Racialized Discourse in Seattle Rental Ad Texts

Abstract

Racial discrimination has been a central driver of residential segregation for many decades, in the Seattle area as well as in the United States as a whole. In addition to redlining and restrictive housing covenants, housing advertisements included explicit racial language until 1968. Since then, housing patterns have remained racialized, despite overt forms of racial language and discrimination becoming less prevalent. In this paper, we use Structural Topic Models (STM) and qualitative analysis to investigate how contemporary rental listings from the Seattle-Tacoma Craigslist page differ in their description based on neighborhood racial composition. Results show that listings from White neighborhoods emphasize trust and connections to neighborhood history and culture, while listings from non-White neighborhoods offer more incentives and focus on transportation and development features, sundering these units from their surroundings. Without explicitly mentioning race, these listings display racialized neighborhood discourse that might impact neighborhood decision-making in ways that contribute to the perpetuation of housing segregation.

While scholars often engage with the housing market through large datasets, most people do so through housing searches, often reading real estate advertisements. These short descriptions of housing units tell home seekers not only about units and their surroundings, but also potentially provide insight into what the landlord or property manager believes is most salient to their potential tenants. Though listings almost certainly include bedroom size, rent, and square footage, they also describe the surrounding neighborhood. The history of housing in the United States is entwined with racial exclusion and motivates our investigation of the degree to which legacies of segregation are present in rental advertisements themselves.

In the era of government-supported racist exclusion through redlining, housing covenants and violence, historical advertisements were often explicitly discriminatory. For instance, an advertisement from The Seattle Times on August 9, 1934 offered four rooms for eighteen dollars ($341 in 2019 dollars) and specified “whites only.” Over time, language became more coded, and by the late 1940s, racial exclusion was described as a “restriction.” One advertisement posted in The Seattle Times on December 23, 1947 read “Property values will continue to increase in this new restricted district on Magnolia Bluff.” This district, now a wealthy Seattle neighborhood, was made restricted through the use of exclusionary housing covenants, discriminatory agreements used across the United States that excluded people of certain races from owning or living in a home (Rothstein 2017). In turn, these covenants prevented non-White households in Seattle and other metropolitan areas from building wealth through homeownership and also limited their access to certain public schools (Denton 1995, Fernald 2019).ⁱ

By 1968, rents had increased so that a three-bedroom home in the South End of Seattle was advertised in the same paper on February 18 of that year for two hundred and fifty dollars a month (over $1,800 in 2019 dollars). The listing was short, and shown in figure 1:

Figure 1.

Rental Advertisement from February 18, 1968. The Seattle Times, 1968, reprinted with permission.

In contrast to prior decades, landlords rarely included such strongly racial language in the late 60s.ⁱⁱ For one thing, most people likely already knew which neighborhoods were friendly for which people. The South End, though, bordered the Central District, where most Black households in Seattle lived. Perhaps the person who posted this listing wanted to make sure that prospective tenants knew who was welcome and may have worried that the neighborhood name might discourage White tenants. At the same time, the writer of this listing might have worried that without this kind of coded exclusion, they would need to turn away prospective Black tenants trying to rent the unit.

The Fair Housing Act of 1968 made overtly discriminatory advertisements illegal, in addition to ending the enforcement of racially restrictive covenants and banning racial discrimination in housing and credit markets. While some listings continued to emphasize their exclusionary covenants into 1969, by 1970 Seattle’s rental advertisements did not include explicitly racial language. However, housing discrimination continues, even as advertisements have moved into the digital sphere. Audit studies show that landlords on Craigslist may discriminate against some potential tenants based on their race (Hogan and Berry 2011, Murchie and Pang 2018), and in 2019 the U.S. Department of Housing and Urban Development sued Facebook for discriminatory housing advertisements because of the way they targeted users based on race.

Rental listings from Seattle’s past show that their content was racialized and help us generate expectations for contemporary advertisements, based on existing theory. A simple, though unlikely, expectation is that contemporary rental housing advertisements follow the economic theoretical ideal as impartial instruments to help home-seekers find the rental unit that maximizes their utility given their budget. Under this post-racial hypothesis, we would expect no systematic racialized language. Alternatively, the growing sociological engagement with Critical Race Theory holds that racial disparities are (re)produced by racialized social systems (Bonilla-Silva 1997, Reskin 2012, Golash-Boza 2016), and that race is implicated in the present realities of urban life and the historical processes, which produced that present (Seamster 2015). This strain of research questions progress narratives (Seamster and Ray 2018) that would label language without explicit racial content as post-racial and attends to the processes that perpetuate racially disparate housing outcomes (Howell and Emerson 2018). Based on this Critical Race perspective, we expect that racialized language was re-codified in the half century since the Fair Housing Act removed the phrases “restricted district” and “whites-only” from the lexicon of rental advertisements in Seattle. In their place, we expect to see language patterns associated with various neighborhood racial compositions.

In this study, we use rental-housing listings from Craigslist in the Seattle–Tacoma–Bellevue, WA metropolitan area as a case study to understand how the description of housing opportunities varies systematically with the racial and ethnic composition of neighborhoods. We first use unsupervised topic modeling, a method that recognizes groups of words that often appear together, to identify a set of topics prevalent across our dataset of rental listings. Next, we estimate regression models to identify how much each topic is associated with neighborhood racial and ethnic composition. Finally, we describe the topics associated with variation across neighborhoods using qualitative analysis and deep reading. We find topical differences in rental listings across neighborhoods that are both statistically and practically significant. We call these observed differences racialized neighborhood discourse. This concept aligns with psychological and geographical work on the racialization of place (Weheliye 2014, Bonds and Inwood 2016, Bonam et al. 2017, Bonam et al. 2020) and the way new White residents often fail to integrate into existing community structures when moving to less White neighborhoods (Walton 2018). Our analysis contributes insight into Seattle’s landscape of racialized neighborhood discourse and explores how variations in language used to describe neighborhood locations could be implicated in the reproduction of residential segregation.

Background

Simultaneously taking account of both the implicit discursive nature of contemporary discrimination and its systemic nature is difficult. Pager and Shepherd (2008) recognize that large-scale studies can only examine unequal outcomes, whereas qualitative studies using interviews and experiments are able to investigate the role of preference. New data sources like advertisements from Craigslist come with new challenges in data processing and validation (Boeing and Waddell 2017, Boeing 2019, Boeing et al. 2020) but also offer a new opportunity to take hold of both the discursive and systemic aspects of contemporary racism.

Online rental advertisements are written by property owners and managers who are interested in finding a tenant for a rental unit. Some landlords have deployed a bevy of tactics of discrimination and tenant manipulation to maximize occupancy by preferred tenants (Rosen 2014, Greif 2018, Korver-Glenn 2018). Yet, these market actors need not have racist intentions in order to reproduce racialized neighborhood discourse. Instead, even in the absence of conscious or unconscious racial intention, there may be racial associations in the way property owners and managers habitually write about neighborhoods with racialized populations. Identifying those discursive associations requires careful empirical examination.

We define racialized discourse as variations in discursive form associated with differences in neighborhood racial composition. The concept of racialized discourse is similar to what Bonilla-Silva (2006); Ian 2015 calls colorblind racism, aligns with Critical Discourse Analysis (CDA) investigations of racism (Van Dijk 1993, Guillem 2018, Van Dijk 2018), and is most similar to what Jones and Jackson (2012, 218) call “discursive redlining,” or the process of discouraging (mostly White) people from visiting (mostly Black) neighborhoods. Colorblind racism, CDA, and discursive redlining all reveal racialized meaning without explicit racial content or racist intent and argue that apparently colorblind language is an essential contributor to the perpetuation of America’s racial order. However, those three methods are limited in how they locate texts and identify possibly racial meanings: the analyst must have strong prior knowledge about what kind of language about, say, neighborhoods, will be racialized. By leveraging the observed association between discursive forms—themes, phrases, ideas, descriptors—and neighborhood racial composition, our approach applies the key thrust of CDA and colorblind racism techniques to a broad corpus.

Neighborhood Discourse and Race

Racialized neighborhood discourse aligns with contemporary sociological theories of racism. While traditional accounts of racial discrimination focus on racially biased treatment of one individual by another, alternatives suggest that discriminatory outcomes can occur without explicit connection to race on the part of the actor (Pager and Shepherd 2008, Reskin 2012, Bracey 2015, Golash-Boza 2016). These new approaches include colorblind racism (Bonilla-Silva 2006) and racial habitus (Emirbayer and Desmond 2015), where racism is enacted and perpetuated in part through discourse and action that is not overtly racial. This view naturally incorporates the way that race, ethnicity and other ascribed characteristics are intertwined in how they attach to people and the way they influence differential outcomes.

The legacy of systemic racism in the United States has produced staggering racial inequality in education, income, wealth, and health (Massey and Denton 1993, Subramanian et al. 2005, Reskin 2012, Sharkey 2013). Racist policies like restrictive housing covenants and redlining intentionally located non-White neighborhoods in less desirable areas, with worse housing stock, and increasing poverty (Rothstein 2017). Those outcomes not only persist in the built environments of contemporary neighborhoods, but also influence perceptions of neighborhoods (Besbris et al. 2015; Bonam et al. 2016; Quillian and Pager 2001, Sampson and Raudenbush 2004). This research provides strong evidence that neighborhood perceptions are racialized in terms of disorder and safety.

Black Feminist theory provides an account of how material reality, especially Black bodies, becomes racialized due to ideology represented in discourse and everyday practices (Spillers 1987, Wynter 2001, Weheliye 2014). Saidiya Hartman (1997) pushes for attention to apparently quotidian activities and language in addition to more explicit events to understand the persistence of racism in the generations after slavery, arguing that while Black communities are not erased by slavery and later instantiations of racialized social systems, they are less visible. Scholars like Junia Howell (2018) and Louise Seamster (2015) have applied these insights to the study of housing and neighborhoods by investigating how race and racialized social systems create neighborhood hierarchies and perpetuate differences across space and race. This theory implies that in addition to issues of disorder and safety already treated in the social science literature, we should also expect to find discussion of neighborhood community and desirability racialized.

Present Study

We use topic modeling (DiMaggio et al. 2013, Egami et al. 2018) to analyze a large corpus of rental housing advertisements scraped from the online platform Craigslist. These text analysis methods have been used to examine public perceptions of pollution (Tvinnereim et al. 2017) and of popular music (Light and Odden 2017). Text analysis techniques require careful implementation, parameter tuning and tests for robustness. We follow Nelson’s (2017) methodology to ensure the validity of our findings using model tests and external data sources.

We explore how apartment listings contain racialized neighborhood discourse with a sample of listings collected from Seattle’s largest online housing market.ⁱⁱⁱ Specifically, we show how descriptions of less White neighborhoods as dangerous and uninteresting conform to expectations based on Critical Race Theory’s and Black Feminist emphasis on processes of racialization. Compared with the past when racial language was explicit but rare, contemporary discourse may be less obvious, but also more pervasive.

Data and Methods

We use a large set of rental-listing texts from the Seattle-Tacoma, WA Craigslist apartment page to conduct a mixed method analysis of discursive differences associated with neighborhood racial and ethnic compositions, as represented by a neighborhood typology. We identify those differences using Structural Topic Modeling (STM) (Roberts et al. 2014). Those analyses test the existence of discursive variation that could reflect racialized neighborhood perceptions. To investigate the content of those perceptions, we select representative texts using STM and subject those texts to qualitative analysis and deep reading.

The raw corpus contains 278,005 rental advertisements posted between March 2017 and September 2018, obtained daily using the Helena web-crawler (Chasins and Bodik 2017).^iv The advertisements are geolocated using scraped addresses and then matched to American Community Survey 2013–2017 five-year estimates at the tract level. Some listings may not be represented in our corpus: those without address data or were posted and removed between our web scraping periods. The Seattle metropolitan area is a particularly apt setting for this case study, given its low levels of racial segregation and high levels of neighborhood change (Krysan and Crowder 2017, Thomas 2017). Areas with higher segregation and less neighborhood change may have more calcified neighborhood perceptions, which could either seem so well known as to be unmentioned or so strong as to dominate the results. Seattle also has a large Craigslist market relative to its metropolitan population, with an average of more than 2000 listings per day.

Duplicate documents weaken the explanatory capability of topic models (Schofield et al. 2017) so we reduce the number of near-duplicate texts in the corpus. In topic models, duplicates tend to coalesce into a single topic. By iteratively fitting STMs, we remove all texts that would appear as duplicates to a human coder.^v In deciding which duplicates to discard, we keep the most recent listing of the ones that are ranked highly similar. This reduces the corpus to 45,358 listings from 848 census tracts in and around Seattle. The study area includes most of the Puget Sound region, including the cities of Olympia, Tacoma, Seattle, and Everett; a full map of the study area with neighborhood typology is included in the digital supplement.

We strip strings of characters that are not part of neighborhood discourse, like URLs and digits, from the texts in the corpus. Additionally, we remove mentions of neighborhoods in the texts, as they are collinear with neighborhood racial makeup. Finally, we implement a neighborhood typology based on that used by Crowder et al. (2012) and Hall et al. (2015) to understand differences in topic prevalence between neighborhoods of varying racial and ethnic compositions. We identify seven neighborhood types (given in order of prevalence): “Predominantly White,” “White Asian,” “White Latinx,” “White Mixed,” “Mixed,” “White Black” and “Majority Non-White.” The largest category, “Predominantly White,” included 246 tracts and 9,991 listings, while the smallest “Majority non-White,” included 14 tracts and 941 listings (see Appendix AII for details).

Structural Topic Modeling

We use STM to quantitatively describe common discursive features of the listings and examine associations between those features and neighborhood racial composition. Topic modeling is an unsupervised technique that generates estimates of topic proportion at the document level. We use regression to assess the association between a particular topic and neighborhood racial and ethnic composition. Estimated coefficients from these models indicate the association between certain patterns of language (i.e., word co-occurrence in STM) and the composition of neighborhood populations in the Seattle metropolitan area.

STM is faster and more reliable than other methods, including Latent Dirichlet Allocation (LDA) topic models (Egami et al. 2018; Roberts et al. 2014). STM uses a more consistent initialization process than LDA and can simultaneously estimate topic proportions and associations with covariates.^vi This estimation method, combined with the train-test split and other robustness checks, helps ensure that the associations observed between quantitative representations of discursive forms and racial proportions are present in the corpus and are not induced by analysis.

Selecting the proper number of topics, k, is an essential part of using Topic Models, including STM, for text analysis. We used STM’s built-in methods to choose a k of 40, but confirm that key topics are robust to the choice of k. We discuss STM model selection and robustness in depth in Appendix AIII. We validate the models on a test set of the same size and report the results from the test set. These results are shown in figures 2, 3, 4, and 7.

Figure 2.

Log-Level Coefficient Plot for Access Topics (N = 22,679). Association between a subset of topics and the neighborhood typology in a log-level model including all unit and neighborhood covariates. Error Bars show 95% Confidence Intervals.

Figure 3.

Log-Level Coefficient Plot for Marketing Topics (N = 22,679). Association between a subset of topics and the neighborhood typology in a log-level model including all unit and neighborhood covariates. Error Bars show 95% Confidence Intervals.

Figure 4.

Log-Level Coefficient Plot for Neighborhood Description Topics (N = 22,679). Association between a subset of topics and the neighborhood typology in a log-level model including all unit and neighborhood covariates. Error Bars show 95% Confidence Intervals.

Figure 7.

Log-Level Coefficient Plot for Unit Description Topics (N = 22,679). Association between a subset of topics and the neighborhood typology in a log-level model including all unit and neighborhood covariates. Error Bars show 95% confidence intervals.

Topic modeling is subject to significant analyst discretion, so we confirm the robustness of these findings in three ways. First, we check the dependence of the model output on the number of topics selected using the “robust LDA” method adjusted for STM to confirm that important topics appear with various selections for the number of topics, which they do (Casas et al. 2018). Second, we perform a simulation test randomizing the census tract of each document and repeating our statistical analysis. We compare the distribution of the coefficients we observed in our model with those produced in the permutation test and find that the observed coefficients would be unlikely to be observed under randomization using a significant Kolmogorov-Smirnov test where . Third, we report results only from a 50% held-out test set of 22,679 documents. Results in the training and test sets are substantively consistent.^vii

We add covariates at the listing level to control for variation in unit quality and at the tract level to account for superficially non-racial neighborhood context. At the listing level, we include the log of the rent and the square footage of the listing. At the tract level, we include the neighborhood typology, the proportion of persons living below the Federal poverty line, log of household median income, population in thousands, proportion of persons 25+ who are college educated, proportion of persons 16+ commuting to work by car, proportion of owner-occupied housing units, proportion of renter-occupied housing units in buildings with 20 or more units, the proportion of renter-occupied housing units in buildings built after 2010, the proportion of renter-occupied housing units built before 1940, log of the Euclidean distance to the nearest light rail station in meters, and log of the Euclidean distance to Seattle city hall in meters.

We fit an STM of 40 topics^viii to the deduplicated training set of 22,679 documents. We use the results of the topic model to investigate the relationship of each neighborhood type and the listing text. To do so, we estimate log-level OLS regressions of each topic on the neighborhood typology and other covariates listed above (40 models). We model the neighborhood typology as a categorical independent variable with the most common category, predominantly White, as the reference category, meaning that the coefficient for any other neighborhood type represents the difference in prevalence, on the log scale, between a predominantly White neighborhood and that neighborhood type, holding other variables unchanged.

In order to understand the content of the topics found to be associated with neighborhood racial composition, we perform qualitative analysis and coding on a subset of representative documents. For each of the 40 topics, we select the ten documents, which have the highest proportional match for that topic, resulting in a set of 400 representative documents. Using ATLAS.ti 8, we code each document for neighborhood and unit features. We pay special attention to features that appear salient in neighborhoods throughout our study area: distance and modes of transportation, neighborhood descriptions, locations described as being nearby the unit, features of the property or development, especially security features, and features of the unit. We also labeled each of the 40 topics at this stage with descriptive titles. This labeling preceded the quantitative analysis so the titles do not depend on the associations between topics and neighborhood types. The full output of the STM model, including the words most associated with each topic, are available in the supplemental materials.

Results

Quantitative Results: Systemic Textual Differences

Across neighborhood types, Craigslist rental listings tend to have a consistent format and present similar types of information. Since the purpose of these texts is to attract renters, the texts include information that home seekers might use to select a place to live. Accordingly, almost all advertisements include details about the unit for rent, like the size, number of bedrooms, amenities, and monthly costs. Based on our readings of historical listings, similar details were present in listings from the 1930s–1960s, which differ from contemporary listings primarily in length because newspapers charged posters by the letter.

Information about the surrounding area is common, but not universal. If a listing includes explicit information about the neighborhood, it tends to focus on nearby places to shop, eat, or visit, and local transportation. We consider it unlikely listings mentions amenities, attractions, or unit features that are not present, as that would be misleading. However, we imagine the opposite case, where details are omitted is quite common. Even if a particular unit or neighborhood feature, say a café or a security system, is present in reality, it may not show up in a text. Results from each step of our analysis focus on how these discursive patterns differ on average for advertisements from neighborhoods with differently racialized populations.

The STM produced 40 topics, defined by a high probability of containing certain groups of words, and a vector of topic proportions for each document. We label each topic by examining the words most associated with it and reading example texts, and refer to the topic by that label and, in parentheses, the numeric ID it was assigned by STM. We focus on how aspects of a listing, like the description of the unit and neighborhood, discourse about commuting and access, and marketing techniques, vary across neighborhood types. Table 1 shows key topics in each of those groupings, the words most associated with those topics and the titles we assigned through qualitative coding. The top words are based on STM’s “FR-EX” measure that uses word frequency and exclusivity within a topic (Roberts et al. 2014).

Table 1.

STM Topic Titles and Top Words

Topic title	Top words using “FR-EX,” stemmed
Access
Short and Central (Topic 7)	unit, plex, apt, triplex, build, coin
Driving and Bus Times (Topic 30)	hospit, safeway, westwood, meyer, colleg, express, mile
Commuting Distance (Topic 39)	burk, trail, gilman, microsoft, campus, googl
Convenience and Ease (Topic 40)	locat, great, conveni, beauti, easi, open
Marketing
Shared Units (Topic 1)	mother, cottag, mil, share, entranc, law
Developments as Communities (Topic 4)	afford, emerg, site, onsit, mainten, paperless
Pools and More (Topic 31)	court, tenni, pool, swim, burn, sauna
Subleases (Topic 32)	someon, subleas, leav, sinc, renew, know
Neighborhood Descriptions
High-Class Surroundings (Topic 8)	cours, golf, jefferson, height, point, sand
Condos with Views (Topic 20)	tower, elliott, skylin, stun, concierg, penthous
Safe and Friendly (Topic 26)	breed, restrict, patrol, weight, select, friday
Vintage Charm (Topic 28)	studio, vintag, classic, summit, brick, rail
Unit Descriptions
Cozy and Comfortable (Topic 16)	winter, warm, morn, nearest, keep, climat
Elegant Homes (Topic 17)	formal, bonus, piec, upstair, master, famili
Tenant Restrictions (Topic 19)	report, incom, reusabl, evict, histori, comprehens
Homes with Personality (Topic 23)	craftsman, basement, driveway, backyard, unfinish, furnac

While the titles of most topics are apparent from the “FR-EX” words, some, like “Short and Central” (Topic 7) relied on the qualitative coding process to produce the title.

We regress the log of the topic proportions for each document on the neighborhood typology and other covariates, listed above, to assess that topic’s association with neighborhood racial composition. We report the results of log-level OLS regression of estimated topic proportions on a held-out test set of documents using predominantly White neighborhoods as the reference category and include covariates with information about the units for rent, the neighborhood, and demographic context, listed above.^ix Log-level coefficients of on a neighborhood type dummy, for example, majority non-White, suggest that a switch from a listing in a predominantly White neighborhood to a majority non-White neighborhood is associated with an increase in the topic proportion of . Figures 2–5 report results from this model. Appearance on the left side of these plots indicates that the neighborhood type is associated with less of that topic compared to predominantly White neighborhoods, while appearing on the right indicates an association with relatively more of that topic.

Figure 5.

Spatial Distribution of “Safe and Friendly” (26) and “Vintage Charm” (28). Showing listings from the area surrounding Seattle, a subset of our study area, and which loaded > .1 on the topic of interest.

For example, results for “Commuting Distance” (Topic 39) and “Convenience and Ease” (Topic 40) are shown in figure 2. The coefficient associating “Commuting Distance” (Topic 39) with majority non-White neighborhoods is 0.51. This means that we expect that, on average, listings for a unit in majority non-White neighborhoods will include this topic or 84.9% more than listings in predominantly White neighborhoods. In contrast to the large effect for “Commuting Distance” (Topic 39), “Convenience and Ease” (Topic 40), a topic which centered on easy access to storage and convenient parking, had only small differences in its prevalence across neighborhood types. This accorded with a general trend that less racialized topics were more focused on unit characteristics than neighborhood or development characteristics.

The quantitative results identify which topics vary with neighborhood racial proportion. Topics pertaining to trust (Topics 1 and 32), personality and other less quantifiable positive qualities, like neighborhood community, (Topics 23, 28), and centrality (Topics 7 and 20) are associated more with predominantly White neighborhoods. Topics associated with travel (Topics 30 and 39), safety (Topic 26), and property amenities (as opposed to unit amenities) which locate community on-site rather than in the neighborhood (Topics 8, 4, and 31) are more associated with less White neighborhoods.

For example, consider two topics concerning trust, shown in figure 3. “Shared Units” (Topic 1) listings advertise units that are attached to the landlord’s home or property, often called accessory units, and “Subleases” (Topic 32) are requests for new tenants to assume a lease or sublet for a short period of time. Both of these arrangements require high levels of trust between the two parties. Our model estimates that, compared to a listing from a predominantly White neighborhood, a listing from a less White tract contains 15.0% (for White Mixed neighborhoods) to 23.7% (for majority non-White neighborhoods) less of “Shared Units” (Topic 1) and from a non-significant 3.1% increase (for majority non-White neighborhoods) to a 13.4% decrease (for White Latinx neighborhoods) of “Subleases” (Topic 32). This suggests that these high-trust arrangements may be most commonly advertised in the Whitest neighborhoods. We can also arrive at this same inference by examining figure 3 and noticing that Topics 1 and 32 are arranged on the left side of the plot, indicating their association with predominantly White neighborhoods.

We also compare regression results and spatial distribution, with an example shown in figures 4 and 5.

We can see from figure 5 that “Vintage Charm” (Topic 28) is clustered in central Seattle. This topic focused on smaller units in older buildings built with natural materials, included words like “vintage,” “charm,” “brick,” “hardwood,” and “studio,” and was associated with predominantly White neighborhoods. By contrast, “Safe and Friendly” (Topic 26) includes a number of central listings but has many more peripheral listings than “Vintage Charm” (Topic 28). In general, topics associated with Whiter neighborhoods are also more central. In part, this reflects the long-standing spatial and racial demographic order in Seattle. The oldest, most central, and most established neighborhoods have been Whiter because of explicitly racially motivated redlining and racial covenants enforced by real estate agents who could have been expelled from the Real Estate Board for non-compliance (Rothstein 2017). These historical patterns have been exacerbated by changes in Seattle over the past decades, as non-White and poorer populations have been pushed out of more desirable central areas by rising rents and evictions (Thomas 2017, Hess 2020). In other words, neighborhoods’ racial composition and their peripheral status are intertwined. The weak racialization of “Condos with Views” (Topic 20) is in line with our assertion that it is specifically neighborhood discourse that is racialized in these texts.

Both “Driving and Bus Times” (Topic 30) and “Commuting Distance” (Topic 39) were more associated with less White neighborhood types (see figure 2). Given that more central neighborhoods have also been Whiter in Seattle, it would be reasonable to consider that this association was a product of less White neighborhoods requiring more talk about transportation options simply because they were further from the center of the city. However, while the spatial distribution of these topics, shown in figure 6, does include listings in peripheral areas, both topics also include significant numbers of listings in central areas, especially less White central areas. While listings including “Vintage Charm” (Topic 28) are more common in Whiter, more central neighborhoods, less White neighborhoods are more likely to mention transportation options—ways to leave the area—regardless of whether they are central or peripheral.

Figure 6.

Spatial Distribution of “Driving and Bus Times” (30) and “Commuting Distance” (39). Showing listings from the area surrounding Seattle, a subset of our study area, and which loaded > .1 on the topic of interest.

This is true especially with mentions of Seattle’s light rail, which occurred in 6.3% of listings from neighborhoods with a Black proportion above the median for the sample, but only 2.0% of other listings. However, the light rail in Seattle passes through more traditionally Black neighborhoods, which could account for some of that difference. To measure that connection, we examined the prevalence of the term “light rail” only for listings geocoded to within one mile of the train’s route, leaving 5,691 listings from above-median Black tracts and 1,774 listings from other tracts. In that subset, 21.5% of listings from high-Black neighborhoods included “light rail,” while only 13.0% of listings from other neighborhoods did, a significantly larger prevalence with . Listings from tracts with higher Black proportion are much more likely to mention the light rail than listings from other areas that are equally close to that transportation option.

Qualitative Results: Community, Value, and Security

Our quantitative analysis shows that differences in discourse are significant and spread across a variety of topics, reflecting racialized neighborhood descriptions, which may be suggestive of racialized perceptions. In the qualitative analysis that follows, we use close readings of listing texts to understand the nature of racialized discourse and explore ways that discourse could influence residential patterns.

Since the model controls for neighborhood median household income and poverty rates, as well as the unit’s advertised rent and square footage, these results are not about the substantial economic differences between White and non-White neighborhoods. In fact, as shown in figures 4 and 7, the “High Class Surroundings” (Topic 8) and topics “Elegant Homes” (Topic 17) were weakly racialized, being slightly more common in certain less White neighborhoods. The qualitative analysis extends this understanding by showing that the important difference is not simply the amenities available in more and less White neighborhoods, but how those amenities and other property and neighborhood features are framed by text writers.

We begin with listings associated with “Vintage Charm” (Topic 28), mapped in figure 5. This example, from the predominantly White and high-profile neighborhood Capitol Hill, represents how listings from that topic describe the connection to their neighborhoods:

The St. Florence is a vintage building with extremely modern conveniences! Located at the intersection of several main streets of Capitol Hill (East Denny, East Olive & Summit) there are endless options for entertainment.

This development emphasizes the care taken with the “vintage” quality of the building, indicating a history worthy of preservation. People who live there have the benefit of connecting to that neighborhood and building its community. It is clear that the “endless options for entertainment” are also features of the neighborhood. By focusing on restoring this building, the development is marking Capitol Hill in both material and discursive terms as a place worth inhabiting and maintaining.

“Vintage Charm” (Topic 28) listings tended to be in central, popular areas, which were not only predominantly White in 2017, but also carried a history of redlining and restrictive housing covenants which excluded non-White residents by law. Furthermore, since our model includes housing stock age, the quantitative association between this topic and less White neighborhoods is not about the age of the buildings, but about what kinds of places have old buildings that count as “vintage” and “charming.”

In contrast, advertisements in less White areas turned inwards, shutting out the surrounding areas in favor of on-site amenities. Consider this “Pools and More” (Topic 31) excerpt from another ad, this one in Bitter Lake, an area where the Black population makes up a large share of all persons in the tract compared to the study area (about 10%):

Not only is this great condo perfectly located, but it also offers everything the busy person needs right in the complex! [The development] offers a gym, pool, spa, trails and beautiful grounds. Step out back and find a nice stretch of private grass at your disposal. And for those Seattle wintry nights, light a cozy fire in the fireplace and kick back. […]* A nice friendly and safe environment.

* 24 Hour security in complex.

While this development is only 300 feet from a public park, and within walking distance to grocery stores and coffee shops, the focus for this and other texts from “Pools and More” (Topic 31) is turned inward. The grass in the development is “private,” and the environment is “safe,” thanks to the 24-hour security.

Texts from “Developments as Communities” (Topic 4), significantly associated with majority non-White neighborhoods and White-Black neighborhoods, were also focused inwards. For example, one listing offered the opportunity to “express yourself in a community of unique apartment homes nestled in between private courtyards and lush Northwest landscaping.” Tenants can relax in the “private” courtyards. The “community” is identified with the unit type, apartment homes, rather than with actual people. It was common for listings in this topic from less White areas to use the phrase “Controlled-Access Community” to highlight that only certain people could enter the community.

These topics show how discourse about community varied in kind between neighborhood types. Community in Whiter neighborhoods was associated with the “vintage” past, and with the neighborhood itself as desirable. Community in less White neighborhoods was located on the development, casting the neighborhood as undesirable and unworthy of mention.

Racial discourse in the United States, whether explicit or implicit, has often focused on security and danger. Like amenities, security appeared in topics associated with both high and low non-White proportions. Similarly, words like ““safety,”“ ““security,”“ ““secure,”“ and ““controlled”“ were present in listings from all neighborhood types. However, security terms were more prevalent in high-Black neighborhoods, and over 1,200 listings in the full dataset mentioned a “Courtesy Patrol.” As the website for one development explains, “[a] Courtesy patrol is on duty every night to ensure that you have a good night’s rest.” Many more mentioned a “Night Patrol,” or a “concierge.” In the STM, “patrol” was a high-loading word for “Safe and Friendly” (Topic 26), which was associated with less White neighborhoods. Compared to predominantly White neighborhoods, the model expects no significant difference for White mixed and White Latinx neighborhoods, 18.8% more of this topic in White Asian neighborhoods, 25.9% more in Mixed neighborhoods, 47.8% more in White Black neighborhoods, and 61.8% more in majority non-White neighborhoods.

It was not only the relative scarcity of security talk that distinguished Whiter listings, but the manner in which that talk was presented. Another text from “Vintage Charm” (Topic 28) demonstrates this:

Quiet street in a safe, very walkable neighborhood, with many cafes, restaurants, and food-shopping options to choose from […] Very centrally located -- wherever you choose to go — Alki beach in West Seattle, hiking in Issaquah, the mall in Bellevue, etc, you're no more than a 20-min car ride away. The apartment is a rare find - an affordable, light-filled, spacious place on a quiet street in a great, safe, walkable and fun neighborhood. Come check it out!

Here safety is not only a feature of the neighborhood but is combined with other aspects of the neighborhood: walkability and fun. This is a marked contrast to the idea of a courtesy patrol in a development otherwise depicted as isolated from its surroundings. Security discourse in less White neighborhoods is more focused on protecting the property with security systems and “Courtesy Patrols.” In predominantly White neighborhoods, by comparison, discussions of safety are less common overall and, perhaps more importantly, point to a safe neighborhood rather than just a protected property.

Discussion

The racialized neighborhood discourse described above is markedly different from the explicit racial exclusion seen in the 1930s or the coded racial language of “restricted districts” seen until 1968. However, that difference cannot be called progress, and it certainly does not suggest the attainment of a post-racial present. Instead, our analysis shows pervasive thematic differences in the way listings in predominantly White neighborhoods treat community, trust, transportation, and safety when compared to less White neighborhoods. White neighborhoods are depicted with ties to history and community, while listings in less White neighborhoods enclose community in developments with their pools and fitness centers. Advertisements for living situations requiring high trust, in which, for example, the renter and landlord live on the same property, are more common in predominantly White neighborhoods. Even accounting for distance to transportation options, listings in less White neighborhoods seem more focused on how to leave a neighborhood than what you can do there. These differences in listing text, which are associated with the racial composition of a neighborhood, constitute racialized neighborhood discourse. Overall, these findings support critical race accounts of a racialized social system insidiously woven into the fabric of American social life (Bonilla-Silva 1997, Reskin 2012, Golash-Boza 2016). Listing texts include racialized discourse in this account not because text writers hold prejudicial beliefs, but because historical forces have produced racialized understandings of neighborhoods as having community or not, as being safe or unsafe, as being sufficient or insufficient. More specifically, the findings about safety confirm previous social science research on neighborhood perception (Quillian and Pager 2001, Sampson and Raudenbush 2004, Besbris et al. 2015, Bonam et al. 2016), while the findings on community, and desirability support Black Feminist theories (Hartman 1997, Seamster 2015, Howell 2018, Howell and Emerson 2018). The findings about trust and transportation are new, further support considering listing texts as part of a racialized social system that contributes to the reproduction of racism.

Racialized neighborhood discourse in online rental texts could influence the perpetuation of racialized housing patterns in at least three ways. While the cross-sectional nature of this study makes it difficult to assess the effects of these linguistic differences, considering them in the context of contemporary theoretical and empirical work on the search for housing and the reproduction of segregation aligns with the recent trend towards process-focused accounts of residential attainment and mobility. Specifically, Krysan and Crowder (2017) propose considering residential attainment in terms of housing search. Taking that view focuses attention on three particular ways the advertisements might relate to the way home seekers sort themselves/are sorted into neighborhoods.

First, for housing searchers with little local information and ample resources who initially include many areas in their search, these differences might lead them to exclude certain neighborhoods, which seem to offer fewer amenities or connections to community. These new movers are a significant portion of the population: in Seattle, a large part of our study area, more than 130,000 people on average moved into the metropolitan area each year from 2013 to 2017. This is the most straightforward way racialized discourse might influence housing dynamics.

Second, these texts both maintain and contribute to shared knowledge about neighborhoods, even among current or long-time residents of the Seattle metropolitan area. Krysan and Crowder (2017) recognize that individuals’ neighborhood perceptions are shaped by many factors, including personal experience in the neighborhood, hearsay, and exposure to media accounts of the neighborhood. Craigslist listings likely contribute to the media-driven portion of the perception-creation process. In the case of security or community discourse, this could serve to reinforce existing biases in long-time residents not only influencing those residents’ own housing search, but also likely influencing the way those residents might suggest or warn against neighborhoods to other home-seekers.

Finally, these differences might emphasize material differences in the neighborhoods that vary with racial composition due to residential segregation. Racialized neighborhood discourse combines distorted perceptions of neighborhoods, like the emphasis on security, with accurate perceptions of neighborhood difference caused by historical discrimination, like the peripheral location of less White neighborhoods. When those combined perceptions are reproduced in discourse like Craigslist listing text, they show up as natural and objective. The discourse then perpetuates the association as normative: making it seem that less White neighborhoods are not only more dangerous and farther from the city center, but also that such a discrepancy is natural, not worth understanding or resisting. In this case racialized neighborhood discourse would not only influence new and long-term resident’s perception of particular neighborhoods but change the space-based racial stereotypes they hold about neighborhoods with racialized residents.

Far from mutually exclusive, these three ways that observed differences in rental texts could influence housing searches likely work synergistically. In addition to existing explicit discrimination and exclusionary tactics by landlords and real estate agents (Greif 2018, Korver-Glenn 2018), home seekers face neighborhoods, which are described differently based on who lives there already. Further work should investigate whether and how those unequal descriptions influence residential mobility and attainment.

Future work should assess the relative contributions of these three pathways to the reproduction of segregation, for instance, by investigating whether varying the level of racialized neighborhood discourse in an advertisement influences a reader’s perception of that neighborhood, and whether the effect is different for readers with different amounts of knowledge about the metropolitan area.

The observed patterns of discursive difference may also correspond to the way people act in neighborhoods after they move there. Walton (2018), through her ethnographic work in stably diverse neighborhoods in Boston, identifies two of what she calls “habits of Whiteness”: (1) anxiety—worry about the security of a neighborhood or if it is a sufficient place to live; and (2) ambivalence—uncertainty about the value of existing neighborhood culture, amenities, and social networks. Both of these habits are present in the discursive patterns observed in less White neighborhoods above. While language associated with predominantly White neighborhoods focuses on connections to neighborhood past and present, language associated with less White neighborhoods seeks to assuage anxiety by emphasizing safety and to counter ambivalence by showing off property features that mirror the neighborhood features of White spaces. This discourse treats White neighborhoods as sufficient places for living, while framing non-White neighborhoods as places that need to be adorned and secured to be suitable for habitation.

By tracing the association between discourse and neighborhood race, we expose a new realm of racialized discourse, which is difficult to access without a large corpus, computational methods, and variation in the racialized contexts of the texts production. Racist language as recognized by colorblind racism (Bonilla-Silva 2006, Ian 2015), Critical Discourse Analysis (CDA) (Van Dijk 1993, Guillem 2018, Van Dijk 2018), or discursive redlining (Jones and Jackson 2012, 218) may not be explicitly racist in language, but still expresses a racist idea, stereotype, or worldview held by the speaker. Racialized neighborhood discourse expands the scope of those theoretical cases, adding cases where no amount of discussion with text writers could reveal racial views. Put bluntly, there may be some cases where landlords use racialized discourse to express racist housing preferences, those instances should be discoverable by both traditional approaches to discourse and our method. There are also, we argue, cases where intentionality is less clear cut and perhaps not essential: when people do not harbor even hidden racist views, but still produce racialized neighborhood discourse. Our perspective leads us to methods that also reveal these cases—though we cannot distinguish between intentional and unintentional cases.

Further work with this perspective may be able to show what processes do the social work of racializing neighborhood discourse and how that racialization can affect housing search, residential attainment, and the perpetuation of segregation. One way of assessing that could be through comparing texts from the same management companies across neighborhoods, which might be possible with an even larger sample of advertisements.

There are a few limitations to this study worth noting. While the validity of these conclusions within the greater Seattle area during the study period is strong, there are potential limitations to these findings’ generalizability to other spaces, times or online rental platforms. These conclusions are based on data from Craigslist, which contains only a subset of online rental advertisements, and what gets advertised online is only a subset of the rental market as a whole. Existing research shows that there are racial differences in housing search method (Desmond 2017), and that may also influence racialized discourse on Craigslist. Recent research has shown that Craigslist listings are unequally distributed across neighborhoods (Boeing 2019, Boeing et al. 2020) and that there are differences between online listing providers (Hess et al. 2019).

Revealing racialized discourse in these texts is possible because of the history of racialized housing policies that imbued the spatial distribution of people and neighborhood qualities with inequality by race. We leverage that unequal distribution to uncover associated variation in listing text. However, in the absence of a corresponding spatial inequity, this method is not useful for investigating gendered language, or language structured around other axes of oppression. Future work would investigate other ways of making this analysis more intersectional, perhaps using more traditional discourse analysis approaches.

We include covariates for class, income, and education in order to account for variations across neighborhoods that might confound differences in discourse associated with racial compositions. However, we do not include adjustment for residual spatial autocorrelation or other neighborhood level covariates. Models including spatial autocorrelation still recover racialized discourse similar to that discussed above, although the apparent racialization of topics with clear spatial dimensions, like “Short and Central” (Topic 7) and “Pools and More” (Topic 31), which is focused in areas of new development, is reduced. We argue that because the history of legally enforced residential segregation in the United States (Massey and Denton 1993, Rothstein 2017) produced spatially structured racial neighborhoods, modeling spatial autocorrelation misattributes the evil residue of racial exclusion as statistical error.

For most of the history of the United States, policy has focused on imposing, rather than reducing, racial disparities in neighborhood outcomes of all kinds. This paper shows that unequal and racialized neighborhood perceptions are part of the present in Seattle. It seems that private markets, like Craigslist, reflect racialized neighborhood discourse and suggests that more equitable results may require robust construction of public housing and assistance with navigating the housing market, with explicit attention to racialized discourse and perception, following Rothstein (2017). The existence of these perceptions suggests that housing policies, including zoning changes, housing vouchers, and the construction of new public housing will not be able to achieve racial integration if they do not face racialization of neighborhood discourse head-on. Communities which have successfully maintained integration, like Oak Park, IL, have done so by carefully managing their image with explicit focus on racial meanings in advertising and through actively opposing false conceptions in counseling with prospective tenants (Krysan and Crowder 2017).

We join a burgeoning community of sociologists (Zuberi 2011, Golash-Boza 2016, Ray et al. 2017, Bonilla-Silva 2019) pushing to incorporate theoretical work from feminist Black Studies exemplified by Hortense Spillers (Spillers 1987), Sylvia Wynter (Wynter 2001), and Saidiya Hartman (Hartman 1997). Those Black Feminists and others push for a focus on processes of racialization—like racialized neighborhood discourse—which could drive apparently non-racial action and discourse (Wynter 2007, Weheliye 2014). From this perspective, and aligning with Reskin (2012), studying language used in Craigslist listings may provide a window into processes, which drive the racialization of space, and influence segregative mobility patterns. We show that racialization of neighborhood discourse takes both expected forms—twisting the presentation of community, desirability and safety of neighborhoods—and unexpected forms—racializing neighborhood trust and transportation options. Sociologists studying race and space should continue to improve our understanding of the pervasiveness and impact of each of these processes of racialization. If we can more accurately trace and understand processes like these that reproduce contemporary segregation, we may be able to learn how to slow down, halt, or reverse them.

Ian Kennedy is a PhD student in the department of sociology at the University of Washington. Their work focuses on how contemporary processes of racism can be exposed using new data sources and methods.

Chris Hess is a postdoctoral fellow in the Rutgers-Camden department of public policy and administration and Center for Urban Research and Education. His research explores the suburbanization of residential segregation by race and ethnicity, disparities in housing cost burden and the role of online housing markets in the housing search process.

Amandalynne Paullada is a PhD candidate in the department of linguistics at the University of Washington whose research focuses on natural language processing. Ongoing work involves applying machine learning methods to information extraction from scientific publications, particularly in the biomedical domain.

Sarah E. Chasins is an assistant professor in the department of electrical engineering and computer sciences at University of California—Berkeley. Her research group develops programming tools for social scientists, data scientists, and other non-technical domain experts. Their goal is to put the power of computation in reach for a broader and more diverse audience.