Journal of Urban Economics 119 (2020) 103278

Contents lists available at ScienceDirect

Journal of Urban Economics

journal homepage: www.elsevier.com/locate/jue

Do short-term rental platforms affect housing markets? Evidence from

Airbnb in Barcelona☆
Miquel-Àngel Garcia-López a, Jordi Jofre-Monseny b,∗, Rodrigo Martínez-Mazza b, Mariona Segú c
a
Universitat Autònoma de Barcelona and Institut d’Economia de Barcelona, Spain
b
Universitat de Barcelona and Institut d’Economia de Barcelona, Spain
c
RITM, Université Paris-Saclay and Institut National d’Études Démographiques (INED), France

a r t i c l e i n f o a b s t r a c t

JEL classification: In this paper, we assess the impact of Airbnb on housing rents and prices in the city of Barcelona. Examining very
R10 detailed data on rents and both transaction and posted prices, we use several econometric approaches that exploit
R20 the exact timing and geography of Airbnb activity in the city. These include i) panel fixed-effects models, where
R31
we run multiple specifications that allow for different forms of heterogeneous time trends across neighborhoods,
Z30
ii) an instrumental variables shift-share approach in which tourist amenities predict where Airbnb listings will
Keywords: locate and Google searches predict when listings appear, iii) event-study designs, and iv) finally, we present
Housing markets evidence from Sagrada Familia, a major tourist amenity that is not found in the city centre. Our main results
Short-term rentals imply that for the average neighborhood, Airbnb activity has increased rents by 1.9%, transaction prices by 4.6%
Airbnb and posted prices by 3.7%. The estimated impact in neighborhoods with high Airbnb activity is substantial. For
neighborhoods in the top decile of Airbnb activity distribution, rents are estimated to have increased by 7%,
while increases in transaction (posted) prices are estimated at 17% (14%).

1. Introduction tralizing tourism within cities. From an economic point of view,

home-sharing platforms can be seen as an efficiency improvement
Tourism has grown enormously in recent decades: between 1990 and in markets where goods are not fully used (Barron et al., 2020). For
2017, the worldwide number of international tourist arrivals increased example, empty apartments during holiday periods are efficiency
from about 400 million to 1300 million (WTO, 2018). This pattern is losses that can be reduced through short-term rentals. However, if
particularly apparent in urban tourism; the number of visitors to the home-sharing platforms are used by owners to permanently shift from
162 most popular cities in the world has increased on average 6.5% long-term to short-term rentals for tourists, the supply of units in the
each year between 2009 and 2018 (MasterCard, 2019). Home-sharing long-term market is reduced, increasing housing prices and rents.
peer-to-peer platforms such as Airbnb have recently entered the mar- Critics of home-sharing platforms emphasize that short-term rental
ket, partly accommodating the increased demand for tourism in cities. units in residential areas might constitute a negative externality for
As a consequence, they have contributed to increasing the overlap be- residents in terms of noise or uncivil behavior and cause displacement
tween tourism and housing markets by allowing owners of residential of long-term residents.1 Complaints about touristification effects and
properties to enter the hospitality sector. nuisances have resulted in local policy implementation that limits the
Proponents of home-sharing platforms argue that short-term rentals expansion of platforms such as Airbnb. Such policy responses include
provide residents with an additional source of income while decen- permit requirements (Barcelona, Berlin, Paris, San Francisco, and Los

☆
We are grateful to the Editor, two anonymous referees, Gabrielle Fack, Apostolos Filippas, Laurent Gobillon, Hans Koster, Miren Lafourcade, Florian Oswald,
Henry Overman, Elisabet Viladecans-Marsal, Xavi Ramos, Rosella Nicolini, Marta Marot, Joel Marot, Maria Janot and participants of UEA (Dusseldorf and New York)
and RUES 2019 (Paris) for their valuable comments. We gratefully acknowledge funding from Ajuntament de Barcelona (18S02330-001), Generalitat de Catalunya
(2017SGR-796 and 2017SGR-1301), Ministerio de Economía (ECO2016-75941R) y Ministerio de Ciencia, Innovación y Universidades (RTI2018-097401-B-I00 and
PID2019-108265RB-I00) and the Fondation Palladio. The Agència Tributària de Catalunya and Daniel del Pozo and David Rey (Idealista) were very helpful and
collaborative with their data.
∗
Corresponding author.
E-mail address: [email protected] (J. Jofre-Monseny).
1
Similarly, the hotel industry views home-sharing platforms as a threat to fair competition. Zervas et al. (2017) empirically studies the effect of Airbnb on hotel
revenues.

https://doi.org/10.1016/j.jue.2020.103278
Received 20 June 2019; Received in revised form 31 August 2020
Available online 9 September 2020
0094-1190/© 2020 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Table 1
Airbnb activity in 2015 in selected cities.

Barcelona New York Los Angeles Paris

Airbnb Listings 16,951 45,260 30,000 35,000

as % of total units 2.06 1.31 0.86 2.56
as % of rented units 6.84 1.92 1.56 4.97
Average Airbnb price/day (€) 71 131 114 81
Long-term rent/day (€) 11 59 75 37
Days/month for same revenue 10 14 20 14

Notes: Data on Barcelona comes from Cadastral Records and INCASOL, data on New York and Los Angeles comes
from US Census Bureau, Zillow Rent Index and airdna, and data for Paris comes from INSEE and OLAP. All Airbnb
data have been obtained through InsideAirbnb.

Angeles), limiting the rental period (Amsterdam, New York, Paris, and To study the effect of Airbnb listings on residential housing markets,
San Francisco), paying a rental tax (Amsterdam and San Francisco), or we combine publicly available web-scraped data on Airbnb listings in
outlawing short-term rentals under some conditions (Berlin and New Barcelona with high-quality data on housing rents and real estate prices.
York). Despite all these local policy responses, we still have limited We have access to i) individual-level data on the universe of transac-
evidence on the effect of home-sharing platforms on housing markets. tions of second-hand apartments sold in the 2009–2017 period and ii)
In this paper, we analyze the effects of the arrival and expansion of all posted ads for rentals and sales from a major real estate website (Ide-
Airbnb in Barcelona. Barcelona is ideal to study the effects of Airbnb alista) that were active each December in the 2007–2017 period. We
on local housing markets for several reasons. First, Barcelona has expe- aggregate the information at the geographical level of small neighbor-
rienced rapid tourism growth. The number of passengers in the city’s hoods, which leaves us with a panel dataset of 221 small geographical
airport increased from 20 to 47 million between 2000 and 2017. It is areas that have an average population of about 7000 inhabitants.
the 7th most visited destination city in Europe, measured by overnight Throughout the empirical analysis, our dependent variable is the
visitors, and the 17th worldwide (MasterCard, 2019). Second, Airbnb ac- average residual resulting from a hedonic regression in which the log
counts for the majority of short-term rental activity in the city, far ahead of housing rents or real estate prices are regressed on time dummies
of its competitors.2 Third, Airbnb penetration in Barcelona is high, with and unit characteristics. In all regressions, we control for neighborhood
Barcelona being Airbnb’s 6th top destination worldwide.3 and time fixed effects. Since Airbnb has grown the most in central parts
Table 1 compares the number of Airbnb listings in Barcelona, New of the city, our main identification concern is that neighborhoods that
York, Los Angeles and Paris in 2015. Despite substantial legal uncertain- experienced higher Airbnb penetration might be simultaneously expe-
ties regarding the use of home-sharing platforms in Barcelona, about riencing processes of urban revival.6 Aside from controlling for time-
2.06% of all housing units are listed on Airbnb.4 This figure is higher varying neighborhood socioeconomic characteristics that are associated
than in New York (1.31%) and Los Angeles (0.86%), and slightly smaller with gentrification processes, we adopt several strategies to account for
than Paris (2.56%). However, if we measure Airbnb listings relative to the potential unobserved confounding effects of urban revival.
the number of rented units, the percentage for Barcelona rises to 6.84%, First, we estimate panel fixed-effects specifications that allow neigh-
a figure significantly higher than the other cities.5 This high penetration borhoods to have different time trends. We run i) specifications that
of Airbnb in Barcelona is likely to be explained by the (large) differ- include interaction terms between baseline neighborhood characteris-
ence between the returns of short-term and long-term rentals. At the tics (including the distance to the city centre) and a time trend (either
bottom of Table 1, we provide estimates of the difference in revenue linear or quadratic), ii) specifications that include interaction terms be-
between Airbnb and long-term rentals. In 2015, the average long-term tween these same neighborhood characteristics and (the log of) aggre-
rental price in Barcelona was € 11 per night (€ 735 per month), while the gate regional GDP, iii) specifications that fit neighborhood-specific time
average Airbnb price (short-term rental) was € 71 per night. An Airbnb trends (either linear or quadratic), and iv) specifications on detrended
listing yields the monthly income of a long-term rental in just 10 days data, where the outcome is measured as the deviation from its pre-2013
of occupancy. extended linear trend.
To guide the empirical analysis and to clarify the underlying mecha- Second, we apply an IV strategy, where the instrument is the inter-
nisms of the Airbnb effect on residential housing markets, we develop a action between i) a measure of proximity to the city’s tourist amenities
stylized model where owners can decide to rent long-term to residents or at the neighborhood level and ii) a Google Trends measure that tracks
short-term to tourists. The model predicts that Airbnb will increase hous- changes in Airbnb activity over time. The proximity to tourist amenities
ing prices and rents, with the effect on rents being larger than for prices. predicts Airbnb listing locations, while searches for Airbnb in Google
In terms of testable implications regarding mechanisms, the model pre- predict when these listings are posted. We indirectly verify the exclusion
dicts that Airbnb reduces the long-term supply of residential housing restriction by showing that proximity to tourist amenities does not pre-
units. dict rent and price growth in the pre-Airbnb period (i.e., before 2013).
Third, we also estimate event-study regressions. Specifically, we es-
2
For Barcelona, Airbnb’s market share is around 70% according to the DataHippo
timate interaction terms between year dummies and a continuous mea-
Project (https://datahippo.org/es/) which collects data from several home-sharing plat- sure of Airbnb activity in 2016. This approach allows us to directly check
forms since 2017. We do not use this data-set because it does not cover the period of if housing markets in neighborhoods that experienced a high Airbnb
study. penetration after 2012 were evolving similarly prior to the expansion of
3
‘You’ll never guess which city has the most Airbnb listings’. Forbes. J Bishop 2017.
4
Airbnb as compared to the rest of the city.
Barcelona’s regulation of short-term rental platforms has not substantially changed
during recent years. A city law passed in 2007 (Housing Rights Act 18/2007) states that
tourist apartments that are neither primary nor secondary residences are required to have
a business activity permit. When Airbnb first arrived in Barcelona around 2009, the short-
term rentals of entire apartments without a permit were illegal. Nevertheless, enforce- 6
For the US, processes of urban revival have been described and studied by Baum-
ment of the law was very low until 2016, when the number of inspections substantially Snow and Hartley (2020); Couture and Handbury (2020), while Behrens et al. (2018) fo-
increased. cus on the changes in local businesses associated with gentrification processes. González-
5
Compared to traditional tourist accommodation, the number of active listings was Pampillón et al. (2019) provide some evidence of gentrification in the city center of
equal to 32% of the total number of beds in hotels in the city in 2017. Barcelona.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Finally, we study rent and price dynamics around one specific lo- fraction of residents with low hosting costs. Almagro and Domínguez-
cation, Sagrada Familia, one of the main tourist attractions in the city. Lino (2020) set up a dynamic spatial equilibrium model of residential
While Airbnb activity is high around Sagrada Familia, it is outside the choice and estimate it with data from Amsterdam. The model features
city centre. The results for Sagrada Familia are less prone to be con- multiple endogenous amenities that include the congestion effects of
founded by urban revival dynamics affecting the most central parts of tourism as well as services provided by monopolistically competitive
the city. firms. The authors emphasize that endogenizing neighborhood amenity
All the empirical approaches yield results that are qualitatively and formation is key to understand the nature and welfare consequences of
quantitatively similar, and are consistent with the predictions of our spatial sorting. Unlike the structural approach of these two studies, our
model. Airbnb has increased both rents and prices. Our preferred speci- focus produces reduced-form estimates that are shown to be robust to
fication results suggest that 54 more active listings in a small neighbor- multiple identification threats.
hood (about the average level in 2016) increase rents by 1.9%, while We contribute to the literature in several ways. First, we focus on
transaction and posted prices increase by 4.6% and 3.7%, respectively. a city where the difference in returns between Airbnb and long-term
However, our estimates imply that local impacts can be substantial in rentals is large, resulting in high levels of Airbnb activity. Second, we
the most touristy parts of the city. Our results imply that an increase provide estimates from four different identification strategies that yield
of 200 listings (the average number of listings in the top decile of the qualitatively and quantitatively similar results that cross-validate each
Airbnb activity distribution in 2016) increase rents by 7% and trans- other. This is particularly true for rental prices, where all our estimates
action and posted prices by 17% and 14%, respectively. We also show indicate that the average impact of Airbnb on rents is between 1% and
that Airbnb listings reduce the number of resident households in the 3%. Third, we have access to multiple high-quality micro-level data sets
neighborhood. to track granular changes in housing rents, and posted and transaction
Despite Airbnb being a recent phenomenon, there are already some prices. These micro-level data sets allow us to measure changes in rents
research papers that estimate the effect of Airbnb on housing markets. and housing prices net of composition changes in rented or sold units,
Barron et al. (2020) and Koster et al. (2020) are the two papers that are which is not possible when working with neighborhood average rents or
the most similar to our study. Barron et al. (2020) look at the impact prices. Fourth, we provide direct evidence on the supply mechanism by
of Airbnb on rents and house prices for all cities in the US.7 Their main showing that Airbnb actually reduces the number of households living
strategy consists of using a ‘shift-share’ instrument, where the time vari- in the neighborhood. Fifth, this is the first study to carefully estimate
ation comes from Google Trends of ‘Airbnb’ searches, while the cross- the effects of Airbnb in the context of a large European city8 This is rel-
sectional variation is a neighborhood ‘touristiness’ index based on the evant given the underlying differences between European and US cities.
location of restaurants. They find that a 1% increase in Airbnb listings For instance, European cities might have less excess capacity, where
increases rents by 0.018% and housing prices by 0.026%. Koster et al. guest houses or basement apartments (below a main house) are virtu-
(2020) study the effects of Airbnb bans implemented by several, but ally nonexistent. For the case of Barcelona, our analysis below shows
not all, local governments in the Los Angeles area. Exploiting changes that only a small proportion of housing units active on Airbnb are pri-
in prices at the administrative border, they find that banning Airbnb mary residences. Despite these differences, the results that we find are
decreases prices by about 5%. remarkably close to those found in Barron et al. (2020).
Compared to Koster et al. (2020), our paper focuses on a different The paper is organized as follows. In Section 2, we develop the
channel through which Airbnb affects housing markets. In Koster et al. stylized model that studies the effects of short-term rentals on residen-
(2020) the use a spatial RD design, which compares changes in prices tial housing markets. Section 3 describes the Airbnb, rents and housing
across municipality borders following Airbnb bans. This neatly identifies prices data and describes the most relevant variables. A description of
the price increase of a property due to the possibility of using Airbnb. our empirical strategies is provided in Section 4. The main results are
However, properties located across a border might be part of the same presented and discussed in Section 5, while Section 6 contains the in-
housing market, and, their spatial RD estimates do not capture changes strumental variables and event-study results. Finally, some concluding
in rents and prices that are caused by supply reductions. As Koster et al. remarks are provided in Section 7.
(2020) point out, rents should be smooth at the border. The supply chan-
nel is of great interest from a policy perspective, as rent increases caused 2. Theoretical framework
by Airbnb are among the main complaints made by critics of short-term
rental platforms. In this section we develop a theoretical framework to understand
Calder-Wang (2020) and Almagro and Domínguez-Lino (2020) adopt how short-term rentals to tourists can affect the residential market for
structural approaches to estimate the effects of Airbnb on the welfare of long-term rentals. The model also guides our empirical analysis in terms
heterogeneous residents. Calder-Wang (2020) develops a model where of model specification, threats to identification, estimation strategies
absentee landlords can choose to rent either short-term (to tourists) or and interpretation of the results.
long-term (to residents) and where residents can also host by temporar-
ily renting their home. In the model, estimated with data from New 2.1. Model set-up
York, Airbnb affects residents’ welfare trough two different channels. A
rent channel, that comes from rent increases, following the reduction There are two neighborhoods: a central neighborhood c with fixed
of long-term supply, and a host channel, since residents are allowed to size C, and a suburban neighborhood s, with a housing supply curve
collect income by renting their homes. Her results show that renters’ net which is not completely inelastic. All units in the city are owned by ab-
welfare effect is negative since the rent channel dominates the host chan- sentee owners. In the centre, owners can rent their units to residents (on
nel. She finds that rent increases are widespread across the city, even a long-term basis) or to tourists (on a short-term basis).9 In contrast, the
in neighbourhoods with low Airbnb activity, due to spillover effects. suburban neighborhood s only hosts residents. The masses of residents
In terms of distributional effects, rent increases are particularly high and tourists have been normalized to one and each individual consumes
for high-income, highly-educated people whose preferences are more one unit of housing.
aligned with tourists. Moreover, host gains are concentrated in a small
8
Although not the main aim in Almagro and Domínguez-Lino (2020), the authors use
a shift-share instrumental variables approach that indicates that Airbnb activity increased
7 rents in Amsterdam.
An earlier contribution is Sheppard and Udell (2016) that focuses on New York City.
9
Their results suggest that doubling Airbnb in a 300-meter circle around a property trans- Note that the model leaves out resident home owners as it focuses on the competition
lates to an increase in its value by 6% to 9%. for housing between tourists and renters.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

2.2. Owner choices owner, 𝑇 − 𝑏∗𝑗 = 𝑄𝑐 , we obtain the share of owners that rent short-term
to tourists in neighborhood c:
Owners in neighborhood c can rent their apartments through a long- (𝐴𝑡 − 𝐴𝑟 ) + 𝐶 − 𝛾(1 − 𝐶)
term rental to a resident and obtain an annual market rent Qc or, al- 𝑏∗𝑗 = (1)
2𝐶 + (1 − 𝛼) + 𝛾(𝐶 )
ternatively, rent short-term to tourists and obtain an annual rent of T.
Each owner j, who owns one unit, faces a cost bj to rent short-term to Eq. 1 indicates that the main driver of the penetration of Airbnb in
tourists, which reflects the legal uncertainties or the costs of running a central neighborhood is the tourists’ valuation of amenities relative to
an Airbnb business. The term bj is heterogeneous across owners since the resident’s valuation (𝐴𝑡 − 𝐴𝑟 ).12
they can differ in their risk aversion towards legal uncertainties or their
access to legal services. If 𝑇 − 𝑏𝑗 > 𝑄𝑐 , the owner rents short-term to 2.5. Rental prices
tourists, while if 𝑇 − 𝑏𝑗 ≤ 𝑄𝑐 , the owner rents long-term to a resident.
The cost bj allows T to exceed Qc in equilibrium, which is a salient fea- The equilibrium price of long-term rentals can be obtained by insert-
ture of the data for the case of Barcelona. In equilibrium, there is a ing the market clearing conditions 𝐶(1 − 𝑏∗𝑗 ) = 1 − 𝑒∗𝑖𝑟 and 𝑄𝑠 = 𝛾 𝑒∗𝑖𝑟 in
marginal owner who is indifferent between renting to residents or to the residents’ willingness to pay function:
tourists, 𝑇 − 𝑏∗𝑗 = 𝑄𝑐 , which implies that owners with 𝑏𝑗 < 𝑏∗𝑗 rent short- 𝑄𝑐 = (1 − 𝐶)(1 + 𝛾) + 𝐴𝑟 + (𝐶 + 𝛾 𝐶 − 𝛼) 𝑏∗𝑗 (2)
term, while those with 𝑏𝑗 ≥ 𝑏∗𝑗 rent long-term. Hence, 𝑏∗𝑗 is the share of
units in neighborhood c that are rented on a short-term basis. Eq. 2 indicates that the number of units in the short-term rental mar-
ket affect long-term rents through three different mechanisms. First, one
additional unit in the short-term market reduces the number of long-
2.3. Resident and tourist choices term residents, which mechanically increases the willingness to pay of
the marginal resident as the market clearing condition reveals. Reduc-
The utility that resident i obtains in neighborhood c is 𝑈𝑖𝑟𝑐 = 𝐴𝑟 − ing the supply of long-term units increases prices. The second term is
𝑄𝑐 − 𝛼𝐹𝑏 (𝑏∗𝑗 ) + 𝑒𝑖𝑟 , where Ar reflects the residents’ valuation of amenities a second order general equilibrium effect. An increase in 𝑏∗𝑗 displaces
of neighborhood c, Qc is the rental price, while 𝛼𝐹𝑏 (𝑏∗𝑗 ) is a term reflect- residents from neighborhood c to s, increasing rental prices in the sub-
ing the negative externality that tourism can impose on residents due urbs as equation 𝑄𝑠 = 𝛾 𝑒∗𝑖𝑟 reveals. Rents can increase in areas with lit-
to noise or uncivil behavior.10 , 11 Finally, eir is an idiosyncratic term re- tle to no Airbnb listings as emphasized in Calder-Wang (2020). A higher
flecting the relative preference of resident i to live in neighborhood c as rental price in neighborhood s makes neighborhood c relatively more at-
opposed to neighborhood s. The utility level that resident i would obtain tractive, further increasing rents in neighborhood c. Finally, a marginal
in neighborhood s is 𝑈𝑖𝑟𝑠 = −𝑄𝑠 , where we normalize to zero the value of increase in 𝑏∗𝑗 means higher negative externalities, which contribute to
amenities in neighborhood s. The willingness to pay of the marginal res- lower long-term rents. Provided that these externalties are not too large,
ident to live in neighborhood c is 𝑄𝑐 (𝑒∗𝑖𝑟 ) = 𝐴𝑟 − 𝛼𝐹𝑏 (𝑏∗𝑗 ) + 𝑒∗𝑖𝑟 + 𝑄𝑠 , with the overall effect of Airbnb on rents will be positive.
everyone with 𝑒𝑖𝑟 > 𝑒∗𝑖𝑟 living in the centre and everyone with 𝑒𝑖𝑟 ≤ 𝑒∗𝑖𝑟
living in the suburbs. 2.6. Housing prices
The utility that tourist i obtains if staying in a short-term rental in
neighborhood c is 𝑈𝑖𝑡𝑐 = 𝐴𝑡 − 𝑇 + 𝑒𝑖𝑡 , where At reflects the tourists’ valu- To relate rents and housing prices, we follow the approach of
ation of amenities in neighborhood c, and eit is an idiosyncratic term re- Barron et al. (2020). The market is assumed to be in a steady state,
flecting the preference of tourist i to stay in a short-term rental in neigh- and the price of a housing unit (Pc ) is given by the present value of
borhood c. As in Almagro and Domínguez-Lino (2020), residents and discounted cash flows to the landlord:
tourists might value amenities differently. Tourists have a fixed reserva- [ ]
tion utility level 𝑈𝑡0 = 0, which could reflect the possibility to stay at a ∑∞ 𝑏∗𝑗
𝑃𝑐 = 𝛿 𝑡 (1 − 𝑏∗𝑗 )𝑄𝑐 + (𝑇 − 𝑏𝑗 )𝑑𝑏𝑗
hotel or visit another city. The willingness to pay of the marginal tourist ∫0
𝑡=1
to stay in a short-term rental in neighborhood c is 𝑇 (𝑒∗𝑖𝑡 ) = 𝐴𝑡 + 𝑒∗𝑖𝑡 , with [ ]
only those tourists with 𝑒𝑖𝑡 > 𝑒∗𝑖𝑡 staying in a short-term rental unit in the 1 (𝑏∗𝑗 )2
= 𝑄𝑐 + (𝑇 − 𝑄𝑐 )𝑏∗𝑗 − (3)
centre. 1−𝛿 2

Assuming 𝛿 as the discount factor, the cash flow in each period re-
2.4. The equilibrium units in short-term rentals flects the fact that 1 − 𝑏∗𝑗 units are rented long-term at price Qc , and 𝑏∗𝑗
units are rented in the short-term market at rate T paying the cost bj .
Without loss of generality, we assume that bj , eir and eit are U ~ (0, Eq. 3 indicates that the effects of Airbnb on housing prices (Pc ) will be
1), which simplifies the market clearing conditions in neighborhood c. larger than those on rents (Qc ), as part of the stock available for rent
1 − 𝑒∗𝑖𝑟 and 1 − 𝑒∗𝑖𝑡 are the shares of residents and tourists who stay in obtains a return of 𝑇 − 𝑏𝑗 that is higher than Qc .
the central neighbourhood c. We write 𝐶𝑏∗𝑗 = 1 − 𝑒∗𝑖𝑡 which ensures that
demand for short-term rentals equals its supply, while market clearing 2.7. Implications for the empirical analysis
for long-term rentals implies 𝐶(1 − 𝑏∗𝑗 ) = 1 − 𝑒∗𝑖𝑟 . In neighborhood s, the
long-term rental price is assumed to be an increasing function of its Eqs. 2 and 3 motivate our empirical analysis consisting of relating
population. Specifically, we posit that 𝑄𝑠 = 𝛾 𝑒∗𝑖𝑟 , with 𝛾 > 0. Combining changes in housing rents or prices with changes in Airbnb activity at the
the market clearing conditions, the willingness to pay of the marginal neighborhood level. From the theoretical framework developed here, we
resident and tourist, 𝑄𝑐 (𝑒∗𝑖𝑟 ) and 𝑇 (𝑒∗𝑖𝑡 ), and the definition of the marginal draw five implications for the empirical analysis.
First, the model predicts that Airbnb activity increases rents and
prices, with the effect on prices being larger in absolute value.
10
Tourism as a negative externality is in line with the local population’s perception of Second, inspecting Eqs. 1–3 reveal the main identification threat
tourism as a negative phenomena in Barcelona. This is documented by an opinion poll
made by local authorities since 2011, which surveys citizen perception of Barcelona’s
faced in the empirical analysis. The effect of Airbnb activity on resi-
most important problems. In this poll, tourism was mentioned, on average, as the city’s dential housing markets will be biased if neighborhoods where Airbnb
fourth largest problem during the entire period, reaching the top ranking in 2017.
11
Almagro and Domínguez-Lino (2020) posit that tourism affects residents through a
direct negative effect and indirectly by changing the availability of non-tradable goods
12
For 0 ≤ 𝑏∗𝑗 ≤ 1, it has to be the case that 𝛾(1 − 𝐶) − 𝐶 ≤ (𝐴𝑡 − 𝐴𝑛 ) ≤ 𝐶 + 1 −
and services such as child care facilities. 𝛼 + 𝛾.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

penetration is high are simultaneously experiencing changes in the resi- activity prior to 2015 (as well as in between the rest of the data points).
dents’ willingness to pay. It could be that neighborhoods where At is high This strategy is supported by the fact that, according to Airbnb, 72%
are becoming increasingly popular among residents. Central neighbor- of guests leave a review.14 We follow Zervas et al. (2017) and consider
hoods where Airbnb activity is higher are going through gentrification that a listing is active in a given quarter if it has received at least one
processes. We will extensively address these concerns in the empirical review during that quarter.15
analysis. The potential consequences of Airbnb might be very different if the
Third, besides the increase in rents (and prices), the model indicates platform is used to rent out excess capacity (home-sharing), or if units
that Airbnb reduces the supply of units in the long-term rental market. are rented short-term through Airbnb all year long. We label listings
As a consequence, the model predicts that Airbnb displaces residents. that correspond to this second category as ‘commercial’. Listings in
To test the model’s main mechanism we will also estimate the effect of Airbnb are entire apartments, private rooms or shared rooms. We consider
Airbnb activity on the number of resident households. multi-hosted properties (host has more than one listing) and single-
Fourth, as we have seen above, Eq. 1 predicts that the penetration of hosted entire apartments with a minimum of 5 reviews per quarter to
Airbnb is (partly) determined by the presence of neighborhood ameni- be commercial. This definition is clearly conservative, as many entire
ties that are more important to tourists than they are to residents. In properties are rented as separate private rooms.16 Despite this, more
Section 4.2 we document that, the proximity to relevant tourist attrac- than 75% of all listings in every single year in our sample correspond
tions is a strong predictor of Airbnb activity at the neighborhood level. to this commercial category. Although some genuine home-sharing
This observation motivates the instrumental variables strategy devel- exists on the platform, Airbnb in Barcelona is mostly a commercial
oped in Section 4.2. activity.
Fifth, our empirical analysis essentially estimates the effect of Airbnb
by comparing changes in rents (or prices) between neighborhoods with
high versus low Airbnb activity. Subtracting Qs from Eq. 2 yields: 3.3. Rents and prices

𝑄𝑐 − 𝑄𝑠 = (1 − 𝐶) + 𝐴𝑟 + (𝐶 − 𝛼) 𝑏∗𝑗 (4) We use two sources of data to obtain information on rents and prices
at a fine spatial level. In particular, we have two measures for prices
Eq. 4 reveals that rent comparisons between neighborhoods with dif-
(transaction prices and posted prices) and one measure for rents (posted
ferent levels of Airbnb penetration provides a lower bound of the total
rents). For transaction prices, we use data from the Catalan Tax Author-
effect of Airbnb. Differences in rents (or prices) net-out the second-order
ity from transaction tax records, which includes the price, exact location,
general equilibrium effects that increase rents and prices throughout the
date of transaction, size of the housing unit, year of construction, and a
city.
variable reflecting the quality of the dwelling. We have the universe of
transactions that occurred in Barcelona during the period 2009–2017.17
3. Data and variables We label this dataset ITP (Impuesto sobre Transmisiones Patrimoniales) or
transaction prices.
3.1. Neighborhood definition For posted rents and prices, we use information from the online real
estate portal Idealista. With more than one million ads and an average of
Our geographical unit of analysis is the Basic Statistical Area (BSA). 17 million weekly views, Idealista is by far the most important Spanish
BSAs are built and used by the Barcelona City Hall for statistical pur- real estate portal. Idealista provided us with all ads that were active for
poses. There are a total of 233 BSAs with an average of 7122 inhabitants, the city of Barcelona in December of every year for the period 2007–
but due to data restrictions we keep 221 BSAs in our sample. We believe 201718 , 19 The data include the exact location, the posted rent or price
that BSAs are the appropriate neighborhood definition, as they are de- and the size of the unit, among other characteristics20
signed to contain population with similar socio-economic characteristics Having two measures of prices is useful because both transaction and
and their size is sufficient to generate meaningful measures of housing posted prices have advantages and disadvantages. Posted prices might
rents and prices for neighborhoods over time. differ from final prices since bargaining is a regular part of the process.
Official transaction prices should, in principle, measure prices more pre-
cisely. However, in practice, the transaction (ITP) data have two limi-
3.2. Airbnb tations. First, there might be a non-negligible time lapse between the

To measure Airbnb activity, we use information extracted directly

from the Airbnb website. InsideAirbnb is a dataset collected at differ- 14
Although some guests do not leave a review, there are no reasons to believe that the
ent points in time by Murray Cox, an independent Internet user who percentage of guests that leave a review changes non-randomly across space or time.
has made it publicly available. For Barcelona, it contains 21 data points 15
An alternative approach would be to use the entry date and assume that listings never
between April 2015 and February 2018.13 Each listing has information exit, which is the preferred method in Barron et al. (2020). In the case of Barcelona, we
consider that this approach is problematic. First, approximately 25% of all listings do not
on the host ID, geographical coordinates, room characteristics, date the
have any reviews at all. Second, the entry date indicates the time when the host registered.
host registered, and date of each guest review. Even though Airbnb is If the host has multiple listings (which is the case for the majority of listings in Barcelona),
not the only home-sharing platform active in the city, we consider that it is not possible to know the entry date of each listing.
its listings are a good proxy for the short-term rental market. Its market 16
This practice increased after July 2016. Short-term rentals of entire apartments without
share is by far the highest among its competitors and most short-term a permit is illegal, but enforcement was very low before July 2016. It is less clear if renting
a private room is also against the law, and in practice, enforcement with respect to private
rentals are advertised through more than one platform simultaneously,
rooms has been low throughout the period we study.
implying that adding listings from a second platform would cause sig- 17
We keep only those sales transactions with a declared value of less than 10,000,000
nificant double counting. euros.
For our purposes, it is crucial to identify a listing’s active period. 18
We have dropped the following data: sales ads with posted prices below 10,000 euros
Even though the information started being scraped in early 2015, by and those of less than 20 square meters, and we drop all ads with monthly rents below
100 euros or above 30,000 euros.
exploiting the date of each review, we are able to reconstruct the listing’s 19
Idealista’s monthly aggregate data for Barcelona show little seasonality, indicating
that using December data should not be a limitation in our context.
20
Other characteristics that are available and that we use are number of floors, num-
13
Details about the website can be accessed through http://insideairbnb. ber of rooms, presence of air-conditioning, lift and boxroom, and whether it is a studio,
com/about.html. penthouse, or duplex.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. 1. Airbnb listings, rents and prices: 2007–2017. Notes: This graph plots the
evolution over time of the BSA averages in Airbnb listings, rents and posted
prices (per square meter) for the period 2007–2017. Rents and prices are nor-
malized to their 2013 value. The dark gray bars represent the average Airbnb
listings for all BSA, while the light gray bars depict the average listings for High
Airbnb Areas (BSAs in the top decile of the Airbnb listings distribution in 2016).
Fig. 2. Airbnb activity across BSAs in 2016. Notes: This graph plots the number
of active Airbnb listings in the fourth quarter of 2016 at the BSA level.

date at which parties agree on a price and the date when the ITP tax is
paid. Second, there is some fraud in the ITP tax that consists of under-
the years 2012 and 2016 for two different samples: all BSAs and High
reporting the ITP price and the tax base. For rents, we cannot compare
Airbnb Areas (BSAs in the top decile of the Airbnb listings distribution
posted to actual rents. However, Chapelle and Eymeoud (2018) show,
in 2016).
in the French context, that bargaining is less of an issue for rents and
that online posted prices are a good measure of actual rents.
4. Empirical strategies
3.4. Descriptive statistics
4.1. Baseline specification
In Fig. 1, we plot the evolution over time of Airbnb activity, together
with that of rents and prices. Airbnb experienced a very rapid increase Our main analysis consists of estimating the following fixed-effects
from its first entrance in 2009 up to 2016, when the growth stopped specification (and variants of it):
because of City Hall’s increased efforts to reduce tourist apartments op-
log(𝑌𝑛,𝑡 ) = 𝛽𝐴𝑖𝑟𝑏𝑛𝑏 𝐶 𝑜𝑢𝑛𝑡𝑛,𝑡 + 𝛾𝑋𝑛,𝑡 + 𝜏𝑡 + 𝜇𝑛 + 𝜀𝑛,𝑡 (5)
erating without a license. In 2016, the average BSA had 54 listings, while
High Airbnb Areas (those BSAs in the top decile) had an average of 200 where Yn,t is our measure of housing rents or prices at the BSA level,
active listings. In these areas of the city, approximately 5% of all hous- Airbnb Countn,t is the number of active listings at time t in BSA n, 𝜏 t are
ing units are listed on Airbnb. The substantial variation in Airbnb activ- time fixed-effects, and 𝜇 n are BSA fixed effects that account for time-
ity across neighborhoods is further explored in Fig. 2, which shows the invariant neighborhood characteristics. Our dependent variable log(Yn,t )
distribution of Airbnb listings across BSAs for the last quarter of 2016. is the average residual at the BSA-time period level of a (micro-level)
Airbnb activity is higher around the city center and, to some extent, regression in which log rents (or log housing prices) are regressed on
along portions of the beach line or around the Sagrada Familia. Airbnb time dummies and unit characteristics.22 This controls for price changes
activity is low in many other parts of the city. across neighborhoods that could be explained by changes in the compo-
The evolution of housing rents and prices in the period we study sition of units rented or sold across BSAs and over time. For example,
is turbulent. Following the financial crisis and the burst of the Spanish it allows us to control for BSAs that might have a growing proportion
housing bubble, rents and prices fell until 2013, when they started re- of high-end apartments being sold or rented over time. Throughout the
covering as the economy started to grow. While prices reached pre-crisis regression analyses, we weight BSA-time cells by the relevant number
levels towards the end of the period, rents surpassed pre-crisis levels of ads or sales. Standard errors are clustered at the BSA level to account
around 2015 and kept rising. During this period, housing affordability for serial correlation within panel units (Bertrand et al., 2004).
was one of the main concerns among residents.21 The recovery of rents Our main concern regarding identification is that neighborhoods
and prices coincides in time with the return of economic growth and with the most Airbnb activity growth during our period of study might
the expansion of Airbnb. We address the endogeneity concerns related be experiencing processes of sociodemographic change, which might
to this coincidence in the empirical section. have a direct impact on housing rents and prices. Airbnb has grown
We complement the data on rents, prices and Airbnb listings with the most in central parts of the city that have also been experiencing
a comprehensive set of socioeconomic BSA characteristics including processes of urban revival in the last two decades. We adopt several
population, age distribution, relative income index, average household
size, unemployment level and the percentage of foreign population. In
Table 2, we report descriptive statistics. We present the BSA means for 22 We construct a panel on the BSA-year (data from Idealista) and BSA-quarter (transac-

tion prices). Unit characteristics in the Idealista database are size, number of floors, num-
ber of rooms, air-conditioning, lift and boxroom, and whether it is a studio, penthouse, or
21
See ‘El acceso a la vivienda, el principal problema de los barceloneses’ Macedo, G., October duplex. In the ITP database, dwelling characteristics are size, year of construction, and a
2010. variable reflecting the quality of the dwelling (with scale 1 to 8).
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Table 2
Descriptive statistics: Variables’ means across BSAs for 2012 and 2016.

2012 2016

All High All High

BSAs Airbnb Areas BSAs Airbnb Areas

Airbnb Count 2.79 13.35 47.84 178.58

Rent (€ /m2 ) 11.83 12.93 16.39 20.19
Posted Price (€ /m2 ) 3250 3338 3753 4282
Transaction Price (€ /m2 ) 2269 2356 2619 3027
Population 6978 7750 6973 7514
Population Density 0.03 0.04 0.03 0.04
Mean Age 43.36 42.10 43.69 42.08
% of Foreign Population 0.18 0.32 0.17 0.33
Household Size 2.47 2.41 2.48 2.41
Unemployment Rate 10.48 10.81 7.80 7.83
Income Index 98.37 96.48 102.78 104.58

Notes: Columns 1 and 3 report the mean for all BSAs in 2012 and 2016. Columns 2 and 4 report the means of High
Airbnb Areas (BSAs in the top decile of the Airbnb activity. distribution in 2016).

strategies to control for the potential confounding effects of gentrifica-

tion.
First, we introduce in Eq. 5 a set of time-varying controls at the BSA
level (Xn,t ); average age, log of population density, average household
size, unemployment rate, relative income, and percentage of foreign res-
idents. Since this equation includes BSA-fixed effects, this allows us to
control for yearly changes in variables associated with processes of gen-
trification. In some specifications, we allow for neighborhoods with dif-
ferent characteristics to have different time trends (linear or quadratic).
We do so by introducing, as additional regressors, interaction terms be-
tween the time trend and the control variables measured in 2012, i.e.,
Xn,2012 × t in the linear case or Xn,2012 × t and Xn,2012 × t2 in the quadratic
case. In this specification, we also include the interaction term between
the time trend and the distance to the city center.23 This would allow,
for instance, more central neighborhoods to have a steeper time trend.
Second, in a more data demanding approach, we include BSA-
specific time trends (linear or quadratic). Specifically, we add 𝜌n × t
interaction terms for the linear case and 𝜌𝑛 × 𝑡 + 𝜓𝑛 × 𝑡2 for the quadratic
case. This is a very flexible specification since it allows each BSA to have
its own time trajectory in housing rents and prices. Here, the variations
that we exploit are deviations from each BSA’s own specific linear (or
quadradic) time trend.
If Airbnb affects not only levels but also trends of these variables,
including BSA-specific time trends would not be appropriate since it Fig. 3. Location of tourist amenities. Notes: Location of tourist amenities across
would capture both the effect of Airbnb and BSA-specific time trends the city, the size of the circles is proportional to the number of reviews. The
(Wolfers, 2006). We resort to a detrending procedure previously applied darker area shows the city’s historical district.
in the taxation (Kleven et al., 2014) and minimum wage (Monras, 2019)
literatures and estimate linear time trends using data prior to 2013 only
4.2. Instrumental variables fixed-effects models
(i.e. the pre-Airbnb period). We estimate the following two equations at
a neighborhood-time level:
As an alternative approach to tackle the endogeneity of Airbnb loca-
tion, we also estimate Eq. 5 through a Two-Stage Least Squares regres-
log(𝑌𝑛,𝑡 ) = 𝜇𝑛 + 𝜏𝑡 + 𝜌𝑛 × 𝑡 + 𝜖𝑛,𝑡 , for 𝑡 ≤ 2012 (6) sion. Our theoretical model indicates that short-term rentals locate in
areas where tourist amenities are high. Following this prediction, we use
̃
𝑙𝑜𝑔( 𝑌𝑛,𝑡 ) = 𝛽𝐴𝑖𝑟𝑏𝑛𝑏 𝑐𝑜𝑢𝑛𝑡𝑛,𝑡 + 𝛾𝑋𝑛,𝑡 + 𝜏𝑡 + 𝜀𝑛,𝑡 , for all 𝑡 (7) a shift-share variable as an instrument that combines i) cross-sectional
variation across BSAs in tourist amenities and ii) aggregate time varia-
tion in Airbnb activity.
The first equation estimates the outcome based on BSA dummies,
For the cross-sectional ‘share’ component of the instrument, we build
time dummies, and BSA specific linear time trends for the years up to
an index that measures proximity to tourist amenities. Our instrument
2012. Based on these OLS coefficients, we predict log(Yn,t ) for the entire
aims to capture the set of amenities that tourists enjoy while not being
̃
sample years and compute the residuals, 𝑙𝑜𝑔( 𝑌𝑛,𝑡 ). In the second stage of particular interest to residents. We use TripAdvisor to produce a com-
(Eq. 7), we regress these detrended residuals against Airbnb listings, plete list of the city’s tourist amenities.24 We geolocate these amenities
time dummies and the time-varying controls (Xn,t ).

24
TripAdvisor is a website that offers tourism-related content. According to the site,
it currently has over 390 million monthly unique visitors. We exclude the more en-
23
Distance to the city center is measured as the distance from Plaça Catalunya (the main dogenous and less historical amenities such as areas known for restaurants, bars or
city square) to the centroid of each BSA. clubs.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. 4. Airbnb activity, tourist amenities and Google Trends searches. Notes: Graph (a) shows BSAs’ Airbnb listings as a function of the tourist amenities index (bins
are deciles of the tourist amenities distribution). Graph (b) shows the time profile of Airbnb listings (solid line, left axis) and Google Trends searches for ‘Airbnb
Barcelona’ (dashed line, right axis).

and collect the number of Google reviews of each attraction. We use the the study period for a reason other than the presence of tourism. If the
number of reviews to weight the relative importance of each site.25 Our instrument is valid, proximity to tourist amenities should not explain
measure of tourist amenities is built as follows: changes in housing rents and prices prior to the arrival of Airbnb. We
∑ 1 address this issue at length below.
𝑇 𝑜𝑢𝑟𝑖𝑠𝑡 𝐴𝑚𝑒𝑛𝑖𝑡𝑖𝑒𝑠𝑛 = × 𝑅𝑒𝑣𝑖𝑒𝑤𝑠𝑘 (8)
𝑘
𝑑𝑖𝑠𝑡𝑛,𝑘
4.3. Event study plots
where k indicates the amenity, distn,k is the distance in meters between
the centroid of each BSA n and amenity k, and Reviewsk is the number We conduct an event study exercise, using the following regressions:
of Google reviews. Fig. 3 shows the location of these amenities, where
the size of each circle is proportional to the number of reviews. ∑
As the ‘share’ component of the instrument, the tourist amenities in- log(𝑌𝑛,𝑡 ) = 𝛿𝑡 × 𝐴𝑖𝑟𝑏𝑛𝑏𝐶𝑜𝑢𝑛𝑡𝑛,2016 + 𝛾𝑋𝑛,𝑡 + 𝜇𝑛 + 𝜏𝑡 + 𝜀𝑛,𝑡 (9)
dex should predict where Airbnb listings will appear. Panel a) in Fig. 4 𝑡≠2012

plots this relationship by binning the data for deciles of the tourist in- where AirbnbCountn,2016 is the number of listings in BSA n in 2016. Like
dex distribution. The graph clearly shows that BSAs that are closer to in previous regressions, we include time and BSA fixed effects and time-
tourist amenities tend to show the highest number of Airbnb listings. varying characteristics (Xn,t ). We estimate AirbnbCountn,2016 × year
This relationship can be rationalized by the model of Section 2. Tourist interactions, leaving 2012 as the base year. This approach allows us to
amenities (At ) increase the tourists willingness to pay which increases estimate the yearly effect of having one additional listing in 2016. Again,
both Airbnb prices (T) and Airbnb activity (𝑏∗𝑗 ). In Fig. A.1 in the Ap- we choose 2012 as the last pre-Airbnb year as, starting in 2013, Airbnb’s
pendix, we show that neighborhoods with high levels of Airbnb activity activity became more significant. This exercise allows us to check if,
have more expensive Airbnb listings. prior to the arrival of Airbnb, areas that will experience higher Airbnb
Turning to the ‘shift’ component of the instrument, we follow activity display similar trends in housing rents and prices compared to
Barron et al. (2020) and use worldwide searches in Google for the term other neighborhoods.
‘Airbnb Barcelona’. This variable is measured at a monthly level and is
normalized to 100 for the month with the highest number of searches. 4.4. Evidence from Sagrada Familia
Panel b) of Fig. 4 shows that the number of Google Trends searches for
‘Airbnb Barcelona’ tracks the time variation in Airbnb activity very well. In our last empirical strategy, we will focus on Sagrada Familia, one of
The rationale behind the instrument works as follows. The proximity the main tourist amenities in the city. It is one of the three major tourist
to tourist amenities predicts where Airbnb listings locate, while searches amenities not found in the city centre as shown in Fig. 3. The other
in Google Trends for the term ‘Airbnb Barcelona’ predict when listings two non-central hotspots are Camp Nou (north-west of city centre) and
appear. Fig. 4 provides suggestive evidence of the relevance of the in- Parc Güell (north-west of Sagrada Familia). Fig. 2 shows that only the
strument. We also test for this in the first stage regressions. area around Sagrada Familia has a high level of Airbnb activity, possibly
As for the exclusion restriction, recent research on shift-share instru- because Camp Nou and Parc Güell are too far from the city centre. Unlike
ments indicates that the main identification threats are related to the the most central parts of the city, the area around Sagrada Familia is
‘share’ component of the instrument (Goldsmith-Pinkham et al., 2020). an upper-middle class residential neighborhood. In 2000, this area was
Since our specifications contain a BSA fixed effect, instrument valid- ranked 14 out of 38 broad neighborhoods by relative family income26 .
ity hinges on the assumption that the cross-sectional ‘share’ component, Its position in this ranking was 13 in 2008, indicating that this area
proximity to tourist amenities, is only correlated with changes in housing was not experiencing gentrification in the pre-Airbnb period27 Sagrada
rents and prices through Airbnb listings. For example, our instrument
would be invalid if residents’ valuation of proximity to tourist ameni-
26
Before 2008, income is only available for 38 broad neighborhoods
ties (or any other BSA characteristic that correlates with it) changes over 27
Similarly, the share of population between 20 and 34 years, which is another indica-
tor associated with gentrification, has also remained stable over the same period in the
25
Although TripAdvisor also provides reviews, Google has more. Sagrada Familia area. It was equivalent to 26% in 2000 and to 27% in 2008.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. 5. Evolution of rents and prices for High Airbnb Area vs. the rest. Notes: Rents, Transaction and Posted prices are expressed in logs. Graphs plot raw averages
and the appropriate confidence intervals. High Airbnb Area are BSAs in the top decile of the Airbnb listings distribution in 2016.

Familia provides us a setting to study the effects of short-term rentals 5. Main results
where concerns regarding the confounding effects of urban revival are
diluted. 5.1. Graphical evidence
First, we estimate Eq. 5 by Two-Stage Least Squares where the instru-
ment is the interaction term between the inverse distance to Sagrada Fa- Before proceeding to the regression results, in Fig. 5, we show graph-
milia (1/distSFn ) and our measure of Google Trends searches. In this case, ical evidence of the effect of Airbnb on housing markets. We plot raw av-
proximity to this particular attraction predicts the location of Airbnb erage (log) prices and rents series over time for High Airbnb Areas (BSAs
listings, while, as before, Google Trends predicts the timing of Airbnb in the top decile of Airbnb listings distribution in 2016) versus the rest.
arrival and expansion. We argue that this is an exogenous instrument In Panel (a) we graph rents, while in Panels (b) and (c), we show corre-
since it is unlikely that residents’ preferences to locate close to Sagrada sponding graphs for transaction prices and posted prices, respectively.
Familia had change during the period 2007–2017 for a reason other than For completeness, in Panel (d), we plot our measure of Airbnb activity.
tourism. The levels of both rents and prices tend to be higher in BSAs with
Second, we also replicate our event-study design but focus on the more Airbnb activity. While the series for the period before 2012 ap-
proximity to Sagrada Familia as a predictor of Airbnb activity. We es- pear fairly parallel, the gaps in rents and prices seem to widen, with the
timate 1/distSFn × year interactions, while controlling for the usual expansion of Airbnb in 2013 and onwards, especially for rents and trans-
neighborhood characteristics and fixed effects. action prices, where the divergence is more noticeable. In the first three
figures, the difference between the two groups is statistically significant
∑ 1 at the end of the period, while this is not the case for the first years.
log(𝑌𝑛,𝑡 ) = 𝛿𝑡 × + 𝛾𝑋𝑛,𝑡 + 𝜇𝑛 + 𝜏𝑡 + 𝜀𝑛,𝑡 (10)
𝑡≠2012
𝑑𝑖𝑠𝑡𝑆𝐹𝑛 Finally, in Panel (d), we report the evolution of the count of Airbnb
listings by group. While the number of listings increased drastically for
This strategy is useful to determine whether BSAs at different dis- the High Airbnb Areas, the increase was very modest for the other BSAs,
tances (measured in kilometers) to Sagrada Familia experience similar reflecting fact that Airbnb is highly concentrated in particular areas of
trends in rents and prices before and after the arrival of Airbnb. the city.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Table 3
Impact of Airbnb on rents and prices - Baseline Specifications.

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Panel A Rents
Airbnb Count (x100) 0.036∗ ∗ ∗ 0.035∗ ∗ ∗ 0.037∗ ∗ ∗ 0.036∗ ∗ ∗ 0.058∗ ∗ ∗ 0.051∗ ∗ ∗ 0.034∗ 0.038∗ ∗ ∗ 0.053∗ ∗ ∗
(0.008) (0.009) (0.009) (0.009) (0.020) (0.024) (0.018) (0.008) (0.010)
N 2.123 2.123 2.123 2.123 2.123 2.123 2.123 2.123 1.920

Panel B Transaction Prices

Airbnb Count (x100) 0.110∗ ∗ ∗ 0.085∗ ∗ ∗ 0.039∗ ∗ 0.040∗ ∗ 0.052∗ ∗ ∗ 0.062∗ ∗ 0.082∗ ∗ ∗ 0.063∗ ∗ ∗ 0.084∗ ∗ ∗
(0.019) (0.016) (0.017) (0.018) (0.020) (0.025) (0.025) (0.017) (0.022)
N 7.901 7.901 7.901 7.901 7.901 7.901 7.901 7.901 7.228

Panel C Posted Prices

Airbnb Count (x100) 0.081∗ ∗ ∗ 0.068∗ ∗ ∗ 0.032∗ ∗ ∗ 0.026∗ ∗ ∗ 0.022 0.005 0.084∗ ∗ ∗ 0.045∗ ∗ ∗ 0.075∗ ∗ ∗
(0.010) (0.009) (0.010) (0.010) (0.019) (0.023) (0.21) (0.010) (0.012)
N 2.229 2.229 2.229 2.229 2.229 2.229 2.229 2.229 2.024
Time FE X X X X X X X X X
BSA FE X X X X X X X X X
Controls - X X X X X X X X
Trends - - Xn,2012 (L) Xn,2012 (Q) BSA (L) BSA (Q) - - -
Detrendend BSA - - - - - - X - -
Xn,2012 × GDP - - - - - - - X -
No Hist Dist - - - - - - - - X

Notes: Significance is indicated by ∗ p<0.1, ∗ ∗ p<0.05, and ∗ ∗ ∗ p<0.01. Standard errors, in parentheses, are clustered at the BSA level. Each column represents a
different specification. Panel A reports the results for rents, while Panels B and C report the corresponding estimates for transaction and posted prices. Outcomes are
average residuals at the BSA-time period level, as explained in the main text. Regressions weighted with the total number of ads (for rents and posted prices) and
of transactions (for prices). The analysis takes place at the BSA-year level for rents and posted prices and BSA-quarter for transaction prices. Controls are: average
age, log of population density, average household size, unemployment rate, average relative income, and percentage of foreign residents. Trends can be either
characteristic-specific (Xn,2012 ) or BSA-specific (BSA) and either linear (L) or quadratic (Q). Characteristic specific trends also include a distance to city center trend.

These graphs are suggestive evidence that neighborhoods with Table 4

higher Airbnb penetration also experienced higher rents and price Impact of Airbnb on the number of households, household size and population.
growth with the arrival and expansion of Airbnb. Since these series (1) (2) (3) (4)
might be affected by other confounding factors that could be biasing the
Panel A Outcome: log(Number of Households)
results, we move to our main empirical strategies described in Section 4.
Airbnb Count (x100) -0.014∗ ∗ ∗ -0.024∗ ∗ ∗ -0.010∗ -0.007∗ ∗
(0.005) (0.006) (0.006) (0.004)
5.2. Baseline results
Panel B Outcome: log(Household Size)
In Table 3, we report our baseline results for the impact of Airbnb Airbnb Count (x100) -0.002 -0.018∗ ∗ ∗ -0.009∗ 0.004
(0.004) (0.005) (0.005) (0.005)
on rents (Panel A) and prices (Panels B and C). As explained above,
throughout the table, the dependent variable is the average BSA-time Panel C Outcome: log(Population)
period residual of a micro regression in which log rents (or log prices) Airbnb Count (x100) -0.016∗ -0.043∗ ∗ ∗ -0.020∗ ∗ 0.002
(0.009) (0.009) (0.008) (0.004)
are regressed on housing characteristics and time dummies. N 2056 2056 2056 2056
In column 1, we regress the outcome of interest against the number Time FE X X X X
of Airbnb listings while controlling only for time and BSA-fixed effects. BSA FE X X X X
Then, in column 2, we add BSA time-varying controls. Coefficients are Controls - X X X
Time Trends - - Xn,2012 (L) BSA (L)
positive and significant for both rents and prices, which implies that an
increase in the number of listings translates into an increase in rents and Notes: Significance is indicated by ∗ p<0.1, ∗ ∗ p<0.05, and ∗ ∗ ∗ p<0.01. Standard
prices. The effects on prices are larger than on rents, especially for trans- errors, in parentheses, are clustered at the BSA level. Each cell represents a
action prices. The presence of contemporaneous controls has no large different regression with the log the number of households (panel A), the log
impact on the estimates for rents, while it slightly decreases coefficients of household size (panel B) and the log of population (panel C). The analysis
for prices, although not in a statistically significant way. Nevertheless, takes place at the BSA-year level for the period 2009–2017. Controls are average
we keep the socioeconomic controls in subsequent specifications for the age, unemployment rate, average relative income, and percentage of foreign
residents. Linear trends can be either characteristic-specific (Xn,2012 (L)) or BSA-
sake of completeness.
specific (BSA (L)). Characteristic specific trends also include a distance to city
In column 3 (4), we include socioeconomic-specific linear (or
center trend.
quadratic) time trends by introducing interaction terms between a linear
(or quadratic) time trend and the control variables measured in 2012 as
detailed in Section 4. The coefficients for prices are somewhat reduced, prices. Nevertheless, by introducing quadratic trends we might be over-
while they remain fairly constant for rents. Then, we report the results of fitting the model as a vast majority of the BSA quadratic trend coef-
specifications that fit BSA-specific time trends. Column 5 shows the re- ficients are non-significant. For posted prices, where results are more
sults for linear trends and the results for quadratic trends are presented sensitive to the inclusion of quadratic trends, the F-test of their joint
in column 6. These allow for both observable and unobservable char- significance is only 1.60.
acteristics to impact neighborhood trends. The inclusion of linear time As mentioned before, one caveat of this approach is that if Airbnb
trends increases the coefficient for rents (though not significantly) and impacts rent and price trends rather than levels, the BSA fixed effects
reduces the coefficients for prices, especially for posted prices where the will absorb part of the Airbnb effect on the outcomes. In column 7, we
coefficient becomes non-significant. As for regressions with quadratic repeat the analysis after detrending the data following the procedure
trends, they do not substantially change the coefficients of rents and described in Section 4. In a first step, the pre-Airbnb data are used to es-
transaction prices but it further decreases the coefficient for posted timate BSA-specific time trends, which are then used to detrend all data
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Table 5
Impact of Airbnb on rents and prices: IV regressions.

First Stage Second Stage First Stage Second Stage

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

Panel A Rents
Airbnb Count (x100) 0.022∗ ∗ 0.033∗ ∗ ∗
(0.011) (0.010)
TouristAmenities × GoogleTrends 0.005∗ ∗ ∗ 0.005∗ ∗ ∗
(0.000) (0.001)
N 2.123 2.123 2.123 2.123
F-stat. excl. inst. 192.2 70.2

Panel B Transaction Prices

Airbnb Count (x100) 0.123∗ ∗ ∗ 0.104∗ ∗ ∗
(0.020) (0.023)
TouristAmenities × GoogleTrends 0.004∗ ∗ ∗ 0.003∗ ∗ ∗
(0.000) (0.000)
N 7.228 7.228 7.228 7.228
F-stat. excl. inst. 217.8 61.3

Panel C Posted Prices

Airbnb Count (x100) 0.074∗ ∗ ∗ 0.047∗ ∗ ∗
(0.014) (0.013)
TouristAmenities × GoogleTrends 0.005∗ ∗ ∗ 0.005∗ ∗ ∗
(0.000) (0.001)
N 2.229 2.229 2.229 2.229
F-stat. excl. inst. 159.0 70.5
Time FE X X X X
BSA FE X X X X
Controls X X X X
Xn,2012 × GDP - - X X

Notes: Significance is indicated by ∗ p<0.1, ∗ ∗ p<0.05, and ∗ ∗ ∗ p<0.01. Standard errors, in parentheses, are clustered at the BSA
level. Panel A reports the results for rents, while Panels B and C report the corresponding estimates for transaction and posted prices.
Outcomes are average residuals at the BSA-time period level, as explained in the main text. Regressions weighted with the total
number of ads (for rents and posted prices) and of transactions (for prices). The analysis takes place at the BSA-year level for rents
and posted prices and BSA-quarter for transaction prices. Controls are: average age, log of population density, average household
size, unemployment rate, average relative income, and percentage of foreign residents. Xn,2012 × GDP also includes distance to city
center times GDP.

points. Here, the coefficient for rents slightly decreases and gets closer plotting the result of multiplying the coefficients obtained in column 2
to the specifications reported in columns 1 and 2. For prices, they both by the Airbnb activity of each BSA in 2016. While the implied effects
increase with respect to columns 3 to 6, and their magnitude becomes are very close to zero for the less central BSAs, our estimates imply some
more similar to each other. While these estimates provide yet another local impacts that are substantial. For the High Airbnb Areas, Airbnb has
robustness test of the results, we acknowledge that the pre-Airbnb pe- increased rents, transaction prices and posted prices by an average of
riod is admittedly short to reliably estimate the pre-trends needed to 7%, 17% and 14%, respectively.
detrend the post-Airbnb data. Across the different specifications, the results indicate that higher
The arrival and expansion of Airbnb coincides with the period of Airbnb penetration leads to increases in both rents and prices, with the
economic recovery that started in 2013. One concern is that economic effects on prices being larger than on rents. These results are an empir-
growth might have impacted rents and prices differently across the city. ical test of the predictions of our theoretical model of Section 2. The
In column 8, we add an interaction term between each control variable results suggest that the net effect of Airbnb activity positively affects
in 2012 (including the distance to the city centre) with the log of re- rents and prices, which implies that the possible negative externalities
gional GDP. This allows us to control for areas reacting differently to associated with Airbnb do not offset its inflationary effects. Since hous-
economic growth. Our coefficients are still positive, strongly significant ing units that are on Airbnb yield, on average, a higher return than
and of a similar magnitude after this inclusion. Finally, in a last sensitiv- those units that are rented to residents, the housing price increase due
ity test, we also show that our findings are not driven by neighborhoods to Airbnb exceeds that of long-term rents.
in the historical city center (Ciutat Vella), characterised by very high The results of specifications with BSA-specific linear time trends in
levels of Airbnb activity. The results reported in column 9 show that column 5 are close to those reported in column 2, which corresponds
coefficients are still significant and of a similar magnitude after BSAs in to a more parsimonious specification with BSA and time fixed effects
the more central parts of the city are excluded. and time-varying control variables. Overall, we consider the estimates
To interpret the economic size of the estimated effects, we focus on in column 2 as our baseline results for two reasons. First, the time pe-
the results in column 2. At face value, our estimates imply that an in- riod before the expansion of Airbnb (i.e., ≤ 2012) might be too short to
crease in 100 Airbnb listings in a given neighbourhood translates to in- obtain robust estimates of BSA-specific time trends. Second, and most
creases of 3.5% in rents, 8.5% in transaction prices and 6.8% in posted importantly, the event-study exercises shown below indicate that the
prices. Given that the average increase in Airbnb activity in the period parallel trends assumption holds before 2013, suggesting that specifica-
2012–2016 is of 54 listings, our estimates imply an average increases of tions that fit neighborhood specific time trends are unnecessary.
1.89% in rents, 4.59% in transaction prices and 3.67% in posted prices. Finally, we perform a formal test á la Oster to assess the robust-
The large degree of heterogeneity in Airbnb activity across BSAs ness of the results to omitted variable bias. The method, developed by
implies that Airbnb has not affected all neighborhoods equally. In Oster (2019) and inspired by Altonji et al. (2005), analyzes how the in-
Fig. A.2 in the Appendix we illustrate these heterogeneous impacts by clusion of controls changes the coefficient of interest and the R-squared
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. 6. Event-study graph for rents and prices Notes: Outcome variables Rents, Transaction and Posted prices are expressed in logs. Event-study regressions following
Eq. 9, in which we interact year dummies with the level of Airbnb activity for 2016.

of the main regression. If including controls increases the predictive ca- are larger than one. This also suggests that gentrification is unlikely to
pacity of the model while not affecting the coefficient of interest, it is less explain the bulk of our effects.
likely that including unobservables would bias the results. One way to
assess this potential bias is to compute the relative importance of unob- 5.3. Alternative Airbnb measures
servables to observables (𝛿) that would be consistent with a coefficient
of interest equal to zero (𝛽 = 0). This is equivalent to asking how impor- In this subsection, we show that the results are robust to alterna-
tant the unobservables would need to be relative to the observables to tive measures of Airbnb activity. So far, our measure of Airbnb activity
eliminate the estimated effect. reflected contemporaneous activity. Each BSA-time cell is matched to
For our baseline specification (column 2 in Table 3), the 𝛿 that the number of Airbnb listings that received a review in that particular
matches 𝛽 = 0 amounts to 6.77 for rents, 1.41 for transaction prices and quarter. In column 2 of Table A.1 in the Appendix, we consider a specifi-
39.7 for posted prices.28 It means that the importance of unobservables cation in which Airbnb activity is measured over a longer time window.
would have to be 6.8, 1.4 and 39.7 times higher than that of the observ- Each BSA-time cell is matched to a moving average (MA) measure of
ables for the coefficients to be null. These high values occur because Airbnb activity that averages contemporaneous activity with that of the
controls have a very large explanatory power while their inclusion has previous three quarters (AbnbCount MA). The purpose of this measure
a small influence on our Airbnb coefficients. This exercise indicates that is to take into account Airbnb seasonality.
concerns about omitted variables bias are limited since the values of 𝛿 Although the BSAs are relatively similar in size, we compute a mea-
sure of Airbnb density by dividing the number of listings over the total
number of housing units (column 3). Finally, in column 4, we take the
log of the number of Airbnb listings to reproduce the log-log specifica-
28
We compare a constrained specification (with time dummies only) with an un-
tion used by Barron et al. (2020). The last row of Table A.1 provides the
constrained specification which also includes BSA fixed effects and all the time
varying controls at the BSA level (Xn,t ). As proposed by Oster (2019), we set average of each of the alternative measures of Airbnb activity to ease
𝑅2max = 1.3 × 𝑅2 . comparability across estimates.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. 7. Event-study graph for rents and prices using inverse distance to Sagrada Familia Notes: Outcome variables Rents, Transaction and Posted prices are expressed
in logs. Event-study regressions following Eq. 10, in which we interact year dummies with the inverse distance to Sagrada Familia.

Overall, our findings are robust to using alternative measures of lation and the number of households. We argue that, while a gentrifica-
Airbnb activity. Despite the underlying differences between the two tion process might reduce population and household size of the gentri-
studies, our results (reported in column 4) are similar in magnitude to fying neighborhood, as new incoming households are richer and have a
Barron et al. (2020) for the US. They find that a 1% increase in Airbnb lower average household size, it should not reduce the number of house-
listings increases housing rents and prices by 0.018% and 0.026%, holds in the neighborhood. Gentrifying processes revitalise neighbour-
respectively. Our estimates are a bit lower for rents (0.0098), while hoods by attracting more households.
Barron et al. (2020)’s estimate for housing prices is in between our es- In Table 4, we report the results of running specifications 1, 2, 3
timates for posted prices (0.017) and transaction prices (0.031). and 5 of Table 3 on the three different outcomes. The results of Panel A
indicate that Airbnb listings have a negative and strongly significant ef-
5.4. Mechanisms fect on the number of households across all four specifications. If we
focus on column 2, the estimates imply that 100 Airbnb listings de-
As explained in the theoretical model of Section 2, the impact of crease the number of households by 2.4%. On the contrary, Panels B
Airbnb on rents comes from the reduction of long-term rental supply and C show Airbnb’s negative effect on household size and population on
caused by owner choice to shift to short-term rentals. To provide di- three of the four specifications. We can decompose the effect of Airbnb
rect evidence of this mechanism, we would ideally look at the number on population into the number of households plus the effect on house-
of units rented to residents. Since this data is not available, we examine hold size.30 Then, we can compute how much of the reduction in pop-
instead the number of households, which includes both owner-occupiers ulation is due to a reduction in the number of households or due to a
and tenant households.29 We also assess the impact on population and lower average household size. In columns 1 to 3, the contribution of the
household size, where the latter is computed as the ratio between popu- number of households to the reduction in population is of 88%, 56%
and 50%, respectively. In column 4, Airbnb only has a significant effect
29
Alternatively, we could look at the number of signed rental agreements from official
records. However, this information is not provided at the BSA level but at the district level 30
The fact that 𝑝𝑜𝑝𝑛 = ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑𝑠𝑛 × (𝑝𝑜𝑝𝑛 ∕ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑𝑠𝑛 ), combined with outcomes are
and only starts in 2013. measured in logs allows for this decomposition.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Table 6
Impact of Airbnb on rents and prices: IV estimates Sagrada Familia.

Rents Transaction Prices Posted Prices

(1) (2) (3) (4) (5) (6)

Airbnb Ln(Rent) Airbnb Ln(Prices) Airbnb Ln(Prices)

Airbnb Count (x100) 0.095∗ ∗ 0.120∗ ∗ 0.101∗ ∗

(0.038) (0.052) (0.044)
Inv Dist SF × GoogleTrends 0.733∗ ∗ ∗ 0.452∗ ∗ ∗ 0.686∗ ∗ ∗
(0.197) (0.112) (0.193)
N 2.138 2.138 7.916 7.916 2.247 2.247
Time FE X X X X X X
BSA FE X X X X X X
Controls X X X X X X
F-stat of exc. inst. 13.8 16.3 12.5

Notes: Significance is indicated by ∗ p<0.1, ∗ ∗ p<0.05, and ∗ ∗ ∗ p<0.01. Standard errors, in parentheses, are clustered at the BSA level.
Each column represents a different specification. Panel A reports the results for rents, while Panels B and C report the corresponding
estimates for transaction and posted prices. Outcomes are average residuals at the BSA-time period level, as explained in the main
text. Regressions weighted with the total number of ads (for rents and posted prices) and of transactions (for prices). The analysis
takes place at the BSA-year level for rents and posted prices and BSA-quarter for transaction prices. Controls are: average age, log of
population density, average household size, unemployment rate, average relative income, and percentage of foreign residents.

on the number of households, suggesting again that Airbnb displaces According to Goldsmith-Pinkham et al. (2020), when discussing the
residents. exogeneity of a shift-share instrument, attention should be paid to the
Overall, these results strongly support the hypothesis that the ‘share’ component of the instrument. In our case, the main concern is
channel behind the impact of Airbnb on housing prices is a supply that BSAs that are close to tourist amenities could be experiencing dif-
reduction of long-term rentals. The results also lend credibility to the ferent trends in housing rents and/or prices for reasons unrelated to
hypothesis that the increases in housing rents and prices that we esti- Airbnb.
mate are caused by Airbnb activity and not by ongoing gentrification To provide some evidence of the exogeneity of our instrument, we
processes. run event-study regressions (as in Eq. 9) where we interact year dum-
mies with an indicator variable for BSAs in the top decile of the tourist
6. Results for alternative empirical strategies index distribution. This exercise attempts to verify whether BSAs that
are closer to relevant tourist amenities were experiencing a different
6.1. Instrumental variables results trend in rents and prices before the arrival of Airbnb. Panel (a) of
Fig. A.3 in the Appendix shows that, before 2013, pretrends in rents
In Table 5, we report first and second-stage results of the instrumen- were not statistically different between the two groups. In 2014, co-
tal variables approach described in Section 4.2. Columns 2 and 4 report inciding with the expansion of Airbnb, the difference becomes signifi-
the second-stage results for rents, transaction prices and posted prices, cant. The results are similar in Panels (b) and (c), hence lending cred-
respectively. The specification corresponds to Eq. 5, where Airbnb ac- ibility to the exogeneity hypothesis of our instrument. The results of
tivity is instrumented with the interaction between the cross-sectional the IV strategy provide a solid robustness test of the validity of our
tourist amenities index (Eq. 8) and the Google Trend searches. In instrumental variables results. The fact that coefficients remain fairly
columns 1 and 2, we control for BSA and time fixed effects as well as the similar and equally significant helps diminishing potential endogeneity
usual control variables; in columns 3 and 4, we also include the interac- concerns.
tion term between the control variables in 2012 (including distance to
the city center) and the regional GDP level.31 6.2. Event-study regression results
Columns 1 and 3 report the first-stage coefficients. To test the rele-
vance of the instrument we provide the F-test of excluded instruments In this subsection, we report the results of the event study regressions
which is well above 10, the standard rule of thumb accepted by prac- (Eq. 9 in Section 4.3). Fig. 6 plots the coefficients of the interaction
titioners (Angrist and Pischke, 2008). The instrument is not weak and terms between Airbnb activity in 2016 and the year dummies for rents
predicts well when and where Airbnb listings appear. Moving to the (a), transaction prices (b) and posted prices (c), where the coefficients
second-stage results, the coefficients remain positive and statistically in 2012 have been normalized to zero.
significant at the 1% significance level. In terms of magnitude, coef- The interaction terms between 2016 Airbnb activity (times 100) and
ficients in column 2 are remarkably similar to their OLS counterparts year dummies are statistically insignificant before 2013, while they are
of column 2 in Table 3, although admittedly the estimated coefficient positive and significant starting in 2014. This indicates that, at the be-
for transaction prices is larger (although not in a statistically significant ginning of the period, when the number of Airbnb listings was low, rents
sense). As for column 4, the inclusion of interaction terms with GDP and prices were not evolving differently in the BSAs that after 2013
increases the coefficient of rents while it decreases the coefficients of had high Airbnb activity. In contrast, between 2014 and 2017, when
prices, just like in Table 3. One concern regarding instrument validity Airbnb’s presence became important, neighborhoods where Airbnb ac-
might be that a large fraction of tourist amenities are located at the tivity was concentrated started to experience higher rents and price
city center. The fact that our results remain stable with the inclusion of growth. In a robustness test that is reported in Fig. A.4 in the Appendix,
distance to the city center times GDP provides evidence that our IV esti- we show that the results are robust to using a binary measure of Airbnb
mates are not biased by a shift in preferences for the city center among activity in 2016, where BSAs are classified into High Airbnb Areas and
residents during the recovery period. other areas.
The coefficients in Fig. 6 can be interpreted as follows: an increase
from zero to 100 listings in 2016 increases rents by 3.8%, transaction
31
These specifications correspond to columns 2 and 8 of Table 3.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

prices by 6.2% and posted prices by 5.6% in that year. These magnitudes Appendix A
are broadly in line with our baseline estimates of Table 3.

6.3. Sagrada Familia results

Table A.1
Finally, in this sub-section we report the Sagrada Familia results.
Impact of Airbnb on rents and prices: Robustness checks.
Table 6 displays the first and second-stage results of estimating Eq. 5 us-
ing the interaction between proximity to Sagrada Familia and the Google Baseline∗ AbnbCount MA∗ AbnbDens log AbnbCount
Trend searches as an instrument. (1) (2) (3) (4)

Columns 1, 3 and 5 report the first stage where our instrument Panel A Rents
is positively and significantly associated with Airbnb activity, which Airbnb 0.035∗ ∗ ∗ 0.029∗ ∗ ∗ 0.0068 0.0098∗ ∗ ∗
(0.009) (0.008) (0.005) (0.003)
suggests that the instrument is relevant. The F-statistics of excluded
N 2.123 2.123 2.123 2.123
instruments are high, which reinforces our claim. Columns 2, 4 and
6 report the second-stage results where the coefficients are positive, Panel B Transaction Prices
statistically significant and of a relatively higher magnitude than in Airbnb 0.085∗ ∗ ∗ 0.086∗ ∗ ∗ 0.030∗ ∗ ∗ 0.031∗ ∗ ∗
(0.016) (0.021) (0.005) (0.006)
our previous regressions, although not in a statistically significant way. N 7.916 7.916 7.916 7.916
Once again, the results of this exercise point in the same direction,
indicating that Airbnb activity had an impact on both rents and prices in Panel C Posted Prices
Airbnb 0.068∗ ∗ ∗ 0.070∗ ∗ ∗ 0.019∗ ∗ ∗ 0.017∗ ∗ ∗
Barcelona.
(0.009) (0.010) (0.004) (0.004)
Finally, Fig. 7 depicts the results of running an event-study regres- N 2.229 2.229 2.229 2.229
sion using proximity to Sagrada Familia as a predictor of Airbnb activity. Time FE X X X X
While the results are less conclusive in this exercise, we can nevertheless BSA FE X X X X
Controls X X X X
observe a similar trend than in the previous event-study where coeffi-
Mean 4Q2016 56 49 1.57% 1.76
cients become positive and significant in 2015. This is true especially
for rents and for posted prices. Yet again, this last strategy suggest that Notes: Significance is indicated by ∗ p<0.1, ∗ ∗ p<0.05, and ∗ ∗ ∗ p<0.01. Stan-
Airbnb had an inflationary effect on housing markets. dard errors, in parentheses, are clustered at the BSA level. Panel A reports the
results for rents, while Panels B and C report the corresponding estimates for
7. Concluding remarks transaction and posted prices. ∗ In the case of Baseline and AbnbCount MA, the
coefficients are multiplied by 100. Outcomes are average residuals at the BSA-
time period level. Regressions are weighted with the total number of ads (for
The rapid expansion of urban tourism and short-term rentals have
rents and posted prices) or transactions (for prices). The analysis takes place at
recently garnered much interest in public opinion and among policy- the BSA-year level for rents and posted prices and BSA-quarter level for prices.
makers, especially in large tourist cities. Concerns about the potential Controls are average age, the log of population density, average household size,
negative consequences of these phenomena have led local administra- unemployment rate, average income, and percentage of foreign residents.
tions to apply a wide range of regulatory measures.
To study how Airbnb affects the city’s housing markets, we examine
high-quality microdata on both rents and prices and combine these data
with information on the location of Airbnb activity within the city. We
apply several regression-based approaches that exploit the timing and
geography of the entry of Airbnb in the city to estimate the effects of
this platform on the city’s housing markets. The results show that Airbnb
activity in Barcelona has led to an increase both in rents and housing
prices, with larger effects for prices than for rents. Our preferred results
indicate that, for a neighborhood with the average Airbnb activity in
the city, rents have increased by 1.9%, while transaction prices have
increased by 4.6% and posted prices by 3.7%.
Although the effects on rents are not small, they cannot explain the
bulk of the high aggregate increases in rents that the city has experi-
enced between 2012 and 2016. In the most touristic parts of the city,
the effects of Airbnb are substantial. In neighborhoods in the top decile
of the Airbnb activity distribution, rents are estimated to have increased
by as much as 7%, while increases in transaction and posted prices are
as high as 17% and 14%, respectively.
Short-term rental platforms such as Airbnb might worsen the hous-
ing affordability problem in cities such as Barcelona, where tourism is
popular and the difference in profitability between renting long-term to
residents or short-term to tourists is high. Our findings can contribute to Fig. A.1. Airbnb Activity and Airbnb Prices. Notes: The graphs shows deciles
of BSAs with respect to their mean Airbnb Count for the fourth quarter of 2016
a more informed debate about the consequences of Airbnb and the desir-
ordered in the y axis. For each decile the mean log Airbnb nightly price for those
ability and design of policies that aim to limit the size of the short-term
active listings is shown in the x axis.
rental market.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. A.2. Implied effects of Airbnb across BSAs. Notes: These maps plot the implied impacts of Airbnb on rents and on transaction (posted) prices. For this we take
the results reported in column 2 of Table 3, and multiply it by Airbnb activity in each BSA at 2016.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. A.3. Instrument pretrend analysis. Notes: Outcome variables Rents, Transaction and Posted prices are expressed in logs. Event-study regressions where we
interact year dummies with an indicator variable for BSAs in the top decile of the tourist index distribution.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Fig. A.4. Event-study graph for rents and prices (top decile). Notes: This graph plots coefficient estimates (and confidence intervals) as in Eq. 9 but the continuous
measure of Airbnb activity has been replaced by a dummy variable for the top decile. Regressions are weighted with total number of ads (for rents) and of transactions
(for prices). Each point represents the difference in rents or prices between BSAs above the 90th percentile of Airbnb listings in 2016 compared to all other BSAs.

Supplementary material Baum-Snow, N., Hartley, D.A., 2020. Accounting for Central Neighborhood Change, 1980–
2010. Journal of Urban Economics. 37 (May), 103228.
Behrens, C., Boualam, B., Martin, J., Mayneris, F., 2018. Gentrification and pioneer busi-
Supplementary material associated with this article can be found, in nesses. CEPR Discussion Paper 13296.
the online version, at doi:10.1016/j.jue.2020.103278. Bertrand, M., Duflo, E., Mullainathan, S., 2004. How Much Should We Trust
Differences-In-Differences Estimates?∗ . Q. J. Econ. 119 (1), 249–275.
CRediT authorship contribution statement doi:10.1162/003355304772839588.
Calder-Wang, S., 2020. The Distributional Impact of the Sharing Economy on the Housing
Market. Harvard University, mimeo.
Miquel-Àngel Garcia-López: Conceptualization, Methodology, For- Chapelle, G., Eymeoud, J., 2018. Can big data increases our knowledge of the rental mar-
mal analysis, Writing - original draft. Jordi Jofre-Monseny: Concep- ket? Sci. Po Mimeo.
Couture, V., Handbury, J., 2020. Urban revival in America. J. Urban Econ. 37, 103267.
tualization, Methodology, Formal analysis, Writing - original draft.
Goldsmith-Pinkham, P., Sorkin, I., Swift, H., 2020. Bartik instruments: What, when, why,
Rodrigo Martínez-Mazza: Conceptualization, Methodology, Formal and how. Am. Econ. Rev. 110 (8), 2586–2624.
analysis, Writing - original draft. Mariona Segú: Conceptualization, González-Pampillón, N., Jofre-Monseny, J., Viladecans-Marsal, E., 2020. Can urban re-
Methodology, Formal analysis, Writing - original draft. newal policies reverse neighborhood ethnic dynamics? J. Econ. Geogr. 20 (2),
419–457.
References Kleven, H.J., Landais, C., Schultz, E., 2014. Migration and wage effects of taxing top
earners: evidence from the Foreigners’ tax scheme in denmark. Q. J. Econ. 129 (1),
333–378.
Almagro, M., Domínguez-Lino, T., 2020. Location Sorting and Endogeneous Amenities: Koster, H., van Ommeren, J., Volkhausen, N., 2020. Short-term Rentals and the Housing
Evidence from Amsterdam. NYU, mimeo. Market: Quasi-experimental Evidence from Airbnb in Los Angeles. CEPR Discussion
Altonji, J.G., Elder, T.E., Taber, C.R., 2005. An evaluation of instrumental variable strate- Paper 13094.
gies for estimating the effects of catholic schooling. J. Hum. Resour. 40 (4), 791–821. MasterCard, 2019. Global Destination Cities Index. Technical Report. MasterCard.
Angrist, J.D., Pischke, J.-S., 2008. Mostly Harmless Econometrics: An Empiricist’s Com- Monras, J., 2019. Minimum Wages and Spatial Equilibrium: Theory and Evidence. J. Labor
panion. Princeton University Press. Econ. 37 (3), 853–904.
Barron, K., Kung, E., Proserpio, D., 2020. The Effect of Home-Sharing on House Prices and Oster, E., 2019. Unobservable selection and coefficient stability: theory and evidence. J.
Rent: Evidence from Airbnb. Marketing Science forthcoming. Bus. Econ. Stat. 37 (2), 187–204.
Miquel-Àngel Garcia-López, J. Jofre-Monseny and R. Martínez-Mazza et al. Journal of Urban Economics 119 (2020) 103278

Sheppard, S., Udell, A., 2016. Do Airbnb Properties Affect House Prices. Williams College WTO, 2018. UNWTO Tourism Highlights. 2018 Edition. Technical Report. United Nations
Department of Economics Working Papers 2016-03. World Tourism Organization.
Wolfers, J., 2006. Did unilateral divorce laws raise divorce rates? a reconciliation and new Zervas, G., Proserpio, D., Byers, J.W., 2017. The rise of the sharing economy: estimating
results. Am. Econ. Rev. 96 (5), 1802–1820. the impact of airbnb on the hotel industry. J. Market. Res. 54 (5), 687–705.