GEORGIOS ZERVAS, DAVIDE PROSERPIO, and JOHN W.

BYERS*

Peer-to-peer markets, collectively known as the sharing economy,

have emerged as alternative suppliers of goods and services traditionally
provided by long-established industries. The authors explore the economic
impact of the sharing economy on incumbent firms by studying the
case of Airbnb, a prominent platform for short-term accommodations.
They analyze Airbnb’s entry into the state of Texas and quantify its
impact on the Texas hotel industry over the subsequent decade. In
Austin, where Airbnb supply is highest, the causal impact on hotel revenue
is in the 8%–10% range; moreover, the impact is nonuniform, with lower-
priced hotels and hotels that do not cater to business travelers being
the most affected. The impact manifests itself primarily through less
aggressive hotel room pricing, benefiting all consumers, not just
participants in the sharing economy. The price response is especially
pronounced during periods of peak demand, such as during the South
by Southwest festival, and is due to a differentiating feature of peer-
to-peer platforms—enabling instantaneous supply to scale to meet
demand.

Keywords: sharing economy, Airbnb, hotel industry, competition, peer-to-

peer markets

The Rise of the Sharing Economy: Estimating

the Impact of Airbnb on the Hotel Industry

The emergence of peer-to-peer platforms, collectively suppliers can readily list and delist the selection of goods or
known as the “sharing economy,” has enabled people services they offer.
to collaboratively make use of underutilized inventory In our work, we focus on the impacts that these peer-to-peer
through fee-based sharing. Consumers have so far en- platforms have on incumbent firms, specifically focusing on
thusiastically adopted the services offered by firms such the case of Airbnb, a provider of travel accommodation and a
as Airbnb, Uber, Lyft, and TaskRabbit. The rapid growth pioneer of the sharing economy. Because Airbnb has served
more than 50 million guests since it was founded in 2008 and
of peer-to-peer platforms has arguably been enabled by
has a market capitalization eclipsing $30 billion, we hypothesize
two key factors: technology innovations and supply-side
that it has a measurable and quantifiable impact on hotel
flexibility. Technology innovations have streamlined the revenue in affected areas (see Airbnb 2015; Farrell and
process of market entry for suppliers, facilitated searchable Bensinger 2016). Our hypothesis is that some stays with
listings for consumers, and kept transaction overheads low. Airbnb serve as a substitute for certain hotel stays, thereby
Supply-side flexibility is another hallmark of these platforms: affecting hotel revenue, and that this impact is differenti-
Uber drivers can add or remove themselves from the available ated by geographic region, by hotel market segment, and by
supply of drivers with a swipe on an app, and similarly other season. Although incumbent firms face higher fixed costs and
offer less personalized products than peer-to-peer platforms,
*Georgios Zervas is Assistant Professor of Marketing, Questrom School of they have only recently begun to view competition from plat-
Business, Boston University (e-mail: [email protected]). Davide Proserpio is As- forms such as Airbnb as a serious threat. For example, hotel
sistant Professor of Marketing, Marshall School of Business, University of executives have publicly issued largely dismissive statements
Southern California (e-mail: [email protected]). John W. Byers is regarding competitors like Airbnb, arguing either that these
Professor of Computer Science, Computer Science Department, Boston
University (e-mail: [email protected]). Coeditor: Randy Bucklin; Associate
peer-to-peer platforms are a niche market or that they target
Editor: Avi Goldfarb. complementary market segments from those targeted by hotel
chains. Notably, Airbnb also appears to espouse this latter

© 2017, American Marketing Association Journal of Marketing Research

ISSN: 0022-2437 (print) Vol. LIV (October 2017), 687–705
1547-7193 (electronic) 687 DOI: 10.1509/jmr.15.0204
688 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

view: according to Airbnb, in many cities, over 70% of Airbnb that the timescale of such a response would occur with a
properties are outside the main hotel districts,1 suggesting multiyear lag.
complementarity of their offerings. Our next set of results develops a more nuanced un-
In this article, we provide empirical evidence to this debate derstanding of the mechanisms behind Airbnb’s impact on
by studying the differentiated impact of Airbnb’s entry in the hotel room revenue by unpacking the effects to study the
Texas hotel market on hotel room revenue. Our study ex- differentiated impacts that Airbnb has had across hotels, cities,
plores the relationship between Airbnb and hotels in the state and time. First, given the nature of rentals currently on Airbnb,
of Texas by estimating monthly hotel room revenue as a which typically provide fewer amenities and services than
function of Airbnb entry in the market. Using data we collected many hotels, we expect hotels that provide more differentiated
from Airbnb, monthly hotel room revenue from approxi- services to be less affected. We examine three such cases in
mately 3,000 hotels in Texas dating back to 2003, and several high-end hotels, chain hotels, and hotels catering to business
other auxiliary data sets to compile controls, we quantify the travelers, each of which provide amenities that a typical Airbnb
extent to which Airbnb’s entry to the accommodation market host does not. First, after segmenting hotels in five industry-
has negatively affected hotel room revenue. standard price tiers (budget, economy, midprice, upscale, and
To identify the causal impact of Airbnb on hotel revenue, we luxury) we find that the impact of Airbnb is gradually mag-
employ a difference in differences (DD) empirical strategy. nified as we move down the price tiers. Then, through a similar
Specifically, because of the significant variability in both the analysis, using conference and meeting room space as a proxy
temporal rate and the spatial density of Airbnb adoption in for the extent to which a hotel caters to business travel, we find
Texas, as well as the geographic specificity of both our hotel that the impact of Airbnb also falls disproportionately on hotels
and Airbnb data sets, we are able to treat Airbnb market entry lacking conference facilities. Finally, we examine Airbnb’s
as a variable intervention in space and time against the hotel differential impact on chain hotels versus independent hotels
room revenue data. Our DD strategy identifies the Airbnb and confirm our expectation that chain hotels will be less
treatment effect by comparing differences in revenue for hotels affected than independents for reasons ranging from chains’
in cities affected by Airbnb before and after Airbnb’s entry larger marketing budgets and stronger brands to their pre-
with a baseline of differences in revenue for hotels in cities dictably consistent service.
unaffected by Airbnb over the same period of time. To perform In our final main result, we study the impact that Airbnb has
the analysis, we regress against two measures of Airbnb during periods of peak demand, leveraging our instantaneous
supply: (1) a cumulative measure that defines supply as all measure of supply. Use of this measure enables us first to
listings appearing prior to a given date in a given city and (2) an confirm that there are significant seasonal fluctuations in city-
instantaneous measure that defines supply as those Airbnb level Airbnb supply that are correlated with periods of peak
demand in those cities. We then study the impacts that Airbnb
listings active within a short (e.g., three-month) period. In all
has exerted, year-over-year, during the highly popular South
our specifications, we include a rich set of controls that vary by
by Southwest (SXSW) festival in Austin and during the Texas
location and over time: population, wages, unemployment,
State Fair in Dallas. Our finding is that Airbnb’s ability to
total hotel room supply in each market, each hotel’s own
flexibly scale instantaneous supply in response to seasonal
capacity over time, airport passenger counts, and the TripAdvisor
demand has significantly limited hotels’ pricing power during
ratings for each hotel as a proxy for quality. In addition to periods of peak demand. Indeed, we argue that accommo-
these measured covariates, we include city-specific trends dating surges in demand through flexible scaling of supply is a
and city-month dummies to account for seasonal variation in defining feature of the sharing economy, and we interpret our
demand across different markets. Using our preferred cu- result as evidence of the power of this capability, which ap-
mulative specification, we find that, in Texas, each additional pears difficult for incumbent firms such as hotels to directly
10% increase in the size of the Airbnb market resulted in counteract.
a .39% decrease in hotel room revenue, with similar but Finally, we conduct several robustness checks to support a
somewhat smaller estimated impacts using the instantaneous causal interpretation of our estimates. First, we show that the
supply measure. These effects are primarily driven by Austin, basic set of controls included in our DD specification (i.e.,
where Airbnb inventory has grown extremely rapidly over hotel fixed effects and time trends) explain approximately 88%
the past few years, resulting in an estimated revenue impact of of the variation in Airbnb supply, whereas time-varying ob-
8%–10% for the most vulnerable hotels in Austin. servables that could potentially drive hotel revenue have al-
We next investigate the market response to Airbnb entry and most no additional explanatory power. Second, we check
study the mechanisms whereby affected hotels might react to whether hotel performance drives Airbnb adoption, which
Airbnb’s market entry both in the short run and in the long run. would indicate that we have confused cause and effect. To the
In the short run, likely responses could take the form of a price contrary, we find that a wide range of pre-Airbnb demographic
response or an occupancy response. Using hotel industry and market characteristics—including, for example, hotel
performance metrics as dependent variables, we find a small room prices, occupancy rates, and hotel room supply per
decrease in occupancy rate and a significant decrease in city, which are all significant predictors of post-Airbnb hotel
hotel room prices. Notably, such a price response benefits room revenue—are not correlated with the patterns of Airbnb
all consumers, not just participants in the sharing economy. adoption we see in our data. Third, we define a measure of
With respect to longer-term responses, such as diminished competing Airbnb supply at a per hotel granularity, accounting
investment or hotel entry and exit, we do not find evidence of for the geographic distance between the hotel and Airbnb
an effect yet, consistent with evidence we present showing inventory. This distance-based analysis shows a magnified
negative impact from Airbnb on hotels as proximity between
1See http://blog.airbnb.com/economic-impact-airbnb/. hotels and Airbnb inventory increases. Fourth, we show that
The Rise of the Sharing Economy 689

our results are robust to alternative measures of Airbnb supply. and Fernandez 2010), and understanding the supply-side labor
Finally, in a separate analysis, we combine DD with coarsened market (Hall and Krueger 2015). In contrast, our work em-
exact matching (CEM; Iacus, King, and Porro 2012). Spe- pirically studies a setting in which a peer-to-peer market
cifically, we match each “treated” hotel affected by Airbnb to a offers a substitute for consumer services supplied by traditional
“control” hotel belonging to the same price tier and sharing the firms.
same affiliation (e.g., an upscale Hilton in Austin where It is in this context that our research contributes to the lit-
Airbnb adoption is high, and an upscale Hilton in Dallas where erature on substitution between peer-to-peer markets and in-
Airbnb penetration is low), discarding hotels that remain cumbent firms, because markets such as Airbnb can be viewed
unmatched. We find that our CEM estimate is similar to our as providing enabling technology that facilitates suppliers of
main analysis. Taken together, these robustness checks provide niche inventory to flexibly bring their products to market.
significant support for the assumptions underlying our DD Unlike traditional markets, Airbnb provides sufficiently low
analysis. We conclude this article by discussing managerial and cost of revenue for people to profitably list remnant inventory
policy implications related to the rapid growth of Airbnb online; moreover, Airbnb provides enhanced reach by re-
specifically and the sharing economy more broadly. ducing consumer search costs (Bakos 1997). As such, our
study can be viewed as investigating the consequences of an
online platform lowering the barrier to entry for suppliers.
RELATED WORK Related work has studied similar examples in other domains.
Relatively few studies have investigated competition between For example, several recent studies have focused on the im-
peer-to-peer markets and incumbent firms offering similar pact of Craigslist—a website featuring free online classi-
goods or services. In one line of recent work, Einav, Farronato, fied ads—on the newspaper industry (Kroft and Pope 2014;
and Levin (2016) discuss the design and regulation of peer-to- Seamans and Zhu 2013).
peer markets and provide theoretical predictions of the effects Finally, our work contributes to the literature focusing on
of competition from these markets on incumbent firms. A key the impact of external shocks on the tourism and the hospitality
prediction they make, which is borne out in our data, is that industry. However, much of the prior work in this area has
peer-to-peer markets can reduce price variability by flexibly centered on demand shocks. For example, O’Connor, Stafford,
scaling supply to accommodate increased demand. As for em- and Gallagher (2008) study the impact of terrorism on tourism
pirical work, a handful of studies have examined the adoption in Ireland; Baker and Coulter (2007) estimate the impact of the
and effects of car sharing; for example, two studies have used 2002 and 2005 terrorist attacks in Bali on the islands’ vendors.
survey analysis methods to find that car sharing is associated Similarly, Kosová and Enz (2012) examine the adverse effects
with significant decreases in miles traveled, gasoline con- of the September 11 attack and the 2008 financial crisis on
sumption, and car ownership (Cervero, Golub, and Nee 2007; hotel performance.
Martin, Shaheen, and Lidicker 2010). In the domain of ac-
commodation sharing, we find numerous opinion pieces in the DATA AND THE AIRBNB PLATFORM
popular press and on blogs, but little in the way of academic For our study, we collect and combine data from various
literature. Our closest comparison point is a set of short studies, sources including the Airbnb website, the Texas Comptroller
commissioned by Airbnb, which claim that the Airbnb business Office, STR, county demographics from the U.S. Census
model is complementary to the hotel industry but primarily Bureau, airport passenger counts from the U.S. Bureau of
focus on arguing for and quantifying the substantial net eco- Transportation Statistics, the Current Population Survey from
nomic benefit to cities that Airbnb travelers provide.2 Although the U.S. Bureau of Labor Statistics, and hotel reviews from
our work is related to these studies, we apply a more sophis- TripAdvisor.
ticated identification strategy, methodology, and segmentation
analysis, resulting in conclusions that are both different and The Airbnb Platform
more nuanced. Notably, recent analyses have confirmed our Much of the data used in our study is collected directly from the
initial findings in Texas in other markets; for example, Credit Airbnb website. Airbnb describes itself as “a trusted com-
Suisse analysts used STR data to estimate that in New York munity marketplace for people to list, discover, and book
City, Airbnb caused January 2015 revenue per hotel room to unique accommodations around the world,” and it exemplifies a
decline by 18.6%, year over year (Phillips 2015). peer-to-peer marketplace in the sharing economy. Prospective
Our work contributes to the growing literature on multisided hosts list their spare rooms or apartments on the Airbnb plat-
platform competition, as Airbnb exemplifies a two-sided form; establish their own nightly, weekly or monthly price; and
platform. Much of this literature has established the eco- offer accommodation to guests. Airbnb derives revenue from
nomic theory of two-sided markets—for example, through both guests and hosts for this service: guests pay a 9%–12%
structural models that establish theories of price structure and service fee for each reservation they make, depending on the
usage (Rochet and Tirole 2003; Rysman 2009; Weyl 2010), length of their stay, and hosts pay a 3% service fee to cover
and models that connect innovations in product design to the cost of processing payments. Since its launch in 2008, the
network effects (Parker and Van Alstyne 2005). Other work, Airbnb online marketplace has experienced very rapid growth,
more closely related to our own, has contributed empirical with more than two million properties worldwide and over 50
results to the literature that try to explain the behavior of firms million guests who have used the service by September 2015
and people in two-sided markets (Jin and Rysman 2012), (Airbnb 2015).
including the role of multihoming (Landsman and Stremersch Airbnb’s business model currently operates with minimal
2011), modeling response to regulation (Valverde, Chakravorti, regulatory controls in most locations, and as a result, both hosts
and guests have incentives to use signaling mechanisms to build
2See https://www.airbnb.com/economic-impact/. trust and maximize the likelihood of a successful booking. To
690 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

reinforce this behavior, Airbnb has built an online reputation monthly hotel room revenue, the data set includes basic
system that enables and encourages participants to rate and information including hotel name, address, and capacity. The
review each completed stay. Guests use star ratings to rate raw data set spans the period between January 2003 and
features of their stay (e.g., cleanliness, location, communication) August 2014.
while both guests and hosts are encouraged to post public re- Notably, according to Texas law, “a hotel is considered to be
views of each stay on the platform. any building in which members of the public rent sleeping
accommodations for 15 or more per day.”3 For this reason,
Airbnb Listings Data revenue from Airbnb properties (as well as various other
To estimate the extent of Airbnb’s market entry, we collected vacation rental options) whose owners are in compliance with
consumer-facing information from Airbnb.com on the com- the Texas tax code is also reported in this data set. This is
plete set of users who had listed their properties in the state of evident from Figure 1, which plots the number of unique tax-
Texas for rental on Airbnb. We refer to these users as “hosts” paying properties in Austin broken down by capacity (i.e.,
and their properties as their “listings.” Each host is associated maximum occupancy). We conjecture that the rapid increase in
with a set of attributes including a photo, a personal statement, low-capacity properties starting in 2008 is related to Airbnb’s
their listings, guest reviews of their properties, and Airbnb- entry into the Texas market at the same time. To exclude non-
certified contact information. Similarly, each listing displays hotel properties from our analysis of impact on hotels, we
attributes including location, price, a brief description, photos, cross-reference the Texas Comptroller data set with the U.S.
capacity, availability, check-in and checkout times, cleaning hotel census data provided to us by STR. The STR census
fees, and security deposits. Our collected data set contains includes all U.S. hotels and contains a rich attribute set for
detailed information on 10,555 distinct hosts and 13,395 each hotel, including its opening date, price segment, capacity,
distinct listings spanning a period from January 2008 to operation type (chain vs. independent), and geographic lo-
August 2014. cation. In total, the STR data set contains information on 3; 747
To conduct our analysis, we must choose an appropriate hotels in Texas metropolitan areas. After linking the STR
level of geographic aggregation. Here, our data are suitably census data set with the Texas tax data set, we obtain high-
granular (with location accuracy to roughly 100 meters) to confidence matches for a panel of 3; 619 properties (96% of
permit analysis at many different scales. Our preferred spec- STR hotels, which account for over 95% of the revenue in our
ification employs city-level granularity and is driven by the data).
observation that a city is the largest geographic unit within Airbnb can affect hotel room revenue through lower oc-
which we reasonably expect to see significant substitution cupancy rates, decreased hotel room prices, or a combi-
patterns between hotels and Airbnb properties. However, nation of these two factors, conventionally reported within
distance-based measures also arguably have operational validity. the hotel and hospitality industry as revenue per available
We discuss these along with our other modeling decisions and room (RevPAR), which is the product of average room price
robustness checks. and occupancy. Because the data we obtained from the Texas
Another central element of our analysis is to accurately Comptroller’s office does not report either occupancy rates or
quantify Airbnb supply; however, this cannot be directly hotel room prices, we obtain additional data on these quantities
inferred from available data and is thus a highly nuanced for a subset of Texas hotels from STR. The room price, also
modeling decision. Indeed, inferring instantaneous Airbnb referred to as average daily rate (ADR), and occupancy rate
supply is a challenging task even for Airbnb itself because data from STR covers a subset of 2; 584 hotels in Texas who
of “stale vacancies” (i.e., Airbnb listings that appear to be part chose to report this information to STR over the same time
of available supply only because the hosts neglected to period (January 2003 through August 2014).
update the availability status of those listings). By analyzing
proprietary Airbnb data, Fradkin (2014) finds that between
Auxiliary Data Sources
21% and 32% of guest requests are rejected as a result of this
effect. We assemble a set of control variables derived from publicly
Despite imperfect information, we do have substantial data available sources. First, for each hotel we collect its entire
with which to construct proxies for supply—namely, the date TripAdvisor review history—a total of 424,583 reviews. We
that hosts became Airbnb members and the date for each then use TripAdvisor star ratings to control for changes in hotel
review of each property. Significantly, Fradkin et al. (2014) quality over our observation period. Second, we collect pas-
report that 67% of Airbnb guests left a review about their stay senger arrival data for all Texas airports from the Bureau of
across their large data set. For market entry, we can estimate Transportation Statistics. We then associate each city in Texas
the (unobservable) entry date of individual listings either by with its nearest airport and use the passenger data to control
using the date their owners became Airbnb members or by the for changes in tourism demand over time that are unrelated to
date of the first review. Similarly, we can construct proxies for Airbnb. The data are a monthly panel of passenger counts,
both cumulative and instantaneous supply by leveraging the in which we exclude passengers connecting through Texas
review histories we compile. We detail and justify our ap- airports. Third, we obtain monthly unemployment and wage
proach in a subsequent section. data at the metropolitan statistical area–level from the Bureau
of Labor Statistics (https://www.bls.gov/). Unemployment
Hotel Data: Revenue, Prices, and Occupancy Rates statistics are updated monthly, while the wage data, which
The main dependent variable we use in our analysis is monthly comes from the Occupational Employment Statistics Survey,
hotel room revenue, which we obtained from public records is updated once a year. Finally, we obtain demographic
furnished by the Texas Comptroller of Public Accounts, in
their capacity as auditors of state tax collection. In addition to 3See https://comptroller.texas.gov/taxes/audit/manuals/hotel/ch2.php.
The Rise of the Sharing Economy 691

Figure 1 correlated with both Airbnb entry and hotel room revenue,
ANNUAL COUNTS OF AUSTIN PROPERTIES THAT PAY HOTEL resulting in endogeneity. Stated differently, we assume that
OCCUPANCY TAX, BROKEN DOWN BY CAPACITY unobserved factors that could potentially jointly affect both
Airbnb adoption and hotel room revenue do not systematically
vary both between different cities and over time. For instance,
1,000 the following unobserved factors are accounted for in our
Capacity estimate and do not bias our estimates: (1) city-specific, time-
invariant differences in adoption rates (e.g., consumers in
Fewer than five rooms Austin overall being more likely to adopt Airbnb than con-
At least five rooms sumers in Dallas); (2) factors that vary arbitrarily over time but
750 do not vary across cities (e.g., a generally increasing awareness
of Airbnb shared across all consumers in Texas over time);
Tax-Paying Properties

and (3) city-specific trends, which allow for unobserved con-

founders that vary both between cities and over time according
to a prespecified functional form (linear or quadratic).
500 Our DD specification takes the following form:
(1) log Hotel Revenueikt = b log Airbnb Supplykt + X0ikt g
+ hi + tt + Cityk × Montht + eikt :
The dependent variable is the log of monthly room revenue of
250 hotel i in city k at time t. Our model includes hotel fixed
effects hi, and time (year-month) fixed effects tt . To im-
plement the DD strategy, we define treated hotels to be those
hotels in cities with an Airbnb presence, and nontreated hotel
to be those hotels in cities with no Airbnb presence. The first
0 difference is taken using the hotel fixed effects, which allow
2003 2005 2007 2009 2011 2013 for time-invariant differences in hotel room revenue between
treated hotels and nontreated hotels. The second difference in
our DD specification is taken over time using year-month
fixed effects tt , which allow for unobserved time-varying
information at the county level from the U.S. Census Bureau revenue differences that are common across different cities.
(https://www.census.gov/).

QUANTIFYING AIRBNB’S IMPACT IN TEXAS Figure 2

GEOGRAPHICAL DISTRIBUTION OF HOTELS AND AIRBNB
Empirical Strategy
LISTINGS IN THE STATE OF TEXAS AND IN TRAVIS COUNTY,
Airbnb has had widely varying degrees of traction within TEXAS IN 2013
different local, regional, and international markets both with
respect to initial market entry and the rate at which it has been
adopted within markets. For example, consider Figure 2, Airbnb Hotels
which depicts the current extent of market penetration both of
Airbnb properties and hotels within the state of Texas and
within the county encompassing the state capital, Austin.
Unlike hotels, which have coverage throughout the state and
Texas

pockets of local density, such as in downtown Austin, Airbnb

has spotty coverage at best throughout the state but broader
coverage across metro areas, including suburbs and exurbs.
Table 1 reveals that, over the past eight years in the ten most
populous cities in Texas, patterns of Airbnb adoption are 0 200 400 km

themselves diverse, with several cities experiencing early

adoption and rapid growth, whereas others experiencing
minimal Airbnb adoption. Our empirical strategy exploits this
variability to identify the impact of Airbnb’s rise on hotel room
Travis County

revenue using a DD identification strategy. Specifically, we

estimate Airbnb’s impact on hotel room revenue by comparing
changes in hotel room revenue before and after Airbnb enters a
specific city with a baseline of changes in hotel room revenue
in cities with no Airbnb presence over the same period of time.
The key identification assumption we must make to
support a causal interpretation of this DD estimate is that there
are no unobserved, time-varying, city-specific factors that are
692 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

Table 1
AIRBNB’S SPATIAL AND TEMPORAL PENETRATION

Houston San Antonio Dallas Austin Fort Worth El Paso Arlington Corpus Christi Plano Laredo
Population (millions) 2.16 1.38 1.24 .84 .78 .67 .38 .31 .27 .24
No. of Airbnb listings in ...
2008 1 9 0 25 0 0 0 0 0 0
2009 6 13 7 146 2 0 1 0 0 0
2010 39 22 23 468 10 0 3 0 1 0
2011 169 72 109 1,862 34 3 19 7 5 1
2012 425 171 271 5,158 68 8 27 24 20 1
2013 695 271 422 7,489 93 23 36 49 33 1
2014 891 346 526 8,575 114 31 52 60 44 2

Notes: This table presents cumulative counts of Airbnb listings per year in the ten most populous Texas cities.

The coefficient of interest is b, which has the usual DD review; thus, listings that receive a review must be on the
interpretation: it is an estimate of the percentage change in market at that time. Moreover, the incidence of reviewing is
hotel room revenue in treated (Airbnb-adopting) cities after high: Fradkin et al. (2014) report that 67% of Airbnb stays in
Airbnb’s entry compared with a baseline of changes in hotel their large data set resulted in a review. Taken together, these
room revenue over the same period in untreated (non- two facts indicate that a time series of Airbnb reviews reflects
adopting) cities. We interpret a statistically significant negative time-varying supply. For each Airbnb listing in our data, we
coefficient on Airbnb supply as indicating that Airbnb listings observe its entire historical record of reviews, which includes
lead to Airbnb bookings that substitute for hotel stays and reviews for the listing as well as reviews for each guest (by
affect hotel room revenue. We interpret a coefficient that is not the host). Using the review data set, we apply the following
statistically significantly different from zero as indicating that heuristic to determine when each Airbnb listing was active:
Airbnb listings having no effect on hotels. We interpret When an Airbnb listing enters the market, we assume that it
a positive coefficient (though implausible) as indicating that remains active for m months, which we refer to as the listing’s
Airbnb listings benefit hotels. Next, we elaborate several time to live (TTL); whenever a listing is reviewed, its TTL is
measures of Airbnb supply that we employ in Equation 1 and extended by m months from the date of the review; if a listing
the various economic impacts each measure can identify. exceeds its TTL, it exits the market; finally, listings become
active again after exiting the market if they receive a new
Modeling Airbnb Supply review.
Our first approach uses a cumulative measure of Airbnb The main advantage of the instantaneous supply measure is
supply, quantified at the granularity of individual cities: for a that it can capture a key differentiating feature of Airbnb: its
given city and date, we count the number of distinct listings ability to scale supply. This measure both has descriptive value
that have cumulatively appeared on Airbnb in that city prior to and enables us to confirm that our results are not driven by our
that date. We approximate the unobservable entry date of choice of a cumulative supply measure. A limitation of the
individual listings by using the displayed date their owners instantaneous supply measure, arising from the way we
became Airbnb members. By construction, a weakness of the construct it, is that it may underestimate Airbnb inventory in
cumulative measure of Airbnb supply is that it ignores listing the low season. During low season, Airbnb listings face lower
exit, which we do not observe in our data. Therefore, our demand, which in turn leads to fewer reviews. Therefore,
estimate of Airbnb’s impact will be consistent if the un- during the off-season, some listings that are available may
observed fraction of active Airbnb listings is not endogenously receive zero reviews and thus be misclassified as unavailable.
correlated with cumulative listing supply and hotel revenue. Figure 3 compares the cumulative and instantaneous Airbnb
To demonstrate when the (observed) cumulative supply and supply measures for the four largest cities in our data. We see
(unobserved) actual monthly supply yield the same consis- that our instantaneous Airbnb supply measure fluctuates sig-
tent estimate, we relate cumulative supply to actual supply nificantly over time, differentiating it from our cumulative
through a set of (unobserved) multipliers fkt 2 ½0, 1, such that supply measure. Moreover, its pattern of variation over time
Actual Airbnb Supplykt = fkt × Cum: Airbnb Supplykt . Here, correlates with periods when we would expect Airbnb supply
fkt is the fraction of Airbnb listings that entered the market prior to be highest, such as March in Austin, when the SXSW
to time t and are still actively in the market at time t. Because festival takes place.
we work with a log-log specification, fkt becomes an unobserved A final issue that pertains to both measures of Airbnb supply
quantity that enters the error term additively. Therefore, only that we must address is that although the unit of analysis is
residual variation in fkt after controlling for observables, fixed hotel monthly room revenue, the treatment, Airbnb adoption,
effects, and trends that is correlated with residual cumulative occurs at the city level. This mismatch in the level at which we
supply, will cause bias. measure our dependent variable compared with the treatment
Our second approach employs an instantaneous proxy variable can result in understating the standard error of the
measure of actual Airbnb supply. To build an instantaneous estimate of Airbnb’s impact, because it is likely that hotel room
measure, we exploit the fact that Airbnb requires guests who revenue is serially correlated over time within a city. We
wish to submit a review to do so within 14 days of a stay and correct for this mismatch by clustering standard errors at the
reports the checkout date (with monthly precision) in each city level, which lets us account for possible serial correlation
The Rise of the Sharing Economy 693

Figure 3
CUMULATIVE VERSUS INSTANTANEOUS AIRBNB

Austin Dallas Houston San Antonio

8,000

Cumulative
6,000
4,000
Airbnb Supply

2,000
0
SXSW

Instantaneous
2,000

1,000

0
2004 2009 2014 2004 2009 2014 2004 2009 2014 2004 2009 2014

Notes: The seasonal peaks of the Austin instantaneous supply curve correspond to SXSW.

in hotel room revenue. In doing so, we follow the standard Second, as we explained previously, our DD estimate will
practice in the literature for analyzing panel data in a DD be biased if there exist unobserved factors that vary across
setting (Bertrand, Duflo, and Mullainathan 2004; Donald and cities and over time and jointly influence Airbnb entry and
Lang 2007). We report standard errors clustered at the city hotel room revenue—most notably, demand for accommo-
level for all subsequent regressions. dation. This type of bias likely works against finding a negative
Airbnb effect: both Airbnb supply and hotel revenue should
respond positively to shifts in accommodation demand, which
Incorporating Controls: Hotel Supply and Quality, Demand implies that if we omit a control for demand, then Airbnb
Shifters, and Demographics supply will absorb its effect and become biased upwards.
An initial identification challenge we face is that increased We use three types of controls to account for variation in
demand for accommodation is likely correlated with increases accommodation demand across different cities. First, we in-
in both Airbnb supply and hotel room supply. Concretely, it is clude quadratic city-specific trends as a control in Xikt. The
plausible that over our decade-long observation period, ho- inclusion of these trends relaxes the DD assumption of no
tel firms have been strategically developing new properties cross-city time-varying unobservables that are correlated with
in areas of anticipated high demand. As high demand could both Airbnb supply and hotel revenue. A concern with the in-
also correlate with increased Airbnb adoption, this pattern of clusion of city-specific time-trends is that they can be confounded
competition could bias our estimation, because city-specific with hotels’ response to Airbnb (Wolfers 2006). Fortunately, our
increases in hotel room supply could decrease per hotel room data set covers a long pre-Airbnb period from 2003 to 2008,
revenue, and this effect could be misattributed to increased allowing us to estimate these trends on a large sample of pre-
Airbnb adoption. To guard against this concern, we construct a treatment observations. Second, we include city-month (e.g.,
control variable Hotel Room Supply−ikt, which measures the Austin-March) fixed effects to control for differences in seasonal
total supply of hotel rooms in the same city as hotel i (but demand patterns across the different cities. For instance, March in
excluding hotel i itself, thus the −i in the subscript) for each Austin is especially popular because of the SXSW festival. The
city-month fixed effects control for such seasonal differences.
time t. To construct this variable, we rely on the same monthly
Finally, we associate each city in our data with the nearest airport
panel of tax reports provided by the Texas Comptroller be-
and use the (log of the) number of passengers disembarking at that
cause, in addition to revenue, taxpayers have to report the
airport as their final destination as a control.
capacity of their properties with each filing. Therefore, Hotel An additional issue relates to the unobserved incentives of
Room Supply−ikt captures changes in competitors’ total room consumers who choose to list their homes on Airbnb. For
supply over time, including changes resulting from hotels example, Airbnb touts the help it provides to struggling or
expanding or shrinking and entering or exiting the market. This unemployed homeowners in paying their mortgage (Primack
control, which we also incorporate in Xikt, allows increases in 2012). Conceivably, an increase in the unemployment rate
the supply of hotel rooms provided by competitors to affect the could simultaneously drive Airbnb adoption and independently
room revenue of each hotel in our data, much as we hypothesize cause demand for hotels to soften. Therefore, failure to control
an increase in Airbnb rooms does. In addition, we control for for cross-city differences in the demographics could potentially
hotel i’s own capacity and quality over time, both of which may bias our estimation. In this case, the bias likely works in favor of
change (e.g., following renovations). We derive hotel capacity finding a negative Airbnb impact. To address this concern, we
from the tax data, and we use TripAdvisor ratings as a proxy for incorporate unemployment rate, the median annual wage, and
quality. population as controls in Xikt.
694 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

Identification Checks Figure 4

Before proceeding with estimation, we conduct a series of CORRELATION BETWEEN AIRBNB SUPPLY AND PRE-AIRBNB
identification checks to assess whether our proposed empirical (2007) CITY CHARACTERISTICS, WITH 95% CONFIDENCE
strategy is capable of recovering Airbnb’s causal impact on INTERVALS
hotel room revenue. Our DD identification strategy relies on
randomness in Airbnb adoption with respect to unobserved
Acc. Sector Empl. (Log) Hotel Occupancy Rate
city-specific time-varying factors (eikt ) that are also correlated
with changes in hotel room revenue (conditional on the control .4
variables we include). As with any study relying on observa- .04
tional data, there is no conclusive test of this assumption. .0
.00
However, we can exploit the richness of our data to check if this –.4
assumption is likely to hold in practice. Similar to Akerman, –.04
–.8
Gaarder, and Mogstad (2015), we perform two checks that
support the basis for our key identification assumption. Hotel Room Price (Log) Hotel Room Supply (Log)
First, we show that most variation in Airbnb adoption is .2
.25
explained by regressing (the log of) Airbnb supply on time-
.1
invariant city-specific factors, time fixed effects, and city- .00
.0
specific trends—all of which are part of the DD model. –.25
These factors explain 88% of the variation in Airbnb adoption, –.1
suggesting that our modeling assumption has a sound basis in
practice. Next, we repeat this regression with the addition Population (Log) Unemployment Rate
of city-specific time-varying observables that could poten- .3
tially be correlated with hotel room revenue: population, un- .1
.2
employment rate, and employment in the accommodation
.1 .0
sector. The inclusion of these factors does not increase the
explanatory power of the regression. .0 –.1
Second, we perform a randomization check by testing
whether pretreatment city characteristics predict future Airbnb 2008 2010 2012 2014 2008 2010 2012 2014
supply, where the time of treatment is taken to be 2008, when
Airbnb entered the Texas market. The idea behind this test is
that, assuming Airbnb adoption is exogenous (with respect to
hotel revenue), it should not be correlated with pretreatment Here, we have shown that various factors potentially af-
factors. To perform this identification check, for each city, we fecting hotel room revenue, including demographic trends as
compute its most recent pretreatment (2007) population, un- well as the structure and performance of the hospitality in-
employment rate, employment in the accommodation sector, dustry across different cities, are not correlated with local
hotel room supply, hotel room prices, and hotel occupancy patterns of Airbnb adoption. These checks increase our con-
rates. We then interact these predetermined factors (Zk,2007 ) fidence that the identification assumptions needed to estimate
with a vector of posttreatment year-month fixed effects (tt ) and Airbnb’s causal impact on hotel room revenue hold in our data.
regress them on Airbnb supply. Concretely, with the units of
analysis being post-2007 city-months, we estimate: Results and Economic Significance

0 We report the results of estimating Equation 1 using our
(2) log Airbnb Supplykt = Cityk + tt × Zk,2007 q + ekt :
cumulative Airbnb supply measure and incorporating the
Each coefficient in the vector of coefficients q is interpreted control variables in the first column of Table 2. We estimate the
as a correlation between a specific pretreatment characteristic coefficient b = −:039 ðp < :01Þ or, equivalently, a 10% in-
and Airbnb adoption in each posttreatment period (from crease in Airbnb listings is associated with a statistically
January 2008 onward). Figure 4 presents the estimated co- significant :39% decrease in monthly hotel room revenue. The
efficients q for each characteristic together with their 95% estimated coefficients for the controls have the signs and
confidence intervals. The only significant association we find magnitudes one would expect (e.g., increased hotel room
is between pre-Airbnb population and subsequent Airbnb supply and unemployment are both associated with decreased
adoption, and, for this reason, we include population as a hotel room revenue), though we note that our estimate for b
control in all our specifications. There also appears to be a without any controls (not presented) is comparable (b = −:035,
weak correlation with pre-Airbnb unemployment rate, further p < :01). As stated previously, we interpret a negative co-
justifying the inclusion of county-level unemployment rates efficient b as indicative of some Airbnb stays substituting for
as a control in Equation 1. Beyond these associations, we find hotel stays in cities with an established Airbnb presence.
no other discernible trend in the remaining coefficients Our estimates are sensitive to the functional form of the city-
(whose 95% confidence intervals always include the zero specific time trends. Table 3 compares models without city-
point, or no effect). It is especially reassuring that the pre- specific trends and with city-specific trends of increasing order.
treatment hotel industry structure—as captured by hotel room Without time trends, we estimate a positive (but insignificant)
supply, occupancy rates, room prices, and accommodation effect, whereas when trends are included, our estimate be-
sector employment in 2007—do not predict Airbnb supply comes negative and significant. To explain this observation,
from 2008 onward. we hypothesize that city-specific demand trends drive both
The Rise of the Sharing Economy 695

Table 2 The economic significance of this estimate is best under-

DD ESTIMATES OF THE IMPACT OF AIRBNB ON HOTEL ROOM stood in the context of Airbnb’s growth. For instance, in
REVENUE USING DIFFERENT MEASURES OF AIRBNB SUPPLY Austin, the Texas city with the highest Airbnb penetration, we
estimate that the impact of Airbnb over the past five years is
(1) (2) (3) approximately 10% of hotel room revenue (this calculation is
Revenue Revenue Revenue based on an increase in cumulative Airbnb supply from ap-
proximately 450 listings in 2010 to over 8,500 listings in 2014,
−.039***
yielding a revenue impact of 1 − ð8, 500/450Þ−:039 ). Consid-
log Cum. Airbnb Supply
(−4.40)
ering the high fixed costs associated with operating a hotel, this
log Inst. Airbnb Supply −.025*** figure could represent a significant fraction of hotel profits.
(TTL 3 mo.) (−2.82)
An alternative way to assess the economic significance of
log Inst. Airbnb Supply −.035*** Airbnb is through a direct comparison of the effects of Airbnb
(TTL 6 mo.) (−3.92) and hotel room supply on hotel revenue. By interpreting the
log Hotel Room Supply −.157*** −.154*** −.156*** coefficient of log Hotel Room Supply in the first column of
(−6.25) (−6.12) (−6.21) Table 2, we find that a 10% increase in the supply of hotel
log Capacity .034 .034 .034 rooms in Texas is associated with an approximately 1:6%
(1.50) (1.50) (1.50) decrease in Texas hotel room revenue, as compared with the
log Median Annual Wage −.212 −.364 −.290 smaller :39% decrease associated with a 10% increase in
(−.60) (−1.01) (−.82) Airbnb supply. It makes intuitive sense that increasing Airbnb
supply has a smaller impact than increasing hotel room supply,
Unemployment Rate −.060*** −.058*** −.058***
(−4.48) (−3.98) (−4.10) because we do not expect all Airbnb stays to substitute for a
hotel room stay. Nevertheless, the two effects are surprisingly
log Population .049 .061 .030
comparable in size: an increase in Airbnb supply has one-
(.33) (.42) (.21)
fourth the negative revenue impact of a corresponding increase
log Airline Passengers .150*** .138*** .148*** in hotel room supply. Taken at face value, this suggests that
(3.24) (2.94) (3.23)
incremental Texas Airbnb inventory does weakly substitute for
Is Reviewed −.057*** −.056*** −.057*** incremental hotel inventory. In addition, although the impact
(−3.03) (−2.94) (−2.98) of additional Airbnb supply is not as large, the significantly
TripAdvisor Star Rating .031*** .031*** .031*** higher costs associated with increasing hotel room supply
(6.93) (6.81) (6.86) implies that hotels are less likely to be able to expand inventory
N 294,383 294,383 294,383 as rapidly, an issue we return to subsequently.
Next, we estimate Equation 1 using our instantaneous
Within R2 .013 .011 .012
Airbnb supply measure. We present these results in the second
*p < .1. and third columns of Table 2. In the second column, we use a
**p < .05. TTL of three months, while in the third column, we use a TTL
***p < .01.
Notes: The dependent variable is log Hotel Revenueikt. Cluster-robust
of six months. In both cases, we obtain negative and significant
t-statistics (at the city level) are shown in parentheses. All specifications estimates, though the three-month TTL estimate for b is
include hotel fixed effects, year-month fixed effects, city-month fixed ef- smaller than the six-month TTL estimate (−:025 vs. −:035,
fects, and a city-specific quadratic time trend. p < :01 for both estimates). Our conclusion is that regressing
on either a cumulative or an instantaneous measure of Airbnb
hotel revenue and Airbnb supply. Therefore, when we omit supply captures a significant effect on hotel revenues due to
city-specific trends from the model, Airbnb supply stands in Airbnb.
for the omitted trend and becomes biased upward. In other This analysis reveals an additional insight regarding
words, increasing Airbnb supply is a sign of increasing de- Airbnb’s economic impact: the significant fluctuation in in-
mand for accommodation. This analysis guides the functional stantaneous Airbnb supply suggests that Airbnb’s impact on
form we choose to control for city-specific trends: because our hotel revenue will vary significantly over time. For instance, in
estimates remain unchanged when we move from quadratic to our data, we estimate that instantaneous Airbnb supply during
cubic trends, we settle for the simpler quadratic form. SXSW has historically been approximately 60% higher than
Table 3
DD ESTIMATES OF THE IMPACT OF AIRBNB ON HOTEL ROOM REVENUE USING CITY-SPECIFIC TRENDS OF INCREASING ORDER

(1) (2) (3) (4)

No Trends Linear Quadratic Cubic
log Cum. Airbnb Supply .009 −.025** −.039*** −.039***
(1.26) (−2.48) (−4.40) (−3.29)
N 294,383 294,383 294,383 294,383
Adj. within R2 .024 .012 .013 .011
*p < .1.
**p < .05.
***p < .01.
696 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

the rest of the year. In turn, this suggests that Airbnb impact on Table 4
hotel revenue is approximately 1.5 percentage points larger DD ESTIMATES OF THE IMPACT OF AIRBNB ON HOTEL
during SXSW (calculated as logð1:6Þ × :035). OCCUPANCY RATES AND PRICES
Variation in instantaneous supply is not the only reason why
Airbnb’s impact could be more pronounced during SXSW or (1) (2)
during other large events. Perhaps it is the case that Airbnb is Occupancy Rate Room Price
especially appealing to SXSW participants but has little or no
log Cum. Airbnb Supply −.005* −.019***
appeal to travelers the rest of the year. However, we find that
(−1.66) (−2.84)
this is not the case: when we censor SXSW from our data, the
elasticity that we estimate is unchanged (b = −:039, p < :05) log Hotel Room Supply −.132*** −.060***
using our cumulative supply measure. This result suggests that (−8.36) (−4.26)
Airbnb’s impact is not solely due to idiosyncratic preferences log Capacity .075*** −.007
of the SXSW demographic. (5.98) (−.43)
In summary, we find evidence that Airbnb’s impact in Texas log Median Annual Wage −.263 −.050
is observable through the lens of both cumulative and in- (−1.65) (−.26)
stantaneous supply measures. We further find that although its Unemployment Rate −.025*** −.016**
impact is most strongly concentrated in Austin and has (−4.50) (−2.47)
maximum impact correlated with periods of peak demand,
log Population −.004 .140**
the impacts are present year-round. Using the instantaneous (−.09) (2.02)
measure, we attributed seasonal variation in impact to a feature
that is unique to the sharing economy: supply flexibility. We log Airline Passengers .012 .044**
(.80) (2.22)
subsequently refine this top-level analysis to study how the
economic impacts are differentiated across different types of Is Reviewed −.060 −.129**
hotels and further unpack the effects of supply flexibility on the (−1.34) (−2.33)
peak pricing power of hotels. TripAdvisor Star Rating .002 .008**
(.86) (2.59)
Hotels’ Responses to Airbnb: Price, Occupancy, Entry,
N 264,172 264,172
and Exit
Within R2 .018 .012
So far, we have measured Airbnb’s impact in terms of hotel
revenue. Next, we turn to the nature of responses by incumbent *p < .1.
hotels to Airbnb market entry. In the short run, hotels could **p < .05.
***p < .01.
plausibly respond to Airbnb market entry through a price Notes: The dependent variable is Occupancy Rateikt in column 1 and log
response, an occupancy response, or both. In the long run, Hotel Room Priceikt in column 2. Cluster-robust t-statistics (at the city level)
Airbnb could cause hotel investments to change course, ul- are in parentheses. All specifications include hotel fixed effects, year-month
timately affecting market entry and exit. All these impacts can fixed effects, city-month fixed effects, and a city-specific quadratic time
be investigated naturally by measuring alternative dependent trend.
variables other than revenue.
Recall that hotel room revenue is the product of two quan- regression has a level-log interpretation.) In the second col-
tities: average occupancy rate within a given time period and umn, we regress against ADR, and we find that a 10% increase
ADR during that same period. A hotel that exerts no response in Airbnb supply is associated with a statistically significant
to a supply shock would exhibit a reduction in occupancy, (p < .01) price decrease of :19%. This suggests that affected
whereas an active manager could alternatively maintain oc- hotels actively respond by lowering their prices. Note that this
cupancy levels through a price response. A notable difference behavior is consistent with basic hotel revenue management
between the two responses is that the latter response, reduced practices, whereby hotels set prices according to the level of
prices, is a net benefit for all consumers seeking accommo- occupancy rates observed. Indeed, the hospitality industry has
dations, whether they use Airbnb or not. high fixed costs and low marginal costs, and therefore the
To estimate these effects, we reestimate the DD specifi- thinking is that it is better to “put a head in a bed”—at a low
cation in Equation 1, substituting the dependent variable first price—than not at all. To understand the economic significance
with occupancy rate and then with the log of ADR, and of these results, we can repeat the same calculation performed in
retaining the controls. Similar to the room revenue analysis, the previous section, which suggests that in Austin, Airbnb
these two quantities vary by hotel and by month. The price and negatively affected hotel prices by approximately 6%.
occupancy data set that we use masks individual hotel iden- Both the price and occupancy effects we investigated
tities; therefore, we cannot link it with the TripAdvisor data constitute immediate responses to Airbnb. In the long run,
on a hotel-by-hotel basis. Instead, we control for changes in Airbnb may also affect hotels’ entry, exit, and investment
hotel quality at the city level using the average hotel rating and decisions. To better understand the decision-making process
fraction of reviewed hotels in each city. We report these results and timetables of hotel development, we assembled a pro-
in Table 4. As reported in the first column of this table, we prietary data set (from STR) that records all ongoing Texas
find a small and weakly significant (p < .1) negative connection hotel projects, including both new construction and renova-
between increased Airbnb listings and occupancy rate. (Note tions (we do not have access to the historical record of
that, in contrast to our other dependent variables, occupancy completed renovations). STR records the dates that projects
rate is already expressed as a percentage and therefore we enter their various phases of development. Using this data set,
do not log-transform it. Therefore, the coefficient of this we computed the average time it takes to transition from one
The Rise of the Sharing Economy 697

Figure 5
AVERAGE TIME BETWEEN VARIOUS STAGES IN THE HOTEL PIPELINE CONSTRUCTION

529 days 100 days 228 days Entered 542 days Projected
Preplanning Planning Final planning
construction opening

phase to the next, which we diagram in Figure 5. The av- hotel types and Airbnb supply to the DD specification in
erage estimated time between preplanning and projected Equation 1:
opening is approximately four years, though there exists (3)
significant variation depending on the project type. There-
fore, hotel projects that were completed or were ongoing log Hotel Revenueikt = b1 log Airbnb Supplykt
during our observation period were likely conceived before + b2 log Airbnb Supplykt × Hotel Typei
Airbnb became a concern for the hotel industry. Indeed, + X0ikt g + ai + tt + eikt :
basic Poisson regressions of hotel entry and exit against
Airbnb supply (not reported here) yielded no correlation. As The coefficient of interest is b2 , which captures the differ-
Airbnb continues to become more established and hotels ential impact of Airbnb on the various segmentations by hotel
have time to incorporate Airbnb in their investment strat- type that we investigate. For our first segmentation, we define
egies, studying the nature of hotels’ long-term response will Hotel Typei as a categorical variable identifying each of the
be worth revisiting. hotel price segments used by STR. In the second and third
segmentations, we define Hotel Typei to be a binary indicator:
whether or not hotel i has conference or meeting space and
VARIATION OF IMPACT ACROSS HOTELS AND whether or not it is a chain, respectively.
ACROSS TIME The results of these analyses appear in Table 5. We start with
Which Hotels Are Most Affected and Why? price segmentation, presented in the first column. We estimate
We have provided evidence that Airbnb has a negative impact Equation 3, interacting hotel price segments with Airbnb
on hotel room revenue in Texas, treating hotels as a homo- supply. Here, we use luxury hotels as a reference category least
geneous set. In this section, we investigate various mecha- affected by Airbnb, motivated by the observation that these
nisms through which Airbnb could exhibit heterogeneous hotels are least comparable to Airbnb based on average room
impacts across different types of hotels and provide supporting price and amenities (e.g., pools, conference rooms, concierge).
empirical evidence. To motivate this analysis, we observe that We find the negative impact of Airbnb increasing as we move
although Airbnb can surely sometimes provide a suitable al- down the price tiers, with statistically significant interaction
ternative to hotels, one can hardly expect it to be a perfect coefficient estimates at the 1% level for each of the three lowest
substitute for all travel needs. Because Airbnb has its roots in tiers (midprice, economy, and budget). In contrast, we find
casual stays, including those involving shared accommoda- only a small negative and insignificant effect for the upscale
tions, we expect it to be a more attractive option for travelers and luxury segments (the latter being the reference level and
on a budget. Conversely, business travelers and vacationers thus its being captured by the main effect). From a managerial
who frequent high-end hotels are examples of consumer standpoint, this result has direct import: even though lower-end
groups we argue are less likely to substitute a hotel stay with an hotels in Texas account for a disproportionately small amount
Airbnb stay. Business travelers in particular are often less of room revenue as compared with upmarket hotels, they
price-sensitive because they are typically reimbursed for their nevertheless bear the brunt of the impact of the market entry
travel; moreover, they also make use of business-related of Airbnb. Our evidence suggests that consumers are increas-
hotel amenities not typically provided by Airbnb proper- ingly substituting Airbnb stays for lower-end hotels in Texas,
ties. Following this logic, we further isolate the impact of possibly identifying the former as offering better value at a
Airbnb on hotel room revenue by partitioning hotels in similar price point. Although this increased competition af-
three ways—each dividing hotels into segments that we fords consumers greater choice, it also places lower-end hotels
expect to be less vulnerable to Airbnb’s entry and other in regions with high Airbnb penetration at greater risk.
segments that we expect to be more vulnerable—then In the second column of Table 5, we report the results of the
estimating the additional interaction effects in our original segmentation of hotels catering to business travelers. We use
DD specification. In our first partition, we segment hotels those hotels having conference and meeting space as the
by price tier, following the STR hotel census, which di- reference category. The estimated coefficient b2 for the in-
vides hotels into five tiers: budget, economy, midprice, teraction between Airbnb supply and the indicator variable
upscale, and luxury. In our second partition, we differ- denoting absence of meeting space is negative and statistically
entiate hotels by their customer base: those that target significant (−:015, p < :01), suggesting that hotels lacking
business travelers versus those that do not. Finally, we business facilities are more affected by Airbnb. These results
consider the differentiated impact on chain hotels versus are consistent with our prior segmentation as well as with
independents. Airbnb’s marketing strategy to date, which has primarily
To estimate heterogeneous treatment effects, we estimate a targeted vacation travel. However, we note that, seeing a
new specification that adds an interaction effect between growth opportunity in the business travel segment, Airbnb
698 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

Table 5
DD ESTIMATES OF HETEROGENEITY IN AIRBNB’S IMPACT ON HOTEL ROOM REVENUE

(1) (2) (3)

Price Segment Meeting Space Operation
log Cum. Airbnb Supply −.016 −.033*** −.038***
(−1.61) (−3.58) (−4.23)
Price segment × log Cum. Airbnb Supply (ref. Luxury) Budget −.039***
(−5.39)

Economy −.031***
(−6.02)
Midprice −.020***
(−5.20)
Upscale −.007
(−1.45)
w/o Meeting Space × log Cum. Airbnb Supply −.015***
(−4.28)
Independent × log Cum. Airbnb Supply −.008**
(−2.53)
log Hotel Room Supply −.158*** −.158*** −.158***
(−6.26) (−6.27) (−6.26)
log Capacity .034 .035 .033
(1.49) (1.53) (1.50)
log Median Annual Wage −.225 −.219 −.215
(−.64) (−.62) (−.61)
Unemployment Rate −.060*** −.060*** −.060***
(−4.46) (−4.46) (−4.47)
log Population .086 .058 .047
(.63) (.39) (.31)
log Airline Passengers .151*** .150*** .150***
(3.28) (3.26) (3.24)
Is Reviewed −.032** −.047*** −.056***
(−2.12) (−2.64) (−2.97)
TripAdvisor Star Rating .026*** .029*** .031***
(7.15) (6.94) (7.00)
N 294383 294383 294383
2
Within R .018 .014 .013
*p < .1.
**p < .05.
***p < .01.
Notes: The dependent variable is log Hotel Revenueikt. Cluster-robust t-statistics (at the city level) are shown in parentheses. All specifications include hotel
fixed effects, year-month fixed effects, city-month fixed effects, and a city-specific quadratic time trend.

recently launched an initiative to attract more business trav- (−:038, p < :01), suggesting that hotels of both operation
elers (Isaac 2014). An interesting open question going forward structures were affected. However, the estimated interaction
is to what extent business travel will continue to differentiate coefficient for independent hotels (−:008, p < :05) is also
the impact of Airbnb on hotels. negative and statistically significant, suggesting that Airbnb has
The third distinction that we explore, which relates primarily indeed had a slightly larger impact on independent hotels.
to hotel operation, is between chain hotels (including fran- Overall, we find that independent hotels, hotels that do not
chises) and independent hotels. Unlike independent hotels, cater to business travelers, and lower-end hotels are all more
chain hotels allocate large marketing budgets to advertising, heavily affected by Airbnb than our respective reference
brand building, guest loyalty programs, and other tactics that categories, hotels without these characteristics. While these
should make them less vulnerable to competition. In addition, results help us better understand the most vulnerable hotel
many chains provide a more predictable standard of service, segments, and are certainly of importance to hoteliers, they
which further differentiates them from both Airbnb and in- also serve as robustness checks to our primary finding, in that
dependent hotels. We present this analysis in the third column the heterogeneous substitution effects they reveal align with
of Table 5, using chain hotels as a reference level. The overall the effects we hypothesized on the basis of the value propo-
effect from Airbnb remains negative and statistically significant sition to consumers that Airbnb offers.
The Rise of the Sharing Economy 699

Airbnb and Peak Pricing Power of Hotels Figure 6

Our analysis so far has focused on quantifying the extent to SXSW INTERACTIVE ATTENDANCE
which Airbnb supply substitutes for hotel room supply and its
differentiated impact across various hotels segments. Next, we
show that Airbnb supply is more than just a partial substitute

SXSW Interactive Attendance

30,000
for low-end hotel supply by proposing and empirically eval-
uating mechanisms whereby changes in Airbnb supply exhibit
fundamental differences from changes in hotel room supply.
In particular, we investigate the ability of Airbnb suppliers to 20,000
exhibit a more flexible response to peak seasonal demand and,
in so doing, crimp operating margins of hotel operators during
these peak periods.
10,000
During localized periods of peak demand, it is well un-
derstood that hotels can respond by raising prices,4 but they
cannot materially increase supply, because of high fixed costs
of new inventory. In contrast, many of the microentrepreneurs 0
providing Airbnb supply can elect to take inventory on and off 1998 2002 2006 2010 2014
the market on very short time scales and with near-zero cost.
Thus, the aggregate decisions of Airbnb providers comprise
both a price response and a supply response. Our subsequent
analysis is therefore motivated by the hypothesis that during which is the difference between year-over-year changes in
localized periods of peak demand, regions with flexible Airbnb high-season prices and low-season prices. Intuitively, double
supply more effectively absorb high seasonal demand than differencing allows us to adjust changes in high-season
regions in which Airbnb is not present. If the hypothesis is pricing (likely related to flexible scaling of Airbnb supply)
operative, the managerial implication is that the hotel industry’s using low-season changes in pricing (likely unrelated to
ability to command high rents during peak periods, which we flexible Airbnb supply) as a baseline. For instance, if year-
refer to as their “peak pricing power,” has become diminished over-year percentage price changes are equal during high and
in regions where Airbnb has actively entered the market, as low seasons, it is unlikely that they are jointly driven by
compared with other locales where Airbnb is less prevalent. Airbnb hosts flexibly scaling supply to accommodate peak
To motivate our definition of peak pricing power, consider demand during specific months of the year; thus, in this case,
that city-specific travel patterns are highly seasonal, and many we estimate DPi,y to be zero.
periods of peak demand predictably recur with an annual To study changes in peak pricing power of hotels in our data
frequency. Therefore, for each hotel-year in our data, we re- set, we considered the impact of two large events that take
fer to peak demand months as the “high season,” and the place annually in Texas: the SXSW festival in Austin in
remaining months as the “low season.” For each hotel i, we March, and the Texas State Fair in Dallas in October. Both
denote high-season prices during year y by pH i,y and low-season
events draw a large number of out-of-town visitors and have a
prices by pLi,y : Given these two quantities, we define hotel i’s substantial impact on area hotels’ bottom line as a result. Both
peak pricing power as events have also grown in popularity in the past decade, though
the much smaller SXSW festival has grown more rapidly in
i,y − log pi,y ,
Pi,y = log pH L
(4) percentage terms. Figure 6 displays attendance for SXSW
which can be interpreted as the percentage increase in prices Interactive. Together with SXSW Film and SXSW Music,
during high season. these are the major components of SXSW. March and October
Because we are interested in understanding changes represent the peak months for demand of hotels in Austin and
Dallas, respectively, measured in terms of both occupancy and
in—rather than absolute levels of—hotel pricing power as
ADR. In both cases, ADR and occupancy range between
Airbnb adoption increases, the quantity we analyze is the
8%–15% above the corresponding values for the rest of the
first difference of peak pricing power:
   
year, consistently over the past decade. However, Airbnb has
DPi,y = log pH grown much faster in Austin than it has in Dallas, suggesting
i,y − log pi,y − log pi,y−1 − log pi,y−1 ,
L H L
(5)
that if Airbnb affects peak pricing power, this effect will be
which can be interpreted as the year-over-year change in a more pronounced in Austin.
hotel’s ability to increase prices during the high season. We begin our analysis by visualizing changes in peak
Rearranging the terms in Equation 5 gives us the more pricing power. Motivated by our previous results, in which we
convenient form: found that Airbnb has a stronger impact on lower-end hotels,
    we segment hotels by price category and consider year-over-
(6) DPi,y = log pH
i,y − log pi,y−1 − log pi,y − log pi,y−1
H L L year changes in pricing power for the high season versus all
other months combined. Following Equation 6, for each hotel,
= D log pH
i,y − D log pi,y ,
L
we compute year-over-year changes in high- and low-season
i,y and D log pi,y ). Figure 7 displays the
prices (i.e., D log pH L

4For example, see evidence of surge pricing coinciding with the annual annual average of these quantities in Dallas for the period
shareholders’ meeting of Berkshire Hathaway in Omaha (The Economist 2010–2014. The gap between the solid line (changes in high
2016). season prices) and the dashed line (changes in low season
700 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

Figure 7
YEAR-OVER-YEAR CHANGES IN DALLAS HOTEL PRICES, BROKEN DOWN BY HOTEL PRICE LEVEL

Budget Economy Midprice Upscale Luxury

YoY Change in Prices

20%

10%

0%

–10%

2010 2015 2010 2015 2010 2015 2010 2015 2010 2015

Texas State Fair Rest of the year

Notes: The solid line displays changes during the State Fair of Texas (October) while the dashed line displays changes for the rest of the year.

prices) can be interpreted as the year-over-year change in power, despite the continued growth of SXSW. This effect
hotel pricing power during periods of peak demand. Vi- is especially pronounced for lower-end hotels, as our pre-
sually, we see little discernible difference between the two vious results would predict. Overall, these visualizations are
lines, with the gap between them always close to zero. This consistent with an explanation of flexible Airbnb supply
suggests that the pricing power of hotels in Dallas during coming online during SXSW to accommodate peak demand,
the state fair does not change significantly compared with thereby crimping the peak pricing power of lower-end hotels
the remainder of the year. specifically.
Next, we consider Austin. With the very rapid growth in As a final step in understanding the statistical significance of
SXSW, one could conjecture that the rate at which peak the effect we visualized, we estimate a descriptive model of
pricing power grows would outstrip that of nonpeak pe- changes in peak pricing power. The dependent variable we
riods. Consider the data plotted in Figure 8, in which we analyze is the seasonal price difference for each hotel i and
depict the year-over-year percentage changes in SXSW year-month t, which is defined as follows:
prices for March (solid line) compared with changes in
(7) =12 log pi,t = log pi,t − log pi,t−12 ,
prices during the remaining months of the year (dashed
line). During the initial period (roughly 2010–2012), vi- where =D is the seasonal difference operator of order D. As
sual evidence suggests the hotel pricing power for SXSW before, the interpretation of this quantity is the percentage
increased faster than during the rest of the year, consistent change in prices for hotel i compared with prices during the
with rapid growth in SXSW. In the second half of the same month the previous year. Unlike our visualization, in
period, 2012–2014, a new phenomenon is at work. The gap which we lumped all low-season months together, here we
between high- and low-season price changes starts to separately difference each month in our data. The model we
narrow as hotels lose the ability to exert the same pricing estimate takes the following triple-differences form:

Figure 8
YEAR-OVER-YEAR CHANGES IN AUSTIN HOTEL PRICES BROKEN DOWN BY HOTEL PRICE LEVEL

Budget Economy Midprice Upscale Luxury

YoY Change in Prices

20%

10%

0%

– 10%

2010 2015 2010 2015 2010 2015 2010 2015 2010 2015

SXSW Rest of the year

Notes: The solid line displays changes during SXSW (March) while the dashed line displays changes for the rest of the year.
The Rise of the Sharing Economy 701

(8) =12 log pi,t = b1 Austini + b2 Marcht + Yeart Figure 9

+ b3 Austini × Marcht + b4 Austini × Yeart PERCENTAGE CHANGES IN YEAR-OVER-YEAR AUSTIN HOTEL
+ b5 Marcht × Yeart PEAK PRICING POWER (SXSW VS. REST OF THE YEAR)
+ b6 Austini × Marcht × Yeart
+ ei,t , 10%
where Marcht is a dummy for March hotel-months, Yeart are
year fixed effects, and Austini is an indicator for hotels in

YoY Changes in SXSW

Hotel Pricing Power
5%
Austin. In addition to these explicit controls, seasonal dif-
ferencing eliminates both hotel fixed effects, as well as hotel-
month-specific linear trends in year-over-year price changes 0%
(such as a specific hotel increasing March prices by 5%
every year, April prices by 2% every year, etc.). The co-
efficients of interest are contained in the vector b6 , and they –5%
can be interpreted as changes in SXSW pricing power.
Intuitively, the model estimates March-specific changes in
pricing power in Austin and then adjusts these estimates for –10%
(1) March-specific changes in pricing power outside Austin
2007 2009 2011 2013 2015
and (2) non-March-specific changes in Austin. Figure 9
displays the coefficients b6 and their associated 95% con-
fidence intervals. Our conclusions here mirror our previous
observations: SXSW pricing power has significantly de-
clined as Airbnb popularity has increased, despite the fact Airbnb supply variable, (2) a specification test using an al-
the SXSW attendance has continued to grow steadily over ternative functional form of Airbnb supply, and (3) a matching
time. method, which we use as a more stringent alternative in de-
Unlike our previous analyses, the results in this section are fining (otherwise similar) treated and untreated properties.
descriptive. When jointly interpreted with our causal estimates
of Airbnb on hotel revenue, they paint a picture of Airbnb Distance-Based Market Definition
reducing hotel pricing power during periods of peak demand,
consistent with our hypothesis that the flexible provisioning In our analyses so far, we have assumed that travelers may
of inventory to accommodate peak demand is a distinguish- substitute between hotels and Airbnb properties within the
ing feature of the sharing economy. Further research should same city irrespective of the distance between properties
create a better understanding of this phenomenon with both (Hollenbeck [2014], who analyzes the same Texas data, also
more sophisticated modeling and data that span other large uses city-level markets to model competition between hotels).
events. Although this seems reasonable for smaller cities in our data
In closing, we compare our observations with another set, it is likely a less reliable approximation of how travelers
sharing economy study that observes flexible supply entering form consideration sets when visiting sprawling cities like
the Uber market during peak periods. Hall, Kendrick, and Houston. To test the sensitivity our results to narrower market
Nosko (2015) study the effectiveness of surge pricing on Uber, definitions, we next analyze a proximity-based Airbnb supply
whereby drivers are incentivized to drive at peak times through measure. Specifically, for each hotel in our data, we measure
higher payment multipliers. The study reports that the surge the cumulative number of Airbnb rooms within a fixed radius
pricing mechanism is effective and leads to reduced wait times at any given point in time. This measure allows different hotels
during periods of peak demand, comparable to levels observed in the same city to face different levels of competition by
in low-demand periods. In comparison, our work shows that Airbnb. To be consistent with this market definition, we also
a similar incentive drives Airbnb suppliers to scale room define hotel competition in the same way: for each hotel, we
supply during periods of peak demand, when they can com- measure the number of hotel rooms within the same fixed
mand higher rents. We witness this effect indirectly, through radial distance.
decreased peak pricing power of hotels in the high season. Using a hotel-specific market definition introduces a new
Whereas Uber directly incentivizes increased supply through type of endogeneity concern that we need to address. For any
central setting of price multipliers, a similar effect arises in given hotel, increased competition by nearby Airbnb prop-
Airbnb without direct control but instead through the collec- erties or hotels is likely correlated with increased demand for
tive, decentralized decision making of its suppliers. Notably, that hotel. In other words, even within the same city, new hotel
Airbnb is moving toward a variable pricing model, in which rooms and Airbnb properties are more likely to be located near
it dynamically adjusts listing prices in response to demand.5 hotels that are facing growing demand by travelers. The city-
specific trends we previously included do not allow for cor-
ROBUSTNESS CHECKS relation between local measures of Airbnb and within-city
We perform three checks to reinforce the causal interpretation hotel revenue variation. Therefore, we estimate a model that
of our DD estimate: (1) a distance-sensitive definition of the includes hotel-specific quadratic trends. This model, known
as the correlated random trends, or random growths, model
5See Financial Times (2015). We thank Avi Goldfarb for pointing out this (Murtazashvili and Wooldridge 2008; Wooldridge 2005,
connection. 2009), allows for correlation between the hotel-specific trend
702 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

component and time-varying observables. The specific model Table 6

we use takes the following form: LOCAL MEASURES OF AIRBNB AND HOTEL ROOM SUPPLY

(9) log Hotel Revenueikt = b log Local Airbnb Supplyikt

(1) (2)
+ X0ikt g + hi + ai1 t + ai2 t2 Within 1 Mile Within 5 Miles
+ tt + Cityk × Montht + eikt : D2 (log Local Cum. Airbnb Supply) −.032** −.025**
(−1.97) (−2.15)
Following standard practice (see, e.g., Wooldridge 2005),
we eliminate the hotel-specific quadratic trends by second- D2 (log Local Hotel Room Supply) .006 .016
differencing our data.6 Second-differencing requires the (.95) (.69)
sacrifice of the first two monthly observations of each hotel. D2 (log Hotel Room Supply) −.005 −.004
Our decade-long panel is sufficient to comfortably accom- (−.19) (−.18)
modate this transformation. The final model we estimate is D2 (log Capacity) −.026 −.026
(−1.18) (−1.18)
(10) D ðlog Hotel Revenueikt Þ = bD ðlog Local Airbnb Supplyikt Þ
2 2

D2 (log Median Annual Wage) −.131 −.129
+ D2 X0ikt g + ai2 + tt + (−.38) (−.37)
Cityk × Montht + eikt , D2 (Unemployment Rate) −.017* −.017*
(−1.88) (−1.86)
where D is the second-difference operator. Note that dif-
2
D2 (log Population) −.156 −.155
ferencing also eliminates the hotel fixed effect hi. The model (−.89) (−.89)
can be estimated using the within transformation to eliminate
the hotel-specific intercepts ai2. We continue to cluster errors D2 (log Airline Passengers) .174*** .176***
(5.49) (5.57)
at the city level.
Table 6 displays our results. In the first column, we display D2 (Is Reviewed) .043*** .043***
(2.65) (2.67)
our results using a radius of one mile around each hotel. We
estimate a significant Airbnb effect with magnitude −:032 D2 (TripAdvisor Star-Rating) −.007* −.007*
(p < :05), similar to our prior estimates. In the second column, (−1.66) (−1.67)
we experiment with a larger radius of five miles. Our estimate N 285187 285187
is again significant; however, it is now smaller in magnitude Within R2 .0011 .0011
(−:025, p < :05). One natural interpretation for the difference
*p < .1.
between these two estimates is that the greater the distance
**p < .05.
between Airbnb listings and hotels within a city, the less likely ***p < .01.
travelers are to substitute between the two. Overall, our results Notes: The dependent variable is D2 (log Hotel Revenueikt), where D2 is the
support our prior hypothesis that Airbnb directly affects hotel second difference operator. Cluster-robust t-statistics (at the city level) are
shown in parentheses. All specifications include hotel fixed effects, year-
revenue, while producing the additional insight that this impact month fixed effects, city-month fixed effects, and a city-specific quadratic
is sensitive to the distance between hotels and Airbnb listings time trend.
within a city.
(11) log Hotel Revenueikt = b1 IðAirbnb Supply 1 − 99Þkt
Alternative Functional Form for Airbnb Supply + b2 IðAirbnb Supply 100 − 999Þkt
The second robustness check we perform guards against a + b3 IðAirbnb Supply 1; 000 + Þkt
functional specification concern in Equation 1: regressing + hi + tt + X0ikt g + eikt ,
the log of Airbnb supply on the log of hotel room revenue
where the Ið:Þ are dummy indicators for the corresponding
implicitly assumes a constant elasticity relationship be-
ranges of Airbnb supply.
tween the two quantities. While this might be a reasonable This model allows the effect of Airbnb to vary depending on
assumption in data with limited variation in Airbnb supply, the number of Airbnb listings present in each city during a
the constant elasticity assumption is likely violated in our given period. In this model, each of three estimated coefficients
setting, as it is implausible that doubling Airbnb supply associated with the three levels of the categorical Airbnb sup-
from 1 to 2 units will have the same effect on hotel room ply variable we use represents a percentage change in hotel
revenue as doubling Airbnb supply from 100 to 200 units. revenue. We estimate this model by replacing Airbnb supply
To ensure that our results are not driven by this modeling with this new categorical variable in Equation 1 using zero
choice, we model Airbnb supply nonparametrically using a Airbnb units as the reference level. We present our results for
categorical variable, which takes on one of the following cumulative Airbnb supply in the first column of Table 7. These
(roughly log-binned) values: 0 Airbnb units, 1–99 Airbnb estimates provide directly interpretable estimates of Airbnb’s
units, 100–999 Airbnb units, and 1,000+ Airbnb units. economic impact. We find that increasing levels of Airbnb
Specifically, we estimate: penetration have proportionally larger impacts on hotel room
revenue, as we would expect. For example, at Airbnb adoption
6To understand this, note that first-differencing transforms the lin-
rates exceeding 1,000 rooms, the estimate (−:085, p < :05), in-
ear trend to a constant, and the quadratic trend to a linear trend:
ai1 t + ai2 t2 − ai1 ðt − 1Þ − ai2 ðt − 1Þ2 = ai1 + ai2 ð2t − 1Þ Taking a second
dicates (because we are now working with a log-level specifi-
difference, we arrive at ai1 + ai2 ð2t − 1Þ − ai1 − ai2 ½2ðt − 1Þ − 1 = 2ai2 , cation) an average impact of 8.5% on hotel room revenue. These
which has ai2 as a hotel-specific intercept. results are in line with our main specification estimates.
The Rise of the Sharing Economy 703

Table 7 globally, and we do not have reason to believe the incumbent

ROBUSTNESS CHECKS hotel industry is better protected from Airbnb in those other
cities. Second, our findings indicate that hotels facing seasonal
(1) (2) tourism demand have a structural susceptibility to Airbnb that is
Revenue Revenue universal, not local to Austin.7
log Cum. Airbnb Supply −.043***
(−4.25) A Matching Estimate Using CEM
Cum. Airbnb Supply (ref. 0) Because Airbnb adoption is clearly not random by design, to
1–99 listings −.020 provide evidence in support of the DD identification assump-
(−1.40) tions, we showed that observed pretreatment demographic and
100–999 listings −.063** market characteristics do not correlate with the patterns of Airbnb
(−2.05) adoption we observe in our data, which is what we would ex-
1,000+ listings −.085**
pect with exogenous Airbnb entry. Here, we combine DD with
(−2.16) matching to further limit the potential for unobserved confounders
biasing our estimates. To explain the matching approach, first
log Hotel Room Supply −.152*** −.151***
(−6.06) (−5.70)
recall our source of identification: roughly speaking, for each
“treated” hotel (i.e., a hotel affected by Airbnb competition), our
log Capacity .034 .075** DD analysis constructs a counterfactual outcome using a set of
(1.50) (2.40)
“untreated” hotels (i.e., hotels unaffected by Airbnb). The in-
log Median Annual Wage −.432 −.246 tuition behind matching is that the more similar treated and un-
(−1.15) (−.66) treated hotels are in their observed characteristics, the less likely
Unemployment Rate −.059*** −.055*** they are to differ in unobserved ways, including bias-inducing
(−3.87) (−3.55) factors. Matching methods aim to reduce endogeneity con-
log Population .128 .152 cerns by ensuring comparability between treated and untreated
(.78) (1.12) units (Heckman and Navarro-Lozano 2004). Although various
log Airline Passengers .127*** .165***
matching methods exist, here we use the CEM procedure (Iacus,
(2.65) (3.52) King, and Porro 2012) because it is intuitive and works well
with categorical data (like most hotel characteristics).
Is Reviewed −.056*** −.050***
Coarsened exact matching occurs in two steps. First, hotels
(−2.93) (−2.72)
are stratified based on observed characteristics; we use price
TripAdvisor Star Rating .031*** .034*** segment (budget to luxury), operation (independent or chain),
(6.79) (6.40) and hotel chain affiliation (e.g., Hilton, Marriott), if any. After
N 294,383 188,818 this first step, each stratum contains hotels that are identical on
Within R 2
.011 .015 the basis of these characteristics. For instance, a single stratum
contains all upscale Marriott hotels, some of which are
CEM sample No Yes
eventually treated and some of which are not. In a setting with a
*p < .1. binary treatment indicator, it is clear which units are eventually
**p < .05. treated. In our case, in which treatment intensity varies, we
***p < .01. make this distinction by defining hotels in cities that have no
Notes: The first column tests an alternative functional form for cumulative
Airbnb supply; the second column estimates Airbnb’s impact using a CEM- Airbnb penetration by the end of our observation period as
matched subset of hotels. The dependent variable is log Hotel Revenueikt. All untreated, and the remaining hotels as treated. One could argue
specifications include hotel fixed effects, year-month fixed effects, city- that this definition of treatment is too permissive; although we
month fixed effects, and a city-specific quadratic time trend. do not present these results for brevity, we found our CEM
It is reassuring that we find no statistically significant effect analysis to be robust to alternative definitions of treated units,
at low levels of Airbnb supply. In fact, this model clarifies that such as hotels in cities that eventually have at least 100 Airbnb
Austin is the primary driver of the Airbnb effect we estimate— listings. In the second step of CEM, we discard strata con-
no other city in our data had more than 1,000 (cumulative) taining only treated or untreated hotels and renormalize
Airbnb listings during our observation period, and therefore the weights of observations in the remaining strata to place equal
8.5% decrease in revenue we estimate for Airbnb penetration at weight on treated and untreated units in each stratum. Ap-
this level is driven by Austin. Indeed, we also find that our plying CEM to our data leaves us with 1; 946 hotels.8 Finally,
estimate between 100–999 listings is also primarily driven by we reestimate the DD specification in Equation 1 on the subset
Austin—deleting Austin from the data and reestimating the of matched hotels using the CEM weights. Conceptually, DD
model gives a negative but statistically insignificant impact. Even
7A recent report by CBRE Hotels’ Americas Research ranks Austin as the
though a few cities in our data experienced Airbnb adoption
13th most vulnerable city to Airbnb in the United States (New York City
at this level, those cities were larger, had more hotel rooms, and ranks first), taking into account both the ratio of Airbnb units to hotel rooms
did so late in our observation period (see Table 1). Therefore, in each market, as well as hotel room and Airbnb prices (see http://rss.
it would be rather surprising had Airbnb exerted a statistically hsyndicate.com/file/152006083.pdf).
8Coarsened exact matching entails a trade-off between matching granu-
significant impact on hotel revenues outside of Austin. Does this
larity and the number of discarded observations. We chose our matching
result suggest that the Airbnb effect is specific to Austin? We criteria to strike a reasonable balance between ensuring that units within each
think this is unlikely for two reasons. First, Airbnb is popular in stratum are similar and discarding too many observations. Our results our
many other destinations outside of Texas both nationally and robust to alternate matching criteria.
704 JOURNAL OF MARKETING RESEARCH, OCTOBER 2017

on the CEM sample estimates a treatment effect within each face from peer-to-peer platforms has several unique features
stratum of comparable treated and untreated hotels, then av- that differentiate it from competition with other firms. First,
erages these treatment effects to arrive at a final estimate. We the Airbnb platform has near-zero marginal cost, in that a
report this estimate in the second column of Table 7. We find new room can be incrementally added to (or removed from)
that the effect of Airbnb on hotel room revenue is robust to the platform with negligible overhead. Because of this,
CEM, attaining a magnitude (b = −:043, p < :01) that is highly Airbnb can scale supply in an almost frictionless manner
comparable to our original estimate (b = −:039) reported in to meet demand, even on short timescales. By contrast,
column 1 of Table 2. increasing hotel room supply involves buildout, causing
significant marginal costs for hotel chains. As we have
DISCUSSION AND CONCLUSIONS shown, this unique feature of Airbnb has already signifi-
The sharing economy has recently emerged as a viable al- cantly affected hotel’s pricing power during periods of
ternative to fulfilling a variety of consumer needs, ranging peak demand. Second, Airbnb offers a much wider range
from prepared meals to cars to overnight accommodations, that of products and services than hotels: Airbnb users can rent
were previously provided primarily by firms rather than en- anything from an apartment to a yurt. More importantly,
trepreneurial individuals. As the size of the sharing economy because Airbnb leverages existing housing inventory, it can
has grown, so has the magnitude of its economic impacts. Our potentially expand supply wherever houses and apartment
work is among the first to provide empirical evidence that the buildings already exist. This is in contrast to hotels, which
sharing economy is significantly changing consumption pat- must be built at locations in accordance with local zoning
terns, as opposed to generating purely incremental economic requirements. Therefore, competition by Airbnb is poten-
activity. Focusing on the case of Airbnb, a pioneer in shared tially harder for incumbents to adapt to, compared with com-
accommodations, we estimate that its entry into the Texas petition by other hotel firms.
market has had a quantifiable negative impact on local hotel Turning to consumers, we show that hotels in areas where
room revenue. The substitution patterns we observe strongly Airbnb has an established presence have responded to in-
suggest that Airbnb provides a viable, but imperfect, alter- creased competition by lowering their prices, which harms
native for certain traditional types of overnight accommoda- their revenues but benefits travelers, even those who do not use
tion. Our analyses pinpoint lower-end hotels, and hotels not Airbnb. In addition to reduced prices, consumers also benefit
catering to business travelers, as those that are most vulnerable from increased variety provided through peer-to-peer plat-
to increased competition from rentals enabled by firms like forms. Furthermore, consumers on the supply side benefit
Airbnb. Moreover, our work provides evidence that Airbnb through additional income generated by providing goods and
supply is differentiated from hotel supply, as shown by Airbnb services through peer-to-peer platforms.
supply-side flexibility and carrying through to the impact on Finally, our results have implications for policy makers.
hotel peak pricing power. Municipal revenues rely in part on tax receipts from well-
Our work has some limitations that could be addressed in regulated industries such as hotels and taxicabs. With demand
further research. First, one must recognize that our findings shifting away from these incumbent firms, and to the extent
are representative of the state of Texas; directly generalizing that regulation and taxation of peer-to-peer platforms proves
them to other markets may not be appropriate given the to be more challenging, the bottom line of cities with an
varying of dynamics of supply and demand for accommo- established Airbnb presence could be hurt in the short run. Of
dation across different regional markets. Additional studies course, peer-to-peer platforms can also bring about increased
that model the impact of Airbnb across these markets could demand, which would provide direct benefit to cities, making
be a useful contribution. A second limitation of work is that we the net impact on cities more difficult to measure. Quantify-
analyze properties listed only on Airbnb, but not properties ing the net impact of peer-to-peer platforms remains a fruitful
available through related vacation rental platforms such as direction for future research.
HomeAway and VRBO. We do not believe that our results are Returning to the thesis that the sharing economy has the
significantly affected by these competitors, because these firms potential to transformatively increase social welfare, as es-
primarily serve the smaller vacation rental market; moreover, poused by Botsman (2012) and others, we assert that a large
they have not experienced the extremely rapid growth of population of people worldwide have indeed benefited from
Airbnb. Nevertheless, researchers could investigate the impact Airbnb—not only hosts who derive incremental income by
of all these firms in aggregate, or individually. A final limi- renting properties through Airbnb and guests who select an
tation of our study pertains to the precise characterization of Airbnb rental as an alternative to a hotel stay but also those
hotels’ response: in this article, we have analyzed two metrics, consumers who benefit from lower prices and increased
price and occupancy rate, that managers can invoke as a re- competition in the accommodation industry. More broadly,
sponse in the short run. On longer time scales, hotels have one can weigh the positive change the sharing economy can
other ways of responding to Airbnb, including alterations bring about not only by providing imperfect substitutes for
to their investment schedules, to their entry and exit de- existing products but also, through an application of Say’s
cisions, and to their marketing campaigns. New promotions, Law, by generating demand that did not previously exist,
advertising campaigns, and even repositioning to provide through the supply of new products and services. Harkening
more personalized Airbnb-like services are all options. back to arguments Airbnb has made, the supply of in-
Work that either informs or interprets the shape of the expensive accommodations can increase travel and tourism
response by hotels in the longer run will address interesting spend overall, and thus, the sharing economy could be a net
open questions. producer of new jobs. However, these positives must be
Our results have direct implications for hotels, travelers, and evaluated against various costs, including those estimated in
policy makers. For hotel managers, the competition their firms this article.
The Rise of the Sharing Economy 705

REFERENCES Isaac, Mike (2014), “Airbnb Expands into Business Travel,” The
New York Times (July 28), bits.blogs.nytimes.com/2014/07/28/
Airbnb (2015), “Airbnb Summer Travel Report: 2015,” http://blog.
airbnWb-expands-into-business-travel.
atairbnb.com/wp-content/uploads/2015/09/Airbnb-Summer-Travel-
Jin, Ginger Zhe, and Marc Rysman (2012), “Platform Pricing at
Report-1.pdf.
Sports Card Conventions,” NBER Working Paper No. 17959,
Akerman, Anders, Ingvil Gaarder, and Magne Mogstad (2015),
National Bureau of Economic Research.
“The Skill Complementarity of Broadband Internet,” Quarterly
Kosová, Renáta, and A. Cathy Enz (2012), “The Terrorist Attacks of
Journal of Economics, 130 (4), 1781–824.
9/11 and the Financial Crisis of 2008 The Impact of External
Baker, Kathleen, and Alex Coulter (2007), “Terrorism and Tourism:
Shocks on US Hotel Performance,” Cornell Hospitality Quar-
The Vulnerability of Beach Vendors’ Livelihoods in Bali,” Jour-
terly, 53 (4), 308–25.
nal of Sustainable Tourism, 15 (3), 249–66.
Kroft, Kory, and G. Devin Pope (2014), “Does Online Search Crowd
Bakos, J. Yannis (1997), “Reducing Buyer Search Costs: Impli-
Out Traditional Search and Improve Matching Efficiency? Evi-
cations for Electronic Marketplaces,” Management Science, 43 (12),
dence from Craigslist,” Journal of Labor Economics, 32 (2),
1676–92.
259–303.
Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan (2004),
Landsman, Vardit, and Stefan Stremersch (2011), “Multihoming in
“How Much Should We Trust Differences-in-Differences Esti-
Two-Sided Markets: An Empirical Inquiry in the Video Game
mates?” Quarterly Journal of Economics, 119 (1), 249–75.
Console Industry,” Journal of Marketing, 75 (6), 39–54.
Botsman, Rachel (2012), “The Currency of the New Economy is
Martin, Elliot, Susan A. Shaheen, and Jeffrey Lidicker (2010),
Trust,” TED, http://www.ted.com/talks/rachel_botsman_the_currency_ “Impact of Carsharing on Household Vehicle Holdings.,” Trans-
of_the_new_economy_is_trust.html. portation Research Record, 2143 (1), 150–58.
Cervero, Robert, Aaron Golub, and Brendan Nee (2007), “City Murtazashvili, Irina, and Jeffrey M. Wooldridge (2008), “Fixed
CarShare: Longer-Term Travel Demand and Car Ownership Effects Instrumental Variables Estimation in Correlated Ran-
Impacts.,” Transportation Research Record, 1992 (1), 70–80. dom Coefficient Panel Data Models,” Journal of Econometrics,
Donald, Stephen G., and Kevin Lang (2007), “Inference with 142 (1), 539–52.
Difference-in-Differences and Other Panel Data,” Review of Eco- O’Connor, Noëlle, Mary Rose Stafford, and Gerry Gallagher
nomics and Statistics, 89 (2), 221–33. (2008), “The Impact of Global Terrorism on Ireland’s Tourism
The Economist (2016), “Buffett’s Revenge,” (January 7), https:// Industry: An Industry Perspective,” Tourism and Hospitality
www.economist.com/news/finance-and-economics/21685502- Research, 8 (4), 351–63.
services-airbnb-are-altering-economics-hotel-business-buffetts. Parker, Geoffrey G., and Marshall W. Van Alstyne (2005), “Two-
Einav, Liran, Chiara Farronato, and D. Jonathan Levin (2016), Sided Network Effects: A Theory of Information Product De-
“Peer-to-Peer Markets,” Annual Review of Economics, 8 (1), sign,” Management Science, 51 (10), 1494–504.
615–35. Phillips, Matt (2015), “New York City Hotel Rooms Are Getting
Farrell, Maureen, and Greg Bensinger (2016), “Airbnb’s Funding Cheaper Thanks to Airbnb,” (February 9), Quartz, http://qz.com/
Round Led by Google Capital,” The Wall Street Journal (Sep- 341292.
tember 22), http://www.wsj.com/articles/airbnb-raises-850-million- Primack, Dan (2012), “How AirBNB Helps Users Save Their
at-30-billion-valuation-1474569670. Homes,” Fortune (August 16), http://fortune.com/2012/08/16/
Financial Times (2015), “Airbnb Shakes Up Pricing Model to Meet how-airbnb-helps-users-save-their-homes.
Surging Demand,” (November 12), http://www.ft.com/cms/s/0/ Rochet, Jean-Charles, and Jean Tirole (2003), “Platform Compe-
bc875c4c-88ee-11e5-9f8c-a8d619fa707c.html. tition in Two-Sided Markets,” Journal of the European Economic
Fradkin, Andrey (2014), “Search Frictions and the Design of Online Association, 1 (4), 990–1029.
Marketplaces,” working paper. Rysman, Marc (2009), “The Economics of Two-Sided Markets,”
Fradkin, Andrey, Elena Grewal, David Holtz, and Matthew Pearson Journal of Economic Perspectives, 23 (3), 125–43.
(2014), “Reporting Bias and Reciprocity in Online Reviews: Seamans, Robert, and Feng Zhu (2013), “Responses to Entry in
Evidence from Field Experiments on Airbnb,” working paper. Multi-Sided Markets: The Impact of Craigslist on Local News-
Hall, Jonathan V., Cory Kendrick, and Chris Nosko (2015), “The papers,” Management Science, 60 (2), 476–93.
Effects of Uber’s Surge Pricing: A Case Study,” working paper. Valverde, Santiago Carbó, Sujit Chakravorti, and Francisco
Hall, Jonathan V., and Alan B. Krueger (2015), “An Analysis of the Rodriguez Fernandez (2010), “Regulating Two-Sided Markets:
Labor Market for Uber’s Driver-Partners in the United States,” An Empirical Investigation,” technical report, Federal Reserve
working paper, Princeton University, http://arks.princeton.edu/ Bank of Chicago.
ark:/88435/dsp010z708z67d. Weyl, Eric Glen (2010), “A Price Theory of Multi-sided Platforms,”
Heckman, James, and Salvador Navarro-Lozano (2004), “Using American Economic Review, 100 (4), 1642–72.
Matching, Instrumental Variables and Control Functions to Es- Wolfers, Justin (2006), “Did Unilateral Divorce Laws Raise Divorce
timate Economic Choice Models,” Review of Economics and Rates? A Reconciliation and New Results,” American Economic
Statistics, 86 (1), 30–57. Review, 96 (5), 1802–20.
Hollenbeck, Brett (2014), “The Spread of Horizontal Chains: Wooldridge, Jeffrey M. (2005), “Fixed-Effects and Related Estimators
Efficiency or Market Power?” working paper, University of Cal- for Correlated Random-Coefficient and Treatment-Effect Panel Data
ifornia, Los Angeles. Models,” Review of Economics and Statistics, 87 (2), 385–90.
Iacus, Stefano M., Gary King, and Giuseppe Porro (2012), “Causal Wooldridge, Jeffrey M. (2009), “Correlated Random Effects
Inference Without Balance Checking: Coarsened Exact Match- Models with Unbalanced Panels,” working paper, Michigan State
ing,” Political Analysis, 20 (1), 1–24. University.