# LongBench v2 / 66f55d66821e116aacb33734

task_id: a442dd8b-8803-5ba9-b73b-4ea8b7f8143f
task_key: train--66f55d66821e116aacb33734
task_revision_id: 3

{"choice_A":"For the homeowners of these houses, the author only used machine learning methods to analyze their account profile pictures and determine characteristics such as race, gender, and age.","choice_B":"This article employs a randomized trial method, selecting 10 experimental areas and creating 20 Airbnb test accounts to randomly book houses listed as \"available\" on the website eight weeks in advance.","choice_C":"The author categorized homeowners into six major groups based on their different responses, focusing primarily on those landlords who requested more information from tenants.","choice_D":"The author collected past guest reviews from homeowners' web pages to ensure the validity of the experiment.","context":"American Economic Journal: Applied Economics 2017, 9(2): 1–22 \nhttps://doi.org/10.1257/app.20160213\n1\nRacial Discrimination in the Sharing Economy: \nEvidence from a Field Experiment†\nBy Benjamin Edelman, Michael Luca, and Dan Svirsky*\nIn an experiment on Airbnb, we find that applications from guests \nwith distinctively African American names are 16 percent less likely \nto be accepted relative to identical guests with distinctively white \nnames. Discrimination occurs among landlords of all sizes, includ-\ning small landlords sharing the property and larger landlords with \nmultiple properties. It is most pronounced among hosts who have \nnever had an African American guest, suggesting only a subset of \nhosts discriminate. While rental markets have achieved significant \nreductions in discrimination in recent decades, our results sug-\ngest that Airbnb’s current design choices facilitate discrimination \nand raise the possibility of erasing some of these civil rights gains. \n(JEL C93, J15, L83)\nO\nver the past 50 years, there have been considerable societal efforts to reduce \nthe level of discrimination against African Americans in the United States. In \nthe context of housing and rental accommodations, antidiscrimination laws have \nsought to eliminate discrimination through regulation. While racial discrimination \ncontinues to exist in rental markets, it has improved in the last two decades (Yinger \n1998, US Department of Housing and Urban Development 2013; compare Zhao, \nOndrich, and Yinger 2005 to Ondrich, Stricker, and Yinger 1999).\nYet in recent years, markets have changed dramatically, with a growing share \nof transactions moving online. In the context of housing, Airbnb has created a new \nmarket for short-term rentals that did not previously exist, allowing small landlords \nto increasingly enter the market. Whereas antidiscrimination laws ban the landlord \nof a large apartment building from discriminating based on race, the prevailing view \namong legal scholars is that such laws likely do not reach many of the smaller land-\nlords using Airbnb (Belzer and Leong forthcoming; Todisco 2015).\nIn this paper, we investigate the existence and extent of racial discrimination \non Airbnb, the canonical example of the sharing economy. Airbnb allows hosts \n* Edelman: Harvard Business School, Morgan 462, 25 Harvard Way, Boston, MA 02163 (e-mail: bedelman@\nhbs.edu); Luca: Harvard Business School, Baker Library 457, 10 Harvard Way, Boston, MA 02163 (e-mail: mluca@\nhbs.edu); Svirsky: Harvard Business School and Harvard University Department of Economics, Baker Library 420A, \n25 Harvard Way, Boston MA 02163 (e-mail: dsvirsky@hbs.edu). We thank Ian Ayres, Larry Katz, Kevin Lang, \nSendhil Mullainathan, Devah Pager, and seminar participants at eBay, Harvard Law School, Hong Kong University \nof Science and Technology, Indiana University, New York University, Northwestern University, Stanford University, \nand University at Albany for valuable feedback. We thank Haruka Uchida for tireless research assistance. Our \nInstitutional Review Board approved our methods before we began collecting data. IRB# 15-2226.\n† Go to https://doi.org/10.1257/app.20160213 to visit the article page for additional materials and author \n \ndisclosure statement(s) or to comment in the online discussion forum.\n\n\n2\t\nAmerican Economic Journal: applied economics\b\napril 2017\nto rent out houses, apartments, or rooms within an apartment. To facilitate these \n­\ntransactions, Airbnb promotes properties to prospective guests, facilitates commu-\nnication, and handles payment and some aspects of customer service. Airbnb allows \nhosts to decide whether to accept or reject a guest after seeing his or her name and \noften a picture—a market design choice that may further enable discrimination.\nTo test for discrimination, we conduct a field experiment in which we inquire about \nthe availability of roughly 6,400 listings on Airbnb across five cities. Specifically, \nwe create guest accounts that differ by name but are otherwise identical. Drawing \non the methodology of a labor market experiment by Bertrand and Mullainathan \n(2004), we select two sets of names—one distinctively African American and the \nother distinctively white.1\nWe find widespread discrimination against guests with distinctively African \nAmerican names. African American guests received a positive response roughly \n42 percent of the time, compared to roughly 50 percent for white guests.2 This \n8 percentage point (roughly 16 percent) penalty for African American guests is par-\nticularly noteworthy when compared to the discrimination-free setting of competing \nshort-term accommodation platforms such as Expedia. The penalty is consistent \nwith the racial gap found in contexts ranging from labor markets to online lending \nto classified ads to taxicabs.3\nCombining our experimental results with observational data from Airbnb’s site, \nwe investigate whether different types of hosts discriminate more, and whether dis-\ncrimination is more common at certain types of properties based on price or local \ndemographics. Our results are remarkably persistent. Both African American and \nwhite hosts discriminate against African American guests; both male and female \nhosts discriminate; both male and female African American guests are discrimi-\nnated against. Effects persist both for hosts that offer an entire property and for \nhosts who share the property with guests. Discrimination persists among experi-\nenced hosts, including those with multiple properties and those with many reviews. \nDiscrimination persists and is of similar magnitude in high- and low-priced units, in \ndiverse and homogeneous neighborhoods.\nBecause hosts’ profile pages contain reviews (and pictures) from recent guests, \nwe can cross-validate our experimental findings using observational data on whether \nthe host has recently had an African American guest. We find that discrimination is \nconcentrated among hosts with no African American guests in their review history. \nWhen we restrict our analysis to hosts who have had an African American guest in \n1 We build on the large literature using audit studies to test for discrimination. Past research considers African \nAmericans and applicants with prison records in the labor market (Pager 2003), immigrants in the labor market \n(Oreopoulos 2011), Arabic job seekers (Carlsson and Rooth 2007), gender (Lahey 2008), long-term unemployment \n(Ghayad 2014), and going to a for-profit college (Deming et al. 2016), among many others. \n2 Some caution is warranted here. We only observe a gap between distinctively white and distinctively African \nAmerican names, which differ not only by suggested ethnicity but also potentially by socioeconomic status (Fryer \nand Levitt 2004). For ease of exposition, we describe our results in terms of differences among the “African \nAmerican guests” or the “white guests,” or use the term “race gap,” without also specifying that our results may \nbetter be described as a “race and socioeconomic status gap.” Section V discusses this issue in more detail. \n3 Doleac and Stein (2013) find a 62 percent to 56 percent gap in offer rates for online classified postings. \nBertrand and Mullainathan (2004) find a 10 percent to 6 percent gap in callback rates for jobs. Pope and Sydnor \n(2011) find a 9 percent to 6 percent gap in lending rates in an online lending market. Ayres, Vars, and Zakariya \n(2005) find a 20 percent to 13 percent gap in how often taxi drivers receive a tip. \n\n\nVol. 9 No. 2\b\n3\nEdelman et al.: Racial Discrimination in the Sharing Economy\nthe recent past, discrimination disappears—reinforcing the external validity of our \nmain results, and suggesting that discrimination is concentrated among a subset of \nhosts.\nTo explore the cost to a host of discriminating, we check whether each listing \nis ultimately rented for the weekend we inquired about. Combining that informa-\ntion with the price of each listing, we estimate that a host incurs a cost of roughly \n$65–$100 in foregone revenue by rejecting an African American guest.\nOverall, our results suggest a cause for concern. While discrimination has \nshrunk in more regulated offline markets, it arises and persists in online markets. \nGovernment agencies at both the federal and state level have routinely conducted \naudit studies to test for racial discrimination since 1955 in offline markets. One \nmight imagine implementing regular audits in online markets as well; indeed, online \naudits might be easier to run at scale due to improved data access and reduced \nimplementation cost.\nOur results also reflect the design choices that Airbnb and other online market-\nplaces use. It is not clear a priori how online markets will affect discrimination. \nTo the extent that online markets can be more anonymous than in-person trans-\nactions, there may actually be less room for discrimination. For example, Ayres \nand Siegelman (1995) find that African American car buyers pay a higher price \nthan white car buyers at dealerships, whereas Morton, Zettelmeyer, and Silva-Risso \n(2003) find no such racial difference in online purchases. Similarly, platforms such \nas Amazon, eBay, and Expedia offer little scope for discrimination, as sellers effec-\ntively ­\npre-commit to accept all buyers regardless of race or ethnicity. However, these \nadvantages are by no means guaranteed, and in fact they depend on design choices \nmade by each online platform. In this situation, Airbnb’s design choices enable \nwidespread discrimination.\nI.  About Airbnb\nAirbnb is a popular online marketplace for short-term rentals. Founded in 2008, \nthe site gained traction quickly and, as of November 2015, it offers 2,000,000 listings \nworldwide.4 This is more than 3 times as many as Marriott’s 535,000 rooms world-\nwide. Airbnb reports serving over 40 million guests in more than 190 countries.\nWhile the traditional hotel industry is dominated by hotels and inns that each \noffer many rooms, Airbnb enables anyone to post even a single room that is vacant \nonly occasionally. Hosts provide a wealth of information about each listing, includ-\ning the type of property (house, apartment, boat, or even castle, of which there are \nover 1,400 listed), the number of bedrooms and bathrooms, the price, and location. \nEach host also posts information about herself. An interested guest can see a host’s \nprofile picture as well as reviews from past guests. Airbnb encourages prospective \nguests to confirm availability by clicking a listing’s “Contact” button to write to the \nhost.5 In our field experiments (described in the next section), we use that method to \nevaluate a host’s receptiveness to a booking from a given guest.\n4 https://www.airbnb.com/about/about-us.\n5 See “How do I know if a listing is available,” https://www.airbnb.com/help/question/137. \n\n\n4\t\nAmerican Economic Journal: applied economics\b\napril 2017\nII.  Experimental Design\nA. Sample and Data Collection\nWe collected data on all properties offered on Airbnb in Baltimore, Dallas, Los \nAngeles, St. Louis, and Washington, DC as of July 2015. Our goal was to collect \ndata from the top 20 metropolitan areas from the 2010 census. We started with these \nfive cities because they had varying levels of Airbnb usage and came from diverse \ngeographic regions. Baltimore, Dallas, and St. Louis offer several hundred listings \neach, while Los Angeles and Washington, DC have several thousand. We stopped \ndata collection after these five cities because Airbnb became increasingly rapid in \nblocking our automated tools which logged into guest accounts and communicated \nwith hosts. (We considered taking steps to conceal our methods from Airbnb, but \nultimately declined to do so.)\nBecause some hosts offer multiple listings, we selected only one listing per host \nusing a random number generator. This helped to reduce the burden on any given \nhost, and it also prevented a single host from receiving multiple identical e-mails. \nEach host was contacted for no more than one transaction in our experiment.\nWe also collected data from each host’s profile page. This allowed us to analyze \nhost characteristics in exceptional detail. First, we saved the host’s profile image. We \nthen employed Mechanical Turk workers to assess each host image for race (white, \nAfrican American, Asian, Hispanic, multiracial, unknown), gender (male, female, \ntwo people of the same gender, two people of different genders, unknown), and age \n(young, middle-aged, old). We hired two Mechanical Turk workers to assess each \nimage, and if the workers disagreed on race or gender, we hired a third to settle the \ndispute. If all three workers disagreed (as happened, for example, for a host whose \nprofile picture was an image of a sea turtle), we manually coded the picture. We \ncoded race as “unknown” when the picture did not show a person. Through this \nprocedure, we roughly categorized hosts by race, gender, and age.\nProfile pages also revealed other variables of interest. We noted the number of \nproperties each host offers on Airbnb, anticipating that professional hosts with mul-\ntiple properties might discriminate less often than others. We retrieved the number \nof reviews the host has received, a rough measure of whether the host is an avid \nAirbnb user or a casual one. We further checked the guests who had previously \nreviewed each host. Airbnb posts the photo of each such guest, so we used Face++, \na ­\nface-detection API, to categorize past guests by race, gender, and age.6 This allows \nus to examine relationships between a host’s prior experience with African American \nguests and the host’s rejection of new African American requests.\nWe also collected information about each listing. We recorded the price of the \nlisting, the number of bedrooms and bathrooms, the cancellation policy, any clean-\ning fee, and the listing’s ratings from past guests. We also measured whether the \n6 In addition to detecting race, gender, and age, Face++ estimates its confidence for each trait. When Face++ \nwas unable to make a match or its confidence was below 95 out of 100, we used Mechanical Turk to categorize the \npast guest via the method described above. \n\n\nVol. 9 No. 2\b\n5\nEdelman et al.: Racial Discrimination in the Sharing Economy\nlisting offered an entire unit versus a room in a larger unit, yielding a proxy for how \nmuch the host interacts with the guest.\nEach listing included a longitude and latitude, which allowed us to link to census \ndemographic data to assess the relationship between neighborhood demographics \nand discrimination. After linking the latitude and longitude to a census tract, we \nused census data on the number of African American, Hispanic, Asian, and white \nindividuals. Table 1 presents summary statistics about the hosts and listings as well \nas balanced treatment tests.\nWe later checked each listing to see whether hosts were ultimately able to fill open-\nings. Our guests inquired about reservations eight weeks in advance. Thus, if a guest \nsent a message on August 1 about the weekend of September 25, we checked on \nFriday, September 24 to see whether the specified listing was still listed as available.\nB. Treatment Groups\nOur analysis used four main treatment groups based on the perceived race and \ngender of the test guest accounts. Hosts were contacted by guests with names that \nsignaled African American males, African American females, white males, and \nwhite females, drawn from Bertrand and Mullainathan (2004). The list was based \non the frequency of names from birth certificates of babies born between 1974 and \n1979 in Massachusetts. Distinctively white names are those that are most likely to \nbe white, conditional on the name, and similarly for distinctively African American \nnames. To validate the list, we conducted a survey in which we asked participants \nto quickly categorize each name as white or African American. With just three sec-\nonds permitted for a response, survey takers had little time to think beyond a gut \nresponse. The survey results, presented in Appendix Table 1, confirm that the names \ncontinue to signal race.7\n7 On a scale of 0 to 1, where 0 is African American, the white female names each had an average survey response \nof 0.90 or above, and the African American female names all had an average score of 0.10 or below. The male \nTable 1— Summary Statistics\nVariables\nMean\nSD\n25th\npercentile\n75th \npercentile\nObservations\nMean, white \naccounts\nMean,\nAfrican \nAmerican \naccounts\np-value\nHost is white\n0.63\n0.48\n0\n1\n6,392\n0.64\n0.63\n0.15\nHost is African American\n0.08\n0.27\n0\n0\n6,392\n0.08\n0.08\n0.97\nHost is female\n0.38\n0.48\n0\n1\n6,392\n0.38\n0.37\n0.44\nHost is male\n0.30\n0.46\n0\n1\n6,392\n0.3\n0.3\n0.90\nPrice ($)\n181.11\n1,280.23\n75\n175\n6,302\n166.43\n195.81\n0.36\nNumber of bedrooms\n3.18\n2.26\n2\n4\n6,242\n3.18\n3.18\n0.96\nNumber of bathrooms\n3.17\n2.26\n2\n4\n6,285\n3.17\n3.17\n0.93\nNumber of reviews\n30.87\n72.51\n2\n29\n6,390\n30.71\n31.03\n0.86\nHost has multiple listings\n0.16\n0.36\n0\n0\n6,392\n0.32\n0.33\n0.45\nHost has 1+ reviews from \n  African American guests\n0.29\n0.45\n0\n1\n6,390\n0.29\n0.28\n0.38\nAirbnb listings per  \n  census tract\n9.51\n9.28\n2\n14\n6,392\n9.49\n9.54\n0.85\nPercent population African\n  American (census tract)\n0.14\n0.2\n0.03\n0.14\n6,378\n0.14\n0.14\n0.92\n\n\n6\t\nAmerican Economic Journal: applied economics\b\napril 2017\nWe then created 20 Airbnb accounts, identical in all respects except for guest \nnames. Our names included ten that are distinctively African American and ten dis-\ntinctively white names, divided into five male and five female names within each \ngroup. To avoid the confounds that would result from pictures, we use only names; \nour Airbnb profiles include no picture of the putative guest. From these 20 guest \naccounts, we sent messages to prospective hosts. Each host was randomly assigned \none of our 20 guest accounts. Figure 1 presents a representative e-mail from one of \nour guests to an Airbnb host. The name and dates changed depending on the mes-\nsage sender and when the message was sent.8 In choosing the dates, we asked hosts \nabout a weekend that was approximately eight weeks distant from when the mes-\nsage was sent. We limited our search to those properties that were listed as available \nduring the weekend in question.\nnames showed slightly more variation but tell the same story: all the white male names scored 0.88 or above, and \nall the African American male names except for Jermaine Jones scored 0.10 or below. The Appendix presents the \nfull results of the survey. \n8 No more than 48 hours elapsed between our first contact to a host in a given city, and the completion of our \ncontacting hosts in that city. Furthermore, no hosts in our sample had listings in more than one of the five cities we \ntested. Hence, it is unlikely that a host contacted later on in the study would have learned about the experiment. \nFigure 1. Sample Treatment\n\n\nVol. 9 No. 2\b\n7\nEdelman et al.: Racial Discrimination in the Sharing Economy\nC. Experimental Procedure\nWe sent roughly 6,400 messages to hosts between July 7, 2015 and July 30, 2015.9 \nEach message inquired about availability during a specific weekend in September. \nWhen a host replied to a guest, we replied to the host with a personal message clar-\nifying that we (as the guest) were still not sure if we would visit the city or if we \nwould need a place to stay. We sent this reply in order to reduce the likelihood of a \nhost holding inventory for one of our hypothetical guests.\nWe tracked host responses over the 30 days that followed each request. A research \nassistant then coded each response into categories. The majority of responses were \nin 1 of 6 groups: “No response” (if the host did not respond within 30 days); “No or \nlisting is unavailable;” “Yes;” “Request for more information” (if the host responded \nwith questions for the guest); “Yes, with questions” (if the host approved the stay \nbut also asked questions); “Check back later for definitive answer;” and “I will get \nback to you.” As these categories show, our initial categorizations used subtle dis-\ntinctions between possible responses. In our analyses below, however, we restrict \nour attention to the simplest response—“Yes”—though all of our results are robust \nto using “No” instead, as well as to ignoring nonresponses or to using broader defi-\nnitions of “Yes.”\nWe collected all data using scrapers we built for this purpose. We sent inquiries to \nAirbnb hosts using web browser automation tools we built for this purpose.\nIII.  Results\nTable 2 presents the main effect. We find that inquiries from guests with \n­\nwhite-sounding names are accepted roughly 50 percent of the time. In contrast, \nguests with African American-sounding names are accepted roughly 42 percent of \nthe time. Columns 2 and 3 introduce additional control variables related to the host \nor the property. The effect stays constant at a roughly 8 percentage point gap across \nthese specifications, controlling for the host’s gender, race, an indicator for whether \nthe host has multiple listings, an indicator for whether the property is shared, host \nexperience (whether the host has more than ten reviews), and the log of the listing \nprice.\nAs noted, we break down hosts’ responses into 11 categories. Figure 2 shows \nthe frequency of each response by race. One might worry that results are driven \nby differences in host responses that are hard to classify, such as conditional “Yes” \nresponses. Similarly, we would be concerned if our findings were driven by differ-\nences in response rate. African American accounts might be more likely to be catego-\nrized as spam, or hosts may believe that African American accounts are more likely \nto be fake, in which case one might expect higher nonresponse rates for African \n9 Our initial goal was to collect roughly 10,000 responses. This was based on a power analysis, which in turn \nused an effect size calculated from Edelman and Luca (2014). To find a similar effect size, we would need a sample \nsize of roughly 3,000 hosts. But, to calculate an effect among a subgroup of hosts, like African American hosts, \nwhich represent roughly 7 percent of the Airbnb population, we would need a sample size closer to 10,000. We fell \nshort of this goal for an exogenous reason: Airbnb shut down the experimental accounts after we collected roughly \n6,400 responses. \n\n\n8\t\nAmerican Economic Journal: applied economics\b\napril 2017\nAmerican accounts. But as Figure 2 shows, the discrimination results occur because \nof differences in simple “Yes” or “No” responses, not because of ­\nnonresponses or \nintermediate responses (like a conditional “Yes”).\nIn the rest of this section, we use the wealth of data available on Airbnb about the \nhost and location for each listing to look for factors that influence the gap between \nTable 2— The Impact of Race on Likelihood of Acceptance\nDependent variable: 1(host accepts)\nGuest is African American\n−0.08\n(0.02)\n−0.08\n(0.02)\n−0.09\n(0.02)\nHost is African American\n \n0.07\n(0.02)\n0.09\n(0.02)\nHost is male\n \n−0.05\n(0.01)\n−0.05\n(0.01)\nHost has multiple listings\n \n \n0.09\n(0.02)\nShared property\n \n \n−0.07\n(0.02)\nHost has 10+ reviews\n \n \n0.12\n(0.01)\nln(price)\n \n \n−0.06\n(0.01)\nConstant\n0.49\n(0.01)\n0.50\n(0.01)\n0.76\n(0.07)\nObservations\n6,235\n6,235\n6,168\nAdjusted R2\n0.006\n0.009\n0.040\nNotes: This table reports coefficients from a regression of a “Yes” response on the guest’s race and \nvarious host and location characteristics. Standard errors are clustered by (guest name) × (city) \nand are reported in parentheses.\nFigure 2. Host Responses by Race\n1,200\n900\n600\n300\n0\nYes\nConditional yes\nNo response\nConditional no\nNo\nGuest is African American\nGuest is white\n\n\nVol. 9 No. 2\b\n9\nEdelman et al.: Racial Discrimination in the Sharing Economy\nwhite and African American names. Does the identity of the host matter? Does the \nlocation of the property matter? Generally, we find that the discrimination is remark-\nably robust.\nA. Effects by Host Characteristics\nWe first check whether our finding changes based on the identity of the host. If \ndiscrimination is driven by homophily (in-group bias), then the host’s race should \nmatter. According to this theory, hosts might simply prefer guests of the same race. \nIf homophily were the primary factor driving differential guest acceptance rates, \nthen African American guests would face higher acceptance rates from African \nAmerican hosts. Table 3 presents regressions that include guest race, host race, and \nan interaction term. Across the entire sample of hosts, the interaction between the \nrace and guest of the host is not significantly different from zero, but the point esti-\nmate is noisy. This result masks heterogeneity across genders. Columns 2 and 3 of \nTable 3 report the same regression limited to male hosts and female hosts, respec-\ntively. Among male hosts, the interaction between the host’s race and guest’s race \nshows a widening of the race gap by 11 percentage points, whereas among females, \nthe race gap narrows by 11 percentage points. Both estimates are noisy; we cannot \nreject coefficients of zero.10\n10 Table 4 explores the effect of the host’s race with more nuance. It shows the proportion of “Yes” responses \nfrom each gender/race cell among hosts in response to each gender/race cell among guests. African American male \nhosts discriminate against African American male and female guests. White hosts of both genders are more likely \nto accept white guests of either gender. African American female hosts are the only exception: they accept African \nAmerican female guests more than any other group. Thus, with the exception of African American females, the data \nTable 3—Race Gap by Race of the Host \nDependent variable: 1(host accepts)\nAll hosts\nMale hosts\nFemale hosts\nOther hosts\nGuest is African American\n−0.08\n(0.02)\n−0.09\n(0.02)\n−0.09\n(0.02)\n−0.07\n(0.03)\nHost is African American\n0.06\n(0.03)\n0.19\n(0.05)\n−0.00\n(0.04)\n0.03\n(0.09)\nHost is African American × guest is\n  African American\n0.01\n(0.05)\n−0.11\n(0.08)\n0.11\n(0.06)\n−0.06\n(0.14)\nConstant\n0.48\n(0.01)\n0.44\n(0.02)\n0.50\n(0.02)\n0.50\n(0.02)\nObservations\n6,235\n1,854\n2,336\n2,045\nAdjusted R2\n0.007\n0.015\n0.007\n0.003\nImplied coefficient on guest is African American + host\n  is African American × guest is African American\n−0.07\n(0.05)\n−0.19\n(0.08)\n0.02\n(0.06)\n−0.12\n(0.14)\nNotes: This table reports coefficients from a regression of a “Yes” response on the guest’s race, the host’s race, and \nthe interaction between the two. Other hosts are hosts we could not classify as male or female. Of the 2,045 host pic-\ntures we could not classify for gender, 972 had a picture of a mixed-gender couple, 259 had a same-gender couple, \n603 had a picture without a human in it, and the rest could not be classified. Standard errors are clustered by (guest \nname) × (city) and are reported in parentheses.\n\n\n10\t\nAmerican Economic Journal: applied economics\b\napril 2017\nDiscrimination may also be influenced by a host’s proximity to the guest. For \nexample, Becker (1957) formalizes racial discrimination as distaste for interactions \nwith individuals of a certain race. On Airbnb, a host must classify each listing as offer-\ning an entire unit, a room within a unit, or a shared room. We classify anything other \nthan an entire unit as a “shared property.” Column 1 of Table 5 shows that the race gap \nis roughly the same whether or not a property is shared. (In unreported results, we find \nthat the race gap stays roughly the same in shared properties with only one bathroom.)\nOne might expect a distinction between casual Airbnb hosts who occasion-\nally rent out their homes, versus professional hosts who offer multiple properties. \nRoughly a sixth of Airbnb hosts manage multiple properties, and roughly 40 percent \nof hosts have at least ten reviews from past guests. Columns 2 and 3 explore the \nextent of discrimination among hosts with multiple locations, and those with more \nthan ten reviews. Across these specifications, the race gap persists with roughly the \nsame magnitude.11\nTo the extent that discrimination rates are changing over time, one might expect \ndiscrimination to be less common among younger hosts. To assess this possibility, \nwe employed Mechanical Turk workers to categorize hosts as young, middle-aged, \nor old. Column 4 shows that discrimination also persists across the age categories \nwith roughly the same magnitude.\nB. Effects by Listing Characteristics\nJust as discrimination was robust across host characteristics, we find that dis-\ncrimination does not vary based on the cost or location of the property. Column 1 of \nTable 6 shows that, overall, listings above the median price are more likely to reject \nis inconsistent with homophily. Table 4 focuses on race/gender subgroups, but we present a more systematic break-\ndown of the raw results in Appendix Table 2. We ultimately focused on race/gender cells for ease of presentation. \n11 Hosts with at least ten reviews still have a race gap, but the acceptance rates for both races are higher among \nthese hosts. Instead of the 50 percent to 42 percent gap we see among all hosts, the race gap among hosts with at \nleast 10 reviews, or hosts with multiple properties, is closer to 60 percent to 52 percent. Hence, the racial gap is the \nsame in terms of percentage points, but not in terms of percent. The same is true in a later specification, where we \nlook at the race gap among hosts with at least one review from an African American guest. In all these specifica-\ntions, the change in the odds ratio is not economically significant. We have insufficient statistical power to reject the \npossibility that the odds ratios remain constant while the gap changes slightly. \nTable 4—Proportion of Positive Responses by Race and Gender\nGuest race/gender\nHost race/gender\nWhite \n \nmale\nAfrican \nAmerican \n \nmale\nWhite \n \nfemale\nAfrican \nAmerican \n \nfemale\nWhite male\n0.42\n0.35\n0.49\n0.32\nAfrican American male\n0.64\n0.40\n0.59\n0.43\nWhite female\n0.46\n0.35\n0.49\n0.44\nAfrican American female\n0.43\n0.38\n0.53\n0.59\nNote: This table shows the proportion of “Yes” responses by hosts of a certain race/gender to \nguests of a certain race/gender.\n\n\nVol. 9 No. 2\b\n11\nEdelman et al.: Racial Discrimination in the Sharing Economy\ninquiries. However, discrimination remains both among more expensive and less \nexpensive listings.\nWe can also check whether the listing was eventually filled (for the nights in \nquestion) to create a proxy for the desirability of the listing. First, we fit a Probit \nmodel to predict the likelihood that the listing was filled, controlling for a fixed city \neffect and a host of covariates.12 Then we assign each listing a probability of being \nfilled. This lets us test whether discrimination changes based on the listing’s desir-\nability.13 It does not.\nWe also hypothesized that the extent of discrimination might vary with the diver-\nsity of a neighborhood. More generally, one might expect that geography matters \nand that discrimination is worse in some areas than others, due to market structure \n12 The covariates are as follows: the host’s race and gender, the price, number of bedrooms, whether the property \nis shared, whether the bathroom is shared, the number of reviews, the age of the host, whether the host operates mul-\ntiple listings, the proportion of white people in the census tract, and the number of Airbnb listings in the census tract. \n13 We thank an anonymous reviewer for suggesting this approach. \nTable 5—Are Effects Driven by Host Characteristics?\nDependent variable: 1(host accepts)\nGuest is African American\n−0.07\n(0.02)\n−0.08\n(0.02)\n−0.09\n(0.02)\n−0.11\n(0.02)\n−0.09\n(0.02)\nShared property\n0.00\n(0.01)\nShared property × guest is African American\n−0.02\n(0.03)\nHost has multiple listings\n0.14\n(0.02)\nHost has multiple listings × guest is African American\n−0.01\n(0.03)\nHost has ten+ reviews\n0.14\n(0.02)\nHost has ten+ reviews × guest is African American\n0.01\n(0.02)\nHost looks young\n−0.03\n(0.02)\nHost looks young × guest is African American\n−0.01\n(0.02)\nHost has 1+ reviews from an African American guest\n0.10\n(0.01)\nHost has 1+ reviews from an African American guest\n  × guest is African American\n0.06\n(0.02)\nConstant\n0.49\n(0.01)\n0.46\n(0.01)\n0.42\n(0.01)\n0.50\n(0.01)\n0.46\n(0.01)\nObservations\n6,235\n6,235\n6,235\n6,235\n6,235\nAdjusted R2\n0.006\n0.014\n0.027\n0.011\n0.019\nImplied coefficient on guest is African American\n  + host trait × guest is African American\n−0.09\n(0.02)\n−0.09\n(0.03)\n−0.08\n(0.02)\n−0.08\n(0.03)\n−0.04\n(0.03)\nNotes: This table reports coefficients from a regression of a “Yes” response on the guest’s race, various host char-\nacteristics, and the interaction between the two. Standard errors are clustered by (guest name) × (city) and are \nreported in parentheses.\n\n\n12\t\nAmerican Economic Journal: applied economics\b\napril 2017\nor underlying rates of discrimination among a population. Merging data on neigh-\nborhoods by census tract, column 2 shows that the extent of discrimination does not \nvary with the proportion of nearby residents who are African American. Column 3 \nshows that discrimination is ubiquitous: it does not vary with the number of Airbnb \nlistings within the census tract. We also find discrimination in all cities in our sam-\nple, as shown in Appendix Table 3.\nC. Robustness—Effects by Name\nTable 7 shows the proportion of positive responses broken down by name. The \neffect is robust across choice of names. For example, the African American female \nname with the most positive responses (Tamika) received fewer positive responses \nthan the white female name with the fewest positive responses (Kristen), though this \ndifference is not statistically significant. Similarly, the African American males with \nthe most positive responses (Darnell and Rasheed) received fewer acceptances than \nthe white male with the fewest positive responses (Brad).\nD. Comparing Experimental Results with Observational Patterns\nEach listing page includes reviews from previous guests, along with profile pic-\ntures for these guests. This allows us to see which hosts previously accepted African \nTable 6— Are Effects Driven by Location Characteristics?\n \nDependent variable = 1(host accepts)\nGuest is African American\n−0.09\n(0.02)\n−0.08\n(0.02)\n−0.09\n(0.02)\n−0.12\n(0.06)\nPrice > median\n−0.07\n(0.02)\n \nGuest is African American × (price > median)\n0.01\n(0.03)\n \nShare of African American population in census tract\n0.05\n(0.05)\n \nGuest is African American × (share of African American\n  population in census tract)\n0.02\n(0.08)\n \nAirbnb listings per census tract\n−0.0007\n(0.0009)\n \nGuest is African American × (Airbnb listings per census tract)\n0.0008\n(0.001)\n \nProbability listing is filled 8 weeks later\n0.56\n(0.08)\nGuest is African American × (probability listing is filled \n  eight weeks later)\n0.09\n(0.12)\nConstant\n0.52\n(0.02)\n0.48\n(0.01)\n0.49\n(0.02)\n0.24\n(0.03)\nObservations\n6,235\n6,223\n6,235\n6,101\nAdjusted R2\n0.01\n0.006\n0.006\n0.030\nNotes: This table reports coefficients from a regression of a “Yes” response on the guest’s race, various location \ncharacteristics, and the interaction between the two. Standard errors are clustered by (guest name) × (city) and are \nreported in parentheses.\n\n\nVol. 9 No. 2\b\n13\nEdelman et al.: Racial Discrimination in the Sharing Economy\nAmerican guests (although not all guests leave reviews and not all guests have pho-\ntos that reveal their race). We use this data to assess the external validity of our \nresults.\nWe collected profile pictures from the ten most recent reviews on each listing \npage. We categorized these past guests by race and gender, finding that 29 percent \nof hosts in our sample had at least one review from an African American guest. We \nthen regressed the likelihood of a host responding positively to our inquiry on the \nrace of the guest, whether the host has at least one recent review from an African \nAmerican guest, and an interaction between these variables. Column 5 of Table 5 \nreports the results. We find that the race gap drops sharply among hosts with at least \none recent review from an African American guest. We cannot reject zero ­\ndifference \nfor requests from our African American test accounts versus requests from our white \ntest accounts, though this result is only significant at the 10 percent level.14\nThis finding reinforces our interpretation of our main effects, including the role \nof race and the interpretation that observed differences reflect racial discrimina-\ntion by Airbnb hosts. Put another way, if our findings are driven by a quirk of our \n14 These findings are robust to alternative specifications of a host’s past guests. The same substantive results \nhold if we look at the raw number of reviews from African Americans, rather than whether there is at least one such \nreview. The same is true if we use the proportion of reviews from African American guests. \nTable 7— Proportion of Positive Responses, by Name\nEntire sample\n0.43 \n(6,390)\nWhite female\nAfrican American female\nAllison Sullivan\n0.49 \n(306)\nLakisha Jones\n0.42 \n(324)\nAnne Murphy\n0.56 \n(344)\nLatonya Robinson\n0.35 \n(331)\nKristen Sullivan\n0.48 \n(325)\nLatoya Williams\n0.43 \n(327)\nLaurie Ryan\n0.50 \n(327)\nTamika Williams\n0.47 \n(339)\nMeredith O’Brien\n0.49 \n(303)\nTanisha Jackson\n0.40 \n (309)\nWhite male\nAfrican American male\nBrad Walsh\n0.41 \n(317)\nDarnell Jackson\n0.38 \n(285)\nBrent Baker\n0.48 \n(332)\nJamal Jones\n0.33 \n(328)\nBrett Walsh\n0.44 \n(279)\nJermaine Jones\n0.36 \n(300)\nGreg O’Brien\n0.45 \n(312)\nRasheed Jackson\n0.38 \n(313)\nTodd McCarthy\n0.43 \n(314)\nTyrone Robinson\n0.36 \n(254)\nNotes: The table reports the proportion of “Yes” responses by name. The number of messages \nsent by each guest name is shown in parentheses. \n\n\n14\t\nAmerican Economic Journal: applied economics\b\napril 2017\n­\nexperimental design, rather than race, then it is difficult to explain why the race \ngap disappears precisely among hosts with a history of accepting African American \nguests.\nE. Importance of Profile Pictures and More Complete Profiles\nA related concern is that we used guest profiles that were relatively bare. A host \nmay hesitate to accept a guest without a profile picture or past reviews. Of course, \nthis alone cannot explain the race gap, since both white and African American guests \nhad bare profiles. But it does raise the question of whether more complete profiles \ncould mitigate discrimination.15\nInternal data from Airbnb and observational data on Airbnb users both suggest that \nprofile pictures alone are unlikely to make much difference. With access to internal \nAirbnb data, Fradkin (2015) looks at roughly 17,000 requests sent to hosts and finds \nthat guests are rejected 49 percent of the time. Notably, these requests from ordinary \nAirbnb users, with typical Airbnb profiles, were rejected at a rate similar to that of \nour guests. In our experiment, as detailed in Appendix Table 4, 44 percent of guests \nwere rejected or received no response. Another 11 percent received a message from \na host requesting more information. The remaining 46 percent were accepted. The \nsimilarity in rejection rates suggests that incompleteness of our guests’ profiles is \nnot likely to be causing a change in the rejection rate, and reinforces the ecological \nvalidity of our experimental design.\nOther methods indicate that profile pictures seem to have little impact on accep-\ntance decisions. In a logistic regression estimating the probability of receiving a \nrejection from a host, again using internal Airbnb data, Fradkin (2015) finds that \nincluding a profile picture has no significant effect. This matches the observational \ndata we collect: in a random selection of Airbnb users, we found that only 44 per-\ncent have a profile picture. The proportion of guests with a profile picture is higher \namong users who have left a review, but nonetheless both analyses indicate that the \nexistence of profile pictures plays a small role in host decision-making. Further, \neven if profile pictures impact rejection rates, it is not clear that the impact should \nbe differential by race. For example, one might expect that pictures would make a \nguest’s race more salient. If our results are driven by race, then our findings would \nbe a lower bound on the true effect.\nOne limitation of our experiment is that we do not observe the effect of past \nreviews on discrimination. If our findings are driven by statistical discrimination, \npositive reviews from previous hosts may reduce the extent of discrimination. \nHowever, three factors suggest that reviews are an incomplete response to a discrim-\nination problem. First, our acceptance rates are similar to overall acceptance rates \n15 Similarly, our experiment does not assess whether discrimination occurs because of race or social class. \nHanson and Hawley (2011) find, in a field experiment on Craigslist’s housing market using similar methodology, \nthat renters with African American names face a penalty, but that the penalty decreases if the e-mail sent to a \nlandlord signals higher social class. Under some specifications, African Americans face a statistically significant \npenalty based on race and an additional penalty for signaling low class, also statistically significant. Under other \nspecifications, the racial gap is not statistically significant when comparing white and African American guests who \nboth signal high social class. On the whole, the paper indicates that social class and race both play a role. \n\n\nVol. 9 No. 2\b\n15\nEdelman et al.: Racial Discrimination in the Sharing Economy\non Airbnb (Fradkin 2015), which indicates that hosts are not treating our test guest \naccounts differently for lack of reviews, meaning that reviews would be unlikely \nto eliminate discrimination. Indeed, for reviews to eliminate discrimination, they \nwould need to provide a 16 percent differential increase in acceptance rates for \nAfrican Americans, relative to white guests. Second, all Airbnb users necessarily \nstart without past reviews, so a review system would not address any initial barri-\ners to entry that guests face. Third, a subjective review system can itself allow or \nfacilitate discrimination. (See, e.g., Goldin and Rouse 2000, finding that visually \nconfirming a musician’s gender may influence an expert’s judgment of her work.) \nWhatever mechanism is causing a lower acceptance rate for the African American \nguests may also cause a worse rating.\nF. How Much Does Discrimination Cost Hosts?\nA host incurs a cost for discriminating when rejecting a guest causes a unit to \nremain empty. The expected cost depends on the likelihood of the property remain-\ning vacant, which in turn depends on the thickness of the market. If a host can easily \nfind a replacement guest, then discrimination is nearly costless for the host. But if a \nproperty remains vacant after the host rejects a guest, then discrimination imposes a \nmore significant cost. In other words, the impact on net revenue from discriminating \ndepends on the likelihood of filling a unit with someone of the host’s preferred race \nafter rejecting a guest of a disfavored race.\nBecause we collect data about each property’s availability after a host declines a \nguest, we can estimate the cost in net revenue from discrimination. Suppose a host \ncharges price p for a listing and pays listing fees f to Airbnb. Let ​\nπ​\nreplace​\n be the prob-\nability of filling the property after rejecting a guest in our study. Then the cost in net \nrevenue of discrimination is as follows:\n\t\nΔNet Revenue = ( \np − f  \n \n) −  ​\nπ​\nreplace​\n ​\n·​\n ( \np − f  \n \n) = (1 −  ​\nπ​\nreplace​\n) · ( \np − f  \n \n).\nThat is, the cost of discrimination, in terms of net revenue, is the revenue that the \nhost forgoes if the listing remains empty multiplied by the probability that the listing \nremains empty.\nIn our data, hosts who rejected or never responded to our inquiries had properties \nwith a median price of $163 and a mean price of $295.16 The numbers are similar \nand slightly higher if we restrict the sample further to those hosts who rejected \nAfrican American guests, or if we expand the sample to hosts who responded “Yes” \nto our accounts.17 Airbnb charges each host a fee equal to 3 percent of the listing \nprice.\n16 In calculating price, we sum the listing price and any cleaning fee. \n17 An anonymous reviewer correctly points out that the host we are interested in is the host on the margin of \ndiscriminating. But there are hosts far from this margin both within the group of hosts who said yes and within the \ngroup of hosts who said no. Nonetheless, our calculations in this section are not sensitive to which group of hosts \nwe include. When including hosts who said yes, the median price drops from $163 to $150, and the probability of \nfinding a replacement guest rises to 64 percent instead of 59.4 percent (excluding disappearing hosts) or 45 percent \ninstead of 37.9 percent (including disappearing hosts). Thus, the cost of discrimination drops by about $10 or $20 \n\n\n16\t\nAmerican Economic Journal: applied economics\b\napril 2017\nAfter our inquiries, roughly 25.9 percent of the listings in our study remained \nvacant on the dates we requested after rejecting or not responding to one of our \nguests. Another 37.9 percent remained listed but were no longer available on those \ndates, suggesting that the host either found another guest or decided to no longer \nmake the property available on the specified dates. The remaining 36.1 percent \nof properties were no longer listed on Airbnb. Because it is unclear whether the \nhosts who exit should be excluded from the sample or treated as not having found a \nreplacement, we develop two estimates.\nIf we exclude these disappearing hosts from our calculation, 59.4 percent of hosts \nfound a replacement guest. Setting p equal to the median price ($163) and fees at \n3 percent of the median price:\n\t\nΔNet Revenue = (1 − 0.594) · ($163 − 0.03 · $163) ≈ $64.19.\nIf we treat disappearing listings as vacancies, in effect assuming that the host of \na dropped listing was not able to find a replacement guest, then only 37.9 percent of \nhosts found a replacement guest. The cost of discrimination rises as a result:\n\t\nΔNet Revenue = (1 − 0.379) · ($163 − 0.03 · $163) ≈ $98.19.\nIn this analysis, we focus on the net revenue, which does not incorporate the \nmarginal cost of each night the listing is rented, since we do not directly observe \ncosts. The cost of hosting includes various types of host effort or wear-and-tear to \nthe property. In principle, hosting also entails a risk of damage by a guest, though \nthroughout the relevant period Airbnb automatically provided all hosts with prop-\nerty insurance, which reduces the risk. Our calculation also excludes unobserved \nbenefits of hosting, such as the possibility that a positive review draws more guests \nin the future and improves the listing position on Airbnb. A full estimate of profit \nwould also need to consider the time cost of looking for new guests after rejecting \nsomeone on the basis of race.18\nWhile these estimates are clearly noisy, they suggest that hosts incur a real cost \nby discriminating. The median host who rejects a guest because of race is turning \ndown between $65 and $100 of revenue.\nIV.  Discussion\nOnline platforms such as Airbnb create new markets by eliminating search fric-\ntions, building trust, and facilitating transactions (Lewis 2011, Luca 2016). With \nthe rise of the sharing economy, however, comes a level of discrimination that \namong hosts who say yes, and therefore either did not discriminate against the African American accounts or did \nnot get a chance to do so. \n18 Our calculation also ignores other factors that cut in both directions. Responding with a “Yes” to a guest does \nnot provide 100 percent certainty of a paid booking; the guest may choose another option or may not make the trip. \nIn that case, our estimates overstate the revenue loss. Similarly, we have imperfect information about whether a \nhost found a replacement guest. Among other complexities, our guests requested two-night stays; we treat a host as \nhaving filled a listing if the host found a replacement guest for at least one of the nights, though a host who filled \nonly one of the nights has nonetheless lost one night of revenue. \n\n\nVol. 9 No. 2\b\n17\nEdelman et al.: Racial Discrimination in the Sharing Economy\nis ­\nimpossible in the online hotel reservations process. Clearly, the manager of a \nHoliday Inn cannot examine names of potential guests and reject them based on race \nor socioeconomic status, or some combination of the two. Yet, this is commonplace \non Airbnb, which now accounts for a growing share of the short-term rental market.\nOur results contribute to a small but growing body of literature suggesting that \ndiscrimination persists—and we argue may even be exacerbated—in online plat-\nforms. Edelman and Luca (2014) show that African American hosts on Airbnb seek \nand receive lower prices than white hosts, controlling for the observable attributes of \neach listing. Pope and Sydnor (2011) find that loan listings with pictures of African \nAmericans on Prosper.com are less likely to be funded than similar listings with pic-\ntures of white borrowers. Doleac and Stein (2013) show that buyers are less likely \nto respond to Craigslist listings showing an iPod held by a Black hand compared to \nan identical ad with a white hand. In contrast, Morton, Zettelmeyer, and Silva-Risso \n(2003) find no difference by race in price paid for cars in online purchases—a sharp \ncontrast to traditional channels (see, e.g., List 2004; Zhao, Ondrich, and Yinger \n2005).\nOne important limitation of our experiment is that we cannot identify the mecha-\nnism causing worse outcomes for guests with distinctively African American names. \nPrior research shows that distinctively African American names are correlated with \nlower socioeconomic status (Fryer and Levitt 2004). Our findings cannot identify \nwhether the discrimination is based on race, socioeconomic status, or a combination \nof these two. That said, we note that discrimination disappears among hosts who \nhave previously accepted African American guests. One might worry that discrimi-\nnation against our test guest accounts results from our choice of names and, hence, \ndoes not represent patterns that affect genuine Airbnb guests. However, we find that \ndiscrimination is limited to hosts who have never had an African American guest, \nwhich suggests that our results are consistent with any broader underlying patterns \nof discrimination.\nSimilarly, our experiment does not provide a sharp test of alternative models \nof discrimination. The theoretical literature on discrimination often distinguishes \nbetween statistical and taste-based discrimination. While our experimental design \ncannot reject either mechanism, our findings suggest a more nuanced story than \neither of the classic models. For one, we find homophily among African American \nfemales, but not among other race/gender combinations. Furthermore, we find that \ndiscrimination is not sensitive to a measure of proximity between the host and guest. \nBoth findings are in tension with pure taste-based discrimination. But we also find \nsome evidence against pure statistical discrimination. As noted above, we find that \nhosts who have had an African American guest in the past exhibit less ­\ndiscrimination \nthan other hosts. This suggests that, at the very least, hosts are using different statis-\ntical models as they evaluate potential guests.\nA. Designing a Discrimination-Free Marketplace\nBecause online platforms choose which information is available to parties during \na transaction, they can prevent the transmission of information that is irrelevant \nor potentially pernicious. Our results highlight a platform’s role in ­\npreventing \n\n\n18\t\nAmerican Economic Journal: applied economics\b\napril 2017\n­\ndiscrimination or facilitating discrimination, as the case may be. If a platform \naspires to provide a discrimination-free environment, its rules must be designed \naccordingly.\nAirbnb has several options to reduce discrimination. For example, it could con-\nceal guest names, just as it already prevents transmission of e-mail addresses and \nphone numbers, so that guests and hosts cannot circumvent Airbnb’s platform and \nits fees. Communications on eBay’s platform have long used pseudonyms and auto-\nmatic salutations, so Airbnb could easily implement that approach.\nAlternatively, Airbnb might further expand its “Instant Book” option, in which \nhosts accept guests without screening them first. Closer to traditional hotels and bed \nand breakfasts, this system would eliminate the opportunity for discrimination. This \nchange also offers convenience benefits for guests, who can count on their booking \nbeing confirmed more quickly and with fewer steps. However, in our sample, only a \nsmall subset of hosts currently allow instant booking. Airbnb could push to expand \nthe use of this feature, which would also serve the company’s broader goal of reduc-\ning search frictions.\nMore generally, our results suggest an important tradeoff for market designers, \nwho set the rules of online platforms, including the pricing mechanisms (Einav et al. \n2013) and the information that is available and actionable at the time of transaction \n(Luca 2016). Market design principles have generally focused on increasing the \ninformation flow within a platform (Bolton et al. 2013, Che and HÖrner 2014, Dai \net al. 2014, Fradkin et al. 2014), but we highlight a situation in which platforms may \nbe providing too much information.\nB. Policy Implications\nBecause the legal system grants considerable protection to online marketplaces, \nAirbnb is unlikely to be held liable for allowing discrimination on its platform. \nWithin the United States, the Civil Rights Act of 1964 prohibits discrimination in \nhotels (and other public accommodations) based on race, color, religion, or national \norigin. But these laws appear to be a poor fit for the informal sharing economy, \nwhere private citizens rent out a room in their home (Belzer and Leong forthcoming; \nTodisco 2015). As discussed in Edelman and Luca (2014), any changes by Airbnb \nwould likely be driven by ethical considerations or public pressure rather than law. \nIn contrast, offline rental markets and hotels have been subject to significant regula-\ntion (as well as audit studies to test for discrimination) for decades. This contributes \nto worry among policymakers that online short-term rental markets like Airbnb may \nbe displacing offline markets, which are more heavily regulated (Schatz, Feinstein, \nand Warren 2016). One clear policy implication is that regulators may want to audit \nAirbnb hosts using an approach based on our paper—much like longstanding efforts \nto reduce discrimination in offline rental markets.\nOne might have hoped that online markets would cure discrimination, and it \nseems a different design might indeed do so. Regrettably, our analysis indicates that \nat Airbnb, this is not yet the case.\n\n\nVol. 9 No. 2\b\n19\nEdelman et al.: Racial Discrimination in the Sharing Economy\nInvited Postscript: Airbnb Implements Market Design Changes\nPrior to this paper, Airbnb repeatedly ignored allegations of discrimination \non the platform (Finley 2016; Larson and Harris 2016). In response to our \nstudy and growing user complaints, the company put together a task force \nincluding former attorney general Eric Holder to propose a set of market \ndesign changes to reduce discrimination on the platform (Benner 2016). \nOn the same day this paper was accepted for publication in this journal, \nAirbnb announced the company’s planned changes. Changes include a \ngoal of increasing the proportion of hosts who offer Instant Book (letting \nguests book instantly, without the host first seeing the guest’s picture or \nname), a reminder to all users of the company’s ­\nanti-discrimination pol-\nicy, increased training for Airbnb staff to assist users who report discrim-\nination, and testing reduced prominence of guests’ photos. However, as of \nthe time of publication, Airbnb continued to reject suggestions to conceal \nguest photos and names before booking.\nAppendix\nTable A1—Results of Survey Testing Races Associated with Names \nWhite female\nAfrican American female\nMeredith O’Brien\n0.93\nTanisha Jackson\n0.03\nAnne Murphy\n0.95\nLakisha Jones\n0.05\nLaurie Ryan\n0.97\nLatoya Williams\n0.05\nAllison Sullivan\n0.98\nLatonya Robinson\n0.07\nKristen Sullivan\n1.00\nTamika Williams\n0.07\nWhite male\nAfrican American male\nGreg O‘Brien\n0.88\nTyrone Robinson\n0.00\nBrent Baker\n0.90\nRasheed Jackson\n0.06\nBrad Walsh\n0.91\nJamal Jones\n0.07\nBrett Walsh\n0.93\nDarnell Jackson\n0.10\nTodd McCarthy\n0.98\nJermaine Jones\n0.26\nNotes: “White” is coded as 1. “African American” is coded as 0. Sample size = 62. \nTable A2—Raw Discrimination across All Race and Gender Groups\nGuest race/gender\nHost race/gender\nWhite  \nmale\n(1)\nAfrican \nAmerican  \nmale\n(2)\nWhite  \nfemale\n(3)\nAfrican \nAmerican  \nfemale\n(4)\nMale\n(5)\nFemale\n(6)\np-value\n(7)\nWhite\n(8)\nAfrican \nAmerican\n(9)\nWhite male\n0.42\n0.35\n0.49\n0.32\n0.39\n0.4\n0.72\n0.45\n0.34\nAfrican American male\n0.64\n0.40\n0.59\n0.43\n0.52\n0.51\n0.99\n0.62\n0.42\nWhite female\n0.46\n0.35\n0.49\n0.44\n0.41\n0.46\n0.06\n0.48\n0.39\nAfrican American female\n0.43\n0.38\n0.53\n0.59\n0.41\n0.56\n0.02\n0.48\n0.50\nWhite\n0.45\n0.36\n0.50\n0.40\n0.41\n0.45\n0.02\n0.47\n0.38\nAfrican American\n0.49\n0.40\n0.58\n0.52\n0.45\n0.55\n0.02\n0.53\n0.46\nOther or uncertain\n0.45\n0.38\n0.51\n0.43\n0.41\n0.47\n0.03\n0.48\n0.40\nMale\n0.43\n0.36\n0.47\n0.35\n0.40\n0.41\n0.80\n0.45\n0.35\nFemale\n0.47\n0.34\n0.51\n0.45\n0.41\n0.48\n0.004\n0.49\n0.40\nOther or uncertain\n0.45\n0.41\n0.54\n0.45\n0.43\n0.50\n0.003\n0.50\n0.43\nNote: This table shows the proportion of “Yes” responses by hosts of a certain race/gender to guests of a certain \nrace/gender.\n\n\n20\t\nAmerican Economic Journal: applied economics\b\napril 2017\nReferences\nAyres, Ian, and Peter Siegelman. 1995. “Race and Gender Discrimination in Bargaining for a New \nCar.” American Economic Review 85 (3): 304–21.\nAyres, Ian, Fredrick E. Vars, and Nasser Zakariya. 2005. “To Insure Prejudice: Racial Disparities in \nTaxicab Tipping.” Yale Law Journal 114 (7): 1613–74.\nBecker, Gary S. 1957. The Economics of Discrimination. Chicago: University of Chicago Press.\nBelzer, Aaron, and Nancy Leong.\u0003\n Forthcoming. “The New Public Accommodations.” Georgetown Law \nJournal.\nBenner, Katie. 2016. “Airbnb Adopts Rules to Fight Discrimination by Its Hosts.” New York Times, \nSeptember 8, A1.\nBertrand, Marianne, and Sendhil Mullainathan. 2004. “Are Emily and Greg More Employable Than \nLakisha and Jamal? A Field Experiment on Labor Market Discrimination.” American Economic \nReview 94 (4): 991–1013.\nTable A3—Discrimination by City\nDependent variable: 1(host accepts)\nAll \ncities\nBaltimore\n(N = 347)\nDallas\n(N = 415)\nLos Angeles\n(N = 3,913)\nSt. Louis\n(N = 151)\nWashington, DC\n(N = 1,559)\nGuest is African American\n−0.08\n−0.07\n(0.02)\n−0.08\n(0.02)\n−0.10\n(0.02)\n−0.08\n(0.03)\n−0.08\n(0.02)\nCity\n—\n0.07\n(0.03)\n0.04\n(0.03)\n−0.00\n(0.03)\n0.02\n(0.05)\n−0.03\n(0.04)\nCity × guest is \n  African American\n—\n−0.12\n(0.05)\n−0.01\n(0.04)\n0.03\n(0.04)\n0.02\n(0.07)\n−0.01\n(0.05)\nConstant\n0.49\n0.48\n(0.01)\n0.49\n(0.01)\n0.49\n(0.02)\n0.49\n(0.01)\n0.50\n(0.01)\nObservations\n6,235\n6,235\n6,235\n6,235\n6,235\n6,235\nAdjusted R2\n0.006\n0.007\n0.006\n0.006\n0.006\n0.007\nImplied coefficient on guest is\n  African American + city\n  × guest is African American\n—\n−0.19\n(0.04)\n−0.09\n(0.04)\n−0.07\n(0.02)\n−0.06\n(0.06)\n−0.09\n(0.05)\nNotes: This table reports coefficients from a regression of a “Yes” response on the guest’s race, a city, and the inter-\naction of city and guest race. Standard errors are clustered by (guest name) × (city) and are reported in parentheses.\nTable A4—Host Responses to Guest Inquiries, by Race of the Guest\nWhite\nguests\nAfrican American\nguests\nYes\n1,152\n940\nYes, but request for more information\n375\n308\nYes, with lower price if booked now\n11\n10\nYes, if guest extends stay\n10\n15\nYes, but in a different property\n18\n8\nYes, at a higher price\n4\n0\nRequest for more information\n339\n323\nNot sure or check back later\n154\n175\nNo response\n429\n423\nNo unless more information is provided\n12\n15\nNo\n663\n873\nNotes: The table reports the frequency of each type of host response to a guest inquiry, by race of \nthe guest. Likelihood-ratio chi-squared = 68.61 ( \np < 0.01). Null hypothesis is that the columns \nwill have equal proportions for each type of response.\n\n\nVol. 9 No. 2\b\n21\nEdelman et al.: Racial Discrimination in the Sharing Economy\nBolton, Gary, Ben Greiner, and Axel Ockenfels. 2013. “Engineering Trust: Reciprocity in the Produc-\ntion of Reputation Information.” Management Science 59 (2): 265–85. \nCarlsson, Magnus, and Dan-Olof Rooth. 2007. “Evidence of ethnic discrimination in the Swedish \nlabor market using experimental data.” Labour Economics 14 (4): 716–29.\nChe, Yeon-Koo, and Johannes Hörner. 2014. “Optimal Design for Social Learning.” http://liberalarts.\nutexas.edu/_files/ms37643/Che-Horner03-04-14.pdf. \nDai, Weijia, Ginger Z. Jin, Jungmin Lee, and Michael Luca. 2014. “Optimal Aggregation of Consumer \nRatings: An Application to Yelp.com.” National Bureau of Economic Research (NBER) Working \nPaper 18567.\nDeming, David J., Noam Yuchtman, Amira Abulafi, Claudia Golding, and Lawrence F. Katz. 2016. \n“The Value of Postsecondary Credentials in the Labour Market: An Experimental Study.” American \nEconomic Review 106 (3): 778–806.\nDoleac, Jennifer L., and Luke C. D. Stein. 2013. “The Visible Hand: Race and Online Market Out-\ncomes.” Economic Journal 123 (572): F469–92.\nEdelman, Benjamin G., and Michael Luca. 2014. “Digital Discrimination: The Case of Airbnb.com.” \nHarvard Business School Working Paper 14-054.\nEdelman, Benjamin, Michael Luca, and Dan Svirsky. 2017. “Racial Discrimination in the Sharing \nEconomy: Evidence from a Field Experiment: Dataset.” American Economic Journal: Applied Eco-\nnomics. https://doi.org/10.1257/app.20160213.\nEinav, Liran, Chiara Farronato, Jonathan D. Levin, and Neel Sundaresan. 2013. “Sales Mechanisms \nin Online Markets: What Happened to Internet Auctions?” National Bureau of Economic Research \n(NBER) Working Paper 19021.\nFinley, Taryn. 2016. “These Airbnb Alternatives Want To Make Travel More Welcoming For Black \nPeople.” Huffington Post, August 18. http://www.huffingtonpost.com/entry/innclusive-noirbnb-\nairbnb-alternatives_us_5768462ae4b0853f8bf1c675.\nFradkin, Audrey. 2015. “Search Frictions and the Design of Online Marketplaces.” http://andreyfradkin.\ncom/assets/SearchFrictions.pdf.\nFradkin, Audrey, Elena Grewal, Dave Holtz, and Matthew Pearson. 2014. “Bias and Reciprocity in \nOnline Reviews: Evidence from Field Experiments on Airbnb.” Unpublished.\nFryer, Roland G., Jr., and Steven D. Levitt. 2004. “The Causes and Consequences of Distinctively \nBlack Names.” Quarterly Journal of Economics 119 (3): 767–805.\nGhayad, Rand. 2014. “The Jobless Trap.” http://www.lexissecuritiesmosaic.com/gateway/FEDRES/\nSPEECHES/ugd_576e9a_f6cf3b6661e44621ad26547112f66691.pdf.\nGoldin, Claudia, and Cecilia Rouse. 2000. “Orchestrating Impartiality: The Impact of ‘Blind’ Audi-\ntions on Female Musicians.” American Economic Review 90 (4): 715–41.\nHanson, Andrew, and Zackary Hawley. 2011. “Do landlords discriminate in the rental housing mar-\nket? Evidence from an internet field experiment in U.S. cities.” Journal of Urban Economics 70 \n(2–3): 99–114. \nLahey, Joanna N. 2008. “Age, Women, and Hiring: An Experimental Study.” Journal of Human \nResources 43 (1): 30–56.\nLarson, Erik, and Andrew M. Harris. 2016. “Airbnb Sued, Accused of Ignoring Hosts’ Race Discrim-\nination.” Bloomberg, May 18. http://www.bloomberg.com/news/articles/2016-05-18/airbnb-sued-\nover-host-s-alleged-discrimination-against-black-man.\nLewis, Gregory. 2011. “Asymmetric Information, Adverse Selection and Online Disclosure: The Case \nof eBay Motors.” American Economic Review 101 (4): 1535–46.\nList, John A. 2004. “The Nature and Extent of Discrimination in the Marketplace: Evidence from the \nField.” Quarterly Journal of Economics 119 (1): 49–89.\nLuca, Michael. \u0003\n2016. “User-Generated Content and Social Media.” In Handbook of Media Economics, \nVol. 1B, edited by Simon Anderson, Joel Waldfogel, and David Strömberg, 563–92. Amsterdam: \nNorth-Holland.\nMorton, Fiona Scott, Florian Zettelmeyer, and Jorge Silva-Risso. 2003. “Consumer Information and \nDiscrimination: Does the Internet Affect the Pricing of New Cars to Women and Minorities?” \nQuantitative Marketing and Economics 1 (1): 65–92.\nOndrich, Jan, Alex Stricker, and John Yinger. 1999. “Do Landlords Discriminate? The Incidence and \nCauses of Racial Discrimination in Rental Housing Markets.” Journal of Housing Economics 8 (3): \n185–204.\nOreopoulos, Philip. 2011. “Why Do Skilled Immigrants Struggle in the Labor Market? A Field Exper-\niment with Thirteen Thousand Resumes.” American Economic Journal: Economic Policy 3 (4): \n148–71.\n\n\n22\t\nAmerican Economic Journal: applied economics\b\napril 2017\nPager, Devah. 2003. “The Mark of a Criminal Record.” American Journal of Sociology 108 (5): 937–75.\nPope, Devon G., and Justin R. Sydnor. 2011. “What’s in a Picture?: Evidence of Discrimination from \nProsper.com.” Journal of Human Resources 46 (1): 53–92.\nSchatz, Brian, Dianne Feinstein, and Elizabeth Warren. 2016. “Letter to Edith Ramirez, Chairwoman \nof the Federal Trade Commission.” http://www.warren.senate.gov/files/documents/2016-7-13-\nletter-to-FTC.pdf.\nTodisco, Michael. 2015. “Share and Share Alike? Considering Racial Discrimination in the Nascent \nRoom-Sharing Economy.” Stanford Law Review Online 67: 121–29.\nU.S. Department of Housing and Urban Development. 2013. Housing Discrimination against Racial \nand Ethnic Minorities 2012. Office of Policy Development and Research. Washington, DC, June.\nYinger, John. 1998. “Evidence on Discrimination in Consumer Markets.” Journal of Economic Per-\nspectives 12 (2): 23–40.\nZhao, Bo, Jon Ondrich, and John Yinger. 2005. “Why Do Real Estate Brokers Continue to Discrimi-\nnate? Evidence from the 2000 Housing Discrimination Study.” Syracuse University Center for Pol-\nicy Research Paper 96.","difficulty":"hard","domain":"Single-Document QA","length":"short","question":"Regarding the experimental methods in this article, the following statement is correct:","sub_domain":"Academic"}

Source: https://huggingface.co/datasets/zai-org/LongBench-v2

initial import

Posting: /agents

GET /api/v1/write?intent=publish&task_id=a442dd8b-8803-5ba9-b73b-4ea8b7f8143f&body={url_encoded_text}&agent_name={optional_name}&nonce={optional_random_id}