Searching for housing in the digital age: Neighborhood representation on internet rental housing platforms across space, platform, and metropolitan segregation

Abstract

Understanding residential mobility, housing affordability, and the geography of neighborhood advantage and disadvantage relies on robust information about housing search processes and housing markets. Existing data about housing markets, especially rental markets, suffer from accuracy issues and a lack of temporal and geographic flexibility. Data collected from online rental platforms that are commonly used can help address these issues and hold considerable promise for better understanding the full distribution of available rental homes. However, realizing this promise requires a careful assessment of potential sources of bias as online rental listing platforms may perpetuate inequalities similar to those found in physical spaces. This paper approaches the production of rental advertisements as a social process driven by both contextual and property level factors. We compare data from two online platforms for the 100 most populated metropolitan areas in the United States to explore inequality in digital rental listing spaces and understand what characteristics are associated with over and underrepresentation of advertisements in certain areas. We find similar associations for socioeconomic measures between platforms and across urban and suburban parts of these metropolitan areas. In contrast, the importance of racial and ethnic composition, as well as broader patterns of segregation, for online representation differs substantially across space and platform. This analysis informs our understanding of how online platforms affect housing search dynamics through their biases and segmentation, and highlights the potential and limits in using the data available on these platforms to produce small area rental estimates.

Introduction

Current, temporally granular, and geographically flexible rental housing market data have historically been unavailable to scholars and policymakers. More recently, the number of digital rental platforms has increased, and these platforms have become a common source of information for households searching for a new place to live. Krysan et al. (2018) find that 67% of recent movers have used internet advertisements to identify potential units to rent and learn about neighborhoods where units are available. Growth in the number of online listings and their use in the housing search process have led to efforts to scrape data from these platforms to deliver in-depth insight into rental housing markets. Information on asking rents and housing characteristics at the unit level have provided opportunities for research on topics including the availability and quality of the housing stock available to lower-income renters; the association between housing cost fluctuations and various types of criminal activity at the micro-level; and the change in rent levels over time at small areas that can be used for rent calculations for various housing programs. However, there has been only limited assessment of the data collected from these digital rental platforms and their potential biases.

In particular, there is growing evidence based on data from Craigslist that certain neighborhoods may be systematically underrepresented among online listings (Boeing, 2020; Boeing et al., 2020a). Further, rental listings in underrepresented neighborhoods may be less descriptive about the unit and include more information about tenant qualification criteria for example (Besbris and Faber, 2017; Besbris et al., 2018; Boeing et al., 2020b). Biases in neighborhood representation have direct implications for the quality of rent estimates and validity of other insights drawn from these listing sources, but there are also important consequences for those who may use these platforms to locate housing opportunities. Extant research points to market segmentation, with some users favoring certain platforms over others as the reliability and quality of the listings on some platforms may not be trusted by certain users (Krysan et al., 2018). To the extent that the opportunities advertised online match the existing geography of inequality in the physical world, listings from these online spaces may require adjustments to generate valid rental estimates and otherwise illustrate limits of the internet for democratizing information about housing vacancies.

Using data collected from two major online rental platforms (Craigslist and Apartments.com) for the 100 most populated metropolitan areas in the United States, we analyze variation in advertisement prevalence across neighborhoods (defined as census tracts) as a function of metropolitan area, online platform, and neighborhood characteristics. The following research questions guide this study:

1. Do the neighborhoods represented online differ between digital rental advertising platforms?

2. Does online neighborhood representation differ between cities and suburbs?

3. Do variations in online representation among urban and suburban neighborhoods vary based on the broader patterns of residential stratification by race, ethnicity, and poverty?

The empirical portion of this study examines the potential and limits of these data for capturing changes in rental markets at a high frequency for small geographic areas, as well as how these platforms may serve to reinforce stratification within the housing search process despite being freely available sources of housing information.

We find relatively comparable associations between neighborhood representation and socioeconomic measures like the share of college-educated persons and prevalence of poverty between platforms and across urban and suburban parts of these metropolitan areas. These similarities in overrepresentation of high socioeconomic status neighborhoods contrast with considerable differences across space and platforms in the importance of racial and ethnic composition, particularly Black and Latinx representation, and broader patterns of segregation by race and ethnicity for online representation. We find that Craigslist, in particular, underrepresents Black and Latinx neighborhoods in metropolitan regions with higher levels of segregation, whereas Apartments.com representation varies based on the socioeconomic composition of a neighborhood. These findings have implications for how researchers and policymakers use online rental listing data to produce rental estimates that can supplement existing survey measures. The findings are also related to the broader theoretical framework about stratification in the housing search process and how unequal access to rental market information can further reinforce residential segregation.

Literature review

Online platforms represent a potential source of information for analyzing the rental market, with an increasing number of articles using data scraped from online listing platforms for understanding housing dynamics and estimating rents for specific geographic areas (e.g. Besbris et al., 2018; Boeing, 2020; Boeing and Waddell, 2017; Boeing et al., 2020a, 2020b; Hess et al., 2019; Kennedy et al., 2020; Olsen, 2019; Palm, 2018). Growing usage in housing and demographic research reflects an understanding that estimates produced from online rental housing market advertisements can improve the timeliness and geographic granularity for estimates of rental market dynamics in a given area. These sources can be particularly valuable in countries that lack regular housing surveys, even at the national level (Peppercorn and Taffin, 2013), but also in countries in which such surveys by public or private providers do not have sufficient sample sizes to produce spatially specific estimates needed to understand local rental market dynamics (Boeing et al., 2020b; Hess et al., 2019). This paper focuses on the United States context, but some of the dynamics described are likely to apply to other countries, although to our knowledge limited work has been done on neighborhood representation on online platforms in other countries despite emerging work on the use of such platforms to analyze housing markets (Chapelle and Eymeoud, 2018).

In the United States, extant data traditionally comes from three types of sources: (1) national surveys such as the American Community Survey (ACS) or the American Housing Survey; (2) local rent board administrative data; and (3) proprietary sources such as landlord associations or market data providers (Boeing et al., 2020a). Importantly, each source suffers from shortcomings that may be resolved through estimates generated from collecting online listings from various rental platforms.

For example, the ACS has a substantial temporal lag between data collection and publication of small area estimates, creating a situation where even the most recent estimates include survey responses about rent payments that are more than 5 years old, for example, the 2015–2019 ACS estimates were released on 10 December 2020, and are partly based on observations collected in the Spring of 2015. Furthermore, because sources such as the ACS include currently leased units, they are of limited utility for characterizing housing opportunities and costs for those currently searching for housing. Local rent board administrative data remain limited to a few jurisdictions and are generally not publically available. And while several private firms collect data and some even share estimates free of charge (e.g. Zillow), these products tend to be available only at aggregate levels and lack transparency in terms of how the estimates are produced, which limits replication.

Even so, questions exist about whether publicly available online platforms such as Craigslist or Apartments.com can provide an effective source of micro data on housing units for rent, given that selective coverage of online rental housing advertisements creates concern about representativeness (Boeing et al., 2020a). In addition, aggregate estimates of rental volumes and costs may be biased if certain units and locations are over or underrepresented on these platforms. Online rental listings have the potential to improve existing rental market estimates by providing high frequency and current small area rental estimates, but the extent to which they capture the overall rental market needs to be understood first. Understanding the nature and extent of these biases is therefore a prerequisite for being able to process the raw data collected from online rental platforms into informative indicators to better understand rental markets.

There is emerging evidence that, in the United States, the data collected from these platforms vary in terms of coverage across unit types and neighborhoods of varying sociodemographic compositions. Boeing et al. use Craigslist data in a number of recent papers analyzing 2014 data (and 2018 data for the most recent paper) for the central cities of the 50 most populated metropolitan regions in the United States (Boeing, 2020; Boeing and Waddell, 2017; Boeing et al., 2020a, 2020b). While these studies’ findings indicate that the rank order of small area rent estimates produced based on Craigslist data are generally similar to estimates based on survey data, there are large differences in rent levels between them that can be explained by differences in temporality—the surveys capture rent for current renter occupied households rather than for units currently on the market. In addition, there appears to be differences in the types of units advertised on Craigslist relative to the overall housing stock (Boeing, 2020; Boeing and Waddell, 2017; Boeing et al., 2020a).

Recent research based on listings from Craigslist and Apartments.com further supports the existence of systematic differences in rent levels between estimates based on online listings and existing sources of rental data (Hess et al., 2019). Data analyzed for Seattle, Washington, San Antonio, Texas, and Fort Lauderdale, Florida, reveals that ZIP codes with online rental listing estimates substantially above or below the U.S. Department of Housing and Urban Development’s range of fair market rents had important differences in their population and housing unit composition. For example, in areas where the online rental estimates are higher than the fair market rents, there are higher poverty rates, vacancy rates, and shares of multifamily units, while also lower homeownership rates (Hess et al., 2019).

Differences in neighborhood representation have apparent importance for estimating typical rents in a market or neighborhood, but also potentially have bearing on search outcomes for households who use these platforms. Housing searches are already complex processes with major long-term consequences for households since neighborhood and housing unit selection has implications for the quality of services and amenities that a household may access such as schools, recreational spaces, and public services, and can impact a household’s commute time, financial situation, and overall well-being (Clark, 2005; Krysan and Crowder, 2017; Ludwig et al., 2012; Newburger et al., 2011). The location of the home can also impact life outcomes for children (Chetty and Hendren, 2018; Chetty et al., 2016). To secure housing, though, households need access to information to identify available units with location, size, and cost characteristics that meet their search criteria. Therefore, the degree to which certain unit types and locations are under or overrepresented on the online listing platforms can have concrete implications on households’ location and housing outcomes.

The process of finding a unit to rent is driven by factors such as preferences, budget constraints, real and perceived discrimination, existing social networks, neighborhood perceptions, and past home searches (Krysan and Crowder, 2017). It is an iterative rather than linear process, but at various stages throughout the process, a crucial factor is the ability to gather information about available units that meet search preference criteria (Walter and Wang, 2016). Identifying a set of potential units is not enough to ensure households are able to rent their preferred unit though. Rental agent and landlord behavior, as well as time pressure and cognitive resource depletion, limits how many units households visit and whether they sign a lease for their most preferred home. Despite this, access to information about a wide range of potential units that can be narrowed down based on household search criteria remains a crucial part of securing desirable housing, and online platforms make it easy to do this without spending a lot of time and money.

Online platforms such as Craigslist have become a major source of information about potentially available units to rent (Boeing et al., 2020a) that are often used in conjunction with other housing search methods such as word of mouth from social networks (family, friends, and colleagues), signs posted in front of the home, and real estate agents (Krysan, 2008; Krysan et al., 2018). These digital spaces have the potential of removing some of the barriers that exist in physical spaces by allowing households to reduce search costs (both time and pecuniary costs) and limiting the potential for discrimination and steering by removing the mediation of real estate agents. The primary barrier left is internet access itself, and though the choice of device is increasingly irrelevant, access and adoption of high-quality broadband internet remain stratified due to uneven availability in lower-income areas (Federal Communications Commission, 2020) and high subscription costs even where available (Horrigan and Duggan, 2015).

Alternatively, under or overrepresentation of certain unit types and neighborhoods in these digital spaces and segmentation across platforms may further contribute to residential stratification rather than alleviate it (Boeing, 2020; Krysan, 2008). The efficacy of relying on online platforms to narrow a search based on what is available depends on the platform’s ability to provide an equal representation of available options. Even if a particular platform is not expected to capture all available listings, biases in over or underrepresentation of certain unit types or locations might affect a household’s ability to identify a unit that meets their need in a location of their choice. Understanding the nature and extent of biases is necessary to understand what elements of the sociocultural structure of rental markets may translate from the physical to the digital rental space.

To further assess differences in the prevalence of online listings across neighborhoods, Boeing (2020) develops a measure of over/underrepresentation based on the number of listings on Craigslist for a given neighborhood (census tract) relative to the number of listings one would expect if Craigslist listings had the same distribution as the tract rental vacancy rate as captured in the ACS 5-year estimates. The rental vacancy rate serves as a proxy for obtaining an expected number of listings. This estimated ratio $(ƛ)$ is equal to one if the number of Craigslist listings in a given tract is proportional to that tract’s share of the city’s total vacant rental units. Boeing (2020) finds that for his sample of 2014 Craigslist listings in the “core municipalities” of the 50 largest metropolitan statistical areas, they were overrepresented in White, wealthier, and more educated census tracts, and particularly underrepresented in census tracts with a higher share of Black and Latinx residents.

Over 50% of metropolitan residents live outside of the core municipalities of metropolitan areas and platforms such as Craigslist cover these suburban markets too, raising a question of how representation among suburban contexts compares to urban ones. Despite some ambiguity in their definition (Forsyth, 2012), suburbs account for a substantial share of all renters in metropolitan areas and increasingly vary in their position within the broader metropolitan social structure (Kneebone and Berube, 2013). Given the salience of socioeconomic status, race and ethnicity for the differences in representation observed in Boeing (2020), trends of rising poverty and increasing racial/ethnic representation in some suburbs suggest that there may be comparable inequality in terms of the neighborhoods advertised through online rental platforms (Allard, 2017; Hess, 2020). Further, residential segregation by race and ethnicity has changed over time to reflect greater suburbanization, with separation between different suburbs accounting for an increasing share of metropolitan segregation since 1990 (Lichter et al., 2015). Similar spatial segregation patterns between cities and suburbs are also present in European countries (Bolt et al., 2008; Quillian and Lagrange, 2016). As such, online housing platforms may play a role in stratifying housing search processes among households moving to the suburbs just as theorized with cities.

Further work is also needed to establish whether findings from previous studies reflect general dynamics across different platforms and metropolitan characteristics, or only localized patterns of usage for Craigslist in particular metropolitan areas. Beyond focusing only on larger municipalities within large housing markets, the existing research does not consider whether features of the broader metropolitan area influence the relative salience of neighborhood characteristics such as racial/ethnic composition for online representation. Additionally, some prior research makes the assumption that unique online posts equate to unique housing opportunities, even when a subset of users may aggressively list the same units with different advertisements, jeopardizing the validity of this data definition. Finally, Craigslist is one of many online listing sources, so the question of whether other prominent sources such as Apartments.com have similar variations in representation looms large.

Using a sample of 100 metropolitan areas and listings from two different online platforms, this paper uses the methodology developed by Boeing (2020) to analyze patterns of over/underrepresentation on online platforms and assess how these might affect housing search processes and reproduce existing inequalities in housing markets. Existing evidence about under/overrepresentation on different rental platforms motivates our investigation of the geography of online rental housing unit platforms to establish their usability in the study of housing markets. This information is crucial for realizing the promise of these platforms as new sources of information about housing market dynamics. Eventually, establishing sources of over/underrepresentation creates the opportunity for producing rental estimates that correct these biases. Such estimates can increase our understanding of social processes such as neighborhood change, gentrification, and suburbanization given the role of the housing markets in these processes. Segregation in terms of housing and tenant search strategies on the part of renters and landlords, respectively, also theoretically underpins variations in platform usage between neighborhoods of varying racial, socioeconomic, or housing unit composition. Overall, this research contributes new insights into the process through which “Big Data” sources such as online rental listing advertisements are generated and provides novel evidence on the extent to which these new digital spaces reproduce those found in physical spaces.

Data

Our sample covers all census tracts from the 100 most populated metropolitan areas in the United States at the time of the 2013–2017 ACS 5-year population estimates. We investigate differences in neighborhood representation for advertisements on the “apartments/housing for rent” section of Craigslist as well as for all types of rental units advertised on Apartments.com. Anyone can post classified ads to Craigslist and they can do so at no cost in most markets, with this platform only requiring landlords to provide a title, rent asked, some description, a ZIP Code and an email address for a new listing. In contrast, Apartments.com requires users to be the owner or a legal actor on behalf of a property to create a listing, and mandates that users provide a physical address and home type for any new listing. Both platforms allow potential tenants to view listings at no cost, though Apartments.com has promoted content where landlords can pay to elevate their listing within the overall index. To collect advertisements from these two platforms, we ran a set of web scraping scripts daily written in Helena and Python. The raw data collected from these programs from January through August in 2019 were then cleaned and processed into a database with cumulative listing tables for each source.

We define unique listings based on their combination of advertised bedroom size, square footage, and location to deduplicate our data such that a given listing theoretically matches a unique housing unit vacancy. We describe this data definition in comparison with alternatives in Part A of our online supplement to provide background on how the deduplication impacts sample summary statistics and the estimated distribution of online representation. After this deduplication process, our sample of tracts covers 847,068 Craigslist listings and 462,073 Apartments.com listings.

We use 2015–2019 ACS 5-year estimates to measure characteristics of neighborhoods, defined as census tracts, across the 100 metropolitan areas with online rental listing data. Although tract data are an imperfect approximation of neighborhood areas, these data represent the most spatially granular information about population and housing composition available for the time period that our scraped listing data cover. Our city sample focuses on tracts whose internal points fall within the boundaries of a principal city, with the suburb sample capturing the remaining tracts in our sample of metropolitan areas. Part B of our online supplement lists and describes the sociodemographic measures used in this study.

Our measure of representation on each platform, lambda $(\lambda )$, uses ACS information on rental vacancies to assess how many ads might be expected compared to this reference distribution. Each platform is treated separately in computing $\lambda$ and its associated quantities since this facilitates comparing differences in representation and distribution of listings between platforms. We use the number of vacant housing units for rent in a tract $(\tau )$ to proportionally allocate the count of listings $(\kappa )$ in the metropolitan area for each tract. We compute the expected number of listings in a tract $(\varphi )$ for a given platform using the following formula:

$${\varphi }_{\text{tract}}=\frac{{\kappa }_{\text{metro}}{\tau }_{\text{tract}}}{{\tau }_{\text{metro}}}$$

This expected number of listings $(\varphi )$ is then compared with the observed number of advertisements $(\kappa )$ to assess how much a tract’s representation differs from what is expected based on the distribution of vacant rental units as measured by the ACS. A given tract’s $\lambda$ is then computed with the following formula:

$${\lambda }_{\text{tract}}=\frac{{\kappa }_{\text{tract}}+1}{{\varphi }_{\text{tract}}+1}$$

When $\lambda$ is 1, the number of listings observed in the tract matches the expected number based on proportional allocation. Lambdas that are between 0 and 1 denote underrepresentation, and $\lambda \text{s}>1$ denote overrepresentation. We present bivariate maps of underrepresented tracts (i.e. $\lambda <1$) on each platform in Part C of our online supplement.

Methods

Between-group differences

We conduct Cohen’s $d$ analyses for each platform to describe differences in neighborhood population and housing compositions between overrepresented and underrepresented tracts, using our overall data and our sample stratified by city and suburb locations. This analysis helps explain how different neighborhoods that are overrepresented on a platform (i.e. $\lambda >1$) and are from those that are underrepresented $\left(\lambda <1\right)$ by expressing the mean difference in terms of pooled standard deviations:

$$d=\frac{{\mu }_o-{\mu }_u}{{\sigma }_p}$$

Positive $d$ estimates imply that levels of a given characteristic are greater in overrepresented neighborhoods, while negative $d$ values correspond to overrepresented neighborhoods having lower values than underrepresented neighborhoods. A set of conventional thresholds aid with interpreting this analysis, with absolute values of $d<0.2$ considered negligible, between 0.2 and 0.5 weak, between 0.5 and 0.8 moderate, and >0.8 strong.

Regression analysis

We include two types of linear models estimated by ordinary least squares (OLS) in this study. The first models describe the importance of neighborhood racial/ethnic composition and socioeconomic composition for differences in representation on Craigslist and Apartments.com.¹ We estimate these models with our overall sample on each platform, as well as for cities and suburbs separately, to understand how urban and suburban neighborhoods potentially contribute to patterns of representation in different ways. The model specification uses the following equation:

$$y={\beta }_0+{\beta }_1{X}_1+ϵ$$

In this formula, $y$ is the log-transformed $\lambda$ for representation on a platform, ${\beta }_0$ is the intercept, $X_1$ is a matrix of tract-level sociodemographic measures, ${\beta }_1$ are the focal coefficients of interest, and $ϵ$ is the model error.

Our second set of models investigate whether associations of neighborhood composition with representation depend on the broader pattern of residential segregation in metropolitan areas. We use an interaction term between dissimilarity indices at the metro level and the respective neighborhood composition measure at the tract level to provide insight into dynamics such as whether the representation of Black neighborhoods in a highly segregated metropolitan area differs from the expected representation of such neighborhoods in a more integrated one. In these models, we include a comparable set of racial/ethnic, socioeconomic, housing, and population composition covariates measured at the metropolitan level to adjust for differences between metropolitan areas other than their patterns of segregation.

Results

Differences in population and housing composition between underrepresented and overrepresented neighborhoods

We start by assessing systematic differences in representation with Cohen’s $d$ analyses using our set of neighborhood population and housing characteristics. Figure 1 presents the standardized effect sizes (Cohen’s $d$) computed between overrepresented and underrepresented neighborhoods using our overall sample, city sample, and suburb sample. In this figure, red circles denote the difference in composition between these groups on Apartments.com and blue triangles indicate the difference between these groups on Craigslist.

Figure 1.

Cohen’s $d$ analyses.

The neighborhood characteristics at the top of Figure 1 show that overrepresented tracts have a different socioeconomic profile than underrepresented tracts, on average—with overall similar patterns across the two platforms. The share of college graduates, median household incomes, and median rents are generally greater in the neighborhoods that appear on Craigslist and Apartments.com. Moreover, the separation between overrepresented and underrepresented neighborhoods is even greater on Apartments.com than on Craigslist. In a similar vein, the largest negative $d$ value is for neighborhood poverty rates, indicating that overrepresented neighborhoods have lower poverty rates than underrepresented tracts. As with the other socioeconomic indicators, Apartments.com appears to have an even greater difference between over- and underrepresented tracts than Craigslist. Finally, the larger median bedroom size in overrepresented tracts is one explanation for higher median rents in overrepresented tracts compared with underrepresented tracts. These observations cumulatively suggest that the disproportionately commercial profile of properties advertised on Apartments.com captures a more socioeconomically selective range of neighborhoods than Craigslist, where both commercial and “mom and pop” landlords are likely to post rental vacancy advertisements.

Figure 1 also shows clear differences between overrepresented and underrepresented neighborhoods in terms of racial and ethnic composition. Neighborhoods with more Craigslist and Apartments.com activity than expected tend to have larger White population shares and have less Black representation than underrepresented neighborhoods. They also have smaller Latinx population shares in the case of Apartments.com but the differences on that dimension are small for Craigslist.

These findings are largely consistent with Boeing (2020), who found similar differences on measures of racial and ethnic composition albeit with slightly larger effect sizes. The difference between studies likely corresponds to the broader swath of neighborhoods considered in this analysis, compared with focusing in particular on urban contexts in the 50 largest metropolitan areas. First, our sample not only includes traditional urban enclaves for Latinx people, but also “new destinations” among suburbs that may be relatively segregated at the block level but well represented because they are located within tracts with other developments that are actively represented (Lichter et al., 2010). Second, including smaller metros where segregation tends to be less acute and non-White populations may be smaller adds over/underrepresented neighborhoods that are not as consistently differentiated by race/ethnicity but instead just socioeconomic status. Finally, racially/ethnically integrated middle/higher-income suburban tracts that may be represented well online and would reduce average differences between over and underrepresented captured by Cohen’s $d$. Taken together, these results are consistent with theoretical expectations drawn from research on racial and ethnic differences in housing search strategies such as those proposed by Krysan and Crowder (2017). To the extent that some metropolitan areas have more entrenched patterns of residential stratification by race and ethnicity, though, we might expect to see differences in the typical racial/ethnic composition of over- and underrepresented neighborhoods based on the context of metropolitan segregation the neighborhood is embedded within.

The remaining differences in Figure 1 of non-negligible size (i.e. $d>0.2$) show that overrepresented neighborhoods have greater values for characteristics such as their share of single-family homes and median owned housing value, and lower values for their share of housing units built prior to 1940 and share of renter households who are housing cost burdened (i.e. paying 30% of more of their household income toward monthly gross rent). There are no substantial differences between platforms across these measures, consistent with the relative consistency between platforms on the other socioeconomic indicators. The $d$ values for internet access and cellular internet-only suggest that both internet access and quality of access are relevant to patterns of neighborhood representation—overrepresented neighborhoods have more households with internet access, while underrepresented neighborhoods have more households whose only connection is a cellular device (i.e. not broadband). Many of the neighborhood characteristics that have negligible differences between over and underrepresented neighborhoods have little to no difference between platforms as well, with a few exceptions (e.g. college student, age 20–34 years, foreign born).

The other facets of Figure 1 disaggregate the prior analysis to understand differences in the composition of over and underrepresented neighborhoods between urban and suburban locations of metropolitan areas. In general, there is a remarkable degree of consistency in the differences between over and underrepresented neighborhoods across metropolitan space and between platforms. In both urban and suburban settings, the socioeconomic characteristics such as the share of persons with a college degree, median household income, and poverty rate have similar $d$ estimates, and across locations, we tend to see larger effect sizes for Apartments.com than Craigslist.

Despite the consistency for measures with the largest effect sizes in the overall analysis, there are still some differences in the composition of neighborhoods that appear online based on their urban/suburban location. For example, the differences for the Black population share, housing cost burdened share, and presence of housing built prior to 1940 between over and underrepresented neighborhoods are smaller among suburbs than cities. This is consistent with the greater presence of newly built housing, relatively less segregated neighborhood conditions and higher average levels of socioeconomic status in suburbs when compared with cities. Many of the characteristics where the sign of the $d$ estimate is different between cities and suburbs fall into the negligible effect size range, suggesting that under and overrepresented neighborhoods are not drastically different on these measures.

While there may not be as stark of a difference in racial and ethnic composition between neighborhoods represented online in suburbs, the findings with respect to the poverty rate show how overrepresented neighborhoods have a similar magnitude of difference in the presence of low-income populations. To the extent that Black and Latinx households are disproportionately represented in poor suburban contexts but still underrepresented in suburbs relative to Whites, the smaller effect size for Black and Latinx population shares may simply reflect a narrower range of average racial/ethnic compositions (in terms of Black and Latinx shares) observed among suburbs.

Multivariable associations of neighborhood composition and online representation

Table 1 presents coefficients from linear regressions estimated using OLS where we model the multivariable associations between neighborhood composition and the log of $\lambda$. We estimate a model for each platform using our overall sample as well as the subsets of tracts in city and suburb areas.

Table 1.

Linear regressions of $\text{log}(\lambda )$ for Craigslist and Apartments.com.

	Overall		City		Suburb
	Craigslist	Apartments.com	Craigslist	Apartments.com	Craigslist	Apartments.com
Neighborhood Racial and Ethnic Composition
Non-Latinx Black	−0.419** (0.134)	0.159 (0.133)	−0.777*** (0.132)	−0.310** (0.115)	−0.420* (0.172)	0.226 (0.165)
Latinx	0.258 (0.187)	0.428** (0.155)	0.108 (0.204)	0.314* (0.136)	0.100 (0.241)	0.191 (0.221)
Non-Latinx Asian/Pac. Islander	0.161 (0.257)	0.159 (0.216)	0.141 (0.294)	0.298 (0.232)	0.092 (0.309)	−0.135 (0.263)
Non-Latinx Native American/Alaska Native	0.213 (0.623)	0.011 (0.663)	−0.276 (1.198)	−0.420 (1.253)	0.305 (0.653)	0.180 (0.655)
Non-Latinx Other Race	1.377** (0.507)	0.693* (0.344)	−0.345 (0.536)	−0.740* (0.367)	2.177*** (0.544)	1.221** (0.454)
Neighborhood Socioeconomic Composition
Poverty Rate	−0.678*** (0.172)	−1.161*** (0.153)	−0.541*** (0.158)	−0.859*** (0.157)	−0.585* (0.272)	−1.269*** (0.213)
College Degree	0.953*** (0.147)	1.209*** (0.1 10)	0.683*** (0.190)	0.919*** (0.198)	0.791*** (0.179)	1.191*** (0.112)
log(Median HH Income)	−0.144 (0.077)	−0.438*** (0.077)	0.055 (0.086)	−0.141 (0.106)	−0.147 (0.107)	−0.496*** (0.086)
Neighborhood Housing Composition
Total Rental HU	−0.216*** (0.044)	−0.056 (0.035)	−0.188*** (0.040)	−0.070 (0.039)	−0.208*** (0.061)	0.015 (0.038)
Vacancy Rate	−15.630*** (0.592)	−12.969*** (0.431)	−16.816*** (0.793)	−14.089*** (0.582)	−15.152*** (0.600)	−12.541*** (0.466)
Built Before 1940	−0.527*** (0.106)	−0.922*** (0.065)	−0.413** (0.135)	−0.741*** (0.062)	−0.660*** (0.117)	−1.138*** (0.097)
Median $N$ Rooms	0.014 (0.031)	0.082*** (0.022)	0.095* (0.041)	0.121*** (0.031)	−0.040 (0.027)	0.036 (0.024)
Median Gross Rent	0.135** (0.046)	0.167*** (0.043)	0.090 (0.056)	0.050 (0.051)	0.120** (0.045)	0.186*** (0.041)
Same Home Last Year	−0.829** (0.295)	−1.786*** (0.258)	−1.043*** (0.314)	−1.873*** (0.318)	−0.459 (0.327)	−1.402*** (0.278)
log(Distance to CBD)	−0.322*** (0.031)	−0.128*** (0.021)	−0.219*** (0.032)	−0.066** (0.024)	−0.344*** (0.046)	−0.123*** (0.027)
Neighborhood Population Composition
Age 20–34	0.185 (0.262)	0.200 (0.183)	0.667* (0.332)	0.491* (0.206)	−0.649* (0.295)	−0.087 (0.320)
Age 65+	0.516 (0.289)	0.071 (0.283)	0.202 (0.333)	−0.176 (0.312)	0.418 (0.332)	0.062 (0.310)
College Student	−0.420* (0.187)	−0.872*** (0.150)	−0.305 (0.251)	−0.644*** (0.193)	−0.337 (0.287)	−0.988*** (0.254)
English Speaking Only	0.972*** (0.196)	0.718*** (0.149)	1.183*** (0.225)	1.144*** (0.162)	0.763** (0.254)	0.341 (0.189)
log(Average HH Size)	0.238** (0.076)	0.100 (0.057)	0.001 (0.088)	−0.018 (0.077)	0.284*** (0.077)	0.1 19 (0.065)
Neighborhood Internet Connectivity
Internet Access	1.399*** (0.282)	1.920*** (0.229)	0.336 (0.274)	0.818*** (0.225)	2147*** (0.316)	2.666*** (0.252)
Cellular Internet Only	−0.358* (0.160)	−0.638*** (0.176)	−0.042 (0.155)	0.020 (0.210)	−0.494* (0.220)	−0.985*** (0.177)
Constant	−0.541 (0.482)	0.960* (0.446)	−0.147 (0.512)	0.781 (0.540)	−0.897 (0.526)	0.852* (0.353)
$N$	44,835	44,835	18,355	18,355	26,480	26,480
$R^2$	0.472	0.491	0.489	0.508	0.472	0.493

HH: household; HU: housing unit; CBD: central business district.

Cluster-robust standard errors in parentheses.

^*p < .05;

^**p < .01;

^***p < .001.

Starting with the neighborhood racial and ethnic composition measures, there is some consistency in the neighborhood populations associated with representation across the two platforms. The one exception to this is Black representation, where there are significant coefficients of opposite signs in the overall sample–while higher levels of percent Black are associated with lower levels of Craigslist representation, the relationship is reversed for Apartments.com. The remaining columns where we model representation among cities and suburbs separately show that there is a strong association between Black representation and lower levels of $\lambda$ across platforms among urban neighborhoods, with the findings noted from the overall sample driven by different associations observed among suburban neighborhoods.

Also notable is that representation of neighborhoods on Craigslist is relatively comparable across levels of Latinx concentration ceteris paribus, in contrast to Apartments.com, which shows higher levels of representation for neighborhoods where greater shares of Latinx persons reside. This dynamic is driven by Apartments.com overrepresenting neighborhoods with greater shares of Latinx persons among urban areas, with both platforms showing no difference in representation of suburban neighborhoods associated with Latinx population shares.

Among suburban neighborhoods, there is also greater representation associated with other race population shares on both platforms. These findings show that all else equal, concentration of racial/ethnic groups itself does not tend to imply lower levels of representation, aside from perhaps Black population composition, and even this relationship may be picking up on broader patterns of residential segregation more than neighborhood composition per se.

Socioeconomic measures remain consistently correlated with representation even after controlling for the influence of other neighborhood characteristics. A greater share of college-educated persons in a neighborhood is associated with a greater degree of representation online, and a higher poverty rate is consistently associated with lower $\lambda$ values. These relationships are consistent across Table 1’s models, with the poverty rate and college degree coefficients the same in terms of sign and significance regardless of sample employed and are consistent with existing findings for central cities (Boeing, 2020). The association for median income with representation is different between Craigslist and Apartments.com, with higher income tracts generally less represented on Apartments.com–a pattern driven mainly by an association among suburban areas. Overall, these associations suggest that higher socioeconomic status neighborhoods are consistently more represented online, regardless of location or platform.

The total count of rental inventory and vacancy rates have inverse relationships with representation, with high inventory and high vacancy in a neighborhood associated with less Craigslist representation and high vacancy associated with less Apartments.com representation. Neighborhoods with relatively older housing stocks are expected to have lower degrees of representation on both platforms, though this dynamic does not necessarily imply these neighborhoods are underrepresented based on the strong salience of other characteristics such as vacancy rates and socioeconomic composition of the population. Higher rent neighborhoods have relatively greater representation online, but this is notably driven by suburban neighborhoods when looking at the spatially disaggregated models. Neighborhoods where there is less turnover in housing tenure are less represented across locations and platforms, and both platforms have greater representation for neighborhoods close to the metropolitan central business district.

The other population composition measures show somewhat varied associations by platform and location. For example, urban neighborhoods with greater shares of younger adults (i.e. age 20–34 years) tend to be overrepresented on both platforms, but all else equal, only Craigslist shows significant underrepresentation of suburban neighborhoods with more younger adults. Similarly, tracts with greater shares of households whose only internet access is with a cellular plan tend to be underrepresented online on both platforms, with this finding in the overall models accounted for mostly by significant associations among suburban areas.

Metropolitan segregation, neighborhood composition, and online representation

The final component of our empirical results focuses on the salience of metropolitan segregation by race, ethnicity, and poverty status for the observed associations between neighborhood composition and $\lambda$ for each platform. We predict $\lambda$ values at varying degrees of neighborhood composition for counterfactual metropolitan areas where segregation exists at levels representative of the 10th (yellow), 50th (gray) and 90th (blue) percentiles of Black–White, Latinx–White, and poverty segregation, with all other population and housing characteristics held at the sample median values.² These figures include 95% confidence intervals based on cluster-robust standard errors for all prediction series.

Figure 2 presents predictions for Black–White segregation using our overall sample as well as the city and suburb specific samples. The differential importance of racial segregation between Craigslist and Apartments.com is apparent when looking at the predictions from the overall model. Whereas greater degrees of Black–White segregation are associated with lower representation of neighborhoods with higher Black concentration on Craigslist, there is a smaller decline in $\lambda$ associated with a neighborhood’s Black composition in metropolitan areas with median levels of segregation and even an increase in $\lambda$ as the Black share increases in a relatively integrated metropolitan area. These results show that the negative coefficient for tract Black share in the models without metropolitan characteristics reflects dynamics in more segregated metropolitan areas where broader patterns of residential stratification disadvantages Black neighborhoods. In contrast, there is little difference in the positive association between the Black share of a neighborhood population and $\lambda$ for Apartments.com across degrees of segregation. All said, these dynamics suggest that Craigslist’s potential to draw in landlords of all sizes makes it more likely to reproduce existing patterns of neighborhood stratification when compared with a commercially oriented platform such as Apartments.com where socioeconomic factors largely drive representation online.

Figure 2.

Adjusted predictions for Black–White segregation and tract % Black.

The middle of Figure 2 investigates the Black–White segregation and Black neighborhood composition dynamics within cities, where we observe lower predicted $\lambda$ values for Craigslist in general and a steep drop off in representation for Black neighborhoods in metropolitan areas with median or greater levels of segregation. Even in more integrated metropolitan areas, there is a slight decline, though the visualized differences across levels of percent Black do not reach statistical significance. This dynamic differs by platform, though, with no significant difference in Apartments.com representation associated with greater Black population shares, regardless of the broader metropolitan context.

Predictions for suburban neighborhoods shown on the right side of Figure 2 continue to highlight the differential importance of segregation between the platforms covered in this study. Although the predicted levels of representation have wider confidence intervals, there is still a significant difference in representation between the high segregation and low segregation predictions in neighborhoods where the Black share of the neighborhood population is 25% or greater. These generally older suburban neighborhoods where Black representation is the greatest difference in their Craigslist representation across levels of racial segregation, while there is again essentially no difference in the association of neighborhood composition with $\lambda$ across degrees of racial metropolitan segregation on Apartments.com.

Figure 3 shows adjusted $\lambda$ predictions for varying degrees of Latinx–White segregation and neighborhood Latinx population shares. There are again important differences in the salience of ethnic segregation between the two platforms covered in this study. Looking at all neighborhoods together, predictions for Craigslist show how greater degrees of Latinx–White segregation in a metropolitan area is associated with lower levels of online representation as the Latinx share of neighborhood populations increases. Greater Latinx presence in a neighborhood relates to essentially no change in $\lambda$ for metropolitan areas with median levels of Latinx–White segregation, and greater levels of representation in metropolitan areas with the lowest levels of Latinx–White segregation. This Craigslist dynamic stands in contrast to the segregation associations observed for Apartments.com, where there is essentially no interaction between broader patterns of segregation and the neighborhood ethnic composition with respect to predicted $\lambda \text{s}$.

Figure 3.

Adjusted predictions for Latinx–White segregation and tract % Latinx.

When looking at associations for urban neighborhoods separately, there is more consistency between platforms to the extent that the Latinx share of a neighborhood’s population relates to essentially no difference (Apartments.com) or slight differences (Craigslist) in online representation depending on metropolitan Latinx–White segregation. Results for suburban locations, however, point to differences between the two platforms, with greater segregation associated with significantly less representation of Latinx neighborhoods on Craigslist but only modest differences for representation of these neighborhoods on Apartments.com. A likely explanation for these differences stems from platforms covering different areas based on the landlords and tenants that use the site. To the extent that the distribution of Craigslist ads is shaped more by landlord’s own housing search strategies than Apartments.com (where socioeconomic factors are most salient), we see greater sensitivity of representation on Craigslist to racial/ethnic segregation.

Finally, Figure 4 shows the interaction between degrees of poverty segregation in a metropolitan area and the neighborhood poverty rate for predicted levels of $\lambda$ on each platform. Unlike the two prior analyses related to racial and ethnic segregation, the overall results show consistency between Craigslist and Apartments.com. In metropolitan areas with lower levels of poverty segregation, there is actually greater salience of neighborhood poverty for levels of online representation. These predictions hold constant racial/ethnic composition and metro segregation, though, so the model still captures how a poorer neighborhood with larger Black or Latinx representation would be particularly underrepresented even if poverty itself is contributing relatively less to the lower $\lambda$.

Figure 4.

Adjusted predictions for poverty segregation and tract poverty rate.

In the spatially disaggregated predictions of Figure 4, there is even more evidence that metropolitan areas with lower levels of poverty segregation tend to represent higher poverty neighborhoods more than contexts where poor households tend to be more segregated from nonpoor households. Among urban neighborhoods, high poverty neighborhoods tend to be slightly more represented in metropolitan contexts with high poverty segregation compared with low poverty segregation. For suburban neighborhoods, there is a decline in representation as to the neighborhood poverty rate increases across the three counterfactual levels of poverty segregation, but the steepest declines are nonetheless in metropolitan areas with lower degrees of separation between poor and nonpoor households. The dynamic noted among suburban neighborhoods appears to account for much of the interaction observed in the overall sample.

Finally, we conducted several important sensitivity checks to ensure that our results were not biased by outliers or purely driven by the use of the ACS vacant for rent count distribution for calculating $\lambda$. First, we estimated our segregation models using robust regression (MM-estimator) and observed substantively similar conclusions to the models estimated by OLS. Second, and perhaps more importantly, we estimated negative binomial models to assess how our neighborhood sociodemographic measures associated with the observed listing count in a given tract. In these count models, we observe comparable dynamics for our focal composition–segregation interactions and thus conclude that these differences related to segregation as measured by $\lambda$ in our focal models are related to variations in user activity first and foremost.

Discussion

The present study underscores the importance of online rental housing platforms for understanding contemporary housing search processes and residential mobility in the United States. We observe substantial differences in the degree to which neighborhoods are represented based on their population and housing composition, though the relevant characteristics of neighborhoods differ across urban and suburban locations, degrees of metropolitan segregation, and the platform in question. Online rental advertisements provide greater insight into residential opportunities in higher State Emergency Service (SES) locations, and in more segregated metropolitan areas, neighborhoods with relatively less Black and Latinx representation.

As a result, White households using these online platforms will be exposed to a disproportionate level of information about units in neighborhoods where Whites are already concentrated and higher-income households will be more likely to be exposed to information about units in higher-income neighborhoods they might already be familiar with. These biases in representation can therefore reinforce existing blinders associated with their existing residential trajectories, social network, and familiar communities (Krysan and Crowder, 2017) even when households search for units that meet their budget constraints but with limited geographic filters. In addition, variations in representation are larger in more segregated metropolitan regions. This means that in regions that are currently experiencing higher levels of segregation, the housing search experience is stratified with households using online platforms less likely to be exposed to units in neighborhoods with greater racial/ethnic representation. At the same time, in less segregated metropolitan regions, households may still have segregated preferences but the platform they rely on for identifying potential units will not a priori rule out neighborhoods with greater racial/ethnic representation.

Since variations in neighborhood representation online reflect existing residential inequalities in physical space, under and overrepresentation in listings on platforms such as Craigslist and Apartments.com contribute to continued stratification during the search process, which can further reinforce segregation by race, ethnicity, and socioeconomic status. Households that do not utilize these platforms may find housing via social networks or physical advertisements that are less available in the neighborhood contexts that tend to be overrepresented online. Stratified information about housing opportunities, digital or otherwise, will only reinforce existing patterns of residential inequality. The fact that online platforms that were expected to remove barriers to access information are unlikely to short-circuit segregated housing search strategies and present a household with housing opportunities in areas they would not otherwise consider represents a challenge in decreasing further segregation. Taken together, online platforms have limited capacity for integration simply through democratizing information, and housing scholars and activists should focus on disrupting how different households search for their next dwelling. Nonetheless, the quality information provided by these sources for high opportunity and low poverty neighborhoods suggests practitioners may find them useful within housing assistance programs where households receive guidance about how and where to search for rental opportunities.

The present study also informs ongoing efforts to use forms of “Big Data” such as online rental advertisements for producing temporally and spatially granular information about housing markets. Studying the sociodemographic and spatial biases within these online platforms (as in the present study) is substantively interesting on its own and this research also plays an important role in informing the potential pitfalls of using Big Data for different academic and policy research applications. The implications of the biases documented here will vary across applications, but highlight that the use of raw data scraped from online sources may produce biased research conclusions and faulty policy decisions when the target population is broader than the platform data were scraped from (e.g. Craigslist). Researchers should see biases noted in this and prior studies as motivation for finding ways to mitigate non-representativeness. With proper adjustment, scraped rental data could be a crucial source of information in rapidly evolving contexts where “ground truth” data simply do not exist but policymakers need to act to help households avoid worsened housing cost burden, housing instability or eviction.

There are some limitations to our study worth noting. Craigslist and Apartments.com are not the only sources of rental housing information on the internet, though they are two of the most well-known. Other platforms for listing rental units may differ in the degree to which socioeconomic factors play such a strong role (e.g. GoSection8), how much prevailing patterns of segregation shape the distribution of advertisements (e.g. Facebook Marketplace), and their use by “mom and pop” landlords in particular (e.g. Nextdoor). Nonetheless, this research provides an important contribution by considering the contours of neighborhood representation online beyond Craigslist alone and demonstrates how comparing different sources presents insights into these platforms’ varying spatial coverage and patterns of use. Additional research extended to other platforms and potentially other countries would provide additional evidence on the extent to which digital rental spaces provide comprehensive coverage of the local rental markets and whether they reproduce the patterns observed in physical rental spaces.

The study is also limited by the reliance on the ACS tract vacancy data to define an expected distribution for creating the $\lambda$ ratio. The vacancy measure captures a stock of units identified as vacant for rent at the time of data collection rather than the flux of vacancy over the observation period. The large share of neighborhoods with zero vacant units for rent is not ideal and likely reflects sampling and definitional issues rather than neighborhoods where there was no vacancy over the course of time covered. Finally, simply counting the number of advertisements within a neighborhood does not provide insight into what these advertisements say and how they describe neighborhoods, which may have its own independent effect on segregating housing search outcomes (Kennedy et al., 2020).

The current results indicate that high SES and low poverty neighborhoods are disproportionately represented on the sites analyzed with some differences exacerbated based on the local patterns of segregation. Further work is needed to explore whether these differences reflect a market segmentation with listings for lower SES and higher poverty neighborhoods concentrated on different platforms or advertised through offline channels, or if there is an overall lower level of publicly available information in these markets, which would further limit opportunities for integrative moves by households unfamiliar with these neighborhoods. Looking at whether neighborhoods undergoing changes in rents are more or less represented and whether the content of the ads is different could provide information on the potential role of those platforms in gentrification processes.

In addition, work should be done to improve our estimates of the expected number of vacancies in a given period by using mobility measures to capture flux rather than stock of vacancies. This would allow us to explore whether the results from this paper and the existing literature (Boeing et al., 2020a, 2020b) are robust to alternative coverage measures. Finally, the biases noted in this study necessitate efforts to adjust observed information from these platforms through methods that incorporate the timeliness and geographic specificity of these scraped data while still building estimates on reference distributions that have more objective validity for covering all types of neighborhood and metropolitan contexts. Poststratification weighting has been used to adjust for compositional biases found on other “Big Data” platforms such as Facebook and may provide a similar solution for housing research (Alexander, 2021).

The representation of different neighborhoods online is a crucial part of understanding how movement of housing search processes into the digital domain may reproduce existing residential inequalities despite information becoming more open in its accessibility. Online rental housing platforms such as Craigslist and Apartments.com have some similarities: high SES and low poverty neighborhoods are disproportionately represented on these sites and both platforms have comparable associations with population and housing characteristics across urban and suburban parts of metropolitan areas. However, these sources of information differ substantively in the salience of neighborhood racial and ethnic diversity for the degree to which we “see” neighborhoods online as much as might be expected. Although online platforms potentially democratize information about housing opportunities, they are nonetheless part of the cycle of segregation and not a panacea for the inequalities that exist outside of these digital spaces.

Supplementary Material

Supplemental material for this article is available online.