Content uploaded by David Neeser
Author content
All content in this area was uploaded by David Neeser on Sep 25, 2015
Content may be subject to copyright.
Máster Universitario en Análisis Económico
2014-2015
Trabajo Fin de Máster
“Does Airbnb Hurt Hotel Business:
Evidence from the Nordic Countries”
Dâvid Neeser
Tutores
Professor Martin Peitz
Professor Jan Stuhler
Madrid 10 de septiembre 2015
Does Airbnb Hurt Hotel Business: Evidence from
the Nordic Countries
Dˆavid Neeser
September 25, 2015
**A special thanks to my supervisors Martin Peitz and Jan Stuhler for their
useful comments, remarks, and time through the learning process of this master
thesis. I also want to thank my colleagues for their marginal input and their
moral support. I especially want to express my gratitude to Morgane Laouenan
without whom I would not have any data to work with. Finally, I want to thank
Morgan Freeman for his constant source of inspiration.
Abstract: This paper wants to measure the impact of Airbnb on hotel rev-
enue in Norway, Finland, and Sweden using a difference-in-differences strategy
with many time periods and different level of treatment. We exploit the richness
of our data to differentiate among Airbnb listings and to identify which type of
hotel costumers Airbnb is more likely to attract.
1 Introduction
In this paper, we want to measure the impact of Airbnb in Norway, Finland,
and Sweden on the hotel industry. More specifically, we look at the revenue
per available hotel room, the price of hotel rooms, the room occupancy rate,
and the composition of hotel guests such as country of residence and purpose
of visit. The main reason why we chose these countries is that we can pool
them together since they should be similarly affected by common shocks. Also,
the data on the accommodation industry was freely available and comparable.
Another reason is that many businesses (photographers, house cleaners, pricing
tools, market analysts, etc.) gravitate around Airbnb in the U.S. and in many
European countries, but not in the Nordic countries yet, at the best of our
knowledge. This is important in our analysis as we only observe the price of
a listing once, and the absence of pricing tools should make it more stable
through time. These countries are also beyond the scope of most studies in this
literature, maybe because they do not try to stop this platform from entering
their country or city, unlike others. Finally, the three countries witnessed a
significant entry of Airbnb. Figure 1 shows the relative size of all listings per
region in May 2015 compared to the average number of room per hotel in the
same region. To give a better idea, figure 2 displays the market share of Airbnb
in each region, also as of May 2015.
1
Peer-to-peer platforms such as Airbnb and Uber are targets to controversy
with regulators and incumbents trying to respond to these new entrants. World-
wide, blogs and editorials attack or defend these business models with arguments
mostly based on theory, without empirical backup. This is an important issue
as many cities are doing everything in their power to block them. For instance,
Uber is completely banned in the state of Nevada and also in India (although it
keeps operating in some cities), and under partial banned in Brussels, Germany,
and the Netherlands for their lower-cost services such as UberPOP and UberX
which employs non-taxi drivers1. Similarly, Paris is performing raids in Airbnb
listings that are suspected to be illegal (cannot be rented out for a short period
of time if owners are not present during their stay). Furthermore, the city of
Paris and San Fransisco have come up with some estimates that Airbnb is con-
tributing to the housing shortage and price increases by removing apartments
from the housing market23.
In any case, truly objective databased research on the topic is still lacking
in the literature as these platforms (mostly privately owned) keep their data for
their in-house publications4. While they have the potential to achieve very high
quality research thanks to the richness of the data they have access to, they are
also likely to bias the results in their own interest.
The reader will find a very complete and recent literature review in Zervas et
al. (2014) covering works on multi-sided platforms, substitution between online
and offline markets, and external shocks on tourism and hospitality industry,
which are all the fields we are contributing to. More recently, a paper by M¨uller
(2014) wants to model the impact of the sharing systems on the social welfare,
but his work is still in progress. Another research in progress from Gutt and
Herrmann (2015) empirically measures Airbnb hosts’ reaction to new reviews on
their pricing behavior using a DD with matching in New-York city. They find
that hosts increase their price by an average of 2.69 euros more than hosts with
comparable listings that has not been rated yet, which is in line with Querbes
(2014)’s simulated model on peer-to-peer platforms interaction.
This paper is closely related to Zervas et al. (2014) where they estimate the
impact of Airbnb’s entry on the hotel industry in Texas. They use a difference-
in-differences (DD) approach with very detailed data on hotels and Airbnb list-
ings at the individual level in Texas’ biggest cities. They find that where Airbnb
entered the most, the impact on the more vulnerable hotels’ revenue is about
8-10% decrease in five years. They manage to differentiate the impact for dif-
ferent market segments, where cheaper independent hotels are competing more
fiercely with this new platform, while hotels focusing more on business travelers
and wealthier customers are reaching a different niche than Airbnb’s typical
1http://www.businessinsider.com/heres-everywhere-uber-is-banned-around-the-world-
2015-4
2http://roadwarriorvoices.com/2015/05/26/paris-officials-go-on-door-to-door-raid-of-
illegal-airbnb-rentals/
3http://www.businessinsider.com/san-francisco-report-blames-airbnb-for-housing-
shortage-airbnb-strikes-back-2015-5
4For instance, see: https://www.airbnb.com/economic-impact/
2
guests.
We contribute to their research by addressing one of the limitations men-
tioned in their paper, namely the specificity of Texas’ accommodation industry.
We closely replicate their methodology to see how their results may change
when we look at countries like Sweden, Norway, and Finland. While we do not
observe hotels individually, we exploit Airbnb’s information details to analyze
the impact of different types of listings. We go one step further by looking at
how the demand for hotels has changed with Airbnb’s market entry in differ-
ent regions using data on hotels’ guests. We will also point out some forces
and weaknesses from their paper that are more easily seen by someone who has
worked with this kind of data.
This paper is organized as follow: section 2 explains the data we used, then
section 3 presents the model, the results are shown in section 4, section 5 presents
our extensions, and section 6 concludes.
2 Data
2.1 Hotel
Data on accommodation was found on Statistics Finland5, Statistics Sweden6,
and Statistics Norway7.
We got monthly data on hotels from January 2004 until May 2015 for Norway
and Finland, and from January 2008 until May 2015 for Sweden. We thus
cover a large period before Airbnb first enters that market (which happened
between August 2008 and April 2009 depending on the country). The lowest
level of aggregation available for all countries was NUTS 3 which corresponds
to 20 regions for Finland, 21 counties for Sweden, and 19 counties for Norway.
Throughout this paper, the words ”region”, ”county”, and ”NUTS 3” will be
used as synonyms. Some data were available at the municipality level but
would have been less accurate and also censored when too few hotels operate in
the municipality. Aland and Svalbard are excluded from the dataset for many
missing observations.
For all counties, we observe the number of establishments, the price of a room
and the occupancy rate of the rooms from which we compute the widely used
revenue per available room (RevPAR) which is simply the product of the last
two. We convert all prices in US dollar to fit Airbnb data. We also observe the
number of rooms for all types of accommodation, as well as the number of nights
spent in hotels per country of residence. Since the countries of origin were not
homogeneously reported across the three databases, and the strict majority of
guests are from the hosting country, we categorize guests from Finland, Sweden,
Norway, Denmark, Russia and the rest of the world.
5http://pxnet2.stat.fi/PXWeb/pxweb/en/StatFin/StatFin lii matk/?tablelist=true
6http://www.statistikdatabasen.scb.se/pxweb/en/ssd/
7https://www.ssb.no/en/transport-og-reiseliv/statistikker/overnatting
3
Norway’s database also includes data on guest nights per purpose of accom-
modation, namely course or conference, occupation, and holiday or recreation.
Sweden provides similar data where they report the number of occupied rooms
per category of guests which are either conference, business, group, or leisure.
An obvious drawback is the fact that the treatment is assumed to affect
all hotels within the same region independently of the distance they are from
Airbnb listings. This is a strong assumption as some regions such as Lapland
cover a large land area from north to south with few inhabitants8. This is why
we also include the population density for each NUTS 3.
Data on population but especially unemployment was difficult to get at the
county level for all countries. For population, we had to extrapolate between
the quarters/years. For Sweden, we observe either monthly unemployment at
the country level, or yearly unemployment at the NUTS 2 level. We decided
to use the first one as it would have been far-fetched to extrapolate yearly
unemployment in monthly observations.
Finally, we also include the variable GDP per capita that we observe at
the country level, since most of hotel guests (between 70 to 90%) are national
tourists. Table 1 reports summary statistics for the main variables.
2.2 Airbnb
Data on Airbnb was very difficult to get and took a lot of time to gather. Al-
though a high level of detail is available for each listing, not all features are
homogeneously reported. This is why we only kept the variables that could
categorize every listings. We thus have the price of a night for each listing, the
type of rental (entire home/apartment, private room, and shared room), the
type of property (house, apartment, cottage, island, etc.), and type of cancella-
tion policy, as we trust these characteristics influence the substitutability with
an hotel room, and thus should have a different impact.
We now describe each characteristic in more details. Listings can be split
in three categories called ”room type” which describes more or less the level of
intimacy of the place: Entire home/apartment, private room, and shared room.
The first two are more likely to attract customers from hotels while the last one
features more of a hostel dorm. A second distinction can be made in the research
filter according to the ”property type”. While most listings are either a house,
an apartment, or a bed & breakfast, you can also find boats, planes, castles,
islands, and many more. Another important aspect of a listing to consider is
the cancellation rule that the host chooses. The main three different options
are ”flexible”, ”moderate”, and ”strict”9. They differ in how long in advance
you must cancel, and how much it will refund you. Important to mention here
is that the additional percentage directly paid to Airbnb cannot be reimbursed.
This is why we consider all these factors in this paper.
8Data at the hotel level is available at a cost from the STR census including most hotels
in these countries.
9See https://www.airbnb.com/home/cancellation policies for more details about cancella-
tion policies
4
Additional information that we have for each listing is reviews from previ-
ous guests, reviews of other listings belonging to the same host if he has more
than one, potential number of guests, and the creation date of the host’s profile.
Remaining information, although interesting, could not be included in this pa-
per as they were much harder to scrape and not available for all listings. This
includes different price schemes (weekly fee, monthly fee), extra fee per addi-
tional guest, different amenities, type of bed, number of bathrooms, minimal
stay, cleaning fee, security deposit, etc.
The paper from Zervas et al. (2014) also obtained a cross-section of all
accommodations listed on Airbnb at the time they scraped the website. They
then proxy the date of entry with the date the user became member of Airbnb.
They mention that one must assume that users with many listings have put
them all online the month they went on Airbnb. More importantly, they do not
observe listings that exited the market. Although this is usually an important
issue as market exit is endogenous, it appears to be less important when market
entry is exponential as they argue Airbnb is. When we scraped the data from
Airbnb, it first took several days just to obtain the url of all available listings. By
the time we used this list to get all information from each listings, many where
no longer available already. This means that exit might be more important than
expected, at least in the countries we are interested in where rural depopulation
could explain part of this phenomenon10.
3 The Model
Following Zervas et al. (2014), our analysis takes advantage of the fact that
Airbnb entered these regions at different points in time, and also grew differently
depending on the geographical area. In figure 4, one can see the different Airbnb
entry patterns in the most populated regions of our sample. This high variability
allows us to look at how hotel room revenue evolved differently where Airbnb
was more present, and thus identify the impact of Airbnb on hotel revenue.
Our main specification for the DD identification strategy looks like this:
logRev P ARit =βlogAirbnbit +δi+τt+X0
itγ+it (1)
where log RevPAR is the log of the average revenue per available room for all
hotel in region iat time t. Log Airbnb is the total number of Airbnb listings in
log, δis a region fixed effect, and τis a month-year time dummy. In Xwe include
all region-specific observables that change with time and could be correlated
with both log RevPAR and log Airbnb such as unemployment rate, population,
GDP per capita, and a region-specific time trend (linear or quadratic)11 . This
allows us to account for unobservables that evolves differently across counties
but follow a structural trend.
10See Martti Lujanen (2004)
11Following Zervas et al. (2014), we use a quadratic region-specific time trend. They men-
tion that the use of such a trend could be problematic without many pre-treatment periods.
5
The main advantage of our approach is that we do not have to worry about
unobservables simultaneously affecting hotel revenues and Airbnb entry that
does not change with time, or change equally across all regions. Indeed, the
first difference uses the region fixed effects to account for time-invariant differ-
ences in average hotel revenues between treated and non-treated counties (i.e.
counties with and counties without Airbnb respectively). In a second time, the
year-month fixed effects account for revenue differences variations that are com-
mon in all regions. For instance, awareness to Airbnb should increase similarly
across the three countries and is thus accounted for. In the same way, if one
region is more attractive to tourists because of its access to the sea or the moun-
tains, this is also implicitly included in our model as it remains constant for the
observed period. The coefficient βis the one we are interested in as it estimates
the percentage change in hotel room revenue in treated regions after treatment,
against the control group. We will consider later the remaining possible endo-
geneity between hotel revenue and Airbnb supply coming from temporary local
demand shocks.
We decided to weigh each region according to its population density for all
specifications. The reason behind this is that some counties such as Oslo are
one big city, while others like Lapland cover a huge but sparsely populated area.
Therefore, if Airbnb enters in Oslo, it should impact a greater proportion of the
hotels in the county than would the same happen in Lapland, as greater distance
between hotels decreases competition12. This is why we want regions densely
populated to weigh more in the regression. All standard errors are robust to
heteroskedasticity and autocorrelation within regions.
4 The Results
Table 3 compares the results for the different controls. Column 1 is the naive
model where we only include region and month-year fixed effects. We hence as-
sume that everything affecting hotel revenue and Airbnb supply is either fixed
in time, or varies equally in all regions. This is clearly a bold assumption since
the coefficient of interest loses its significance when we introduce observables
varying differently across regions or countries such as the log of population, the
unemployment rate or the log of GDP per capita. We then want to control for
unobservables that evolve following some trend that can differ from a region to
another. This is often done in DD papers by including a linear or a quadratic
region-specific time trend13 . Column 3 and 4 reports the results with a lin-
ear and a quadratic region-specific time trend respectively. Whilst βhas the
expected negative sign in all our specifications, it is only significant when we
abstract from any additional control. In Zervas et al. (2014), they only report
estimates with the quadratic city-specific trend since their results were appar-
ently not really affected by this choice. The impact they find is three times
higher than the one we have in column 4, and is significant at the 1% level. In
12See Balaguer & Pernias (2013)
13See Zervas et al. (2014) p.4
6
what follows, we explore different specifications to see if our coefficient remains
insignificant.
Because we observe Sweden for a shorter pre-treatment period, we have an
unbalanced panel data. Column 2 in table 4 uses the same specification that the
last column of table 3 (which we reproduced in column 1 of table 4 for easier
comparison) but we excluded Sweden from the sample to obtain a balanced
panel data. One can see that the βcoefficient becomes more negative (although
still not significant). This could indicate that Airbnb’s impact is heterogeneous
across the countries which casts doubt on the relevance of pooling them. It is
true that Sweden differs from its neighbors with respect to its housing market14.
In its most important agglomerations, regulation to keep prices for housing low
is such that demand is not fully supplied and part of Airbnb listings might just
be using this to rent their place for a much higher price, and should have no
impact on hotels. If this is truly happening, the observed increase in magnitude
of βwhen excluding Sweden should be more important when we weigh regions
with respect to their population density than when we do not. To give the
intuition behind this, remember that the shortage of housing in Sweden is only
present in its most important cities, thus in its most densely populated counties.
Hence, when we give more weight to these regions, including Sweden in the
sample should logically push the coefficient more toward zero since Airbnb would
compete relatively much less under this assumption. To test this hypothesis,
we report in column 3 and 4 the equivalent of column 1 and 2 respectively
when we give the same weight to every regions in the sample. To put it more
clearly, under the null hypothesis that Sweden is comparable to Finland and
Norway, |(1) −(2)|≯|(3) −(4)|, where (x) designs table 4’s column x. Since
0.007 <0.01062, we cannot reject the null hypothesis and must find another
explanation for what we observe15
What we believe is that this effect comes from the short pre-treatment period
available for Sweden. Figure 3 shows the evolution of log hotel revenue per
available room for all counties (seasonally adjusted). As we can see, hotels
have experienced an important drop in their revenue between 2008 and 2010,
which coincides with the whole pre-treatment periods for Swedish counties where
Airbnb entered the earliest. Even though we control for the GDP per capita, it
is likely that we do not completely overcome this issue. Provided that this is the
only thing that causes our estimates to differ, we can take the results from the
subsample and extend them to Sweden. From this point we will always report
the results with and without Sweden16
14See Martti Lujanen 2004
15The conclusion of the test is robust to different controls, although not reported here.
16One thing to consider though is that we always cluster at the region level so that removing
Sweden leaves us with only 38 regions. This could be considered as a few clusters problem
with the possibility that we underestimate the standard errors. Since our standard errors do
not get smaller, we do not address this potential issue. See Cameron and Miller (2015) for
more on clusters.
7
Adressing endogeneity
As we mentioned earlier, our model does not account for temporal local demand
shocks. In Zervas et al. (2014), they claim to control for that by including
the log of hotel rooms supply in a city excluding hotel i. They explain that
a demand shock anticipated by both hotels and Airbnb could cause them to
change their capacity in the same direction, increasing the impact on hotel i’s
revenue. We cannot replicate this strategy in our work since we do not have
observations at the hotel level. On the other hand, we also have our doubts
on the effectiveness of their method for two reasons. First of all, this assumes
that hotels use quantity rather than price to adjust for short-term demand
fluctuations. Intuitively, unless they close during low seasons, hotels should be
more likely to adjust their price or their staff instead of closing some rooms.
Second, if Airbnb has an effect on hotel rooms supply, then the model will not
assess the net impact of Airbnb on hotel revenue, but only the direct one.
Instead, we try to tackle the problem by reducing the frequency of observa-
tions to quarterly or biannual data. That way we should be able to ignore local
demand shocks that are temporary enough and thus solve our main endogene-
ity concerns. The results are reported in table 5 where the first two columns
report the previous estimations with monthly data from before for comparison,
and even column numbers exclude Sweden from the sample. In column 3 and
4, we used quarterly data for the analysis, and biannual data for the last two
columns. We still do not find any significant impact, so that if this successfully
solves the endogeneity issue, we must conclude that Airbnb did not significantly
affect hotel revenue per available room since it first arrived in these countries.
The revenue per available room is the product of the price of a room and the
occupancy rate of the rooms. We thus repeat the same exercise with occupancy
rate and the log of price as independent variable. Table 6 reports the estimates
for the log of the price of a room. First thing to notice is that the impact becomes
significant at the 5% level when we exclude Sweden. The impact also increases in
column 3 and 4 as we use quarterly observations, pointing toward the possibility
that this solves the endogeneity problem. The estimate even becomes slightly
significant in the full sample when using biannual data (column 5), although
column 6 shows a lower impact than column 417. Estimates for occupancy rate
are not reported since the coefficients never became significant (although always
negative as anticipated).
These results let us believe that hotels tend to adjust to the market using
the price rather than the quantity. Hotels might prefer to reduce their price to
maintain a certain level of occupation in order to cover costs that are fixed in
the short run. The estimate tells us that whenever Airbnb listings increase by
10%, hotel room monthly revenue decrease by 0.111%18 (to be conservative) in
Finland or Norway, which we could extrapolate to Sweden.
We also explore the possibility that the impact from Airbnb’s entry may take
17One must be careful in interpreting the estimates from column 6 since the number of
observations is much lower.
181−(1.1−0.0117) = 0.001114
8
some time to be felt. Intuitively, most people traveling plan their trip in advance
so that listings entering the market at the time of traveling do not impact their
choice of accommodation. We compare both specifications in table 7. βis
indeed of higher magnitude when we take one month lagged value of Airbnb
listings, but only for the non significant coefficient, while it loses significance in
column 6. The results are thus inconclusive.
As opposed to Zervas et al. (2014) we did not find that Airbnb had a
significant impact on hotel revenue per available room when extending their
methodology to the Nordic countries. We present here some possible reasons
why our results differ from their analysis.
First of all, they individually observed all hotels while we only had the
average values at the region level. This difference has three major implications.
First, we measure the impact of Airbnb supply on a much greater area while
they get a much more local effect which is likely to be of greater magnitude.
Second, they can focus on big cities while we must also include rural areas in our
sample in which hotels compete differently19. Finally, the impact we measure
could still be underestimated because we do not observe hotels individually. For
instance, if Airbnb has a strong impact on the lower cost hotels to the point that
they put them out of business, the net impact on the average price or revenue
would be upward biased.
Alternatively, Texas and our countries of interest are different in many ways
which we cannot account for. For instance, maybe hosts join Airbnb in Texas
because they need it to pay their rent, whilst in Norway or Finland they do it
to accommodate tourists. In other words, it could be the case that people use
Airbnb to absorb individual shocks when the government provides less insur-
ance (people in Texas receive less benefits from the State when they become
unemployed than people in the Nordic countries). We would thus be faced with
more endogeneity in our sample, hence underestimating Airbnb’s impact. A last
hypothesis that we cannot test is whether most hotels in the Nordic countries
are only a front to bleach money from illegal activities. If that would be the
case, revenue per hotel room should not correlate with Airbnb supply as they
do not compete against each others, which would explain our results.
5 Some Extensions
Differentiating Listings
Now that we have a good idea of the average impact of any additional listings
on hotel revenue, we can dig deeper to see how listing’s characteristics matter.
As they advertise it themselves, Airbnb lists all sort of accommodations, from a
house or an apartment to some more extravagant ones such as a tipi, a boat, a
lighthouse, or even an island. As we mentioned earlier, all these characteristics
should influence how closely they can replace a hotel room, and thus have a
different impact on their revenue. To verify this, we regressed the log of Airbnb
19See Balaguer & Pernias (2013)
9
supply on revenue, occupancy rate, and hotel room price, excluding certain
types of listings. For instance, we first took out listings categorized as ”shared
room”. One can argue that people searching for a shared room are looking
for a different experience than what hotels typically provide. As expected, the
coefficient does increase from the original settings.
We then kept only listings that are not categorized as tent (including yurts
and igloos) or as a mean of transportation like boats or RV. We also exclude
”others” as it was hard to find out what it was without looking manually into
each of them. Surprisingly, found a lesser impact than without that filter. Ap-
parently, at least in these countries, these sort of ”into the wild” accommodation
do compete with hotels somehow.
The last feature we use to differentiate listings is the cancellation policy.
This is important as uncertainty is always part of a trip, and most hotels on
booking sites have relatively flexible cancellation rules. On Airbnb, flexible and
moderate cancellation rules are comparable with standard hotels, while all rules
beyond that should make the listing less substitutable. This is why we try and
exclude these from the sample and find slightly higher β. To avoid too much
redundancy, we do not report each of these specifications, but these results show
that shared rooms and strict cancellation policies capture less costumers than
other types of listings. To show it more explicitly, we have excluded both types
from table 8, as well as Sweden otherwise no coefficient would be significant.
Now a 10% increase in this type of listing reduces the price of a hotel room by
0.12% in Norway and Finland.
Country of origin
An interesting question that will be addressed here is what kind of hotel guests
Airbnb is capturing. For the sake of data availability, we restrict our analysis
on the country of residence and the category of the guests (business, conference,
holiday). Table 1 shows that the majority of hotel guests (80% in average)
come from the same country than the hotel is located in, which we will call
national guests. These guests should be relatively less likely to use Airbnb since
its main appeal, besides its relatively low price, is the ”unique experience to live
like a local” which should be more attractive to outsiders. We should therefore
expect to see a positive impact of Airbnb supply on national guests’ share of
total hotel guests. Using the same model than equation 1 but substituting log of
RevPAR by share of national guests, we find in table 9 what we were expecting.
In column 1, we look at the impact on national guests’ share. Column 2 takes
the proportion of guests coming from Sweden, Finland, and Norway (Nordic
guests), and column 3 adds Denmark and Russia in the share since they are
neighbors (Nordic guests + neighbors). It is interesting to compare the three
columns to find out that not only locals are relatively less interested in using
Airbnb when traveling within their own country, but also within those three
countries as the coefficient becomes slightly more positive for Nordic guests.
Furthermore, including Danish and Russians reduces the impact, implying that
they substitute hotels with Airbnb relatively more that Norwegians, Finnish,
10
and Swedish. The absolute magnitude of the coefficient is of little interest since
it tells nothing about the total number of guests.
GDP per capita is obviously positive and highly significant since a strong
majority of hotel clients are national guests, and their ability to travel depends
a lot on their income.
Category of guests
We repeat the same procedure but differentiating guests with respect to the
purpose of accommodation. Table 10 regresses Airbnb supply on the fraction of
guests sleeping in an hotel for leisure, for a conference or a course, and for work
related purpose in column 1, 2, and 3 respectively. Although none of them is
significant at the 5% level, the sign does coincide with Zervas et al. (2014)’s
results where they find that Airbnb affected more hotels without a meeting
space, meaning that people traveling for work are less likely to turn to Airbnb
for their accommodation, at least at the time of this study. Indeed, Airbnb
has lunched last year a campaign targeting business travelers20 which should
increase further their impact on hotels.
6 Conclusion
The purpose of this paper was to measure the impact of Airbnb on the hotel
industry in Norway, Finland, and Sweden. The motivation for this was to
empirically contribute to the actual heated debate taking place worldwide by
looking at a different market than those more typically covered. We found that
Airbnb did not significantly affect hotel’s revenue per available room in average,
but did contribute to a reduction in the average price of a room where Airbnb
entered the most. We also found evidence that Airbnb’s ”cultural experience”
makes it relatively more attractive for foreigners than locals.
The long term impact of peer-to-peer platforms is uncertain as companies
such as Airbnb are receiving a lot of pressure from people who invested huge
amounts of money in it and are expecting the same high returns they have
achieved before due to their exponential growth worldwide. They have recently
started to reposition themselves toward a more standardized experience in term
of quality and have hired Chip Conley (founder of Joie de Vivre Hospitality) as
the Head of Global Hospitality21.
One aspect concerning Airbnb that we have briefly mentioned in our paper
but have not given the appropriate attention is the impact that Airbnb has on
housing markets. According to a report by the city of San Fransisco, Airbnb
could be taking about 40% of potential rentals off the market22. Similar worries
20See http://techcrunch.com/2014/07/28/airbnb-concur/
21https://www.airbnb.com/press/news/airbnb-names-chip-conley-as-head-of-global-
hospitality
22http://www.businessinsider.com/san-francisco-report-blames-airbnb-for-housing-
shortage-airbnb-strikes-back-2015-5
11
have begun to emerge in other big cities around the world such as New-York,
Vancouver, and Berlin. They fear that Airbnb reduces supply for long-term
accommodation, which increases the price, or put people on the street where
prices are regulated. Although some studies have tried to empirically verify
these claims, they either lack of good data or objectivity. We believe that future
research should focus on that aspect if we want to get a better idea about this
type of peer-to-peer platform’s impact on total welfare.
References
[1] A. Colin Cameron and Douglas L. Miller (Spring 2015), ”A Practitioner’s
Guide to Cluster-Robust Inference”, Journal of Human Resources, Vol.50,
No.
[2] Ascolli, A., Scamoni, C., de Lezo, A., de Lezo, S., Lefebre, F., Mckenna,
C., Heimz, R.-R. (2007). CMS Guide to Principle Hotel-Specific Laws in
Europe.
[3] Balaguer, J., & Pernias, J. C. (2013). Relationship Between Spatial Ag-
glomeration and Hotel Prices. Evidence from Business and Tourism Con-
sumers. Tourism Management. Vol 36, 391-400.
[4] Carb´o Valverde, Santiago, Sujit Chakravorti, Francisco Rodriguez Fernan-
dez. (2010). Regulating Two-Sided Markets: An Empirical Investigation.
Tech. rep., Working Paper, Federal Reserve Bank of Chicago.
[5] Cheng, Denise. (2014) Is sharing really caring? A nuanced introduction to
the peer economy. MIT Center for Civic Media
[6] Donald, Stephen G, Kevin Lang. (2007). Inference with Difference-in-
Differences and Other Panel Data. The Review of Economics and Statistics
89(2) 221-233.
[7] Fremstad, A. (2014). Gains From Sharing: Sticky Norms, Endogenous Pref-
erences and the Economics of Shareable Goods. Working Paper, University
of Massachusetts Amherst.
[8] Gutt and Herrmann (2015) Sharing means caring? Hosts’ price reactions
to rating visibility.
[9] Hagiu, A., & Wright, J. (2015). Multi-Sided Platforms. International Jour-
nal of Industrial Organization.
[10] Kroft, K., & Pope, K. G. (2012). Does Online Search Crowd Out Tradi-
tional Search and Improve Matching Efficiency? Evidence from Craigslist.
[11] Lewis, G., & Wang, A. (2013). Who benefits from improved search in plat-
form markets?
12
[12] Martti Lujanen. (2004). Housing and Housing Policy in the Nordic Coun-
tries. Nordic Council of Ministers, - 328 pages
[13] M¨uller, M. P. (2014). An economic analysis of online sharing systems’ im-
plication on social welfare. University of Innsbruck. (Research in progress)
[14] Querbes, Adrien. (2014). Banned from the Sharing Economy: An Agent-
based Model of the Peer-to-Peer Distribution of Consumer Goods. Depart-
ment of Social and Decision Sciences, Carnegie Mellon University.
[15] Seamans, R., & Zhu, F. (2013). Responses to Entry in Multi-Sided Markets:
The Impact of Craigslist on Local Newspapers.
[16] Zervas, G., Proserpio, D., & Byers, J. (2014). The rise of the sharing econ-
omy: Estimating the impact of Airbnb on the hotel industry. Boston U.
School of Management Research Paper, (2013-16).
[17] Zervas, Georgios and Proserpio, Davide and Byers, John. (January 28,
2015). A First Look at Online Reputation on Airbnb, Where Every Stay is
Above Average. Available at SSRN: http://ssrn.com/abstract=2554500
13
7 Appendix
Table 1: Principal summary statistics
Variable Mean Std. Dev. Min. Max. N
Log RevPAR 3.642 0.303 2.442 4.643 6959
Occupancy rate of room 0.47 0.112 0.158 0.891 6959
Log of room price 4.426 0.12 3.996 4.942 6959
National guests 0.803 0.122 0.204 0.982 6977
Nordic guests 0.836 0.115 0.23 0.985 6977
Nordic guests + neighbors 0.876 0.097 0.267 0.989 6977
Holiday travelers 0.41 0.173 0.101 0.944 4493
Table 2: Number of Airbnb listings depending on different search filters
Filter Number Percentage
No filter 16688 100.00
No shared room 16417 98.38
No camping & others 16487 98.80
No strict cancellation 15618 93.59
No shared room & no strict cancellation 15367 92.08
14
Figure 1: Comparison of number of rooms per hotel and total number of Airbnb
listings as of May 2015, by county
15
Figure 2: Market share of Airbnb measured by available rooms as of May 2015,
by county
16
Figure 3: Evolution of log revenue per available hotel room (RevPAR), season-
ally adjusted and biannual observations
Figure 4: Evolution of number of Airbnb listings relative to the number of hotel
rooms before Airbnb entry for the 10 most populated regions
17
Table 3: DD estimates of the impact of Airbnb on hotel room revenue, with
different controls.
(1) (2) (3) (4)
Revenue Revenue Revenue Revenue
log Airbnb supply -0.0138∗∗ -0.00384 -0.00569 -0.0102
(0.00579) (0.00657) (0.0101) (0.0110)
log population -0.581∗5.080∗∗ 2.680
(0.293) (2.013) (3.074)
unemployment -0.00861∗∗ -0.00292 -0.00253
(0.00339) (0.00307) (0.00249)
log GDP per capita 0.152 -0.0775 0.0105
(0.285) (0.281) (0.326)
region-specific trend No No linear quadratic
Observations 6959 6959 6959 6959
Adjusted R20.396 0.400 0.429 0.440
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Figure 5: Evolution of log RevPAR for regions with low Airbnb presence com-
pared to regions with high Airbnb presence, seasonally-adjusted biannual ob-
servations
18
Table 4: DD estimates of the impact of Airbnb on hotel room revenue, weighed
and unweighed regressions.
weights no weights
(1) (2) (3) (4)
Revenue Revenue Revenue Revenue
log Airbnb supply -0.0102 -0.0168 -0.00651 -0.0164
(0.0110) (0.0112) (0.00995) (0.0109)
log population 2.680 -2.393 4.729∗-0.500
(3.074) (3.068) (2.692) (2.768)
unemployment -0.00253 -0.00298 -0.00357 -0.00469
(0.00249) (0.00283) (0.00312) (0.00353)
log GDP per capita 0.0105 -0.364 -0.301 -0.981
(0.326) (0.588) (0.499) (0.824)
Sweden Yes No Yes No
Observations 6959 5069 6959 5069
Adjusted R20.440 0.476 0.357 0.397
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
19
Table 5: DD estimates of the impact of Airbnb on hotel room revenue at different
frequency level.
Monthly Quarterly Half yearly
(1) (2) (3) (4) (5) (6)
Revenue Revenue Revenue Revenue Revenue Revenue
log Airbnb supply -0.0102 -0.0168 -0.0135 -0.0203 -0.0124 -0.0141
(0.0110) (0.0112) (0.0116) (0.0124) (0.0106) (0.0107)
log population 2.680 -2.393 2.911 -1.806 2.517 -2.293
(3.074) (3.068) (3.131) (3.142) (2.926) (3.089)
unemployment -0.00253 -0.00298 -0.00312 -0.00384 -0.00103 -0.00124
(0.00249) (0.00283) (0.00261) (0.00290) (0.00260) (0.00280)
log GDP per capita 0.0105 -0.364 0.0396 -0.407 0.0419 -0.0541
(0.326) (0.588) (0.278) (0.545) (0.231) (0.370)
Sweden Yes No Yes No Yes No
Observations 6959 5069 2324 1694 1159 844
Adjusted R20.440 0.476 0.420 0.479 0.553 0.611
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
20
Table 6: DD estimates of the impact of Airbnb on hotel room price at different
frequency level.
Monthly Quarterly Half yearly
(1) (2) (3) (4) (5) (6)
Room price Room price Room price Room price Room price Room price
log Airbnb supply -0.00608 -0.0117∗∗ -0.00758 -0.0142∗∗ -0.00816∗-0.0116∗∗
(0.00580) (0.00567) (0.00606) (0.00597) (0.00480) (0.00502)
log population 0.952 0.356 1.018 0.669 0.962 0.608
(1.289) (1.599) (1.307) (1.614) (1.255) (1.618)
unemployment -0.000460 -0.000661 -0.000758 -0.00102 0.000610 0.000502
(0.00112) (0.00121) (0.00115) (0.00120) (0.00117) (0.00121)
log GDP per capita 0.119 0.0417 0.125 0.00391 0.117 0.0486
(0.108) (0.204) (0.0931) (0.183) (0.0812) (0.146)
Sweden Yes No Yes No Yes No
Observations 6959 5069 2324 1694 1159 844
Adjusted R20.283 0.300 0.366 0.429 0.576 0.619
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
21
Table 7: Try with different lags
(1) (2) (3) (4) (5) (6)
Revenue Occupancy rate Price room Revenue Occupancy rate Price room
log Airbnb supply -0.0102 -0.00210 -0.00608 -0.0168 -0.00251 -0.0117∗∗
(0.0110) (0.00404) (0.00580) (0.0112) (0.00437) (0.00567)
log population 2.680 0.293 0.952 -2.393 -2.229 0.356
(3.074) (1.358) (1.289) (3.068) (1.393) (1.599)
unemployment -0.00253 -0.00128 -0.000460 -0.00298 -0.00147 -0.000661
(0.00249) (0.000861) (0.00112) (0.00283) (0.000937) (0.00121)
log GDP per capita 0.0105 -0.0690 0.119 -0.364 -0.271 0.0417
(0.326) (0.114) (0.108) (0.588) (0.196) (0.204)
Observations 6959 6959 6959 5069 5069 5069
Adjusted R20.440 0.592 0.283 0.476 0.615 0.300
Standard errors in parentheses
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
(1) (2) (3) (4) (5) (6)
Revenue Occupancy rate Price room Revenue Occupancy rate Price room
log Airbnb supplyt−1-0.0126 -0.00353 -0.00624 -0.0184 -0.00339 -0.0116∗
(0.0108) (0.00394) (0.00598) (0.0109) (0.00418) (0.00595)
log population 2.688 0.304 0.944 -2.369 -2.202 0.330
(3.032) (1.328) (1.290) (3.040) (1.378) (1.594)
unemployment -0.00251 -0.00126 -0.000458 -0.00297 -0.00146 -0.000659
(0.00249) (0.000858) (0.00112) (0.00283) (0.000936) (0.00122)
log GDP per capita 0.00900 -0.0699 0.119 -0.364 -0.271 0.0430
(0.327) (0.115) (0.108) (0.589) (0.196) (0.204)
Observations 6959 6959 6959 5069 5069 5069
Adjusted R20.440 0.592 0.284 0.476 0.615 0.300
Standard errors in parentheses
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
22
Table 8: DD estimates of the impact of Airbnb’s closest hotel substitutes on
hotel room revenue, occupancy rates, and prices.
(1) (2) (3) (4) (5) (6)
Revenue Occupancy rate Price room Revenue Occupancy rate Price room
log Airbnb supply -0.0181 -0.00225 -0.0126∗∗ -0.0168 -0.00251 -0.0117∗∗
(0.0120) (0.00454) (0.00577) (0.0112) (0.00437) (0.00567)
log population -2.584 -2.268 0.224 -2.393 -2.229 0.356
(3.132) (1.410) (1.662) (3.068) (1.393) (1.599)
unemployment -0.00292 -0.00146 -0.000619 -0.00298 -0.00147 -0.000661
(0.00282) (0.000935) (0.00121) (0.00283) (0.000937) (0.00121)
log GDP per capita -0.367 -0.271 0.0394 -0.364 -0.271 0.0417
(0.588) (0.196) (0.204) (0.588) (0.196) (0.204)
All listings No No No Yes Yes Yes
Observations 5069 5069 5069 5069 5069 5069
Adjusted R20.476 0.615 0.300 0.476 0.615 0.300
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
23
Table 9: DD estimates of the impact of Airbnb on hotel’s guests composition
by country of origin.
(1) (2) (3)
National guests Nordic guests Nordic guests + neighbors
log Airbnb supply 0.00466∗∗ 0.00486∗∗ 0.00342∗∗∗
(0.00202) (0.00190) (0.00110)
log population -0.876 -0.855 -0.166
(0.585) (0.596) (0.390)
unemployment 0.00129 0.00102 0.000305
(0.000917) (0.000864) (0.000641)
log GDP per capita 0.391∗∗∗ 0.411∗∗∗ 0.411∗∗∗
(0.119) (0.116) (0.0936)
Observations 6977 6977 6977
Adjusted R20.293 0.291 0.356
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
24
Table 10: DD estimates of the impact of Airbnb on hotel’s guests composition
by purpose of stay.
(1) (2) (3)
Holiday, recreation Course, conference Occupation
log Airbnb supply -0.00442 0.00410∗0.000335
(0.00352) (0.00222) (0.00310)
log population -0.751 0.0175 0.740
(0.822) (0.682) (0.759)
unemployment 0.00512∗0.000200 -0.00529∗∗
(0.00265) (0.00102) (0.00215)
log GDP per capita -0.0764 0.0305 0.0464
(0.0898) (0.0437) (0.0753)
Observations 4493 4493 4493
Adjusted R20.767 0.630 0.636
Cluster-robust standard errors in parentheses.
All specifications include region and time fixed effects.
All specifications include a quadratic region-specific time trend.
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
25
Table 11: Experimental 2SLS model instrumenting Airbnb supply with region-
specific time trend.
(1) (2) (3) (4)
Revenue Revenue Revenue Revenue
log Airbnb supply -0.0334∗∗∗ -0.0681∗∗∗ -0.0169∗∗ 0.0126
(0.0107) (0.0138) (0.00835) (0.0105)
log population 4.632∗∗ 4.508∗∗ 6.615∗∗∗ 6.291∗∗∗
(1.944) (1.952) (1.064) (1.066)
unemployment -0.00358 -0.00359 -0.00497∗∗ -0.00472∗∗
(0.00257) (0.00258) (0.00230) (0.00230)
log GDP per capita -0.324∗∗∗ -0.353∗∗∗ -0.334∗∗∗ -0.333∗∗∗
(0.115) (0.116) (0.102) (0.102)
region-specific trend quadratic quadratic linear linear
Observations 6959 6959 6959 6959
Instrument degree 4 degree 3 degree 3 degree 2
Standard errors in parentheses
∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
26