Munich Personal RePEc Archive The Visible Host: Does Race guide Airbnb rental rates in San Francisco? Kakar, Venoo and Franco, Julisa and Voelz, Joel and Wu, Julia San Francisco State University 10 March 2016 Online at https://mpra.ub.uni-muenchen.de/78275/ MPRA Paper No. 78275, posted 14 May 2017 16:35 UTC The Visible Host: Does Race guide Airbnb rental rates in San Francisco? Venoo Kakar ∗ Joel Voelz † Julia Wu ‡ Julisa Franco § June 18, 2016 Abstract ThesurgeinPeertoPeere-commercehasincreasinglybeencharacterizedbychang- ingtheonlinemarketplacetoamorepersonalizedenvironmentforthebuyerandseller. This personalization involves revealing information on buyer reviews, pictures and bi- ographical information on the sellers to reduce the perceived “purchase risk” or to facilitate trust with the buyers. However, this personalization has generated possibili- ties for discrimination in the online marketplace. In this paper, we examine the effect of host information available online (race, gender and sexual orientation etc.) on price listings on Airbnb.com in San Francisco. We find that Hispanic hosts and Asian hosts, on average, have a 9.6% and 9.3% lower list price relative to their White counterparts, after controlling for neighborhood property values, user reviews and rental unit char- acteristics. We don’t find any significant impact of gender and sexual orientation on price listings. Overall, our findings corroborate the presence of racial discrimination in the online marketplace. Keywords: Airbnb, Discrimination, Race, Online marketplace JEL: D40, D47, J15, J71 ∗Department of Economics, San Francisco State University, E-mail: [email protected] †Department of Economics, San Francisco State University, E-mail: [email protected] ‡Department of Economics, San Francisco State University, E-mail: [email protected] §Department of Economics, San Francisco State University, E-mail: [email protected] 1 Introduction Internet commerce has grown dramatically over the past decade, moving from an inter- esting niche to a mainstream component of both business and consumer markets with 2013 figures totaling over $700 Billion of consumer purchases in the U.S. A growing sub-market is the area of Peer-to-Peer or P2P commerce. P2P e-commerce involves consumers or small craftspersons acting as sellers and buyers offering and buying everything from used goods (EBay/Craigslist) to new craft products (Etsy) to personal services (TaskRabbit) to rooms 1 for rent (Airbnb). Accompanying this growth and emergence of the P2P e-commerce mar- ket has been an evolution from the early anonymous arms-length transaction environment of the internet to a more personalized environment and purchase process where the goal is to make a personal connection between the buyer and seller. This personalization involves techniques such as buyer reviews, pictures and biographical information on the sellers to give potential buyers more information on the seller. These techniques attempt to reduce perceived purchase risk and create a personal social connection making the purchase from a stranger more palatable. However, as P2P commerce has become more personal, it becomes less anonymous and so opens the possibility of various forms of discrimination by both buyers and sellers. This can occur because the race and gender of participants are frequently revealed through photos and biographical information. So, while P2P e-commerce has opened opportunities for minorities to participate in a growing market it has also generated questions about and possibilities for discrimination similar to face-to-face markets. Buyers now have the information to bypass e-commerce sellers based on race or gender in a manner similar to bypassing a brick and mortar store. In this paper we address the question of whether there is evidence that information regarding the race of the Airbnb host affects the listing price of rooms in the San Francisco market. The assumption is that this could indicate potential price discrimination based on race against the hosts. Airbnb incorporates multiple rating techniques to help increase buyer confidence including reviews by previous guests and available social media links of the hosts. To personalize listings, they permit sellers offering room listings (the hosts) to provide both a picture and biographical/listing information This allows potential renters to identify both the race and sex of the host. A representative sample of Airbnb listings in San Francisco was analyzed and the race of each host identified visually through their posted picture. After controlling for common factors, we find evidence of statistically significant price differences for racial minorities. We find that, White hosts are able to charge 9.2% more than Asian hosts and 9.6% more than Hispanic hosts holding all other listing factors constant such as location, room amenities and ratings. These findings raise questions concerning the migration of racial discrimination to on-line markets, possible differing pricing strategies or business objectives of minorities, and policy issues regarding potential liability of companies such as Airbnb where business practices enable potential discrimination in on-line commerce. 1At the end of 2014, Airbnb had 925,000 listings and over 25 million customers, Todisco (2015). 0 1.1 Related Literature 1.1.1 Discrimination in Traditional Housing markets Discrimination against minorities in the rental and housing market is certainly not new. Turner(2013)foundevidenceofsignificantdiscriminationagainstminoritiesrelatedtorental offerings and availability. The study conducted more than 8,000 paired tests in a nationally representative sample of 28 metropolitan areas. As reported, Hispanic renters learned about 12.5 percent fewer available units and were shown 7.5 percent fewer units than Whites. Asians learned about 9.8 percent fewer available units and were shown 6.6 percent fewer units than Whites. Also, Bayer et al. (2012) analyzed panel data covering over two million repeat-sales housing transactions from four metropolitan areas and found that Black and Hispanic homebuyers pay premiums of about three percent on average across the four cities, differences that are not explained by variation in buyer income, wealth or access to credit. While we did not use the techniques of these early studies, they provide a backdrop against which our study and others evaluating the emerging internet commerce space are evaluated. 1.1.2 Discrimination in the emerging P2P Commerce Market The emergence of the internet and on-line commerce has created both opportunities and pitfalls. Early Internet commerce provided anonymity to both buyer and seller. This lack of personal information increased the arms length nature of the transaction and removed many opportunities to practice discrimination against a now unknown” buyer or seller. However, in recent years the trend in internet commerce - and especially in the rapidly growingP2Pmarkethasbeentoincreasethepersonalizationofbuyerandsellerinanattempt to reduce the perceived risk of dealing with individuals instead of a commercial concern. An earlier study, Pope and Sydnor (2011) examined the effect a loan applicant’s personal information, including a picture on their acceptance rate for a personal loan at the P2P lending site Prospero. The study found that Black applicants had a 2.4-3.2 percentage points lower chance of getting funded, other factors held constant. The study further compared this to the average probability of getting funded, 9.3%. Blacks had a 30% decrease in the likelihood of funding. Additionally, they found that the interest rates offered to Blacks were 60-80 basis points higher than Whites with similar credit profiles. This was an example of supply side discrimination. Another recent study, Doleac and Stein (2013) posted classified advertisements offering iPod Nano music players for sale on several hundred locally focused websites throughout the US. They observed the effect of the seller’s skin color on outcomes such as bid price and preference for a face-to-face delivery vs having the product mailed. They included a photograph of a dark -skinned (Black) or light-skinned (White) hand holding the item for sale and were able to vary the apparent race of the seller while fixing other sales and market characteristics. In addition to race they examined the effect of a social signal communicated by a tattoo on the seller’s hand. They used this option as a suspicious’ White control group. Controlling for all common factors it was found that Black visible hand seller ads had 13% fewer responses and 18% fewer offers. The offers made to Black-handed sellers were on average 11% lower than those made to White-handed sellers. Black-handed sellers also garnered lower trust as they were 17% less likely to have the bidder’s name included in their 1 initial offer or inquiry. This study shows the impact of visual information that identifies the race of the seller, although it cannot conclude whether the captured effect is due to taste-based discrimination. This is similar to our study’s examination of the use of host pictures on Airbnb listings. In this case, it was an example of demand side discrimination as potential buyers bid lower on products being sold by minorities or life style groups (such as those with tattoos) that are discriminated against. 1.1.3 Discrimination in the P2P Room Rental Market of Airbnb Two recent studies have focused on potential pricing effects of racial discrimination against minority hosts of Airbnb, the P2P commerce site for short term room rentals. The initial study by Edelman and Luca (2014) using the set of all Airbnb listings in New York City, identified the race of the hosts using their on-line listing photo. Controlling for all visi- ble information on Airbnb listings regarding the quality of the listing they found statistically significantly evidence that non-Black hosts charge approximately 12% more than Black hosts for an equivalent rental. They point out the possible unintended consequences of offering hosts the ability to post pictures for increased personalization on enabling discrimination. All 3752 New York listings were downloaded and the offered rental price as well as all other listing variables coded. Amazon Mechanical Turk was used to categorize a host’s race based on their on-line photo as well as individually rate each listing for quality on 1 2 seven-point scale. Using the Airbnb rating variables (cleanliness, location, communication, check-in, etc) as well as the host race and listing quality, they find evidence of differences in listing price based on race. Their findings showed a statistically significant difference between posted rent prices of Black hosts vs non-Black hosts. They also found that non-Black hosts earn roughly 12% more for a similar rental relative to Black hosts. Although Black hosts’ listings tended to be in less desirable locations and therefore could be expected on average to have lower prices, their study controlled for all attributes that are viewable by customers in making a rental decision. However, they were not able to determine the extent of taste-based vs statistical discrimination taking place. This study did not explicitly try to control for the quality of the property or the value of the neighborhood as a factor affecting the listing price and it focused only on Black/Non- Black hosts. Our study has tried to improve on this by using a normalized factor for property values in each neighborhood where a listing is located in addition to race, gender, couple and sexual orientation variables. A follow up study to Edelman and Luca (2014) focused on potential racial discrimination against Asian Airbnb hosts in the city of Oakland in the California market, Wang et al. (2015). Although a minority group, Asians have a very different social, economic, and educational profile than Blacks. They have the highest median income of all racial groups, as well as the highest average test scores for college admission. Despite polling data where Asians indicated they were more satisfied than the general public with life, finances and the country’s direction there are indicators of negative bias against Asians. The focus of their study was to answer the question of whether Asian minority group might experience covert racial discrimination on Airbnb. Using a sample of Airbnb listings 2Amazon Mechanical Turk is a service that provides workers to do large scale repetitive IT-related tasks 2 fromaraciallyandeconomicallybalancedneighborhoodtheyfollowedamethodologysimilar to the previous study in New York City. In addition to Airbnb listing variables, a host’s race was determined from inspection of the listing photo. However, only two alternatives were used in their dataset: Asian or White. Any hosts that were not either Asian or White (or could not be determined) were excluded. Additionally, the number of observations were very small, 101, and as a result they needed to manipulate the form of the independent variables in order to compensate for strong right- skewed distributions. Basing their listing price on a week’s stay vs a single night to reduce price variability, they found a statistically significant price difference of 20% less per week for Asian hosts vs White hosts with similar listings. As with the New York study, the authors cannot say that the cause of this price difference is discrimination or whether it is based on differing price strategies or goals for Asian vs White hosts. They also acknowledge that an extension of their study would be to control for the quality of the listing location, something that was attempted in our study of the San Francisco market. Another recent paper, also by Edelman et al. (2015) once again looked at the the Airbnb market for potential racial discrimination but this time focusing on the supply side. In contrast to their earlier study where they looked at discrimination aimed at Airbnb hosts, manifested in a lower listing price, presumably due to lack of demand at a price equivalent to non-minority listings - here they examined the rates of Airbnb host acceptance of a request from Black and White potential guests. Using 6400 Airbnb accounts in five different cities they sent guest requests for a room using identical data except for a distinctly White or Black name. Their results show that Blacks received a positive response 42% of the time comparedtoroughly50%forWhiteguests. Thistranslatesintoa16%differencebetweenthe two groups, consistent with the racial gap found in several other markets including taxicab rides, labor markets and on-line lending (earlier cited Prospero study). This recent work adds evidence to a growing structure and operation of P2P commerce markets. 2 Data Sources and Structure WefocusonthemarketforAirbnblistingsinSanFrancisco. Ourdatasetconsistsofcross sectional data that was available from Inside Airbnb, a data collection project independent of Airbnb that compiles information of Airbnb listings for public use. The raw dataset was comprised of approximately 6,000 listings as of September 2, 2015. Since we utilized several guest rating categories as explanatory variables, we applied adjustments to the complete dataset to increase the accuracy of our sampling process and regression analysis. First, we restricted the listings to hosts with a verified identity (per Airbnb processes) and profile picture - a picture being essential to determining the host’s gender and race. Second, to ensure that the user ratings of the Airbnb rentals and hosts were as reliable as possible, we only included listings with a minimum of 5 reviews (since we used several of the Airbnb review values as explanatory variables). These restrictions reduced the Inside Airbnb data set to 2,772 listings belonging to 2,161 unique hosts. Approximately 85 percent of hosts have only one listing, with a small number of hosts having 10 or more listings. This distribution of listings per host closely matched an earlier study by the San Francisco Chronicle using Airbnb data from May 2014, Said (2014). 3 Since we needed to individually verify the race and gender of the host for each listing, we created a smaller sample data set from the full Inside Airbnb dataset. Because of resource limitations, we drew a proportionally weighted sample of 800 observations from the Inside Airbnb dataset based on the number of listings per neighborhood based on zip code in the full dataset. Listings were randomly selected without replacement to create a sample with the same weighted neighborhood listing composition as the complete dataset. This was done to ensure that our sample reflects the distribution of Airbnb listings across the various San Francisco neighborhoods. For each listing, we manually viewed the Airbnb host’s online profile page and, based on the included picture of the host and biographical information, categorized the host’s gender, race, whether the host was a couple, and the host’s sexual orientation (gay or not), if it could be reasonably determined based on host’s biography. For interracial couple hosts where one partner was White, we categorized the listing under the race of the non-White host. Listings of interracial couples where both partners were non-White were uncommon and were removed from our sample. When the host’s characteristics were unidentifiable (e.g., the host’s photo did not include any people), the listing was removed from the sample. After this categorization process, we were left with 715 listings belonging to 588 unique hosts. The dummy variables female, Black, Asian, Hispanic, couple and Gay were created using the values we obtained from a direct observation of the host’s listing site. To incorporate the quality of the neighborhood as a factor influencing the listing price, both in terms of customer desirability and in terms of owner costs that could be reflected in pricing, we used data on the price per square foot (cost) based on recent property sales in each neighborhood. Property costs/sq ft were obtained from Trulia.com This was done for all property types in each Airbnb listing neighborhood. See Airbnb neighborhood list in Appendix. We used the value of all property types and not just single family homes or condominiums to reflect the fact that Airbnb listings can include single family homes, con- dominiums and commercial buildings. For each neighborhood in our sample, we calculated a z-score to show how many standard deviations each neighborhood was above or below the mean average cost per square foot. This z-score was then used as an explanatory variable in our model for each listing/neighborhood. The variables used in our paper are specified in Tables 1 and 2. The baseline race category was White; baseline gender category was male. Single was the baseline category for the couple dummy variable, and non-gay/not-specified for the gay dummy variable. For the dummy variables created from the Inside Airbnb data, Shared Room’ was the baseline room type (as opposed to Private Room or Whole Apartment) and non-superhost was the default for the Superhost variable. Using one of the possible race, gender, couple/single and gay categories as the base case or control eliminated the potential of perfect multicollinearity for those variables. 4 3 Model 3.1 Dependent Variable We used log of price per unit listing as our dependent variable. This enabled us to relate the effects of our independent variables to percentage changes in the listing price. This is significantly more revealing since a listing price will vary due to the number of rooms, bathrooms, etc and using log of price allows us to see the effect of the race variables on listing price in percentage terms, vs absolute terms. 3.2 Independent Variables We categorize our independent variables into four major categories, namely, 1. Host features: consisting of dichotomous variables on the host’s race (White, Hispanic, Black, other); gender (male, female); whether the host was a couple, and the host’s sexual orientation (gay or not). 2. Rental listing features: We expected the listing price to be directly affected by specific attributes of the rental which would objectively add or subtract to the perceived value of the unit. These are variables quantifying the number of bedrooms, bathrooms, whether rental is a whole apartment, has a private room and maximum number of guests accommodated. They all seemed likely to have a direct and positive impact on listing price. 3. User Reviews: The other Airbnb variables included in the model are categorical involv- ing reviews from previous guests and therefore less reliable or subject to interpretation. However, it seems that they are still likely to have an impact on a potential guests eval- uationofthelistingandqualityoftheirstay. Consequentlyuserreviewsforcleanliness, communication(responsivenessofthehosttotheguest)andoverallvaluewereincluded in the model. The final Airbnb variable included was the designation of Superhost by Airbnb. This captures a signal by Airbnb of a minimum level of quality of the host, and we expected it to influence a rental decision or act as a filter for potential guests. All the user review variables and the Superhost designation were expected to have positive effects on the listing price. 4. Neighborhood value: As foreshadowed, this variable is a z-score constructed to reflect the neighborhood value and captures how many standard deviations each neighbor- hood was above or below the mean average cost per square foot. This can proxy for customer desirability or demand and owner costs. The listing price should be influ- enced by the value of the neighborhood both on the renters side, such that a nicer neighborhood would be more valuable to the renter, and from the host’s perspective, a moreexpensive neighborhood might incurhigher costs, both, to purchaseand maintain the property and therefore require a higher rental rate. Our final model includes both a neighborhood value variable and the square of the value. 5 We estimate the following specification: log(pricei) = α+δHi +βRi +τUi +ζNi +ǫi (1) The H vector contains information on the features of host i; the R vector is composed of rental listing features; the U vector contains categorical information on user reviews and the N vector contains the z-score and the squared value of the z-score reflecting neighborhood values as described above. Table 3 presents the summary statistics for all the variables used in the estimation. 4 Results Table 4 represents the estimation results from four different models were estimated. 1. Model 1: includes host’s biographical information as explanatory variables 2. Model 2: includes Model 1 and user reviews variables 3. Model 3: includes Model 2 and rental unit features 4. Model 4: includes Model 3 and neighborhood values Our initial parsimonious models 1 and 2 do not perform well in explaining the variation in listing prices on Airbnb as reflected in their very low R-squared of 1% and 7%. Model 3 has an R-squared of 68%, which makes it a good predictive model for Airbnb listings in San Francisco. We suspect that this is strongly related to the suggested price feature on Airbnb’s website, which suggests a listing price to the host based on the listing information that the host enters about the rental unit features. Model 4 has an R-squared of approximately 73% which means that all our explanatory variables are able to explain 73% of the variation in listing prices on Airbnb. With the exception of some of the race or gender variables, all of our independent variables are statistically significant. For the discussion below, we will refer to Model 4. Our expectation was that we would find a pricing differential in listing prices based on race. In fact, our model does predict a statistically significant 9.6% lower list price for Hispanic hosts and a 9.3% lower list price for Asians vs the control group of White, single, male hosts, while keeping all other explanatory factors constant. While the coefficient for Blacks indicated a 2.3% lower listing price (the sign being as expected) vs the White control group, it was not statistically significant. Our assumption is that this occurred due to the low number of observations of Black hosts in our data set (only 12) which accounted for a mere 1.6% of the observations compared to the San Francisco population percentage of 6.1%. While the Black race of the host was not a significant explanatory variable, it did predict a negative price effect as in Edelman and Luca (2014) on Airbnb listings. However, their study was focused specifically on potential discrimination against Blacks and only used two race categories, Black vs non-Black. Our model included more racial groups and other host features and was based on a sample size of 715 observations. In addition, we chose San Francisco as the focus of the study because it has very different racial demographic than New 6 York City (San Francisco is 6.1% Black; New York City is 15% Black) and also because we were interested in examining the possible effect on listing prices of other minority categories 3 as well as possible gender effects. The signs of the coefficients for many of our race and gender variables reflected our expected results. We generally expected minorities (e.g. Asians and Hispanics) to price their listings below the market price of non-minority hosts, which was in fact the result we found. Our model’s predicted effect on listing price of being an Asian host of -9.3% is consistent with but less in absolute values than that found by the recent Oakland Airbnb study, Wang et al. (2015). Their study found that Asian hosts earn on average 20% less than White hosts per week. However, there are significant differences between our data set and methodologies. The Oakland study used a very small sample size (100 observations), in a carefully selected neighborhood with a balanced Asian and White population and based listing price on a weekly cost vs a daily cost in an attempt to reduce price volatility. Our studyusedalargerneighborhoodweightedsampleof715observationsrepresentingtheentire San Francisco market, daily prices, and included variables to reflect the quality of specific neighborhoods. Additionally, the Oakland study, like the New York City study, only looked at two race categories. The Oakland study only used Asian/White as a variable of interest without the additional race, gender, couple and Gay variables in our model. Both gender and sexual orientation were found to be insignificant. Female hosts repre- sented a fairly large number of total observations, 265 of 715 or 37% and gay host listings wereonly35or4.8%ofthesamplesetvsarecentlyreportedgaypopulationforSanFrancisco of 6.2%, Newport and Gates (2015). We speculated that a female host may have an effect on listing price as a result of female guests having a preference for staying with a female host. However, since we have no data to observe the demand side of the market, it is unknown what the distribution of potential guests is based on gender who are shopping Airbnb listings or actually renting a listing. Therefore, any potential gender effects are difficult to identify. While the number of gay host listings is close to being representative of the San Francisco demographics (4.8% vs 6.2%) it is still a small number of observations which may be the reason its insignificance. The other Airbnb variables such as bedrooms, bathrooms, and being titled a “Superhost” proved to be consistent with our intuition and were found to positively affect the listing price. Only the overall review score value, although being statistically significant, had a sign opposite of what we expected, negative vs positive and is puzzling. This could be occurring because it is seen as too subjective to be reliable or that the term value may somehow have a low quality or cheap connotation. Since neither of the previous Airbnb studies - New York City or Oakland, included this variable in their model, we do not have a reference point against which to judge this result. Our use of the neighborhood value as reflected by the constructed z-scores and a their squared value proved to be highly significant. Our intuition was that the quality or value of the neighborhood as expressed by real estate prices should have an impact on listing price. This could have an impact on price from both the buyer and sellers side. On the buyers side, a higher quality neighborhood should be seen as more desirable and therefore justify a higher rental price and on the host side, a higher-priced neighborhood probably 3Population Demographics for New York 2016 and 2015 7