.9 The average vacancy rate for neighborhoods in our sample
is 4.2 percent.
Table 1 also shows the gross discrimination rate against African Americans as 13.1
percent which is the percentage of audits where the landlord only responded to the white
potential client. The gross discrimination rate against whites, or the percentage of audits where
the landlord only responded to the African American potential client, is 7.6 percent. The
indicator variable for central city is unity for housing units lying inside an urbanized area.10
8 The median rent data come from HUD. The ACS provides a self-reported gross rent available at the county level
for 2010. Using either data produces the same result.
9 The vacancy data come from Census. The ACS provides the number of for rent properties and the number of
vacant for rent properties. The vacancy rate is a simple ratio of these two measures.
10 An urbanized area is defined by the Census Bureau as an area with 50,000 or more people and a densely settled
core of census tracts or blocks along with adjacent territory with lower population density.
9
Table 2 summarizes how each of the neighborhood characteristics we examine relates to
the Yinger hypotheses on the causes of discrimination. The simplest hypothesis is the landlord
taste-based discrimination. The taste hypothesis states the level of discrimination should not vary
with the characteristics of the neighborhood but only due to the characteristic of the home-
seeker. The resulting expectation is constant discrimination against African American home-
seekers and no discrimination against white home-seekers.
The second hypothesis tested is the current or future customer’s prejudice. Under this
hypothesis, landlords should discriminate more against one race when the surrounding
population is of a different race. If the belief of the landlord is that current or future tenants who
live in more homogenous neighborhoods want to continue to live in such neighborhoods, then
landlords may discriminate more when the race of the home-seeker does not match the current
demographic. An additional hypothesis that provides a very similar story is the tipping
phenomenon. Tipping may appear as a desire for integration at very high percentage majority
population followed by a rapid increase in discrimination against minority home-seekers as the
neighborhood approaches the tipping point. This behavior is consistent with the current
population trying to ‘save’ the neighborhood. After the tipping point is achieved, the hypothesis
predicts a decline in discrimination rates against minority clients as the majority leaves and the
neighborhood becomes fully integrated.
The final hypothesis is the perceived preference or budget of a perspective client. If
perceived customer preferences are driving landlord discrimination, we should find
discrimination against whites in majority-minority neighborhoods, and discrimination against
African Americans only in vast majority white neighborhoods. Landlords in neighborhoods with
very high percentages of white residents may not believe that an African American would truly
10
want to live in their unit. The email inquiry may not be seen as genuine interest and thus the
landlord does not respond. The perception is that each race may prefer a small amount of
integration, but in neighborhoods with virtually no integration landlords may not see an inquiry
as genuine interest.
The primary method for analyzing the audit data is a local polynomial smoothing
technique. This estimation strategy combines the simplicity of a parametric method, Weighted
Least Squares (WLS), with the flexibility of a non-linear regression technique. The estimation
fits multiple WLS regressions to ‘localized’ bins of the data building a more flexible result that is
determined by the variation within the data, point-by-point.
2
1
min ( )
n
ii i
i
wy x
β
=
−
∑
The data within the bin that surrounds the point of interest are weighted with increased
importance given the ‘closer’ data. The same estimation is applied for all data points in the
distribution, shifting ‘localized’ bins accordingly. The resulting smooth function, f, is a
compilation of each point’s WLS predicted value.
()
i ii
y fx
ε
= +
The local polynomial smoothing method, introduced by Cleveland (1979) and refined by
Cleveland and Devlin (1988), places the resulting fitted values into a plot or functional form that
is not possible to estimate with parametric techniques. The biggest benefit to using smoothing as
opposed to standard linear regression is that it does not require a global functional form
assumption of the underlying distribution of the data, only for each bin of the data. An additional
benefit of the smoothing is that many of the parameters such as the degree of polynomial,
bandwidth (size of the ‘localized’ bin), or weights, are flexible in the estimation. We choose
11
local polynomial over other techniques, such as spline estimation, because it allows the data to
determine where (or if any) structural breaks occur.
The local polynomial smoothing technique also allows for confidence interval
calculation. We use the confidence intervals to determine significant differences in the
discrimination rates against either race along the distribution of the characteristic. These
differences inform the plausibility of the underlying causes.
Using local polynomial smoothing requires us to make three choices- the degree of
polynomial estimate within bins, the bandwidth around each point, and the weights assigned to
other points around each data point. We implement local linear smoothing for all of the results
shown. Cleveland (1979) discusses choosing the degree of polynomial and finds linear
smoothing is almost always an adequate balance between flexibility and computational ease.
The choice of higher degree polynomials does not change the results and only makes the
intuition of the method more cumbersome. We must also choose a bandwidth around each
observation to define how much of the data to use in each regression, we use the “rule of thumb”
bandwidth.11 The appendix examines sensitivity of our results to this choice. The weights are
kernel distributed.12 DiNardo and Tobias (2001) point out that, in general, non-parametric
estimates are not sensitive to the choice of how nearby observations are weighted.
The local polynomial smoothing method is not a new empirical method in urban
economics. Meese and Wallace (1991) use the nonparametric technique to evaluate hedonic
price models and residential housing price indices. McMillen (1996) applies the locally weighted
estimates to land value data from Chicago over the past 150 years. The flexibility of the model
11 The lpoly statistical package in STATA 12 provides a “rule of thumb” method for bandwidth choice. We follow
this choice as a starting point and provide robustness checks for this choice.
12 We use the Epanechnikov kernel which is standard and applies increased weights to the observations as the
distance from the point of interest falls to zero.
12
helps provide insights on the polycentric nature of Chicago. More recently, McMillen and
Redfearn (2010) show how hypothesis testing can be done with the local polynomial smoothing
method. We are advancing this line of literature by using this flexible empirical method to
analyze the location of discrimination across urban areas.
V. Results
Percent of White Residents
Figure 1 shows the discrimination rates across the racial composition of the neighborhood
surrounding rental properties in our data. The discrimination rate against African American
home-seekers is significantly higher than against white home-seekers across the full range of
neighborhoods. Discrimination against white home-seekers does not fluctuate across
neighborhood composition, which suggests racial composition is not important in determining
the level of discrimination against whites.
If we start at a location where the neighborhood comprises 100 percent white residents
and slowly add minority households, we see the discrimination rate against African American
home-seekers rises steadily. The current residents may be fearful of their neighborhood tipping,
as the discrimination rate reaches a plateau at around 80 percent white residents. This result is
consistent with Card, Mas, and Rothstein (2008) as most of the neighborhoods showed tipping
between 10 and 20 percent minority. There is no drop off in discrimination rates between 80
percent and 30 percent white residents which is not consistent with the expectation of tipping
concerns. African American home-seekers face similar discrimination rates for a large range of
mixed neighborhoods. Below 30 percent white residents, the discrimination rate falls, implying
residents in neighborhoods with a large percentage of minority residents do not discriminate
against African Americans as much as the landlords in mixed neighborhoods. This figure is not
13
consistent with current or future customer prejudice, since the expectation under this hypothesis
is higher levels of discrimination in largely white or African American neighborhoods. It is also
not consistent with landlord taste-based discrimination as the rate of discrimination does
significantly change with the racial composition.
Figures 1.1-1.4 show examples of different city specific cases for discrimination by the
percent of white residents in the surrounding neighborhood, as discrimination may differ across
the distribution in each city.13 In Atlanta, the discrimination rate against white home-seekers
falls as the percentage of white residents becomes larger. This is consistent with customer
prejudice against living with white residents. Additionally, there is only a small significant
difference in discrimination rates between African Americans and whites for Atlanta, and the
location of this difference is consistent with a tipping concern story. As the minority population
grows, so does the discrimination rate, but the rate begins to decline at around 75 percent white
residents.
Figure 1.2 describes the Boston sub-case. There seems to be a constant discrimination
rate difference against African Americans in Boston. Over a wide range of neighborhood
compositions, below 85 percent white, the predicted difference is stable.14 Above 85 or 90
percent, the discrimination rate against African Americans does fall slightly this is again
suggestive of a neighborhood tipping concern story. The Chicago sub-case tells a different
story. Figure 1.3 shows a desire for integration in neighborhoods that are highly white
concentrated, as the rate of discrimination is not statistically different between white or African
American home-seekers. However, as the minority share continues to climb, the difference
13 We picked examples that were representative. The Atlanta sub-case is similar to Dallas and Houston. The Boston
sub-case resembles Seattle and New York. The Chicago sub-case is comparable to San Francisco and Los Angeles.
14 Even though the confidence intervals begin to expand at around 20 percent white residents, the predicted values
are still relatively flat. This is due to the relatively small sample size in those neighborhood types for Boston.
14
becomes significant and peaks around 50 percent. This pattern may be consistent with perceived
preference discrimination. Landlords may believe African Americans want to live in either
mainly white neighborhoods or largely minority neighborhoods with less desire to be in the
middle.
Figure 1.4 presents the District of Columbia sub-case. The highest rates of
discrimination appear in the white concentrated neighborhoods and falls off steadily with the
percentage of white residents. This city’s sub-case is most consistent with the customer
prejudice hypothesis. Landlords may believe that current neighbors or future clients will not
appreciate integration in their neighborhoods and these preferences can produce a picture like the
D.C. sub-case.
Distance to City Center
With the stark difference in housing structures and social interaction within urban and
suburban neighborhoods, examining discrimination rates by distance from center of the city may
provide insight into how these factors influence discrimination. Figure 2 shows the
discrimination rate against African American home-seekers is always higher than whites across
the entire distance distribution, although the rates do pinch together slightly at around 20 to 30
miles from the city center. Again, since the individual cities vary greatly on the types of
neighborhoods at various distances, we look at the individual sub-cases to reveal more about
potential causes.15
Figure 2.1 shows the Atlanta sub-case. This figure shows a preference for integration
very close to the city center as well as in the suburbs between 15 and 30 miles away.
Interestingly, there are two ranges of distances for which the discrimination rates are
15 These locations are representative of different patterns for the cities in our data. Atlanta is similar to Chicago,
Seattle, and San Francisco, where Boston is aligned with District of Columbia, Los Angeles, New York, Dallas and
Houston.
15
significantly different. In the urban setting from about 2 miles to 15 miles out and in the outer
suburb ring (between 30 and 40 miles), where the discrimination rate against African American
home-seekers is statistically larger than against white home-seekers. This type of pattern may
suggest customer prejudice as the landlords are likely aware of the clientele of those
neighborhoods.16 The Chicago sub-case is similar as shown in Figure 2.3.
Figure 2.2 describes the discrimination rate difference across distance for the Boston sub-
case. In this case there does not seem to be a preference for integration near the city center but
as the distance is increased the difference in discrimination rate goes away. Figure 2.4 shows the
District of Columbia sub-case which is similar to Boston in that there is a statistically significant
difference very close to the city center. This difference disappears as the distance increases, but
unlike Boston, in the further suburbs the difference reemerges.
Advertised Rental Rate
Figure 3 presents the discrimination rates over the advertised rent relative to the median
rent. At low rents relative to the median, there is no statistical difference in differential treatment
against white or African American home-seekers. There is a large increase in discrimination
against white home-seekers in the lower range. It may be that landlords do not believe the
typical white customer will want to live in a low-rent area, and discriminates against them due to
perceived preference.
When the rent ratio reaches around 0.75, or at rents above 75 percent of the median,
landlords start to discriminate more against African Americans than their white counterparts.
This difference in discrimination rates continues to rise and peaks at around 1.5, or at rents that
16 In Atlanta, distance interacts with racial composition of neighborhoods. A basic description of the city is that there
is an inner ring of African American majority neighborhoods surrounded by the heavily mixed urban area. Further
out is a ring of more heavily African American populated suburbs encompassed by a ring of white populated
neighborhoods.
16
are 150% of the median. This type of result is suggestive of statistical discrimination, if the
landlord has a perception of ability to pay from the racial group. Interestingly, as the rent ratio
continues to increase the discrimination rates become closer and eventually not statistically
different from each other. This suggests in very expensive areas landlords are not favoring
African Americans or whites disproportionately, at least through initial email contact. This does
not occur until very high levels of rent, where even inquiring about a rental unit may send a
strong signal about ability to pay.
In the city-specific cases, Atlanta (Figure 3.1) and Boston (3.2) are similar to the national
average. The Chicago (Figure 3.3) and D.C. (Figure 3.4) sub-cases, there is statistically
significant discrimination at the lower part of the rent ratio distribution that does not exist in the
national sample. Although it is not entirely clear what is driving these differences, part of the
reason may be the presence of other minority populations in Chicago and D.C. in low rent areas.
Neighborhood Vacancy Rate
Figure 4 examines discrimination across neighborhood vacancy rates. After a slow initial
decline, the discrimination rate against African Americans rises sharply in neighborhoods
starting with about 12 percent vacant units. This type of result is consistent with personal
prejudice rather than other statistical discrimination hypotheses, as a profit maximizing landlord
should rent to any client where marginal benefit exceeds marginal cost, and a client paying rent
should be superior to leaving a unit empty in nearly all cases.
The Chicago sub-case (Figure 4.3) shows a similar pattern to the national average, but all
of the other city specific figures are substantially different. In Atlanta (Figure 4.1) and the
District of Columbia (Figure 4.4), we find statistically meaningful discrimination against African
Americans at low vacancy levels, and no discrimination at higher levels of vacancy- consistent
17
with a statistical explanation, and not with landlord tastes, contrary to the national sample.
Boston (Figure 4.2) has the odd pattern of no discrimination at either high or low vacancy rates,
but fairly constant discrimination in mid-level vacancy areas, although the vacancy distribution
as a whole is shifted to the left.
Multivariate Parametric Results
We use a multivariate nonlinear parametric model using simple polynomial to check that
the nonlinear relationships shown in the bivariate nonparametric estimation stay consistent. This
robustness check increases the confidence of our estimation strategy and results. Table 3 shows
the nonlinear parametric results. We estimate a linear probability model with additional
neighborhood and unit characteristic control variables such as number of bedrooms, number of
bathrooms, and the percentage of college graduates in the census tract. The dependent variable is
a binary response variable with unity implying a response was received. All of the previous unit
and neighborhood characteristics are used as independent variables both together (column 1) and
separately (columns 3-6).
The parametric results show strong support for the local polynomial smoothing
estimation strategy. The direction and magnitude of the results buttress the non-parametric
figures. For example, the percentage of white residents in the neighborhood has a positive and
strongly significant impact on the probability of a response controlling for race and the
interaction of race and white residence percentage. This shows the discrimination rate against
African Americans is much higher as the percentage of white residents in the neighborhood
increases (from zero to one). The square of the percent of white residents shows the effect of
additional white residents falls for neighborhoods that are more than 50 percent white. This
18
result mirrors the finding in Figure 1, while controlling for other covariates that may affect
landlord response.
The remaining parametric results also follow the nonparametric findings; however, some
are not quite as statistically precise (although most remain statistically meaningful at the ten
percent level). Higher rents lead to higher rates of discrimination against African Americans but
at a decreasing rate around the critical value of 1.5 times the median rent. Likewise, very low or
high vacancy rates lead to increases in the discrimination rate against African Americans around
the critical value of 15 percent. These patterns are similar to the nonparametric figures. The
distance measure does not show statistical significance in any specification but the sign and
magnitude are congruent with the nonparametric results.
VII. Conclusion
We use a non-parametric estimation strategy to examine racial discrimination across
neighborhood and housing unit characteristics. We examine how discrimination rates fluctuate
within the distribution of racial composition, distance to city center, affordability, and vacancy
rates. African American home-seekers are discriminated against at a higher rate than their whites
in neighborhoods that are racially mixed. They also face larger discrimination rates in areas with
rents near or above the median rent for the city, as well as in areas with low or very high vacancy
rates. Higher discrimination rates against African Americans are also observed in neighborhoods
that are located very close to the center of the city or in the first ring of suburbs.
While it is extremely difficult to determine the exact cause of discrimination, the results
do suggest that landlord taste is not the central cause of discrimination rates against home-
seekers. In general, the perceived preference or current/future customer prejudice cause is much
more plausible. Both of these hypotheses of discrimination are in line with statistical
19
discrimination. While still illegal, the results suggest it is the landlords’ intent to maximize
profits by treating the marginal home-seeker like a perceived typical group member.
This study provides potential hot-spots for discrimination such as neighborhoods in the
‘tipping’ range, with low or very high vacancy rates, and with rents that are near or above the
median rent for the city. This knowledge of where discrimination occurs is helpful for targeting
future enforcement activities or for informing the choice of locations for public sessions on how
to spot and report discriminatory behavior. We should note, however, that it is possible that a
different sample of cities would produce different results, even if those cities have similar
demographic and economic characteristics as our sample. It is possible that other factors about
metropolitan areas such as segregation, commuting patterns, immigration, or local public
services would play an important role that we do not capture in our data.
Our work does suggest that studies of discrimination should consider taking into account
the types of neighborhoods in the sample selection process, as oversampling from certain
neighborhoods may be consistent with finding above or below average discrimination rates. The
literature and enforcement of Fair Housing Laws should also consider how access to specific
types of neighborhoods may have a differential impact on housing outcomes than a general level
of discrimination.
References
Ahmed, Ali M., Lina Andersson, and Mats Hammarstedt (2010) “Can discrimination in the
housing market be reduced by increasing the information about the applicants?” Land
Economics, 86 (1), 79-90.
Ahmed, Ali M. and Mats Hammarstedt (2009) “Detecting discrimination against homosexuals:
evidence from a field experiment on the internet,” Economica, 76, 588-597.
Ahmed, Ali M., Lina Andersson, and Mats Hammarstedt (2008) “Are lesbians discriminated
against in the rental housing market? Evidence from a correspondence testing
experiment,” Journal of Housing Economics, 17 (3), 234-238.
Ahmed, Ali M. and Mats Hammarstedt (2008) “Discrimination in the rental housing market: a
field experiment on the internet,” Journal of Urban Economics, 64 (2), 362-372.
Alesina, Alberto, Reza Baqir, and William Easterly (1999) “Public goods and ethnic divisions,”
Quarterly Journal of Economics, 114 (4), 1243-1284.
Bosch, Mariano, Angeles Carnero, and Lidia Farre (2010) “Information and discrimination in the
rental housing market: evidence from a field experiment,” Regional Science and Urban
Economics, 40 (1), 11-19.
Card, David, Alexandre Mas, and Jesse Rothstein (2008) “Tipping and the dynamics of
Segregation,” Quarterly Journal of Economics, 123 (1), 177-218.
Carpusor, Adrian G., and William E. Loges (2006) “Rental discrimination and ethnicity in
names,” Journal of Applied Social Psychology, 36 (4), 934-952.
Choi, Seok Joon, Jan Ondrich, and John Yinger (2005) “Do rental agents discriminate against
minority customers? Evidence from the 2000 housing discrimination study,” Journal of
Housing Economics, 14 (1), 1-26.
Cleveland, William S. (1979) "Robust locally weighted regression and smoothing
scatterplots," Journal of the American Statistical Association, 74, 829-836.
Cleveland, William S. and Susan J. Devlin (1988) "Locally weighted regression: an approach to
regression analysis by local fitting," Journal of the American Statistical Association, 83,
596-610.
DiNardo, John, and Justin L. Tobias (2001) "Nonparametric Density and Regression
Estimation," Journal of Economic Perspectives, 15(4), 11–28.
Ewans, Michael, Bryan Tomlin, and Liang Choon Wang (2012) “Statistical Discrimination or
Prejudice? A Large Sample Field Experiment,” Review of Economics and Statistics,
Forthcoming.
Hanson, Andrew and Zackary Hawley (2011) “Do landlords discriminate in the rental housing
market? Evidence from an internet field experiment in US cities,” Journal of Urban
Economics, 70 (2-3), 99-114.
Hanson, Andrew, Zackary Hawley, and Aryn Taylor (2011) “Subtle discrimination in the rental
housing market: evidence from e-mail correspondence with landlords,” Journal of
Housing Economics, 20, 276-284.
Heckman, James J. (1998) “Detecting discrimination,” Journal of Economic Perspectives, 12 (2),
101-116.
Heckman, James J. and Peter Siegelman, “The urban institute audit studies: their methods and
findings,” Clear and Convincing Evidence: Measurement of Discrimination in America,
M. Fix and R. Struyk, (eds.) Chapter 5, 187-258.
McMillen, Daniel P. (1996) “One hundred fifty years of land values in Chicago: a nonparametric
approach,” Journal of Urban Economics, 40 (1), 100-124.
McMillen, Daniel P. and Christian L. Redfearn (2010) “Estimation and hypothesis testing for
nonparametric hedonic house price functions,” Journal of Regional Science, 50 (3), 712-
733.
Meese, Richard and Nancy Wallace (1991) “Nonparametric estimation of dynamic hedonic price
models and the construction of residential housing price indices,” Real Estate Economics,
19 (3), 308-332.
Ondrich, J., Alex Stricker, and John Yinger (1998) “Do real estate brokers choose to
discriminate? Evidence from the 1989 housing discrimination study,” Southern Economic
Journal, 64 (4), 880-901.
Ondrich, Jan, Alex Stricker, and John Yinger (1999) “Do landlords discriminate? The incidence
and causes of racial discrimination in rental housing markets,” Journal of Housing
Economics, 8 (3), 185-204.
Ondrich, Jan, Steven L. Ross, and John Yinger (2000) “How common is housing discrimination?
Improving on traditional measures,” Journal of Urban Economics, 47 (3), 470-500.
Ondrich, Jan, Steven L. Ross, and John Yinger (2003) “Now you see it, now you don’t: why do
real estate agents withhold houses from black customers?” Review of Economics and
Statistics, 85 (4), 854-873.
Page, Marianne (1995) “Racial and ethnic discrimination in urban housing markets: Evidence
from a recent audit study,” Journal of Urban Economics, 38 (2), 183-206.
Phelps, Edmund (1972) “The statistical theory of racism and sexism,” American Economic
Review, 62 (4), 659-661.
Roychoudhury, Canopy and Allen C. Goodman (1992) “An ordered probit model for estimating
racial discrimination through fair housing audits,” Journal of Housing Economics, 2 (4),
358-373.
Yinger, John (1986) “Measuring racial discrimination with fair housing audits: caught in the
act,” American Economic Review, 76 (5), 881-893.
Yinger, John (1991) “Acts of discrimination: evidence from the 1989 housing discrimination
study,” Journal of Housing Economics, 1 (4), 318-346.
Zhao, Bo (2005) “Does the number of houses a broker shows depend on a homeseeker’s race?”
Journal of Urban Economics, 57 (1), 128-147.
Zhao, Bo, Jan Ondrich, and John Yinger (2006) “Why do real estate brokers continue to
discriminate? Evidence from the 2000 housing discrimination study,” Journal of Urban
Economics, 59 (3), 394-419.
Appendix: Robustness of Estimates to Bandwidth Choice
The choice of bandwidth may be the most commonly criticized parameter selection for
local smoothing techniques. We test how sensitive our primary results are to changing the rule of
thumb bandwidth in the local polynomial regression. We examine how choosing a smaller (one-
half the size) and larger (twice the size) bandwidth changes our results.
Figure A1 shows the discrimination rate across the percent of white residents distribution
using half and double the rule-of-thumb bandwidth, respectively. As Figure A1 shows, the
results are quite similar to the results for the standard rule of thumb bandwidth. Mixed
neighborhoods still retain the higher discrimination rate differences with lower discrimination
rates against African Americans in the largely majority neighborhood and the largely minority
neighborhoods. Similarly, Figure A2 illustrates the discrimination rates with half and double the
rule-of-thumb bandwidth, against African American and white home-seekers along the distance
to the city center distribution. Again the choice of bandwidth does not seem to affect the pattern
that we observe.
Figure A3 and A4 show the bandwidth robustness checks across the rent ratio and
vacancy rate distribution. We find that the bandwidth choice does not change the distribution of
discrimination substantially across any of these attributes. We also check the robustness of the
results with the degree of polynomial and the kernel function choice for the local smoothing
estimation. The choice of these parameters does not significantly alter the main conclusions
described above.
Variable Description Obs. Mean St. Dev. Min Max
Percentage of White Residents The percentage of white residents. 2029 0.618 0.236 0.005 1
Distance to City Center
The straight line distance to the city
center (proxied by tallest building) in
miles.
2207 10.605 10.881 0.136 113.495
Rental Rate
The advertised rental rate set by the
landlord.
2009 1488.04 769.37 250 8750
Rental Rate Ratio The ratio of the advertised rental rate
divided by the median rent for the city. 2009 1.080 0.437 0.169 4.714
Vacancy Rate The percentage of rental properties
that are for rent and vacant. 2026 0.042 0.037 00.591
Gross Discrimination Rate Against
Whites
The percentage of landlords who only
responded to the African American
client in the audit pair.
2207 7.612 26.525 0100
Gross Discrimination Rate Against
African Americans
The percentage of landlords who only
responded to the White client in the
audit pair.
2207 13.14 33.791 0100
Central City =1 if rental location is within the
urbanized area 2207 0.977 0.147 0 1
Table 1: Summary Statistics
Landlord Taste Current or Future Customer Prevent Tipping Perceived Preference/Budget
Discrimination against Afrian
Americans everywhere
Decreasing discrimination against African
Americans as white population declines
Discriminate against African
Americans in neighborhoods
rises sharply in tipping range,
falls on either side of this range
Discriminate against African
Americans in neighborhoods that are
vast majority white
No discrimination against
whites
Increasing discrimination against whites,
as white population declines
Discriminate against whites in
neighborhoods rises sharply in
tipping range, falls on either
side of this range
Increasing discrimination against
whites, as white population declines
Discrimination against Afrian
Americans everywhere
Discriminate against African Americans
in neighborhoods further from the city
center
Not Relevant
Discriminate against African
Americans in neighborhoods further
from the city center
No discrimination against
whites
Discriminate against whites in
neighborhoods closer to city center
Not Relevant
Discriminate against whites in
neighborhoods closer to city center
Discrimination against Afrian
Americans everywhere
Discriminate against African Americans
more as rental rate increases
Not Relevant
Discriminate against African
Americans more as rental rate
increases
No discrimination against
whites
No discrimination against whites Not Relevant
No discrimination against whites, or
discrimination at lowest rents
Discrimination against Afrian
Americans everywhere
Reduced discrimination against African
Americans as vacancy rate rises
Not Relevant
Discrimination against African
Americans as vacancy rate falls
No discrimination against
whites
Less discrimination against whites as
vacancy rates rise
Not Relevant
More discrimination against whites
as vacancy rates rise
Table 2: Cause of Discrimination Hypotheses by Neighborhood Characteristic
Racial Composition
Distance to City Center
Rental Rate
Vacancy Rate
(1) (2) (3) (4) (5) (6)
Combined
Effect
African American
Only Effect
Percentage of White
Residents Effect
Distance Effect
Median Rent
Ratio Effect
Vacancy Rate
Effect
African American (AA) 0.0106 -0.0542*** 0.0131 -0.0530** -0.0126 -0.0838***
[0.10] [-3.49] [0.18] [-2.03] [-0.19] [-2.92]
Percentage of White Residents in the Neighborhood 0.7137*** 0.7517***
[3.52] [3.71]
Square of Percentage of White Residents -0.5993*** -0.5783***
[-3.23] [-3.12]
AA * Percentage of White Residents -0.2842 -0.2866
[-0.99] [-1.01]
AA * Square of Percentage of White Residents 0.2742 0.2511
[1.05] [0.96]
Distance to City Center 0.0008 0.0017
[0.36] [0.82]
Square of Distance to C ity Center 0.0000 0.0000
[0.68] [0.46]
AA * Distance to City C enter 0.0000 -0.0001
[0.02] [-0.02]
AA * Square of Distance to City Center -0.0000 -0.0000
[-0.11] [-0.05]
Percentage of Median Rent in the City 0.1518** 0.1240*
[2.08] [1.80]
Square of Percentage of Median Rent in the City -0.0523** -0.0475**
[-2.43] [-2.25]
AA * Percentage of Median Rent in the City -0.0525 -0.0558
[-0.52] [-0.58]
AA * Square of Percentage of Median Rent in the City 0.0122 0.0135
[0.40] [0.45]
Vacancy Rate in the Neighborhood -1.1315** -1.3643***
[-2.16] [-2.73]
Square of Vacancy Rate in the Neighborhood 3.6450** 4.0341***
[2.35] [2.65]
AA * Vacancy Rate in the Neighborhood 0.8717 0.8206
[1.18] [1.18]
AA * Square of Vacancy Rate in the Neighborhood -1.4698 -1.3510
[-0.67] [-0.63]
Bedrooms -0.0133 -0.0117 -0.0109 -0.0133 -0.0119 -0.0138
[-1.42] [-1.25] [-1.17] [-1.43] [-1.27] [-1.48]
Bathrooms -0.0022 -0.0025 -0.0013 -0.0029 -0.0013 -0.0034
[-0.22] [-0.25] [-0.13] [-0.29] [-0.13] [-0.34]
Single Family 0.0301 0.0421* 0.0424* 0.0286 0.0423* 0.0357
[1.28] [1.83] [1.84] [1.22] [1.83] [1.54]
Apartment -0.0605*** -0.0647*** -0.0625*** -0.0624*** -0.0635*** -0.0663***
[-3.29] [-3.53] [-3.42] [-3.40] [-3.47] [-3.62]
Percentage College Educated in Neighborhood -0.0651 -0.0223 -0.1088* 0.0085 -0.0125 -0.0343
[-1.01] [-0.43] [-1.85] [0.16] [-0.22] [-0.65]
Constant 0.4039*** 0.6375*** 0.4465*** 0.6124*** 0.5644*** 0.6927***
[4.94] [22.81] [7.93] [19.02] [10.66] [19.99]
Observations 4,042 4,060 4,058 4,050 4,060 4,052
R-squared 0.021 0.009 0.015 0.012 0.011 0.012
F Statistic 3.955 6.160 6.258 4.769 4.709 4.864
Table 3: Response Rate by Unit and Neighborhood Characteristics
Notes: The dependent variable is a binary respons e, where unity implies an email response received. Neighborhood is defined as a ce nsus tract. Distance is the straight line distance in num ber of miles to the
tallest building in the city. Median rent is at the city level by number of bedrooms; percentage of median rent is the s imple ratio of reported rent to the median rent. P-values reported in brackets. ***, **,
and * denote significance at the 1%, 5%, and 10% lev el respectively.
