NBER WORKING PAPER SERIES
SAFETY REVIEWS ON AIRBNB:
AN INFORMATION TALE
Aron Culotta
Ginger Zhe Jin
Yidan Sun
Liad Wagman
Working Paper 31855
http://www.nber.org/papers/w31855
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
November 2023, Revised January 2024
We are grateful to AirDNA for providing the data and to our home universities for financial
support. Marshall Van Alstyne, Matthias Hunold, Xiang Hui, Meng Liu, Peter Coles, Francine
Lafontaine, Ying Fan, Juan Pablo Atal, Juan Camilo Castillo, Sophie Calder-Wang, Yufeng
Huang, Zhe Yuan, Jisu Cao, and participants at the Luohan Academy Webinar, Washington
University at St. Louis, Boston University, University of Michigan, University of Pennsylvania,
the 2022 MaCCI annual conference, the 2022 IIOC annual conference, the 2022 INFORMS
Marketing Science and virtual conference, and the 2023 Strategy and Economics of Digital
Markets conference have provided constructive comments. Tejas Nazare, Nour Ben Ltaifa, and
Hunter Petrik provided excellent research assistance. The content and analyses in this paper
reflect the authors' own work and do not relate to any institution or organization with whom the
authors are affiliated. None of us has a financial relationship with Airbnb or competing short term
rental platforms. All rights reserved. All errors are our own. The views expressed herein are those
of the authors and do not necessarily reflect the views of the National Bureau of Economic
Research.
NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.
© 2023 by Aron Culotta, Ginger Zhe Jin, Yidan Sun, and Liad Wagman. All rights reserved.
Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission
provided that full credit, including © notice, is given to the source.
Safety Reviews on Airbnb: An Information Tale
Aron Culotta, Ginger Zhe Jin, Yidan Sun, and Liad Wagman
NBER Working Paper No. 31855
November 2023, Revised January 2024
JEL No. D83,L15,R3
ABSTRACT
Consumer reviews, especially those expressing concerns of product quality, are crucial for the
credibility of online platforms. However, reviews that criticize a product or service may also
dissuade buyers from using the platform, creating an incentive to blur the visibility of critical
reviews. Using Airbnb and official crime data in five major US cities, we find that both reviews
and self experiences concerning the safety of a listing's vicinity decrease guest bookings on the
platform. Counterfactual simulations suggest that a complete removal of vicinity safety reviews
(VSR) would hurt guests but increase revenues from reservations on Airbnb, with positive sorting
towards listings formerly with VSR. Conversely, incorporating VSR in a listing's overall ratings
or highlighting VSR could generate opposite effects. Either way, the interests of consumers are
not always aligned with the interests of a revenue-centric platform. Because VSR are more
closely correlated with official crime statistics in low-income and minority neighborhoods, our
findings suggest that suppressing or highlighting VSR would have different effects on different
neighborhoods.
Aron Culotta
School of Science and Engineering
Tulane University
New Orleans, LA 70118
United States
Ginger Zhe Jin
University of Maryland
Department of Economics
College Park, MD 20742-7211
and NBER
Yidan Sun
Illinois Institute of Technology
565 W. Adams St.
Chicago, IL 60661
Liad Wagman
Illinois Institute of Technology
565 W Adams, 4th Floor
Chicago, IL 60661
1 Introduction
Information design is crucial for online platforms. Take consumer feedbac
k as an example: not only does
it allow future buyers to discern high- and low-quality sellers, it also encourages good sellers to maintain
high quality and motivates poor-performing sellers to improve quality. Arguably, reputation mechanism
is even more effective online than offline, because online plat f orms can gather consumer feedback in a
standardized format, make it available globally, and aggregate it in a way t h at is sali ent and easy to
digest and search, if they so choose (see reviews by Einav et al., 2016; Tadel i s, 2016; Luca, 2017).
1
Critical consumer feedback (i.e., reviews criticizing a product or service) is part i cu lar l y important for
online reputation systems, whether i t is a simple choice of positive/neutral/negative, a star rating, or
detailed reviews with free-flowing text, photos, or even videos. It is likely t h at an e-commerce website
that continuously lists all products or services as rated five stars or with 100% positive feedback would
quickly lose credibili ty. Indeed, the literature has shown that c onsumers respond significantly to critical
feedback, although consumers tend to under-report critical feedback. Many platforms try to encourage
consumer feedback—including criti cal feedback—by offering status, coupons, and merchandise discounts.
Some p l at for ms also encourage sellers to respond to consumer rev i ews.
However, platforms face mixed incentives regarding critical consumer feedback. On the one hand,
future buyers may compare listings on the focal platform and prefer those with no or l ess criti cal feedback.
Such within-platform sorting may benefit the platform, as high-quality sellers often charge a higher price
or enjoy a h i gher probability of selling. On the other hand, buyers always have an outside good in mind;
observing many listings with critical feedback on a platform may dissuade buyers from using the platform
at all. The d an ger of losing a potential buyer could motivate a plat f or m to blur the visibility of critical
consumer feedback, by keeping it private to the platform, d el et in g it after a short time of posting, or
making it difficul t to find despite public posting.
More specifically, crit i cal consumer feedback may generate three information externalities on a digital
platform: first, buyer A’s critical feedback on product listing X may deter herself and other buyers from
buying X in the future. This “within-listing-across-buyer” effect is typical in a reputation system and is
well-studied. Second, critical feed b ack regarding X may lead other buyers to infer that listings similar
to X may have similar quality concerns even if they have not themselves received such feedback. This
is a “cross-listing-cross-buyer” effect. Third, a poor experien ce with listing X may motivate buyer A
to give critical feedback to X and reassess other buyers’ critical feedback towards other listings or even
the whole feedback system. This “cross-listing-within-buyer” effect is often omitted because Bayesian
updating assumes that learning from others’ experience is the same as learning from self experience if the
information has the same accuracy. However, in practice, self-experience can be much more salient to an
1
Recent examples include YouTu be, which has adopted a policy of hiding dislike counts on shared videos (see, e.g.,
https://rb.gy/xhhqnd), an d Instagram, which has given users the option of hiding likes (see, e.g., https://rb.gy/tacuj5).
2
individual. Few researchers have quantified the second and third ext er nal i t i es explicit ly ; one exception is
Nosko and Tadelis (2015), who show that buyers that have bought from a more reputable seller on eBay
are more likely to retu rn t o the platform to transact with any sellers, above and beyond the likelihood
to t r ansact with the same seller that created t h at good experience.
In this paper, we use user reviews about vicinity safety of short-term rentals to demonstrat e the
importance of these information externalities. Safety around a listi n g’s vicinity is an important dimension
of quality given the listing’s physical location . The host of a listin g cannot do much to change its vicinity
safety but prior guests may comment on it in their reviews. Such reviews may inform other gu est s of
the vi ci ni ty safety risk for nearby listings, which is a built-in cross-listing externality. Consumer reviews
regarding vi ci n ity safety are often of a critical tone because guests that have chosen to stay at a d welling
owned or managed by an anonymous host usually assume the neighborhoo d is reasonably safe. At the
same time, almost no hosts would volunteer to discuss neighborhood safety in their li st i ng descriptions,
because any mention (even the phrase of “perfectly safe”) may call guest attention to safety concerns.
Using all Airbnb listings and their reviews in five major US cities (Atlanta, Chicago, Los Angeles,
New Orleans, and New York City) from 2015/5 to 2019/12, we use a Lexicon ap pr oach to identify safety
reviews posted by Airbnb guests. We find that 0.51% of the 4.8 mil l ion guest reviews express concerns
of safety, among which 48.08% are about safety issues near but outside the focal property (referred to as
vicinity safety reviews, or VSR) rather than safety issues inside the property (referred to as listing safety
reviews, or LSR). Further sentiment analysis suggests that VSR and LSR identified by our algorithm are
significantly more negative in sentiments than an average review. Alth ou gh VSR and LSR only account
for a tiny fraction of guest reviews, 4.43% of listing-months ever have any VSR since May 2015, and
8.49% of listing-months ever have any VSR or LSR. These facts imply that safety concerns are not iced
by guests, regardless of whet h er they relate to the actual d welling or its nearby surroundings. At the
same time, th e low occurrence of VSR and LSR makes learning through self-experience a lengthy process.
Thus, guests with safety concerns mostly rely on the platform’s online review system and/or external
information.
Since guest feedback may reflect guests’ subjective opinion of their stay experience, we obtain (local)
government-reported crime statistics for the five sample cities, by zip code and month. Th e data suggest
that, as VSR accumulate slowly on Airbnb, the rank correlation between the normalized total count of
VSR in a zip code up to a month t and the normalized official cr i me statistics of that zip code-month is
increasing over time. For low-income or minority zip codes, the rank correlation can be as h i gh as 0.75
by the end of our sample period (2019/12). This suggests that the VSR, though noisy and subj ec t ive, do
reflect real safety risks in the related zip codes t o some degree.
As expected, when we f oll ow the same listings before and after they receive any VSR or LSR, there is
a sign ifi cant drop in the listing’s monthly occupancy rate as well as its average paid price per night. The
effect is stronger for LSR (-2.58% in occupancy and -1.52% in price) than VSR (-1.82% in occupancy
3
and -1.48% in price), but all are statistically significant with 99% confidence. These findings suggest that
prospective guests are concerned about both listing and vicinity safety, and have different sensitiviti es
to changes in t h ese two types of safety reviews. In addition to this classical “within-listing-cross-buyer”
effect in listing reputation, we also find a significant negative effect from VSR of nearby listings, where
nearby listings are defined as those wi t h in 0.3 mile radius of the focal listing accordi n g to Airbnb’s proxy
longitude and latitude of each listing. This “cross-listing-cross-buyer” effect corresponds to the second
information externality as mentioned above.
To document the third externali ty, we zoom into the gu est s that wrote VSR on Airbnb (ref er red to as
VS guests). Compar ed to the guests that have used Airbnb with similar frequencies and booked similar
listings (in terms of crime and VSR) but never write any VSR in our dataset, VS guests are 60.07% less
likely to book future stays on Air b nb after posting the VSR. And when they do book on Airbnb, they
tend to book in areas with fewer official crimes, fewer overall VSR, and a lower percentage of listings with
any VSR. The learning is weaker if the focal listing that t r i gger ed the VS guest’s VSR h ad previously
received any VSR from other guests, but even in this case, the VS guests are still 51.62% less likely to
book future stays on Airbnb after posting their own VSR. This suggests that self experience is much
more salient than reading other guests’ VSR; thus, the online review system is not fully effective as far
as conveying all the information emb ed ded in VSR.
Platform wide, we argue that these information external it i es —especially VSR spillovers to nearby
listings (the second externality) and VS guests’ strong reactions to thei r own vicinity safety exper ien ces
(the third externality)—may undermine a platform’s incentives to post an d highlight VSR as critical
feedback. Interestingly, in a recent policy change effective December 11, 2019, Ai r bnb announced that,
going forward, guest reviews about listings that include “content that refers to circumstances entirely
outside of another’s control” may be removed by the platform.
2
This policy change, if strictly enforced,
could imply that VS R are discouraged and may be subject to deletion by the platform while LSR are still
permitted. To be clear, we find no evidence suggesting that Airbnb has omitted or deleted VSR in any
systematic way after the 2019/12 policy. But the announcement itself suggests that Airbnb has broad
discretion regarding the collection, posting, or removal of c ons umer reviews, especially those that include
contents that t h e platform believes to be irrelevant or useless. Our analysis of VSR aims to shed light on
the potential economic incentives behind a platform’s review policy.
To do so, we must incorporate listing competition becau se within - and cross-platform sortings have
different implications for the platf or m. To account for listing competition, we obtain a dataset of compet-
ing entire-home VRBO listings and use a discrete choice model to estimate consumer utility from Airbnb
entire-home l ist i n gs, while treating VRBO listings in the same zip code-month as the outside good. We
then use the structural estimates to quantify consumer surplus under the status quo of our sample (i . e.,
VSR are largely permitted) versus three counterfactual information regimes: elimin at i ng all VSR (“no
2
See, e.g. , https://rb.gy/0pu5ck and https://rb.gy/9y6bum .
4
disclosure”), adjusting the rating of each listing to account for the number of VSR of the list i ng itsel f and
nearby listings (“VSR-adjusted ratings”), or alerting all guests to the existing VSR an d making them as
informed as those that have writt en VSR themselves (“high alert”).
Compared to the status quo, we find t h at no disclosure of VSR would decrease consumer surplus
in the market by 0.032% and increase revenues from reservations on Airbnb by 0.041%, with positive
sorting towards list in gs formerly with VSR. Conversely, VSR-adjusted ratings would increase the market
consumer surplus by 0.004% but decrease Airbnb’s GBV by 0.142%. High alert would increase the market
consumer surplus by 3.065% to 4.144% and change Airbnb’s GBV by +0.301% (+$10.1 millions) to -
1.304% (-$44 millions), depen d in g on whether we allow listin g price to change by 1% in response and
whether we assume the hi gh alert on vicinity safety also applies to the VSR for nearby listings. In short,
the interests of consumers and Airbnb are not always aligned, because guest sorting from Airbnb to
off-Airbnb alternatives would hurt Ai rb nb’s GBV with certainty but the within-Airbnb sort i ng between
listings with and without VSR may increase or decrease Airbnb’s GBV depending on how sensitive guests
are t o pricing and perceived vicinity safety of listings.
Although the overall welfare effects are moderate (because VSR is rare in the data), they mask l arge
distributional effects: more VSR transpar en cy benefits the l i st in gs without VSR at the cost of the listings
with VSR. Because l i st i ngs with VSR are more likely to locate in low-income or mi nor i ty neighborhoods,
consumer sorting upon VSR transparency would generate sizable GBV shifts across hosts in different
neighborhoods. These effects highlight a tradeoff as far as generating high er revenues and attracting
hosts in low-income and minority ar eas on the one hand, which can enhance the economic impact of the
platform in und erser ved neighborhoods, an d providing additional value t o guests on the other.
As detailed below, we contribute to the empirical literatu re of online feedback and sell er reputation,
and the risin g literature of information design in online plat f orms. As an information intermediary, online
platforms have more incentives than a traditional seller to alleviate information asymmetries between
buyers and sellers. But they are still inherently different from a social planner, because t h ey may put
more weight on their own business interests than on the welfare of buyers and sellers on t he platform, and
they may not fu l l y internalize the impact of their polici es on competing plat for ms. Our empirical findings
highlight these differences. We also document how the impact of a platform’s review policy may vary for
neighborhoods of different incomes or with different minority representation, as being inclusive could be
important for the platform or the social planner. These findings can help facilitate ongoing discussions
as to what role and r esponsibility digital platforms should have as far as collecting and disseminating
quality-related information online.
The rest of t he paper is organized as follows. Sect i on 2 reviews the related literature. Section 3
provides background regard i ng Airbnb’s review sy st em. Section 4 describes the dataset, defines VSR and
LSR, and provides summary statistics. Section 5 reports reduced-form evidence for the three information
externalities of safety reviews. Section 6 incorporates all of these externali t i es into a structural demand
5
model and predi ct s how listings’ GBV and consumer su r pl u s would change under three counterfactual
scenarios. Sect i on 7 discusses the implications of our findi n gs and future research directions.
2 Related Literature
Safety review is a type of buyer-to-seller feedback; thus, ou r stud
y is directly related to the literature on
online feedback and seller reputation.
The efficacy of online reputation depends on how consumers respond to buyer feedback posted on the
platform. Researchers have shown that consumers are more likely to purchase from sellers with better
buyer feedback and, conditional on purchase, are wi l l in g to pay more to reputable sellers (see reviews
by Bajari and Hortacsu , 2004; Tadelis, 2016; Einav et al., 2016). Consistently, we find that having any
safety reviews associated with a listing tends to negatively impact the occupancy and price of the li st i ng
because safety reviews dampen the li sti n g’ s rep ut at i on on Air bnb. Th e magnit u de of this effect on the
occupancy rate is comparable to a 70.18% red uct i on in the listin g’ s average guest ratings, confirming
the finding in Chak ravar ty et al. (2010) that consumers are more responsive to critical feedback than to
positive feedback.
Beyond the classical within-listing-across-b uyer effect, we are one of the few that attempt t o quantify
the spillover effect s of critical feedback. By definition, vicinity safety reviews may generate spillovers
among listings in nearby geographies, should guests infer the overall safety of the vicinity from multiple
nearby list i ngs. We find that for a focal listing, a higher percentage of other nearby listings with VSR
is negatively associated with the focal listing’s occupancy rate, as well as its price. This cross-listing-
cross-buyer spillover h as different implications for hosts and guests: hosts without VSR may suffer from
the negat i ve externality of nearby listings with VSR; but from a pr ospective guest’s perspective, this is
a positive information externality that coul d help them make more informed choices ex ante. Hence, the
information design optimal to the hosts or the platform can be different from that optimal to guests, a
key point we examine in th e counterfactual analysis.
While the cross-listing-cross-buyer spillover is specific to the nature of vicinity safety, we argue the
cross-listing-within-buyer spillover of critical feedback is more generalizable to other onli n e platforms.
As shown by Nosko and Tadelis (2015), buyers that have had a good experience from a reputable seller
on eBay are more likely to return to eBay for sales with any sellers. Similarly, we show that h aving a
negative vi ci ni ty safety experience tends to motivate a guest to avoid booking any listings on Airbnb in
our sample cities and, if she books again at all, to avoid both the listings and the areas that have any
VSR. Compared to Nosko and Tadelis (2015), we show that the cross-listing-within-buyer spillover is not
only limited to the extensive margin (whether to return to the platform for future transactions); but it
also motivates the experienced buyer to adju st how she interprets the presence of VSR in other listings.
Using a structural approach, we take a deep dive into t he implications of these spillover effects for the
6
welfare of guests, the revenue of hosts and the platform, and the distributional changes across different
types of neighborhoods.
Most of the aforementioned liter at ure of seller reputation is conditional on buyer feedback that online
platforms aggregate and present to consumers. However, buyer feedback is under-provided partly because
reviewers are not compensated for submitting reviews. For example, 64% of eBay transactions are rated
by buyers in the sample studied by Hui et al. (2021), and 73.5% of New York City UberX trips are
rated by passengers (Liu et al., 2021). In comparison, 44.6% of Airbnb trips in our sample have received
feedback from guests, which is in line with the guest r ev i ew rate repor ted by Fradkin et al. (2021) based
on ear li er Airbnb data in 2014.
Since accurate feedback is a public good subject to under-provision, many platforms attempt to
encourage buyer feedback by offering status, coupons, and merchandise discounts (Li and Xiao, 2014;
Cabral and Li, 2015; Li et al., 2020; Fradkin et al., 2015; Fradki n and Holtz, 2023). Some even encourage
sellers to respond to consumer reviews. Proserpio and Zervas (2017) find that hotels responding to user
online reviews enjoy 0.12-star increase in ratings and a 12% increase in review volume. When hotels
start responding, they tend to receive fewer but longer negative reviews because unsatisfi ed consumers
become less likely to leave short ind efen si bl e reviews when hotels are likely to scrutinize them. Similarly,
Chevalier et al. (2018) find that managerial responses stimulate consumers’ reviewing activity, especially
the negative revi ews that are seen as more impactful. This effect is reinforced by the fact that managers
respond more frequently and in more detail to negative reviews. These findings suggest that allowing
managerial response can be viewed as a platform policy that effectively highlights and addresses critical
feedback. In contrast, the 2019 Airbnb policy that motivates this study, if fully implemented, could
discourage buyers from providing critical feed b ack on certain quality dimensions such as vicinity safety,
and thus exacerbate the public good problem of critical feedback.
The imperfect review rate is particularly problematic as far as critical feedback is concerned. Studies
have shown that buyers tend to under-report b ad experiences, with potential explanations that include
fear of retaliation (Dellarocas and Wood, 2008), preference to leave the platform after a b ad experi-
ence (Nosko and Tadelis, 2015) , pressure to provide above-average ratings (Barach et al. , 2020), and
social connections to the rated sellers (Fradkin et al., 2015). For arguab l y rare, bad experiences such as
safety, th e prob ab ili ty of observi n g per t i nent feedb ack from prior buyers could be fur th er red uced , simply
because the chance of exp er i en cin g a safety i ssu e is small in absolute terms, even i f a neighborhood has
safety risks. A platform policy that discourages VSR could r ei nfor c e an existing bias against critical
feedback.
Another consequence of the bias against critical feedback is that saf ety r ev i ews on any Airbnb listi n g
accumulate slowly over time. This could affect the overall informativeness of safety revie ws. As shown
below, between 2015 and 2019, we observe a growing rank correlation between a zip code’s normalized
cumulative VSR count and the zip code’s normalized official crime stat i sti cs in low income and minority
7
areas. This suggests that cumulative VSR do contain useful informati on regarding a zip code’s actual
safety status, with informativeness that may increase over time. The rare occurrence of VSR further
highlights the importance of cross-listing-cross-buyer and cross-list i n g-wit h i n-b u yer spillovers, because
they magnify the impact of the rare experien ces and thus make t h e gradual accumulation of VSR more
informative. In comparison, a few studies argue that online feedback systems may b ecome less informative
over time because of the aforementioned feedback bias reasons (Barach et al., 2020; Klein et al., 2009;
Hui et al., 2021). Most of these stud ies infer feedback informativeness fr om the content of feedback or
policy variations within the feedback system. Our approach is different, as we compare online feedback
with a completely independent data source.
More broadly, our study contributes to the growing literature of inf or mati on d esi gn in online platf orms.
Because feedback is under-provided and there is a selection against critical feedback, researchers have
studied the design of feedback systems in terms of who is allowed to provide feedback (Klein et al., 2016;
Mayzlin et al., 2014; Zervas et al., 2021), h ow to improve the authenticity of feedback (Wagman and
Conitzer, 2008; Conitzer et al., 2010; Conitzer and Wagman, 2014), when the feedb ack i s revealed to the
public (Bolton et al., 2013; Fradkin et al., 2021), what kind of feedback is shown to the public, and how
to aggr egat e historical feedback (Staats et al., 2017; Dai et al., 2018).
Interestingly, some p lat f or ms highlight criti cal consumer feedback, so t hat future consumers are aware
of potential risks associated with the target seller or target product. An economic reason to do so is
that many consumers on online platforms tend to be more responsi ve to criti cal feed back than to positive
feedback (Chakravarty et al., 2010). Highlighting such f eedb ack may hurt the sellers with critical feedback
but d i vert buyers towards ot h er sellers on the same plat f or m with zero or not as much crit ical feedback.
If this sorting effect generates more revenue for the platform or reinforces the platform’s re put at i on as
far as honesty and transparency, the platform would have an incentive to highlight critical feedback.
In our setting, we offer a counterexample where a platform’s review policy has the potential to
discourage buyers from providing a specific type of critical feedback. The discouragement can occur
when a platform hides, obfuscates, or deletes critical feedback. To be clear, there are legitimate reasons
to do so in some situati on s: for example, a platform may find certain feedback fake, abusive, or misleading
ex post; omitting such feedback could make the information system more authentic and informative for
both buyers and sellers (Luca and Zervas, 2016; Chevalier and Mayzlin, 2006).
At the same time, theories have shown that platforms may be strategically motivated t o omit certain
information, includi n g critical feedback. For instance, Kovbasyuk and S pagn ol o (2018) explain why
sometimes plat for ms seek to erase some historical bad records of sellers, in order to increase matching
rates. Romanyuk and Smolin (2019) show that platforms such as Uber may seek t o hide some buyer
information (say, destination) prior to completing a buyer-seller match, because doing so would avoid
sellers waiting for a specific typ e of next buyer which would reduce the overall matching rate on the
platform. These two papers differ in the direction of information withholding: the former withholds
8
seller-relevant information from futur e buyers, while the latter withholds buyer-relevant information from
future sellers. Both suggest that the party from whom the information is kept hidden may be worse off
and the platform has an incentive to trade off their welfare loss against the welfare gain of the other side
of t he platform and the platform’s overall matching rate.
As shown in our counterfactual analysis, the plat f orm may have economic incentives to downplay
vicinity safety reviews, because the more guests are alerted ab ou t vicinity safety, the lower the matching
rate for the whole platform. In theory, such incentives could be dominated by a sorting effect, if posting
or highlighting VSR could dir ect buyers towards safer listings on t h e same platform and motivate the
safer listings to increase their prices sufficiently high to compensate for the pl at f orm’ s loss from a lower
matching rate. Our counterfactual analysis suggests that this i s not the case.
Finally, we are not the first to study safety issues regarding online short-term rental platforms. Suess
et al. (2020) find that non-host i n g residents with a higher emotional solidarity with Airbnb visitors are
more supportive of Airbnb hosts, and reside nts hold different views about safety (“stranger danger”) and
Airbnb dependin g on whether t h ey have children i n the household. Local plan ner s pay attention to the
impact of online sh ort -t er m rentals on neighborhood noise, congestion, safety, and local housing markets
(Gurran and Phibbs, 2017; Nieuwland and Van Melik, 2020; Kim et al., 2017). Zhang et al. (2021)
shows that regulations that negatively affect Uber/Lyft services may also n egati vely affect the demand
for Airbnb. Han and Wang (2019) d ocument a positive association between commercial house-sharing
and the rise of crime rate in a city, whil e non-commercial house-sharing does not have this association.
A number of studies find that an increase in Airbnb listings — but not reviews — relates to mor e
neighborhood crimes in later years (Xu et al., 2019; Maldonado-Guzm´an, 2020; Roth, 2021; Han et al. ,
2020; Fil i er i et al., 2021). More specifically, Airbnb clusters are found to correlate positively with property
crimes such as robbery and motor vehicle theft, but negatively with violent crimes such as murder and
rape. Also, Airbnb listings of the type in which guests may share a room wit h other unrelated guests
are found t o be more related to cri mes (Xu et al., 2019; Maldonado-G u zm´an, 2020) and to skirting local
regulations (Jia and Wagman, 2020).
Our study complements this growing literature, by highlighting safety reviews, distinguishing vicinity
and list i ng safety reviews, and documenting consumer responses to safety reviews or experiencing safety
issues. Although we cannot identify the effect of Airbnb on local crime rates, our work helps quantify
guest preferences regarding safety, and clarify how the interests of guests, different hosts and the platform
diverge with r espect to the disclosure of VSR. As shown in our counterfactuals, disclosing and highlighting
VSR could encourage guests to shy away from potentially unsafe listings and disproportionately affect
hosts in certain areas.
9
3 Background of Airbnb’s Review System
Over the past decade, short-term vacation rental markets have quickly
expanded worldwide. Airbnb, the
leading home-sharing marketplace, now offers 6.6 million active listings from over 4 million hosts in more
than 220 countries and regions.
3
As with any lodging accommodation, the specific location of a listing
can affect the experience of its gu est s. For instance, if a property is lo cat ed in a relatively unsafe area,
crimes such as carjacking or burglary may be more likely. In Los Angeles, the number of victims t o crimes
such as theft or burglary at short-term rental l odgings reportedly increased by 555% in 2017-2019.
4
As is
common in th e lodging indu st ry, guests, who may b e traveling outside their home towns and are theref or e
less familiar with local neighborhoods, are responsible for their own safety in the areas in which they
choose to stay. In particular, as with hotels, guests receive little to no protection fr om rental platforms
as far as crimes they may experience in a listing’s vicinity.
5
However, prior to making a reservation, potential guests may refer to a number of sources to gauge
the safety of a listing’s area — these sources include local news, crime maps, websites that summarize
neighborhoods
6
, and perhaps most readily linked to each li st i n g, the list i n g’ s reviews from prior guests.
7
Airbnb enables guests and hosts to blind l y review each other after a guest’s stay.
8
In an effort to appease
hosts, and perhaps to encourage more listings acr oss a larger number and variety of neighborhoods, a
recent Airb nb policy effective December 11, 2019 announced that, going forward, guest reviews about a
listing that include “content that refers to ci r cumst an ces entirely outside of another’s control” may be
irrelevant and subject to removal.
9
This policy change implies that reviews about the safety of a li st in g’ s
vicinity (“vicinity safety reviews” or VS R) may be deemed irrelevant and subject to removal, since such
a safety aspect is outside the control of the host. As detailed below, we compare th e frequency of VSR
(as observed on Airbnb) from mid 2015 to the end of 2020 but find no evidence indicating that Airbnb
has enforced this policy post 2019/12 as far as vicinity safety is concerned. However, anecdotes suggest
that some reviews that touched on neighborhood safety have been removed.
10
The policy does not apply
3
See Airbnb’s official statistics as of December 31, 202 2 available at https://news.airbnb.com/about-us/#:
~
:text=
Airbnb%20was%20born%20in%202007,every%20country%20across%20the%20globe.
4
See, e.g. , https://rb.gy/1eohbw .
5
See, e.g. , https://rb.gy/nwetrv and https://rb.gy/wrqvy4 .
6
See, e.g. , https://www.neighborhoodscout.com/.
7
Reviews have been well establish ed as having a potential effect on buyer decisio n s an d sellers’ repu t a ti o n s, p a rti c u la rl y
in the tourism in d u st ry (Schuckert et al., 2015). The lit erat u re also suggests that cri ti c a l information in reviews in particular
can have an effect on guest decisions and be useful to platforms in disting u i sh in g seller and produc t quality (Jia et al., 2021) .
8
If on e side does not review the other, the other’s review becomes visible aft er 14 days.
9
See, fo r example, https://rb.gy/0pu5ck and https://rb.gy/9y6bum .
10
For example, o n Jan. 27 , 2020, a tweet from “PatrickR0820” wrote “I used @Airbnb when we went to Atlanta fo r
the Panthers game. In my review I left numerous things that could be fixed as well as ‘the area that it is located in, is
pretty sketchy.’ My review and 4 ot h er simi la r rec ent reviews were deleted because it wasn’t relevant.” Another tweet by
“AveryBrii” on May 18, 2021 stated : “@Airbnb is such a joke!!! we litera l ly had a car stolen at the place we stayed at,
didn’t get refunded (which wahtever) & then i try to leave a review to inform others that it clearly was not a safe area
(cops told us this & other info that i tried to include) & they didn’t post.” A journalist also describes his experience on
Bloomberg Opinion: “Airbnb Took Down My Negative Review. Why?” (May 26, 2021 by Timothy L. O’Brien), accessed
at https://rb.gy/dxfkxw , on November 26, 2021.
10
to “listing safety reviews” (LSR), because these reviews are about the safety wit h in the listed property,
which presumably can be more r ead il y controlled and improved by the listing’s host.
It is difficult to pin down exact l y why Airbnb adopted this new review policy in 2019/12. If Airbnb
believes that the main role of online reviews is to motivate hosts to provide high-quality servi ce s t o guest s,
review content regarding something outside the host’s control may not help in that regard. Anecdot es
suggest that hosts have complained about the harm they suffer f r om “irrelevant” reviews about the
vicinity of their li st in gs,
11
and this policy change could be a way t o address these complaints. Another
reason might be the concern of review accuracy: arguably, vicinity safety is a subjective feeling subject to
the reviewer’s prior and interpretation, and it is often difficult to prove correct or wrong. However, similar
accuracy concerns cou l d apply to other review content, though the degree of objectiveness may vary. A
third reason may have somethin g to do with the aspiration of being inclusive. Airbnb has advocated
for in cl usi ve design, whi ch is defined as “consciously designing products, servi ces, and environments that
don’t create barriers to belonging.”
12
The same aspiration may have motivated Airbnb to adopt an anti-
discrimination policy, establish a permanent anti-discrimination team, and encourage designs and services
friendly to users with disabilities. To the extent that vicinity saf ety reviews are more present in low-
income or minority neighborhoods, the new review policy could be another effort to make the platform
more friendly to hosts in economically disadvantaged neighborhoods. The key question we address in
this paper is how the new poli cy, if fully implemented as far as VSR is concer ned , would redistribu t e the
economic benefits and costs among hosts, guests, and the plat for m.
To be clear, Airbnb has adopted other methods to address neighborhood safety directly. For example,
Airbnb introduced a neighborhood support hotline in 2019/12
13
, around the same time as Airbnb adopted
the new review policy. This hotline is primarily intended to be a means for neighbors of Airbnb listings to
contact the platform in certain situations (e.g., in the event of a party taki n g place at a listed proper ty).
In addition, since our main analysis sample ends in 2019/12 and we do not know how many guests that
left VSR in our sample would have used the hotline should the hotline exist at the time of the review,
we cannot predict how the hotline could counter some of the effects shown in our analy si s. That being
said, hotli n e usage is ex post and is not visible to future guest s, hence its impact on guests can be
fundamentally different from the impact of rev i ews visible under each listing on Airbnb.
Airbnb’s review system also allows guests to leave a 1-5 star rating by specific categories (cleanliness,
accuracy, check-in, communication, location, and value), in addition to leaving an overall rating and de-
tailed review. According to Airbnb’s response to a host’s question, location rating is meant to “help future
guests get a sense of the area and t en ds to reflect proximity to nearby destinations.”
14
Hence, location
11
Nina Medvedeva, “Airbnb’s Location Ratings as Anti-Black Spatial Disinvestment in Washin g t o n D.C.” Platypus: The
CASTAC Blog (March 16, 2021) accessed at https://rb.gy/ottzf9 .
12
See, e.g. , https://rb.gy/eq7ltv .
13
See, e.g. , https://rb.gy/sykoim .
14
See, e.g. , https://rb.gy/qs13gh .
11
rating could capt u re many location-speci fi c aspects such as local transit, nearby stores, neighborhood
walkability and noise, and may not be directly related to vicinity safety.
4 Data
Data of shor t-term rental listings. T
he main dataset we use has information on the set of short-term
rental list i n gs that had been advertised on Airbnb fr om 2015/5 to 2019/12, and on VRBO from 2017/6
to 2019/12, in five US cities (Atlanta, Chicago, Los Angeles, New Orleans, and New York). The data was
acquired from AirDNA, a company that specializes in collecting Airbnb and VRBO data. For Airbnb
listings, this dataset includes the textual contents of al l Airbnb listing reviews in those cities. We have
no access to reviews on VRBO. The original data from AirDNA extends to 2020/12 but demand for
short-term rentals has changed dramatically because of the COVID-19 pandemic, so our main analysis
uses data up to 2019/12 but we use data til l 2020 to infer Ai rb nb’s (lack of) enforcement of its 2019/12
policy beyond 2019.
Each listing is identified by a unique property ID and comes with time-invariant characteristics such
as the listing zip code, listing’s property type (entire home, private room, shared room, or hotel room)
as well as the host’s unique identifier. Listings also have time-variant characteristics, including average
daily rate,
15
the number of reservations, days that are reserved by guests, occupancy rate,
16
number
of reviews, overall rating scores,
17
the listing’s Superhost status,
18
the listing’s guest-facing cancellation
policy,
19
the average number of words in the listing’s reviews, the number of listings in the same zip code,
and whether the listing is cross-listed on VRBO.
20
Our unit of observat i on is listing-month. We focus on “active listings” (listings whose calendars ar e
not indi cat ed as ‘blocked’ in the dataset for an entire month), and exclude observations with an average
daily rate (ADR akin price per ni ght) over $1000, as some hosts may set their rates prohibit i vely h i gh in
lieu of blocking their calendars. We use regular monthly scrapes between 2015/5 and 2019/12 on Airbnb
(2017/6 to 2019/12 for VRBO). In total, the sample comprises 2,866,238 listing-months observations on
Airbnb, and 201,718 listing-months observations on VRBO.
Definition of safety reviews on Airbnb. We define two different types of safety reviews —
15
Average daily rate (ADR) is calculated by dividing the total revenue, including both nightly rates and cleaning fees,
earned by the host from reservations over a gi ven month by the total number of nights in that month’s reservations.
16
Occupancy rate is ca lc u l a ted by dividing the number of booked nights by the sum of the available nights an d booked
nights.
17
Overall rating scores are normalized to 0-10 range. Our dataset also includes location star ratings. Adding it as an extra
control variable does not change our main results, so we do not report it in this paper. Results are available upon request.
18
Superhost refers to a status badge rel a ted to metrics concerning a listing’ s performance. Hosts who meet the following
criteria, evaluated quarterly, receive a Superhost design a t io n : (i) Completed at least 10 reservations in the past 12 months;
(ii) maintained a high response rate and low response time; (iii) rec eived p ri m ari ly 5-star reviews; (iv) did not cancel guest
reservations in the past 12 months.
19
Cancellation policy could be strict, moderate, fl exi b le. For simplicity, we use a dummy variable to indicate whether a
listing’s cancellation p o li c y is strict or not.
20
Only l ist in gs with entire home that could be bot h listed on Airbnb and VRBO.
12
listing safety reviews (LSR) and vicinity safety reviews (VSR). LSR are those reviews that describe
issues pertaining to safety within a listing (e.g., “the listing is unsafe because there are fire hazards”,
“the listing is unsafe because of the slippery tub”, or “we saw mice in the kitchen three times during
our stay”). VSR contain information pertaini ng to the safety of the nearby vicinity or neighbor hood
of the listing (e.g., “the neighborhood is not safe”, “shady neighborhood”, or “unsafe ar ea”) . While
there i s considerable research regarding the use of machine learning for automated content analysis, these
methods typically require a large number of hand-labeled examples for training. We instead use a lexicon
approach due to its simplicity and transparency. Lexicons are also f oun d to have high levels of precision
as compared to machine learning approaches (Zhang et al., 2014; Hutto and Gilbert, 2014), and have
been used exte nsively in the literature (Monroe et al., 2008; Dhaoui et al., 2017).
To identify a suitable set of keywords, we use an it er at ive approach, startin g with terms such as
“unsafe,” “dangerous,” and “scary” and all of their synonyms, to obtain an initial keyword set; next, we
manually inspect reviews containing such keywords so as to identify additional keywords. We then select
keywords based on the accuracy of safety reviews.
More specifically, we conduct two iterations of manual labelin g. In the first iterati on, three research
assistants (comprising both male and female and different races) labeled 1.4K reviews that were generated
from the Lexicon approach algorithm wi t h the initial keyword set for both LSR and VSR. W hi l e labeling,
for each review the reviewers identified (i) whether the review pertains to neighborhood and/or listing
safety, (ii) whether the review has a negative sentiment with respect to neighborhood an d/or listing
safety, and (iii) three specific keywords that suppor t ed the reviewer’s decision in (i) and (ii). With these
human-labeled keywords, we obtain an updat ed list of vicinity and listing safety keywords such that the
percentage of critical reviews regarding vicinity safety (listing safety) in the 1.3K sample with such a
human-selected keyword is greater than 0% (10%).
In the second iteration of labeling, two research assistants (male and female) of d i ffer ent races labeled
3.1K reviews that were generated from the Lexicon approach algorithm wit h the updated keyword set
for both LSR and VSR, such that 5 reviews associated with each keyword were randomly selected. In
this iteration, reviewers labeled wheth er each r ev i ew pertains to negative sentiment about vicinity and/or
listing safety. The final set of keywords is the one where each vicinity safety (listing safety) keyword has
a percentage of negative-sentiment vicinity safety (listing safety) reviews greater than or eq ual to 60%
from both reviewers’ second-iter at ion labeling results. After two iterations, we expanded the list to 41
vicinity safety keywords and 50 listing safety keywords, as delineated in Appendix Table A1.
21
The keyword lists developed above are not the only inputs we use to define vicinity or listing safety
reviews. As far as VSR, to improve precision and to ensure that the text is indeed describing issues
21
Most of the keywords appear relatively infrequently, and removing any one of them alone has little effect on the results.
For example, one may argue t h a t “government housing” sug gest s a low-income area rather than vicinity safety issues.
Including it in our vicinity safety keyword list would only identify th ree more vicinity safety reviews and removing the
keyword has no qualitative impact on the results.
13
pertaining to the safety of a listing’s vicinity and not other aspects of a list i ng, we identified a list of 24
location keywords that tend to indicate a statement about the surrounding area (e.g., “neighborhood”,
“area”, “outside”) in Appendix Table A1. We then categorized the matching reviews into those in which
the vicinity safety keyword occurred within 20 words of a location keyword as vicinity safety reviews,
and those in whi ch the listing safety keyword occurred outside of the 20-word context as listing safety
reviews.
22
Next, we selected 13 ‘negative’ keywords, and fi l t ered out double-n egat i ve reviews where the
keyword occurs within 5 words of a safety keyword.
Overall, our approach resulted in 11.8k matched VSR and 12.8k mat ched LSR across the 5 sample
cities. In total, they account for 0.25% and 0.27% of all the observed Airbnb reviews respectively. From
2015/5 to 2019/12, only 4.43% of listings ever had any VSR, and only 8.49% of listing ever had any safety
reviews (VSR or LSR).
As shown in Appendix Figures A1 and A2, the top matching vicinity safety keywords are “unsafe”
(4,519), “homeless” (3,398), “yelling” (854), and “uneasy” (733), and the top matching listing safety
keywords are “worst” (1,803), “mold” (1,350), “stained” (1,172), and “filthy” (1,135). As an additional
validation check, we sampled several thousand matches at random, and manually labeled them as relevant
or not, finding 78.21% and 75. 64% accuracy for vicinity safety keywords and listing safety keywords,
respectively.
23
The mislabeled data often used figurative language (“scary how perfect this neighborhood
is”) or used safety words in other contexts (e.g., “watched a scary movie on Netflix”). While any such
method will be imperfect, we did not find any evidence suggesting that the error rat es were systematically
biased for some neighborhoods over others. However, we did restrict our keywords to English, so the
method will be less effective in areas with many non-English reviews.
To check whether the safety reviews defined above are indeed critical feedback as we intend to identify,
we employ a pre-trained NLP mo d el from Hugging Face to determine the sentiment score of all reviews
24
. According to the analysis, the overall average sentiment score across all available reviews is 0.79.
Specifically, VSR show a relatively neutral average sentiment score of 0.06, while sentences containing
VSR safety keywords tend to have a negative average sentiment score of -0.31. In contrast, the non-VSR
reviews have an average sentiment scor e matching the overall average of 0.79. On t h e other hand, LSR
demonstrate a comparatively lower average sentiment score of -0.41, and sentences with safety keywords
within the LSR category have the most negative average sentiment score of -0.76. In comparison, the
22
While the 20-word window is arbitrary, a sensitivity analysis suggests no qualitative difference when using a slightly
longer or shorter window. Moreover, the average review had roughly 50 words, so this seemed to restrict to the 1-2 sentences
around the keyword match.
23
This indicates a 21.79% fa l se-positive error rate for vicinity safety revi ews (24.36% for listing safety reviews). Since our
lexicon approach aims to minimize the false- positive rate while allowing false negatives, the safety reviews identified by this
approach tends to make the estimated impact of safety reviews more c o n servative than the true effect.
24
The utilized model is a fine-tuned checkpoint of DistilBERT-base-uncased, accessible at https://huggingface.co/
distilbert-base-uncased-finetuned-sst-2-english. It demonstrates a noteworthy accuracy of 91.3% on th e develop-
ment set. The sentiment scoring system ran ges from -1 to 1, where a score of -1 indicates an extremely negative sentiment,
and a score of 1 indicates an extremely positive sentiment.
14
non-LSR reviews have an average sentiment score again aligning with the overall average of 0.79. These
patterns suggest that our Lexicon ap pr oach has successfully captured the negative sentiment when guests
comment on listing or vicinity safety issues during their stay.
Official crime and demographic statistics. A second dataset we collect covers official c rime
records from databases tracking cr i mes in Chicago
25
, New Orleans
26
, New York City
27
, Atlanta
28
, and
Los Angeles.
29
These databases cover d i fferent types of crimes, including property-related crimes and
violent crimes. In terms of the geographical granularity of crimes, we consider crime events at the zip
code level. We also obtain median income and other demogr aph ic information at the zip code level from
2014, one year before our Airbnb sample period begins, from the Uni t ed States Census Bureau
30
. We
make the assumption that the income and demograp hi c in f ormat i on di d not change significantly over
our sample period. Throughout the paper, we ref er to a zip code as high-income (H) or low-income (L)
according to whether it s average income is above or below the median of the city it lo c ate s in. Similarly,
we r ef er to a zip code as minority (M) or white (W) according to whether i ts percentage of minorities in
population is below or above the city median.
Variable Definition. Ab ove al l , Appendix Table A2 defines the key variables used i n this paper,
including listing att r i bu t es (such as price, occupancy rate, safety reviews, and ratings) and neighbor h ood
attributes (such as income, population, and crime statist ics by zip code).
Summary of VSR and LSR on Airbnb. Tab le 1 summarizes the data at the listing-month level,
where vicinity safety (VS) Airbnb listings are defined as observations that have a positive number of
vicinity safety reviews (VSR) before the reporting month, while “normal” Airbnb listings do not h ave
any VSR before the reporting month. As the table indicates, about 4% of the total observations are
VS listings. O n average, VS l i st i ngs have higher occupancy rates, a higher number of reservations, a
higher fraction of Superhosts, and a higher number of reviews than normal listings. In contrast, the
nightly rates and overall rating of VS listings are lower on average than normal listings. The mean
number of cumulative VSR (aggregat ed up to the report i ng month) is 0.06 across all Airbnb listings,
and the mean number of cumulative listing safety reviews (LSR) is 0.06. Appendi x F i gur es A3 an d
A4 demonstrate t h e distribution of VS keywords for four groups of zip codes (high-income, low-income,
white, and minority). Compar i n g high-income with low-income (and white with minority) groups, it
appears that t he low-income (minority) group dominates the volume of VSR.
Did Airbnb enforce its new review policy after 2019/12? To infer whether Airbnb has enforced
its 2019/12 policy as far as vicini ty and listing safety is concerned, Figure 1 displays the percentage of
VSR and LSR on Airbnb, as id entified by our Lexicon method, from 2015/7 to 2020/12. While both VSR
25
Official crime data in Chicago: https://rb.gy/atjsss .
26
Official crime data in New Orleans: https://rb.gy/4vue82 .
27
Official crime data in New York City: https://rb.gy/iwrwp2 .
28
Official crime data in Atlanta: https://rb.gy/96txbl .
29
Official crime data in Los Angeles: https://rb.gy/tebnla .
30
See, e.g. , https://www.census.gov/data.html.
15
All listings VS listings Normal listings
(N=2,866,238) (N=126,868) (N=2,739,370)
VARIABLES mean p50 mean p50 mean p50
occupancyrate 0.56 0.64 0.68 0.78 0.56 0.64
occupancyrate du mmy 0.85 1.00 0.95 1.00 0.85 1.00
adr 164.69 125.51 134.15 106.31 166.10 126.67
No. of reservations 3.77 3.00 5.76 5.00 3.68 3.00
No. of reservationdays 14.16 14.00 18.56 21.00 13.95 14.00
lag VSR cumu dummy 0.04 0.00 1.00 1.00 0.00 0.00
lag LSR cumu dummy 0.05 0.00 0.20 0.00 0.04 0.00
lag VSR cumu 0.06 0.00 1.34 1.00 0.00 0.00
lag LSR cumu 0.06 0.00 0.26 0.00 0.05 0.00
lag VS listing radius pct 0.07 0.04 0.10 0.07 0.07 0.03
safety score (1-10, constructed by us) 4.96 5.09 2.83 2.33 5.06 5.23
ratingoverall (1-10) 9.18 9.60 9.09 9.20 9.18 9.60
review utd 33.71 15.00 93.02 70.00 30.96 14.00
No. of listing zip 540.67 449.00 554.66 481.00 540.02 447.00
cross listing 0.02 0.00 0.03 0.00 0.02 0.00
superhost 0.23 0.00 0.26 0.00 0.23 0.00
strict cp 0.50 0.00 0.58 1.00 0.49 0.00
ave wordcount cumu review 53.83 50.43 57.49 53.91 53.66 50.20
median income zip 57,187 50,943 42,645 34,432 57,861 51,427
population zip 48,158 45,747 42,514 36,654 48,419 46,025
white pct zip 0.53 0.59 0.41 0.38 0.53 0.60
h zip 0.52 1.00 0.29 0.00 0.53 1.00
w zip 0.60 1.00 0.44 0.00 0.61 1.00
crime cumu 19,435 9,650 31,230 14,205 18,889 9,475
Table 1: Summary Statistics of Airbnb Listings (2015/7-2019/12, unit of observation=listing-month)
and LSR have increased drastical l y after 2020/3, neither shows any discontinuous jumps f rom 2019/12
to 2020/2 as compared to the month-to-month fluctuation b ef or e 2019/12. The increase post 2020/3 is
likely driven by guests’ high attention to safety issues due to the pandemic rather than Airbnb’s change
of review policy in 2019/12, because that policy, if significantly enforced, should have led to a differential
reduction of VSR relative to LSR.
To double check, we have also examined the number of VSR/LSR removed in each quarter, by
comparing the reviews available on Airbnb from time to time. We find that almost all of the removed
VSR/LSR were from inactive listings. In short, we conclude that no eviden ce suggests Airbnb has enforced
its 2019/12 policy for VSR up to the end of 2020.
How do VSR correlate with official cr ime statistics? We also test the rank correlation between
the official crime records and VSR. Speci fi call y, we use the percentile rank of normalized crime records
in each zip code-month with i n each city — calculated as the number of reported crime cases in a month,
divided by the size of the population in that zip code. For each month, we rank the normalized crime
data within each city, and determine t he percentile crime rank of the zip code for that month. For
VSR, we use the percentile rank of the number of cumulative VSR in the zip code up t o the reporting
16
Figure 1: Percentage of Vicinity Safety Reviews (VSR) and List in g Saf ety Rev i ews ( LS R) on Ai r bnb
Over Time
month.
31
We then test the percentile ran k correlation index between the crime records and VSR in each
month, resulting in the time-series correlation trends depicted in Figure 2, which illustrates the correlation
trends for the f ou r different groups of zip codes (high-income, low-income, whit e, and minority). Figure 2
indicates that the correlation in low-income and min or ity groups exhibits an increasing trend, suggesting
that the percentile rank of VSR in a zip code is more likely to reflect t h e actual crime reports in the zip
code over time in these areas.
5 Reduced-form Effects of Safety Reviews
This section presents r edu ced -for m evidence from listing-level
and guest-level analyses. Listing-level
analysis documents the within-listing-cross-buyer effects of safety reviews as well as the cross-listin g-
cross-buyer effects of VSR. Guest-level analysis captures the cross-listing-within-buyer effects of VSR.
5.1 Listing-Level Ana ly s is
Baseline results. W
e begin by assessing the effects of VSR and LSR by listing-month. Our hypothesis
is that if potential guests view VSR and LSR as a proxy for safety around or withi n a listing, such reviews
would reduce the guests’ willingness to book the listing. Our base specification is gi ven by:
y
j,t
= α
j
+ α
k,t
+ δX
j,t
+ β
1
Crime
j,t−1
+
β
2
LSR
j,t−1
+ β
3
V SR
j,t−1
+ β
4
V SRADIUS
j,t−1
+ ǫ
j,t
, (1)
31
Due to data limitations, we assume that both records begin with clean slate (0 records) as of the beginning of our
dataset.
17
Figure 2: Correlation between the rank of normalized crime flow and the normalized total VSR
where j denotes a listing j-month t observation, Crime
j,t−1
is a log transformed variable that indicates
the normalized number of cumulative official crime reports since the start of the sample period for the
zip code where listing j is located, LSR
j,t−1
and V SR
j,t−1
are two dummy variables that equal 1 if the
listing has at least one LSR and VSR, respectively, before month t, V SRADIUS
j,t−1
is t he percentage of
listings that have at least one VSR within a 0.3-mile radius of listing j p r i or to month t, X
j,t
are listing-
level controls (logged except for dummy var iab l es) , includin g the number of reviews, overall ratings,
cancellation policy, number of listing in the same zip code, cross-listing status (i.e., whether the listing
is also listed on VRBO) , and whether the listing is host ed by a Superhost. The dependent variable y
j,t
is either the l og of listing j’s average daily rate (ADR) in month t, or the log of listing j’s monthly
occupancy rate (cal cu l at ed as log of 1 plus th e occupancy rate).
32
Listing and City–year-month fixed
effects are denoted by α
j
and α
k,t
, respectively, where the ci ty of listing i is denoted by k. Standard
errors are clustered by Airbnb property ID. The primary assumption is that, within a listing, the pr esence
and timing of safety reviews are correlated with the true safety condition around or inside the listing
and do not reflect selective reporting, fake reviews, or other strategic reasons once we control for other
time-varying listing attributes.
Our main specifications in Table 2 indicate that both VSR and LSR significantly decrease a listing’s
price (ADR) and occupancy. Speci fi call y, for an average Airbnb listing in our sample, having any VSR is
associated with a 1.82% reduct i on in the listin g’ s monthly occupancy rate and a 1.48% reduction in its
32
Some listing-mo nth observations have an occupancy rate of 0 and consequently are missing an average reserved daily
rate in the dataset for tho se months, though the dataset does offer a separat e “listing price” (i.e., a base rate) for those
listings. To extrapolate the ADR of th ese listings in the months in which they are missing, we calculate the mean ratio of
their ADR to their listing price in the months in which they are available, and multiply this average by the listing price in
the m issi ng months (if available, or by using the listin g price from the nearest month in which it is available).
18
average price per reserved night; having an LSR is associated with a 2.58% drop in occupancy and 1.52% in
price. LSR thus have a large r effect on price and occupancy than VSR, possibly because some prospective
guests have a specific geograph ic area (e.g., neighborhood) in mind, regardless of safety issues concerning
that area, whereas LSR describe safety issues that per t ai n to t h e listing itself. The pe rcentage of listings
with VSR within a 0.3-mile rad i us is associated with lower prices and lower occupancy, suggesting that
guests may also infer v i ci ni ty safety from the VSR of nearby listings.
(1) (2)
SAMPLE whole whole
MODEL OLS OLS
VARIABLES log occupan cy rate log adr
lag VSR cumu dummy -0.0182*** -0.0148***
(0.00140) (0.00219)
lag LSR cumu dummy -0.0258*** -0.0152***
(0.00135) (0.00210)
lag
VS listing radius pct -0.00859*** -0.00872**
(0.00253) (0.00390)
lag log crime cumu norm 0.0693*** -0.0508***
(0.00826) (0.0130)
lag
log review utd 0.00420*** 0.0117***
(0.000415) (0.000678)
log No. of listing zip -0.0212*** 0.0146***
(0.00185) (0.00289)
log
rating overall 0.0257*** -0.00240
(0.00128) (0.00200)
superhost 0.0175*** 0.00817***
(0.000586) (0.000845)
cross
listing 0.0311*** -0.00564
(0.00278) (0.00384)
strict cp 0.000601 0.0123***
(0.000803) (0.00126)
(0.0119) (0.0185)
Observations 2,866,238 2,866,238
R-squared 0.559 0.928
Note: *** p < 0.01, ** p < 0.05, All regressions control Time*City FE and Property ID FE, with st an dar d errors
clustered by Property ID. The variable crime cumu is normalized by the population.
Table 2: Main Results of Reduced-form Listing-level Analysis of Airbnb Listings
In contrast, normalized official crime records is associated with lower prices but higher occupancy. A
potential explanation is that hosts are aware of s afe ty issues in the areas of their listings, and proactively
lower their rates when t h ei r listings are located in relatively unsafe areas. These lower prices attract
more guest bookings, p er hap s either because guests tend not to seek information about crimes in the
neighborhood or because they prioritize price. In particular, for the average Airbnb listing in our sample,
given a 1% increase in the normalized official crime records, the daily rate is 0.05% lower whereas the
occupancy rate i s 0.07% higher.
Robustness. Our first robustness check tries to separate the extensive and intensive margins. Col-
19
umn 1 of Table 3 considers as the dependent variable a dummy t hat equals 1 when a listing’s occupancy
rate is posit ive and 0 othe rw is e. It reports a positive coefficient on Crime
j,t−1
, suggesting that the
variable Crime
j,t−1
not only describes the relative crime status of a zip code, but may also capture the
relative guest traffic to the area, where areas with relatively high guest traffic (e.g., downtown areas) tend
to h ave a higher number of reported (normalized) crimes.
Comparing th e coefficients on VSR and LSR for the whole-sample specifications (Table 2) to the
conditional sample with posit i ve occupancy rates (Col u mns 2 and 6 of Table 3), we find that the coeffi-
cients are similar but have somewhat higher magnitudes for the whole sample. One exception is that the
coefficient on Crime
j,t−1
becomes negat i ve after we condition the sample on listings wit h any positive
occupancy r at e, suggesting that the positive coefficient on this variable in the whole sample is driven by
the extensive margin only, whereas the intensive margin is consistent with the prior that bookings tend
to d ecl in e for listings located in a zip code with higher crime statistics over time.
(1) (2) (3) (4) (5) (6) (7)
SAMPLE whole occ>0 review utd<=13 review utd>13 whole whole who l e
MODEL OLS OLS OLS OLS OLS OLS OLS
VARIABLES
occupa n c y
rate
dummy
log
occup an c y
rate
log
occupa n c y
rate
log
occupa n c y
rate
log
occupa n c y
rate
log
occupa n c y
rate
log
occupa n c y
rate
lag VSR cumu dummy -0.0132** * -0.0129*** -0.0212*** -0.0100*** -0.0180*** -0.0155*** -0.0152***
(0.00155) (0.00119) (0.00541) (0.00146) (0.00140) (0.00140) (0.00139)
lag LSR c u mu dummy -0.0131*** -0.0213*** -0.0362*** -0.0173*** -0.0251*** -0.0228*** -0.0224***
(0.00153) (0.00114) (0.00458) (0.00142) (0.00135) (0.00136) (0.00136)
lag
VS listing radius pct -0.0100** -0.00575** -0.00760** -0.00334 -0.00864*** -0.00848*** -0.00 8 4 7 * * *
(0.00416) (0.00238) (0.00378) (0.00342) (0.00253) (0.00263) (0.00263)
lag log crime cumu norm 0.180*** -0.0167** 0.219*** 0.0118 0.0693*** 0.0652*** 0.0653***
(0.0123) (0.00734) (0.0150) (0.0106) (0.00826) (0.00874) (0.00874)
lag
log ave wordcount cumu review -0.00890***
(0.000744)
lag r sentiL cumu ave 0.0251***
(0.00215)
lag r sentiN cumu ave 0.0228***
(0.00120)
R-squared 0.420 0.499 0.565 0.522 0.559 0.560 0.560
VARIABLES log adr log adr log adr log adr log adr log adr
lag VSR cumu dummy -0.0126*** -0.00411 -0.0110*** -0.0150*** -0.0145*** -0.0146***
(0.00201) (0.00726) (0.00231) (0.0021 9 ) (0.00219) (0.00219 )
lag LSR c u mu dummy -0.0112*** -0.00152 -0.0124*** -0.0158*** -0.0150*** -0.0152***
(0.00189) (0.00705) (0.00218) (0.0021 0 ) (0.00210) (0.00210 )
lag
VS listing radius pct -0.00848** 0.00263 -0.0140*** -0.00868** -0.00993** -0.00992**
(0.00337) (0.00620) (0.00478) (0.0039 0 ) (0.00387) (0.00387 )
lag log crime cumu norm -0.000974 -0.0600** -0.0242 -0.0508*** -0.0490*** -0.0490***
(0.0120) (0 . 0 2 40 ) (0.0168) (0.0130) (0.0135) (0.0135)
lag
log ave wordcount cumu review 0.00645***
(0.00106)
lag r sentiL cumu ave 0.00572*
(0.00293)
lag r sentiN cumu ave -0.00124
(0.00163)
R-squared 0.943 0.931 0.937 0.928 0.931 0.931
Observations 2,866,238 2,441,566 1 , 3 7 0 ,6 5 5 1,495 ,5 8 3 2,866,238 2,655,504 2,655,504
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control Time*City FE and Prop erty ID FE, with standard errors
clustered by Property ID.
Table 3: Robustness Checks for Reduced-form Listing Level Analysis of Air
bnb Listings
We conduct a number of additional checks. First, we split t he sample by whether a listing has an
20
above- or below-median number of rev i ews in a given month (median is 12), as a proxy for whether the
listing is in its early or later “stage” of taking guest reservations, since only staying guests can post a
review.
33
Another motivation for this partition is that prospective guests are more likely to notice safety
reviews (both VSR and LSR) when listings have a lower number of reviews. Indeed, Columns 3 and 4 of
Table 3 report that in the subsample of listings with 13 or fewer reviews, the negative effects of having any
VSR and LSR on occupancy rate (2.12% for VSR and 3.62% for LSR) are higher than the corresponding
negative effects for list in gs with more than 13 reviews (1.00% for VSR and 1.73% for LSR). However,
Columns 7 and 8 indicate that as far as listings’ dail y rates are concerned, this comparison is reversed,
possibly because hosts of newer listings may still be in the process of identifying their pricing for those
listings. Second, we add additional controls for the average word count of a listin g’s reviews.
34
As
Columns 5 and 9 of Table 3 indicate, the results do not qualitat i vely change from our main specifications
when incorporating the addit i on al control.
Heterogeneous effects. We next explore a number of he t erogen eou s effects. Table A3 provides
summary statistics based on the type or area of a listing. In particular, the table reports different
normalized zip code crime levels for listings in these categories. We proceed with a similar empirical
methodology as in (Equation 1), but with different subsamples.
We begin by analyzing four groups of zip codes separately (high-i ncome, low-income, white, and
minority). Table 4 shows that VSR have negative effects on occupancy rates across all four subsamples.
The negative effects of having any VSR on occupancy rates have h i gher magnit u des in high -in come and
white zip codes (1.76% and 1.89%) than in low-income and minority zip codes (1.72% an d 1.75%). A
similar comparison holds for LSR. One potential explanation is that guests may have different prior beliefs
and different sensitivities to safety issues, and perhaps more so if their search targets a specific area that
they believe is r el ati vely safe. Hosts in different areas may also react differently to VSR and LSR, based
on h ow they gauge guest perception and guest preferences.
We next consider subsamples comprising different listing types (entire home, private room, shared
room, and hotel room). Additional heterogeneous effects may arise here because, for instance, for guests
who seek parti al spaces (private room, shared space) within a dwelling, safety issues may be more salient.
The results in Appendix Table A4 indeed show that the magnitude of the negative effects from having
any VSR and LSR on occupancy are lar ger for private rooms and shared spaces (2.10% and 3.01% for
VSR and 3.08% and 2.89% for LSR, respectively) in comparison with entire-home listings (1.61% for
VSR an d 2.36% for LSR).
33
To be clear, the same listing may be in both subsamples over time, but belong to only one of the subsamples in any
given month.
34
Host responses to safety reviews are not observed in our data
21
(1) (2) (3) (4) (5) (6) (7) (8)
SAMPLE H L W M H L W M
MODEL OLS OLS OLS OLS OLS OLS OLS OLS
VAR I A B LES
log
occ u pa n c y
rate
log
occ u p a nc y
rate
log
occu p a n c y
rate
log
occ up a n c y
rate
log adr log adr log adr log adr
lag VSR cumu dummy -0.0176*** -0.0172*** -0.0189*** -0.0175*** -0.0163*** -0.0138*** -0.0153*** -0.0136** *
(0.00257) (0.00168) (0.00215) (0.00185) (0.00389) (0.00267) (0.00330) (0.00295)
lag LSR cumu dummy -0.0263*** -0.0247*** -0.0251*** -0.0265*** -0.0181*** -0.0123*** -0.0177*** -0.0114***
(0.00196) (0.00187) (0.00178) (0.00207) (0.00284) (0.00307) (0.00269) (0.00335)
lag VS listing radius pct -0.0117*** -0 . 0 0 44 9 -0.00780** -0.00942*** -0.00261 -0.0126** -0.00308 -0.0122**
(0.00370) (0.00346) (0.00385) (0.00335) (0.00564) (0.00535) (0.00589) (0.00516)
lag
log crime cumu norm 0.0512*** 0.171*** 0.0427*** 0. 17 0 * * * -0.0496*** -0.0561*** -0.0478*** -0.0625**
(0.0111) (0.0137) (0.00950) (0.0168) (0.0179) (0.0213) (0.0150) (0.0265)
Observations 1,484,474 1,381,764 1,716,774 1,149,464 1,484,474 1,381,764 1,716,774 1,149,464
R-squared 0.552 0.569 0.551 0.573 0.921 0.924 0.919 0.925
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control Time*City FE and Prop erty ID FE, with standard errors
clustered by Property ID.
Table 4: Reduced-form Listing-level Analysis of Airbnb Listings By Four Are
a Typ es
5.2 Guest-Level Analys is
We conduct guest-level analyses to test whether guests who leave any
VSR (hencef ort h , VS guests) act
differently befor e and after they post their first VSR in comparison to otherwise similar gu ests who did
not leave any VSR. This aims to capture the cross-listi n g-wi th i n-b u yer effect of VSR. To that end, we
assume that the first VSR that a VS guest posts for one of the listings in our sample (i.e., covering Airbnb
listings in the five cities we consider, with reviews beginning in May 2015) is the first VSR that this guest
posted. To reiterate, any such guest s who have ever posted VSR in our sample are considered VS guests;
otherwise, they are treated as “normal ” users. To ensur e that the VS users have had some experience on
Airbnb prior to leaving their first VSR, we focus on the subset of VS users th at left at least two reviews
in the five sample cities before leaving their first VSR.
In order to match VS users with n or mal users, we use a K-nearest neighbor (KNN) method to select
the two most simil ar control (normal) users for each tr eat ment (VS) user. The user characteristics used
in the KNN method (as of the time of the treat ment user’s first VSR) are the user’s number of prior
reviews, the average normalized crime reports in the cities in which the user stayed (based on their prior
reviews), the average number of VSR for listings for which the user left reviews, the average percentage
of overall VS listings in the same zip codes as well as i n the 0.3-mile radius area as listings for which the
user had prev iou sl y left reviews, and the average number of words f or the reviews that the user posted
before. The matching is done for each month (i.e., based on new treatment users in each month). The
same “treatment month” is applied (hypothetically) to control users that are matched with a treatment
(VS) user, based on the latter’s timing of their fir st VSR.
To assess if the treatment and control users have the same tendency to post VSR, we also calculate
the propensity score for each user in our matched sample. In particular, we regress the dummy of a user
22
Figure 3: Distribution of Propensity Scores for VS users ( t r eat ed) and Normal users (control)
being a VS user on the number of reservations she had made on Airbnb before the treatment time, the
average zip code-wide crime rate of these reservat i on s at the time of reservation, the average number of
VSR in these reservations, the percent of listings with any VS R in the zip co d e as well as in the 0.3-mile
radius area of these reservations, and the average number of words for the reviews that t h e user posted
before. For a treated user, the treat ment time is when she wrote her first VSR in our sample. For a
control user, the treatment time is when the treatment user she is paired with wrote her first VSR in
our sample. Table 5 reports that the t r eat ment and control users are si mi lar as far as the characteristics
considered in the KNN method; t h e two user groups also have similar propensity scores, as shown in
Figure 3.
Panel A: VS users Panel B: Normal users
VARIABLES mean p50 sd N mean p50 sd N
reservation pre 2.76 2.00 1.51 2,252 2.72 2.00 1.43 4,504
log ave crime cumu norm pre 0.93 0.28 1.95 2,252 0.81 0.27 1.44 4,504
ave
vsr cumu pre 0.63 0.50 0.44 2,252 0.64 0.50 0.43 4,504
ave vs listing zip pct pre 0.06 0.05 0.04 2,252 0.06 0.05 0.04 4,504
ave vs listing radius pct pre 0.09 0. 07 0.07 2,252 0. 08 0.07 0.06 4,504
log ave wordcount cumu review pre 4.37 4.39 0.64 2,252 4.36 4.39 0.63 4,504
propensity
score 0.74 0.72 0.15 2,252 0.73 0.71 0.15 4,504
Table 5: Summary Statistics by VS and Normal Users in the DID Sample
We first test whether VS users behave differently in terms of su b sequ ent reservations on Airbnb after
23
their first VSR (as exhibited by their subsequent listing reviews). We use a difference-i n -di ffer en ces
methodology (DID) as follows:
y
it
= α
t
+ α
p
+ β · V S
user
i
+ γ · V S user
i
× post
t
+ ǫ
i,t
, (2)
where the subscrip t p denotes the treatment-control pair identified in the sample construction.
We const r uct several measures for the dependent variable y
it
: the first is the number of reviews th at
user i wrote in month t. We use it as a proxy of user i’s Airbnb reservations in t, which can be zero and
thus captures both the extensive and intensive margins. Because it is a count variable, we use a Poisson
regression instead of ordinary least squar es. The other measures of y
it
include the normalized cumulative
count of officially reported crimes in the zip codes of user i reserved listings in month t, the number of
VSR in i reserved listings, the percentage of VS listings in the zip codes as well as i n the 0. 3-radi u s area
of the i reserved listings, and whether the reserved listings have any VSR. These variables capture t he
types of listings that i books on Airbnb conditional on her booking at all (the intensive margin). The
dummy V S
user
i
equals 1 for VS users and 0 otherwise, and the dummy post
t
equals 1 if t is after the
time of the first VSR of VS user i. The key variable is the interaction between V S user
i
and post
t
while
post
t
alone is absorbed in year-month fixed effects α
t
. Tr eat ment-control pairs fixed effects are denoted
by α
p
. Standard errors are robust and clustered by treatment-control pairs.
In Panel A of Table 6, Column 1 reports results fr om a Poisson mo d el based on an unbalanced monthly
panel data, in di cat i ng that VS users tend to book fewer reservations (as evi d enced by subsequent revi ews)
after posting their first VSR. In particular, the average monthly number of subsequent reviews is expected
to be 60.07% lower for VS users in comparison with normal users.
35
Columns 2-6 assess whether VS users
are more sen sit i ve to safety information when booking subseq uent Airbnb listings after posting their first
VSR. Results suggest that the subsequent listings chosen by VS user s tend to locate in zip codes that
have fewer normalized crime reports, are less likely to have VSR, and are less likely to locate in zip codes
that have a higher overall percentage of VSR or a higher percentage of other listings with VSR. This
suggests that VS users, relative to normal users, are more sensitive to safety information after posting
their first VSR.
One may argue that the extent of learning t hr ou gh self-experience would depend on a guest’s prior
about vicinity safety. Unfortunately, we have no data on each guest’s home town and therefore cannot
approximate their prior with the type of vicinity they normally live in. Nevertheless, some VS users may
have seen some VSR left by a listing’s previous guests, and that listing eventually triggered their own
VSR, and therefore would not respond as vigorously to their own vici ni ty safety experience as other VS
users. To test this, we create a dummy (First-Is-First) indicati ng whether a VS user’s own VSR was the
35
This is not the coefficient of the treatment dummy (-0.918) because we use a Poisson model for this regression, i.e., th e
applicable percentage is 1 − e
−.918
.
24
(1) (2) (3) (4) (5) (6)
SAMPLE
monthly
reservation
reserved
property
reserved
property
reserved
property
reserved
property
reserved
property
MODEL Poission Poission Logit OLS OLS OLS
VARIABLES
reservation
monthly
VSR
cumu
VSR cumu
dummy
crime cumu
norm
VS listing
pct zip
VS listing
pct radiu s
Panel A: Full sample
VS user × post -0.918*** -0.697*** -0.490*** -0.927*** -0.0250*** -0.0247***
(0.0601) (0.135) (0.113) (0 .1 1 2 ) (0.00267) (0.00505)
Observations 254,056 22, 2 6 5 22,237 22,415 22,415 22,415
Panel B: Subsample VS use’ s 1st VSR is the 1st VSR of the listing
VS user × post -0.961*** -0.793*** -0.696*** -0.961*** -0.0280*** -0.0275***
(0.0667) (0.146) (0.129) (0 .1 2 7 ) (0.00271) (0.00551)
Observations 202,262 17, 7 4 3 17,726 17,893 17,893 17,893
Panel C: Subsample VS use’ s 1st VSR is not the 1st VSR of the listing
VS user × post -0.726*** -0.372 0.256 -0.710*** -0.00872 -0.00854
(0.139) (0.298) (0.239) (0.228) (0.00838) (0.0129)
Observations 51,794 4,522 4,511 4,522 4,522 4,522
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control treatment-control pair ID FE with standard errors
clustered by pair ID.
Table 6: Reduced-form Guest-level Analysis: DID for VS users (treated
) an d Normal Users (control)
first VSR on the fo cal list in g. About 79.6% of VS users have First-Is-First = 1. We then rerun the DID
analysis for the subsamples of First-Is-First = 1 and First-Is-First = 0, r espectively. Each subsample
includes the VS users with the specific value of First-Is-First and their matched normal users. Regression
results are reported in Panels B and C of Tab l e 6. If t h e above prediction is correct, the VS users with
First-Is-First = 1 should demonstrate greater changes post their own VSR experience, as compared to
those with Fir st -Is-Fi r st = 0.
Indeed, the coefficients reported in Panel B of Table 6 (for First-Is-First = 1) are of a l ar ger magnitude
than those in Panel C (for First-Is-First = 0). The estimates in Panel C are noisier and sometimes
insignificant, in part because only 20.4% of VS users may have seen previous VSR on the focal listing
before posting their own VSR. That being said, even these VS users demonstrate a strong decline of
Airbnb bookings post their own VS experience (-51.62% Column 1) as compared to -61.75% for VS users
with First-Is-First = 1 and -60.07% for all VS users, suggest in g that the VSR left on the focal listings
before our VS users’ own VSR experience have a limited influence on their prior of vicin ity safety before
booking the focal listing and one’s own vicinity safety experience is a still a salient shock ex post. This
points to a significant cross-listing-within-buyer effect of VSR.
We further examine whether VS users subsequently act differently as a function of th e area (high-
income, low-income, minority or white) in which they posted t hei r first VSR. To do so, we gr oup VS users
according to the zip code of the listing for which they posted their first VSR, and proceed to conduct the
DID analysis separ at ely for each of the four su b sampl es.
25
(1) (2) (3) (4)
SAMPLE 1st
vsr h zip 1st vsr l zip 1st vsr w zip 1st vsr m zip
MODEL Logit Logit Logit Logit
VARIABLES h zip h zip w zip w zip
Panel A: Full sample
VS user × post -0.351** 0.316*** -0.628*** 0.682***
(0.160) ( 0. 0990) (0.135) (0.105)
Observations 6,205 14,830 8,880 12,815
Panel B: Subsample VS use’s 1st VSR is the 1st VSR of the l i st in g
VS user × post -0.287* 0.370*** -0.646*** 0.729***
(0.169) (0.111) (0.149) (0.117)
Observations 5,539 11,377 7,181 10,113
Panel C: Subsample VS use’s 1st VSR is not the 1st VSR of the listing
VS user × post -0.887* 0.143 -0.545* 0.494**
(0.496) (0.218) (0.327) (0.247)
Observations 666 3,453 1,699 2,702
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control treatment-control pair ID FE with standard errors
clustered by pair ID. Use the booking sample for users whose 1sr VSR is posted on a property listing in a H, L, W, or M
area, for regressions of (1), (2), (3), or (4), respectively.
Table 7: Reduced-form Guest-level Analysis: DID for VS users by the zi
p code of their VSR bookings
From the interaction term in Panel A of Table 7, it is apparent that VS users exhibi t a positive effect
on subsequent reservations in opposite type of zip codes (Columns 2 and 4) and a negative effect in the
same type of zip codes (Columns 1 and 3). One explanation is that VS users expect a certain level of
safety in the area of their booking, and when they encounter a negative shock, they prefer to avoid that
type of area in subsequent stays.
One may argue that the tendency to avoid the same type of areas is driven by mean reversion rather
than active learning. To address this, we repeat the exercise for the sub sam pl e s with First-Is-First=1 and
First-Is-First=0 separately. Results are reported in Panels B and C of Table 7. Three of the four columns
(Columns 2-4) are consistent with the argument that learning through self-experience is stronger when
the VS user d i d not see any other VSR on the focal listing b efor e her own VSR. The only exception is
when the self VSR is in a high-income zip code (Column 1). In th at case, both VS users of First-Is-First
equal to 1 or 0 decrease their likelihood of booking future Airbnb stays in high-income zip codes (which
is consi st ent with learning) but the coefficient on t h e DID interaction term i s of a larger magnitude for
those with First-Is-First = 0, though the difference is not statisti call y significant. Compared with other
columns, this column has less statistical power because VSR are rarer in high-income zip codes. Overall,
we conclude that the tendency to avoid the type of zip code that triggered VS users’ own VSR is partly
driven by learning from one’s own VSR experience.
To push it further, we reorganize our DID sample into another eight subsamples depending on whether
a VS user has previously had Airbnb stays in the same type of area that triggered her own VSR exper ien ce.
26
For example, if her own VSR experience was in a low-income (L) area, she may or may not have had
Airbnb stays in low-income areas before. This gives us the sub sampl es of HL, LL, LH, and HH, where the
second letter indicates t h e income type of the area that triggered the VS user’s own VSR, and the first
letter represents the income type area of her prior experience. Similarly, we can create the su bs amp l es
of WM, MM, MW, and WW depending on whether the area is primarily white or minority. All matched
normal users belon g to the same subsample as t h e VS users with whom they are paired.
Results are reported in Appendix Tables A5 and A6. In the raw data, we know VSR are more likely
to o ccur i n low-income and minority areas, but listings in these ar eas also account for over 60% of all
Airbnb bookings; thus, the sample sizes of LH and LL are larger than those of HL and HH and the sample
sizes of MW an d MM are larger than those of WM and WW. If we focus on Column 1 of Table A5, the
VS users in LH are the most ‘surprised’ (in terms of reducing future reservations on Airbnb) among the
four L/H groups and the VS users in MW are the most surprised among the four M/W groups. This is
intuitive b ecau se the VS users with at least one L or M stay before t h ei r own VSR experience in H or
W may have high vicinity safety expectations in H or W and are consequently most disappointed when
vicinity safety issues arise in those areas. In contrast, the VS users in HL or WM are not as sur pri sed
(Column 1), likely because they had a lower prior for vicinity safety i n the L or M areas. Nevertheless,
conditional on book i ng on Airbnb, they tend to book listings with fewer VSR following their own VSR.
These p at t ern s confirm the cross-listing-within-buyer effect of self-experience with vicinity safety.
6 Structural Estimatio n and Counterfactual Analysis
So far, reduced-form evidence supports all three information extern
alities of VSR: having any VSR on a
listing may reduce the listing’s price and occupancy, li kely bec aus e VSR discourage future buyers from
booking on that listing (a within -l i st in g-acr oss-bu yer effect); VSR on nearby listings reduce bookings and
price on the focal listing even if listing itself does not have any VSR (a cross-listing-cross-buyer effect),
and a guest that wrote VSR t en ds to avoid other listings/vicinities with any VSR in f u t ur e booking or
avoid booking on Airbnb at all (a cross-listing-within-b u yer effect).
However, it is difficult to use these reduced-form estimates to understand the implications of the
externalities f or hosts and platforms, because they do not address listing competition. In par ti cu l ar,
listings with and without VSR may compete against each other on Airbnb, and all Airbnb list i ngs compete
with the prospect guest’s outside options (includi ng listings on competing short-term-rental platforms,
hotels, bed and breakfasts, a friend’s couch in the destin at ion city, or no t r avel at all). To address
this shortcomi n g, we resort to a structural model that describes how guests choose among competing
short-term-rental listings.
27
6.1 Demand Model and Estimation
We define the market as online short-term entire-home rentals in each z
ip code-month, where Airb nb
and VRBO — the two largest short-term-rental platforms in the US — are assumed to be the only two
platforms that supply this mar ket. Each guest chooses among all Airbnb entire-home listings available
in t he target zip code-month, with the pool of VRBO-only l i st in gs in the same zip code-month as the
outside good.
36
We focus on entire-home listings because only entire-home listings are available on VRBO.
Since our VRBO data period is from 2017/6 to 2019/12, our analysis in this subsection consi d er s Airbnb
entire-home listings from 2017/6 to 2019/12 only.
Following Berry ( 1994) , we assume that each prospective guest chooses an Airbnb entire-home listing
or the outside good (VRBO) so as to maximize her utility from the listing, wh ere the utility associated
with listing j in zip code z of city k and month t can be written as:
U
j,t
= EU
j,t
+ ǫ
j,t
= α
j
+ α
k,t
+ δ · X
j,t
+ β
0
· log( ADR
j,t
) + β
1
· Crime
z,t−1
+ β
2
· LSR
j,t−1
+ β
3
· V SR
j,t−1
+ β
4
· V SRADIU S
j,t−1
+ ǫ
j,t
.
(3)
If ǫ
j,t
conforms to the logistic di str i b ut i on, we can express the market share of listing j at time t as
s
j,t
=
exp(EU
j,t
)
1+
P
m
exp(EU
m,t
)
. Thus:
ln(s
j,t
) − ln(s
0,t
) = EU
j,t
(4)
This is equivalent to regressing t he difference of log market share b etween listing j and the outside
good (ln(s
j,t
) − ln(s
0,t
)) on the attributes of listing j in month t.
37
The right-hand side of Equation 4
is similar to Equation 1 except for two changes: first, we exclude the number of Airbnb listings in the
zip code-month because the discrete choice model already accounts for the size of the choice set; second,
we includ e the log of the listing’s ADR (price). To the extent that log(ADR) might b e endogenous,
we instrument it by the average attributes of entire-home listings within a 0.3-mile radius of the focal
listing in the same zip code-month, following Berr y et al. (1995) . The underlying assumption is that
these so-called ”BLP” instruments are correlated with price becau se of horizontal competition (whereby
competitors’ attributes affect mar gin s) but are excluded because they do not affect the focal listing’s utility
directly. As shown in th e bottom of Table 8, the instruments are strongly correlated with log(ADR),
leading to a first stage F-statistics as high as 288.5.
The O LS and IV estimat ion results of the utility function are report ed in the two columns of Table
8. They suggest that guest reservat ion s are sensiti ve to price, and guests dislike listings with any VSR
36
Listings that co-l ist on Airbnb and VRBO are treated as Airbnb listin g s, thus inside goods.
37
Following Berry (1994), we attemp t ed an alternative specification o f nested logit, where all Airbnb listings in the market
belong t o one nest and the outside good VRBO belongs to another nest. This estimation (with and without instruments)
produces a nesting p a ra met er outside the theoretical range of (0, 1), which leads us to conclude that the nested logit
specification is no better than the simple logit in our setting.
28
(1) (2)
SAMPLE Entire home Entire home
MODEL OLS IV
VARIABLES utility utility
log adr -1.100***
(0.00903)
log
adr iv -6.735***
(1.609)
lag VSR cumu dummy -0.0914*** -0.147***
(0.0121) (0.0199)
lag
LSR cumu dummy -0.101*** -0.204***
(0.0107) (0.0315)
lag VS ehlisting radius pct -0.129** -0.240***
(0.0549) (0.0634)
lag
log crime cumu norm 0.249*** 0.102
(0.0932) (0.105)
lag log review utd 0.0220*** 0.132***
(0.00347) (0. 0317)
log
rating overall 0.304*** 0.373***
(0.0279) (0.0339)
superhost 0.0651*** 0.159***
(0.00424) (0. 0272)
cross
listing 0.0570*** 0.0211
(0.0144) (0.0180)
strict cp -0.0405*** -0.0151*
(0.00534) (0.00905)
Observations 1,014,301 1,014,301
R-squared 0.800 0.789
First stage F statistics 288.5
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control Time*City FE and Prop erty ID FE, with standard errors
clustered by Property ID. The 1st stage of regression uses the average attribute of entire home li st in g s with i n a 0 .3 - mi le
radius area on Airbnb. The attributes include review utd, rating overall, superhost, cross listing, and strict cp. The 1st
stage OLS regression controls Time*City FE and Property ID FE, wi th standard errors clustered by Property ID.
Table 8: Utility estimation
or
LSR, everything else being equal . Based on the IV esti mat es, the guest’s disutility from a listing with
any VSR (as compared to no VSR) is equivalent to 2.2% of average daily rate ($164.7), namely $3.62.
38
And this disutility is comparable to the guest’s disut i l i ty from observing any VSR in 61.25% of all listings
within a 0.3 mile radius of the focal listing.
39
Consistent with our reduc ed -f orm results, these structural
estimates confirm the existence of the “within-listing-cross-buyer” and “cross-li st in g-cr oss-bu yer” effects
of VS R.
Note that the coefficients of V SR
j,t−1
and V SRADIUS
j,t−1
capture how an average prospective
guest in our sample perceives the vicinity safety around listing i at the time of choice. Because VSR
only account for 0.25% of al l guest reviews, the vast majority of the guests may have not experienced
any vicinity safety issues on Airbnb (or have ex perienced but never reported it in a user review) before
38
We derive it by −0.147/(−6.735) · $164.7 = $3.62.
39
We derive it by −0.147/(−0.240) = 61.25%.
29
t. Indeed, if we rerun Equat i on 4 excluding t he VS users identified in our reduced-form analysis, the
estimated coefficients barely change. This means that estimates from Equation 4 can only y iel d reliable
estimates for the within-listing-cross-buyer and cross-listing-cross-buyer effects, bu t not t he cr oss-l i st in g-
within-buyer effect driven by VS users learnin g from their own VSR experience.
Such learning has al read y been captured in our reduced-form guest-level analysis; thus, a key question
is how to incorporate the DID estimate into the structural framework. This is important not only because
this extra “cross-listing-within-buyer” effect is in addition to t h e other two effects we can identify directly
from the vast majority of Airbnb guests but also because self experi en ce sheds light on the guest’s realized
utility when she stays in a listing with vicinity safety issues. Although such realized utility, as reported in
guest r evi ews, has only occurred to a t i ny fraction of Airbnb stays, a fully informed guest should expect
the realized u ti l i ty when she chooses where to stay. As documented by Jin and Sorensen (2006); Allcott
(2011); Train (2015); Rei mers and Waldfogel (2021), the difference between realized and perceived uti l ity
is essential for evaluating how consumer surplus changes under different information regimes.
In parti cu l ar, Figure 4 illustrates the role of perceived and realized utility in consumer surplus.
Consider two demand curves: the inner one represents demand for Airbnb listings under high alert of
VSR while the outer one represents demand under less information about VSR. When prospective guests
perceive the listings to be safe r th an th ey actu al l y are, the market clears at a higher price and more
bookings than in high alert (i.e. P
less info
> P
full info
and Q
less info
> Q
full info
). Those who book under less
information consist of two types of guests: some have a high willingness to pay and would have booked
on Airbnb even i f they know the high alert ex ante, their realized consumer surplus is area A; the others
have a relative low willingness to pay and would not book on Airbnb should t hey know the listings are
actually less safe than they appear, their realized consumer surpl us turns out negative (area B). Hence
the total realized consumer surplus is A-B under less information. In comparison, un d er high alert, the
realized consumer surplus is A+C where C r epr esents the extra consumer surplus that fully -i nf or med
guests could obtain via a lower equilibrium price.
There is another way to arrive at the same conclusion. Let us denote the white trapezoid between
the two demand curves as area D. Under less information, the perceived consume r surplus is A+D while
the realized consumer surplu s is A-B; under fu ll information, both the perceived an d realized consumer
surplus are A+C. Thus, the difference between the reali zed con su mer su rpl u s in full and less information,
(A+C)-(A-B)=C+B, can also be written as the difference between their perceived consumer surplus plus
an adjustment that reflects the shift of the demand curve for all consumers that would purchase under less
information, namely (A+C)-(A+D)+(D+B)=C+B. We will use this alternative expression to compute
consumer surplus changes in the counterfactual analysis.
As shown below, our counterfactual analysis assumes the listin g choices made by VS users after they
wrote VSR on Airbnb reflects their updated interpretation of VSR on all potential listings. Since this
updated interpretation incorporates their true experience in the stay that triggered thei r VSR, we assume
30
!
!"##$%&!'
"
!"##$%&!'
"
#())$%&!'
!
#())$%&!'
!
"
!"
!"##$%&'()$*"+,"%-".$
#$%$&$'
!"
!"##$%&'()$+"/!%0".
$#$$%$($)
!"
'1!!$%&'()$*"+,"%-".$
#$!"
'1!!$%&'()$+"/!%0".
$$#$$%$&$!
$
*!"$#$!"
'1!!$%&'()$+"/!%0".
$($!"
!"##$%&'()$+"/!%0".
$#$!$&$)
$#$!"
'1!!$%&'()$*"+,"%-".
$($!"
!"##$%&'()$*"+,"%-".$
&*+,-.$
$#$/%&!0$1$/%&'0$&$/'&)0
#
$
%
&'()*+,
!"##$%&!'
&'()*+,
#())$%&!'
&
Figure 4: Consumer Surplus Under high alert of VSR (Realized) Or Less Inf or
mation of VSR (Perceived)
it captures the realized utility from VSR. This means th at VS users would have a new β
3
in the utility
function up on their own VSR experience. Changes in their β
3
would capture the difference between
perceived and realized utility driven by VSR.
We calibr at e a new β
3
that would explain why the number of Airbnb bookings of VS users has dropped
60.07% post their own VS experience, according to our guest-level DID analysis ( Table 6 Column 1).
Following the procedure described in Appendix B, our calibration suggests that the VS users must have
changed their β
3
by -2.17, which is more than 14 times of the estimated β
3
of the whole sample (-0.147).
This suggests t hat the cross-listing-wit h in -b uyer effect of VSR – based on a guest’s own VSR experience
– is strong and would have an impact on booking decisions should all normal users interpret the VSR
the same way as these VS users.
6.2 Counterfactual Analy s is
We consider three counterfactual information regimes as compared to the s
tatus quo. The first coun-
terfactual regime is “no disclosure”, where all the VSR are removed in our data. Conceptually, this is
equivalent to an extreme interpretation and implementation of Airbnb’s 2019/12 policy on VSR. The
other extreme regime is “high alert”, where we assume all users react to VSR as much as VS users react
to t h ei r own reported VSR.
Between the two extremes, we explore a third scenario that incorporates VSR so that each listing’s
overall rating is adju st ed to account for the number of VSR of the listing itself as well as listings in a 0. 3-
31
mile radiu s area (“VSR-adjusted ratings”). In particular, we compute a safety score for each observation
by using the reversed percentile of the number of VSR of the listing itself and listings in a 0.3-mile
radius area for each city-month, normalizing it on a r ange from 0 to 10 with a uniform distribution, and
then adjusting the new overall rating as a weighted average of the overall rating and the safety score,
where the overall rating has a weight of 6/7, to account for the 6 ratings originally included by Airbnb of
communication, accuracy, cleanliness, check-in, location, and value. Admittedly, the uniform distribution
assumption is ad hoc. Because the existing distribution of a listing’s overall rating is skewed towards
the high end, this adjustment pulls down the ratings of the listings with VS R for themselves or nearby.
We choose to do so in order to highlight how the platfor m and platform users gain or lose when VSR-
adjusted ratings not only highlight the relative difference between VS and nor mal listings on Airbnb but
also u nd er min e the overall v i ci ni ty safety perception of Airbnb listings relative to the outside good.
We now elaborate how we calculate consumer welfare under each scenario. For the status quo, we
use the IV results in Column 2 of Table 8 to calculate EU
j,t
for each Ai r b nb listing-month, and then
normalize it into US dollars. By definition, this is the guest’s perceived utility. Foll owing Small and Rosen
(1981) and McFad den (2001), in a simple logit model as ours, a consumer’s expected utility from her
utility-maximizing choice depends on the inclusive value of the choice set (namely ln(1 +
P
k
exp(EU
j,t
)).
As depicted in Figure 4, a consumer’s perceived utility may guide her choice of listing ex ante, but
her realized ut i l i ty may deviate from her perceived utility. To measure the realized utility, we use the
calibrated change of β
3
to update β
3
in the utility function (whi le taking everything else unchanged) and
recompute the ut i li ty.
For the counterfactual of no-disclosure, we set all VSR as zero in the (perceived) utility function,
recompute EU
i,t
for each Airbnb entire-home listing, and simulate its market share. This assumes ev-
erything else remain s the same when the platform removes all VS R, which could be violated if listings
adjust prices after t he regime shift. Since the vast majority of our data precede Airbnb’s new review
policy and Ai r b nb seems far from fully implementing the policy, we cannot observe such price adjustments
directly. The reduced-form regressions in Table 2 associate the presence of VSR in VS listings wi t h a
1.48% difference in price. Hence in an alternative calculation, we assume the no-disclosure regime may
enable a 1% price increase in VS listings while the price of normal listings remains unchanged. Thi s gives
us a comparison between no disclosure with price changes versus no disclosure without pr i ce changes.
The high alert counterfactual is equivalent to assuming the guests have full information and therefore
their perceived utility is the same as th e realized utility calculated above. As in the no disclosure regime,
we first simulate the high alert regime without price changes and then introduce an ad-hoc price change
(-1% for VS l i sti n gs) to illust rat e how price changes may alleviate the impact of making all users highly
alert to VSR.
Under t he counterfactual of VSR-adjusted ratin gs, we adjust each listing’s overall rating to account
for the number of VSR of the listi n g itself as well as listings in a 0.3-mile radius area. This calculation
32
assumes the platform has one additional safety rating dimension in addition to the existing 6 rating
dimensions (cleanliness, accuracy, check-in, communication, location, and value). Since we d o not know
how much prices would adjust with such a rating change, we assume an ad-hoc price change (-1% for VS
listings) and si mulate market shares with and without price changes under this counterfactual.
After we compute the perceived and reali zed utility under each scenario, we can quantify changes
in consumer surplus from the status quo to any counterfactual. Defining each zip code-month (z, t) as
a unique market, our analysis includes 9,440 markets in total. We further define market size Mz, t as
the total reserved days for all listings in the market (z, t). Following Reimers and Waldfogel (2021) and
Figure 4, the consumer surplus changes in a single market from the stat us quo to the high alert scenario
can be computed as:
∆CS
z,t
=
M
z,t
β
0
·
ln(1 +
X
j
exp(U
jt,highalert
)) − ln(1 +
X
j
exp(U
jt,perceived
))
+
X
j
((U
jt,perceived
− U
jt,highalert
) · s
j
)
.
(5)
Similar calculations are per f ormed for scenarios of VSR-Adjusted Ratings and No Disclosure.
Table 9 reports the consumer surplus result s under different counterfactuals.
40
The first two rows
refer to no disclosure with and without price changes; the next two rows refer to VSR-adjusted ratings
with and without price changes; the last four rows refer to high al ert with and wit h out price changes
and with and without a “rad iu s effect,” where the radius effect allows for the same updated distaste
for VSR to apply to the VSR in nearby listings as well. To quantify the radius effect, we assume that
the estimat ed coefficient of V SRADIUS (β
4
) would increase in the same proportion as the calibrated
coefficient of V SR (β
3
), should guests extrapolate the high alert of vicinity safety concerns to nearby
VSR in the same way as a listing’s own VSR. This extreme scenari o is designed to illustrate the potential
consequences in case prospective guests become sensitive to any VSR under high aler t.
Table 9 indicates that, under high alert without price changes and without a radius effect, overall
consumer surplus under high alert (without a radius effect) increases by roughly 3.218%, which is slightly
less if we incorporate the hypothetical 1% price drop of VS listings (3.065%) and slightly higher if we
allow a radi u s effect in high alert (4.144% and 3.988%), because high alert h el ps guests reduce their stays
in relatively unsafe listings.
Consumer surplus und er no disclosure declines as compared to the status quo (by 0.032% without
price change and 0.013% with price changes) because consumers cannot use VSR to sort between listings.
Consumer surplus under VSR-adju st ed-r at i ng increases slightly as compared to the status quo (by 0.004%
40
The consumer surplu s reported in Table 9 is for an average user in an average reservation d ay across all 9,940 zip
code-months.
33
∆ Consumer Surplus (versus Status quo)
All
Listings
$
All
Listings
%
No Disclosure w/o P change -2.2 K -0.032%
No Disclosure w/ P change -0.9 K -0.013%
VSR-Adjusted Ratings w/o P change 0.3 K 0.004%
VSR-Adjusted Ratings w/ P change 0.6 K 0.008%
High Alert w/o P change & w/o radi us effect 218.2 K 3.218%
High Alert w/ P change & w/o radi us effect 207.8 K 3.065%
High Alert w/o P change & w/ radi us effect 281.0 K 4.144%
High Alert w/ P change & w/ radius effect 270.4 K 3.988%
Table 9: Counterfactual Analysis: Simulated Changes in Consumer Surp lus in the market
without price change and 0.008% with price changes), because the VSR-ad j us t ed ratings shed light on
VSR, though this change is milder than in the high alert counterfactual.
These estimated changes in consumer surplus are conservative, in part because our definition identifies
only 0.25% of all Airbnb reviews as VSR, and only 4.43% of listings ever had any VSR in our 2015-2019
sample. Because of this, no disclosure only moves 0.74% of market share from VRBO and normal
Airbnb listings to VS listings ( before we take i nto account any price change), and the VSR-adjusted
ratings only move 0.32% of market share away from VS listings. In comparison, the dramatic high alert
counterfactual would move 5.05% of market share away from VS listings, leaving less than 1% of users
choosing VS listings (with or w it h out price change).
Table 10 reports GBV changes based on simulated market shares in each scenario. No disclosure
generates 0.127% more GBV for entire home listings on Airbnb in our sample (assuming the price for
VS listings increases by 1%). Conversely, VSR-adjusted ratings could decrease Airbnb’s GBV by 0.224%
(assuming t he price for VS listing decreases by 1%). In both counterfacturals, the interests of Airbnb and
consumers are misaligned: consumers would prefer more tran sp aren cy through VSR-adjusted ratings but
a GBV-ce ntric Airbnb would prefer no disclosure.
Interestingly, the high alert regime could change Airbnb’s entire-home GBV positi vely or ne gat ively,
ranging from +0.301% ($10.1 million) to -1.304% ($-44 million), depending on whether we allow the price
of VS li st in gs to drop by 1% in response and whether we assume the high alert also applies to t h e VSR for
nearby l i st i ngs through the radius effect. These overall effects are driven by an enormous redistri b ut i on
of GBV from a 82.91-84.38% drop in VS listings to a 3.58-5.38% gain in normal listings and a 5.3-18.18%
gain in VRBO listings. The overall impact on Airbnb GBV could be positive or negative because the
within-platform sorting from VS to normal listings tends to have positive GBV effects for Airbnb but the
cross-platform sorting from Airbnb t o VRBO li st i ngs has negative effects for Airbnb. The radius effects
under high alert strengthens the latter, and thus further hurts Airb nb’s GBV. Similarly, VSR-adjusted
ratings pull down the overall ratings of most Airbnb listings by const ruction. As a result, the cross-
platform sorting generates a negative effect that exceeds the pot ential positive effect of within-platform
34
sorting.
To explore the distributional effects of our counterfactual regi mes, Table 11 breaks down the coun-
terfactual GBV changes in Air bnb listings by the four types of zip codes. Since VS listings are more
likely to locate in low-income and minority zi p codes, no disclosure benefits Airbnb listings in these zip
codes, at the cost of high -i ncome and white zip co d es.
41
In the counterfactual of VSR-adjusted ratings,
Airbnb listings lose GBV in all f our types of zip codes because the way we construct safety scores in the
VSR-adjusted r at i ngs brings down the overall ratings of most Airbnb listings, which motivates guests
to switch away from Ai r bnb. This effect is even stronger for high-income and white zip codes than for
low-income and mi n ori ty zip codes, likely because price tends to be h igh er in high-income and white zip
codes.
∆GBV (versus Status quo)
Airbnb
VS
Listings
%
Airbnb
Normal
Listings
%
Airbnb
Listings
%
VRBO
Listings
%
All
Listings
%
No Disclosure w/o P change 12.360% -0.696% 0.041% -1.471% -0.203%
No Disclosure w/ P change 7.745% -0.329% 0.127% -1.152% -0.080%
VSR-Adjusted Ratings w/o P change -0.682% -0.110% -0.142% 1.208% 0.076%
VSR-Adjusted Ratings w/ P change 3.669% -0.457% -0.224% 0.891% -0. 044%
High Alert w/o P change w/o radius effect -84.382% 5.380% 0.312% 5.298% 1.116%
High Alert w/ P change w/o radius effe ct -83.569% 5.320% 0.301% 5.222% 1.095%
High Alert w/o P change w/ radius effe ct -83.721% 3.635% -1.298% 18.183% 1.846%
High Alert w/ P change w/ radius effect -82.913% 3.580% -1.304% 18.062% 1.821%
Market Share
Status quo 5.984% 83.069% 89.053% 10.947% 100%
No Disclosure w/o P change 6.723% 82.491% 89.214% 10.786% 100%
No Disclosure w/ P change 6.383% 82.796% 89.179% 10.821% 100%
VSR-Adjusted Ratings w/o P change 5.943% 82.978% 88.921% 11.079% 100%
VSR-Adjusted Ratings w/ P change 6.266% 82.690% 88.955% 11.045% 100%
High Alert w/o P change & w/o radi us effect 0.935% 87.539% 88.473% 11.527% 100%
High Alert w/ P change & w/o radiu s effect 0.993% 87.488% 88.481% 11.519% 100%
High Alert w/o P change & w/ radiu s effect 0.974% 86.088% 87.063% 12.937% 100%
High Alert w/ P change & w/ radius effect 1.033% 86.043% 87.076% 12.924% 100%
Table 10: Counterfactual Analysis: Simulated Market Shares and Chan ges in GMV
In contrast, high alert generates strong sorting away from low-income and minority zip codes towards
high-income and white zip codes. When the radius effect is allowed, the GBV of Airbnb listings in white
zip codes also declines because some of t he listi n gs in these zip codes, even if they do not have VSR
themselves, are susceptible to VSR in nearby listings.
Another effort to highlight the distri b ut i on al effects of the counterfactual regimes is zooming into th e
10 zip code-month markets that have the worst VSR in our data. In particular, for each city in th e five
sampled ci t i es, we list all zip code-months that have at least 10 Airbnb listings, sort them by counts of
VSR, and take the top two. Our final sample of 10 worst VSR markets consists of two months of zip
41
If we allow 1% price change, whit e zip codes also enjoy a mod era t e increase of GBV because some VS listings are located
in wh it e zip codes.
35
Airbnb Listings
∆GBV (versus Status quo) H L W M
No Disclosure w/o P change -0.240% 0.472% -0.105% 0.381%
No Disclosure w/ P change -0.047% 0.393% 0.037% 0.337%
VSR-Adjusted Ratings w/o P change -0.130% -0.161% -0.136% -0.157%
VSR-Adjusted Ratings w/ P change -0.313% -0.088% -0.270% -0.117%
High Alert w/o P change w/o radius effect 2.239% -2.653% 1.315% -2.029%
High Alert w/ P change w/o radius effe ct 2.209% -2.636% 1.294% -2.017%
High Alert w/o P change w/ radius effe ct 0.577% -4.184% -0.322% -3.576%
High Alert w/ P change w/ radius effect 0.553% -4.161% -0.338% -3.559%
Table 11: Counterfactual Analysis: Changes in GBV By Four Area Types
∆ Consumer Surplus (versus Status quo)
All
Listings
(Sampled)
$
All
Listings
(Sampled)
%
No Disclosure w/o P change -13 -0.141%
No Disclosure w/ P change -6 -0.062%
VSR-Adjusted Ratings w/o P change 1 0.008%
VSR-Adjusted Ratings w/ P change 2 0.021%
High Alert w/o P change & w/o radi us effect 1173 12.836%
High Alert w/ P change & w/o radius effect 1116 12.210%
High Alert w/o P change & w/ radius effect 1832 20.047%
High Alert w/ P change & w/ radi us effect 1772 19.390%
Table 12: Counterfactual Analysis: Simulated Consumer Surplus changes
(10 samp l e zip code months only)
codes 60624 (Chicago), 10454 (New York City), 90011 (Los Angeles), 70116 ( New Orleans), and 30303
(Atlanta). All of them are located in low-income areas, with 80% of them in minority areas.
As shown in Table 13, no disclosure would increase the market share of VS listings in t hese markets
from 37.81% to 41.41% (without price change), at the expense of normal and VRBO listings, resul t i ng
in a 0.141% decline in consumer surplus. Interestingly, the total Airbnb GBV also decl i nes in these local
markets despite an incr ease in the overall market share of Airbnb listings, because it encourages sorting
from more expensive normal listings to less expensive VS listings. For the same reason, in the high alert
counterfactual, the sorting goes the other way from VS listings to normal and VRBO list i n gs, resulting
in a 12.210%-20.047% gain of consumer su rp l us and a 8.72-20.21% gain of Airbnb GBV in these markets.
Both extremes suggest that for the markets with the worst VSR, the financial interests of Airbnb could
be aligned wit h consumers towards more transparency of VSR in the review sy st em.
In comparison, as shown in Tables 12 and 13, the counterfactual of VSR-adjusted ratings demonstrates
a misalignment between consumers and Airbnb. It generat es slightly more consumer surplus (0.008-
0.021%) but lowers Airbnb’s total GBV from these local markets (-0.198% to -1.348%). The misalignment
occurs because the VSR-adjusted rat i n gs, as we construct them, pull down the average ratings of most
Airbnb listings; as a result, the GBV loss from cross-platform sorting (from Airbnb to VRBO) dominates
the potential GBV increas e from the within-p lat f or m sorting between VS and normal listings. Thi s
36
∆GBV (versus Status quo)
Airbnb
VS
Listing
(sample
zip-month)
%
Airbnb
Normal
Listing
(sample
zip-month)
%
Airbnb
Listing
(sample
zip-month)
%
VRBO
Listing
(sample
zip-month)
%
All
Listing
(sample
zip-month)
%
No Disclosure w/o P change 9.523% -5.488% -1.989% -8.223% -3.207%
No Disclosure w/ P change 6.837% -3.142% -0.816% -6.512% -1.929%
VSR-Adjusted Ratings w/o P change -0.405% -0.135% -0.198% 3.327% 0.491%
VSR-Adjusted Ratings w/ P change 2.268% -2.447% -1.348% 1.544% -0.783%
High Alert w/o P change w/o radius effect -79.532% 50. 523% 20.212% 30.800% 22.281%
High Alert w/ P change w/o radius effect -78.495% 49.716% 19.834% 30.402% 21.899%
High Alert w/o P change w/ radius effect -81.209% 36.372% 8.967% 154.272% 37.359%
High Alert w/ P change w/ radius effect -80.315% 35.777% 8.720% 153.045% 36.920%
Market Share
Status quo 37.808% 55.336% 93.144% 6.856% 100%
No Disclosure w/o P change 41.409% 52.299% 93.708% 6.292% 100%
No Disclosure w/ P change 39.993% 53.598% 93.591% 6.409% 100%
VSR-Adjusted Ratings w/o P change 37.655% 55.261% 92.916% 7.084% 100%
VSR-Adjusted Ratings w/ P change 39.056% 53.982% 93.038% 6.962% 100%
High Alert w/o P change & w/o radius effect 7.739% 83.294% 91.032% 8. 968% 100%
High Alert w/ P change & w/o radius effect 8.213% 82. 847% 91.060% 8.940% 100%
High Alert w/o P change & w/ radius effect 7.105% 75. 463% 82.567% 17.433% 100%
High Alert w/ P change & w/ radius effect 7.518% 75.134% 82.651% 17.349% 100%
Table 13: Counterfactual Analysis: Simulated Market Shares and GM V Changes
(10 samp l e zip code months only)
misalignment is similar to what we have seen across all markets in the data ( Tables 9 and 10).
7 Conclusion
Taking vicin ity safety reviews as an example of critical feedback on Airb
nb, we show that VSR not only
have the classical effect of guiding future buyers towards listings without VSR, but they also generate
spillovers for nearby listings and motivate guests that have written VSR to lear n and update their
understanding of VSR on other listings. As a result, they are less likely to book future st ays on Airbnb,
and when they do book, they tend to book listings with out VSR and in areas with fewer official crime
reports and fewer VSR.
Using a structural approach to account for listing competition on and off Airbnb, we show that
expanding the disclosure of VSR may disproportionately affect hosts in low i ncome and minority areas,
and that a GBV-centric platform may prefer to limit the disclosure of VSR altogether, even thou gh the
aggregate surplus of guests appears to increase when the VSR are instead emphasized to alert prospective
guests.
Combined, our findings highlight the economic incentives behind a platform’s information policy
regarding critical feedback. On the one hand, the platform’s general booking value (GBV) may stand
to decrease under the high alert of VSR if the alert ed guests become sensitive to all VSR; and listings
37
in low-income and minority zip co d es may stand to lose a dispropor t i onat e share of revenues than their
counterparts in high-income and white zip codes. On the other hand, consumer surplus under the high
alert regime is higher than under the status quo and the no-disclosure regimes. The platform thus faces
a tradeoff as far as generatin g higher revenues an d attracting hosts in low-income and minority areas on
the one hand, and providing additional value to its buyers on the other.
To the extent that being inclusive is one motivation behind Airbnb’s new review policy, our findings
suggest that the policy, if fully implemented, may have some unintended consequences on consumers and
listings without VSR. How to balance the economic i nterests of all users is a challenge to platforms and
policy makers that strive t o maximi ze social welfare. One potential solution is that the platform may
import external information about vicinity safety an d p r esent i t as an alternative to VSR for each listing.
Unfortunately, crime statistics (when available) may not fully capture all of the safety concerns a guest
may have in mind at the ti me of booking. Another alternati ve is to incorporate VSR into the overall
ratings of a listing (as in our “adjusted rating” counterfactual); our counterfactual analysis suggests that
this moderate regime may have a smal l negative effect on Airbnb G BV while slightly boost i n g consumer
surplus. The magnitudes of these predictions depend on how we normalize VSR and weigh them in the
overall ratings. How to adju st ratings in line with the platform or the social planner’s objective certainly
merits future research.
There are a number of limitations to our analyses. First, guest reviews in our data d o not include
potential responses from hosts. Second, in the guest-level analysis, we only observe a guest’s reservation
provided that they have made any Airbnb r eservations in the five maj or US cities we consider and posted
a review on Airbnb. If VS users are more vocal and thus more likely to post subsequent reviews after
their first VSR, then our findings underestimate the negative effects on their subsequent b ooking activity;
if, however, VS users are less likely to post subsequent reviews, then our findings overestimate the effects.
Third, we do not have listing reviews for VRBO listings nor do we have hotel booking data to explicitly
consider hotels as an outside option in our utility estimation.
42
Fourth, our main analysis ends in 2019/12,
the same month when Airbnb announced its new review policy. Because we do not know exactly how
Airbnb implements its new policy,our counterfactual simulations are hypothetical.
These limitations suggest additi onal directions for future work. In particul ar , VRBO does not have
a policy of discouraging reviews about the vicinity of listings, as Airbnb introduced in 2019/12. This
may facilitate an interesting comparison between VRBO and Airbnb listings in the same locales, given a
sample period that encompasses Airbnb’s introduction of its new review policy. In addition, one welfare
aspect that is difficult to quantify but may be relevant for Airbnb is the long-run entry and exit of
users. As shown in our counterfactual analysis, a policy that encourages and highlights VSR could
disproportionately hurt Airbnb hosts in relatively unsafe neighborhoods. In the long run, this could l ead
42
Hotels, in particular, may offer enhanced safety measures to their guests through security arra n g ements and by having
door and security staff.
38
to a smaller choice set for guests, drive away some types of hosts and guests, and affect the economic
parity across different neighborhoods.
References
Allcott, H. (
2011): “Consumer s’ Perceptions and Misperceptions of Energy Cost s, ” American Eco-
nomic Review, 101, 98–104.
Bajari, P. and A. Hortacsu (2004): “Economic Insights from Internet Auctions, ” Journal of Eco-
nomic Literature, 42, 457–486.
Barach, M. A., J. M. Golde n, and J. J. Horton (2020): “Steering in online markets: th e role of
platform incentives and credibility,” Management Sci ence, 66, 4047–4070.
Berry, S. (1994): “Estimating Discrete Choice Models of Product Differentiation,” RAND Journal of
Economics, 25, 242–262.
Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile Prices in Market Equilibrium,” Econo-
metrica, 63, 841–890.
Bolton, G., B. Greiner, and A. Ock e nfe ls (2013): “Engineerin g trust : reciprocity in the produc-
tion of reputati on information, ” Management science, 59, 265–285.
Cabral, L. and L. Li (2015): “A dollar for your thoughts: Feedb ack-conditional rebates on eBay,”
Management Science, 61, 2052–2063.
Chakravarty, A., Y. Liu, and T. Mazumdar ( 2010) : “The differential effects of online word-of-
mouth and critics’ reviews on pre-release movie evaluation,” Journal of Interactive Marketing, 24,
185–197.
Chevalier, J. A., Y. Dover, and D. Mayzlin (2018): “Channels of Impact: User Revi ews When
Quality Is Dynamic and Managers Respond,” Marketing Science, 37, 688–709.
Chevalier, J. A. and D. Mayzlin (2006): “The effect of word of mouth on sales: Online book reviews,”
Journal of marketing research, 43, 345–354.
Conitzer, V., N. Immorlica, J. Letchford, K. Munagala, and L. Wagman (2010): “False-
name-proofness in social networks,” in International Workshop on Internet and Network Economics,
Springer, 209–221.
Conitzer, V. and L. Wagman (2014): “False-name-proof voting over two alternatives,” International
Journal of Game Theory, 43, 599–618.
Dai, W. D., G. Jin, J. Lee, and M. Luca (2018): “Aggregation of consumer ratings: an application
to Yel p. com,” Quantitative Marketing and Economics, 16, 289–339.
Dellarocas, C. and C. A. Wood (2008): “The sound of si len ce in online feedback: Estimating trading
risks in the presence of reporting bias,” Management Science, 54, 460–476.
39
Dhaoui, C., C. M. Webster, and L. P. Tan (2017): “Social media sentiment analysis: lexicon versus
machine learning,” Journal of Consumer Marketing.
Einav, L., C. Farronato, and J. Levin (2016): “Peer-to-peer markets,” Annual Review of Economics,
8, 615–635.
Filieri, R., E. Raguseo, and C. Vitari ( 2021) : “Extremely negative ratings and online consumer
review helpfulness: the moderating role of product quality signals,” Journal of Travel Research, 60,
699–717.
Fradkin, A., E. Grewal, and D. Holtz (2021): “Reciprocity and unveiling in two-sided reputation
systems: Evid ence from an experiment on airbnb,” Marketing Science.
Fradkin, A., E. Grewal, D. Holtz, and M. Pearson (2015): “Bias and Reciprocity i n Online
Reviews: Eviden ce From Field Experiments on Airbnb.” EC, 15, 15–19.
Fradkin, A. and D. Holtz (2023): “Do In centives to Review Help the Market? Evidence from a Field
Experiment on Airbnb,” Marketing Science, forthcoming.
Gurran, N. and P. Phibbs (2017): “When tourists move in: how should urban planners respond to
Airbnb?” Journal of the American planning association, 83, 80–92.
Han, W. and X. Wang (2019): “Does Home Sharing Impact Crime Rate? A Tale of Two Cities.” in
ICIS.
Han, W., X. Wang, M. Ahsen, and S. Wattal (2020): “Do es Home Sharing Impact Crime Rate?
An E mpi r ical Investigation,” An Empirical Investigation (January 16, 2020).
Hui, X., T. J. Klein, K. Stahl, et al. (2021): “When and Why Do Buyers Rate in Onli ne Markets?”
Tech. rep., University of Bonn and University of Mannheim, Germany.
Hutto, C. and E. Gilbert (2014) : “Vader: A parsimonious rule-based model for sentiment analysis
of social media text,” in Proceedings of the International AAAI Conference on Web and Social Media,
vol. 8.
Jia, J., G. Z. Jin, and L. Wagman (2021): “Platform as a Rule Maker: Evi den ce from Airbnb’s
Cancellation Policies,” Working Paper.
Jia, J. and L. Wagman (2020): “Platform, Anonymity, and Illegal Actors: Evidence of Whac-a-Mole
Enforcement from Airbnb,” The Journal of Law and Economics, 63, 729–761.
Jin, G. Z. and A. Sorensen (2006): “Information and Consumer Choice: The Value of Publici ze d
Health Plan Ratings,” Journal of Health Economics, 25, 248–275.
Kim, J.-H., T. C. Leung, and L. Wagman ( 2017) : “Can restricting property u s e be value enhancing?
Evidence from shor t -t erm rental regulation,” The Journal of Law and Economics, 60, 309–334.
Klein, T. J., C. Lambertz, G. Spagnolo, and K. O. Stahl (2009): “The actual structur e of
eBay’s feedback mechanism and early evidence on the effects of recent changes,” International Journal
of Electronic Business, 7, 301–320.
Klein, T. J., C. Lambertz, and K. O. Stahl (2016): “Market transparency, adverse selection, and
moral hazard,” Journal of Political Economy, 124, 1677–1713.
40
Kovbasyuk, S. and G. Spagnolo (2018): “Memory and markets,” Available at SSRN 2756540.
Li, L., S. Tadelis, and X. Zhou (2020): “Buying reputat i on as a signal of quality: Evidence from an
online marketplace,” The RAND Journal of Economics, 51, 965–988.
Li, L. and E. Xiao (2014): “Money talk s: Rebate mechanisms in reputation sy st em design,” Manage-
ment Science, 60, 2054–2072.
Liu, M., E. Brynjolfsson, and J. Dowlatabadi (2021): “Do digital platforms reduce moral hazard?
The case of Uber and taxis,” Management Science.
Luca, M . (2017): “Designing Online Marketplaces: Trust and Reput at i on Mechanisms,” Innovation
Policy and the Economy, edited by Shane Greenstein, Josh Lerner and Scott Stern, 17.
Luca, M. and G. Zervas ( 2016): “Fake it till you make it: Reputation, competi t i on, and Yelp review
fraud,” Management Science, 62, 3412–3427.
Maldonado-Guzm
´
an, D. J. (2020): “Airbnb and crime in Barcelona (Spain): testing the relationship
using a geographically weighted regression,” Annals of GIS, 1–14.
Mayzlin, D., Y. Dover, and J. Chevalier (2014): “Promot i onal reviews: An empirical investigation
of on li n e review manipulation,” American Economic Review, 104, 2421–55.
McFadden, D. (2001): “Economic Choices,” American Economic Review, 91, 351–378.
Monroe, B. L., M. P. Colaresi, and K. M. Quinn (2008): “Fightin’words: Lexical feature selection
and evaluation for identifying the content of political conflict,” Political Analysis, 16, 372–403.
Nieuwland, S. and R. Van Melik (2020): “Regulating Airbnb: how cities deal with perceived negative
externalities of short-term rentals,” Current Issues in Tourism , 23, 811–825.
Nosko, C. and S. Tadelis (2015): “The limits of reputation in platform markets: An empirical analysis
and field exper iment,” Tech. rep., National Bureau of Economic Research.
Proserpio, D. and G. Zervas (2017): “Online Reputation Management: Estimating the Impact of
Management Responses on Consumer Reviews,” Marketing Science, 36, 645–665.
Reimers, I. and J. Waldfogel (2021): “Digitization and pre-purchase information: the causal and
welfare impacts of reviews and crowd ratings,” American Economic Review, 111, 1944–71.
Romanyuk, G. and A. Smolin (2019): “Cream skimming and information design in matching markets,”
American Economic Journal: Microeconomics, 11, 250–76.
Roth, J. J. (2021): “Home Sharing and Crime Across Neighborhoods: An Analysis of Austin, Texas,”
Criminal j u s tice review, 46, 40–52.
Schuckert, M., X. Liu, and R. Law (2015): “Hospitality and tourism online reviews: Recent trends
and future direct i on s,” Journal of Travel & Tourism Marketing, 32, 608–621.
Small, K. A. and H. A. Rosen (1981): “Applied Welf are Economics with Discrete Choice M odels,”
Econometrica, 49, 105–130.
41
Staats, B. R., H. Dai, D. Hofmann, and K. L. Milkman (2017): “Motivating process compliance
through ind iv i d ual electronic monitoring: An empirical examination of hand hygiene in healthcare,”
Management Science, 63, 1563–1585.
Suess, C., K. M. Woosnam, and E. Erul (2020): “Stranger-danger? Understandin g the moderating
effects of children in the household on non-hosting residents’ emotional solidarity with Airbnb visitors,
feeling safe, and support for Airbnb,” Tourism Management, 77, 103952.
Tadelis, S. (2016): “Reputation and feed back systems in online platform markets,” Annual Review of
Economics, 8, 321–340.
Train, K. (2015): “Welfare Calcu lat i on s in Discrete Choice Models When Anticipated and Experienced
Attributes Differ: A Guide with Examples, ” Journal of Choice Modelli ng, 16, 15–22.
Wagman, L. and V . Conitzer (2008): “Optimal Fal se-Nam e-Pr oof Voti ng Rules with Costly Voting.”
in AAAI, 190–195.
Xu, Y.-H., L. Pennington-Gray, and J. Kim (2019): “The shar i ng economy: A geographically
weighted regression approach to examine crime and the shared lodging sec t or, ” Journal of travel re-
search, 58, 1193–1208.
Zervas, G., D. Proserpio, and J. W. Byers (2021): “A first look at online reputation on Airbnb,
where every stay is above average,” Marketing Letters, 32, 1–16.
Zhang, H., W. Gan, and B. Jiang (2014): “Machine learning and lexicon based methods f or sentiment
classification: A survey,” in 2014 11th web inform ati on system and application conference, IEEE, 262–
265.
Zhang, S., D. Lee, P. Singh, and T. Mukhopadhyay (2021): “EXPRESS: Demand Interactions in
Sharing Economies: Evidence from a Natural Experiment Involving Airbnb and Uber/Ly ft , ” Journal
of Marketing Research, 00222437211062172.
Online Appendix A: Supplemental Figures and Tables
42
Figure A1: Distribution for keywords of vicinity safety review
Figure A2: Distribution for keywords of listing safety review
43
Figure A3: Distribution for keywords of vicinity safety review in H & L zip codes
Figure A4: Distribution for keywords of vicinity safety review in W & M zip codes
44
Vicinity safety keywords:
‘abuse’, ‘ally way’, ‘and run’, ‘appalling’, ‘assaulted’,
‘bad neighborhood’, ‘bit scary’, ‘blighted’, ‘burglar bars’,
‘creepy’, ‘dangerous neighborhood’, ‘not safe’,
‘dicey’, ‘do drugs’, ‘drug addict’, ‘drug dealers’, ‘dr ug use’,
‘drug users’, ‘drugs’, ‘extremely dangerous’, ‘fights’, ‘gang’,
‘government housing’, ‘gunpoi nt’, ‘harassed’,
‘homeless’, ‘incred i bl y unsafe’, ‘loud music’,
‘meth’, ‘mugged’, ‘pretty dangerous’, ‘rough area’,
‘run down’, ‘shady characters’, ‘shady neighborhood’,
‘shooting up’, ‘tenement area’, ‘uneasy’,
‘unsafe’, ‘very sketchy’, ‘yelling’
Listing safety keywords:
‘alarming’, ‘t hr eat en in g’ , ‘brown stains’, ‘cigarettes’,
‘dangerous’, ‘dan gli n g’ , ‘peril’, ‘disgusted’, ‘disgustingly’,
‘drugs’, ‘dump ’, ‘excrement’, ‘exposed pipe’,
‘felt violated ’ , ‘filthy’, ‘fire hazards’,
‘something fishy’, ‘very poor’, ‘mold’, ‘grime’,
‘not maintained’, ‘gross’, ‘harass’ ‘hazard’, ‘hazards’,
‘highly uncomfor t abl e’ , ‘safety concern’, ‘illegally’, ‘infested’,
‘inhospitable’, ‘loosely attached’, ‘meth’, ‘mice’, ‘naked’,
‘no instruction s’ , ‘not provided’, ‘scam’, ‘unhygienic’,
‘roaches’, ‘sanitation issues’, ‘shocked’,
‘slippery tub ’, ‘squalid’, ‘stained’, ‘sticky’, ‘terrible condition’,
‘threatened’, ‘ u nan n oun ced’ , ‘unlocked door’, ‘worst’,
Vicinity locati on keywords:
‘neighborhood’, ‘area’, ‘feel’, ‘felt’, ‘night’, ‘location’,
‘walking’, ‘people’, ‘seemed’, ‘outside’, ‘looked’, ‘looks’, ‘late’,
‘surrounding’, ‘located’, ‘neighb ou r hood’ , ‘walked’, ‘areas’,
‘feeling’, ‘st reet s’ , ‘street’, ‘outside’, ‘parking’, ‘neighbors’
Negative keywords:
‘hardly’, ‘n ever’, ‘scarcely’, ‘seldom’, ‘barely’, ‘no’, ‘not’,
‘without’, ‘ not h i ng’ , ‘nobody’, ‘neither’, ‘nor’, ‘none’
Table A1: Vicinity and listing safety review keywords
45
VARIABLES Explanation
occupancy rate
Occupancy Rate = Count of Reservation Days / (Count of Reservation Days + Count of Available Days).
occupancy rate dummy occupancy r at e du mmy = 1 of occupancy rate is greater than 0, otherwise equals to 0.
adr
Price per night in USD. ADR = Total Revenue / Booked Nights. Includes cleaning fees.
No.
of reservations # of reservations in a property list i ng i n a month.
No. of reservationdays # of days reserved in a property listing in a month.
lag VSR cumu dummy Lag of the VSR dummy, which indicates whether a li st i ng ever have any VSR bef or e a report in g month.
lag
LSR cumu dummy Lag of the LSR dummy, which indicates whether a listing ever have any LSR before a reporting month.
lag VSR cumu Lag of t h e VS R cumu, which indicates the number of cumulative sum of VSRs before a reporting month.
lag LSR cumu Lag of the VSR cumu, which i n di cat es t h e number of cumulative sum of VSRs before a reporting month.
lag VS listing radius pct Lag of the percentage of listings that ever have any VSR in a 0.3 mile radius area before a reporting month.
safety
score
Calculated by using the reversed percentile of th e number of VSR of the listing it sel f an d l i st i ngs
in a 0.3-mile radius area for each city-month,normalizing it on a range from 0 to 10 with a unif orm distribution.
ratingoverall The overall rat i ng scor e shown to the users. Normalized on a range from 0 to 10 with a uniform distribution.
ratingcommunication The commucation rating score shown to the use rs . Normalized on a range from 0 to 10 with a uniform distribut i on.
ratingaccuracy The accuracy rat in g scor e s hown to the users. Normalized on a range from 0 to 10 wi th a uniform distribution.
ratingcleanliness The cleanliness rating score shown to the users. Normalized on a range from 0 to 10 with a uniform distribution.
ratingcheckin The check-in rating score shown to the users. Normalized on a range from 0 to 10 with a uniform distribution.
ratinglocation The location rating score shown to the users. Normalized on a range from 0 to 10 with a uniform distribution.
ratingvalue The value rating score shown to the users. Normalized on a range from 0 to 10 with a unif or m distribution.
review
utd # of r ev i ews i n a property listing until a reporting month.
No. of listing zip # of listings in a zip code.
cross listing Dummy variable indicates whether the property listing is also listed on VRBO in a reporting month.
superhost Dummy variable indicates whether the property listing is hosted by a superhost in a reporting month.
strict
cp Dummy variable indicates whether the property listing has a strict cancellat i on positl i ty in a reporting month.
ave wordcount cumu review Average word counts for cumulative reviews in a reporting month.
median income zip Median income for hou seh old s i n a zip code (using 2014 census data).
population zip # of population in a zip code (using 2014 census data).
white
pct zi p Percentage of white population in a zip code (using 2014 census data).
h zip
Dummy variable shows if the median i ncome of a zip code exceeds the city’s median income level.
w zip
Dummy variable signifies if the white population percentage in a zip code surpasses the city’s median white population percentage.
crime cumu # of cumulative crime until a reporting month in a zip code.
crime
cumu norm Normalized # of cumulative crime until a reporting month in a zip code by the population in th e same zip code.
r sentiL cumu ave
The average sentiment score is calculated for cu mulative reviews up to a reporting month, using the Lexicon approach.
r sentiN cumu ave
The average sentiment score is calculated for cu mulative reviews up to a reporting month, using the NLP app roach.
Table A2: Variable Definition
46
Panel A: H Panel B: L Panel C: W Panel D: M
VARIABLES mean p50 N mean p50 N mean p50 N mean p50 N
occupancy rate 0.56 0.64 1,484,474 0.57 0.65 1,381,764 0.56 0.64 1,716,77 4 0.57 0.65 1,149,464
occupancy rate dummy 0.84 1.00 1,484,474 0.86 1.00 1,381,764 0.85 1.00 1,716,774 0.85 1.00 1,149,464
adr 190.38 149.00 1,484,474 137.08 103.62 1,381,76 4 188.33 147.48 1,716,774 129.37 98.07 1,149,464
No. of reservations 3.65 3.00 1,484,474 3.91 3.00 1,381,764 3.71 3.00 1,716,774 3.87 3.00 1,149,464
No. of reservationdays 13.86 14.00 1,484,4 7 4 14.48 15.00 1,381,764 13.99 14.00 1,716,774 14.41 15.00 1,149,464
lag VSR cumu dummy 0.02 0.00 1,484,474 0.07 0.00 1,381,764 0.03 0.00 1,716,774 0.06 0.00 1,149,464
lag LSR cumu dummy 0.05 0.00 1,484,474 0.05 0.00 1,381,764 0.05 0.00 1,716,774 0.05 0.00 1,149,464
lag VSR cumu 0.03 0.00 1,484,474 0.09 0.00 1,381,764 0.04 0.00 1,716,774 0.08 0.00 1,149,464
lag LSR cumu 0.05 0.00 1,484,474 0.06 0.00 1,381,764 0.06 0.00 1,716,774 0.06 0.00 1,149,464
lag VS listing radius pct 0.05 0.02 1,484,474 0.09 0.05 1,381,764 0.05 0.03 1,716,774 0.09 0.05 1,149,464
safety sco re 5.62 6.15 1,484,474 4.26 3.67 1,381,764 5.31 5.76 1,716,774 4.45 3 .8 7 1,149,464
ratingoverall 9.23 9.60 1,484,474 9.13 9.50 1,381,764 9.23 9.60 1,716,774 9.11 9.50 1,149,464
ratingcommunication 9.57 10.00 1,484 , 4 7 4 9.52 10.00 1,381,764 9.57 10.00 1,716,7 7 4 9.52 10.00 1,149,464
ratingaccuracy 9.41 10.00 1,484,474 9.35 10.00 1,381,764 9.41 10.00 1,716,774 9.34 10.00 1,149,464
ratingcleanliness 9.21 10.00 1,484,474 9.11 10.00 1,381,764 9.21 10.00 1,716,774 9.09 10.00 1,149,464
ratingcheckin 9.55 10.00 1,484,474 9.51 10.00 1,381,764 9.55 10.00 1,716,774 9.50 10.00 1,149,464
ratinglocation 9.53 10.00 1, 4 84 , 4 7 4 9.08 9.00 1,381,764 9.52 10.00 1,716,774 9.01 9.00 1,149,464
ratingvalue 9.22 10.00 1,484,474 9.17 10.00 1,381,764 9.22 10.00 1,716,774 9.16 10.00 1,149,464
review utd 33.07 14.00 1,48 4 ,4 7 4 34.40 15.00 1,381,764 33.65 15.00 1,716,774 33.80 15.00 1,149,464
No. of listing zip 502.69 463.00 1,484,474 581.47 428.00 1,381,764 609.92 512.00 1,716,774 437.24 372.00 1, 1 49 , 4 6 4
cross listing 0.02 0.00 1,484,474 0.02 0.00 1,381,764 0.03 0.00 1,716,774 0.02 0.00 1,149,464
superhost 0.25 0.00 1,484,474 0.22 0.00 1,381,764 0.24 0.00 1,716,774 0.22 0.00 1,149,464
strict cp 0.50 0.00 1,484,474 0.49 0.00 1,381,764 0.51 1.00 1,716,774 0.48 0.00 1,149,464
ave wordcount cumu review 53.93 50.67 1,484,474 53.7 2 50.15 1,381,764 54.40 51.18 1,716,774 52.98 49.25 1,149,464
median incom e zip 75,865 71,278 1,484,474 37,121 35,112 1,381,76 4 69,745 68,346 1,716,774 38,432 37,116 1,149,46 4
population zip 43,535 41,453 1,484,474 53,124 51,791 1,381,764 44,706 38,752 1,716,774 53,313 54,440 1, 14 9 , 4 6 4
white pct zip 0.68 0.72 1,484,474 0.37 0.32 1,381,764 0.69 0.72 1,716,774 0.28 0.30 1,149,464
h zip 1.00 1.00 1,484,474 0.00 0.00 1,381,764 0.78 1.00 1,716,774 0.13 0.00 1,149,464
w zip 0.90 1.00 1,484,474 0.27 0.00 1,381,764 1.00 1.0 0 1,716,774 0.00 0.00 1,149,464
crime cumu 14,737 7,735 1,484,474 24,483 12,205 1, 38 1 , 7 6 4 20,925 8,347 1,716,774 17,211 1 1 ,6 3 7 1, 1 4 9 ,4 6 4
Table A3: Summary Statistics of Airbnb Listings by Four Area Types
47
Panel E: Entire Home Panel F: Private Room Panel G: Shared Room Panel H: Hotel Room
VARIABLES mean p50 N mean p50 N mean p50 N mean p50 N
occupancy rate 0.58 0.67 1,745,432 0.55 0.63 1,016,553 0.44 0.41 94,722 0.46 0.43 9,531
occupancy rate dummy 0.87 1.00 1,745,432 0.83 1.00 1,016,553 0.77 1.00 94,722 0.87 1.00 9,531
adr 212.81 170.46 1,745,432 91.67 76.25 1,016,553 58.23 39.36 94,722 197.16 153.87 9,531
No. of reservations 3.86 3.00 1,745,432 3.65 3.00 1,016,553 3.31 2.00 94,722 5.79 5.00 9,531
No. of reservationdays 14.45 15.00 1,745,4 3 2 13.93 14.00 1,016,553 11.44 9.00 94,722 12.53 11.00 9,531
lag VSR cumu dummy 0.05 0.00 1,745,432 0.04 0.00 1,016,553 0.04 0.00 94,722 0.06 0.00 9,531
lag LSR cumu dummy 0.06 0.00 1,745,432 0.04 0.00 1,016,553 0.03 0.00 94,722 0.03 0.00 9,531
lag VSR cumu 0.06 0.00 1,745,432 0.06 0.00 1,016,553 0.06 0.00 94,722 0.11 0.00 9,531
lag LSR cumu 0.07 0.00 1,745,432 0.05 0.00 1,016,553 0.03 0.00 94,722 0.04 0.00 9,531
lag VS listing radius pct 0.06 0.03 1,745,432 0.07 0.04 1,016,553 0.08 0.05 94,722 0.06 0 . 0 4 9,531
safety sco re 4.90 5.05 1,745,432 5.10 5.23 1,016,553 4.83 4.87 94,722 4.52 4.38 9,531
ratingoverall 9.26 9.60 1,745,432 9.09 9.50 1,016,553 8.74 9.30 94,722 9.03 9.40 9,531
ratingcommunication 9.61 10.00 1,745 , 4 3 2 9.48 10.00 1,016,553 9.18 10.00 94,722 9.38 10.00 9,531
ratingaccuracy 9.47 10.00 1,745,432 9.29 10.00 1,016,553 8.91 10.00 94,722 9.26 10.00 9,531
ratingcleanliness 9.27 10.00 1,745,432 9.02 10.00 1,016,553 8.64 9.00 94,722 9.24 10.00 9,531
ratingcheckin 9.59 10.00 1,745,432 9.47 10.00 1,016,553 9.16 10.00 94,722 9.47 10.00 9,531
ratinglocation 9.43 10.00 1, 7 45 , 4 3 2 9.15 10.00 1,016,553 8.87 9.00 94,722 9.46 10.00 9,531
ratingvalue 9.25 10.00 1,745,432 9.14 10.00 1,016,553 8.86 9.00 94,722 9.01 9.00 9,531
review utd 34.28 16.00 1,74 5 ,4 3 2 34.20 14.00 1,016,553 19.24 8.00 94,722 21.77 7.00 9,531
No. of listing zip 562.51 481.00 1,745,432 513.44 401.00 1,016,553 433.98 339.00 94,722 504.51 449.00 9,531
cross listing 0.04 0.00 1,745,432 0.00 0.00 1,016,553 0.00 0.00 94,722 0.00 0.00 9,531
superhost 0.25 0.00 1,745,432 0.22 0.00 1,016,553 0.11 0.00 94,722 0.13 0.00 9,531
strict cp 0.53 1.00 1,745,432 0.43 0.00 1,016,553 0.52 1.00 94,722 0.41 0.00 9,531
ave wordcount cumu review 55.03 51.60 1,745,432 52.9 4 49.48 1,016,553 42.84 38.70 94,722 37.27 33.14 9,531
median incom e zip 59,726 54,023 1,745,432 53,568 47,050 1,016,55 3 48,929 40,873 94,722 60,291 56,337 9,531
population zip 44,465 38,752 1,745,432 54,260 54,440 1,016,553 52,003 48,852 94,722 35,315 30,648 9,531
white pct zip 0.56 0.61 1,745,432 0.48 0.46 1,016,553 0.45 0.44 94,722 0.55 0.60 9,531
h zip 0.59 1.00 1,745,432 0.41 0.00 1,016,553 0.37 0.00 94,722 0.59 1.00 9,531
w zip 0.68 1.00 1,745,432 0.48 0.00 1,016,553 0.43 0.0 0 94,722 0.74 1.00 9,531
crime cumu 22,154 9,569 1,745,432 15,269 9,806 1,016,553 13,058 9,2 6 0 94,722 29,232 13,756 9,531
Table A3: Summary Statistics of Airbnb Listings By Fou r Listing Types
48
(1) (2) (3) (4) (5) (6) (7) (8)
SAMPLE EH PR SR HR EH PR SR HR
MODEL OLS OLS OLS OLS OLS OLS OLS OLS
VARIABLES
log
occupancy
rate
log
occupancy
rate
log
occupancy
rate
log
occupancy
rate
log adr log adr log adr log adr
lag VSR cumu dummy -0.01 6 1* * * -0.0210*** -0.0301*** -0.0372** -0.0131*** -0.0171*** -0.023 5 * -0.0164
(0.00171) (0.00249 ) (0.00915) (0.0181) (0.00275) (0.003 7 6 ) (0.0130) (0.0408)
lag
LSR cumu dummy -0. 0 2 36 * * * -0 . 0 3 08 * * * -0 . 0 28 9 * * * -0.0621** -0.0201*** -0.00809** -0.0102 0.0269
(0.00158) (0.00265 ) (0.0108) (0.0285) (0.00248) (0.00406) (0.0168) (0.0410)
lag VS listing radius pct - 0 .0 1 0 7 * * * -0.00419 -0.0282** 0.0154 -0.00485 -0.0115** -0.0154 0.0310
(0.00338) (0.00396 ) (0.0130) (0.143) (0.00540) (0.00571) (0.0277) (0.2 0 5 )
lag
log crime cumu norm 0.0439*** 0.120** * 0.169*** -0.411** -0.06 9 4 * * * -0.00732 -0.302** 0.969***
(0.00996) (0.0153) (0.0617) (0.167) (0.0156) (0.0233) (0.136) (0.320)
Observations 1,745,432 1,01 6 ,5 5 3 94,722 9,531 1,745,432 1,016,553 94,722 9,531
R-squared 0.540 0.581 0.593 0.621 0.887 0.858 0.894 0.918
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control Time*City FE and Property ID FE, with standard errors clustered by Property ID.
Table A4: Reduced-form Listing-level Analysis of Airbnb Listings By Four Lis
ting Types
49
(1) (2) (3) (4) (5) (6)
MODEL Poission Poission Logit OLS OLS OLS
VARIABLES
reservation
monthly
VSR cumu
VSR cumu
dummy
crime cumu
norm
VS listing
pct zip
VS listing
pct radius
Sample Monthly Reservation Reserved Proper ty Reserved Property Reserved Property Reserved Property Reserved Property
Subsample 1st VSR booking in H, at least 1 booking b ef ore 1st VSR in L
VS user × post -1.215*** -0.796*** -0.598*** -0.945*** -0.0292*** -0.0378***
(0.0990) (0.247) ( 0. 227) (0.200) (0.00394) (0.00604)
Observations 69,417 6,101 6,101 6,155 6,155 6,155
Subsample 1st VSR booking in L, at least 1 booking bef ore 1st VSR in L
VS
user × post -0.806*** -0.618*** -0.421*** -0.926*** -0. 0224*** -0.0175***
(0.0740) (0.162) ( 0. 135) (0.134) (0.00340) (0.00675)
Observations 175,770 15,556 15,533 15,648 15,648 15,648
Subsample 1st VSR booking in L, all booking before 1st VSR in H
VS user × post -0.730** -1.445** -1.105 -0.190 -0.00572 -0.0326
(0.342) (0.642) (0.711) (1.055) (0.0216) (0.0239)
Observations 6,612 462 462 462 462 462
Subsample 1st VSR booking in H, all booking before 1st VSR in H
VS user × post -0.730 -1.474 -0.720 -1.949** -0.0567*** -0.0717***
(0.445) (1.312) (0.941) (0.759) (0.0164) (0.0185)
Observations 2,135 146 141 146 146 146
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control treatment-control pair ID FE with standard errors clustered by pair ID. The subsamples
are defined by whether the VS users’ 1st VSR post is in the L area and whether VS users have bookings in the L ar ea before their 1st VSR post.
Table A5: Reduced -f orm Guest-level Analysis: DID for the VS Users whose 1s
t VSR booking and bookings before 1st VSR are in areas of H
or L.
50
(1) (2) (3) (4) (5) (6)
MODEL Poission Poission Logit OLS OLS OLS
VARIABLES
reservation
monthly
VSR cumu
VSR cumu
dummy
crime cumu
norm
VS listing
pct zip
VS listing
pct radius
Sample Monthly Reservation Reserved Proper ty Reserved Property Reserved Property Reserved Property Reserved Property
Subsample 1st VSR booking in W, at least 1 booking b efor e 1st VSR in M
VS user × post -1.012*** -0.877*** -0.727*** -1.502*** -0.0261*** -0.0337***
(0.0911) (0.211) ( 0. 191) (0.185) (0.00398) (0.00596)
Observations 104,313 8,959 8,959 9,031 9,031 9,031
Subsample 1st VSR booking in M, at least 1 booking before 1st VSR in M
VS
user × post -0.869*** -0.546*** -0.306** -0.550*** -0.0247*** -0.0176**
(0.0814) (0.175) ( 0. 146) (0.135) (0.00360) (0.00759)
Observations 140,874 12,698 12,675 12,772 12,772 12,772
Subsample 1st VSR booking in M, all booking before 1st VSR in W
VS user × post -0.595* -1.774** -1.298* -0.664 -0.00956 -0.0406**
(0.331) (0.757) (0.740) (1.153) (0.0212) (0.0193)
Observations 5,656 383 383 383 383 383
Subsample 1st VSR booking in W, all booking before 1st VSR in W
VS user × post -0.994** -0.848 -0.454 -1.039** -0.0353 -0.0433
(0.433) (1.110) (0.829) (0.430) (0.0240) (0.0373)
Observations 3,091 225 220 225 225 225
Note: *** p < 0.01, ** p < 0.05, * p < 0.1. All regressions control treatment-control pair ID FE with standard errors clustered by pair ID. The subsamples
are defined by whether the VS users’ 1st VSR post is in the M area and whether VS users have bookings in the M area before their 1st VSR post.
Table A6: Reduced-form Guest-level Analysis: DID for the VS Users wh ose 1s
t VSR booking an d book i ngs before 1st VSR are in areas of W
or M .
51
Online Appendix B: Calibra t e the cross-listing-with-buyer effect of VSR
for the structural model
This appendix explains how we use the reduced-form DID results of VS
users to calibrate t he coefficient
of h aving any VSR in the structural demand mo del .
According to Table 6 Column 1, the DID coefficient on treated × post is -0.918 in a Poisson regression
of the number of Airbnb reservations, which is a 60.07% decrease for VS users relative to normal users.
Given the average number of reservations per month for a singl e VS user in our sample is 0.1092 and
review rate is 44.56%, a VS user’s Airbnb reservations are (0.1092 × 60. 07%/0. 4456 = 0.147 fewer than
a nor mal user after she has reported a VS issue in her first VSR. Thi s can be expressed by:
[#Airbnbbooking
V S user,aft
− #Airbnbbooking
V S user,bef
]
−[#Airbnbbooking
NM user,aft
− #Airbnbbooking
NM user,bef
] = −0.147
(6)
Recall that we define an average guest’s utility from listing j as:
U
j,t
= EU
j,t
+ ǫ
j,t
= α
j
+ α
k,t
+ δ · X
j,t
+ β
0
· log( ADR
j,t
) + β
1
· Crime
z,t−1
+ β
2
· LSR
j,t−1
+ β
3
· V SR
j,t−1
+ β
4
· V SRADIU S
j,t−1
+ ǫ
j,t
.
If we assume self experience of vicinity safety issues only changes β
3
, we can write:
β
3
= β
3,NM
+ ∆β
3
· [i = VS User]
where β
3,NM
indicates normal users’ sensitiv i ty to observing any V SR in a listing, [i = VS User] is a
dummy equal to one f or VS users, and thus β
3,NM
+∆β
3
indicates VS users’ updated sensitivity to V SR.
Assuming VS and normal users have the same t en d ency to book short-term rentals, the DID results
can be rewrit ten as user i’s market share for all Airbnb choices
P
j∈Airbnb
s
ij
:
∂
P
j∈Airbnb
s
ij
∂V SR
i=V S user
−
∂
P
j∈Airbnb
s
ij
∂V SR
i=NM user
= −0.147 (7)
The mar ket share of all Airbnb reservations is:
X
j∈Airbnb
s
ij
= 1 − s
i,V RBO
= 1 −
1
1 +
P
j∈Airbnb
exp(U
ij
)
(8)
Then:
∂
P
j∈Airbnb
s
ij
∂V SR
i=NM user
= β
3,NM
· s
NM user,V RBO
·
X
j∈Airbnb & V SR
s
NM user,j
(9)
∂
P
j∈Airbnb
s
ij
∂V SR
i=V S user
= (β
3,NM
+ ∆β
3
) · s
V S user,V RBO
·
X
j∈Airbnb & V SR
s
V S user,j
(10)
Denote a user’s total probability of choosing any Airbnb listings with any VSR as:
s
NM user,Airbnb & V SR
=
X
j∈Airbnb & V SR
s
NM user,j
(11)
52
s
V S user,Airbnb & V SR
=
X
j∈Airbnb & V SR
s
V S user,j
(12)
The D ID results can be written as:
(β
3,NM
+ ∆β
3
) · s
V S user,V RBO
· s
V S user,Airbnb & V SR
−β
3,NM
· s
NM user,V RBO
· s
NM user,Airbnb & V SR
= −0.147
(13)
Note that we observe normal users’ market shares in the data because almost all users are normal
users, but we do not observe V S users’ market shares because we cannot track V S user s in all Airbnb
and V RBO bookings. However, the utility framework spells out how these two types of users should
differ. More specifically, the model implies:
s
NM user,V RBO
s
V S user,V RBO
= s
NM user,V RBO
+ s
NM user,Airbnb & V SR=0
+ exp(∆β
3
) · s
NM user,Airbnb & V SR
.
(14)
This implies:
s
V S user,V RBO
=
s
NM user,V RBO
s
NM user,V RBO
+ s
NM user,Airbnb & V SR=0
+ exp(∆β
3
) · s
NM user,Airbnb & V SR
(15)
Similarly:
s
NM user,Airbnb & V
SR
s
V S user,Airbnb & V SR
= exp(∆β
3
) · (s
NM user,V RBO
+ s
NM user,Airbnb & V SR=0
+ exp(∆β
3
)
·s
NM user,Airbnb & V SR
)
(16)
This implies:
s
V S user,Airbnb & V SR
=
1
exp(∆β
3
)
·
s
NM user,Airbnb & V SR
s
NM user,V RBO
+s
NM user,Airbnb & V SR=0
+exp(∆β
3
)·s
NM user,Airbnb & V SR
(17)
P
lug these into the DID results:
(β
3,NM
+ ∆β
3
) · s
V S user,V RBO
· s
V S user,Airbnb & V SR
−β
3,NM
· s
V S user,V RBO
· s
V S user,Airbnb & V SR
= −0.147
(18)
Because almost all users are normal users, the data gives us s
NM user,V RBO
(market share of V RBO),
s
NM user,Airbnb & V SR=0
(total market share of all normal Airbnb listings), as well as s
NM user,Airbnb & V SR
(total market share of all Airbnb VS listings). We also know β
3,NM
from the utility regression. Thus,
the only unknown in the above equation is ∆β
3
. We can readily solve for it and obtain ∆β
3
= −2.17.
53
