The Impact of Seller Transaction Unreliability: Evidence from Airbnb

Transcription

1 The Impact of Seller Transaction Unreliability: Evidence from Airbnb Jian Jia Liad Wagman February, 2018 Abstract A seller s transactional reliability refers to the seller s reputation for successfully completing a contracted future transaction. We model how information about transactional reliability impacts sellers, and show that the e ect depends on the search process that consumers endogenously follow. We empirically test our model s predictions using data from Airbnb listings in Manhattan, one of the most active short-term rental housing markets. We demonstrate evidence suggesting that consumers search for redeeming information about sellers who are perceived as transactionally unreliable, and that such sellers face significant consequences in terms of reduced sales and prices. Keywords: Reliability; reputation; sharing economy; Airbnb JEL Classifications: D81, D83, L14, L15 Stuart School of Business, Illinois Institute of Technology. jjia5@hawk.iit.edu. 565 W Adams St, 4th Flr, Chicago, IL Tel: Corresponding author. Stuart School of Business, Illinois Institute of Technology. lwagman@stuart.iit.edu. 565 W Adams St, 4th Flr, Chicago, IL Tel:

2 1 Introduction Sellers regularly contract with buyers for transactions that will take place at some point in the future. Examples include airlines, hotels, suppliers, among other numerous contractors. Sometimes, contractors fail to follow through on contracted obligations. Airlines oversell seats, hotels overbook rooms, suppliers underdeliver product units, and contractors in construction, consulting, carpentry, roof repair, among others, may fail to follow through on agreed upon projects. The impact such failures have on future transactions, once these failures are released to the public domain, for instance on the news or on review platforms, is difficult to assess. Of particular importance when studying the effect of broken contractual promises is to compare sellers, and their transaction failures, ceteris paribus. In this paper, we offer an empirical analysis of the impact of perceived transactional unreliability by using a unique data set of short-term home rentals. Airbnb, a sharing economy platform in the hospitality industry, enables residents to offer their homes as short-term rental accommodations. Airbnb landlords (henceforth, hosts ) can provide three different accommodation types: an entire home, a private room, and shared space. An important component of the Airbnb platform is the reciprocal reputation system it facilitates for guests and hosts within a 14-day deadline after a guest s stay, both guest and host (blindly) review each other. If one side does not review the other, the other s review becomes visible after the 14-day deadline. The literature has already shown that user reviews can play an important role in decisionmaking and purchasing behavior, 1 but one unique element of Airbnb s reputation system is that the platform also provides an automated system review, which looks similar to any other review, for listings whose hosts cancel a confirmed reservation prior to the guest s arrival. 2 These cancellation reviews signal to travelers that there may be a higher than usual 1 See, for instance, Bakos (1997), Chevalier and Mayzlin (2006), O Connor (2010), Moe and Schweidel (2011), Mayzlin et al. (2014), and Xiang et al. (2015). 2 The automated cancellation review format is: The host canceled this reservation X days before arrival. This is an automated posting, where X 1 is as stated. For same-day cancellations, guests can still post 2

3 probability that their lodging plans could fall through at some point prior to their arrival a potential situation that, especially in locales that are in high demand for temporary accommodations, can be quite costly. In addition, Airbnb as a platform, recognizing the impact that cancellations may have on travelers, discourages hosts from cancelling. 3 A distinguishing feature of these automatic reviews, since they are system generated and are posted only upon a confirmed cancellation by a host, is that they are credible and non-manipulable. In this paper, we study the resultant costs to sellers from being perceived as transactionally unreliable. We construct a model where consumers assess Airbnb hosts based on information such as reviews, and then proceed with booking a reservation if they consider a listing to be a good match for their needs. To identify whether a listing is a match, consumers endogenously follow one of two search approaches: purchasing after finding information that qualifies a listing to be a match, versus purchasing after not finding information that would disqualify a listing from being a match. We generate predictions, depending on the type of search, on how negative information about reliability in the form of cancellation reviews impacts sellers in equilibrium. We test our predictions against data from Airbnb listings in Manhattan, one of the most active short-term rental markets. Our focus on Manhattan reinforces the weight consumers may give to automated cancellation reviews because of the high demand for temporary lodging in the area (in the event a host cancels, it may be difficult to find alternate accommodations). We demonstrate evidence suggesting that consumers tend to use the approach of searching for information that qualifies a listing as a match prior to purchasing, as opposed to purchasing after not finding information that would disqualify a listing. We then demonstrate significant negative consequences for sellers who are perceived as transactionally a (non-automated) review. Prior to August 2015, the format was: The reservation was canceled X days before arrival. This is an automated posting. 3 Aside from receiving an automated cancellation review, hosts forfeit eligibility for Superhost status on Airbnb for a year, a status related to metrics concerning host and listing reliability, and which we will later show can have monetary value to hosts. Hosts may also incur direct monetary punishments from Airbnb in the form of a reduction in the amount of a future payout. 3

4 unreliable. There are multiple benefits to looking at system-generated cancellation reviews as a measure for negative information about seller transaction reliability. First, they are credible, non-manipulable, and demonstrably negative. Second, while prior works that study usergenerated reviews tend to focus on products such as goods, hotels or restaurants (including Mayzlin et al., 2014 and Luca and Zervas, 2016), Airbnb reviews are much more personal and rate an experience in another individual s dwelling. As a result, reviews on Airbnb are overwhelmingly positive (Zervas et al., 2015), which may grant further weight to the negative information implied by automated cancellation reviews. Third, Airbnb does not show individual guest scoring of a listing but only averages, making it less clear-cut to objectively identify negative guest reviews in a data set a non-issue for automated cancellation reviews. 1.1 Related Literature This paper contributes to the literature on review informativeness and reputation systems in online marketplaces (Senecal et al., 2004; Cox et al., 2009; Hu et al., 2009; Zhang and Sarvary, 2015; Dai et al., 2018). Works in this literature also demonstrate that reviewers can have strategic incentives to manipulate reviews, which may result in underreporting of negative reviews, particularly when users fear retaliation on platforms with reciprocal review systems (Bolton et al., 2013; Fradkin et al., 2017). Reviewers can also suffer from selection bias, where consumers are more likely to purchase and review products and services with which they are a priori satisfied (Li and Hitt, 2008; Masterov et al., 2015). Moreover, some reviewers may in fact be businesses leaving promotional or voluntary content (or even damaging the content of competitors) to artificially inflate their online reputations (Mayzlin et al., 2014; Benderson et al., 2018). Reviews that can be generated anonymously, even at a cost, may encourage manipulation (Conitzer and Wagman, 2014), and hosts on Airbnb can in fact boost the ratings of their listings by renting nights to friends or family, or to their own alternate accounts at minimum cost. 4

5 The automated cancellation reviews we study do not suffer from these potential manipulations. Our study thus adds to this literature by focusing on what is objectively negative system reviews, which can only be triggered by seller actions. This allows us to avoid issues concerning review authenticity, and to focus on equilibrium behavior as a function of seller reliability as driven by these cancellation reviews. The benefit of doing so is significant, because even the sheer possibility of review manipulation may impact the beliefs and actions of both buyers and sellers, which may result in different equilibrium behavior (Dellarocas, 2006; Anderson and Simester, 2014). This paper also contributes to the growing literature on the sharing economy and Airbnb. 4 Lee et al. (2015) point out that host reputation, including the number of reviews, host responsiveness, and host tenure, can impact the price of a listing. Zervas et al. (2015) indicate that Airbnb listings have higher average ratings compared to the hotel industry. Wang and Nicolau (2017) document that host attributes are the most important price determinants of Airbnb listings. Our work complements the above by shedding some light on how consumers incorporate information about seller reliability into their search and by providing empirical evidence that suggests that perceived unreliability is associated with significant costs for sellers. These costs include lower prices, less quantity sold, and ineligibility for a Superhost reputation badge on the platform which we show can have significant monetary benefits. The remainder of the paper is organized as follows. Section 2 presents the model and its equilibrium predictions. Section 3 describes the empirical methodology and data set. Section 4 reports our baseline empirical findings and Section 5 shows that our results are robust to other specifications. Section 6 concludes. 4 Recent works include Zervas et al. (2016), Edelman et al. (2017), Fradkin (2017), Fradkin et al. (2017), Kim et al. (2017), and Jia and Wagman (2017). 5

6 2 Model The majority of the Airbnb listings we consider have at or above 4-star reviews. The degree to which potential guests scrutinize a listing that matches particular search criteria can often be driven by the content of the listing s reviews. That is, consumer search within a listing is a critical element in the decision-making of a consumer, and directly connects with our focus on the impact of transactional reliability, as measured through sellers past cancellations a measure that is only observable once a listing is viewed. We thus focus our framework to model within-listing search, which takes as an input a seller s cancellations as a measure of reliability. Consider a market with sellers (hosts) and buyers (guests). Hosts advertise their listings for rent at a price per night, p. Listings are divided into categories (e.g., according to their number of bedrooms, bathrooms, their neighborhoods, etc, as observed on the platform). Let q denote the ex-ante category of a listing as it appears on the platform s search results, and let v(q) denote the value to a potential guest from staying at an accommodation of category q. Some listings, once viewed, may be perceived as riskier to book, due to negative information about reliability (i.e., whether the host will follow through with the reservation after a potential guest observes a cancellation in the listing s reviews, where we use higher riskiness to refer to lower seller reliability). Let us differentiate among listings within a given category according to their hosts reliability or riskiness, where this risk is modeled as a probability of additional costs for guests. That is, guests who book riskier (safer) listings are more (less) likely to incur high additional costs. We denote the type of cost outcome a guest may incur from booking a listing as either H or L, representing high and low expected additional costs, respectively. The prior that a listing is of type H is denoted by λ θ, where λ θ is higher for riskier listings and θ represents an ascending index of riskiness. Our assumption is that once viewed, guests gain an assessment of a listing s riskiness by reading 6

7 its reviews. A guest who books an accommodation of category q then incurs expected costs c {c L (q), c H (q)}, where c H (q) > c L (q). To account for traffic to a listing as a function of its price, we let D q (p) represent the aggregate guest traffic that views a listing of category q that offers a price p, with D q(p) < 0. A listing that is viewed by a potential guest is assumed to be a match for the guest s needs in terms of its category. Once potential guests view a listing and assess its riskiness, they can then scrutinize the finer details of the listing for any initially unobserved characteristics. This step is important, because the impact of perceived unreliability or riskiness is through its influence on how guests decide whether or not to book a listing. A guest s information acquisition process for this step is modeled by way of the guest choosing a screening intensity, α, representing the degree to which the guest scrutinizes the listing before moving forward, where α gives the probability with which the guest receives an informative signal about the listing s unobservable attributes. Additional scrutiny (higher α), which raises the likelihood of an informative signal, is increasingly costly. For tractability, we consider quadratic cost κ(α) = k α 2 for achieving search intensity α, with the scaling parameter k > 0 ensuring an interior equilibrium. Thus, when a guest acquires information about a listing, he receives a signal s, where: c with probability α s = with probability 1-α That is, with probability α the guest receives an informative signal, learning the H or L type that would be associated with reserving this listing, and with probability 1 α the guest receives an empty signal and is left with his prior belief. Different guests may associate different match values with the same listing. The implicit assumption is that all listings and their hosts can be seen as possibly being a bad match by potential guests. That is, hosts do not know whether a particular guest will perceive them as type H or L. Hosts, however, do know their perceived reliability type, as captured by θ, and set prices accordingly. 7

8 We simplify by normalizing the variables such that 1 > c H (q) > v(q) > c L (q) 0 and set the marginal cost of hosting to zero; assuming a non-zero cost does not change the qualitative nature of the results. The timeline of the game is as follows: Hosts first set prices for their listings, and each listing then receives some traffic from potential guests according to its price and category. Once guests view a listing, they assess its reliability type θ, and acquire information to decide whether or not to proceed with booking a reservation. 2.1 Guest Decision Following an informative search, a guest will choose to book a listing that is revealed to be a low-cost type and reject a listing revealed to be a high-cost type. Following an uninformative search, for θ {r, s}, a guest will book a listing if its value exceeds its expected cost, i.e., if: v(q) p q,θ λ θ c H (q) + (1 λ θ )c L (q). Equivalently, the price p q must satisfy p q,θ v(q) λ θ c H (q) (1 λ θ )c L (q). (1) If the rent offered by the host satisfies (1), the guest will book the listing following an uninformative search; in other words, the guest is searching for disqualifying information about the listing, and unless such information is found, the guest will proceed with the reservation. Let α q,θ denote the guest s search intensity in this case. Then the guest s problem is given by: max α q,θ (1 λ θ )(v(q) p q,θ c L (q)) + λ θ (1 α q,θ )(v(q) p q,θ c H (q)) kα 2 q,θ. 8

9 Solving the above yields: α q,θ (p q,θ ) = λ θ(c H (q) v(q) + p q,θ ). 2k 2.2 Host Decision Hosts with listings of category q and reliability type θ set a price p q,θ to maximize: Π q,θ (p q,θ ) = D q (p q,θ )(1 λ θ α q,b (p q,θ ))p q,θ (2) Anticipating the choice of screening intensity by guests in the proceeding stage, hosts solve for p q,θ that satisfies the following first-order condition D q(p q,θ )p q,θ [2k λ 2 θ(c H + p q,θ v(q))] + D q (p q,θ )[2k λ 2 θ(c H + 2p q,θ v(q))] = 0 (3) Rearranging gives ɛ q,θ (p q,θ )[2k λ 2 θ(c H + p q,θ v(q))] + 2k λ 2 θ(c H + 2p q,θ v(q)) = 0, (4) where ɛ q,θ (p q,θ ) is the price elasticity of listing traffic at price point p q,θ. Equilibrium price is then implicitly defined by: p q,θ = 1 + ɛ q,θ(p q,θ ) 2 + ɛ q,θ (p q,θ ) 2k λ 2 θ (c H v(q)), (5) λ 2 θ The right-hand side of (5) is increasing in ɛ q,θ (p q,θ ). Moreover, an increase in θ (and thus in λ θ ) decreases the second term on the right-hand side and thus decreases price. However, any decrease in price is coupled with a corresponding increase in ɛ q,θ (p q,θ ), and thus an increase in the first term on the right-hand side of (5). That is, while price indeed decreases in the riskiness level θ (i.e., when reliability decreases), the increase in demand that results from 9

10 the price change acts as a moderating effect. 5 To proceed with closed-form solutions, we henceforth make the following assumption: Assumption 1 Demand exhibits constant price elasticity over the price range of interest. Armed with Assumption 1, we have the following result. Proposition 1 When consumers search for disqualifying information, prices are lower for less reliable listings; however, less reliable listings exhibit lower vacancy rates than more reliable listings, ceteris paribus. Formal proofs are relegated to the Appendix. The result in Proposition 1 has empirical implications. Specifically, when buyers employ the approach of searching for information that would disqualify a listing: Hypothesis 1: Less reliable listings charge lower prices, ceteris paribus. Hypothesis 2: Less reliable listings exhibit lower vacancy rates, ceteris paribus. Hypothesis 2 may initially seem counter intuitive. However, it is simply the result of less reliable sellers (higher θ) having to doubly compensate buyers in terms of offering lower prices. One component of the price decrease is to compensate buyers for taking a greater risk when booking listings that are more likely to result in host cancellations. The second component of the price decrease is to incentivize buyers to proceed with a costly screening of such listings (i.e., to spend the time and the effort, rather than simply considering a different listing altogether). Both components are necessary in order for buyers to move forward with booking less reliable listings. In equilibrium, it is profit-maximizing for such sellers to pursue a strategy of volume, selling more (lower vacancy) but generating less profit per stay. 5 A technical condition required for the equilibrium to be well-specified is that ɛ q,θ (p q,θ ) ( 1, 0] or ɛ q,θ (p q,θ ) < 2 holds at the equilibrium price. This is because the right-hand side of (5) does not give a well-defined price for ɛ q,θ (p q,θ ) [ 2, 1]. 10

11 While this approach of searching for disqualifying information prior to proceeding with booking is theoretically viable and may in fact hold in some markets (e.g., Burke et al., 2012; Kim and Wagman, 2015), the next subsection presents an alternate approach that may arise in equilibrium. 2.3 Searching for Qualifying Information If the price of a listing is higher than specified by (1), then a potential guest will only proceed with a reservation if the listing is revealed to be a low-cost type. In other words, the guest does not allow for false positives the guest is searching for information that would definitively qualify the listing prior to moving forward with a reservation. The guest then chooses a search intensity as follows: Solving this maximization problem gives: max α q,θ (1 λ θ )α q,θ (v(q) p q,θ c L (q)) kα 2 q,θ α q,θ (p q,θ ) = (1 λ θ)(v(q) p q,θ c L (q)) 2k From the perspective of a host, when the price does not satisfy (1), expected profit is: We have the following result: Π gn q (p q,θ ) = D q (p q,θ )(1 λ θ )α q,θ (p q,θ )p q,θ. (6) Proposition 2 When consumers search for qualifying information, prices are lower for less reliable listings; moreover, less reliable listings exhibit higher vacancy rates, ceteris paribus. The result in Proposition 2 also has empirical implications. The first implication lines up with the former search approach, i.e., less reliable listings will tend to charge lower prices, ceteris paribus. The second implication is in direct contrast to the former search approach and provides an alternative hypothesis. Specifically, when buyers employ the approach of searching for information that would qualify a purchase: 11

12 Hypothesis 2-Alt: Less reliable listings exhibit higher vacancy rates, ceteris paribus. The reason for the alternate hypothesis is that, despite the increase in traffic following price reductions, less reliable hosts are significantly less likely to secure reservations. This is because buyers only book once they have qualified a listing as a definitive match for their needs. This is a criterion that is more difficult for unreliable listings to satisfy, since buyers are not as willing to invest time and effort into screening them. We next aim to test these hypotheses with data. Our objective is simple to identify how listings which may be perceived as less reliable, as measured by prior cancellations, compare in terms of prices and vacancies with more reliable listings. 3 Methodology and Data We collected all consumer-facing information and review content on the complete set of hosts who had advertised their listings in Manhattan on Airbnb from January 2015 until July This dataset has monthly scrapes at slightly irregular intervals. 7 Each listing is identified by a unique identifier and comes with time-invariant characteristics such as the host s unique identifier, neighborhood, approximate locale (latitude and longitude positioning in a six-digit decimal format that indicates the approximate location of a listing), and property type (entire apartment, private room, or shared space; we omit the latter due to its relatively low numbers approximately 300 listings on average and focus on the former two listing types, which are much more densely dispersed). 8 The listing information also contains some time-variant characteristics such as listing price, 9 the number of days during which the property is available for bookings over the next 6 New York City comprises one of the largest Airbnb markets in the world. The data is publicly available and was obtained from: 7 There are 17 scrapes in total including Jan, Mar, April, May, Jun, Aug, Sep, Oct, Nov(two scrapes), Dec in 2015; and Jan, Feb, April, May, June, July in The neighborhood of a listing is either mentioned in the actual listing or obtained from its coordinates. 9 Hosts may adjust prices of individual days. The listing price represents the base price chosen by the host for the listing, i.e., the price for days that are not specifically edited by the host. It is also the price 12

13 30, 60, or 90 days, whether the host has a so-called Superhost badge, 10 number of reviews, review rating, cancellation policy, minimum nights per stay, the maximum number of guests, a measure of the host s experience (number of days since the host s first listing was created), review gap (number of days since the latest review), whether the listing is offered for Instant Booking (i.e., without requiring host approval), and the host s average response time and response rate to guest inquiries. We focus on short-term rentals, thus, we concentrate on listings that are offered for rent for less than 30 days, omitting observations with minimum nights greater than 30. We also exclude observations with listing prices per night that exceed $1000 because some hosts may set their rates prohibitively high in lieu of blocking their calendars. We exclude observations before August 2015 because such observations do not contain important controls, such as cancellation policies, whether Instant Book was offered, and host response time and rate. We use the roughly monthly scrapes between August 2015 and July 2016, initially comprising 132,031 observations. Next, we construct a measure of competition for each listing by using geographical mapping software to tally up the total number of other listings of the same type that are located in close proximity. We define close proximity by forming a geographic circle of radius 0.1 or 0.3 miles around each listing based on its approximate coordinates. This calculation is repeated for each time period, so these count measures are time-varying. The next step in constructing our dataset is tallying up host cancellations for each listing. To do so, we take advantage of the previously-mentioned formatting of these reviews, e.g., The host canceled this reservation X days before arrival. This is an automated posting. We count the number of cancellation reviews for each listing in each time period. This number links to the parameter θ in our theoretical model. In particular, listings with more potential guests observe when they do not enter specific dates. 10 Hosts who meet the following criteria receive a Superhost designation, which indicates high reliability: (i) Hosted at least 10 guests in the past year; (ii) maintained a high response rate and low response time; (iii) received primarily 5-star reviews; (iv) did not cancel guest reservations in the past year. 13

14 cancellations correspond to riskier listings with lower seller reliability, which potential guests may associate with higher probabilities of additional costs. We also calculate the number of days that are vacant for each listing in the period of 30-to-60 days ahead of each data time period. We focus on this timeframe for two reasons. First, some guest reservations for this time window are likely to still be forthcoming and thus may be affected by host cancellations that occurred since the previous data period (i.e., days 0-to-30 are more likely to have been previously booked by guests). Second, 30-to-60 days is not too far in the future, making it more likely that potential reservations fall or partially fall in those dates, which gives a meaningful comparison among listings. While hosts may block certain days on their listings calendars (e.g., when they are traveling), we believe that any such behavior is independent of the number of cancellations a listing has. Table 1 gives descriptive statistics, separating entire-home rentals and private-room rentals in Manhattan from August 2015 to July 2016, as well as separating public listing information, the counts of cancellations, and the measures of competition. The average review rating is quite high, consistent with the literature. The bottom two rows give the average number of same-type competitors in a 0.1-mile and 0.3-mile geographical radius. They indicate that there are, on average, 41 entire-home competitors in a 0.1-mile radius and 298 entire-home competitors in a 0.3-mile radius. Although the average number of cancellation reviews seems low, approximately 27% of listings have at least one cancellation at the end of July Figure 1 depicts the distribution of cancellations across listings in the last time period of our sample, July Figure 2 depicts trends of variables of interest for the two different listing types over time. Figures 2(a) and 2(e) show that the number of entire-home listings stays roughly the same overall but the number of private-room listings increases. The average prices and number of cancellation reviews for both listing types share a similar pattern in Figures 2(b) and 2(d). The average number of vacant days in the period of 30-to-60 days ahead of each data pull is also similar for both listing types in Figure 2(c). Both listing types also have high 14

15 average review ratings in Figure 2(f) of about 4.5 (in comparison, Zervas et al., 2015, report an average hotel rating of 3.8). 4 Empirical Evidence We test how the number of cancellation reviews affects listing vacancy and price for the entire sample, and then run subsample analysis, separating entire-home and private-room rentals. Our baseline specifications are as follows: ln(v acancy i,t ) = α + βcancel i,t + δx + Listing i + Month t + ε it. (7) ln(p rice i,t ) = α + βcancel i,t 1 + δx + Listing i + Month t + ε it. (8) In specification (7), the dependent variable is the logarithm of vacancy of listing i in days 30-to-60 from time t. In specification (8), the dependent variable is the logarithm of the price of listing i at time t. The parameter Cancel i,t, the total number of cancellations for listing i at time t, is our explanatory variable in both specifications. Controls in X include listing attributes such as the number of bedrooms, beds, bathrooms, the number of reviews, overall review rating, host response rate, an indicator for Superhost status, an indicator for Instant Book being enabled, cancellation policy, the number of total listings the host has, maximum guests per stay, minimum nights per stay, and neighborhood (indicating one of 32 neighborhoods to which the listing belongs). In specification (7), controls also include the listing price at time t, to control for listings where the previous listing price may drive the changes in vacancy. In specification (8), controls include vacancy in days 0-to-30, to control for listings where immediate-term vacancy may drive price reductions. Listing i captures fixed effects for listing i, and Month i is a time dummy variable indicating each respective month. The parameter Cancel i,t 1 is lagged in specification (8) in order to consider potential price responses to a new cancellation, similarly 15

16 to how vacancy is treated. Both specifications are estimated using ordinary least squares. Table 2 shows the results for the vacancy regressions. Columns (1), (3), and (5) give the results without considering listing and time fixed effects. The coefficient on the number of cancellation reviews is similar in all three specifications. In particular, in the entire-home subgroup considering listing and time fixed effects, column (4) suggests that vacancy in days 30-to-60 is higher by 14.15% (about 4 days) for a listing that has one standard deviation above the mean number of cancellations relative to comparable listings without cancellations. In comparison, column (6) suggests a corresponding vacancy increase of 7.1% for privateroom rentals. It is important to emphasize that these vacant days may yet be reserved, as the timeframe under consideration is 30-to-60 days ahead of each time period. However, the results demonstrate significant cancellation-driven slowdowns in reservations. Enabling Instant Book does not seem to benefit entire-home listings in comparison with private-room rentals, which exhibit a 9.36% increase in booked nights over days 30-to-60. It is worth noting that the number of reviews does not play a significant role in driving vacancies, despite the coefficients being significant in all specifications. However, higher review ratings and Superhost qualifications do tend to lower vacancies. Table 3 reports the results of listing price regressions. The coefficient on the number of cancellation reviews, while negative, is small. For instance, column (4) suggests that listing prices tend to decrease by 0.38% for listings that have one standard deviation higher than the mean number of cancellations. Those negative effects may be relatively small because hosts are able to adjust the prices of individual nights without needing to alter the listing s base price. For instance, hosts who recently cancelled guests reservations may reduce prices in the near future in order to encourage guest bookings and receive additional (non-cancellation) reviews for their listings. The proceeding sections demonstrate a more pronounced effect on listing price when accounting for other relevant factors that differentiate listings. Table 3 also indicates that the number of reviews has a negligible effect on listing price, whereas review rating has some effect, although due to high average ratings, it is more likely 16

17 to be a punishing effect for listings with low review ratings. Listings hosted by Superhosts have listing prices that are approximately 10% higher. We also control for time-invariant listing attributes in both Tables 2 and 3. While these results are not reported, they relate to the variable v(q) in our theoretical model, which represents the value guests associate with a listing of category q. Of notable interest is the observation that the number of bedrooms and bathrooms play a major role in listing price and vacancy. For entire-home rentals in particular, listing price increases by 19.6% and vacancy decreases by 18.7% with an additional bathroom, and listing price increases by 9.4% and vacancy decreases by 11.3% with an additional bedroom. 5 Robustness 5.1 Accounting for Cancellation Attributes While the number of cancellations clearly interacts with price and vacancy, it does not reflect all of the information that is presented to buyers. More specifically, these systemgenerated cancellation reviews also reveal how many days ahead of guests arrival the host canceled their reservations. Moreover, the recency of a cancellation may matter a more recent cancellation that appears at the top of the listing s review stack may be both more noticeable and possibly perceived as more indicative of the host s unreliability than one that is buried beneath dozens of other reviews. In addition, the percentage of total reviews that are cancellations could clearly affect the host s perceived reliability. As an example, consider two listings, each with one cancellation, but one has a cancellation that is at the top of its small review stack and took place a day before guest arrival, while the other is on page 5 of its reviews (each review page shows 7 reviews on the desktop version of the platform) and took place 30 days before guest arrival. The listings are otherwise identical. In the previous subsection, both listings would have been treated identically, but they are unlikely to be treated as such by potential guests. By accounting for these 17

18 cancellation attributes, we can better align the empirics with the reliability parameter θ from our theoretical model. We record for each cancellation review (i) its location in the listing s review stack, and (ii) the number of days prior to guest arrival that the cancellation took place. We propose a risk index that would make it simple to incorporate these cancellation attributes into our specifications: [ RiskIndex i,t = ln ( + N i,t D1 P 1 1 D2 P ] 1 Dni,t P ) n i,t, (9) where N i,t is the total number of reviews listing i has at time t, n i,t represents the number of cancellations that listing i has at time t, D j, j = 1, 2,..., n i,t, D j 1, represents the number of days before guest arrival for each cancellation review, and P j, j = 1, 2,..., n i,t, P j 1, gives the page number of each system cancellation review. While fewer reviews are shown per page on the mobile version of the platform, their number is still proportional to the page number on the desktop version. We have alternatively used the review s position in the review stack in lieu of page number and obtained similar results. The risk index therefore captures the extent of a cancellation notice guests received and the recency of cancellations in each time period. We normalize by the number of reviews to reduce bias, while the square roots prevent the denominators from being overspread. Table 4 shows the summary statistics of the risk index for the three different groups. Entire-home rentals exhibit a higher average than private-room listings although the magnitudes are relatively small. Figure 3 depicts risk-index averages of entire-home and private-room listings over time. Tables 5 and 6 report the effect of the risk index in analogous vacancy and listing price regressions. The coefficient on the risk index is larger for vacancies in days 30-to-60 of entire homes than of private rooms. Column (4) of table 5 indicates that vacancies in days 30-to-60 increase by 15.67% (about 4.64 days) for listings with risk indices one standard deviation above the mean. In comparison, column (6) suggests a corresponding 7.59% (about

19 days) increase in vacancies for private-room listings. Both cases are similar to our previous specifications. From Table 6, the risk index also has a similar though somewhat larger negative effect on listing price compared to the previous specifications. Prices of entire-home (private-room) listings decrease 0.83% (1.03%) for listings with a risk index one standard deviation above the mean. Instant book, review number, and review rating exhibit similar effects on vacancies and price as previously. As before, the Superhost badge, which is akin to a certification of host reliability by the platform, carries monetary benefits, whereby the loss of eligibility for Superhost status by those hosts who cancel reservations has significant monetary implications. Superhosts of entire-home (private-room) listings have less vacancies, exhibiting 12.71% (6.31%) more occupancy in days 30-to-60, and they set listing prices that are about 10% higher. The results thus far are consistent with our Hypothesis 1 and 2-Alt of our theoretical model. That is, hosts whose listings are perceived as less reliable, i.e., listings that are associated with higher levels of θ in our theoretical framework, tend to have more vacancies and offer lower prices. The alignment with Hypothesis 2-Alt suggests that for a given listing category, as captured by our various controls, potential guests are wary of unreliable hosts, even when those guests do not encounter any other negative signals when reviewing a listing. That is, cancellations raise a sufficiently red flag that guests look to find positive redeeming information before moving forward with a reservation. The proceeding analyses reinforce the alignment with Hypotheses 1 and 2-Alt when accounting for other market factors. 5.2 Accounting for Supply Rental listings within the same neighborhood may face different degrees of competition. When a cancellation takes place, its effect on a listing may depend on the extent of nearby short-term rental supply. We use a small geographic radius around the coordinates of each listing to tally up the number of nearby competing listings of each type in each period. We 19

20 use these measures of competition as additional controls in our regressions. We report results for a 0.1-mile radius, and obtain similar findings when using a 0.3-mile radius. We hypothesize that the competitive effect from having more listings in close proximity has a positive impact on vacancy and a negative impact on price. We estimate the following fixed-effect specifications: ln(v acancy i,t ) = α + βriskindex i,t + ζcomp i,t + δx (10) +ηriskindex i,t Comp i,t + Listing i + Month t + ε it ln(listing P rice i,t ) = α + βriskindex i,t 1 + ζcomp i,t + δx (11) +ηriskindex i,t 1 Comp i,t + Listing i + Month t + ε it The dependent variables for both specifications are the same as previously. The variable Comp i,t gives the number of competitors in a 0.1-mile radius of listing i in period t. We also include interactions of the RiskIndex and Comp and mean center them both to better interpret the coefficients relative to average conditions, since both variables are continuous. Specifications (10) and (11) are estimated using ordinary least squares with standard errors clustered at the neighborhood level. Tables 7 and 8 report the results. For entire-home listings, column (2) of Table 7 indicates that given the average number of competitors in a 0.1-mile radius, there is a 16.35% increase in 30-to-60 day vacancy for listings with cancellations a standard deviation higher than the mean. Under the risk-index specification, vacancy would increase by 20.52%. In comparison, column (6) suggests that private-room listings, in the average competitive environment, have a corresponding increase of 8.2% in 30-to-60 day vacancy for listings with a number of cancellations a standard deviation higher than the mean, and their vacancy would similarly increase by 8.4% under the risk-index specification. 20

21 Under either specification, the effect on listing price is more negative in comparison to the same specifications but without accounting for the number of competitors. For entire-home (private-room) listings that have risk indices a standard deviation higher than the mean, in the average competitive environment, listing prices decrease by 2.47% (1.07%). Moreover, both Tables 7 and 8 suggest that vacancies (price) tend to increase (decrease) in the number of competitors, everything else held constant. The number of competitors in a 0.1-mile radius does not appear to impact the vacancies of entire-home listings as significantly as it does private-room rentals. One possible explanation is that due to regulatory issues and rising enforcement, the number of entire-home listings in Manhattan has remained depressed relative to demand. In contrast, the number of private rooms, which operate in less of a legal grey zone, has steadily increased, as suggested in Figure 2(a). That is, private-room hosts face fewer legal and enforcement hurdles than entire-home hosts, and, as this issue has received significant attention in the press, would-be entire-home hosts may have been discouraged from listing their properties. 5.3 Accounting for Demand While we accounted for the number of competitors, one potential concern is that vacancy is also a function of demand for a particular unit. To examine this issue, we implement a subgroup analysis. We use vacancy in days from each data-pull period, a future timeframe for which high-demand listings are more likely to have reservations, as a means to separate high-demand listings from low-demand listings. In each month, we separate listings into 5 groups based on their days 60-to-90 vacancies. We then calculate the effect of cancellations on vacancies and prices for each subgroup using both approaches (i.e. the number of cancellations and our constructed risk index). Tables 9 and 10 report the results. Column (1) of Table 9 shows that for high-demand listings with cancellations that number a standard deviation higher than the mean, there is a 17.22% increase in days 30-to-60 vacancy. Under the risk-index specification, vacancy 21

22 increases by as much as 18.32%. The corresponding increase in vacancy is approximately 9% for high-demand private-room listings. Tables 9 and 10 also indicate that high-demand listings suffer more from cancellations. One possible reason is that all attributes (location, amenities, service, etc) that made these listings popular are each individually harmed by an increase in the perceived unreliability of these listings hosts. Coupled with the rise in cancellation-driven vacancies, under either specification, the effect on listing price is similarly more pronounced, with high-demand listings incurring a larger price decrease than low-demand listings. In particular, high-demand entire-home (private-room) listings with cancellations a standard deviation higher than the mean incur a 2.21% (1.34%) price decrease relative to listings without cancellations. Under the risk-index, entire-home (private-room) high-demand listings similarly lower prices by 2.66% (2.08%). 6 Limitations and Conclusions We used automated cancellation reviews to form a measure of seller reliability and applied this measure to an analysis of Airbnb listings in Manhattan. We demonstrated evidence suggesting that such cancellation reviews may indeed be factored into the decision making of consumers, and sellers who are perceived as less transactionally reliable may suffer significant costs in terms of increased vacancies, lower prices, and loss of reputation by way of being disqualified from Superhost status. This evidence is consistent with our theoretical model. Our study suffers from a number of limitations. First, we are only able to account for the base price of listings and not the price of individual nights. Second, we are unable to account for blocked calendar days that may arise due to, e.g., reservations from other platforms (or due to the actual cancellations the platform automatically blocks those calendar dates). Third, we do not know whether the placement of listings in search results is affected (most likely negatively) by host cancellations. However, these limitations in a sense strengthen our findings, because they imply that our results represent lower bounds on the effects on listing 22

23 prices and vacancies. There are a number of promising directions for future work, including extending our analysis to other geographic regions in both the U.S. and in other countries. Since norms, density, and the legal landscapes, among other considerations, are likely to be locale and culture specific, equilibrium behaviors with respect to transaction reliability could differ from what we see in Manhattan. Our analysis can also be extended to other sharing-economy platforms. References [1] Anderson, E.T., and D.I. Simester (2014): Reviews without a purchase: Low ratings, loyal customers, and deception, Journal of Marketing Research, 51(3), [2] Bakos, J.Y. (1997): Reducing buyer search costs: Implications for electronic marketplace, Management Science, 43(12), [3] Bederson, B., G.Z. Jin, P. Leslie, A.J. Quinn, and B. Zou (2018): Incomplete Disclosure: Evidence of Signaling and Countersignaling, American Economic Journal: Microeconomics, 10(1), [4] Bolton, G., B. Greiner, and A. Ockenfels (2013): Engineering trust: Reciprocity in the production of reputation information, Management Science, 59(2), [5] Burke, J., C. Taylor, and L. Wagman (2012): Information acquisition in competitive markets: An application to the US mortgage market, American Economic Journal: Microeconomics, 4(4), [6] Chevalier, J., and D. Mayzlin (2006): The effect of word of mouth on sales: Online book reviews, Journal of Marketing Research, 43, [7] Conitzer, V. and L. Wagman (2014): False-name-proof voting over two alternatives, International Journal of Game Theory, 43(3),

24 [8] Cox, C., S. Burgess, C. Sellitto, and J. Buultjens (2009): The role of usergenerated content in tourists travel planning behavior, Journal of Hospitality Marketing & Management, 8(2), [9] Dai, W., G.Z. Jin, J. Lee, and M. Luca (2018): Aggregation of consumer ratings: an application to Yelp.com, Quantitative Marketing and Economics, forthcoming. [10] Dellarocas, C. (2006): Strategic manipulation of Internet opinion forums: Implications for consumers and firms, Management Science, 52(10), [11] Edelman, B., M. Luca, and D. Svirsky (2017): Racial discrimination in the sharing economy: Evidence from a field experiment, American Economic Journal: Applied Economics, 9(2), [12] Fradkin, A. (2017): Searching, matching, and the role of digital marketplace design in enabling trade: Evidence from Airbnb. MIT Sloan School of Management and Airbnb working paper. [13] Fradkin, A., E. Grewal, and D. Holtz (2017): The determinants of online review informativeness: Evidence from field experiments on Airbnb. MIT Sloan School of Management and Airbnb working paper. [14] Hu, N., J. Zhang, P. Pavlou (2009): Overcoming the J-shaped distribution of product reviews, Communications of the ACM, 52(10), [15] Jia, J. and L. Wagman (2017): Platforms, Anonymity, and Illegal Actors: Evidence from Airbnb. Working paper. [16] Kim, J.-H., T. Leung, and L. Wagman (2017): Can restricting property use be value enhancing? Evidence from short-term rental regulation, Journal of Law & Economics. Forthcoming. [17] Kim, J.-H., and L. Wagman (2015): Screening incentives and privacy protection in financial markets: A theoretical and empirical analysis, RAND Journal of Economics, 46(1), [18] Lee, D., W. Hyun, J. Ryu, W.J. Lee, W. Rhee, and B. Suh (2015): An analysis of social features associated with room sales of Airbnb, In Proceedings of the 18th ACM Conference Companion on Computer Supported Cooperative Work & Social Computing,

25 [19] Li, X., and L.M. Hitt (2008): Self-selection and information role of online product reviews, Information Systems Research, 19(4), [20] Luca, M., and G. Zervas (2016): Fake it till you make it: Reputation, competition, and Yelp review fraud, Management Science, 62(12), [21] Masterov, D., U.F. Mayer, and S. Tadelis (2015): Canary in the e-commerce coal mine: Detecting and predicting poor experiences using buyer-to-seller messages, In Proceedings of the Sixteenth ACM Conference on Economics and Computation, [22] Mayzlin, D., Y. Dover, and J. Chevalier (2014): Promotional reviews: An empirical investigation of online review manipulation, American Economic Review, 104(8), [23] Moe, W.W., and D.A. Schweidel (2011): Online Product Opinions: Incidence, Evaluation, and Evolution, Marketing Science, 31(3), [24] O Connor, P. (2010): Managing a hotel s image on TripAdvisor, Journal of Hospitality Marketing & Management, 19(7), [25] Senecal, S., and J. Nantel (2004): The influence of online product recommendations on consumers online choices, Journal of Retailing, 80(8), [26] Wang, D. and J.L. Nicolau (2017): Price determinants of sharing economy based accommodation rental: A study of listings from 33 cities on Airbnb.com, International Journal of Hospitality Management, 62, [27] Xiang, Z., V.P. Magnini, and D.R. Fesenmaier (2015): Information technology and consumer behavior in travel and tourism: Insights from travel planning using the internet, Journal of Retailing and Consumer Services, 22, [28] Zervas, G., D. Proserpio, and J. Byers (2015): A first look at online reputation on Airbnb, where every stay is above average. Working Paper. [29] Zervas, G., D. Proserpio, and J. Byers (2016): The rise of the sharing economy: Estimating the impact of Airbnb on the hotel industry, Journal of Marketing Research, Forthcoming. [30] Zhang, K., and M. Sarvary (2015): Differentiation with user-generated content, Management Science, 61(4),