NBER WORKING PAPER SERIES
THE WELFARE EFFECTS OF PEER ENTRY IN THE ACCOMMODATION MARKET:
THE CASE OF AIRBNB
Chiara Farronato
Andrey Fradkin
Working Paper 24361
http://www.nber.org/papers/w24361
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
February 2018
Katie Marlowe and Max Yixuan provided outstanding research assistance. We thank Nikhil
Agarwal, Susan Athey, Matt Backus, Liran Einav, Christopher Knittel, Jonathan Levin, Greg
Lewis, Chris Nosko, Debi Mohapatra, Ariel Pakes, Paulo Somaini, Sonny Tambe, Dan
Waldinger, Ken Wilbur, Kevin Williams, Georgios Zervas, and numerous seminar participants
for feedback. We are indebted to Airbnb’s employees, in particular Peter Coles, Mike Egesdal,
Riley Newman, and Igor Popov, for sharing data and insights. We also thank Duane Vinson at
STR and Sergey Shebalov at Sabre for sharing valuable data insights. Airbnb reviewed the paper
to make sure that required confidential information was reported accurately. STR reviewed the
paper to verify that all data and information provided by STR and the STR SHARE Center were
correctly cited. Farronato has no material financial relationship with entities related to this
research. Fradkin was previously an employee of Airbnb, Inc. and holds stock that may constitute
a material financial position. The views expressed are those of the authors and do not necessarily
reflect the views of the National Bureau of Economic Research.
NBER working papers are circulated for discussion and comment purposes. They have not been
peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies
official NBER publications.
© 2018 by Chiara Farronato and Andrey Fradkin. All rights reserved. Short sections of text, not
to exceed two paragraphs, may be quoted without explicit permission provided that full credit,
including © notice, is given to the source.
The Welfare Effects of Peer Entry in the Accommodation Market: The Case of Airbnb
Chiara Farronato and Andrey Fradkin
NBER Working Paper No. 24361
February 2018
JEL No. D4,D6,L1,L22,L23,L85,L86
ABSTRACT
We study the effects of enabling peer supply through Airbnb in the accommodation industry. We
present a model of competition between flexible and dedicated sellers - peer hosts and hotels -
who provide differentiated products. We estimate this model using data from major US cities and
quantify the welfare effects of Airbnb on travelers, hosts, and hotels. The welfare gains are
concentrated in locations (New York) and times (New Years Eve) when hotels are capacity
constrained. This occurs because peer hosts are responsive to market conditions, expand supply
as hotels fill up, and keep hotel prices down as a result.
Chiara Farronato
Harvard Business School
Morgan Hall 427
Soldiers Field
Boston, MA 02163
and NBER
cfarronato@hbs.edu
Andrey Fradkin
MIT Sloan School of Management
E62 Room 412
30 Memorial Dr.
Cambridge, MA 02142
afradkin@gmail.com
A data appendix is available at http://www.nber.org/data-appendix/w24361
1 Introduction
The Internet has greatly reduced entry and advertising costs across a variety of industries.
As an example, peer-to-peer marketplaces such as Airbnb, Uber, and Etsy currently provide
a platform for small and part-time peer providers to sell their goods and services. Several
of these marketplaces have grown quickly and become widely known brands. In this pa-
per, we study the determinants and e↵ects of peer production in the market for short-term
accommodation, where Airbnb is the main peer-to-peer platform and hotels are incumbent
suppliers.
We present a theoretical model of competition between incumbent hotels and peer hosts.
We then use data from top US cities to test the model hypotheses about the entry of peer
supply, and to quantify the e↵ects of this entry on travelers, incumbent hotels, and peer
hosts. We find that Airbnb generated $41 of consumer surplus per room-night and $26 of
host surplus while reducing variable hotel profits from accommodations by up to 3.7%. This
resulted in a total welfare gain of $137 million in 2014 from Airbnb in these cities and this
e↵ect was concentrated in locations (New York) and times (New Years Eve) where hotel
capacity was constrained.
Since its founding in 2008, Airbnb has grown to list more rooms than any hotel group in
the world. Yet Airbnb’s growth across cities and over time has been highly heterogeneous,
with supply shares ranging from over 15% to less than 1% across major US cities at the end
of 2014. Airbnb’s entry has also prompted policy discussion and varied regulation in many
cities across the world. In order to understand Airbnb’s growth and its e↵ects, we propose
a simple demand and supply framework where accommodations can be provided by either
dedicated or flexible supply – hotels vs peer hosts.
The role of Airbnb in our framework is to lower entry costs for peer hosts. This reduction
in entry costs is similar across cities but the benefits of hosting travelers vary. Prices and
occupancy rates, as well as marginal costs a↵ect the benefits of hosting travelers. In the
long run, our model predicts higher entry of peer supply in cities with higher prices and
occupancy rates, and lower peers’ marginal costs. Prices and occupancy rates are in turn
determined by the trend and variability in the number of travelers, as well as geographic and
bureaucratic constraints to the expansion of hotel capacity. Marginal costs are determined
by the perceived risk of hosting strangers, which is higher for families with children than for
unmarried and childless adults. We confirm that these predictions hold in the 50 largest US
cities in terms of hotel rooms. The entry of flexible supply is higher in cities like New York,
where demand is growing and highly variable, where hotels are constrained from expanding
room capacity, and where peer hosts have lower marginal costs than in cities like Atlanta.
2
In the short-run, peer producers decide whether to host on a particular day. Because
of the flexible nature of their supply, we hypothesize that these producers will be highly
responsive to market conditions, hosting travelers when prices are high, and using accommo-
dation for private use when prices are low. In contrast, because hotels have a fixed number
of rooms dedicated to travelers’ accommodation, they will typically choose to transact even
when demand is relatively low, while they won’t be able to expand capacity during peaks
in demand. These di↵erences imply that peer supply elasticity should be higher than ho-
tels’ supply elasticity on average. We validate this prediction by estimating a peer supply
elasticity that is twice as high as hotels’ elasticity.
The heterogeneous entry of peer hosts across cities and over time has surplus implications.
We estimate our short-run equilibrium model to quantify the e↵ect of Airbnb on total welfare
and its distribution across travelers, peer hosts, and hotels. Travelers benefit from Airbnb
for two reasons. First, flexible sellers o↵er a di↵erentiated product relative to hotels. Second,
they also compete with hotels by expanding the number of rooms available. This second e↵ect
is particularly important in periods of high demand when hotels are capacity constrained
and can thus charge higher prices. Consequently, we find that the increase in consumer
surplus from Airbnb is concentrated in city-days of peak demand, which the accommodation
industry defines as compression nights. In those cities and periods, flexible sellers allow more
travelers to stay in a city without greatly a↵ecting the number of travelers staying at hotels.
Our data mainly come from two sources: proprietary data from Airbnb, and data from
STR, which tracks supply and demand data for the hotel industry. We obtain data on
average prices and rooms sold at a city, day, and accommodation type level between 2011
and 2014 for the 50 largest US cities.1 We first document heterogeneity in the number of
Airbnb listings across cities and over time. Cities like New York and Los Angeles have grown
more quickly, reaching supply shares exceeding 15% and 5% respectively in 2014, while cities
like Oklahoma City and Memphis have grown more slowly, with less than 1% supply shares
at the of 2014. Within each city over time, the number of available rooms is higher during
peak travel times such as Christmas and the summer. The geographic and time heterogeneity
suggests that hosts flexibly choose when to list their rooms for rent on Airbnb, and are more
likely to do so in cities and times when the returns to hosting are highest.
In Section 2, we incorporate this intuition into a model of the market for accommoda-
tions. In this model, rooms for accommodations can be provided by dedicated or flexible
sellers, and products are di↵erentiated. We include two time-horizons. The long-run hori-
zon is characterized by the entry decision of flexible sellers given the new Airbnb platform.
We model the decision of flexible sellers to join the platform as dependent on the expected
1The 50 largest US cities were selected on the basis of their total number of hotel rooms.
3
returns from hosting, which in turn depend on competition from hotels and expected de-
mand levels. The short-run horizon focuses on daily prices and quantities of rooms sold.
We define the short-run horizon as one day in one city. In the short-run, the capacity of
flexible and dedicated sellers is fixed, and overall demand level is realized. Travelers choose
an accommodation option among di↵erentiated hotel and Airbnb rooms. Hotels maximize
profits subject to their capacity constraints, while peer hosts take prices as given.
The model o↵ers testable predictions. The long-run share of flexible sellers should di↵er
across cities. Entry should be largest in cities where hotel investment costs are high, flexible
sellers’ marginal costs are low, and demand variability is high so that there are periods of
high prices. In the short-run, flexible sellers should increase competition: they will reduce
prices and occupancy rates of hotels, and the e↵ects will be largest in cities where hotel
capacity is low relative to demand. We describe those cities as having constrained hotel
capacity. In capacity-constrained cities, the model predicts that Airbnb reduces prices more
than occupancy rates relative to non-capacity-constrained cities.
In Section 3, we confirm that these model predictions hold in the data. We first look
at the long-run patterns. We show that peer supply as a share of total supply is larger in
cities where hotel prices are higher. These high prices are associated with the di□culty of
building hotels due to regulatory or geographic constraints. Peer supply is also larger in
cities where residents tend to be single and have no children. These residents likely have
lower costs of hosting strangers in their homes. Another factor influencing peer supply is
the volatility of demand. A city can experience periods of high and low demand due to
seasonality, festivals, or sporting events. When the di↵erence in peaks and troughs is large,
the provision of accommodation exclusively by hotels can be ine□ciently low. We show that
Airbnb’s supply share is larger precisely in cities with high demand volatility, and, perhaps
more intuitively, in cities where demand growth is high.
We then test the predictions of the model on short-run hotel outcomes. We do this by
estimating regressions of hotel performance on Airbnb supply using two types of instruments
as well as controls for aggregate demand shocks. We find that the negative e↵ect of Airbnb
on hotel revenues is larger in cities with constrained hotel capacity, and that compared to
other cities, hotels here experience a bigger reduction in prices than occupancy rates. The
heterogeneity in estimates is due to di↵erences in both the size of Airbnb and the e↵ects of
Airbnb across markets conditional on that size.
In Section 4, we describe our estimation strategy for recovering the primitives of our
model. We proceed in three steps. First, we estimate a random coe□cient multinomial
logit demand model (Berry et al. (1995)). We augment our estimation with survey data
regarding the preferred second choices of Airbnb travelers, which helps us identify substitu-
4
tion between Airbnb and hotel options. Second, we estimate hotels’ cost functions assuming
Cournot competition between hotels of the same scale. In order to take into account the
fact that prices steeply increase when occupancy approaches 100%, we follow Ryan (2012)
and rationalize these price changes with marginal costs that start increasing when hotels
are close to their capacity constraint. Third, we estimate the cost distribution of peer hosts
assuming that they are price takers. Together, these estimates allow us to measure consumer
and peer producer surplus, as well as to quantify how surplus would change in the absence
of peer supply.
Section 5 presents our results for the top 10 US cities. We find that consumers’ utility for
Airbnb is lower than for hotels, but that preferences for Airbnb increase between 2013 and
2014. By the end of the sample period, the mean utility from top quality Airbnb listings is
close to the mean utility of economy and midscale hotels. We find that peer hosts have higher
marginal costs than hotels on average, and that consistent with our model, the distribution
of peer costs makes peer supply highly elastic.
In the absence of Airbnb, total welfare would be lower, travelers and peer producers
would be worse o↵, while hotels would benefit from less competition. In the top 10 US cities,
total welfare would decrease by $137 million and consumer surplus would decrease by $276
million if Airbnb did not exist in 2014.
This corresponds to a consumer surplus of $41 per night for every Airbnb booking. The
reduction in consumer surplus if Airbnb did not exist occurs because fewer travelers would
book rooms, and travelers who end up booking hotel rooms would pay higher prices. As
it turns out, because of the elastic peer supply, actual Airbnb bookings and thus surplus
gains disproportionately occur in cities (New York) and times (New Year’s Eve) when hotel
capacity constraints bind. This implies that in the absence of peer supply, travelers could
not easily find a substitute hotel room because hotels would be fully booked. Indeed, we
find that around half of Airbnb bookings would not have been hotel stays had Airbnb not
existed.
The concentration of Airbnb bookings in cities and periods of peak demand suggests that
in the absence of Airbnb, hotels would be limited in their ability to increase the number of
booked rooms – they were already operating at or close to full capacity – but instead would
be able to increase prices. Revenues for hotels would increase by 1.54% if Airbnb did not
exist, and profits would likely increase by a larger percent. In fact, our estimates suggest
that the variable profits of hotels could increase by up to 3.69% across the 10 cities in our
sample.
We contribute to the growing empirical literature on online peer-to-peer platforms. A
limited number of papers have looked at the e↵ect of online platforms on incumbents, in
5
particular Zervas et al. (2015) for Airbnb, Seamans and Zhu (2014) and Kroft and Pope
(2014) for Craigslist, and Aguiar and Waldfogel (2015) for Spotify. In this paper we not
only estimate the e↵ects on incumbent firms, but also on consumers and new producers. In
addition, we highlight important dimensions of heterogeneity of the e↵ects of Airbnb across
cities and over time. A complementary paper to ours is Cohen et al. (2016), which uses
discontinuities in Uber’s surge pricing policy to estimate the consumer surplus from ride
sharing. Both of our papers find that successful peer-to-peer platforms generate substantial
consumer surplus. However, while Cohen et al. (2016) assume that incumbents do not change
their behavior, we incorporate capacity constraints and allow for hotel prices to adjust in
the absence of Airbnb. This is important for our setting because even travelers who book
hotel rooms benefit from Airbnb through lower prices. Similarly to us, Lam and Liu (2017)
estimate a model of competition between Uber, Lyft, and taxis using data from New York.
Another related stream of research studies the role of peer-to-peer markets in enabling
rental markets for durable goods. Horton and Zeckhauser (2016) derive a theoretical equilib-
rium model for ownership and rental of durable goods, and make predictions on the existence
and size of rental markets across di↵erent product categories. Fraiberger and Sundararajan
(2015) calibrate a model of car usage and quantify the expected reduction in car ownership
as a result of peer-to-peer rental markets.
Our paper is also complementary to existing studies of labor supply and market design
on peer-to-peer platforms. We find that host supply is highly elastic on the margin. This is
consistent with analysis of suppliers on Taskrabbit (Cullen and Farronato (2017)) and Uber
(Hall et al. (2016), Chen and Sheldon (2015)). Other work on peer-to-peer markets has
focused on the market design aspects of reputation systems (Fradkin et al. (2017), Nosko
and Tadelis (2015), Bolton et al. (2012)), search (Fradkin (2017), Horton (2016)), and pricing
(Einav et al. (Forthcoming), Hall et al. (2016)). Specifically focused on the hotel industry,
Lewis and Zervas (2016) study the welfare e↵ects of online reviews. Finally, in our analysis of
growth heterogeneity across cities, we contribute to the predominantly theoretical literature
on technology adoption and di↵usion (e.g. Bass (1969) and Griliches (1957)).
The paper is structured as follows. In the next section, we present the data and document
geographic and time heterogeneity in the size of Airbnb, which motivates our theoretical
framework of competition between dedicated and flexible sellers (Section 2.1). In Section 3
we test the predictions of our model on the long- and short-run elasticities of flexible supply,
and on the competitive e↵ects of Airbnb on hotels. Section 4 presents our empirical strategy
for structurally estimating the short-run equilibrium of our model. We discuss the estimation
results in Section 5 and conclude in Section 6.
6
2 Motivation and Theoretical Framework
Airbnb describes itself as a trusted community marketplace for people to list, discover,
and book unique accommodations around the world — online or from a mobile phone.
The marketplace was founded in 2008 and has at least doubled in total transaction volume
during every subsequent year in our sample. Airbnb has created a market for a previously
rare transaction: the short-term rental of an apartment or room to strangers. In the past,
these transactions were not commonly handled by single individuals because there were large
costs to finding a match, securely exchanging money, and ensuring safety. While Airbnb is
not the only company serving this market, it is the dominant platform in most US cities.2
Therefore, we use Airbnb data to study the drivers and the e↵ects of facilitating peer entry
in the accommodation market.
Airbnb room supply has grown quickly in the aggregate, but the growth has been highly
heterogeneous across geographies. Figure 1 plots the size of Airbnb measured as the daily
share of available Airbnb listings out of all rooms available for short-term accommodation.3
Even among the top 10 cities in terms of listings, there are high growth markets like San
Francisco and New York, as well as slow growth markets like Chicago and DC. This increase
in available rooms is specific to the peer-to-peer sector and does not represent a broader
growth of the supply of short-term accommodation (see Figure A1).
Within a city over time, there is also heterogeneity in the size of Airbnb relative to the
size of the hotel sector. The fluctuations are especially prominent in New York in Figure
1, which experiences large spikes in available rooms during New Year’s Eve, and in Austin
during the South by Southwest festival. The figure suggests that market conditions during
these spikes are especially suited to peer-to-peer transactions. These facts motivate our
theoretical model, in which we distinguish between dedicated sellers (hotels) and flexible
sellers (peer hosts).
2.1 Theoretical Framework
In this section, we introduce a theoretical model for understanding market structure with
dedicated supply (hotels) and flexible supply (peer hosts) in the accommodation industry.
We will test the predictions of this model in Section 3, and structurally estimate the short-run
2The most prominent competitor is Homeaway/VRBO, a subsidiary of Expedia. Its business has his-
torically been concentrated in rentals of entire homes in vacation destinations, such as beach and skiing
resorts.
3The total number of available rooms is the sum of available hotel rooms and listings available on Airbnb.
The same heterogeneity is apparent if we adjust for capacity, or if we divide the number of Airbnb listings
by the number of total housing units within an MSA.
7
component in Section 4.
In our model, hosting services can be provided by dedicated and flexible sellers, who
o↵er di↵erentiated products. The model has a short and long-run component. The short-
run equilibrium consists of daily prices and rooms sold of each accommodation type as a
function of the overall demand level and the respective capacities of dedicated and flexible
suppliers. We assume hotels are competing against a fringe of flexible sellers. The long-
run component determines the entry condition of flexible sellers as a function of fixed hotel
capacity and the distribution of demand states.
We start with the short-run equilibrium representing daily market outcomes. We simplify
the exposition by assuming that there is one single hotel and one undi↵erentiated type
of Airbnb listings. In the empirical counterpart of this model in Section 4 we relax this
assumption. Let K
h
denote the mass of existing dedicated capacity (number of hotel rooms),
and K
a
the existing flexible capacity (Airbnb rooms). Demand state, d, is drawn from a
distribution F , which can be interpreted as the distribution of demand states over the course
of a year. Hotel rooms and Airbnb rooms are di↵erentiated products. Qd
i
(p
i
,p
j
) is the residual
demand for product i as a function of its price and the price of the other product. Qd
i
(p
i
,p
j
)
is assumed to be increasing in d and p
j
, and decreasing in its own price p
i
.
The short-run sequence of events is as follows. Capacity K
h
and K
a
are given, demand
state d is realized, the hotel sets prices and at the same time Airbnb sellers choose whether
to host at the prevailing prices. We assume that the hotel faces marginal cost c
h
to book one
room for one night, and it sets its price to maximize profits subject to its capacity constraint:
Max
ph
Qd
h
(p
h
,p
a
)(p
h
□c
h
)
s.t.
Qd
h
(p
h
,p
a
)  K
h
(1)
Flexible sellers have unit capacity and variable marginal costs of renting their room. We
assume that marginal costs of peers are randomly drawn from a known distribution.
When choosing whether to rent out their room for a night, flexible producers take prices
as given, and sell their unit if and only if the market clearing price is greater than their cost.
The choices of individual hosts are aggregated to determine the total number of flexible
rooms rented:
Qd
a
(p
a
,p
h
) = K
a
Pr(c  p
a
),
(2)
where K
a
is the mass of peer hosts, and Pr(c  p
a
) is the share of hosts with costs lower
than p
a
.
The market equilibrium consists of prices and quantities for hotel rooms and peer rooms
(p
h
,p
a
,q
h
,q
a
) that equate flexible and dedicated room demand with flexible and dedicated
8
supply.
The short-run model already o↵ers some comparative statics predictions, listed below
and proven in Appendix A. Under standard conditions, hotel profits per available room,
as well as both prices and occupancy rates, are lower if K
a
is higher. The separate e↵ect
of an increase in K
a
on hotel prices is higher if hotel capacity constraints are more often
binding, but the opposite is true for the e↵ect on occupancy. Intuitively, this occurs because
the increase in flexible capacity a↵ects hotels through a reduction in their residual demand
(Figure 2), and when hotels are capacity constrained, their supply curve is vertical (Figure
2b). A marginal downward shift in residual demand will have no e↵ect on quantity and a
large e↵ect on price if supply is perfectly inelastic.
In the long-run, entry of flexible suppliers is endogenous. We assume that K
h
was opti-
mally set knowing the distribution of demand states and not expecting that Airbnb would
lower entry costs of flexible sellers. Holding demand fixed, if investing in hotel capacity is
more costly, optimal dedicated capacity is lower and expected profits per unit of capacity
are higher.
A peer-to-peer platform enables the entry of flexible sellers. Flexible sellers decide
whether to join the peer-to-peer platform and start producing as a function of expected
demand and expected marginal costs. We assume that flexible sellers face a one-time cost
C of joining the platform, which is randomly drawn from a given distribution. We let
v
a
=
R
d
E
c
□
max{0,pd
a
□c}
□
dF (d) be the expected daily benefit of joining the platform.
The expression E
c
□
max{0,pd
a
□c}
□
denotes the expected profit of a flexible seller given
demand state d, where the expectation is taken over the distribution of marginal costs.
A flexible seller joins the peer-to-peer platform if the expected benefits are higher than the
entry cost. If expected daily profits v
a
are higher, more flexible sellers will join the platform
and start producing, and the share of flexible supply out of total supply will be higher. What
a↵ects v
a
? The first element is the distribution of marginal costs c. Holding everything else
constant, if the distribution of costs decreases in the sense of first order stochastic dominance,
more peers will enter and start hosting. The second element is pd
a
, itself a function of K
h
and the distribution of demand F (d). All else equal, a lower K
h
will increase equilibrium
prices whenever capacity constraints bind, so it will increase the distribution of pd
a
in the first
order stochastic dominance sense. Clearly, a higher level of demand in every state is more
attractive, but, perhaps less obviously, also an increase in demand variability is attractive for
flexible suppliers. To explain why, we can think of a simple mean-preserving spread of two
demand states. In the low demand state, flexible suppliers host very few travelers in either
case because hotels are not capacity constrained, leading to low equilibrium prices. The
di↵erence occurs in high demand states. If the high demand state doubles, prices increase
9
steeply, especially if hotel capacity constraints are hit, making it very attractive for flexible
suppliers to host in periods of high demand. Appendix A formally states the assumptions
and proofs for these results. Section 3.2 confirms that these comparative statics predictions
hold in the data.
Before presenting our data, we should note that our model does not allow hotels to adjust
dedicated capacity K
h
in response to peer entry. In the long-run, peer entry could partially
crowd out dedicated sellers. Since our data only spans the first few years of Airbnb di↵usion
and hotel construction projects take between 3 and 5 years to complete, we are unable to
empirically capture hotels’ capacity adjustments. Exploring the entry and exit decisions of
dedicated producers would be a valuable extension of our work.
3 Data and Tests of the Model
In this section, we describe our data on Airbnb and hotels and document how it confirms the
predictions of the theoretical framework. Our proprietary Airbnb data consists of information
aggregated at the level of listing types. The variables we observe include the number of
bookings, active and available listings, as well as average listed and transacted prices. An
available listing is defined as one that is either booked through Airbnb or is open to being
booked on the date of stay according to a host’s calendar. An active room is defined as a
listing that is available to be booked or is already booked for at least one date in the future.
Average transacted prices are calculated among all booked rooms on a given date, regardless
of the time of booking. Analogously, average listed prices are calculated among all available
rooms on the day of stay, regardless of prior price changes.
We categorize Airbnb listings into four types: ‘Airbnb Luxury’, ‘Airbnb Upscale’, ‘Airbnb
Midscale’, and ‘Airbnb Economy’.4 Listing types are defined using the following algorithm.
We first run a city level hedonic regression of nightly price on listing fixed e↵ects, date
fixed e↵ects, and bins for the number of five-star reviews.5 Second, we extract the listing
fixed e↵ects and use Bayesian shrinkage to shrink fixed e↵ects towards the mean. Third, we
compute quartiles of listing quality and categorize a listing in a given quartile if its fixed e↵ect
plus review coe□cient falls into the appropriate range. This procedure allows us to account
for heterogeneity in Airbnb listing types without specifically modeling detailed geographic
and room type characteristics at a city level.
The hotel data come from STR, an accommodation industry data provider that tracks
4These categories are defined solely for the purpose of this paper and do not correspond to any metric
used by Airbnb itself.
5The bins for the number of reviews are: 0, 1, 2-3, 4-5, 6-10, 11-25, 26-50, 51-100, □101 .
10
over 161,000 hotels. Our sample contains daily prices and occupancy rates for the 50 largest
US cities for the period between January 2011 and December 2014.6 STR obtains its infor-
mation by running a periodic survey of hotels, to which they ask daily revenue attributable
to the sale of hotel rooms, total rooms sold, and total rooms available. For the 50 largest
markets, 68% of properties are surveyed, covering 81% of available rooms. STR uses sup-
plementary data on similar hotels to impute outcomes for the remaining hotels which are
in their census but do not participate in the survey. The data is then aggregated to six
hotel scales, from luxury to economy, which indicate the quality and amenities of the hotels.
So the data tell us, for example, the average price and the total number of rooms sold on
January 10th, 2013, of midscale hotels in San Francisco.
Table 1 shows city-level descriptive statistics regarding hotels and Airbnb. In the average
city, hotels charge $108 per room and their occupancy rate is 66%. Perhaps surprisingly,
Airbnb has very similar transacted prices ($109) and much lower occupancy rates (15%). The
within-city standard deviation of these outcomes varies greatly across cities. For example,
the city at the 25th percentile has a standard deviation of hotel prices of $10 ($22 for Airbnb
prices), while the city at the 75th percentile has a standard deviation of $21 ($34 for Airbnb
prices). This indicates that markets di↵er not only in levels but in the extent to which
conditions fluctuate within a year and over time.
During our sample period, Airbnb comprises a small share of the overall market as a
percentage of total rooms available for short-term accommodation. The average Airbnb
share of available rooms in the last quarter of 2014 is 2%, and in most cities it is between
1% and 3% (25th and 75th percentiles). Two other normalizations confirm that Airbnb
was still small in most US cities by the end of our sample period. Across all cities, Airbnb
rooms represent 4% of all guests and represent less than 1% of total housing units for all
metropolitan statistical areas (MSAs) in our sample.
3.1 The Long-Run: Determinants of Peer Entry
In this section, we verify the theoretical predictions regarding the long-run growth of peer
supply from Section 2.1. Although the theoretical model assumes that entry decisions are
made instantaneously and jointly for all flexible sellers, in practice awareness about the
Airbnb platform has grown between 2011 and 2014. We assume that the last quarter in
2014, the end of our sample, provides a valid proxy for the long-run share of peer supply
derived in our model.
6The cities are ranked based on the absolute number of hotel rooms in 2014. See Census Database: http:
//www.str.com/products/census-database and STR Trend Reports: http://www.str.com/products/
trend-reports
11
Figure 3 shows the relationship between Airbnb share of room capacity and hotel daily
revenues per available room. Not surprisingly, the size of Airbnb is positively correlated with
the average revenue per available room in a city, with New York being both the city with
the highest hotel revenues and the one with the highest penetration of peer hosts.
Our theory predicts that if hotels have high investment costs or peer hosts have low
marginal costs, profitability for peer hosts will be high. So we should expect more peer
entry in cities with high hotels’ fixed costs and low peers’ marginal costs. We use two
proxies for hotel fixed costs. The first is the share of undevelopable area constructed by Saiz
(2010). The index measures the share of a metropolitan area that is undevelopable due to
geographic constraints, e.g. bodies of water or steep mountains. The second index is the
Wharton Residential Land Use Regulatory Index (WRLURI), which measures regulation
related to land use in each metropolitan area and is based on a nationwide survey described
in Gyourko et al. (2008).7 Figures 4a and A2 confirm that constraints to hotel capacity are
correlated with the share of peer supply.8
The second cost factor influencing the viability of peer production is the marginal cost
of peers. Although many factors a↵ect the costs of hosting, we focus on those related to
demographics.9 Households vary in their propensities to host strangers in their homes. For
example, an unmarried 30-year-old professional will likely be more open to hosting strangers
than a family with children. This occurs for at least two reasons. First, children increase
a host’s perceived risk of the transaction. Second, unmarried professionals are more likely
to travel, creating vacant space to be rented on Airbnb. Figure 4b plots the share of peer
supply at the end of 2014 against the percentage of unmarried adults, while Figure A2 uses
the percentage of children. The figures confirm that cities with more unmarried adults and
fewer children are those where Airbnb represents a higher share of accommodation rooms.
In addition to cost factors, our model predicts that travelers’ demand a↵ects peer entry.
This is due to two related reasons. First, hotels typically do not have enough dedicated
capacity to absorb all potential travelers in times of peak demand. In contrast, flexible sellers
are able to provide additional supply during peak times, when their rooms are especially
valuable to travelers. Second, since hotels must pre-commit to capacity and any adjustment
in the form of new hotel buildings takes 3 to 5 years, unforeseen growth in demand will
7Saiz (2010) uses these two measures to calculate the housing supply elasticity at the level of a metropoli-
tan area.
8Building restrictions also a↵ect Airbnb supply through another channel, the cost of residential housing.
There are greater incentives to monetize a spare bedroom when the costs of housing are higher, especially
for liquidity constrained households. Figure A2 in the Appendix confirms a positive relationship between
the share of household income used to pay rent in 2010 and the size of Airbnb in 2014.
9Other potential shifters of the returns to hosting include household liquidity constraints, building regu-
lation and enforcement of short-term rentals, and the ease of vacating an apartment in high demand periods.
12
create an ine□ciently low dedicated supply and will induce entry by flexible sellers.
We use data from air travelers to proxy for accommodation demand trends and fluctua-
tions at the city-month level. We measure these demand characteristics in 2011, the earliest
year in our sample, in order to reduce the risk that peer entry influences demand rather
than vice versa. Our data come from Sabre Travel Solutions, the largest Global Distribution
Systems provider for air bookings in the US. We isolate trips entering a city as part of a
round trip from a di↵erent city in order to measure the potential demand for short-term
stays.10 Figure 5a confirms the intuition that year-on-year demand growth results in greater
peer entry by showing that the 2012-2011 growth rate in travelers to a city is positively
related to Airbnb penetration in 2014. Figure 5b plots the standard deviation of demand in
2011 and confirms that by the end of 2014 Airbnb is bigger in cities where the fluctuations
in the number of arriving travelers are larger.
To conclude this section, we combine all the descriptive results into a regression. Table
2 displays the summary statistics for the cost and demand factors described above. Table 3
displays results from a regression where the dependent variable is the size of Airbnb in the last
quarter of 2014 and the explanatory variables are combinations of the measures of relative
costs, demand growth, and demand variability described above. We also control for market
size in order to isolate the component of the standard deviation of demand which is due to
demand variability. Despite the small sample size, column (1) shows that all factors a↵ect
the size of Airbnb in the expected direction, and two coe□cients are statistically significant
- peers’ marginal costs, and demand volatility. Column (2) adds additional and potentially
redundant proxies for costs and demand. The coe□cients are in the expected direction for
all proxies.
In the last column of Table 3 we add the 2011 average revenue per hotel room as an
additional control. The coe□cient on revenue per available room is positive and statistically
significant, while the coe□cients on the demand and hotel investment cost proxies decrease
in magnitude and become insignificant. This result suggests that demand proxies and hotel
investment costs a↵ect peer entry mostly through price and occupancy as expected. Taken
altogether, our proxies for the determinants of long-run peer supply explain between 60%
and 75% of the variation across our cross-section of US cities.
3.2 The Short-Run: E↵ects of Peer Entry on Hotels
In the previous section we have tested the long-run predictions of our theoretical model,
those related to the entry of peer producers. Here, we take entry as given, and focus on the
10Data from Sabre include monthly number of passengers by origin and destination airport. We aggregate
these observations to an MSA-month measure of air travelers.
13
short-run drivers of peer supply to test our model hypotheses on the e↵ects of peer supply
on hotels. The awareness and di↵usion process of Airbnb and its variation across cities help
us test these hypotheses.
First, we show how to properly measure the size of Airbnb, and how the short-run
elasticity of Airbnb supply is twice as large as that of hotels. Then, we use an instrumental
variable approach to study the reduction in hotel revenues after the entry of Airbnb, and its
heterogeneity across cities and hotel scales.
Measuring Airbnb Supply
We start by demonstrating how to properly measure Airbnb supply and studying how hosts
flexibly respond to fluctuations in market-level demand over time. Figure 6 displays four
measures of the size of Airbnb plotted over time: active listings, two measures of available
listings, and booked listings. This figure displays three important facts. First, the share
of active or available listings that are booked varies greatly over time. The booking rate is
especially high during periods of high demand such as New Year’s Eve and the summer. What
we will show just below is that this is the result of a highly elastic peer supply. Second, the
gap between active listings and available listings is increasing over time, suggesting attrition
in active listings. Therefore, the meaning of an active listing does not stay constant over the
entire period of study.
The third and most relevant fact from Figure 6 is that the number of unadjusted available
listings (blue line) actually decreases during periods of high demand, most notably on New
Year’s Eve. The main reason for this is that calendar updating behavior responds to room
demand. Many hosts do not pro-actively take the e↵ort to block a date on their calendar
when they are unavailable (see Fradkin (2017) for evidence). However, when they receive a
request to book a room, they often reject the guest and update their calendar accordingly.
Since a larger share of listings receives inquiries during high demand periods, the calendar is
also more accurate during those times. Therefore, the naively calculated availability measure
su↵ers from endogeneity and is even counter-cyclical – high when demand is low, and low
otherwise.
Since we need a measure of the size of Airbnb that stays stable over time, we create an
adjusted measure of available listings. This measure includes any rooms which were listed
as available for a given date or were sent an inquiry for a given date and later became
unavailable. Therefore, it does not su↵er from the problem of demand-induced calendar
updating. It does overstate the “true” number of available rooms in the market, but as
long as it overestimates true availability consistently over time we consider it to be the
best measure of Airbnb size. Figure 6 displays our proposed measure (red line) against the
14
naive measure of available listings (blue line). The new measure does not su↵er from drops
in availability during high demand periods. Throughout the rest of the paper we use the
adjusted number of available listings as the size of Airbnb supply unless otherwise noted.
Peers’ Responses to Demand Fluctuations
From Figure 6 it is clear that Airbnb bookings fluctuate over time: more rooms are booked
during the peak season than in other periods. In this section, we use 2SLS to document that
flexible suppliers are almost twice as elastic as dedicated suppliers.
We estimate the average supply elasticity of hotel and Airbnb rooms with respect to their
prices using the following equation:
log(Q
mt
) = □log(K
mt
) + log(p
mt
) + µ
mt
+ ✏
mt
,
(3)
where Q
mt
is the number of (hotel or Airbnb) bookings in city m and day t, K denotes
capacity, and p is the average transacted price. The equation is estimated separately for
hotels and Airbnb.  is the elasticity of supply with respect to prices, and will be di↵erent
between flexible and dedicated supply. µ
mt
includes city, seasonality (month-year), and day
of week fixed e↵ects to control for the fact that costs might change by city or over time (e.g.,
due to average di↵erences in costs over cities or due to particular periods where hosts are
less likely to occupy their residences).
Equation 3 su↵ers from standard simultaneity bias because the price of accommodations
is correlated with demand, and with unobserved fluctuations in marginal costs. Furthermore,
in the case of Airbnb, the number of available rooms K
mt
is itself endogenous because hosts
may list their room as available precisely during high demand periods.11
We discuss each concern in order. We instrument for price with plausibly exogenous
demand fluctuations which are typically caused by holidays or special events in a city. We
use two instruments. The first is the number of arriving (not returning) flight travelers
in a city-month, which we used in Section 3.1. The second comes from Google Trends,
which provides a normalized measure of weekly search volume for a given query on Google.
Our query of interest is “hotel(s) c”, where c is the name of a US city in our sample. We
de-trend each city’s Google Trends series using a common linear trend to remove long-run
changes in overall search behavior on Google. We use the one-week lagged search volume
as an instrument. Using other lags or the contemporaneous search volume yields similar
estimates.
11The same endogeneity issue is not important for hotels because hotel capacity is typically fixed in a
4-year interval, our sample period. However, instrumenting for hotel capacity with a quadratic time trend,
as we do for Airbnb, does not change our results.
15
To control for the fact that room availability on Airbnb is endogenous to demand, we
instrument for the number of available listings with a city-specific quadratic time trend. It is
reasonable to believe that this instrument captures the long-run di↵usion process of Airbnb
while being uncorrelated with contemporaneous idiosyncratic shocks to supply. We use this
same instrumentation strategy below to test for the e↵ect of Airbnb supply on hotel revenues
predicted by our model.
Table 4 contains our estimates of Equation 3 for Airbnb and hotels separately. Turning
first to column 1, a 1% increase in the average hotel daily rate increases hotel bookings by
1.1%. This elasticity is half as large as that of Airbnb (column 2), whose estimated elasticity
is 2.2. An important implication of this result is that smaller fluctuations in prices are needed
for Airbnb supply to adjust upward or downward.
We have shown that the Airbnb supply is highly responsive to price, more so than hotels:
a small price increase due to high demand greatly increases the number of booked rooms on
Airbnb, and this increase is twice as large as for hotels. The lower elasticity of hotel supply
has a simple explanation. To the extent that hotels have a constant marginal cost and a fixed
supply, hotel bookings cannot increase in response to increases in demand when demand is
su□ciently high. The higher elasticity of flexible supply implies that there are many hosts
willing to rent their rooms when prices are high, but prefer not to host when prices are just
a little lower. Our structural model in Section 4 rationalizes this result by estimating that
there is a large mass of peers with costs close to the market clearing prices.
E↵ects of Peer Entry on Hotel Revenue
In this section, we test our model hypotheses on the e↵ects of peer entry on hotels’ revenue,
occupancy rates, and prices using linear specifications. Before describing our empirical strat-
egy, we discuss the two most important challenges to identifying the e↵ect of Airbnb. To do
this, we consider the hypothetical scenario where Airbnb supply grows randomly across cities
and over time. In this scenario, regressing the outcomes of hotels on the Airbnb supply would
yield an unbiased estimate of the causal e↵ect of Airbnb supply. However, as highlighted
above, Airbnb supply does not grow randomly. In fact, Airbnb supply is larger in cities with
high hotel revenues, and during periods of high demand within each city. Observables like
the number of arriving flight travelers, city fixed e↵ects, and seasonality fixed e↵ects, help
us control for this selection.
We instrument for the currently available Airbnb supply with a city-specific quadratic
time trend, which isolates the predictable di↵usion of Airbnb over-time from the endogenous
short-run Airbnb supply responses to demand shocks during peak days such as holidays.
The key assumption behind this identification strategy is that there are no other city-specific
16
trends conditional on observables.12 We remove this assumption in our structural estimation
in Section 5, which relies on plausibly exogenous components of price variation and market
structure to identify the e↵ects of interest.
Our baseline regression specification is:
y
mt
=↵ log(airbnb
mt
) + □log(gtrend
mt
) + □log(travelers
mt
) + ✓
mt
+ ⌫
mt
.
(4)
Here y
mt
is one of three hotel outcomes (log revenue per available room, log price, occupancy
rate) in a city m on day t, airbnb
mt
is the number of available Airbnb listings, gtrend
mt
is the
one-week lag of Google searches for hotels in the city, travelers
mt
is the number of arriving
air passengers, and ✓
mt
includes city, quarter-year, and day of week fixed e↵ects. Impor-
tantly, the Google metric captures demand shocks at the week level, while the number of
incoming air passengers captures monthly fluctuations in demand. The fixed e↵ects capture
seasonality, di↵erences across the days of the week, and time-invariant city characteristics
that a↵ect both the size of Airbnb and hotel revenue.
The e↵ect of interest is ↵, which is the average short-run elasticity of hotel outcomes to
peers’ supply over our sample period. The coe□cient is identified o↵ of two types of variation.
First, there is variation across cities and over time in the number of available listings due
to increasing awareness of Airbnb. Second, there is variation in the availability of listings
due to hosts’ daily costs of hosting, which we assume are uncorrelated with residual daily
demand for accommodation within the city.
Table 5 displays the results of the baseline specification. The coe□cient on Airbnb size in
column (1) is statistically significant and the estimated elasticity for hotel revenue is -.033.
This coe□cient implies that a 10% increase in available listings decreases the revenue per
hotel room by 0.33%. The coe□cient estimates for our demand proxies, Google trends and
arriving air travelers, are of the correct sign and statistically significant. Once we break
down the e↵ect into a reduction in occupancy rates (column 2) and a reduction in prices
(column 3), we see that on average Airbnb has a larger e↵ect on prices than on occupancy
rates.
Recall that our model predicts that, holding fixed Airbnb supply, in days and cities
when hotels are not capacity-constrained, Airbnb should have a relatively bigger e↵ect on
occupancy than on price. The opposite is true when hotels are capacity-constrained: on
those days, Airbnb should have a relatively bigger e↵ect on price than on occupancy. We
also predicted and confirmed empirically that there will be more Airbnb rooms available in
cities where hotels are often capacity constrained. Consequently, our theory predicts that
12In Appendix B we conduct robustness checks to demonstrate that these controls and instruments likely
capture potential sources of endogeneity.
17
we will find larger e↵ects of Airbnb in constrained cities and that the price rather than
occupancy channel will be more important in these cities than in non-constrained cities.
To test this prediction, we divide our cities into two groups and explore the heterogeneity
of the e↵ect of Airbnb across cities. Saiz (2010) uses the WRLURI and the share of unde-
velopable area described in Section 3.1 to estimate the housing supply elasticity at the city
level. We take that supply elasticity as a proxy for the elasticity of hotel construction, and
split our sample of cities at the median level of Saiz’s estimates for the cities in our sample.
Table 6 displays the estimates of Equation 4 separately for the two groups of cities. Columns
(1) and (4) display the estimates of the e↵ect on revenue per available hotel room. Both
coe□cients on Airbnb are statistically insignificant. When we break the outcomes into prices
and occupancy rates, we see that the statistically significant e↵ect of Airbnb is a reduction in
hotel prices in the cities where hotels are capacity constrained (column 3). This is consistent
with the fact that binding capacity constraints lead to spikes in hotel prices, which in turn
attract more competition from Airbnb. When Airbnb enters, hotels are often fully booked,
so they face more pressure on prices than on occupancy. Table A4 separates the e↵ect in
constrained cities by hotel scale and detects e↵ects on all hotel scales other than Luxury.
Di↵erences in the e↵ect of Airbnb on hotels across constrained and unconstrained cities
occur for two reasons. First, for the same level of Airbnb and hotel capacity, the e↵ect of
Airbnb is relatively larger on prices if hotel capacity constraints are more often binding (due
to higher levels of demand). Second, for the same level of demand and hotel capacity, the
e↵ect on hotel revenues is larger if there are more Airbnb listings. Intuitively, the elasticity of
hotel revenues with respect to the size of Airbnb should be higher, the higher the Airbnb share
of supply because a 1 percent increase in Airbnb size is a much bigger share of market supply
when Airbnb penetration is 3% then when it is 1%. Both conditions are true when we split
our cities. Indeed, in December 2014 the average Airbnb supply share in hotel-constrained
cities was 4.3% while it was only in 1.4% in unconstrained cities. At the same time, the
average hotel occupancy rate was 61% in constrained cities and only 53% in unconstrained
cities.
Before concluding this section, one caveat is in order. In these specifications we cannot
take advantage of exogenous changes in price that would allow for a valid causal estimate of
the e↵ect of Airbnb on hotel performance. For that analysis we refer the reader to Section
5. However, this exercise helped us confirm that the main predictions of our model from
Section 2 hold in the data. We documented that the entry of peer hosts is responsive to
long-run supply and demand characteristics. Peer supply is more likely to enter in cities
where hotels’ fixed costs are high, where peers marginal costs are low, and where demand
is increasing and highly variable. We have also shown that flexible supply is highly elastic,
18
and twice as elastic as dedicated supply. Finally, we have shown that the entry of flexible
supply has negative spillovers on the revenue of dedicated suppliers. This negative e↵ect is
higher in cities with binding hotel capacity constraints and has a relatively larger impact on
prices in those cities. In the rest of the paper, we structurally estimate our short-run model
in order to measure the welfare e↵ects of Airbnb on consumers, peer hosts, and hotels.
4 Model and Estimation Strategy
In this section, we describe the fully specified short-run model that we estimate. This extends
the theoretical model from Section 2 to multiple hotel and Airbnb listing types. A market
n is defined by day t and city m. On the demand side, our model is a random coe□cients
logit model (Petrin (2002) and Berry et al. (1995)), where rooms are di↵erentiated across
hotel scales and Airbnb listing types. On the supply side, we assume that hotels engage in
Cournot competition with di↵erentiated products across scales. Within a scale, each hotel
is undi↵erentiated. Airbnb hosts are price takers with randomly drawn marginal costs.
Consumer Demand
Consumers make a discrete choice between hotel scales, Airbnb listing types, and an outside
option for a given night. Consumer i has the following utility for room option j in market
n:
u
ijn
= µ
ijn
□↵
i
(1 + ⌧
jn
)p
jn
+ ✏
ijn
.
(5)
For consumer i, µ
ijn
represents a market-specific mean utility for accommodation j. The
price of an accommodation is denoted p
jn
, while the lodging tax rate is ⌧
jn
. Finally, ✏
ijn
is an
idiosyncratic component with a type I extreme value distribution. We normalize the value
of the outside option to 0 for all markets. This demand specification yields the following
quantities for each accommodation type:
Q
jn
(p
jn
,p□jn) = Dn
Z
eµijn□↵i(1+⌧jn)pjn
1 +
P
j
0 eµij0n□↵i(1+⌧jn)pj0n
dH(µ
ijn
,↵
i
),
(6)
where D
n
is the market size, and H is the joint distribution of consumer heterogeneity in µ
ijn
and ↵
i
. We allow for consumer heterogeneity in how travelers value the inside options (hotels
and Airbnb), the high-end hotel scales (luxury and upper upscale), and accommodation
prices. We assume that the distribution of consumer heterogeneity is multivariate normal
with a mean and variance matrix to be estimated. We do not allow for correlation across
distinct components of consumer heterogeneity.
19
Hotel Supply
Each hotel competes with other hotels of the same scale, hotels of di↵erent scales, and peer
supply. We assume that this competition takes the form of a Cournot equilibrium. Hotels
of type h, where h 2 {luxury, upper-upscale, upscale, upper-midscale, midscale, economy},
have aggregate room capacity K
hn
. Since there are multiple hotels within each scale, we
need to distinguish between scale-level and hotel-level quantities. We let Q
hn
denote the
scale-level number of rooms sold. We assume no di↵erentiation in room quality within scale,
so the number of rooms sold by each hotel, denoted q
hn
, is the ratio of aggregate quantity
divided by the number of hotels. Analogously, scale-level capacity is denoted K
hn
, while
hotel-level capacity is k
hn
.
We must also match the fact that prices increase sharply as the number of rooms sold
approaches the number of available rooms. In practice, occupancy rates never reach 100% at
the scale level, but prices start increasing before then (Figure 7). This is because, although
we model hotels as homogeneous within each scale, some individual hotels may sell out before
others and this may result in sharply increasing scale-level prices. In addition, if hotels face
uncertainty about the actual level of demand when setting prices, increases in expected
demand will increase the probability of hitting capacity constraints, thus increasing prices
before realized demand reaches 100%. We allow our model to fit this increasing price profile
by estimating an increasing cost function for hotels that kicks in as soon as hotel occupancy
is at least 85% within a scale. The estimation of increasing marginal costs as production
approaches capacity constraints was previously used by Ryan (2012) to estimate the cost
structure of the cement industry.
We assume that hotels’ variable costs are made of two parts: a constant marginal cost c
hn
,
and an increasing marginal cost □
hn
(q
hn
□⌫k
hn
), which starts binding as quantity approaches
the capacity constraint. Given the above discussion, we set ⌫ = 0.85. So, instead of solving
a maximization problem subject to a capacity constraint as in Equation 1, each hotel selects
its quantity to maximize the following profit function:
Max
qhn
q
hn
p
hn
(Q
hn
,Q□hn,Qan)□qhnchn □
□
hn
2
1(q
hn
> ⌫k
hn
)(q
hn
□⌫k
hn
)2.
We assume that hotels observe all components of demand and competitors’ costs, so that
there is no uncertainty about whether q
hn
> ⌫k
hn
or not. Letting N
hn
denote the number
of hotels within scale h, we have that q
hn
= Qhn
Nhn
. Taking advantage of the implicit function
20
theorem, the optimization problem gives rise to the following first order condition:13
p
hn
=□1
N
hn
Q
hn
Q0
hn
+ c
hn
+ □
hn
1(q
hn
> ⌫k
hn
)(q
hn
□⌫k
hn
),
(7)
where Q
hn
is scale-level room demand from Equation 6, and Q0
hn
is the derivative with
respect to its own price.
Peer Supply
Peers of each quality type a, where a 2 {Airbnb luxury, Airbnb upscale, Airbnb midscale,
Airbnb economy}, with total available listings K
an
, take prices as given. Hosts draw marginal
costs from a normal distribution with parameters !
an
and □
an
. Each draw is iid across hosts
and time. Hosts of type a choose to host only if the price p
an
is greater than their cost.
Therefore, the quantity supplied will be determined by the following equation:
Q
an
(p
an
,p□an,phn) = KanPr(c  pan) = Kan□
✓
p
an
□!
an
□
an
◆
.
(8)
Equilibrium
The market equilibrium consists of prices and quantities for hotels and peer hosts (p
hn
,p
an
,
Q
hn
,Q
an
) such that consumers, hotels, and peer hosts make decisions to maximize their
surplus, and their optimal choices are consistent with one another.
4.1 Estimation Strategy
We estimate the demand, hotel supply, and peer supply separately. For demand, the high-
level choices are the market size, the moments to match, and the instruments used.
Starting first with demand, we need to make a normalization. Since Airbnb listings are
on average bigger than hotel rooms and can host more guests, we adjust quantities so that
room capacity is comparable across Airbnb listings and hotel rooms. To do this, we take
advantage of the fact that we have information on the average number of guests for Airbnb
transactions.
In addition, lower quality Airbnb listings are typically private rooms with
similar capacity as standard hotel rooms. For this reason, we assume that each hotel room
is occupied by as many people as the average number of occupants of Airbnb listings in the
midscale quality category in the same city. Given this adjustment, our quantities, prices, and
estimates should be interpreted as referring to room-nights with standard hotel occupancy.
13The objective function is not di↵erentiable at qhn = ⌫khn, but otherwise the first order condition holds
everywhere else.
21
We use data on the 10 largest cities in terms of the share of Airbnb bookings in our
sample. Our initial estimation sample includes all days in 2013 and 2014. We restrict the
sample to 10 cities and 2 years of data for two practical reasons. First, in other cities and time
periods market shares of Airbnb are often close to zero, which complicates our estimation.
Second, there will not be much of an e↵ect of Airbnb when market shares are close to 0. For
the same reason, we also drop Airbnb options if their share of available rooms is less than
0.5% on a given day and city.
One key choice we must make in the estimation is D
n
, the total number of consumers
considering to book accommodations. The choice of D
n
will a↵ect market shares for hotels
and Airbnb, as well as the share of potential travelers choosing to stay home, to travel to
other locations, or to stay in alternative accommodations, e.g. friends and family. We set D
n
equal to three times the average number of rooms booked in the corresponding month in each
city in 2012. This assumption allows the potential number of travelers to vary seasonally
across cities, and it allows for both substitution from hotels – hotel travelers switching to
Airbnb –, and market expansion – travelers switching from the outside option to Airbnb.
We rationalize any remaining variation over time in the total number of travelers booking
accommodations with mean utilities for inside options that vary as a function of observable
characteristics such as Google search trends, and unobservables.
The second key choice is the set of moments that we match to the data. We construct
two types of moments for the demand estimation: moments to match predicted and realized
market shares (market share moments) and one moment to match predicted hotel-Airbnb
substitution with substitution obtained from survey responses that Airbnb conducted (sub-
stitution moment).
Our market share moments are
m1jn =
h
□
jn
□□
jn
i
Z
jn
,
(9)
where □
jn
is the realized mean utility from accommodation j in market n that rationalizes the
observed market shares, and□
jn
is the mean utility predicted from the vector of parameters
to be estimated.□
jn
is the component of utility from Equation 5 that does not di↵er across
individual travelers, and is a function of observable and unobservable shifters of demand for
di↵erent types of accommodations. The observable shifters include city-scale fixed e↵ects,
city-month fixed e↵ects (to account for market specific seasonality), city-specific and Airbnb-
city-specific time trends, the log of Google searches and its square, and the log of Airline
passengers and its square.
The substitution moment comes from survey data on alternative accommodation choices
22
of travelers booking on Airbnb. Airbnb has conducted surveys of guests in four of the
sampled cities during 2013 and 2014. The surveys asked the question: “If Airbnb had
not been available, what would you have done?”. Between 58% and 81% of guests across
cities said that they would have stayed in a hotel. A simple average across cities yields a
substitution share of 68% towards hotels.14
We match the survey moment in our model by computing the share of Airbnb travelers
who would have wanted to book a hotel at the observed prices had Airbnb not been available.
To predict the share of Airbnb travelers choosing hotels in the absence of Airbnb, we first note
that aggregate hotels’ market share in market n is s
hotels,n
=
P
j2hotel
R
e
µijn□↵i(1+⌧jn)pjn
1+
P
j0 e
µij0n□↵i(1+⌧jn)pj0n
dH(µ
ijt
,↵
i
). Airbnb market share, denoted s
airbnb,n
, is similarly computed. If Airbnb listings
were not available, hotels’ market share would be s
hotels,n
⇤ =
P
j2hotel
R
e
µijn□↵i(1+⌧jn)pjn
1+
P
j02hotel e
µij0n□↵i(1+⌧jn)pj0n
dH(µ
ijt
,↵
i
). Aggregating over all markets gives us the following moment:
m2n =
X
n

D
n
✓
s
h,n
⇤ □s
h,n
s
a,n
□s
survey
◆□
.
(10)
The final key choice is the set of instruments for the market share moments in Equation
9. We first generate instruments to predict after-tax prices. We use a series of cost-shifters
that a↵ect prices and are unlikely to be correlated with demand shocks. These cost-shifters
include hotel and Airbnb tax rates,15 and specifically for Airbnb, the number of residents
traveling out of a city. We also construct instruments using the total number of available
rooms. Since for Airbnb room availability responds to demand, the number of available rooms
is predicted with city-specific quadratic time trends. We interact the predicted number of
available Airbnb rooms with an indicator variable for hotel options, and we do the same for
hotel rooms and an indicator variable for Airbnb options. This set of instruments proxy for
the level of competition within a market. A hotel scale’s own number of available rooms can
also be used to construct an instrument because it a↵ects whether hotel capacity constraints
are likely to bind. To use it as an instrument, we interact its inverse, and its inverse squared,
with the Google search trend, and its square. We use the ratio because capacity constraints
a↵ect prices more when demand is higher. Since we have a number of instruments that are
14In 2015, Morgan Stanley and AlphaWise conducted a representative survey of 4,116 adults in the US,
UK, France, and Germany. In the survey, they asked respondents about their travel patterns. 12% of
respondents had used Airbnb within the past year and when asked which travel alternative Airbnb replaced,
42% of respondents answered a hotel. See Nowak et al. (2015). We think that the major reason for the
di↵erences between the Airbnb and Morgan Stanley surveys is that Morgan Stanley sampled guests to
all types of destinations including resorts and European cities. There are typically more non-Airbnb and
non-hotel options for guests in these locations.
15Airbnb started collecting occupancy taxes in Portland, OR on July 1, 2014, and began collecting taxes
in San Francisco on October 1, 2014.
23
potentially weak and correlated with each other, we take the principal components of all
the instruments, and keep the components that account for 95% of the variation (Carrasco
(2012)).
Once we have predicted prices from the set of instruments presented above, we follow
Gandhi and Houde (2016) to construct additional instruments that measure the distance in
characteristic space between di↵erent accommodation options. The relevant characteristics
in our model include the ordering of scales from luxury to economy, and prices. We construct
the following instruments: the di↵erence and square of the di↵erence between the predicted
price of an option and the predicted price of its closest alternatives – for midscale hotels,
the closest alternatives are upper midscale and economy –, the total number of alternatives
whose predicted price is within a standard deviation of an option predicted price, and the
sum and sum of squares of the di↵erence between the predicted price of a hotel option and
the predicted prices of all other hotel options. Since these instruments are highly correlated,
we keep the principal components accounting for 75% of the variation for the estimation.
Appendix C describes these instruments in greater detail. With this list of instruments,
we use the moments in equations 9 and 10 to estimate the set of demand parameters by
generalized method of moments.
It is useful to give an intuition for how the variation in the data allows us to estimate
our demand parameters. Our descriptive statistics show that the prices of hotels and Airbnb
options, unadjusted for di↵erent number of occupants, are similar. This fact, together with
the relatively high substitution rate between hotels and Airbnb rooms derived from survey
responses, suggests that the mean utilities of hotels and Airbnb options should be fairly
similar. However, in practice we also observe very di↵erent market shares, much higher
for hotels than for Airbnb options. The market share and substitution moments helps us
rationalize these two patterns in the data. On one hand, the substitution moment helps
us identify the random coe□cient on the inside option, which is common across hotels and
Airbnb listings. On the other, di↵erences in market shares rationalize mean utilities that
will be higher for hotels than for Airbnb options. Finally, di↵erences in market shares when
the price of competitors change help us identify consumer preference heterogeneity across
options and prices.
Once we obtain demand estimates that let us compute Q
hn
and its derivative, we estimate
the supply function from equation 7 using a linear IV approach:
p
hn
+
1
N
hn
Q
hn
Q0
hn
= ✓X
hn
+ □
hn
1(q
hn
> ⌫k
hn
)(q
hn
□⌫k
hn
) + ✏
hn
.
X
hn
includes city-scale fixed e↵ects and city-specific linear time trends. We allow □
hn
to vary
24
by city and scale.
We instrument for the increasing cost component using interactions of the Google search
trend with city and hotel fixed e↵ects. We use these instruments because they a↵ect hotel
prices only by increasing the likelihood that capacity constraints bind. We estimate the
supply equation separately for each day of the week to allow for the possibility that sta□ng
requirements and hotels’ response to capacity constraints vary across days of the week.
Finally, the supply of Airbnb can be estimated separately using another set of linear IV
regressions for the same sample period. Equation 8 implies that □□1
⇣
Qan
Kan
⌘
= !an
□an
+ 1
□an
p
an
,
where the left-hand side is the inverse of a standard normal cumulative distribution function
calculated at a value equal to the share of booked rooms out of all Airbnb listings. We
estimate this equation separately for each listing type using the specification
□□1
✓
Q
an
K
an
◆
= □
a
p
an
+ □
a
X
an
+ ✏
an
,
where K
an
is the number of active Airbnb listings, p
an
is the average transacted price of
Airbnb type a in market n, and X
an
include year-month fixed e↵ects, city fixed e↵ects, and
city-specific linear time trends. Here we choose K
an
to be the number of active listings
because the decision of being available is a choice of the host that depends on his own
marginal cost draw. We instrument for the transacted price with log of (de-trended) Google
search trends and the log of incoming air passengers.
After estimating the above equation, we can transform the coe□cients into the following
peer cost parameters:
□
an
=
1
□
a
, !
an
=
□
a
X
an
+ ✏
an
□
a
.
5 Results
In this section, we discuss the results of our estimation. We first go over our estimated
parameters. Then we discuss the e↵ects of Airbnb on consumer surplus. Lastly, we study
the e↵ects on hotels’ and hosts’ bookings, revenues, and surplus. For each of these outcomes,
we discuss heterogeneity across cities and time periods.
5.1 Parameter Estimates
Table 7 displays the estimates of demand parameters that are common across cities and
accommodation options. We first discuss the parameters governing the distribution of price
sensitivity across travelers. The mean price coe□cient is -.033 and the standard deviation is
25
.006. The standard deviation is imprecisely estimated, but our estimates are consistent with
existing work on hotel demand (Koulayev (2014)). Google search trends and incoming airline
travelers are estimated to have a positive e↵ect on demand for the range of values observed
in the data. We also estimate large and statistically significant heterogeneity in preferences
for high-end hotel options – luxury and upper upscale – and a more limited but equally
significant level of heterogeneity in preferences between traveling or not. This last estimate
of the random coe□cient on the inside option is especially important because it governs the
extent to which Airbnb travelers would substitute towards hotels or the outside option if
Airbnb were not available. The model achieves a value for the objective function of 170
with a relatively parsimonious specification. The total number of observations in the data
is 65,172 while we estimate 375 linear parameters and 3 non-linear parameters determining
the variances of the random coe□cients.
Next, we consider the mean utilities across accommodation options. Figure 8 displays the
mean willingness to pay per night for each option and city at the end of 2014. The fact that
some values are negative reflects our choice of a market size that’s three times the average
number of booked rooms in a city-month. When looking at the mean utilities in relative
terms, our estimates show that willingness to pay tends to be decreasing between upscale and
economy hotels and between Airbnb luxury and economy listings. The extremely low mean
utilities for luxury and upper upscale hotels are rationalized by the fact that preferences for
these options are highly heterogeneous, as highlighted in Table 7.
The value of the top Airbnb option is lower than the value of the lowest hotel option across
all cities, with some variation in the relative di↵erences. We cannot distinguish between
alternative explanations for this di↵erence, but one likely possibility is that not all travelers
actually consider Airbnb as a viable option for their travel plans, yet. Another possibility
is that Airbnb listings are typically located farther away from the city tourist and business
centers, which might lower travelers’ willingness to pay relative to hotels.
Table A5 shows the city-specific elasticities of demand for di↵erent accommodations with
respect to their own price and Table A6 shows the average cross-price elasticities. We find
that demand for accommodations is elastic on average. For example, in New York, the
demand elasticities range between -8.54 for luxury hotels and -2.93 for the lowest quality
of Airbnb listings. There is also substantial variation across cities in demand elasticities,
ranging between -2.65 in Portland and -6.24 in New York for midscale hotels.
Next, we turn to the estimates of hotel cost parameters. Our parameter estimates are
precise and the estimation procedure explains most of the variation with an R-squared of
0.83. The interquartile range for the errors is -$9.9 to $12.2.16 Figure 9 plots the marginal
cost curves for di↵erent hotel scales and di↵erent cities at the end of 2014. We find that
26
the constant components of hotels’ marginal costs have the expected relationship with hotel
quality. The marginal cost for luxury hotels in New York city are $324 on average. We
want to be careful in not interpreting these costs as actual expenditures per night-booked.
Research by Kalnins (2006) suggests that due to reputational concerns, hotels tend to have
a price threshold below which they will not go, and this threshold is typically higher than
the cost of an additional maid- or clerk-hour. We view our estimates as reflecting this price
threshold. The figure also plots the increasing component of hotels’ marginal costs. We find
that for all but one of the city and hotel combinations, marginal costs increase relatively
steeply with quantity when hotel occupancy reaches 85%. This increasing cost reflects the
fact that regardless of the level of competition, hotels will increase their prices as they
approach full capacity. Tables A7 and A8 report the full set of city-scale cost estimates.
Finally, Figure 10 displays the mean costs over time for listings in New York City. Costs
vary over the course of the year, and there is a slightly increasing trend for all room options.
In New York but also in other cities, costs increase monotonically in listing quality, and the
mean costs exceed the mean transacted prices. These relatively high costs stem from the
fact that fewer than 50% of active listings on Airbnb typically transact (Table 1). With
R-squared values ranging from .34 to .46, the variation in our data can explain a little
less of Airbnb costs than hotels’ costs. However, we estimate economically and statistically
significant dispersion in the cost distribution for all listing types, which explains the high
supply elasticity of Airbnb accommodations. Interestingly, costs go up during New Year’s
Eve and other high seasons. This may reflect the fact that travelers during those times are
costlier to host, maybe because of their higher likelihood to be disruptive. Alternatively, it
may mean that during those periods hosts prefer to stay in their apartments with friends
and family rather than strangers. Table A9 displays the full set of estimates of Airbnb costs
by listing type and city. The next section uses the estimates presented above to compare
realized welfare to a world without Airbnb.
5.2 Counterfactual Analysis
We present two counterfactual scenarios without Airbnb, and describe the e↵ect that the
absence of Airbnb would have for consumers, hotels, and peer hosts. The first counterfactual
scenario ignores hotel capacity constraints and how hotel prices would change if Airbnb
were not available. In this scenario, travelers who booked on Airbnb are allowed to book
any hotel option at the prevailing prices, regardless of actual room availability. The second
counterfactual scenario allows hotels to adjust prices in response to the absence of competing
16We also conducted a 10-folds cross-validation and found that the mean absolute error is similar between
estimation sample and hold-out sample, suggesting that we’re not over-fitting.
27
accommodations on Airbnb. This counterfactual requires computing new Cournot equilibria
for each market with demand and hotel cost parameters taken from our estimates.17 We
refer to the first counterfactual as the ‘unconstrained’ scenario, and to the second as the
scenario with ‘price adjustment’.
We discuss consumer surplus first. Table 8 presents how consumer surplus would change
in 2014 without Airbnb. The table shows that there would be a $143 million loss in consumer
surplus if we could ignore hotels’ capacity constraints and price adjustments. This loss
corresponds to $21 per room-night, about 15% of the purchase price. The consumer surplus
loss in this scenario only measures one channel through which Airbnb benefits consumers,
i.e. product di↵erentiation.
Relative to the ‘unconstrained’ counterfactual, there are two additional mechanisms
through which the ‘price adjustment’ counterfactual hurts consumers. First, travelers who
booked on Airbnb now face higher hotel prices. Second, travelers who previously booked
hotel accommodations also face higher prices. The consumer surplus loss in this scenario
almost doubles, rising to $276 million. This loss corresponds to $41 per room-night.
We can also look at how consumer surplus losses vary across cities and time periods.
The total surplus from Airbnb is primarily determined by the number of Airbnb bookings in
that city. This means that the greatest aggregate losses from the removal of Airbnb are in
New York, Los Angeles, and San Francisco. On a per room-night basis, the surplus change
is greatest for Seattle and lowest for Miami.
We also split our sample into ‘compression nights’, defined by the lodging industry to
be nights when hotels reach 95% occupancy, and non-compression nights. The surplus per
night booked is bigger on compression nights ($57) than on non-compression nights ($36).
This e↵ect is not due to travelers liking Airbnb more on these nights – the surplus per
night in the unconstrained counterfactual is nearly identical between these periods. Instead,
the e↵ect comes from the supply side. When comparing the surplus per night between
the counterfactual with and without price adjustments, there is a $35 di↵erence during
compression nights versus $14 for non-compression nights. This occurs because hotels can
charge high prices during compression nights and the elastic host supply keep prices down.
We now turn to the e↵ects of Airbnb on hotels. If hotels did not have capacity constraints,
Table 10 shows that hotels would increase the number of sold rooms by 2%. However, when
17Our estimation procedure doesn’t guarantee that, in the absence of Airbnb, the equilibrium quantities
remain below hotel capacity. We add an additional, although rarely binding, constraint in the price ad-
justment equilibrium which ensures that hotel quantities do not exceed an appropriately defined measure
of hotel capacity for each hotel type and period. We define the threshold to be the maximum between
95% occupancy, the 99th percentile if occupancy for the city-scale and the observed occupancy level for a
particular hotel scale in a given city-day.
28
we account for capacity constraints and the increase in hotel prices, sold rooms would only
increase by 1.3%, for a revenue increase of 1.5%. This corresponds to a gain of $412 million
in revenue for hotels.
The e↵ects of Airbnb on hotel revenues are heterogeneous across cities and over time.
For example, in New York, bookings would increase by 2% and revenue would increase by
2.24% without Airbnb. In contrast, in San Jose, a city where hotels are less likely to hit
their capacity constraints, bookings would increase by .63% and revenues would increase
by .72%. There is also heterogeneity in e↵ects across high and low demand days. Hotel
bookings increase by just .99% on compression nights but revenues increase by 1.7%. On the
rest of the nights, bookings and revenues would both increase by about 1.4%. The di↵erence
across nights demonstrates that the e↵ect of Airbnb on hotels is more concentrated on prices
than quantities during high-demand days relative to low-demand days. If we take our cost
estimates seriously, we can also look at the e↵ect of Airbnb on hotel profits, which we
calculate as hotel revenue minus the non-increasing part of the cost function. We find that
profits would increase by 3.69% on average across all of the cities in the sample, with the
largest increase happening in New York.
We should note that our estimates give us a lower bound for hotel variable profit loss – or
an upper bound on the percent change. This is because the profit change does not directly
correspond to hotel surplus for at least three reasons. First, hotels earn additional revenues
through complimentary services such as conferences and food sales, but also incur additional
costs. Second, there are fixed costs involved in operating a hotel which we do not model in
this exercise. If competition from Airbnb is strong enough, then some hotels may close down
or new hotels may not be built. Third, our marginal cost estimates correspond in part to
reputation costs rather than ‘true’ marginal costs. These additional costs and revenues do
not allow us to state with certainty whether hotel surplus is larger or smaller than our profit
estimate.18
Before moving to the surplus of peer hosts, we consider the extent to which Airbnb
expands the market versus cannibalizes hotel demand. Table 11 displays results on the share
of Airbnb travelers who would have booked a hotel room in the absence of Airbnb. In the
‘unconstrained’ scenario, between 27% and 34% of Airbnb bookings would not have resulted
in a hotel booking, which is broadly consistent with the survey moment used to estimate
demand. The market expansion e↵ect becomes much bigger when we account for capacity
constraints and hotels’ price responses. The share of Airbnb travelers who would have not
18In Appendix Table A10 we display the results assuming an alternative measure of costs for hotels imputed
from the wage bill of hotels in the STR data and trends in the wages of maids across cities and over time.
This is likely a lower bound on the true marginal cost of hotels.
29
in fact booked a hotel room ranges between 42% in Portland to 63% in New York.
Not surprisingly, peer hosts would lose without Airbnb. We use the estimated cost
distributions of hosts to back out the surplus that they receive from hosting on Airbnb. The
surplus for each day can be calculated using the following expression PS
an
=
R
pan
□Inf (pan □
max(c, 0))dF
an
(c) , where we censor the cost distribution at 0. Note that this expression
ignores the variable costs of being listed for a given day, which are likely to be negligible,
and the fixed costs of entry into the platform.
Table 12 displays average surplus per room-night and total surplus in 2014. The typical
surplus per night ranges between $23 in Miami and $30 in Austin. Across all bookings in
these cities, the average surplus is around $26 and does not vary much across high and low
demand days. In the aggregate, peer hosts enjoy $19 million in producer surplus, with hosts
in New York City, Los Angeles, and San Francisco getting the majority of it. Almost half of
this surplus is concentrated on compression nights, which comprise only 20% of all nights.
6 Conclusion
We have studied the economics of peer production in the lodging industry. We first doc-
umented the determinants of peer supply and showed how market-specific factors such as
supply constraints and the costs of hosting a↵ect whether peer production is viable in a
given city. We then documented that peer supply is twice as elastic as hotel supply in the
short-run.
The highly elastic host supply implies that the largest e↵ects of Airbnb occur in markets
where hotels are often near full capacity. We presented a simple model of competition
between peer supply and hotels, and tested its implications using data from 50 major US
cities. We confirm that the entry of Airbnb negatively a↵ects hotel revenues in cities where
hotels are more likely to be capacity-constrained, and that the e↵ect is more concentrated
on price than on quantity, at least compared to non-capacity-constrained cities.
Next, we estimated our short-run equilibrium model to study the surplus and market
expansion e↵ects of Airbnb. The availability of peer hosts generates $41 of surplus per
room-night in 2014. This surplus comes both from new bookings generated on Airbnb and
from lower prices paid by hotel travelers. In total, Airbnb generates $276 million in consumer
surplus in 2014 for the 10 largest US cities.
We showed that Airbnb has also a↵ected producers. Without Airbnb, hotel revenues
would be 1.5% higher. Nonetheless, between 42% and 63% of nights booked on Airbnb
would not have resulted in a hotel booking in the absence of Airbnb. These travelers would
have instead chosen the outside option, which could represent staying with friends or family,
30
staying at a non-hotel accommodation, booking fewer nights, or not traveling to the city at
all. Lastly, peer hosts also benefit from the introduction of Airbnb. Their average surplus
per night from hosting is $26, which totals $19 million in the estimation sample.
Our data only extend through the end of 2014. Since then, Airbnb has continued its
rapid growth in both active listings and global awareness. While we cannot say how large
its e↵ects have been since then, our paper documents two fundamental reasons why peer
production is valuable in the accommodation industry. First, peers o↵er a di↵erentiated
product that is not a perfect substitute to hotel rooms and is valued by consumers. Second,
the hotel sector in many cities is frequently constrained by a limited number of available
rooms, which lead to high prices during demand peaks because hotels cannot accommodate
all potential travelers. Peer production expands available supply at exactly these times of
peak demand, thus reducing hotel pricing power and increasing consumer surplus.
Our paper also informs the active policy debate regarding whether and how to regulate
peer-to-peer accommodations. Proposed policies include fees and taxes, mandated registra-
tions, quotas, caps on the number of nights per hosts, and bans.19 While we cannot conduct
a full welfare analysis of these policies due to the other markets a↵ected, our analysis sug-
gests that Airbnb is especially beneficial to consumer and host welfare during peak demand
periods in hotel constrained cities.
We have focused on the short-run e↵ects of a peer-to-peer platform on the agents directly
involved – hotels, peer hosts, and travelers – without highlighting the platform’s own costs
and revenues. In the longer run, the number of hotel rooms may also adjust to peer entry.
Peer production can have externalities and spillovers into other markets, including the labor
and housing markets (Filippas and Horton (2017), Barron et al. (2017)). We leave the study
of these e↵ects for future work.
References
Aguiar, Luis, and Joel Waldfogel. 2015. “Streaming Reaches Flood Stage: Does Spotify
Stimulate or Depress Music Sales?” National Bureau of Economic Research Working
Paper 21653.
Barron, Kyle, Edward Kung, and Davide Proserpio. 2017. “The Sharing Economy
and Housing A↵ordability: Evidence from Airbnb.”
Bass, Frank. 1969. “A New Product Growth for Model Consumer Durables.” Management
Science, 15(5): 215–227.
19See https://www.airbnb.com/help/article/1376/responsible-hosting-in-the-united-states.
31
Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. “Automobile Prices in Market
Equilibrium.” Econometrica, 63(4): 841–890.
Bolton, Gary, Ben Greiner, and Axel Ockenfels. 2012. “Engineering Trust: Reci-
procity in the Production of Reputation Information.” Management Science, 59(2): 265–
285.
Bulow, Jeremy I., John D. Geanakoplos, and Paul D. Klemperer. 1985. “Multi-
market Oligopoly: Strategic Substitutes and Complements.” Journal of Political Economy,
93(3): 488–511.
Carrasco, Marine. 2012. “A Regularization Approach to the Many Instruments Problem.”
Journal of Econometrics, 170(2): 383–398.
Chen, K. M., and Michael Sheldon. 2015. “Dynamic Pricing in a Labor Market: Surge
Pricing and Flexible Work on the Uber Platform.” Mimeo, UCLA.
Cohen, Peter, Robert Hahn, Jonathan Hall, Steven Levitt, and Robert Metcalfe.
2016. “Using Big Data to Estimate Consumer Surplus: The Case of Uber.” National
Bureau of Economic Research Working Paper 22627.
Cullen, Zoë, and Chiara Farronato. 2017. “Outsourcing Tasks Online: Matching Supply
and Demand on Peer-to-Peer Internet Platforms.”
Einav, Liran, Chiara Farronato, Jonathan Levin, and Neel Sundaresan. Forthcom-
ing. “Auctions versus Posted Prices in Online Markets.” Journal of Political Economy.
Filippas, Apostolos, and John J. Horton. 2017. “The Tragedy of Your Upstairs Neigh-
bors: Is the Airbnb Negative Externality Internalized?” ArXiv e-prints, arXiv:1611.05688.
Fradkin, Andrey. 2017. “Search, Matching, and the Role of Digital Marketplace Design in
Enabling Trade: Evidence from Airbnb.”
Fradkin, Andrey, Elena Grewal, David Holtz, and Matthew Pearson. 2017. “The
Determinants of Online Review Informativeness: Evidence from Field Experiments on
Airbnb.”
Fraiberger, Samuel P., and Arun Sundararajan. 2015. “Peer-to-Peer Rental Markets
in the Sharing Economy.” Social Science Research Network SSRN Scholarly Paper ID
2574337, Rochester, NY.
32
Gandhi, Amit, and Jean-François Houde. 2016. “Measuring Substitution Patterns
in Di↵erentiated Products Industries.” University of Wisconsin-Madison and Wharton
School.
Griliches, Zvi. 1957. “Hybrid Corn: An Exploration in the Economics of Technological
Change.” Econometrica, 25(4): 501–522.
Gyourko, J., A. Saiz, and A. Summers. 2008. “A New Measure of the Local Regulatory
Environment for Housing Markets: The Wharton Residential Land Use Regulatory Index.”
Urban Studies, 45(3): 693–729.
Hall, Jonathan, Cory Kendrick, and Chris Nosko. 2016. “The E↵ects of Uber’s Surge
Pricing: A Case Study.”
Horton, John J. 2016. “Misdirected Search E↵ort in a Matching Market: Causes, Conse-
quences and a Partial Solution.” Working Paper.
Horton, John J., and Richard J. Zeckhauser. 2016. “Owning, Using and Renting: Some
Simple Economics of the” Sharing Economy”.” National Bureau of Economic Research.
Kalnins, Arturs. 2006. “Markets: The U.S. Lodging Industry.” The Journal of Economic
Perspectives, 20(4).
Koulayev, Sergei. 2014. “Search for Di↵erentiated Products: Identification and Estima-
tion.” The RAND Journal of Economics, 45(3): 553–575.
Kroft, Kory, and Devin G. Pope. 2014. “Does Online Search Crowd Out Traditional
Search and Improve Matching E□ciency? Evidence from Craigslist.” Journal of Labor
Economics, 32(2): 259–303.
Lam, Tom, and Meng Liu. 2017. “Demand and Consumer Surplus in the On-Demand
Economy: The Case of Ride Sharing.”
Lewis, Gregory, and Georgios Zervas. 2016. “The Welfare Impact of Consumer Reviews:
A Case Study of the Hotel Industry.”
Nosko, Chris, and Steven Tadelis. 2015. “The Limits of Reputation in Platform Markets:
An Empirical Analysis and Field Experiment.”
Nowak, Brian, Thomas Allen, Jamie Rollo, Vaughan Lewis, Lin He, Ananda
Chen, Wilson W. Ng, Michael Costantini, Owen Hyde, Kevin Liu, Mark
33
Savino, Baset A. Chadhry, Anne M. Grube, and Ed Young. 2015. “Global Insight:
Who Will Airbnb Hurt More - Hotels or OTAs?” Morgan Stanley.
Petrin, Amil. 2002. “Quantifying the benefits of new products: The case of the minivan.”
Journal of political Economy, 110(4): 705–729.
Ryan, Stephen P. 2012. “The Costs of Environmental Regulation in a Concentrated In-
dustry.” Econometrica, 80(3): 1019–1061.
Saiz, Albert. 2010. “The Geographic Determinants of Housing Supply.” The Quarterly
Journal of Economics, 125(3): 1253–1296.
Seamans, Robert, and Feng Zhu. 2014. “Responses to Entry in Multi-Sided Markets:
The Impact of Craigslist on Local Newspapers.” Management Science, 60(2): 476–493.
Zervas, Georgios, Davide Proserpio, and John Byers. 2015. “The Rise of the Shar-
ing Economy: Estimating the Impact of Airbnb on the Hotel Industry.” Social Science
Research Network SSRN Scholarly Paper ID 2366898, Rochester, NY.
34
Figures
Figure 1: Growth of Airbnb
Chicago
DC
San Diego
Seattle
Boston
Austin
Miami
LA
SF
NY
0.05
0.10
0.15
2012
2013
2014
2015
Date
Airbnb Share of RoomsThe figure plots the size of Airbnb over time in 10 selected cities. The y-axis is the monthly
average of the daily share of Airbnb listings out of all (hotel and Airbnb) rooms available
for short-term accommodation. The 10 selected cities are those with the largest number of
listings on Airbnb as of December 2014 among the 50 US major cities.
35
Figure 2: Predictions on the E↵ect of Peer Supply on Hotels
(a) Unconstrained Equilibrium
(b) Constrained Equilibrium
The figures plot the supply and demand curve for hotel rooms in two scenarios. The hotel
supply curve is drawn holding constant the price of peer rooms p
a
, varying the demand state
d, and letting the hotel set the price to maximize its profits as in equation 1. The left panel
displays an unconstrained equilibrium, while the right panel displays an equilibrium where
the hotel capacity constraint is binding. Peer entry represents a downward shift in demand
for hotel rooms. This downward shift will a↵ect hotel quantity relatively more when the hotel
supply curve is more elastic. The opposite is true for the e↵ect on hotel prices, which is
higher in the capacity-constrained equilibrium.
36
Figure 3: Peer Production and Hotel Revenues
This figure plots the supply share of Airbnb against the average revenue per available room
in each respective city.
37
Figure 4: Peer Production and Fixed or Marginal Costs
(a) Hotel Supply Constraints
(b) Peers’ Marginal Costs
The figures plot the size of Airbnb against a proxy for hotel investment costs (left panel)
and a proxy for peers’ marginal costs (right panel). The proxy for the constraints to the
construction of new hotels is the share of undevelopable area developed by Saiz (2010). This
index measures the share of a city that is undevelopable due to geographic constraints, like
steep mountains or the ocean. The proxy for peers’ marginal costs is the share of unmarried
adults in the MSA. The size of Airbnb is measured as the average share of available listings
in the last quarter of 2014. Figure A2 in the Appendix confirms that other proxies such
as regulatory constraints, the share of children, and the rent to income ratio are also good
predictors of peer entry.
38
Figure 5: Peer Production and Demand Characteristics
(a) Demand Growth
(b) Demand Variability
The figures plot the size of Airbnb against the growth rate in incoming air passengers to
an MSA between June 2011 and June 2012 (left) and against the standard deviation of
incoming air passengers (right). The standard deviation of air travelers is measured using
2011 monthly data on arriving (not returning) passengers at major US airports. We focus
on data from 2011-2012, when Airbnb was very small relative to the accommodation market,
to limit the possibility that the availability of Airbnb hosts could generate such growth or
variability in demand. The size of Airbnb is measured as the average share of available
listings in the last quarter of 2014.
39
Figure 6: Measures of Airbnb Supply: Demand-induced Calendar Updates
Booked Listings
Available Listings
(Unadjusted)
Available Listings
Active Listings
25
50
75
100
2012
2013
2014
2015
Date
Size of Airbnb (Linear Scale)This figure plots four measures of the size of Airbnb. An active listing is defined as a listing
available to be booked or booked for any future date. An (unadjusted) available listing is one
that is either booked or has an open calendar slot on the date of stay. Available listings
augment the unadjusted measure with listings that were contacted for a particular date of
stay and were later updated to be unavailable for that date. A booked listing is one that has
been booked for that date.
40
Figure 7: Prices and Occupancy Rates
This figure plots prices and occupancy rates of upscale hotels in New York in 2014.
Figure 8: Estimated Utilities for Accommodation Options Across Cities
This figure plots the estimated mean utilities for accommodation options across the 10 cities
used in our estimation. The values are computed as averages over the last month in our
data. The negative values of some of the parameters reflect the fact that our normalization
of the outside option means that most people choose the outside option.
41
Figure 9: Estimated Hotel Costs
(a) Costs by City – Midscale Hotels
(b) Costs by Hotel Scale – New York City
These figures plot the estimated marginal cost curves of hotels across cities (left panel) and
across scales (right panel). The values are computed as averages over the last month in our
data. Appendix Tables A7 and A8 display the cost estimates by city and hotel scale.
Figure 10: Mean Costs of Airbnb Hosts in New York City
The figures plot the estimated mean costs of Airbnb hosts in New York over time. Appendix
Table A9 displays all the estimated means and standard deviations.
42
Tables
Table 1: Descriptive Statistics on Hotel and Airbnb Outcomes
Statistic
N
Mean
St. Dev.
Pctl(25)
Median
Pctl(75)
Mean Hotel Occupancy
50
0.66
0.07
0.61
0.65
0.70
Std Dev Hotel Occupancy
50
0.13
0.03
0.12
0.14
0.15
Mean Hotel Price in $
50
107.94
35.04
86.29
98.69
121.01
Std Dev Hotel Price
50
16.34
9.34
9.77
13.00
21.02
Mean Hotel Revenue (Thousand $)
50
3,785.34
3,588.86
1,575.12
2,488.65
4,856.56
Airbnb Share of Available Rooms (Q4 2014)
50
0.02
0.03
0.005
0.01
0.03
Airbnb Share of Potential Guests (Q4 2014)
50
0.04
0.04
0.01
0.02
0.05
Airbnb Share of Housing Units (Q4 2014)
50
0.001
0.001
0.0002
0.0005
0.001
Mean Airbnb Occupancy
50
0.15
0.06
0.11
0.13
0.18
Std Dev Airbnb Occupancy
50
0.09
0.02
0.07
0.08
0.10
Mean Airbnb Price in $
50
108.63
24.67
92.01
99.76
120.87
Std Dev Airbnb Price
50
31.24
13.17
21.78
28.50
34.29
Mean Airbnb to Hotel Price Ratio
50
1.06
0.28
0.91
1.00
1.14
Std Dev Price Ratio
50
0.33
0.18
0.21
0.29
0.39
This table shows hotel and Airbnb descriptive statistics for the 50 cities in our sample. For each city, we compute the mean
and standard deviation of daily occupancy rate and price for hotels and Airbnb listings. The Airbnb share of available rooms
is computed as the average of daily share of rooms in the last quarter, i.e. October - December 2014. The Airbnb share of
potential guests is computed as the quarterly average of rooms adjusted for their realized capacity, assuming that the typical
hotel has the same number of average guests as a ‘Midscale’ Airbnb listing. This number is larger than the Airbnb share of
rooms because Airbnb listings typically have higher capacity than hotel rooms.
Table 2: Descriptive Statistics on Market Characteristics
Statistic
N
Mean
St. Dev.
Pctl(25)
Median
Pctl(75)
WRLURI
50
0.31
0.82
□0.30
0.20
0.84
Share of Undevelopable Area
46
0.30
0.24
0.09
0.23
0.43
Percent Never Married
48
0.33
0.03
0.31
0.33
0.36
Share of Children
48
0.31
0.02
0.30
0.31
0.32
Rent to Income Ratio
50
0.18
0.03
0.15
0.17
0.20
Std Dev of Google Trend (2011)
50
12.05
4.22
9.62
11.51
13.70
Std Dev of Incoming Passengers (2011) / 10,000
50
6.95
6.63
1.89
4.17
11.30
Passengers’ Growth (2012-2011)
50
0.02
0.06
□0.02
0.01
0.04
The table shows descriptive statistics on market characteristics for the 50 cities in our sample. The WRLURI and Saiz’s share
of undevelopable area are proxies for constraints to hotel supply. The share of children and unmarried adults proxy for the
availability of Airbnb hosts. The standard deviation of Google trends and incoming passengers are two measures of demand
volatility.
43