Does the real estate market for Logistics assets benefit from the growth of E-commerce? : hedonic pricing modelling in Europe

(1)

Does the real estate market for

Logistics assets benefit from the

growth of E-commerce? Hedonic

pricing modelling in Europe

Master Thesis

Author: Barbora Houfková Supervisor: Marcel Theebe Student Number: 11374438 Month and Year: July 2017

(2)

Declaration of Authorship

This document is written by Student Barbora Houfková 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 University of Amsterdam is responsible solely for the supervision of comple-tion of the work, not for the contents.

(3)

Acknowledgments

I am grateful to my supervisor Marcel Theebe for his kind guidance and helpful advice.

This thesis was written within an internship at Patrizia Immobilien – Logistics Management Europe B.V. I would hereby like to thank all my colleagues for pro-viding me with valuable experience and especially Alfio Shkreta for his interesting comments and overall support.

(4)

Abstract

E-commerce is considered to be one of the biggest drivers of logistics real es-tate demand. Nevertheless, this claim have not been proven by any empirical research. Therefore, the aim of this thesis is to investigate the potential causal relationship between E-commerce and transaction prices of logistics properties.

Real estate transactions from 2007 until 2016 from selected European coun-tries were pooled into one dataset consisting of 6482 observations. To analyse the effect, OLS estimation was used to estimate a hedonic pricing model. The choice of control variables was driven by previous literature including physical, locational and market characteristics in the model.

The main conclusion is following: "There is no evidence of causal relationship between E-commerce and price for examined dataset." Nevertheless, the interac-tion term of E-commerce and Size was estimated as a significant factor. It implies that with increasing size of the property, the E-commerce effect is decreasing. There is no evidence that E-commerce effect changed in 2016 compared to 2017. Moreover, the findings about the effects of market characteristics on industrial property prices were confirmed.

JEL Classification R30, R33, R39, G11

Keywords Commercial Real Estate, Industrial, Logistics, He-donic Pricing Model, E-commerce

(5)

List of Tables

2.1 Literature Review Summary . . . 8

3.1 Market Independent Variables . . . 16

3.2 Summary Statistics . . . 21

3.3 Dummy Variables . . . 21

3.4 Price per Square Metre . . . 23

5.1 Estimation Results . . . 36

5.2 Estimation with Interactions . . . 39

6.1 Robustness Check . . . 44 A.1 Descriptive Statistics with Outliers . . . I A.2 Cross-Correlation Table . . . II A.3 Estimation Results . . . IV A.4 E-commerce Change . . . V

(8)

List of Figures

3.1 E-commerce Changes . . . 17

3.2 E-commerce Development . . . 18

3.3 E-commerce Change Development . . . 19

3.4 Composition of Transactions . . . 22

5.1 P-P Plot . . . 31

5.2 Heterogeneous Effects of 10% Increase in E-commerce . . . 42 A.1 Accessibility Ranking . . . VI

(9)

Chapter 1 Introduction

E-commerce is attracting widespread interest due to its impact on real estate in-dustry in general. However, it is not a new phenomenon. E-commerce sales in Europe were growing at average rate of 14% per year since 2010 (BNP, 2016) and are expected to reach 598 billion euros in 2017 (News, 2016). Many studies fo-cused on consequences for the retail sector. However, do the logistics real estate assets benefit from the growth of E-commerce?

Logistics investment volumes were boosted in 2016. Moreover, gross initial yields in this particular sector of industrial real estate substantially declined from 8.5% to 6.5% for small and large logistics since 2014 (RCA, 2016b), showing the increasing investor’s interest and establishment of the logistics real estate sector as a safe investment opportunity.

Online sales play a big role in this phenomenon and E-commerce is widely considered to be one of the biggest drivers of the logistics market. Moreover, it is said that "E-commerce became the single largest disrupter to the logistics indus-try, changing the way we think about industrial real estate" (CBRE, 2017).

Topic background is based on three main fundamentals: increased construc-tion, increased tenant demand and current investment market.

1.1 Increased Constructions

Until now, the supply side of industrial real estate was not that responsive mean-ing that new completions have been less than long term average whereas net absorption has been greater than long-term average. Nevertheless, it begun to evolve. In Europe, year 2016 was forecasted to be one of the strongest years for development of industrial real estate (CBRE, 2016).

(10)

1. Introduction 2

1.2 Increased Tenant Demand

There is also a shift in demand. While demand for logistics remain strong, there is increasing interest in smaller warehouse spaces to meet consumer’s expecta-tion and quick delivery. In the US, so called light industrial properties account for a third of all buildings under construction (CBRE, 2016). There is also de-mand for different type of facilities such as collection points, processing centres for returned items and sortation centres (JLL, 2013). In Europe, take-up generally increased by 28% in 22 cities (BNP, 2016).

1.3 Investment Market

On average, yields of logistics in Europe experienced inward movements, espe-cially for smaller units (less than 10 000 square metres). Lower vacancy rates are pushing rents higher for example in Berlin or Barcelona. In the Netherlands, de-mand remains strong but due to constrained supply, there is an upward pressure on rent levels. By comparing year on year in terms of transactions, industrial sec-tor was number one. In overall, there was 21% YOY decline whereas industrial sector experienced 1% YOY increase, which can be taken as outperformance. Par-ticularly, logistics volumes were boosted in 2016 (RCA, 2016b).

∗ ∗ ∗

Despite general perception that there is a positive effect of E-commerce on prices of logistics real estate, there is still some controversy surrounding this claim. Hence, both sides will be discussed in the literature part.

How can we assess the performance of logistics assets within the sector? Are the property values increasing over the time due to E-commerce or due to market sentiment? To capture the impact of E-commerce, transaction prices of proper-ties will be analysed. The research countries are the United Kingdom, Germany, France, Sweden, the Czech Republic, Spain, Belgium and the Netherlands.

The aim of the thesis is to reveal a potential causal relationship between E-commerce and logistics prices and the main research question is: "Does E-E-commerce have a significant effect on prices of logistics properties?" This thesis studies the main determinants that influence property values in the logistics sector over the last 10 years. A longer time period is preferred to study the actual long-term trend in capital values. The importance of E-commerce for industrial real estate is

(11)

ana-1. Introduction 3

lyzed while controlling for physical, locational, financial and economic variables. In particular, this study will use hedonic pricing modelling to explain differences in capital values of logistics buildings in selected European countries and discuss the role of E-commerce.

This topic is relevant for real estate professionals to show an outlook of the investment attractiveness of industrial real estate properties in a world where E-commerce is further on the rise.

The structure of the thesis is following. Chapter 1 gives a brief introduction into the topic. Chapter 2 presents relevant literature related to the research ques-tion and is divided into two secques-tions. The first part is devoted to the determani-nats of industrial real estate while the second part is devoted to the link between real estate and E-commerce. Chapter 3 provides the description of examined dataset and the main descriptive statistics. A proposed methodology is outlined in Chapter 4. The next chapter discusses the assumptions for the model and in-vestigates the main hypothesis. Furthermore, Chapter 6 provides a robustness check for previous models. Conclusions are drawn in the final section together with limitations, implications and extentions of the thesis.

(12)

Chapter 2 Literature Review

Despite increasing importance of E-commerce, no one to the best of our knowl-edge has directly studied the relationship between E-commerce and logistics real estate, showing the necessity of future research. In particular, concerns about E-commerce have arisen with regard to the retail property sector whereas positive consequences for logistics real estate are expected but still remain unclear. In or-der to address this topic thoroughly, it is important to elaborate on the previous research from two different perspectives. Therefore, the literature review is di-vided into two parts. The first part summarizes articles studying determinants of industrial real estate and the second part is devoted to the relationship between E-commerce and real estate in general.

2.1 Determinants of Industrial Real Estate

Numerous authors have written about the main factors influencing values of in-dustrial real estate such as Ambrose (1990) who explained variations in asking prices and rents or Hoag (1980) who focused on property characteristics using actual sale prices. The studies of Fehribach et al. (1993), Miles et al. (1990) and Beekmans et al. (2014) are deemed to be a starting point for the thesis because they are focused on values of industrial properties using a hedonic pricing model. The selection of explanatory variables is crucial for our model to sufficiently explain the variation in prices and simultaneously avoid omitted variable bias. The characteristics influencing prices of real estate in general can be divided into three groups: physical, locational and market variables.

(13)

2. Literature Review 5

2.1.1 Property Characteristics

Property characteristics are important to analyse real estate in general (Hoag (1980) or Ambrose (1990)). Size is considered to be one of these that cannot be ne-glected in the analysis and its importance was highlighted for example by Am-brose (1990), Fehribach et al. (1993) or Ryan (2005). Beekmans et al. (2014) ar-gue that same implies for age whereas Ambrose (1990) or Fehribach et al. (1993) did not find the same results. Besides age and size, many authors included in the model other property factors such as office share (Lockwood and Rutherford, 1996), number of doors (Fehribach et al., 1993), height (Buttimer et al., 1997), type of property (Beekmans et al., 2014), land size (Jackson, 2001), free standing (Miles

et al., 1990) or vacancy of the building (Ryan, 2005).

2.1.2 Locational Characteristics

While assessing real estate values, location is of great importance (Hoag, 1980) and it can be seen from different perspectives. While some authors found lo-cation close to motorways to be disadvantage in some areas in terms of rents (Ryan, 2005), other authors argue that proximity to ports and roads, significantly increases industrial property values (Dunse et al., 2005).

Accessibility can be also explained by the distance from CBD. While residen-tial and office buildings often benefit from the centre location, different views can be seen in case of industrial real estate. On the one hand, location in CBD usually increases the value or rent due to agglomeration effect (Dunse et al., 2005), mean-ing that there is a negative relationship between prices/rents and distance from CBD. On the other hand, being more distant from CBD does not have to neces-sarily imply lower price of the property in case of industrial real estate due to less congestion and better accessibility (Lockwood and Rutherford, 1996). Greater distance from CBD can therefore increase value of industrial property. Some studies did not find any significant link between rents and distance to CBD (Ryan, 2005) whereas some studies pointed out that logistics centres tend to be in pe-ripheries rather than in neighbourhoods (Gorczynski and Kooijman, 2015).

Furthermore, Lockwood and Rutherford (1996) found a negative relationship between distance from the airport and industrial values whereas positive rela-tionship between distance from the central business district and industrial val-ues. Nevertheless, the findings can be distorted by multicollinearity problems.

(14)

forward-looking concept of future cities – Aerotropolis as a wave of new devel-opment. He argues that gateway airports will become important to business ac-tivity in 21st century as automobiles were in the 20th century. Consequently, he suggests that investors could benefit from this trend by focusing on logistics real estate investments in the near proximity to airports (Kasarda, 2000).

Beekmans et al. (2014) test hedonic pricing model for industrial sites in the Netherlands during the time span from 1997 to 2008. The study indicates that hedonic pricing model can be used in a meaningful way in case of industrial real estate. Moreover, they claim that despite heterogeneity of properties, industrial sites are sufficiently homogenous to apply hedonic model. They found other sig-nificant variables related to location – sea port, accessibility and the presence of water. By contrast, variable centrality showed to have no influence on average property values.

Moreover, most of the studies found variable Country or Region significant, implying different attractiveness of areas or countries for industrial real estate.

2.1.3 Market Characteristics

Among literature, contradictions about inclusion of market characteristics can be seen. In 1980, Hoag pointed out that economic variables capturing supply and demand for real estate should be not be neglected. He highlighted that national economy is positively correlated with industrial values and he found important significant variable influencing prices of industrial real estate – volume of indus-trial sales.

Miles et al. (1990) found two important market characteristics influencing transaction prices of industrial warehouses – population change and income per capita. Both have a positive influence on industrial prices. By contrast, they did not find any significant relationship in case of change of wholesale earnings. Moreover, the capitalization rate was not included in the model for industrial warehouses. By contrast, in the study of Fehribach et al. (1993), capitalization rate of industrial properties showed to have negative effect on industrial values.

Further research of Buttimer et al. (1997) showed that employment change have a positive significant effect on industrial rent rates. Lockwood and Ruther-ford (1996) used LISREL model to confirm a positive effect of employment change on industrial real estate prices. Furthermore, they showed the importance of in-clusion GDP with a positive effect whereas they found no significant relationship between property industrial values and interest rates, namely bond yields,

(15)

indus-2. Literature Review 7

trial yields and industrial conventional loan rate. The positive effect of household income was confirmed by Ryan (2005).

The study of Lieser and Groh (2014) examined the determinants of commer-cial real estate investment using market level data of 47 countries and random effect panel regression. They conducted two analyses – between and within es-timation. According to their findings, GDP per capita and size of economy were found to be significant factors in both cases. Nevertheless, it needs to be noted that study uses market level data and analyses real estate investments from dif-ferent perspective. The importance of the size of economy was also confirmed by Chen and Hobbs (2003) or DiPasquale and Wheaton (1992).

Lastly, Beekmans et al. (2014) found the variable density (number of addresses per ha) as a significant factor influencing industrial property values while urban-isation rate not (in a form of dummy variables).

To sum up, while some authors found market characteristics to be an impor-tant input to explain selling prices (Hoag (1980), Lockwood and Rutherford (1996) or Beekmans et al. (2014)), some authors did not take them into account while studying industrial prices (Ambrose (1990) or McDonald and Yurova (2007)). Many studies pointed out the presence of multicollinearity problems (Beekmans et al. (2014) or Lockwood and Rutherford (1996)). Therefore, the results should be treated with the utmost caution.

For better illustration, Table 2.1 summarizes the literature review about deter-minants of industrial real estate.

(16)

2. Literature Review 8 T ab le 2.1 : L iter atur e R evie w S u mma ry A uthor H o a g Amb rose M il es et al. F ehr ibac h et al . B utt im er et al. Lo c kw ood R u ther fo rd Jack so n R y an D unse et al. B eekman s et a l. Y ea r 19 80 1 99 0 199 0 19 93 1 99 7 19 96 20 01 2 005 20 05 2 014 D e pendent V ar ia b le P rice P rice , R ent P ri c e V alu e R ent P ri c e P ri c e R ent R en t V a lue Ind e pendent V ar ia b les Ag e -A ge -Age -A ge -Ag e -Ag e -S iz e + S iz e + Inc o m e + S iz e + S iz e + S iz e + S iz e + S iz e + -T y p e Offi ce + B u ilding s + Offi c e + Offic e + -Offic e + Offic e + V acan cy -S ea p or t -V ol ume of in du st rial sale s + Do ors + F re e st an ding + Doors + H ei g ht -Land + L a nd + A ccessibili ty + H ei g ht + D ist an ce air p or t -D ist an ce motor w ay + D ist anc e motor way -D ist anc e m o tor way + P op u lat ion cha nge + S in gle ten ant + D ist anc e CB D + D ist anc e S tation -D ista nc e CB D -W ater + Income p er c apit a + D ist anc e Airpor t -Income + In come +-E mplo yment ch ang e + E mp lo ymen t R ate + D ensity + C ap italiz at io n rate -N c as es 46 3 5 7 102 17 0 8 48 30 8 12 2 11 65 42 9 2 7 14 1 N ote : Th is table pr o vi d e s an o v er v iew of th e li ter a tur e rev iew abou t in du st rial re a l est ate sinc e 198 0 u n til 20 14. In th e table , th e n ame of the auth o r, y ear , dep e n dent v ar iab le , indep enden t v ar iables an d n u mb e r o f obser v at ion s ca n b e seen. T he d e -pen dent v a riable s u se d in studies ar e pr ic e , re n t or v alu e . In each c o lumn, in depen-den t v ar ia bl es tha t w er e fou n d as sign ifican t a re pr esent e d w ith th e sign o f estimat ed coeffi cient .

(17)

2.2 Real Estate and E-commerce

At first glance, the relationship between E-commerce and logistics can seem straight-forward, meaning that increasing online sales require more distribution space and demand is therefore higher. There are many studies or reports supporting this claim such as Fernie et al. (2010) or JLL (2013). Although in reality the link be-tween these two is more complex. In fact, E-commerce does not have influence only on logistics but also on retail sector. More importantly, E-commerce, logis-tics and retail are highly interconnected. The reason is that logislogis-tics do not serve only to households, but also to many shops, whose future stays unclear. Hence, E-commerce effects can divert from general perception and higher online sales do not have to necessarily mean higher demand for logistics real estate.

In the US, there is also a shift in occupier demand. The focus changed from big-box facilities towards smaller light-industrial properties with the size smaller than 20 000 square metres. This pattern is a consequence of substitution of logis-tics for retail with logislogis-tics for E-commerce. To fully address this topic, one needs to first understand the relationship between E-commerce and retail sector. Since there is still much controversy surrounding these two, a small part of literature will be devoted to this point.

First attempts to outline the relationship between E-commerce and commer-cial real estate were made for example by Graham and Marvin (2002), Worzala and McCarthy (2001) or Bardhan et al. (2000). Worzala and McCarthy (2001) were one of the first ones examining demand for retail space and impact of internet on shopping centre’s value. At that time, it was too early for predictions, nevertheless, preliminary conclusions were drawn. Based on survey, retailers of standardized goods were more concerned about the future of their retail business whereas re-tailers with specialized goods did not perceive Internet as a threat.

Bardhan et al. (2000) pointed out that all goods need space for storage. Con-sequently, the growth of E-commerce entails increased demand for warehouse space. In case of retail sector, first impression could be that retail is the most vul-nerable sector when E-commerce is on the rise. Nevertheless, the process is more complex and retail tends to be rather complement of online sales than being re-placed by them. Either way, the expansion of E-commerce would imply higher demand for warehouse space not depending on whether retail sector suffers or not (Bardhan et al., 2000). Moreover, the study does not consider the case when retail suffers more than could online sales compensate for. Apart from quantify-ing the demand, there is also shift from storage place to pass through locations

(18)

suggesting a shift in demand. Furthermore, some studies have shown that stores and internet sales can reinforce each other by integrating different channels (Mc-Goldrick and Collins (2007) and Williams (2009)).

By contrast, research from the Chinese market suggested that traditional retail sector is likely to suffer from the growth of E-commerce and that each type of commercial real estate is differently affected by E-commerce. Nevertheless, the study used a statistical comparison to investigate the impact of E-commerce on commercial real estate by looking at trends and development of vacancy rates, number of sales, stores, etc. rather than conducting regression analysis (Zhang

et al., 2016).

Another studies suggesting that E-commerce can bring serious consequences for traditional retail are for example Baen (2000) or Williams (2009). Weltevreden and Rietbergen (2007) studied this relationship on data from the Netherlands and found that more than 20% online buyers made fewer trips to brick and mortar stores due to E-shopping.

In the US, it was suggested that despite the fact that Internet will never com-pletely replace in-store experience, E-commerce is seen as the biggest threat for retail. In theory, retailers could exploit E-commerce to boost their revenue by us-ing stores and Internet, nevertheless, they are often not able to meet customers’ demand in real life. One example of retail suffering is decreasing store footprint (Carney and Tolliver, 2015).

Consequently, the impact of E-commerce on logistics is questionable. Re-cent study of Gorczynski and Kooijman (2015) investigated consequences of E-commerce for supermarkets in the Netherlands. It showed the interdependency of E-commerce, logistics and supermarket sector. The findings were based on questionnaires from eight supermarket chains in the Netherlands. Surprisingly, only one supermarket chain implemented specialized logistics real estate due to E-commerce.

Furthermore, respondents said that online retailing does not have any signif-icant effect on size, layout or location of distribution centres in the short run and they do not have a tendency to expand amount of square metres of distribution centres. Potential adverse effects on logistics can be explained by existing model of distribution centres and unwillingness to implement a new strategy. Never-theless, the situation is expected to be different in the long run. One conclusion claims that after reaching certain level of online sales, retail sector will be detri-mentally affected and other conclusion claims that once online sales change

(19)

dras-2. Literature Review 11

tically, the number of warehouses per store increases because of higher demand for pick-up centres. Although in certain countries, the shift from stores to logis-tics is not established yet. Moreover, there is hardly effect on division between warehouse and sales area in the short run. This finding was found interesting, therefore, attention will be devoted to differentiate between short run and long run while stating the hypotheses.

∗ ∗ ∗

Although the growth of E-commerce has attracted attention over the last years, logistics was rarely the research subject in journals. Available literature mainly fo-cused on the impact of E-commerce on retail sector, however, consequences were not revealed in detail. Also, the findings cannot be translated into logistics easily because of its particular features.

The research area was limited to the US. While investigating Europe, articles were focused only on few countries such as the Netherlands or the UK. Hence, he thesis will therefore analyse eight countries based on criteria that will be dis-cussed later.

Furthermore, studies mainly went into how E-commerce is connected with real estate through questionnaires. Hence, the thesis will contribute to current literature by performing a quantitative approach using property-level data.

In comparison to previous research, this thesis is mainly focused on logis-tics assets. This study aims to reveal a potential causal relationship between E-commerce and capital values of logistics assets in selected European countries. By expanding the data sample about more observations, estimates of coefficients are likely to be more accurate. Even though examined data set does not include all information about physical characteristics of the properties, as many as possi-ble control variapossi-bles will be considered to avoid omitted variapossi-ble bias. Moreover, to control for economic trends, transaction date will be included in the model as it was indicted in the study of Beekmans et al. (2014) or Ryan (2005).

(20)

Chapter 3 Data

In this part, Section 3.1 describes the sources of the data, Section 3.2 gives a brief overview of all variables used for the analysis – including its specification and form and Section 3.3 presents descriptive statistics about E-commerce.

3.1 Data Source

The dataset is a combination of primary source data specific for each property transaction taken from the database Real Capital Analytics (RCA) and secondary source data from different websites.

The Real Capital Analytics (RCA) database provided the information about in-dustrial European real estate transactions. "RCA is privately held company and authority on the deals, the players and trends driving the commercial real es-tate markets". The database covers more than $18 trillion of transactions in 172 countries linked to 200 000 investors. "Since 2000, RCA is considered to contain the most timely and reliable transaction data with no other competitor providing such detailed information on financing. RCA uses a vast network of independent sources, objective methodology and specialized team of researchers to provide the most trusted data in the industry." Their reports are frequently quoted in in-dustry news or in academic research –the founder of RCA, Robert M. White, Jr. has published numerous articles and is frequently cited in the press (RCA, 2016a). The standardized data sheets include the information about transaction prices, physical, locational and deal characteristics of properties.

The properties have been filtered according to the time period, selected coun-tries and the type of real estate. In the beginning, the dataset included 7463 trans-actions. For the purpose of analysis, only transactions with available information

(21)

3. Data 13

about the price and the size could be used. This fact reduced the number of ob-servations resulting in the final dataset of 6482 obob-servations.

The secondary data are taken from various sources: Eurostat, The World Bank or OECD. The data were obtained separately for each country, merged together and modified.

A period of 10 years (01-01-2007 until 31-12-2016) is chosen to study a long-term relationship. The properties are well diversified geographically (8 countries) and by industrial type (flexible and warehouse).

The final research countries are Belgium, the Czech Republic, France, Ger-many, the Netherlands, Spain, Sweden and the United Kingdom. As it was men-tioned before, previous studies on industrial real estate are mostly limited to the US. Hence, the thesis is focus on European countries. The inclusion of coun-tries with significant change of E-commerce over the time period allows study-ing potential E-commerce effects. Furthermore, countries with different logistics performance index were included. Some countries were excluded due to unavail-ability of E-commerce data such as Luxembourg or Switzerland.

The transactions are unequally spread over the countries with different num-ber of transactions. For instance, the Czech Republic contains 58 observations whereas the United Kingdom contains 2717 observations. From this reason, all countries are pooled together to estimate the final model.

In general, logistics deals belong to long term investments. During the exam-ined study period, each property was sold only once, therefore, there are different units in each year. The data sample has cross-sectional and time dimension.

3.2 Description of Variables

In this part, each variable used in the analysis is described in a detail. The selec-tion of the control variables is driven by the previous findings in literature.

3.2.1 Dependent Variable

l_Price: A natural logarithm of the transaction price of property at given date in

(22)

3. Data 14

3.2.2 Independent Variables

Physical Characteristics:

Flex: A dummy variable that takes value 1 if the property is flexible, otherwise it

takes value 0. Flexible properties can be used for industrial and office purposes (so called light industrial uses). The base omitted group is Warehouse. The defini-tion of warehouse is a property used for manufacturing, storage or redistribudefini-tion purposes.

Age: Age of the building calculated as the difference between 2017 and the year

built. The information about the age is available for 2898 properties.

l_Size: A natural logarithm of the size of property in square metres.

Year: The year when the transaction was executed. The omitted group is year

2007. Disregarding one element of the set of dummy variables is crucial to avoid the dummy variable trap. The model specification with all dummy variables tends to violate the "no perfect multicollinearity assumption" of the OLS (Wooldridge, 2015).

Locational Characteristics:

Country: A dummy variable that takes value 1 if the property is located in certain

country, otherwise it takes value 0. There are eight countries in the data sample.

Access: The location was assessed based on industrial European ranking of cities

according to the accessibility and infrastructure (Colliers, 2013). The dummy variable Access takes value 1 if the city was ranked among the most accessible cities in the ranking. The ranking is based on following criteria: quality of in-frastructure, air freight capacity of airports, capacity of seaports or the degree of accessibility by rail. The base group is Access_bad. The list of cities was modified and can be seen in Appendix – Figure A.1.

Market Characteristics:

(23)

3. Data 15

E-commerce: The percentage of individuals that made purchases online in the

past 12 months in a given country.

E-commerce =individuals that made purchases online

all individuals .

E-commerce_chng: The difference in E-commerce between two consecutive years

in a given country expressed in basis points.

GDP_lagged: One year lagged gross domestic product at market prices in a given

country and year. GDP includes output, expenditure and income capturing eco-nomic activity in a given country. The data was taken from Eurostat in millions of EUR.

Inflation_chng: Total annual growth rate of inflation expressed in percentages in

a given country – taken from the OECD.

Unemployment_chng: Unemployment change expressed as a difference in

un-employment rates between two consecutive years in a given country expressed in basis points. The unemployment rate is expressed as % of total labor and was taken from the OECD.

Density_chng: The difference in the number of people per square km of land area

between two consecutive years in a given country – taken from the World Bank. Density represents a proxy for urbanisation rate.

Population_chng: The difference in the number of inhabitants between two

con-secutive years in a given country – taken from the Eurostat.

Properties_ratio: Variable Properties represents the average number of sold

prop-erties in one quarter (due to availability of quarterly data). Therefore, Propprop-erties ratio is the average number of sold properties in one quarter in a given country divided by the average number of sold properties in one quarter in all European countries in a given year – taken from RCA.

Industrial: The average industrial cap rate in a given year and country – taken

(24)

3. Data 16 T a bl e 3.1 : M ar ket Indep e n den t V ar iab les V a ri a ble D e fini ti o n U nits S ign Sour c e E-comme rce Th e per c ent age of in div idual s tha t mad e pu rch ases online in the p ast 12 m o n th s in a g iv en count ry . % E ur ost at E-comme rce_c h n g Th e diff e renc e of E -commer ce v ar iab le bet w ee n two con secu tiv e y e a rs in a giv en cou n tr y. basis p oi n ts E ur ost at GDP_lagged O n e y ear la gged g ross d o m e stic pr odu c t in a giv en c o unt ry . mill ion eur o + E ur ost at In flation _chng Inflat io n ch an ge expr essed as to tal ann u a l g ro wth rate . % + OECD U ne mplo ymen t_c h n g Th e diff e renc e in u n emplo y m ent rat es betw e en two consec u ti v e y ears in a giv en c ou n tr y. basis p oi n ts -OECD D ensi ty _ch n g Th e diff e renc e in th e number o f people per squ a re km of lan d ar e a betw een two people/ squar e km + W orl d B an k con secu tiv e y ears in a g iv en cou n tr y. P opu lat ion_c h n g Th e diff e renc e in th e number o f inh abitan ts betw een tw o c o n secutiv e y ears in a g iv e n count ry people + E ur ost at P rope rt ies _r ati o Th e av er a ge number of sold pr oper ti es in o n e qu ar ter in a g iv en cou n tr y divided b y the ratio + R C A av er age n u mb e r of sol d p roper ties in one qu a rter in all E ur opean cou n tr ies . In d u str ial Th e av er a ge industr ia l c apita liz a tion rate in a giv en y ear an d count ry . decimal n u mbers -R C A N ote : Th is table pr o vi d e s a n o v e rv iew of mar ket indep enden t v ar ia bl es used in re-gr ession models . Th e table pr esent s th e n ame of th e v ar iable , th e defi nition, u n it s an d th e so ur c e of th e d ata. In c ase of con tr o l v ar iab les , exp e c ted sig ns of c o effi cient ar e p resent ed acc o rding to th e p revi ous liter at u re or economic sen se . T her e ar e sev e n con tr o l v ar iables an d tw o v a riable s of in ter e st for th e an aly sis . Th e sou rce of th e dat a is E ur ost at, W orl d B an k o r OECD .

(25)

3. Data 17

3.3 E-commerce

Since E-commerce is the variable of interest, it is important to describe it in a detail.

The data are collected annually by the National Statistical Institute through questionnaires by households and individuals. For the purpose of the analysis, the percentage of all individuals that made purchases online in the past 12 months is used. When studying effects of E-commerce, the variable "Internet purchases by individuals" was for instance chosen in a recent study of Anvari and Norouzi (2016).

Figure 3.1: E-commerce Changes

0 10 20 30 40 50 60 70 80 90

Sweden Germany Netherlands Spain Czech Republic

United Kingdom

Belgium France European Union 2006 2016

Note: This figure shows the percentage of individuals that made purchases online in the past 12 month compared to all individuals. The first column shows the percent-ages in 2006 and the second column shows the percentpercent-ages in 2016. In all countries the percentage of individuals that made purchases online in the past 12 month in 2016 was higher than in 2006. There are eight countries presented in the figure. The last two columns belong to the average of European Union. Some countries (the Czech Republic, Belgium or Spain) have lower percentages of online individuals in 2006 than the average of European Union while some countries (Germany or the Netherlands) have higher percentage of online individuals in 2006 compared to the average of Eu-ropean Union. The same implies for year 2016. The countries are ranked according to the nominal change of E-commerce over the time period from the smallest to the biggest change.

Source: Eurostat, 2017

According to available sources of data, E-commerce accounted for 3,4% of total sales in the end of 2007 and since then the development differs across the countries in Europe (E-commerceland, 2000). While some countries experienced a sudden rise of internet purchases since 2006 (Czech Republic or Belgium), E-commerce activities did not change substantially for example in case of Germany.

(26)

3. Data 18

E-commerce was increasing in all countries over the examined time period. Therefore, it is more logical to study changes of E-commerce than nominal val-ues to see the effect on prices. Figure 3.1 shows the valval-ues of variable Ecommerce for two different years – 2016 compared to 2006. As it can be seen, France ex-perienced the biggest change of E-commerce activities (changes in basis points) while Sweden experienced the smallest change of E-commerce activities (chan-ges in basis points) over the time period.

Figure 3.2 shows how E-commerce changed over the time span in the Czech Republic, Belgium, the United Kingdom and France (countries with a significant change of E-commerce over the time) compared to the rest of countries.

Furthermore, Figure 3.3 shows also the development of E-commerce but ex-pressed in nominal changes over the time span for France, the Czech Republic, Belgium and the United Kingdom.

Figure 3.2: E-commerce Development

0 10 20 30 40 50 60 70 80 90 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

Czech Republic Belgium United Kingdom France

Note: This figure shows the development of E-commerce over the time span (from 2006 until 2016). The countries included in the figure experienced the biggest nomi-nal change of E-commerce over the time span – France, the Czech Republic, Belgium and the United Kingdom. The inclusion of this countries with significant increase of E-commerce allows to study a potential causal effect of E-commerce on price. Source: Eurostat, 2017

(27)

3. Data 19

Figure 3.3: E-commerce Change Development

-400 -200 0 200 400 600 800 1000 1200 1400 1600 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016

E

-c

om

m

er

ce

c

ha

ng

es

Belgium Czech Republic France United Kingdom

Note: This figure shows the development of E-commerce changes in nominal val-ues over the time span. The countries included in the figure experienced the biggest nominal change (in basis points) of E-commerce over the time span – France, the Czech Republic, Belgium and the United Kingdom. The inclusion of these countries with significant increase of E-commerce allows to study a potential causal effect of E-commerce on price.

(28)

3. Data 20

3.4 Descriptive Statistics

3.4.1 Correction for Outliers

Before analysing the data, it is important to correct for outliers in order to have unbiased and consistent coefficients. One way how to deal with the presence of outliers is the imputation of the values by winsorizing. Firstly, the cap was set at the 5th and 95th percentiles and extreme values were transformed. Even though this procedure ignores the uniqueness of the dataset, the number of observations stays at the same level and regression results are not disrupted by outliers. Two variables – Price and Size were winsorized. The descriptive statistics with outliers can be seen in Table A.1.

3.4.2 Descriptive Statistics

Table 3.2 provides the main statistics about the data (mean, standard deviation, min, max and number of observations). The average price of the property is 14 582 000 EUR and median price is 8 521 000 EUR. It implies that there are more properties with lower price and few properties with high price increasing the mean. The same implies for size. The average size of the property is 19 233 square metres while median size is 11 499 square metres. The distinction be-tween smaller and larger industrial properties usually lies around 10 000 square metres. In the data sample, there is approximately 45% small properties with the area lower or equal to 10 000 square metres.

The variable Age ranges from -1 to 255. A negative number suggests that the building is finished one year after the transaction is done. High numbers show that there are many buildings in the sample constructed a long time ago – for example in year 1833.

The sample includes two types of logistics buildings. There is approximately 19% flexible properties and 81% pure industrial warehouses. The average price of flexible property is 11 022 000 EUR with the average size of 12 717 square metres while the average price of warehouse is 15 412 000 EUR with the average size of 20 753 square metres. The maximum age of flexible property is 184 years. Only 11% observations are considered to have good accessibility according to the ranking.

The properties are located in 8 European countries. The biggest representa-tive is the United Kingdom with 42% properties. The Figure 3.4 illustrates the

(29)

3. Data 21

Table 3.2: Summary Statistics

Variable Mean Std. Dev. Min. Max. N

Price (w) 14 582 15 072 1 799 57 500 6482 Size (w) 19 233 19 541 2 423 75 253 6482 Age 25 23.8 -1 255 2898 Ecommerce_chng 298 286 -400 1500 6482 GDP_lagged 1 660 561 850372 123 743 3 032 820 6482 Inflation_chng 1.647 1.217 -0.5 6.34 6482 Unemployment_chng -30.9 81.6 -243 662 6482 Density_chng 1 1.5 -4 5 6482 Population_chng 229 797 424861 -1 580 192 978 147 6482 Properties_ratio 0.206 0.138 0.011 0.5 6482 Industrial 0.076 0.008 0.057 0.101 6482

Note: This table contains the descriptive statistics of the dependent variable and main continuous independent variables. The mean value of price and size suggests that on average the price is 14 582 000 EUR and the average size is 19 233 square metres. The average change of E-commerce is 2,98%, the average change of Inflation is 1,6%, the average change of Density is 1 person per square kilometre, the average change of Population is 229 797 inhabitants and the average change of Unemployment is -0,3%. All variables are expressed yearly. The mean of Industrial cap rate is 7,6% and the mean of GDP is 1 660 561 million EUR.

composition of transactions according to the countries and the type of the prop-erties. The observations are spread across the whole time span, nevertheless, the biggest number of transactions was observed in 2015 with 14% of observations. The amount of transactions significantly differs among countries.

Table 3.3: Dummy Variables

Stats Flex Access

mean .189 .112

sum 1226 728

Note: This table provides an overview of dummy variables. It can be seen that 18,9% of properties are flexible while 81,1% properties are pure industrial buildings. In total, 1226 properties are flexible out of 6482. Furthermore, 11% properties have a good accessibiity and infrastructure resulting in 728 properties.

(30)

3. Data 22

Figure 3.4: Composition of Transactions

139 ₅₈ 661 935 894 178 900 2717 0 500 1000 1500 2000 2500 3000 Belgium Czech

Republic France Germany Netherlands Spain Sweden KingdomUnited

Flex Warehouse

Note: This figure shows the composition of properties according to the country and the type. As it can be seen, most of the properties are warehouses. The biggest rep-resentative is the United Kingdom with 2717 transactions and the lowest amount of transactions was observed in the Czech Republic.

Source: RCA, 2017

The highest average price is observed in the Czech Republic despite the low-est price level index (OECD, 2016). Nevertheless, the bigglow-est average size of the property is located in the Czech Republic as well. Therefore, it is meaningful to look at price per square metre.

Table 3.4 shows the average price per square metre in each country. Column 1 includes raw data and column 2 includes windsorized data according to the country. It is logical to first look at unmodified data. As expected, the highest average price per square metre is observed in the United Kingdom with the value of 1211 EUR and the lowest average price per square metre is observed in the Czech Republic with the value of 571 EUR (OECD, 2016). The average price per square metre in the whole data sample is 952 EUR.

Furthermore, the highest average price per square metre was 1046 EUR in 2016. In year 2012, standard deviation of price per square metre was the highest with the minimum value of 33 EUR and the maximum value of 30 667. Therefore, it is important to correct for outliers because such values could distort the results. In case of windsorized data, the highest average price per square metre is also observed in the United Kingdom with the value of 1158 EUR and the lowest av-erage price per square metre is observed in the Czech Republic with the value of 539 EUR. The average price per square metre in the whole data sample is 923 EUR.

(31)

3. Data 23

Furthermore, the highest average price per square metre was 1027 EUR in 2016. Table 3.4: Price per Square Metre

Country Price per Sqm Price per Sqm (w)

Belgium 730 726 Czech Republic 571 539 France 740 711 Germany 755 752 Netherlands 721 723 Spain 951 820 Sweden 819 822 United Kingdom 1211 1158 Total 952 923

Note: This table shows price per square metre according to the countries for raw data and windsorized data. The highest price per square metre was observed in the United Kingdom and the lowest price per square metre was observed in the Czech Republic. Source: RCA, 2017

Moreover, the Table 3.2 presents range and mean of market variables. A posi-tive mean in case of E-commerce, Inflation, Density and Population suggests that these variables were growing on average during the time span. Furthermore, Un-employment was decreasing on average. The average change of E-commerce is 2,98%, the average change of Inflation is 1,6%, the average change of Density is 1 person per square kilometre, the average change of Population is 229 797 in-habitants and the average change of Unemployment is -0,3%. All variables are expressed yearly. The mean of Industrial cap rate is 7,6% and the mean of GDP is 1 660 561 million EUR.

The average values will be taken into account while interpreting the results of the regressions. For example, it makes sense to interpret E-commerce in 1% change (100 basis points), GDP in millions of EUR, population changes in 100

(32)

3. Data 24

000 inhabitants or unemployment changes in 1% change (100 basis points). The variables are rescaled and estimated this way in regression tables.

(33)

Chapter 4 Methodology

In the previous section, the topic background, relevant literature overview and data description were introduced. Next part will focus on the specification of the model and hypotheses. There are different views how the impact of E-commerce can be quantified in the context of properties. The most common approach is studying the development of rents or capital values (Malpezzi et al., 2003). Ac-cording to Phillips (1988), residential rents and values are interrelated and can be expressed by capitalization rate. The same implies for commercial real estate and the relationship can be summarized by following way (Geltner et al., 2001):

capitalization rate =net operating income

property value .

Hence, the choice of dependent variable is about rent or value. Whereas the first impression can be that rents and values move in the same direction, it is not always the case. One possible example is following. Property prices are based on supply and demand. When office sector becomes more attractive for investors than others, industrial property prices may experience decline due to lower de-mand for this type of real estate. Nevertheless, rental prices of industrial proper-ties may stay at the same level supposing that there is still need to store the goods somewhere because sales are on the rise.

Studying rents can be cumbersome because "market rent" may significantly differ from "contract rent"due to concessions (rent abatement, moving expense allowance or so forth). Landlords often offer free rent period or below market rent in order to get tenants to sign long-term contracts (Geltner et al., 2001). There is therefore need to control for lease terms or contract conditions (Malpezzi et al., 2003). Nevertheless, the information about contracts is not always revealed.

(34)

More-4. Methodology 26

over, one needs to take into account also the fact whether user costs are a part of passing rent or not (Sivitanidou and Wheaton, 1992).

When estimating regression with values, one needs to be aware of difference between terms "price" and "value" because these two are often interchanged. It is important to mention that price and value are equal only under special con-ditions. Market value is a theoretical concept that applies at each point of time and represents expected selling price under the assumption that the property is sold. Transaction price is an empirical observation that can be applied only if the property is sold. Whereas market value can be applied to every property, transac-tion price is known only for a small sample of properties that have been sold. As a consequence, these two values may differ in reality because transaction prices take into account negotiations (Fisher et al., 1994). In the thesis, the difference is neglected supposing that actual price is the best approximation of capital value as it was done in previous studies.

Some studies used appraisal values instead of actual realized prices. The use of observed transaction prices brings obvious advantages such as greater preci-sion and less potential bias of estimates (Malpezzi et al., 2003). On that account, the main variable of interest is transaction price. The use of realized prices en-ables to avoid any swaps not reflecting the price paid (Hoag, 1980).

4.1 Hedonic Pricing Model

According to theory, people perceive real estate units as bundles of attributes and their utility depends on different combinations of these attributes. Therefore, the price of property can be expressed as:

Pi= P(x1i, x2i, ..., xni, l1i, l2i, ..., lki, m1i, m2i, ..., mpi),

where

X = (x1, x2, ..., xn), L = (l1, l2, ..., lk), M = (m1, m2, ..., mp)

is vector of individual physical characteristics, locational and market characteris-tics including financial and economic variables (Fehribach et al., 1993).

The hedonic pricing model is often used to approximate the values based on people’s decisions and is revealed preference method of valuation (Gundimeda, 2005). The value of this approach is in separation of different attributes that

(35)

influ-4. Methodology 27

ence the selling price and simultaneously controlling for other variables (Malpezzi

et al., 2003).

Despite the difficulties to obtain sufficient number of transactions in case of industrial real estate, the hedonic pricing model has been widely applied in many studies for industrial properties (for example Tonin and Turvani (2010) or Beek-mans et al. (2014)). Nevertheless, the suggested approach is very data intensive. In order to estimate equation, a lot of information including important charac-teristics of properties is needed as well as sufficient amount of observations. For example, it was necessary to delete observations with missing selling prices or size to be able to conduct the analysis as they were found essential to conduct the analysis (Gundimeda, 2005).

The application of hedonic pricing model assumes that relative coefficients are constant over the time. Due to sparsity of the transaction data, this approach does not allow to estimate the coefficients for each year. Therefore, the regres-sion analysis will be conducted for a whole data sample and the effect of different markets (countries) and time will be captured by inclusion of dummy variables (Fisher et al., 1994). Some variables include both dimensions–cross-sectional and time-series as it can be seen in the equation.

Since the hedonic pricing model is sensitive to omitted variables, we have to carefully specify the accurate form of the model. There is no preferred functional form of the equation. Each functional form entails certain advantages and disad-vantages and is suitable for different purposes. According to Cropper et al. (1988), linear or quadratic functions provide the most accurate estimates of marginal prices. On the contrary, Malpezzi et al. (2003) shows the importance of log models due to extra advantages.

Firstly, the estimates provide a meaningful interpretation for real estate prices. For example, with linear model, the value added by a third door to a small prop-erty is the same as the value added by a third door to a big propprop-erty whereas with semi-log model the value added to the property varies proportionally with the size. Moreover, it is convenient to interpret results in percentage changes of prices given a unit change in independent variables (Malpezzi et al., 2003).

Secondly, semi-log model helps to minimize the problem of heteroscedastic-ity. As it was suggested for example by Stevenson (2004)., the heteroscedasticity is a common problem while estimating hedonic pricing model. Therefore, miti-gation of this problem is helpful in the analysis and is discussed in Section 5.1.

(36)

4. Methodology 28

form or Box- Cox that are supposed to be more flexible than semi-log models (Halvorsen and Pollakowski, 1981). Nevertheless, these models are more suitable for estimation structural equations of demand and supply. Furthermore, the flex-ibility will be added to a semi-log model by including dummy variables in the right-side of equation. Because of these reasons, the transaction price is included in the model in the form of natural logarithm (semi-log model).

Studying logistics sector can produce some potential weaknesses because in-dustrial transactions are not that frequent as other type of commercial real estate. One way how to deal with sparse data is using independently pooled cross sec-tions to increase the sample size.

By analysing random samples of properties at different points in time, the es-timators are more precise and statistical tests gain more power compared to sim-ple cross sections. Furthermore, the advantage of pooled cross sections is their independence that eliminates correlation in the error terms across different ob-servations (Wooldridge, 2015).

The following model is employed:

ln(price)_{i ,t}= α0+α1Ecommercei t+ n X j =1 βjXj ,i+ k X h=1 βhLh,i+ p X l =1 βlMl ,i t+ 8 X m=1 Country_m,i+ 10 X t =1 Yeari ,t

where X is a vector of physical characteristics, L is a vector of locational char-acteristics and M is a vector of market charchar-acteristics. The equation is expressed generally. In this analysis,

n = 3, k = 1 and p = 7 ,

because physical characteristics include Age, Flex and Size, locational charac-teristics include Access and market characcharac-teristics include GDP lagged, Inflation change, Unemployment change, Density change, Population change, Properties ratio and Industrial.

By estimating this equation, OLS estimation will be applied to study the im-pact of E-commerce on transaction prices. Specifically, independently pooled cross sections of two dimensional data (cross-sectional and time-series) will be analysed with different units (properties) every year.

While for the objective of this research we are mainly interested in coeffi-cients of E-commerce, the inclusion of property, locational and market charac-teristics is important (Lockwood and Rutherford, 1996). Pooling also allows the

(37)

4. Methodology 29

inclusion of fixed effects (time invariant factors) in the equation to study the lationship between the dependent variable and independent variables that re-main constant over time such as physical and locational characteristics. Fur-thermore, the method enables studying the effect of time and how relationships changed(Wooldridge, 2015).

By contrast, using pooled cross sections brings also few issues. As it was men-tioned earlier, one needs to control for time effects. This can be easily accom-plished by including different time dummy variables in the equation except for the base year which is usually the earliest year. Furthermore, the model enables the inclusion of interactions in the equation. For instance, the interaction of a year dummy variable with the key E-commerce variable will be included to see if the effect of E-commerce had changed over studied time period.

Based on previously mentioned literature (Table 2.1), the hypotheses are:

Hypothesis 1: E-commerce has a significant positive effect on the transaction

price of logistics property.

Hypothesis 2: The effect of E-commerce increases with decreasing size of the

property.

(38)

Chapter 5 Results

In this section, different models are estimated. For the interpretation, if it is not stated differently, 5% significance level is used. According to Wooldridge (2015), the results of pooled cross sectional model are valid only if the Multiple Linear Regression assumptions are satisfied. Hence, a brief discussion of the assump-tions (linearity, multicollinearity and heteroscedasticity) is presented before the estimation.

5.1 Assumptions

5.1.1 Linearity and Normality

The relationship between the independent and dependent variables needs to be linear in parameters. To satisfy this assumption, it is important to check for out-liers as it was described in Subsection 3.4.1. The variables Price and Size were winsorized at 5% significance level as described in Subsection 3.4.1. The P-P plot presented in Figure 5.1 was used to assess whether the data are approximately normally distributed. The cumulative probabilities are not significantly different from the straight line.

5.1.2 Homoscedasticity

The Breusch Pagan test was conducted after each regression. In most cases, the null hypothesis of homoscedasticity was rejected, showing the presence of het-eroscedasticity. Heteroscedasticity does not cause bias or inconsistency in the

(39)

5. Results 31 Figure 5.1: P-P Plot 0 .0 0 0 .2 5 0 .5 0 0 .7 5 1 .0 0 N o rm a l F [( P ric e_ w -m )/ s] 0.00 0.25 0.50 0.75 1.00

Empirical P[i] = i/(N+1)

Note: This figure shows P-P Plot of windsorized variable Price. P-P Plot is a graphical method to asses whether the data are approximately normally distributed. The data are plotted against a normal distribution. Deviation from the straight line implies de-viation from normality. In this case, the cumulative probabilities are not significantly different from the straight line.

(40)

5. Results 32

OLS estimator, nevertheless, standard errors and t statistics are not valid anymore (Wooldridge, 2015). To correct for heteroscedasticity, robust standard errors were used.

5.1.3 Multicollinearity

With a presence of perfect linear relationship among the independent variables, the coefficients may not be reliable. To detect multicollinearity, variance infla-tion factor (VIF) was calculated after each regression. As a rule of thumb, the VIF should not exceed the value 10 (Thompson et al., 2017).

Also, correlation matrix between examined independent variables was calcu-lated and can be seen in Appendix –Table A.2. In case of dummy variables, it is important to avoid multicollinearity among the explanatory variables by exclud-ing the base group from the regression because of the dummy variable trap.

Multicollinearity can result in unstable estimates of coefficients in the model. One way how to mitigate the potential problems introduced by the presence of multicollinearity is centring the variables prior to the formation of product terms (Jaccard and Turrisi, 2003). Nevertheless, such transformation does not have to be sufficient to interpret interaction effects and is therefore left for further research.

5.2 Estimation

5.2.1 Main Hypothesis

The final models are presented in Table 5.1. Different specifications of the model were tested. For instance, the variable Size was transformed into natural loga-rithm. Using logarithms on both sides of the equation ensures linearity in param-eters which is one of the crucial assumptions of the OLS. The model is therefore useful when the relationship between these two variables is nonlinear in param-eters.

Moreover, the interpretation of coefficients is in percentages. Both specifica-tions (log-level and log-log) provided a significant positive effect for the variable Size, nevertheless, the R squared almost tripled in case of log-log model and Root MSE was significantly lower. Since the specification between Price and Size is log-log, the interpretation is following (Wooldridge, 2015):

(41)

5. Results 33

The specification between other independent variables and dependent variable is in a form of semi-log model. Consequently, the interpretation is following (Wooldridge, 2015):

%∆Pr ice = 100βi %∆Xi

While comparing the models, R2 and adjusted R2are relevant. It is impor-tant to check whether all models have fairly high value of R2. It means that the model explains a decent part of variation of the dependent variable around its mean. After adding more independent variables, R2never decreases (Kumar and Srivastava, 2004). Therefore, the significance of new added variables is important as well as adjusted R2. However, most of the models are estimated with robust standard errors that are generally attributed to (White, 1980). Adjusted R2shows the same metric but is adjusted by the number of variables and observations. The more observations and less variables, the closer adjusted R2and R2are.

In Table 5.1, column 1 presents model including physical and locational char-acteristics and the variable E-commerce. Adjusted R2is approximately 66% , which is sufficient in comparison to other mentioned studies (for example Beekmans

et al. (2014), Adj R2= 37% or Miles et al. (1990). Adj R2= 68 %). The coefficient of E-commerce is 0,007 meaning that the increase in one unit of independent vari-able, increases the transaction price about 0,7%. One unit of E-commerce can be interpreted as 1% increase in E-commerce (100 basis points) because it is mea-sured in hundreds of basis points. The coefficient is statistically significant at 1% significance level. Other significant variable in the model is Size with coefficient 0,821. The increase in Size about 10% increases the price about approximately 8,21%. The effect of E-commerce is lower than the effect of size but quite high. For instance, 0,7% increase of average price means the increase about 100 000 EUR. Nevertheless, it is likely that the model suffers from omitted variable bias due to lack of control variables. The same regression omitting E-commerce can be seen in Appendix. In the model including E-commerce, the adjusted R2is about 0,05% higher.

Despite having cross sections, the data span covers 10 time periods. To take this fact into account, time dummy variables are added into the regression in column 2. The adjusted R2slightly increased compared to the first model. The omitted category is year 2007. The significance and value of variables’ coeffi-cients remained almost the same. Most of the year variables are significant ex-cept for 2008, 2015 and 2016. All signs of coefficients estimated for time dummy variables are negative. In year 2008, the prices decreased by about 5,6%

(42)

com-5. Results 34

pared to 2007. However, a positive time trend is observed since 2011 because prices were increasing with a diminishing effect. Column 3 shows the same model with the inclusion of the variable of interest, commerce. The inclusion of E-commerce in column 3 causes the increase of adjusted R2. Furthermore, the ef-fect of E-commerce is positive and significant even at 1% significance level with lower magnitude than in column 1. 1% increase (100 basis points) of individuals purchasing online pushes the price up approximately by 0,015%. The variables Size and Flex are still not statistically significant.

One way how to rearrange the model is replacing time dummy variables by market characteristics. These independent variables control for the time effect because they are different for each country every year. Therefore, they represent a proxy variable for the market situation. The inclusion of both metrics – time and market variables could result in multicollinearity, which is not desirable. In Chapter 3, market characteristics were presented and their choice and specifica-tion were driven by previous findings - Table 3.1.

To avoid multicollinearity problem, the correlation matrix was calculated and is presented in Table A.2. In the final model specification, no variable was omit-ted due to high multicollinearity. The highest coefficient of correlation can be seen in case of Population and Density change with the value of 68%. By looking at pairwise correlations, we can analyse the relationships between independent variables.

For instance, lagged GDP is negatively correlated with unemployment growth. This fact can be described by Okun’s Law. Huang and Yeh (2013) confirmed this negative relationship both in short run and long run using country and state-level data.

Hoag (1980) used the variable Volumes of industrial sales in his study. In this thesis, the variable Properties is used instead because of easily available data for selected countries. Properties and Volumes are similar metrics representing the total amount of sold properties in a given year and country–either in units or eu-ros. For the purpose of analysis, variable Properties (the average of sold properties in one quarter) was modified as a ratio. Properties ratio is the average number of sold properties in one quarter in a given country divided by the average number of sold properties in one quarter in all European countries (Section 3.2).

Different specifications of market characteristics were tested. Thereafter, an appropriate interpretation, percentage changes, nominal changes or ratios were chosen. All variables are described in Table 3.1 in detail. Column 4 shows the

Does the real estate market for Logistics assets benefit from the growth of E-commerce? : hedonic pricing modelling in Europe