Market segmentation in Airbnb : an empirical study of competition based on Amsterdam’s neighbourhoods and property values

(1)

Faculty of Economics and Business (FEB) MSc. Economics

Track: Industrial Organization, Regulation and Competition Policy

Market segmentation in Airbnb:

an empirical study of competition based on Amsterdam’s

neighbourhoods and property values

Alejandro Zerain 10318968

Supervisor: Dr. Andras Kiss

July 15th, 2016 Amsterdam

(2)

Statement of Originality

This document is written by Student Alejandro Zerain Collet who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

The disruptive effects of the Airbnb platform and whether it should be regulated and how, have been the topic of an on-going debate in most major cities. The municipality of Amsterdam, as many other local governments, have adopted measures to regulate Airbnb and set a legal framework. However, to this point, the lack of empirical research on Airbnb has led policy-makers to address concerns and enable policies with a normative framework. The purpose of the research at hand is to empirically evaluate the dynamics of the Airbnb platform and conduct an exploratory study on the Airbnb market for the city of Amsterdam. In particular, this study aims to elucidate how Airbnb prices are set by answering the following questions:

• To what extent is competition between Airbnb listings localised at neighbourhood level?

• To what extent is competition between Airbnb listings localised to listings with similar property value/quality?

By answering these questions, this paper sheds light on the demand determinants of Airbnb listings in Amsterdam and how this market is constituted. In order to do so, we use data retrieved from the Airbnb platform and the Amsterdam municipality to employ a one level nested logit model on two different set-ups: one where the neighbourhoods are the nests, and another in which the listings have been clustered according to their property value. In both settings we use as instruments the number of listings within their segments and the sum of of the other listings’ characteristics within their segment.

Our results show that the segmentation parameters in both models are close to zero and statistically insignificant. This indicates that the degree of correlation between consumer preferences for apartments in the same neighbourhood or property value range was extremely low. In other words, the competition between Airbnb listings is not localised at the neighbourhood level or between properties with the same value. We conclude that other mechanisms are at play when it comes to product differentiation and consumer preference, and we propose proximity of listings to major tourist attractions or to the city centre, access to public transport or competitors (either other Airbnb listings or hotels) within a certain distance, as factors to take into account for future research.

(5)

5

The following paper is structured as follows. Section 2 elaborates on the previous literature carried out on the Airbnb platform and on the nested-logit model. Section 3 describes the theoretical model used in this research, the nested logit model, and its main assumptions. Section 4 provides data description of the sample used in this research and a preliminary analysis of the listings per neighbourhood. The results from the nested-logit models for both set-us are presented in section 5. The results are discussed and interpreted in section 6, including its limitations and recommendations for future research. Lastly, section 7 provides a brief conclusion.

2. Literature review

There are two forms of literature that are particularly relevant for the research at hand: the work that focused on discrete-choice models of product differentiation, which serves as a theoretical model for our econometric analysis; and the empirical research carried out on the price determinants of the Airbnb market and in attempting to understand better the dynamics of this market. In this section, both strains will be explained as they are relevant to this paper.

There is a significant amount of research done on markets with differentiated products, which serve as a framework for this paper. Berry (1994), Brenkers and Verboven (2004, 2006) use a nested-logit model to specify a market with differentiated products, a methodology that will be to a large extent replicated in this paper. Berry (1994) considers methods for estimating product differentiation models in the presence of unobserved product characteristics. The author elaborates on the theoretical framework for the nested-logit model and performs a Monte Carlo simulation to empirically show how this model reduces the significant bias of unobserved characteristics. In his model, Berry uses a one-nested logit model, which we consider in both of our settings.

The work of Brenkers and Verboven (2004, 2006) has also extensively used the nested-logit model to study the European car market. In both of their papers, the authors employ a two-level model by dividing the car market into segments (domestic vs. foreign) and in sub-segments (e.g. compact, luxury, intermediate, etc.). In one of their papers (2006), the authors quantify the competitive effects of liberalizing the distribution system in the car market for

(6)

6

Europe by using this two nested-logit model as an oligopoly pricing model with differentiated markets. In their paper, the authors consider both a restricted and a flexible version of the model to quantify such competitive effects. The restricted model only has one segmentation parameter for all nests, whereas the flexible model has one segmentation parameter for each nest. In other words, the restricted model assumes all nests have the same level the segmentation whereas the flexible model allows for variation per nest.

In a different work from Brenkers and Verboven (2004), the authors solely employ the flexible nested logit model; and subsequently perform cross-price elasticities analyses across segments and sub-segments, and SSNIP-tests to define relevant markets based on the previously specified nests. Their results show that each car segment appears to constitute a separate market at the country-level with the exception of the mini-van segment. Similar to their work (2004, 2006), in our paper we also consider a two-level nested model in which the sub-segments are the renting conditions; and the segments vary according to our setting: in one we use the neighbourhoods as nests and for the second we cluster listings based on their property value. The main differences in our paper is that we only have one-level nests as Berry (1994) and that we focus only on a restricted model instead of a flexible one

The literature on the Airbnb market is less substantial and there has, to this point, no work been done that attempts to answer the same questions that have been raised in this paper. However, there have been some contributions to the understanding of the Airbnb platform that are worth mentioning. Gutt and Hermann (2015), for instance, study the effect of online reviews in determining listing’s prices on Airbnb, showing that indeed positive rating has an effect on pricing. Edelman & Luca (2014) found that hosts’ race had an impact on the price these could charge for their Airbnb listings, being significantly lower (12% less) for black hosts than for non-black hosts. Previous literature also suggests that listings’ locations is a determinant factor in explaining price variations for Airbnb listings. Quatttrone et al. (2016) use data of the city of London to determine which factors determine the supply of Airbnb listings and, by matching such listing information with census and hotel data. Their results show that supply and demand is determined by the attractiveness of the neighbourhood (in terms of attractions, nightlife, among others) and the socio-demographics of residents (student and lower income neighbourhoods tended to offer more listings). Tang & Sangani (2015) performed a neighbourhood and price prediction for San Francisco’s Airbnb listings. The authors created clusters of neighbourhoods and prices based on listings’ descriptions and particular features,

(7)

7

finding a significant overlap between the features that predicted neighbourhood and price; and suggested that future research should focus on the correlation between these two.

It is indeed this relationship that this study aims to explore and provide some insight. By applying a methodology appropriate to markets with heterogeneous products, this study aims to answer to what extent neighbourhoods and property values determine listings’ prices and delimit markets within the Airbnb platform.

3. The Model: the one nested-logit

For the nested logit model, the products are grouped into G + 1 sets which are exhaustive and mutually exclusive, 𝑔 = 0,1, … , 𝐺. There is an outside good, 𝑗 = 0, which is in this case the option not to purchase. The outside good is the only product of group 0. For each other product 𝑗 in which 𝑗 ∈ 𝔍_- , the utility of the consumer 𝑖 is:

𝑢₀₁ = 𝑥₁𝛽 − 𝛼𝑝₁+ 𝜉₁ + [𝜁_0-+ 1 − 𝜎 𝜀₀₁] (1)

where 𝑥₁ is a vector of observed characteristics of the products (Airbnb listings), in this case being: number of guests allowed, number of bathrooms and the number of beds. 𝜉₁ denotes unobserved product characteristics. 𝜎 represents the within group correlation (0 < 𝜎 < 1) and the variable 𝜁 is common to all products in group 𝑔. The term [𝜁0-+ 1 − 𝜎 𝜀01] has an extreme

value distribution if 𝜀₀₁ is extreme value distributed as well. If the parameter 𝜎 approaches one, then the within group correlation of utility levels approaches one; and if 𝜎 approaches zero, the within group correlation of utility levels is close to zero. The previous equation can be rearranged in the following form:

𝑢₀₁ = 𝛿₁ − _-[𝑑_1-𝜁_0-] + 1 − 𝜎 𝜀₀₁ (2)

in which 𝛿₁ = 𝑥₁𝛽 − 𝛼𝑝₁+ 𝜉₁ and 𝑑_1- is a dummy variable equal to one if 𝑗 ∈ 𝔍_-. This specification shows how this model allows to model correlation between groups of similar products.

It is required for a further analysis with the nested-logit model to determine the market share of each product 𝑗 and its market share as a fraction of its group (if 𝑗 ∈ 𝔍_-). If 𝑗 ∈ 𝔍_- then the market share of product 𝑗 as a fraction of its group is the probability of choosing product j

(8)

8 given its group g:

𝑠

C D

=

E FC GHI J_D (3) where 𝐷_- 𝑖𝑠

𝐷

_-

≡

_1∈𝔍

𝑒

GHIFC D (4)

On the other hand, the probability of picking any given product from group g is

𝑠

_-

=

JD(GHI)

J_D(GHI)

D

(5)

Combining these two probabilities give us the market share of product 𝑗 in group g, as it is the product of the probability of choosing group g, and the probability of choosing product j given group g. Thus, by combining (3) and (5) we obtain:

𝑠

₁

= 𝑠

_-

𝑠

C D

=

E FC GHI J_D

J_D(GHI) J_D(GHI) D

=

EGHIFC J_DI _DJ_D(GHI) (6)

The outside good, with 𝛿_O ≡ 0 , 𝐷_O = 1 and being the only product in group zero (𝑔 = 0) has a market share of

𝑠

_O

=

P

J_D(GHI)

D

(7)

Thus, (6) can also be rearranged as

𝑠

₁

= 𝑠

_-

𝑠

C D

=

E FC GHI J_DI _DJ_D(GHI)

=

EGHIFC J_DI

𝑠

O (8)

In order to derive an expression for mean utility levels, we take the logs of market shares1:

1_{ln 𝑠} 1 − ln 𝑠O = ln E FC GHI J_DI _J D(GHI) D − ln P J_DGHI D = ST

PUV− σ ln(DY) − ln -𝐷-PUZ + ln -𝐷-PUZ

∴ ln 𝑠₁ − ln 𝑠_O = δ]

(9)

9

ln 𝑠1 − ln 𝑠O =_PUVST − σ ln(𝐷-) (9)

This expression, however, is problematic because of the unknown value of D_Y. If we take the log of (5), we obtain: ln(𝑠-) = ln J_D(GHI) J_D(GHI) D = 1 − 𝜎 ln(𝐷-) − ln 𝐷 -(PUZ) - (10)

Notice that the second term in this equation is equal to ln 𝑠_O , therefore it can be rewritten as ln(𝑠_-) = 1 − 𝜎 ln(𝐷_-) + ln(𝑠_O)

or

ln 𝐷_- = ^_(`D)U^_ `a

PUZ (11)

By substituting (11) into equation (9), we obtain

ln 𝑠₁ − ln 𝑠_O = δ] 1 − σ− σ ln(𝑠_-) − ln 𝑠_O 1 − 𝜎 ∴ ln 𝑠₁ − ln 𝑠_O = 𝛿₁ + 𝜎 ln 𝑠1 - (12)

As 𝛿₁ = 𝑥₁𝛽 − 𝛼𝑝₁+ 𝜉₁ , then this expression can be written as

∴ ln 𝑠₁ − ln 𝑠_O = 𝑥₁𝛽 − 𝛼𝑝₁+ 𝜎 ln 𝑠1

- + 𝜉1 (13)

Based on this expression, the parameters of 𝛽, 𝛼 and 𝜎 can be estimated with a linear regression of differences in log market shares on observed product characteristics, prices, and the log of the within group share. However, from economic theory it is known that both the prices and the market shares in this equation are endogenous. To resolve these, we will use a two-stage least square (TSLS) regression model taking both listing prices and market shares as endogenous.

4. Data description and preliminary analysis

The data used in this research stems from three different sources. Each subsection will focus on a part of the data that is intended for specific purposes. Namely, section 4.a describes the largest part of the data which consist of the apartments listed in Airbnb and their main characteristics, hereby presented by neighbourhood in order to understand properly the

(10)

10

categorization (nests) implemented in our first nested logit model further on. Section 4.b elaborates on the retrieved calendar data and how a demand estimator has been constructed based on this information. Finally, section 4.c describes the data obtained on property values (WOZ) for the city of Amsterdam, which is used to determine the nests used in our second nested logit model.

a. Listings and characteristics per neighbourhood

The data available on the listings has been retrieved from the ‘Inside Airbnb project’, a platform that publishes Airbnb data for research and to encourage debate. The data has been retrieved at four different periods of time: April 2015, August 2015, September 2015 and January 2016. For the current study only the data retrieved in two subsequent months (August and September 2015) is considered, due to the demand estimator that has been devised for this paper, explained in the following section.

As a result, the data consist of 3,068 observations from both August and September 2015 (1,814 listings in August and 1,250 in September). There are in total 20 neighbourhoods/areas in which the apartments are distributed. The data includes prices, the date on which the data has been retrieved, room types, host ID and neighbourhoods/areas in which the apartments are located. Furthermore, this data has been complemented with information on the listings’ characteristics retrieved directly from the Airbnb platform in May 2016. It was possible to merge the listing’s characteristics with the information previously obtained using a unique ID number given to each listing by Airbnb itself. These characteristics consist of the number of guest allowed in an apartment, the number of beds and bathrooms. These features will be included in our nested logit model as instrumental variables on the log market shares of the listings and their price.

The neighbourhood categorization which will serve as nests in our model is based on the classification provided by Airbnb and ‘Inside Airbnb’ to filter their data. The reason why such categorization for the nests is the most appropriate has also to do with the fact that Airbnb provides filters on their search results based on these neighbourhood distributions. Thus, as the listings are already differentiated by Airbnb in their platform based on their location, we found appropriate to follow such distribution. In appendix A we provide a summary of statistics of

(11)

11

the price, quantity demanded and the listings’ characteristics per neighbourhood for both periods (August and September 2015).

This preliminary analysis already provides important insight regarding the Airbnb market in Amsterdam.2 From the comparison between both periods, there seems to be a seasonality effect: average prices tend to be higher for all neighbourhoods in the month of August than in September, and so is the supply of listings. This result would be consistent considering that in the month of August there is a larger flow of visitors and thus, higher demand for Airbnb listings explaining both the increase in price and supply. This analysis also shows how skewed the supply of Airbnb listings is based on their neighbourhoods. The neighbourhoods with the largest number of apartments were those within the city centre, such as Centrum-West and Centrum Oost; and areas with a reputation for being trendy, close to attractions and vibrant nightlife such as De Pijp/Rivierenbuurt or Westerpark. The more distant areas are from the centre, the lower the prices and supply of Airbnb listings appears to be (See Gaasperdam, Bijlmer-Centrum or Bijlmer-Oost, for instance). This price dispersion and difference in the supply of Airbnb listings already suggest that there is pronounced product differentiation between neighbourhoods.

b. Calendar data and demand estimations

One of the most common and important limitations on the previous research carried out on Airbnb is the lack of available data on Airbnb demand. For that matter, most of the analyses previously done focus on the supply-side. In our particular case, all the accessed data also provides information on the supply-side; which poses an even larger problem for our nested logit model as it requires data on market shares. However, one solution has been found to provide a reliable demand estimator for the Airbnb listings.

The data from the platform ‘Inside Airbnb’ offers detailed data on the calendar provided by the Airbnb platform. This calendar is shown when a reservation is being made and displays which days of the month the apartment is available and at what price. We will use this information to determine how availability of nights within a month changes over time, and define this as the demanded quantity. As the data has been retrieved in several periods of time there are in

(12)

12

particular two subsequent periods for which the change in availability can be easily captured: August and September 2015. Therefore, a demand estimator is defined as:

𝑄₀ = ∆𝑁₀ = (𝑁_0P− 𝑁_0e) (14)

where N denotes the number of nights available to book within a month. The subscript i indicates each listing, and the subscript numbers refer to the period in time in which the data was retrieved: 1 indicates August 2015, and 2 indicates September 2015. Thus, quantity demanded (𝑄₀₁) is defined as the change from August to September on the number of nights a listing is available for each of the two following months (September and October).

One of the strengths of this estimator is that it accounts for the nights that listings are unavailable for different reasons than previously booked. Consider a listing that was available for the month of September when the data was observed in August; if it appears unavailable a month later it is very likely that the reason is because the listing has been booked. Whereas if an apartment was unavailable from the beginning, it is likely that it was for other reasons besides being booked already.

Another strength of this estimator is that it allows for a definition of the “outside good” that is consistent with the nested logit model. The outside good, which is another option for which the consumer can pick but receive no utility, is usually defined as a not-to-purchase option within a certain market. Our demand estimator allows for a simple definition of the outside good, that being the number of nights for which an apartment was available and remained unreserved at the second point in time when the data was retrieved.

c. Property value data

Information on property value of properties (Woningwaarde WOZ) has been retrieved online from a publicly available database provided by the Municipality of Amsterdam.3 The information published by the municipality includes geographical coordinates of the properties

3_{Gemeente Amsterdam – Belastingen (January 2015) Retrieved from:} http://maps.amsterdam.nl/woningwaarde_woz

(13)

13

which allows us to match such information with the Airbnb listings.4 This average WOZ property value of an Airbnb listing is assigned by matching the coordinates of the listing to those of to the nearest block of houses presented in the Municipality’s dataset, measured by Euclidean distance in meters. This process does not come however with its limitations. Due to the lack of preciseness in the matching procedure of the geographical coordinates, it is the case that for some of the listings the coordinates do not match precisely within both datasets. As a measure of inaccuracy the mean distance between the Airbnb listing’s coordinates and the municipality’s information is of 24.92 meters, which is rather large.

Nevertheless, the use of this information allows for a different classification which might be more in line with the actual quality of the apartments offered in the Airbnb platform and possible segmentation based on their characteristics. The WOZ property value is an annual estimation carried out by the municipality for taxing purposes such as property tax, income tax or water services. As the assessment of the Municipality reflects the state and quality of the property, this perhaps reflects a better classification in terms of apartment’s characteristics and features rather than using neighbourhoods. To further substantiate this claim, the WOZ property values in Amsterdam shows that these values are still similar within each neighbourhood/area.5 Thus, the use of the classification provided by the Municipality in terms of property value may not only reflect the appeal of being located within a certain neighbourhood, as reflected in the Airbnb classification; but also segmentation based on the quality of the apartment itself as reflected in its property value.

5. Data Analysis

a. Demand estimation with nested logit model

Based on the nested logit model specified before, using at two-stage least square we will determine the following equation6:

4_{It is important to note that the WOZ property value does not directly reflect the actual value of the} property but is in itself an average of the housing block in which the property is located. In this way, the Municipality of Amsterdam ensures the anonymity of the property owners. (Gemeente Amsterdam – Belastingen, January 2015). Retrieved from: http://maps.amsterdam.nl/woningwaarde_woz

5_{See visualization in Gemeente Amsterdam – Belastingen (January 2015). Woningwaarde WOZ. Retrieved} from http://maps.amsterdam.nl/woningwaarde_woz

6_{Income (𝑦) reflects the GDP per capita in the Netherlands (2015) per month as 4036 USD. Source: IMF}

(14)

14 ln 𝑠₁ − ln 𝑠_O = 𝑥₁𝛽 − 𝛼 gC

h + 𝜎 ln 𝑠1 _- + 𝜉1 (15)

As we have discussed beforehand, the price and the market shares of the apartments are endogenous. Finding supply-side variables that affect costs and market shares is rather difficult, so instead we will use as instrumental variables demand-related variables. These instrumental variables will be the following:

• The number of listings present in their nest (neighbourhood/area) at that given period • The sum of the values of the same characteristics of products offered within the same

neighbourhood/area by other hosts. These characteristics are the number of guests allowed, the number of beds available and the number of bathrooms.

A dummy variable for the months has been included to control for time fixed effects, such as the seasonality effect mentioned early.

b. Nests based on neighbourhood/areas

The results in Table 1 reflect different regression models using as nests the neighbourhoods/areas that are employed by Airbnb. Column (1) shows a simple OLS nested logit model, whereas the other two columns use a TSLS regression based on the instruments previously specified, and allow for a better analysis of segmentation based on neighbourhoods. Column (2) shows the results for a one-level nested logit and column (3) refers to a simple nested logit model where 𝜎 = 0. For all models we have controlled for time-fixed effects by adding a dummy variable.

(15)

15

Table 1 - Regression results with nests as neighbourhoods/areas

(1) OLS (2) Nested Logit (3) Logit

y y y princ 9.6512*** -3.9165 -4.2080 (16.02) (-0.64) (-1.92) ln_Sjg 0.4842*** 0.0025 (54.91) (0.05) Bathrooms -0.0303 0.0742 0.0770 (-1.01) (1.00) (1.51) Guests -0.0660*** 0.0072 0.0091 (-5.43) (0.16) (0.40) Beds -0.0476*** -0.0228 -0.0221 (-3.83) (-1.02) (-1.20) Sep-dummy -0.6050*** -0.4967*** -0.4966*** (-28.22) (-16.47) (-16.41) _cons -7.1789*** -9.7489*** -9.7612*** (-120.65) (-39.47) (-178.88) N 3068 3068 3068 t-statistics in parentheses. * p<0.05, ** p<0.01, *** p<0.001

The segmentation parameter (𝜎) for the OLS model in column (1) appears to be rather strong at 0.4842. At first sight this would indicate strong correlation for listings with similar characteristics, these being located in the same area. The results, however, also indicate problems with the model specifications. All of the listings’ characteristics appear to have a negative impact on the demand which is counterintuitive: one would expect that the apartments of higher quality (better features) would have a higher demand than the rest. The results are statistically significant for the number of beds and guests allowed. Moreover, against the model’s assumptions, α> 0 strengthening the argument for a model misspecification.

The TSLS regression of the nested logit model in column (2) confirms the existence of endogeneity problems which are expected from such a model. The number of listings within a

(16)

16

listings’ neighbourhood and the sum of the characteristics of other listings, have been instrumented on the price and the log of the within group market shares (ln S_]|Y). In this model, the impact of the listings’ characteristics on the demand changes considerably. The listings’ characteristics do not longer have a negative effect on demand, as it can be seen with the number of guests allowed and the number of bathrooms, but most importantly these effects are no longer statistically significant. The results from the first stage regression on price and the log of the within group market shares (ln S_]|Y) are presented in appendix B. These show that the number of competing listings in the neighbourhood and the characteristics of these other listings impact negatively the price and the market share of the apartment.7 These results are not only coherent with economic theory but are also robust statistically. Post-estimation tests confirm the specification of this model: the Wooldridge’s score test and a regression-based test of exogeneity both indicate that the instrumented variables are endogenous.8 Moreover, a Wooldridge’s robust score test of overidentifying restrictions also confirms that the instruments are uncorrelated with the error term and the equation is not wrongly specified.

In contrast with the OLS model, the segmentation parameter (𝜎) is very weak and close to zero. This suggest that even if there is product differentiation between all the listings, there is no clustering at the neighbourhood level, meaning that it is unlikely that the cross price elasticity between products of the same neighbourhood to be larger than with apartments in other neighbourhoods. In fact, the segmentation parameter (𝜎) is not only close to zero but also statistically insignificant, thus it cannot be discarded that 𝜎 = 0. Thus, there is very little correlation for demand between apartments located in the same neighbourhood in comparison to the rest of areas. Column (3) presents a simple logit model which is the equivalent of a nested logit model in which the segmentation parameter 𝜎 = 0. The results are similar to the nested logit model in column (2), illustrating the fact that there is very small segmentation in the Airbnb market: there appears to be non-localised competition based on the neighbourhood results.

c. Nests based on WOZ property value

The previous section showed that even as the Airbnb market for the city of Amsterdam is heterogeneous, there is little segmentation in the Airbnb market for the city of Amsterdam

7

See appendix B 8_{See appendix C}

(17)

17

when the nests are defined based on the neighbourhoods in which these listings are located. A possible explanation for these results could be that listings located within neighbourhoods are different between each other and may not share similar characteristics that justify this classification, despite having the same location. This section considers the nested logit model in which the nests have been defined based on WOZ property value rather than by neighbourhood/areas.

Table 2 considers different regression models in which the listings have been nested based on WOZ property value. Similar to the previous section, column (1) shows a simple OLS model, whereas the other three columns use a TSLS regression model, using similar instruments as before with the difference that the nests are not the same. Column (2) shows a one-level nested logit model and column (3) presents a simple logit model in which 𝜎 = 0, for comparison.

Table 2 - Regression results with nests based on WOZ property value

(1) OLS (2) Nested Logit (3) Logit

y y y princ 1.8656*** -5.2933 -5.6413 (4.14) (-1.17) (-1.27) ln_Sjg 0.6927*** 0.0094 (84.57) (0.33) Bathrooms -0.0160 0.0877 0.0916 (-0.69) (1.32) (1.37) Guests -0.0168 0.0168 0.0190 (-1.79) (0.47) (0.54) Beds -0.0252** -0.0194 -0.0187 (-2.63) (-0.94) (-0.90) Sep-15 -0.7355*** -0.5026*** -0.5000*** (-44.01) (-15.32) (-15.69) _cons -5.5336*** -9.6990*** -9.7555*** (-96.42) (-53.46) (-167.89) N 3068 3068 3068 t-statistics in parentheses. * p<0.05, ** p<0.01, *** p<0.001

(18)

18

The results in Table 2 are similar to those presented in the previous section. The OLS model (column 1) shows a strong segmentation parameter (0.6927), however once instrumental variables are included for the log of market shares and the price, the segmentation parameter decreases significantly. In the nested logit model (column 2) the segmentation parameter is remarkably small and not significantly different from zero. Comparing the nested logit model to the simple logit model in which the segmentation parameter is assumed to be equal to zero, the results are very similar again. None of the listings’ characteristics have a strong effect on demand as their coefficients are small and statistically insignificant; and the assumption of no segmentation in logit model does not change these results. Thus, similar to the neighbourhood nesting, there is a clear indication that there is very little correlation for demand between apartments with the same WOZ property value. This suggest that even if there is product differentiation in terms of quality and value between all the listings, there is no clustering based on property value.

Post-estimation tests confirm that all of these models are not wrongly specified and do not violate the assumptions of a two-stage least square. The Wooldridge’s score test and a regression-based test of exogeneity both indicate that the instrumented variables are endogenous.9 Moreover, a Wooldridge’s robust score test of overidentifying restrictions also confirms that the instruments are uncorrelated with the error term and the equation is not wrongly specified

6. Discussion and limitations

Before interpreting the results of this paper it is fundamental to understand what the segmentation parameter (𝜎) conveys. This parameter measures the degree of correlation of consumer preferences for apartments that belong to the same group. If the consumer preferences are correlated between listings in the same groups, this allows us to evaluate whether substitution between products is driven by how similar the listings are. If 𝜎 is close to 1, preferences for apartments within the same group are perfectly correlated, which translate as these listings being perfect substitutes. Thus, one would expect that when the price of one apartment increases, consumers would replace it with another apartment in the same

(19)

19

neighbourhood or property value/quality. Similarly, if 𝜎 goes to zero, all apartments are uncorrelated with each other despite their similarities, resulting in a simple logit model. That is to say, if 𝜎 = 0 there is non-localised competition in the Airbnb market. Therefore, the estimation of these parameters allow us to establish the extent to which competition is contained between listings from the same segment.

In our case, we have presented two set-ups in which the listings have been categorised differently. The first consisted of clustering at the neighbourhood level and the second one of clustering based on ranges of WOZ property values as groups. The results from the neighbourhood model using a nested-logit model show that there is very little correlation of consumer’s preferences between listings in the same neighbourhood. As such, substitution patterns do not seem to be driven by how similar the apartments are in terms of their location: if the price of one apartment increases, consumers will not resort more to different apartment within the same neighbourhood.

The other interpretation evaluates whether is possible that competition within the Airbnb platform in Amsterdam is contained in different segments based on the quality of the apartment, denoted by its WOZ property value. This interpretation assumes that a higher WOZ property value implies that the property is in better condition and expensive properties probably offer a more luxurious accommodation. The nested logit model results for this interpretation do not show a strong segmentation parameter that supports this hypothesis. The group parameter (categories based on WOZ property value) is also close to zero, indicating that consumer preferences are not at all correlated by how similar the property value of the accommodation is. Thus, competition in the Airbnb market for the city of Amsterdam is not localised either on different segments that perhaps resemble more the housing market.

Based on these results, it is not necessary to estimate the cross-price elasticities or perform a SSNIP-test such as Brenkers and Verboven (2004) to determine whether each neighbourhood, or each segment of the WOZ property value categories represent a separate market. The lack of correlation in consumer preferences already suggest that substitution patterns are dependant only on the market shares of the substitute listing, rather than on how similar they are. These results are strikingly different to those observed in the European car market by Brenkers and Verboven (2004) mainly due to the reason in which the market is constituted. Brenkers and Verboven’s model ought to study an oligopolistic market. The European car market consist of

(20)

20

a limited number of firms which have a substantial market share in each country. Moreover, the barriers to entry are high due to the large investment costs and the technology required to produce vehicles.

On the other hand, the Airbnb platform more closely resembles a market with perfect competition. Besides the large number of hosts participating in the market; there are two factors that limit the capacity and the supply of listings, disenabling hosts from capturing a larger market share. Due to government regulation, each host can only offer their apartment on Airbnb a limited number of nights per year, thus even if demand is high the host is restricted to provide accommodation the same number of nights as other hosts. Secondly, the ability of hosts to offer multiple accommodations on the Airbnb platform is limited by the fixed costs of purchasing a second property and the existing regulations both of Airbnb and the government. Thus, due to these conditions of the Airbnb market the market power of hosts is rather small and the substitution patterns are hardly correlated within groups.

Despite the large amount of hosts competing in the Airbnb platform, by taking a closer look at the Airbnb supply by neighbourhood one can observe that for some of the neighbourhoods there are very few apartments competing in their segment. Before disregarding the use of a nested logit model it would be relevant for further research to relax the segmentation parameters assumption and allow for each group to have a segmentation parameter of its own. By using a restricted nested logit model, one of our strongest assumptions is that all groups share the same segmentation parameter (𝜎). In implementing a flexible model, we would be able to test whether the correlation of consumer preferences for apartments within the same group differs significantly across each group.

Moreover, the weak segmentation parameters and the lack of correlation in consumer preference within group could be partly attributed to the heterogeneity of listings within groups in our classification. For that reason, it would be relevant as an extension of our work to further classify the listings into sub-groups based on certain features that the apartments may share. An important one, for instance, are the different renting conditions that Airbnb allows in their platform: either renting an entire apartment or sharing with the owner). Taking into account these conditions may resemble more appropriately consumers’ preferences and show higher correlation within these sub-groups. Thus, we suggest this aspect to be considered in a two-level nested logit model.

(21)

21

There are other limitations that are worth considering when assessing the results of this paper and that should be taken into account for further research. One of them is the assumption made in this paper in which Airbnb platform constitute a separate market to the hotel industry in the city of Amsterdam. This assumption was made on a qualitative basis: due to the significant differences in the service they provide (i.e. room service, breakfast, among others), it was assumed that the products they offer are different and constitute two separate markets. Nevertheless, in light of the poor results hereby obtained with the nested logit model, it could be the case that other mechanisms are at play in this market, such as the hotel concentration. The assumption to exclude the hotel industry is rather strong and lacks empirical evidence to support it. It is therefore recommended for future research to assess whether the Airbnb platform and the hotel industry are part of the same market.

Similar to the case of the hotel industry, the proximity to major tourist attractions in the city of Amsterdam has not been taken into account. The lack of correlation in consumer preference between listings of the same property value suggest that the Airbnb market does not act similarly as Amsterdam’s housing market. An alternative explanation for this result is that consumers value more the proximity of their accommodation to the city centre, major tourist attractions or access to public transport. Thus, using the coordinates available in our dataset a different model including geospatial information could provide more insight on the mechanisms that shape the price and demand in the Airbnb market.

Lastly, limitations regarding the data collection process and the quality of the estimators should also be mentioned. Even when this paper provides a valid demand estimator considering the lack of reliable data available on quantity demanded, there is room for improvement. The dates in which the calendar data information was retrieved by the Inside Airbnb project were taken as given due to the complexity of collecting this information directly and also to the short timeframe to carry out this study. A better solution would be to use the same estimator but retrieve this information directly from the Airbnb platform by doing a weekly or by-weekly follow up on the platform availability for a prolonged period of time. This would provide a more consistent picture of the changes in availability and demand than what has been used in this paper. It would allow to consider more months and perhaps evaluate better the seasonality effects that have been previously mentioned. Moreover, a continuous follow-up would allow to check more accurately changes in availability reflecting a superior demand estimator.

(22)

22

7. Conclusion

This paper assessed to what extent is the competition between Airbnb accommodations contained within neighbourhoods; or alternatively, whether competition in the Airbnb platform is higher between apartments with similar property value/quality. In order to do so, we employed a one-level nested logit model with two different categorisations. One established neighbourhoods/areas as nests; and the other grouped listings with similar property value as nests.

In the case of the neighbourhood/areas definition, the results of the nested-logit model indicate that the degree of correlation between consumer preferences for apartments in the same neighbourhood was extremely low. This suggest that competition is not localised at the neighbourhood level but rather that listings compete fiercely in the city as a whole. In other words, the listings’ neighbourhood does not play an important role in differentiating the product from the rest of the market. Similarly, the results from clustering listings based on property value also showed that the consumer preferences were hardly correlated between apartments that had a similar property value/quality. Therefore, contrary to the housing market, competition in the Airbnb platform is not higher between listings that share similar value or quality compared to others.

The weak segmentation parameters in our nested logit-models suggest that there are other mechanisms at play when it comes to product differentiation, consumer preferences and therefore as determinants of the demand in the Airbnb market. Alternatives that we have discussed include the different Airbnb’s renting conditions, proximity of listings to major tourist attractions or to the city centre, access to public transport or competitors (either other Airbnb listings or hotels) within a certain distance. We recommend that these aspects are considered in future empirical research regarding the Airbnb platform.

(23)

23

Bibliography

Berry, S. T. (1994). Estimating discrete-choice models of product differentiation. The RAND

Journal of Economics, 242-262.

Brenkers, R., & Verboven, F. (2004). Market definition with differentiated products-lessons from the car market. mimeo, Catholic University of Leuven.

Brenkers, R., & Verboven, F. (2006). Liberalizing a distribution system: the European car market. Journal of the European Economic Association, 4(1), 216-251.

Edelman, B. G., & Luca, M. (2014). Digital discrimination: The case of airbnb.com. Harvard

Business School NOM Unit Working Paper, (14-054).

Gemeente Amsterdam – Belastingen (January 2015). Woningwaarde WOZ. Retrieved from http://maps.amsterdam.nl/woningwaarde_woz

Gutt, D., & Herrmann, P. (2015). Sharing Means Caring? Hosts' Price Reaction to Rating Visibility.

International Monetary Fund (2016) World Economic Outlook Database 2016. Retrieved from: http://www.imf.org/external/pubs/ft/weo/2016/01/weodata/index.aspx

Inside Airbnb (January 2016). Inside Airbnb: Adding data to the debate – Amsterdam. Retrieved from http://www.insideairbnb.com/get-the-data.html

Quattrone, G., Proserpio, D., Quercia, D., Capra, L., & Musolesi, M. (2016, April). Who Benefits from the Sharing Economy of Airbnb?. In Proceedings of the 25th International

Conference on World Wide Web (pp. 1385-1394). International World Wide Web

Conferences Steering Committee.

Tang, E., & Sangani, K. (2015) Neighborhood and Price Prediction for San Francisco Airbnb Listings.

(24)

24

Appendix

A. Summary Statistics per neighbourhood/area

Neighbourhood: Bijlmer Centrum

August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 4 63.94 20.56 45 90 2 47.71 5.85 43.57 51.84 Quantity 4 3.75 3.40 1 8 2 5 0 5 5 Guests 4 2.5 1 2 4 2 2 0 2 2 Bathrooms 4 1 0 1 1 2 1 0 1 1 Bedrooms 4 1 0 1 1 2 1 0 1 1 Beds 4 1.5 1 1 3 2 1 0 1 1 Neighbourhood: Bijlmer-Oost August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 5 65.74 26.28 37.33 109 3 72.67 31.94 49 109 Quantity 5 6.8 4.71 3 15 3 4.67 2.52 2 7 Guests 5 2.6 1.34 2 5 3 3 1.73 2 5 Beds 5 1.8 0.84 1 3 3 2 1 1 3 Bedrooms 5 1.2 0.45 1 2 3 1.33 0.58 1 2 Bathrooms 5 1.1 0.55 0.5 2 3 0.83 0.29 0.5 1 Neighbourhood: Bos en Lommer

N Mean SD Min Max N Mean SD Min Max Price 81 110.09 61.92 25 532.47 55 94.88 32.17 25 170.3 Quantity 81 5.38 4.29 1 20 55 7.24 7.95 1 31 Guests 81 2.70 1.10 2 8 55 2.75 0.97 2 6 Beds 81 1.57 0.91 1 4 55 1.6 0.91 1 4 Bedrooms 81 1.19 0.81 0 5 55 1.09 0.73 0 3 Bathrooms 81 1.01 0.19 0.5 2 55 1.03 0.22 0.5 2 Neighbourhood: Buitenveldert - Zuidas

N Mean SD Min Max N Mean SD Min Max Price 27 94.97 39.23 44 189.08 14 93.66 49.47 45 228.6 Quantity 27 5.30 5.57 1 29 14 5.43 4.65 1 15 Guests 27 2.67 1.18 1 6 14 2.86 1.35 1 6 Beds 27 1.70 1.44 1 6 14 1.79 1.37 1 6 Bedrooms 27 1.30 0.61 1 3 14 1.36 0.63 1 3 Bathrooms 27 1.09 0.28 0.5 2 14 1.07 0.33 0.5 2

(25)

25

Neighbourhood: Centrum-Oost

N Mean SD Min Max N Mean SD Min Max Price 204 183.45 110.14 54 850 164 167.46 87.86 54.45 632.71 Quantity 204 5.53 4.90 1 30 164 5.68 4.38 1 21 Guests _{204 2.98} _1.33 ₁ ₈ _{164 3.10} _1.27 ₁ ₈ Beds 204 1.93 1.32 1 9 164 1.99 1.29 1 9 Bedrooms 204 1.27 .88 0 6 164 1.24 .86 0 6 Bathrooms 204 1.11 .41 .5 5 164 1.09 .41 0 5 Neighbourhood: Centrum-West August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 362 188.52 100.15 45 937 260 178.07 93.07 45 937 Quantity 362 4.93 4.11 1 22 260 6.14 5.42 1 31 Guests _{362 3.29} _1.88 ₁ ₁₆ _{260 3.18} _1.83 ₁ ₁₆ Beds 362 2.13 1.48 1 10 260 2.06 1.47 1 10 Bedrooms 362 1.29 0.78 0 6 260 1.26 0.78 0 5 Bathrooms 362 1.15 0.39 0 3.5 260 1.14 0.38 .5 3.5 Neighbourhood: De Aker - Nieuw Sloten

N Mean SD Min Max N Mean SD Min Max Price 10 142.93 149.06 50.5 500 4 182.6 214.78 55 504 Quantity 10 5.3 3.56 1 11 4 2.75 0.96 2 4 Guests ₁₀ _3.8 _3.08 ₁ ₁₂ ₄ _5.25 _4.57 ₂ ₁₂ Beds 10 2.9 3.38 1 12 4 1.38 0.48 1 2 Bedrooms 10 1.1 0.57 0 2 4 1.25 0.5 1 2 Bathrooms 10 1.25 0.42 1 2 4 4.5 5.07 1 12 Neighbourhood: De Baarsjes - Oud-West

N Mean SD Min Max N Mean SD Min Max Price 236 138.12 68.39 30 400 182 131.71 68.30 30 400 Quantity 236 5.89 4.92 1 28 182 6.20 5.88 1 31 Guests _{236 2.92} _1.29 ₁ ₁₀ _{182 2.92} _1.34 ₂ ₁₀ Beds 236 1.90 1.25 1 10 182 1.93 1.35 1 10 Bedrooms 236 1.31 0.73 0 4 182 1.29 0.73 0 4 Bathrooms 236 1.10 0.57 0 7.5 182 1.11 0.62 0 7,5

(26)

26

Neighbourhood: De Pijp - Rivierenbuurt

N Mean SD Min Max N Mean SD Min Max Price 172 136.81 57.06 45 354.37 113 132.21 59.15 45 370 Quantity 172 5.42 4.57 1 25 113 6.60 6.79 1 31 Guests _{172 2.83} _1.17 ₁ ₈ _{113 2.84} _1.28 ₁ ₈ Beds 172 1.87 1.21 1 6 113 1.92 1.30 1 6 Bedrooms 172 1.28 0.71 0 3 113 1.27 0.71 0 3 Bathrooms 172 1.08 0.25 0.5 2.5 113 1.09 0.27 0.5 2.5 Neighbourhood: Gaasperdam - Driemond

N Mean SD Min Max N Mean SD Min Max Price 6 69.04 20.82 38.42 85 5 65.86 22.42 36 85 Quantity 6 7 6.69 1 18 5 4.8 4.09 2 12 Guests ₆ _2.33 _0.52 ₂ ₃ ₅ _2.2 _0.45 ₂ ₃ Beds 6 1.33 0.52 1 2 5 1.2 0.45 1 2 Bedrooms 6 1.17 0.41 1 2 5 1 0 1 1 Bathrooms 6 1.08 0.20 1 1.5 5 1.1 0.22 1 1.5 Neighbourhood: Geuzenveld - Slotermeer

N Mean SD Min Max N Mean SD Min Max Price 33 95,41 52,49 29 240 21 94,81 55,86 29 240 Quantity 33 5,52 3,83 1 16 21 8,95 7,49 2 30 Guests ₃₃ _3,03 _1,93 ₁ ₁₀ ₂₁ _3,33 _2,27 ₁ ₁₀ Beds 33 1,91 1,18 1 6 21 2,05 1,32 1 6 Bedrooms 33 1,45 0,83 1 4 21 1,57 0,98 1 4 Bathrooms 33 1,09 0,26 0,5 1,5 21 1,14 0,23 1 1,5 Neighbourhood: IJburg - Zeeburgereiland

N Mean SD Min Max N Mean SD Min Max Price 31 124,32 76,30 47 325 21 112,41 60,90 47,5 219 Quantity 31 5,13 4,64 1 25 21 6,10 6,57 1 31 Guests ₃₁ _3,65 _1,56 ₂ ₇ ₂₁ _3,57 _1,63 ₂ ₆ Beds 31 2,81 1,68 1 7 21 2,43 1,60 1 6 Bedrooms 31 1,45 0,72 1 3 21 1,48 0,93 0 4 Bathrooms 31 1,21 0,42 0 2 21 1,21 0,37 1 2

(27)

27

Neighbourhood: Noord-Oost

N Mean SD Min Max N Mean SD Min Max Price 15 77,11 21,33 46 125 10 67,18 15,41 45 90 Quantity 15 4,87 3,64 1 11 10 4,70 4,03 1 15 Guests ₁₅ _2,93 _1,03 ₁ ₄ ₁₀ _2,50 _0,97 ₁ ₄ Beds 15 2,00 1,07 1 4 10 1,80 1,14 1 4 Bedrooms 15 1,07 0,59 0 2 10 1,00 0,47 0 2 Bathrooms 15 1,03 0,13 1 1,5 10 1,05 0,16 1 1,5 Neighbourhood: Noord-West August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 36 91,33 38,20 35 200 20 86,97 28,25 45 153 Quantity 36 4,92 3,78 1 19 20 5,15 4,08 1 15 Guests ₃₆ _2,94 _1,47 ₁ ₈ ₂₀ _3,05 _1,43 ₂ ₈ Beds 36 2,17 1,42 1 6 20 1,95 1,10 1 5 Bedrooms 36 1,42 0,94 0 4 20 1,40 0,88 0 4 Bathrooms 36 1,08 0,25 1 2 20 1,08 0,18 1 1,5 Neighbourhood: Oostelijk Havengebied - Indische Buurt

N Mean SD Min Max N Mean SD Min Max Price 71 112,06 48,91 42 291,8 49 107,90 41,06 42 260 Quantity 71 5,24 4,28 1 20 49 7,55 8,61 1 31 Guests ₇₁ _2,76 _1,19 ₁ ₆ ₄₉ _2,78 _1,21 ₁ ₆ Beds 71 1,72 0,97 1 5 49 1,76 0,95 1 4 Bedrooms 71 1,21 0,53 0 3 49 1,27 0,57 0 3 Bathrooms 71 1,09 0,24 0,5 2 49 1,07 0,20 1 2 Neighbourhood: Osdorp August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 20 98,78 44,27 43,75 197,2 11 78,17 26,64 50 140 Quantity 20 6,95 4,82 1 20 11 7,27 8,89 1 31 Guests ₂₀ _2,85 _1,35 ₁ ₆ ₁₁ _2,55 _1,37 ₁ ₆ Beds 20 1,90 1,33 1 6 11 1,27 0,47 1 2 Bedrooms 20 1,35 0,67 1 3 11 1,09 0,30 1 2 Bathrooms 20 1,08 0,24 0,5 1,5 11 0,95 0,27 0,5 1,5

(28)

28

Neighbourhood: Oud-Noord

N Mean SD Min Max N Mean SD Min Max Price 68 131,26 113,31 40 700 34 113,57 108,62 40 700 Quantity 68 6,15 5,13 1 24 34 6,24 4,88 1 19 Guests ₆₈ _3,47 _2,22 ₂ ₁₄ ₃₄ _3,59 _2,43 ₂ ₁₄ Beds 68 2,37 1,98 1 14 34 2,59 2,43 1 14 Bedrooms 68 1,46 1,04 0 7 34 1,50 1,29 0 7 Bathrooms 68 0,99 0,35 0 2 34 1,00 0,43 0 2 Neighbourhood: Oud-Oost August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 89 128,36 54,62 45 400 54 115,95 45,27 43 295 Quantity 89 5,73 4,84 1 25 54 5,78 6,11 1 31 Guests ₈₉ _2,78 _1,28 ₁ ₈ ₅₄ _2,81 _1,27 ₁ ₆ Beds 89 1,67 1,12 1 8 54 1,70 1,27 1 8 Bedrooms 89 1,33 0,72 0 5 54 1,33 0,80 0 5 Bathrooms 89 1,07 0,24 0,5 2 54 1,02 0,19 0,5 2 Neighbourhood: Slotervaart August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 36 107,87 54,35 50 265 31 98,85 51,26 50 250 Quantity 36 7,39 7,21 1 30 31 7,52 8,23 1 31 Guests ₃₆ _3,00 _1,26 ₁ ₆ ₃₁ _3,13 _1,18 ₁ ₆ Beds 36 1,67 1,07 1 6 31 1,81 1,11 1 6 Bedrooms ₃₆ _1,42 _0,77 ₀ ₃ ₃₁ _1,45 _0,72 ₀ ₃ Bathrooms 36 1,11 0,21 1 1,5 31 1,11 0,25 1 2 Neighbourhood: Watergraafsmeer August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price ₃₀ _117,70 _54,70 _49,09 _277,17 ₂₂ _103,66 _40,03 _49,4 ₁₉₅ Quantity 30 6,27 4,60 1 17 22 7,36 7,99 1 31 Guests ₃₀ _2,83 _0,95 ₂ ₄ ₂₂ _2,55 _0,86 ₂ ₄ Beds 30 1,97 1,10 1 5 22 1,73 0,83 1 3 Bedrooms 30 1,27 0,78 0 3 22 1,11 0,31 1 2 Bathrooms 30 1,05 0,24 0,50 2 22 1,05 0,58 0 2

(29)

29

Neighbourhood: Westerpark

N Mean SD Min Max N Mean SD Min Max Price 153 154,93 93,46 49 775 102 138,7 76,85 50,1 528,6 Quantity 153 6,43 5,31 1 29 102 7,04 7,52 1 31 Guests ₁₅₃ _3,08 _1,89 ₁ ₁₆ ₁₀₂ _3,01 _2,18 ₂ ₁₆ Beds 153 2,02 2,01 1 16 102 2,08 2,34 1 16 Bedrooms 153 1,40 1,00 0 10 102 1,28 0,84 0 6 Bathrooms 153 1,10 0,34 0 3 102 1,12 0,37 0,5 3 Neighbourhood: Zuid August 2015 September 2015

N Mean SD Min Max N Mean SD Min Max Price 125 162,99 99,90 28 633,33 73 155,23 103,95 51,46 640 Quantity 125 6,10 5,44 1 29 73 7,19 6,97 1 31 Guests ₁₂₅ _3,11 _1,77 ₁ ₁₄ ₇₃ _3,07 _1,50 ₁ ₁₀ Beds 125 2,03 1,72 1 14 73 2,07 1,49 1 7 Bedrooms 125 1,51 0,89 0 5 73 1,52 0,90 0 4 Bathrooms 125 1,14 0,31 1 2,5 73 1,13 0,31 1 2,5

(30)

30 B. First-stage regressions a. Table 1, Column (2) Number of observations = 3068 F (8,3059) = 152.20 Prob>F = 0.000 R-Squared = 0.5138 Adjusted R-Squared= 0.5126 Root MSE = 0.0166

princ Coef. Robust Std. _Err. t P>t [95% Confidence Interval] Bathrooms .0090899 .0020764 4.38 0.000 .0050186 .0131612 Guests .0066115 .0006567 10.07 0.000 .0053239 .007899 Beds .0025153 .0005923 4.25 0.000 .0013538 .0036767 Sept.15 -.0027201 .0005896 -4.61 0.000 -.0038762 -.0015639 ins_Bathrooms_2 -.0001466 .000037 -3.96 0.000 -.0002192 -.0000741 ins_Guests_2 -.0000239 .0000197 -1.21 0.226 -.0000625 .0000148 ins_Beds_2 -.000011 .000029 -0.38 0.705 -.0000679 .0000459 nlistings .0001905 .0000254 7.50 0.000 .0001407 .0002403 _cons -.0042295 .0022872 -1.85 0.065 -.0087142 .0002552 Number of observations = 3068 F (8,3059) = 326.84 Prob>F = 0.000 R-Squared = 0.4943 Adjusted R-Squared= 0.4930 Root MSE = 0.8790

ln_Sjg Coef. Robust Std. _Err. t P>t [95% Confidence Interval] Bathrooms .0345185 .0436775 0.79 0.429 -.0511218 .1201588 Guests -.0116287 .0170349 -0.68 0.495 -.0450297 .0217724 Beds -.0337367 .0184609 -1.83 0.068 -.0699336 .0024602 Sep-15 .347409 .0323873 10.73 0.000 .283906 .410912 ins_Bathrooms_2 -.0168158 .0019358 -8.69 0.000 -.0206114 -.0130202 ins_Guests_2 .0127565 .0011928 10.69 0.000 .0104176 .0150953 ins_Beds_2 -.0073441 .0016767 -4.38 0.000 -.0106317 -.0040565 nlistings -.0079367 .0012727 -6.24 0.000 -.0104321 -.0054413 _cons -4.117136 .062445 -65.93 0.000 -4.239574 -3.994698

(31)

31 b. Table 1, Column (3) Number of observations = 3068 F (8,3059) = 152.2 Prob>F = 0.000 R-Squared = 0.5138 Adjusted R-Squared= 0.5126 Root MSE = 0.0166

princ Coef. Robust Std. Err. t P>t [95% Confidence Interval] Bathrooms .0090899 .0020764 4.38 0.000 .0050186 .0131612 Guests .0066115 .0006567 10.07 0.000 .0053239 .007899 Beds .0025153 .0005923 4.25 0.000 .0013538 .0036767 Sept.15 -.0027201 .0005896 -4.61 0.000 -.0038762 -.0015639 ins_Bathrooms_2 -.0001466 .000037 -3.96 0.000 -.0002192 -.0000741 ins_Guests_2 -.0000239 .0000197 -1.21 0.226 -.0000625 .0000148 ins_Beds_2 -.000011 .000029 -0.38 0.705 -.0000679 .0000459 nlistings .0001905 .0000254 7.50 0.000 .0001407 .0002403 _cons -.0042295 .0022872 -1.85 0.065 -.0087142 .0002552 c. Table 2 – Column (2) Number of observations = 3068 F (8,3059) = 107.05 Prob>F = 0.000 R-Squared = 0.4545 Adjusted R-Squared= 0.4531 Root MSE = 0.0176

princ Coef. Robust Std.

Err. t P>t [95% Confidence Interval] Bathrooms .009736 .0021348 4.56 0.000 .0055501 .0139218 Guests .0066464 .0006807 9.76 0.000 .0053117 .007981 Beds .0025745 .0006078 4.24 0.000 .0013827 .0037663 Sep-15 -.0023935 .0006271 -3.82 0.000 -.0036231 -.001164 ins_Bathrooms_2 .0001449 .0000529 2.74 0.006 .0000411 .0002487 ins_Guests_2 -.0002051 .0000334 -6.15 0.000 -.0002705 -.0001396 ins_Beds_2 .0002329 .0000252 9.23 0.000 .0001835 .0002824 nlistings 7.48e-06 2.01e-06 3.72 0.000 3.54e-06 .0000114 _cons .0024365 .0025143 0.97 0.333 -.0024934 .0073664 Number of observations = 3068 F (8,3059) = 141.41 Prob>F = 0.000 R-Squared = 0.3057 Adjusted R-Squared= 0.3039 Root MSE = 0.8364

(32)

32

ln_Sjg Coef. Robust Std.

Err. t P>t [95% Confidence Interval] Bathrooms .0238286 .0420937 0.57 0.571 -.0587062 .1063633 Guests -.0219204 .0164734 -1.33 0.183 -.0542204 .0103796 Beds -.0267662 .0175899 -1.52 0.128 -.0612555 .0077231 Sep-15 .3619546 .0306834 11.80 0.000 .3017924 .4221168 ins_Bathrooms_2 -.0063695 .0023937 -2.66 0.008 -.0110629 -.0016761 ins_Guests_2 .0005232 .0015937 0.33 0.743 -.0026016 .0036481 ins_Beds_2 .0020294 .0012862 1.58 0.115 -.0004925 .0045513 nlistings -.0020618 .0000937 -22.02 0.000 -.0022455 -.0018782 _cons -4.559599 .0734813 -62.05 0.000 -4.703677 -4.415521 d. Table 2 – Column (3) Number of observations = 3068 F (8,3059) = 107.05 Prob>F = 0.000 R-Squared = 0.4545 Adjusted R-Squared= 0.4531 Root MSE = 0.0176

princ Coef. Robust Std.

Err. t P>t [95% Confidence Interval] Bathrooms .009736 .0021348 4.56 0.000 .0055501 .0139218 Guests .0066464 .0006807 9.76 0.000 .0053117 .007981 Beds .0025745 .0006078 4.24 0.000 .0013827 .0037663 Sep-15 -.0023935 .0006271 -3.82 0.000 -.0036231 -.001164 ins_Bathrooms_2 .0001449 .0000529 2.74 0.006 .0000411 .0002487 ins_Guests_2 -.0002051 .0000334 -6.15 0.000 -.0002705 -.0001396 ins_Beds_2 .0002329 .0000252 9.23 0.000 .0001835 .0002824 nlistings 7.48e-06 2.01e-06 3.72 0.000 3.54e-06 .0000114 _cons .0024365 .0025143 0.97 0.333 -.0024934 .0073664

(33)

33

C. Post-estimation tests

a. Table 1, Column (2)

Test of overidentifying restrictions (stata code: estat overid)

• Score chi2(2) = 3.1066 (p = 0.2115)

Tests of endogeneity (stata code: estat endogenous) Ho: variables are exogenous

• Robust score chi2(2) = 991.975 (p = 0.0000)

• Robust regression F (2, 3059) = 4038.22 (p = 0.0000)

b. Table 1, Column (3)

• Score chi2(3) = 3.08849 (p = 0.3782) Tests of endogeneity (stata code: estat endogenous) Ho: variables are exogenous

• Robust regression F (1, 3061) = 8.36562 (p = 0.0039)

c. Table 2, Column (2)

(34)

34

d. Table 2, Column (3)

Market segmentation in Airbnb : an empirical study of competition based on Amsterdam’s neighbourhoods and property values