The Impact of Short-Time Rentals in the Demography of Touristic - - PDF document

1

The Impact of Short-Time Rentals in the Demography of Touristic Neighborhoods: the case of Barcelona

Joan SALES-FAVÀ (jsales@ced.uab.es) Antonio LÓPEZ-GAY (tlopez@ced.uab.es) Juan Antonio MÓDENES (juanantonio.modenes@uab.cat) Centre d’Estudis Demogràfics and Department of Geography (Universitat Autònoma de Barcelona). THIS IS A PRELIMINAR VERSION PREPARED FOR THE INTERNATIONAL POPULATION CONGRESS 2017.WORK-IN-PROGRESS PAPER. Abstract: Cities are becoming a preferred tourist destination and have recently experienced the emergence of new sources of accommodation. Besides the general increase in the number

• f urban hotel rooms, the change in the uses of dwellings (from residential to tourist) is

generating a significant impact on the housing system of the neighborhoods under high touristic pressure. These shifts could be strengthening new sociodemographic processes in the most touristic areas. Up to now this field is driving the attention of different social sciences that commonly put population displacement in the spotlight. In this paper we aim to explore the impact of short-term rentals, and the emergence of Airbnb, in Barcelona and its neighborhoods. Our main goal is to measure the direct displacement of households. This article analyses the relationship between households’ dynamics and their determinants, taking account space. For this reason, a Geographically Weighted Regression technique is employed to explore this relationship. The outcome of the paper provides evidence that the Short Term Rentals are significantly and negatively related to the evolution of the number

• f households. Other structural variables are introduced in the model in order to control

this parameter. Key Words: Tourism, Population, Households, Geographically Weighted Regression, Barcelona.

• 1. Introduction

In the last years, tourism has become an important economic sector. The consumption of services that provide leisure have become an increasingly important part of mass consumption (Bauman, 1988). As a result of this rapid growth we could find some evidence

• f the first externalities in some crowded touristy places. The literature about the

development of tourism and population dynamics have been focusing about the repercussions of this economic activity in rural areas and in low-density population areas. In these contexts authors usually associated the tourism expansion with population growth, demographic changes and relapse in ageing (Greenwood, 1972; Loukissas, 1982; Getz, 1986; Gill and Williams, 1994). Nevertheless, more recent studies have found some

2 negatives externalities of tourism activity on rural population (Smith and Krannich, 1998; Heberlein et al. 2002; Hines, 2010). Although the contemporary development of cities, at urbanistic and population levels, has been shaped by the rise of tourism there are only a few articles about the repercussions on population dynamics (Bock, 2015). Olds (1989) explained how hotel owners during the 1980's in Vancouver removed 2.000 housing in the process of converting them from residential to tourist use. This implied that hundreds of long-term resident were evicted. Additionally, we shall raise here the contribution of Gotham (2005) and Garcia Herrera et

• al. (2007) that explained the population changes of a city under high tourist pressure.
• 2. Short-Term rentals externalities

From an economic point of view, there are some relevant findings on the effect of Short- Term rentals in the hotel sector and in the housing market (Zervas et al. 2013; Sheppard and Udell, 2016). The expansion of new online platforms, such as Airbnb or HomeAway - which are pooling in apartments, or parts of these, to the tourist accommodation supply- is generating new dynamics in the housing market of the most touristic areas. The profitability

• f holiday accommodations leads to a shortage of housing supply for residential uses and

may be causing an increase in housing prices (Lee 2016; Schafer and Braun, 2016). These transformations could be strengthening new sociodemographic processes in the neighborhoods under high tourist pressure. Such areas are often experiencing ongoing gentrification processes (Gotham, 2005; García-Herrera et al., 2007; Quattrone et al., 2016). Gotham’s contribution on this issue appears to be specially relevant: “I develop and apply the concept of tourism gentrification as a heuristic device to explain the transformation of a middle-class neighbourhood into a relatively affluent and exclusive enclave marked by a proliferation of corporate entertainment and tourism venues” (Gotham, 2005; page 1102). In fact, the Short-Term rentals (STRs) appears to be already an old activity in some Spanish beach holidays resorts. But it is a recent phenomenon in cities like Barcelona, Madrid or Bilbao (linked with Airbnb emergence). The expansion of this kind of establishment has given rise to competition in the tourist market (Zervas et al. 2016). In addition, the STR's have affected the housing market and especially the rental market (Duatis et al., 2016; Consultora EY, 2016; Lee, 2016; Llop, 2016). From a theoretical point of view, STRs inside in to different ways into demographic reduction (from a household or individual level). From a household perspective: any given dwelling that was previously the residence of a family could now end up being a tourist flat

• ffered at Airbnb. Thus, the home would be been eliminated from the conventional

residence stock. On the other hand, from an individual scope: the owner or the tenant of a dwelling that is seeking a flat mate has an economic stimulus when he or she rents a room at Airbnb, which is greater than doing so with a neighbor from town (Lee, 2016). The main goal of this article is to measure the first case, the reduction of the number of households induced by the Airbnb expansion.

3

• 3. Tourism, housing and households: recent trends in Barcelona.

This paper focuses on the case of Barcelona, the third city in Europe by number of listings in Airbnb, and one of the most visited1, with almost 32 million tourist overnight stays in 20162 (20 times their population). In ten years, the number of overnight stays have tripled. The dramatic increase of tourism has raised complaints and even adverse reactions. In 2016, according to a local council survey on the resident’s evaluation of public services in Barcelona (Ajuntament de Barcelona, 2016), tourism was said to be the second main problem of the city (following unemployment). Moreover, in 16 of the 73 neighborhoods of the city, tourism ranks #1 as the biggest concern for the residents (especially in central areas). Figure 1 shows the number of tourists that have visited Barcelona since 2001 (by the number of visitors and number of overnights) and the evolution of the tourist

• accommodations. Since 2001 and until 2007 the number of hosting beds in Barcelona has

increased by 48%. Barcelona saw a significant increase in visitors and tourist supply after the entrance of low cost carriers (Jones Lang Lasalle, 2006). However, it has been between 2009 and 2016, and particularly in 2012, 2013 and 2014 when the number of accommodations had increased swifter. In this last period, the supply has grown by 74,521

• places. This growth was stimulated by an increase in the number of tourists. Barcelona

received 6,480,051 tourists in 2009, seven years later the number increased up to 12,715,000 (2016). Since 2001, Barcelona have experienced high increase of the housing stock. 67,293 dwellings have built until 2016 (Figure 2). The dwelling stock has grown progressively during the period 2001-2016. Additionally, the number of households had risen rapidly until 2010, at the same time as the international arrivals wave (Bayona i Carrasco, 2007). In

1 According to the MasterCard Global Destinations Cities Index (2016), Barcelona ranks 12th in the World and

5th in Europe (following London, Paris, Berlin and Rome) on international overnight visitors.

2 According to the official Tourism Barcelona Agency (www.barcelonaturisme.com) and Hotel Occupancy Survey

from INE (National Institute of Statistics). 36,000 72,000 108,000 144,000 8,000,000 16,000,000 24,000,000 32,000,000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 Beds Travellers and overnights stays Year

Travellers Overnights stays Places Figure 1: Tourist activity in Barcelona since 2001. Beds Source: Inside Airbnb, Tourism register of Catalonia and Xarxa Catalana d'Instal·lacions Juvenils. Travellers and overnight stays Source: Hotel Occupancy Survey and Business Survey of the Hotel Sector.

4 the period 2010-2016 the number of households has started to decrease, even though housing units had increased. Hence, after 2010 we are in a new phase where the classical paradigm between dwellings and households has been broken down. Therefore, we have two simultaneous processes at the end of the decade: The decrease of households stock and the rise of the tourism activity. As a consequence of the emergence of companies like Airbnb, we should expect an impact on the family dwellings and the number

• f households.

We present Figure 3 as an example of this relation between households and Airbnb. It shows the relative evolution of households stock and the increase of the Short-Term Rentals3 in each one of the 233 Basic Statistical Areas tracts (AEB)4. Thick black boundaries correspond to the 10 districts of the city, each one with a reference number. The strong pink colour represents the tracts where tourism pressure is intense and where a large drop of households has been registered since 2010. This colour appears to be predominant in the inner city (1) and has a high representation in the rest of tourist districts of the city (2, 3 and 6), except in district 10 and some sections of district 2 and 3. The later areas could have not been losing households since their holiday dwellings were set up in substitution of old

• ffices. In the case of district 10 the conventional housing construction was accompanied

with the expansion of the tourist market in an area where there was plenty of vacant land, as a result of the change of use from industrial to residential and urban densification. From a general point of view, decline in household numbers and rise of STR's happen to meet in the same tracts. Almost 50% of the tracts with high pressure of tourist dwellings (values

3 For more information about these variables see below, in the chapter 5.3. Data. 4 These are the intermediate level between neighbourhoods and census tracts. In Barcelora there are 233 AEB.

The AEB is the geographical unit of analysis of our study 590,000 605,000 620,000 635,000 650,000 665,000 775,000 787,000 799,000 811,000 823,000 835,000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 Households Dwellings Year Dwellings Households

Figure 2: Evolution of households and Dwellings, 2001-2016, City of Barcelona. Source: Spanish Population Register and Cadastre.

5 higher than 2.1%5) have lost more than 1.4% of households, and only 20% of tracts with high pressure increase 0.5% or higher their households. In Figure 46, one can see the relation between the households evolution and housing units variation between 2010 and 2016 by AEB, with dots colour according to Short Term-Rentals pressure (above 2.1% is high). The horizontal axis shows the relative evolution of

• households. The Vertical axis displays the relative growth of dwellings stock. General trend

exposes a positive relation between both variables, household growth rate tends to be higher in those areas where dwelling stock have risen. Even though most part of areas have increased dwellings units, there are some areas that have lost households (top left box). Among all areas that have increased the dwellings stock, 48% of areas have decreased the number of households. In areas with high tourist pressure this percentage swell until 64%. In the rest of areas without high pressure (green triangles) only the 40% have lost households and have increased dwellings. In addition, among those areas that have reduced the number of households, 63% have a high pressure and 37% low pressure of tourism7. To summarise this section; the period 2010-2016 have seen an increase of tourism. The number of visitors has risen and the tourist market has expanded (triggered by Airbnb). At

5 Above these value there are one thirds of the total tracts. 6 The scatter plot shows only the values above -10 and below 10. 12 AEB tracts are out of this plot. 7 After apply the weights of two thirds to the high pressure area and one third to the low pressure

area. Figure 3: Bivariate map between households evolution and tourist dwellings pressure in Barcelona (2010-2016). Source: Population register, Cadastre and insideairbnb.

6 the same time, in 2010 Barcelona had finished their phase of household expansion and since 2010 has started to decrease the number of family units, especially in those areas with more Airbnb pressure. In this relation seems that housing plays an important role but is interfered by their uses (residential or tourist).

• 4. Objectives

A major objective of this paper is to make a contribution to the empirical literature about tourism externalities on population. In the paper, we aim to explore the impact of the diffusion/increase of STRs on population stock. In order to calibrate the STRs strength we examine the relation between different housing variables. Being Barcelona a known case of high pressure of STR’s, the article will contribute to the expanding knowledge on: (i) The new relationships, as in many European cities, between tourism and population. (ii) The impact of tourism in household decline and direct population displacement. (iii) We will try to demonstrate if the population register is able to capture these types of processes at high geographical detail.

Figure 4: Scatter plot between relative evolution of households and relative evolution

• f dwellings coloured by Airbnb pressure in Barcelona (2010-2016). Source:

Population register, Cadastre and insideairbnb.

7

• 5. Methodology and Data

5.1 Study site: Barcelona is the Spanish city with more Airbnb listings. According to inAtlas data8 during the weekend of 25-26 of March (2017), there were 55.768 beds in available listings (gathered in 15,831 listings), of which 49,000 were booked. Examining Airbnb data9 we will see that a large part of the listings offered in the city of Barcelona are Short-Term rentals of whole apartments (more than 50% of the listings). In addition, the users that control 2 or more listings are the 53.7% of all the users that offer an apartment in Airbnb. 5.2 Spatially varying relations: The relation between household evolution rate and underlying determinants is important in order to understand the influence that play STRs. Our study assumes the Tobler's first law of geography (1971) "everything is related to everything else, but near things are more related than distant things". According to this, we suppose an analysis with spatial data requires specific geographical techniques. Multiple regression analysis is based on the assumption of independence in observations, and it usually fails when is put to work with spatial data. In order to find a solution for this, Fotheringham et al. (2002) developed a local regression technique: geographically weighted regression (GWR). This Method, designed to explore spatial non-stationarity, defined when the nature and significance of relationships between variables differ from location to location. GWR is a local modelling approach that explicitly allows parameter estimates to vary over space (Brunsdon et al. 1996). Rather than specifying a single model to characterize the entire housing market, GWR estimates a separate model for each household growth rate level polygon and weights observations by their distance to this polygon. Geographically weighted regression is often compared to Ordinary least squares (OLS) because it illustrates the benefits of using a spatial non-stationarity approach in statistical

• models. In contrast to the OLS approach, GWR can construct a separate OLS equation for

every location in the data set, which incorporates the dependent and explanatory variables

• f locations falling within the bandwidth of each target location. In estimating each region’s
• wn regression, the characteristics of the individual areas included in the sub-sample are

weighted by their spatial proximity (Gilbert and Chakraborty, 2010). Despite its advantages, the statistical method applied, GWR, also has some limitations that need to be raised. Wheeler and Tiefelsdorf (2005) found that GWR is weak in the ability to differentiate between spatially stationary processes and varying ones, calling into question the values of the local estimators. In addition, the results are highly dependent on the

8 In Atlas is a Company that crawled Airbnb listings every week, amongst other things. They have published on

twitter a summary of the Airbnb listings where is possible to see results of occupancy, active offer and inactive

• ffer. They did this every weekend during the last year.

9 The listings are publicly accessible (without requiring an account) on Inside Airbnb (released by Murray Cox)

for many major cities scraped Airbnb listing data, including some individual information about the hosts, reviews, location of each listing. The data of Airbnb website was crawled 6 times (between 2015 and 2016) for Barcelona.

8 selection of the bandwidth for the weighting function. The choice of an adaptive or fixed bandwidth is a decisive decision that could greatly influence the results. The analytical utility of GWR has been demonstrated in different fields of social and environmental sciences: In essays about migration flows (Nelson, 2008; Helbich and Leitner, 2009; Jivraj et al., 2013; Villarraga et al., 2014), public policy (Gutiérrez-Puebla et al., 2012), housing market prices (Bitter et al., 2007), climatology studies (Ivajnsic et al., 2014) or even ecologic disasters (Oliveira et al., 2014). In an urban context, where the location of services, housing and households is relevant, it is important to assume the spatial non-stationary in relationships between household evolution and the explanatory variables. 5.3 Data: The proposed work is mainly based on data from the Spanish Population Continuous Register (INE and Statistics Department of Barcelona’s city council) from 2010 to 2016. The Municipal Register is the administrative registry where the municipality neighbours are

• recorded. Its data is the proof of residence in the municipality and usual domicile in the
• same. We have used the information about the number of households. For the number of

dwellings we have used the land registry (cadaster). These show the number of the premises dwellings according to antiquity. All these data are available at AEB level. 5.3.1 Dependent Variable Household growth rate (ℎℎ_𝑓𝑤𝑝𝑗): The relative evolution of households stock between 2010 and 2016 was calculated and used as the dependent variable. ℎℎ_𝑓𝑤𝑝𝑗 Refers to the evolution of households that has occurred in a specific area i during the period 2010-2016. 𝐼2016𝑗 is the total households in 2016 in area i and 𝐼2010𝑗 is the total households in 2010 in the same area. k is the constant term (100). ℎℎ_𝑓𝑤𝑝𝑗 = 𝐼2016𝑗 − 𝐼2010𝑗 𝐼2010𝑗 ∗ 𝑙 The series of the Spanish Population Register was corrected for 2010 in a specific area of the city centre where the Statistical Department of Barcelona office is established. In this area, there was an overrepresentation of population and households without permanent address between 2005 and 2011 (Bayona i Carrasco, 2007; López-Gay and Cócola Gant, 2016). Martí (2016) and López-Gay and Cócola Gant (2016) corrected this anomaly, in the same way, with a good resource that identifies the over representation of households with 9 or more members as part of this over registration. 5.3.2 Explanatory variables Short-Term Rentals rate (𝑇𝑈𝑆𝑡): The data about STRs are deployed from Inside Airbnb (www.insideairbnb.com). Inside Airbnb have crawled the listings of Airbnb for Barcelona 6

9 times10 (between 2015 and 2016). Airbnb listings are georeferenced. Nevertheless, location information for listings is anonymized by Airbnb. This means that the location for a listing

• f the data will be up to 150 meters away from the real address. It is thus impossible to

gather points at the level of census tract. Usually, a census tract has 1.28 hectares and the listing error location could be up to 2.25 hectares. Because of it, the AEB, that has a mean area of 11.5 hectares, fits better to gather georeferenced points of Airbnb. Our assumption on bookings distinguishes casual from “commercial” short-term rental hosts for entire units. A commercial host is the one that practices short-term renting as a business instead of listing a unit on the long-term rental market. In our study commercial hosts are those renting their housing unit more than 3 times every month and who received more than 6 bookings11. In total there are 6,706 commercial entire unit listings at Airbnb and 5,255 casual hosts for entire units that in some moment rented their flat during the specified period. We only take into account those apartments that have started their activity in Airbnb between June of 2010 and June of 2016. Why do we use this data and not the official register? In 2011 the municipality of Barcelona stopped granting licenses for holiday dwellings in the city centre and in 2014 in the whole city. This ban activated a substantial increase of STRs black market, especially in the city centre. Moreover, before the first January of 2010 (the starting point for the study) STRs market had already started to be popular in Barcelona12. The official register does not allow to know the temporal moment that each Tourist Dwelling started their activity. For this reason, we shall take account listings of Inside Airbnb in order to detect parts of the black market and parts of the official holiday dwellings that started their activity after 2010. In our study we are taking into account those 6,706 commercial entire unit listings in airbnb that have started their activity after June of 2010 and before June of 2016. At the theoretical level this is the main variable of the study. 𝐵𝑗𝑠𝑐𝑜𝑐𝑗 refers to the number of tourist apartments in a specific area, i. 𝐸𝑋

2016𝑗 is the total number of dwellings in 2016 in the same

area, i. k is the constant term (100). 𝑇𝑈𝑆𝑡𝑗 = 𝐵𝑗𝑠𝑐𝑜𝑐𝑗 𝐸𝑋

2016𝑗

∗ 𝑙 New-Build Housing rate (𝑂𝐶𝐼): This variable counts the relative evolution of the new build housing market. The ratio of new housing explains well the arrivals of families in a specific

• area. 𝑂𝐸𝑋

𝑗 refers to the housing stock built during the period 2010-2016 in a specific area,

• i. 𝐸𝑋

2010𝑗 is the total number of dwellings in 2010 in the same area, i.

𝑂𝐶𝐼𝑗 = 𝑂𝐸𝑋

𝑗

𝐸𝑋

2010𝑗

∗ 𝑙 Old-Build Housing rate (𝑃𝐶𝐼): This variable explains the relative evolution of the old build housing market. 𝑃𝐸𝑋

𝑗 refers to the housing stock evolution built before 2010 in a specific

10 The listings in six moments; April 2015, July 2015, September 2015, October 2015, January 2016 and

December 2016.

11 According to correction weight using reviews rate of 30.5 percent was applied on the number of reviews per

listing reported by Budget and Legislative Analyst’s Office (2015).

12 At 2008 there was already 2,537 tourist dwellings (Ajuntament de Barcelona, 2008)

10 area, i. 𝐸𝑋

2010𝑗 is the total number of dwellings in 2010 in the same area, i. Negative values

indicates housing demolition, positive values indicates subdivision of houses. 𝑃𝐶𝐼𝑗 = 𝑃𝐸𝑋

𝑗

𝐸𝑋

2010𝑗

∗ 𝑙 5.4 Regression methodology: The starting point to applicate GWR model or Semiparametric Geographically weighted regression (SGWR) is the OLS multiple regression equation. The OLS expresses the relationship between the dependent variable and a combination of independent variables simultaneously in a single model. An ordinary linear regression model can be expressed by: 𝑧𝑗 = 𝛾0 + ∑ 𝛾𝑙 ∗ 𝑦𝑙

𝑛 𝑙

+ 𝜁 with 𝑧𝑗 as a vector of the dependent variable, 𝑦𝑙 as a vector of the 𝑙 independent variable and 𝜁 as a vector of the error term. In a semiparametric Geographically Weighted Regression some of the global regression coefficients are replaced by local parameters: 𝑧𝑗 = 𝛾0(𝑣𝑗, 𝑤𝑗) + ∑ 𝛾𝑙(𝑣𝑗, 𝑤𝑗)𝑦𝑙𝑗

𝑛 𝑙

+ ∑ 𝛿𝑚𝑨𝑚𝑗 + 𝜁𝑗

𝑛 𝑚

where 𝑧𝑗, 𝑦𝑙𝑗 and 𝜁𝑗 are, respectively, dependent variable, kth independent variable, and the error at location i; (𝑣𝑗, 𝑤𝑗) is the x/y coordinate of the i location; and coefficients 𝛾𝑙(𝑣𝑗, 𝑤𝑗) are varying conditionals on the location (Fotheringham, 2002). 5.5 Bandwidth selection for Geographical Weighting: The results of SGWR are sensitive to the choice of bandwidth, as weighting procedures that specify a wide bandwidth and allow for only minimal distance decay will produce results that are similar to OLS model (Pineda Jaimes et al., 2010). The contrary, if the bandwidth is narrow only points in close proximity will be considered, which will lead to high variances in the estimators (Fotheringham et al., 2002). An adaptive bandwidth was chosen for the current analysis, among other reasons, because some polygons are spatially dispersed. We chose to use an adaptive spatial kernel that allows the bandwidth to vary based on the density of AEB polygons, thus encapsulating a smaller area where there are a lot of polygons and a larger area where data the polygons are sparse. The number of observations to retain within the weighting kernel is irrespective

• f distance. Adaptive models are able to adapt themselves in size to variations in the density
• f the data so that the kernels have larger bandwidths where the data are sparse and have

smaller ones where data are plentiful. In Bi-squares models when the distance is greater than the bandwidth, the weight is zero. The Akaike Information Criterion correction (AICc) criterion, which optimizes the choice, was used to select the bandwidth. The resulting bandwidth is 81 AEB. The AICc estimation can also be used to compare whether the results

11 from SGWR present a better fit than the global model or GWR normal model, taking models’ degrees of freedom into consideration (Fotheringham et al. 2002).

• 6. Comparison between OLS and GWR models

Before executing the SGWR model we have developed an OLS model (Table 1) using hh_evo as a dependent variable and all the variables inside "explanatory variables" section as independent variables (STRs, NBH and OBH). Inside brackets we have the estimates values after standardization by the standard score formula (z-scores)13. This allows the comparison between parameters. The multicollinearity test of explanatory variables, the Variance Inflation Factor Test (VIF), proved that the predictors were not redundant. This indicates that each of the predictors has a different influence on the dependent variable. The diagnostics of the OLS model show that overall the OLS model is statistically significant. Table 1 includes the R-squared and AICc values from the OLS model as well as parameter

• ther indicators. Moreover, around 68% of the variation in the household growth rate is

explained by this model according to the adjusted R², a reasonable global fit. The t-statistics and the p-value of the estimated parameters indicate that all explanatory variables are

• significant. But the effect and sign level appears to be different. STRs have a negative impact
• n the increase of households, ceteris paribus. Meanwhile, an increase in the NBH, OBH

variables has a positive impact in hh_evo, ceteris paribus. The variable that have the most influence on the dependent variable is NBH. However, is expected to find spatial autocorrelation between OLS model residuals. This was additionally confirmed with the statistically significant Global Moran's Index test, which is a stronger indicator for spatial non-stationarity among predictors and response variables (Anselin, 1988).

13 All the values are standardized by subtracting the mean to each specific parameter and diving the result by

the standard deviation. The result is the z-score and tells us how many standard deviations we have from our mean score.

Table 1 Diagnosis of the OLS analysis OLS regression results Variable Estimate Standard error t(EST/SE) p value VIF Intercept

• 1.862

0.287

• 6.473

5.77e-10*** STRs

• 0.418 (-0.561)

0.161

• 2.812

0.008** 1.12 NBH 1.162 ( 5.223) 0.053 21.534 < 2e-16*** 1.96 OBH 0.563 ( 1.406) 0.097 5.801 2.17e-08*** 1.31 OLS Diagnostic information R2 0.680 p-value < 2e-16*** Adjusted R2 0.676 Observations 233 AICc 1233.052 Residuals Moran’s Index 0.051 Moran Index Probability (p-value) 0.000*** Koenker (BP) Statistic Probability (p-value) 0.343

• Signif. Codes (p<): 0.001‘***’ 0.01‘**’ 0.05‘*’

12 All variables were tested for geographical variability test according to AICc. GWR4 software (Nakaya et al., 2016) reports if the explanatory variables have or do not have spatial

• variability. These will indicate which coefficients have spatial heterogeneity and which have

spatial stationarity. When "Diff of Criterion" is positive, it suggests that there is a no spatial

• variability. On the contrary, when this indicator shows negative values, it denotes spatial

heterogeneity.

Table 2 Geographical variability tests of local coefficients Variable F DOF for F test DIFF of Criterion Intercept 3.352 5.721 209.091

• 6.416

STRs 1.129 1.958 209.091 4.123 NBH 5.941 5.981 209.091

• 21.919

OBH 5.552 5.265 209.091

• 17.547

Note: A positive value of diff-Criterion (AICc, AIC, BIC/MDL or CV) suggests no spatial variability in terms of model selection criteria. F test: in the case of no spatial variability, the F statistics follows the F distribution of DOF for F test.

As seen in Table 2 STRs exhibited no spatial variability levels. Hence, NBH and OBH variables have local variability. Following this, as a result of the variability test, we developed a semiparametric GWR method where NBH, and OBH worked as local coefficients. On the

• ther hand, STRs were selected as global independent variable. And this was in contrast

with the GWR where all the independent variables have a local incidence. Table 3 presents a comparison between OLS, GWR and SGWR models. The SGWR shows significant improvement over the OLS and GWR models. First, the AICc indicates that the SGWR model has a much better goodness-of-fit than the OLS and GWR models. Second, the increase of adjusted R2 suggests the SGWR model has much better performance in exploring the relationships between the dependent variable and explanatory variables than the global model and the GWR model.

Table 3 Summary statistics for OLS, GWR and SGWR models Model Bandwidth size AICc R2 Adjusted R2 OLS NA 1233.052 0.680 0.676 GWR 81 AEB 1189.036 0.782 0.749 SGWR 81 AEB 1185.279 0.784 0.751

• 7. SGWR results

The SGWR generates a set of parameter estimates of explanatory variables (that have a local incidence) for each land use sample point, which can be used to analyse spatial variations

• f the effects of households decrease. In addition, a t-statistic is also calculated to indicate

the significance of the parameters, which can be obtained by dividing a parameter estimate

13 by its standard error14 (Fotheringham et al., 2002). Local estimated coefficients and the local R2 are presented in the next figures (5a, 6a and 6b). The SGWR results are quite different, in comparison with OLS results. For example, OLS regression estimated the parameter for the New-Build Housing (NBH) to be significant and equal to 1.162 for the whole of Barcelona (Table 4). Nevertheless, the local parameter for this variable range from -0.132 to 1.525 across AEB of Barcelona, with a mean of 0.946. This variability in the coefficient suggests that the relationship between the NBH and variation

• f household stock is not stationary. Similar results can be observed in Table 4 for our other

model coefficients.

Table 4 Summary of results and comparison between OLS and SGWR Variable OLS Coefficient SGWR Coefficients Percent of census tracts by significance (95% level) of t-statistic Min Mean Max t ≤ -1.96*

• 1.96 < t < 1.96

t ≥ 1.96* Intercept

• 1.862***
• 2.757
• 1.510

0.285 69.52% 30.47% 0.00% NBH 1.162***

• 0.132

0.946 1.525 0.00% 23.17% 76.82% OBH 0.563***

• 0.477

1.031 2.687 0.00% 56.22% 43.77% STRs

• 0.418**
• Signif. Codes (p<): 0.001‘***’ 0.01‘**’ 0.05‘*’

The local R2 show a dissimilar distribution within each region (Figure 5a). Geographic variations in R2 demonstrate how the combined statistical effect of our explanatory variables on hh_evo differs across the city. Higher values are concentrated in two spots. One is located between the district 1, 2 and 10. The other one is located in the northern part of the city (mountain side) and affects the district 6, 7and 8. Inversely, the main spot with lower values is concentrated in some parts of the district 3, 4 and 5 (the lighter colours). The spatial distribution of hh_evo is depicted in Figure 5b. It seems that there is no geographical concordance between the variables represented in Figure 5a and Figure 5b. We can suppose that SGWR model explains similarly the reduction and increase of household variation. Low local R2 values coincide with some of the areas where residential flow dynamics are more intensive15 and there are more housing units used for non- residential uses (such as offices, studies, etc.). Both elements could distort the model capacity for detecting household movements. Among the three variables, STRs is the variable with lesser power to explain the changes of hh_evo. But even so has a significant and negative relation with the dependent variable, notwithstanding, in comparison with the rest of independent variables STRs does not have a spatial heterogeneity. According to SGWR model the influence of STRs is global.

14 Concerning the individual significance of the parameters, although they are all statistically significant in the

global regression, there are several non-significant parameters in the SGWR models. A pseudo t-test can be put in practice dividing each local estimate by the corresponding local standard error of the estimate. Standard t- values of ±1.98 are used to represent the 95% level threshold for negatively and positively significant relationships.

15 The highest unregistered citizens values are concentrated, in general, in the areas with low R2. In these areas

the percentage of unregistered inhabitants in their housing are higher than the mean of Barcelona (except for some specific neighborhoods). All this information is based on data of the local council survey of 2016 (Ajuntament de Barcelona, 2016).

14 In order to illustrate the spatial variation in the effects of the housing variables, the SGWR coefficients are mapped in the figure 6a and 6b. For the colour-scheme, greens were used to indicate negatives parameters and brownish colours for positive parameters. The darkest tonalities indicate values above the estimator given by the OLS model. Areas where pseudo t-value are not significant (at 95%) are covered in white. There is a positive relationship between NBH and the dependent variable. The positive effect is especially strong in parts of district 1, 2 and in the whole district 10. On the contrary, it is not significant in some parts

• f the district 2, 3 5 and 6 and in the district 4. Nevertheless, in general terms the

construction of new housing units is related to households increases. Figure 6b shows the importance of old housing stock on households increase and decline. Some areas of Barcelona have experienced demolition of old housing stock and another areas have subdivided part of the housing stock. As expected, the estimates are positive and significant in some parts of Barcelona. The highest OBH estimates are found within the city

• center. In this part of the city the subdivision of old dwellings have been a reiterated activity

by the real estate agencies and individual owners. There are also high values in the northeast of the city (districts 8, 9 and 10). This area has characterized to have poor-quality housing buildings and some were demolished in the period 2010-2016.

Figure 5: a) SGWR model performance in Barcelona distribution of local R squared values by AEB. b) hh_evo distribution at AEB level.

15

• 8. Conclusions

As stated previously, the purpose of this paper is to analyse the local geography of the relationship between the evolution of the number of households and their determinants in Barcelona, especially the Short-Term Rentals. This paper established that the results of the OLS model were insufficient for drawing conclusions about the relationship between household growth rate and their causes. The OLS model has not the capacity to take into account the geographical context within which properties are located. To correct this, our further analysis determined that a spatial model would be a better tool for analysing the

• relationship. Therefore, the SGWR model was employed to analyse more directly the

relationship locally. This paper has explored the spatial variation in the effect of Short-Term Rentals expansion, new-build housing and old-build housing evolution stock on household number evolution from AEB in Barcelona using a standard Ordinary Least Squares regression model and a Geographically Weighted Regression model. We found that SGWR model can significantly improve the global regression model. SGWR has a much better goodness-of-fit than OLS, and has better performance in exploring the relationships between the evolution of the number of households and explanatory variables than the global model. The results provided an improved explanation of data variance compared with the global model (OLS), going from 0.671% to 0.751% in the SGWR model and solves the problems related to the autocorrelation of residuals. Moreover, we found that our local R2 has a high variation level. As it has been explained above, low local R2 values correspond to areas with intensive population flows and a high percentage of housing units used as offices or studies. The SGWR allows the model estimators to vary across Barcelona, which provides deep insights into the spatial variations of decrease and increase of households. It has been

Figure 6: a) Spatial variation in the effect of NBH on hh_evo: SGWR coefficients b) Spatial variation in the effect of OBH on hh_evo: SGWR coefficients

16 demonstrated that the spatial variability of each factor influences the dependent variable. The conclusions of this paper are confined to an urban densified context where the housing is a limited and valuable good. In terms of its overall findings, the results of the SGWR model demonstrated that new-build housing played an important role to make the households arrivals possible. The results of the analyses also indicated that the effect of the new housing varies across Barcelona, but almost always is significant and always appears as positive. The results of the SGWR model clearly demonstrated that old housing evolution rate and evolution of households are positively correlated (housing demolition implies households decrease). Besides this, the relationship varies across the geographical area. Our main hypothesis in this study was that Short Term Rentals appearance have reduced or have limited the number of households, after controlling other structural variables. For this reason the findings of our research match with our initial hypothesis. In Conclusion, it could be said that the appearance on the urban scene of this new kind of tourist accommodation affects the number of households, and this affect does not has local divergences. Our investigation is a relevant in the study of externalities of tourism pressure in cities from a quantitative data and geographical methodology perspective. It also proves that it is possible to study population dynamics, with the Continuous Register, and their determinants, at a high geographical detail. Even so, for future research is important to study how have changed the population structure of the neighbourhoods with high tourist

• pressure. Along with its contribution to knowledge in the field of tourism externalities in

the urban context, these results could be relevant to those who are in charge of tourism policies in Barcelona.

• 9. References:

Ajuntament de Barcelona (2016). Enquesta de Serveis Municipals 2016, Informe general. Departament d’Estudis d’Opinió. Gabinet Tècnic de Programació. Anselin, L., (1988). Spatial Econometrics: Methods and Models. Kluwer Academic, Dordrecht, the Netherlands. Bauman, Z. (1988). Sociology and postmodernity. Sociological Review, 36, 790-813. Bayona-Carrasco, J. (2007). La segregación residencial de la población extranjera en Barcelona:¿ una segregación fragmentada?. Scripta Nova. Revista electrónica de geografía y ciencias sociales, 11(229- 255). Bitter, C., Mulligan, G. F., & Dall’erba, S. (2007). Incorporating spatial variation in housing attribute prices: a comparison of geographically weighted regression and the spatial expansion method. Journal

• f Geographical Systems, 9(1), 7-27.

Bock, K. (2015). The changing nature of city tourism and its possible implications for the future of cities. European Journal of Futures Research, 3(1), 20. Brunsdon C, Fotheringham S, Charlton M (1996). Geographically weighted regression: A method for exploring spatial nonstationarity. Geographical Analysis 28: 281-298. Consultora EY (2016). Impactos derivados del exponencial crecimiento de los alojamientos turísticos en viviendas de alquilar en España, impulsado por los modelos y plataformes de comercialización P2P. Estudio encargado por Exceltur. Duatis, J., Buhigas, M., & Cruz, H. (2016). Impacte del lloguer vacacional en el mercat de lloguer residencial de Barcelona. Pla Estratègic de Turisme Barcelona 2020. Direcció de Turisme Gerència d’Empresa i Turisme. Setembre 2016.

17

Feijten, P., & Mulder, C. H. (2002). The timing of household events and housing events in the Netherlands: A longitudinal perspective. Housing studies, 17(5), 773-792. Fischer, M. (2006). Spatial analysis in geography. In M. Fischer (Ed.), Spatial analysis and

• GeoComputation. Selected essays (pp. 17–28). Berlin: Springer.

Fotheringham, A.S., Brunsdon, C., Charlton, M.E., (2002). Geographically Weighted Regression: The Analysis of Spatially Varying Relationships. John Wiley & Sons Ltd., Chichester, UK. García Herrera, L. M., Smith, N., Mejías Vera, M. A. (2007). Gentrification, displacement, and tourism in Santa Cruz de Tenerife, Urban Geography, 28/3, 276-298. Getz, D. (1986). Tourism and population change. Scottish Geographical Magazine, 102(2), 113–126. Gilbert, A., & Chakraborty, J. (2011). Using geographically weighted regression for environmental justice analysis: Cumulative cancer risks from air toxics in Florida. Social Science Research, 40(1), 273- 286. Gill, A. & Williams, P. (1994). Managing growth in mountain tourism communities. Tourism Management, 15, 212–220. Gotham, K.F. (2005). Tourism gentrification: The case of New Orleans' vieux carre (French Quarter), Urban Studies, 42/7, 1099-1121.

• Greenwood. D.J. (1972) Tourism as an Agent of Change: A Spanish Basque Case. Ethnology 11(1), 80-

91. Gutiérrez-Puebla, J., García-Palomares, J., & Cardozo, O. (2012). Regresión geográficamente ponderada (GWR) y estimación de la demanda de las estaciones del Metro de Madrid. In Ponencia presentada en el XV Congreso Nacional de Tecnologías de la Información Geográfica (pp. 19-21). Heberlein, T. A., Fredman, P., & Vuorio, T. (2002). Current tourism patterns in the Swedish mountain

• region. Mountain Research and Development, 22(2), 142–149.

Helbich, M., & Leitner, M. (2009). Spatial Analysis of the Urban-to-Rural Migration Determinants in the Viennese Metropolitan Area. A Transition from Suburbia to Postsuburbia? Applied spatial analysis and policy, 2(3), 237–260. Hines, J. D. (2010). Rural gentrification as permanent tourism: the creation of the ‘New’West Archipelago as postindustrial cultural space. Environment and Planning D: Society and Space, 28(3), 509-525. Ivajnšič, D., Kaligarič, M., & Žiberna, I. (2014). Geographically weighted regression of the urban heat island of a small city. Applied Geography, 53, 341-353. Jivraj, S., Brown, M., & Finney, N. (2013). Modelling Spatial Variation in the Determinants of Neighbourhood Family Migration in England with Geographically Weighted Regression. Applied Spatial Analysis and Policy, 6(4), 285-304. Jones Lang Lasalle (2006). Low Cost Carriers and their Impact on European Tourism. Kulu, H. (2008). Fertility and spatial mobility in the life course: evidence from Austria. Environment and planning A, 40(3), 632-652. Lambea Llop, N. (2017). A policy approach to the impact of tourist dwellings in condominiums and neighbourhoods in Barcelona. Urban Research & Practice, 10(1), 120-129 Lee, D. (2016). How Airbnb Short-Term Rentals Exacerbate Los Angeles's Affordable Housing Crisis: Analysis and Policy Recommendations. Harv. L. & Pol'y Rev., 10, 229. López-Gay, A., & Cócola Gant, A. (2016). Cambios demográficos en entornos urbanos bajo presión turística: el caso del Barri Gòtic de Barcelona. Población y territorio en la encrucijada de las ciencias sociales, 380-394. Loukissas, P. J. (1982). Tourism's regional development impacts: A comparative analysis of the Greek

• islands. Annals of Tourism Research, 9(4), 523-541.

Martí-Comas, P. (2016). Competition between tourism and residential uses in Ciutat Vella. Master

• thesis. Polytechnic University of Catalonia.

Mastercard (2016). MasterCard Global Destination Cities Index. Referred on, 5, 2016. Nakaya, T., Charlton, M., Fotheringham, S., & Brundson, C. (2016). GWR4 version 4.0.80. Application for geographically weighted regression modelling. GWR 4 Development Team. 24 March 2016.

18

Nelson, P. (2008). Life-course influences on nonearnings income migration in the United States. Environment and Planning A, 40, 2149–2168. Olds, K. (1989). Mass evictions in Vancouver: the human toll of Expo ‘86. Canadian Housing, 6(1), 49- 53. Oliveira, S., Pereira, J. M., San-Miguel-Ayanz, J., & Lourenço, L. (2014). Exploring the spatial patterns

• f fire density in Southern Europe using Geographically Weighted Regression. Applied Geography, 51,

143-157. Pineda Jaimes, N. B., Bosque Sendra, J., Gómez Delgado, M., & Franco Plata, R. (2010). Exploring the driving forces behind deforestation in the state of Mexico (Mexico) using geographically weighted

• regression. Applied Geography, 30, 576–591.

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 (1385-1394). Schäfer, P. & Braun, N. (2016). Misuse through short-term rentals on the Berlin housing market, International Journal of Housing Markets and Analysis, 9/2, 287-311. Schäfer, P., Schäfer, P., Braun, N., & Braun, N. (2016). Misuse through short-term rentals on the Berlin housing market. International Journal of Housing Markets and Analysis, 9(2), 287-311. Sheppard, S., & Udell, A. (2016). Do Airbnb Properties Affect House Prices? (No. 2016-03). Smith, M., & Krannich, R. (1998). Tourism dependence and resident attitudes. Annals of Tourism Research, 25(4), 783–802. Tobler, W.R., 1970. A computer movie simulating urban growth in the Detroit region. Economic Geography 46, 234–240. Villarraga, H. G., Sabater, A., & Módenes, J. A. (2014). Modelling the spatial nature of household residential mobility within municipalities in Colombia. Applied spatial analysis and policy, 7(3), 203- 223. Wheeler, D., & Tiefelsdorf, M. (2005). Multicollinearity and correlation among local regression coefficients in geographically weighted regression. Journal of Geographical Systems (7), 161-187. Zervas, G., Proserpio, D., & Byers, J. W. (2014). The rise of the sharing economy: Estimating the impact

• f Airbnb on the hotel industry. Journal of Marketing Research.