E-commerce: the end of offline retail?

(1)

Andreas Vikström Spring 2019 Master I, 15 ECTS Master’s in Economics

E-commerce: the end of offline retail?

A quantitative study on how online shopping habits affect the

structure of the retail market.

Name: Andreas Vikström

(2)

I

Acknowledgements

Writing this thesis has been really interesting and educational. I would like to thank my thesis supervisor Jurate Jaraite-Kazukauske. She was always keen to help whenever a problem arose. I would also like to express my profound gratitude to my family and friends for providing me with unfailing support during the last couple of months. This accomplishment would have been difficult without you, so thank you.

With that said, I would like to wish you a continued pleasant read.

Sincerely,

(3)

II

Abstract

The internet has provided us with a lot of things. One major thing is the convenience of buying goods online and get them delivered to us. Not only does the internet contain everything one might desire, but also where one can acquire it for the lowest possible price. In this thesis, this growing phenomenon, called e-commerce, is empirically studied. The main purpose is to examine the effect of e-commerce on the retail market in Scandinavia (Sweden, Norway and Denmark). The thesis includes fixed and random effects models that describe the effect of e-commerce on the number of offline enterprises and fixed effect models that describes the effect of e-commerce on the average size of offline enterprises.

(4)

III

1. Introduction

In this section, the subject of this study – e-commerce – will be presented. There will be given a brief background on how e-commerce has developed over the years and how the structure of the retail market has changed over the same period. The final part of the section will consist of a description of the purpose of the study and the research question.

The internet has provided us with a lot of things. One major thing is the convenience of buying goods online and get them delivered to us. Not only does the internet contain everything one might desire, but also where one can acquire it for the lowest possible price. This thesis will empirically study this growing phenomenon, called e-commerce, and the effect of it on the retail market structure in Scandinavia. This will be done in two ways. First, the effect of e-commerce on the number of retail enterprises will be examined, and secondly, the effect of e-commerce on the average size of the retail enterprises. All this will be done with fixed and random effects regression models using regional data for the period 2009 – 2016.

The layout of this thesis will look like follows; First, a brief background on how e-commerce and the retail market structure has developed and changed over the years. Thereafter, the theoretical background and a review of previous studies will be given in section 2. This is followed by a description of the method and data material in section 3. Furthermore, the results in section 4 are presented, followed by discussion and conclusions in section 5.

1.1 The e-commerce development in Scandinavia

Every year since 2008, the Swedish postal company PostNord writes a report about the e-commerce market in the Nordic countries; Sweden, Norway, Denmark, and Finland. In these reports, one can read about how the e-commerce market is developing in each country. The information provided in this section is mainly gathered from these reports.

(7)

2

Sweden is the leading country in the Nordics when it comes to online shopping. In 2016, approximately 67 per cent of the population was purchasing something online every month. The total turnover added up to 84.8 billion SEK. In 2009, that number was 28.1 billion SEK (201.8 per cent increase). When it comes to what the Swedish people buy online, clothes and shoes are in the lead (34 per cent). In second place is media products (30 per cent), and in third, beauty and health products (25 per cent). This is for the year 2016.

Denmark is, like Sweden, a mature e-commerce nation. Around 63 per cent of the population was purchasing goods online every month in 2016. The total turnover was 51 billion SEK, which is an increase of 107.3 per cent from 2009. Just like in Sweden, clothes and shoes are the most popular online-shopping products (28 per cent). After that, it is electronic products (23 per cent), and media products (22 per cent).

Not surprisingly, Norwegian people are very similar to their Scandinavian neighbours when it comes to online shopping. Norway is a nation with a lot of purchasing power. On average, each online consumer spends 1730 SEK per month (2016). That was more than both Swedish and Danish consumers. This gave a total turnover of 51.3 billion SEK, a 108.5 per cent increase from 2009. Clothes and shoes are the most popular products (28 per cent). Media products come in second place (27 per cent), and electronic products in third (21 per cent).

The reasons why Scandinavian people choose to order their goods online is also presented in these reports. First and foremost, it is the convenience of being able to shop whenever they can that brings the people online. It saves the customers time, which is appreciated by many. The lower price that e-commerce brings is not surprisingly a big reason, and of course the much larger and better range of products.

1.2 The retail market structure

(8)

3

meaning an increase in the number of enterprises would be natural. The changes in the market structure differ across different sub-sectors within retail. For example, in the clothing and shoe sector, a reduction in the number of enterprises was seen in all countries. In Denmark, the count went down from 13257 to 12253, in Sweden from 22731 to 21647, and in Norway from 13760 to 12299. Regarding the number of people employed, an increase was seen between 2009 and 2016. If one looks at it in relative terms to the population in the age span 15-74, however, this increase was not too large. In the clothing and shoe sector, the change was almost non-existent and even negative in Sweden and Norway.

In Figure 1, Figure 2 and Figure 3 presented below, the trends that are described above are shown. In the left charts, Employment is on the LHS Y-axis, while on the right charts, Number

of enterprises is. Internet purchase is on the RHS Y-axis in both charts. The variables in the

figure’s description are as follow.

EMPLOYMENT = The total number of people employed in the retail sector.

INTERNET PURCHASE = Percentage of the population that purchased something online in the last 12 months.

(9)

4 Figure 1 – Denmark Figure 2 – Sweden Figure 3 – Norway 50 60 70 80 90 300000 305000 310000 315000 320000 325000 330000 335000 2009 2010 2011 2012 2013 2014 2015 2016 IN TE R N ET PURCHA SE EM PLO YME N T YEAR

EMPLOYMENT INTERNET PURCHASE

50 60 70 80 90 150000 170000 190000 210000 230000 250000 2009 2010 2011 2012 2013 2014 2015 2016 IN TE R N ET PURCHA SE EM PLO YME N T YEAR

50 60 70 80 90 210000 215000 220000 225000 230000 235000 2009 2010 2011 2012 2013 2014 2015 2016 IN TE R N ET PURCHA SE EM PL O YM EN T YEAR

50 60 70 80 90 2009 2010 2011 2012 2013 2014 2015 2016 29000 29500 30000 30500 31000 31500 IN TER N ET P UR SH A CE YEAR N UM BE R O F E N TE R PRIS ES

ENTERPRISES INTERNET PURCHASE

50 60 70 80 90 68000 69000 70000 71000 72000 2009 2010 2011 2012 2013 2014 2015 2016 IN TER N ET P UR SH A CE N UM BE R O F E N TE R PRIS ES YEAR

ENTERPRISES INTERNET PURCHASE

50 60 70 80 90 36500 37000 37500 38000 38500 2009 2010 2011 2012 2013 2014 2015 2016 IN TER N ET P UR SH A CE N UM BE R E N TE R PRIS ES YEAR

(10)

5

As one can see, the trends are a bit different across the countries. In Denmark and Norway, the trend for the number of enterprises is negative, while it is positive in Sweden. The trend for employment in the retail sector is positive for Norway and Denmark, but negative in Sweden. Note that these trends are different depending on which sub-sector are examined.

1.3 Purpose of the study

The purpose of this study is to empirically examine the effects of e-commerce on the retail market in Scandinavia. It should be treated as an extension of the previous work on the matter, and as a strategic guideline for companies that might get affected by the growth of online shopping. The study will include models that describe the effect of e-commerce on the number of retail enterprises and models that describes the effect on the average size of retail enterprises. This will be done for the different sub-sectors in retail.

1.3.1 Research question

What is the effect of the growth in e-commerce on the number of enterprises and the average size of enterprises in the retail sector and its sub-sectors in Scandinavia?

(11)

6

2. Theoretical background

In this section, the theoretical background and some earlier studies regarding this topic will be presented. The section is divided into three parts. The first part will cover the effects of e-commerce on prices of goods, the second will cover the effects on cost, and the last will cover the effects on the market structure.

2.1 Effect of e-commerce on prices

There are plenty of earlier studies regarding this topic. If one looks at it from a theoretical perspective it is quite easy to conclude that e-commerce has a negative effect on prices of goods. The reason for this is simply the reduction in the firm's costs as they no longer need physical stores in order to sell their goods. A lower cost means a downward shift of the marginal cost curve, which in turn is leading to a lower profit-maximizing price. Although this is the case under very strong assumptions, many studies have shown that many markets in real life are somewhat characterized by this behaviour. Clay, Krishnan, and Wolff (2001) did a study about the effect of e-commerce on prices on books. Their study shows that prices drop due to the introduction of online books stores. A similar result was shown by Brynjolfsson and Smith (2000). It is not only the prices of books that are affected by the introduction of internet sales. According to Sengupta and Wiggins (2006) study, airline ticket prices are affected negatively by online sales. Also, the online service could help consumers to search for and purchase cars for the price which, on average, is lower than the price paid by consumers that do not use the internet when buying a car, which is shown by Scott Morton, Zettelmeyer, and Silva-Risso (2001).

According to Lieber and Syverson (2011), the consequences of e-commerce is a more competitive market. In their paper, they discuss how this can be beneficial for firms with lower marginal costs than their competitors, simply because they can afford a lower price. They also discuss how the competitor firms act in order to not get forced out of the market. In fact, many firms with higher costs than their competitors are focusing on emphasizing the brand and bundled services instead of the price of the good.

(12)

7

well developed, price dispersion still exists in the most online market. This is shown by Baye, Morgan, and Scholten (2007). The two authors that were mention earlier, Smith and Brynjolfsson, also made a study about the online book market in 2001. With data from an online price comparison site, they found that brand is important for consumers when deciding where to buy their books. Consumers are on average willing to pay more in order to purchase from a well-known retailer, for example, Amazon. The reason is simply the reliability of the quality, and shipping times.

Why does the brand matter for the consumer? The answer is the information asymmetries that comes with the online market. In an offline world, the consumer can easily walk into a store and physically examine the good they intend to buy. With online stores, this is not possible, and hence uncertainty arises. Consumers may, therefore, choose the online retailers that they trust. This was experimentally tested by Resnick et al. (2006). They let one seller with a high rating and one with no rating sell the matched lots of postcards. It turned out such that the winning bids were around 8 per cent higher for the high rated seller.

2.2 Effect of e-commerce on cost

(13)

8

It is not only the cost for firms that are reduced with e-commerce. Since the internet is providing consumers with information web sites, price comparison sites and review sites, the consumer search cost is reduced. Instead of travel to all the stores that are selling the goods demanded, the consumer can simply just spend a couple of minutes online and find all the information he/she needs (Lieber and Syverson, 2011).

2.3 Effect of e-commerce on market structure

To summarize part 2.1 and 2.2 above, according to earlier studies, e-commerce has a negative effect on prices. Price is however not the only thing that customers are looking at when making purchasing decisions. Brand and shipping times are important as well and many are willing to pay more for this. Both online firms and customers that are using the internet when purchasing goods are facing a lower cost. For firms, it is distribution cost that is reduced and for customers it the search cost that is reduced. All these outcomes that are presented above will in turn influence the market structure. This is discussed by Lieber and Syverson (2011). As mentioned in part 2.2, with the help of the internet, consumers can easily find the goods they are looking for to the lowest price. This means that firms that can afford to push down the prices, meaning firms with lower costs, will gain market shares. In turn, competitive firms with higher costs will lose market shares. Some firms will even be forced out of the market.

Goldmanis et al. (2010) did a study on this matter. In their model, firm heterogeneity comes from differences in marginal costs1. Their results are consistent with previous studies, meaning that if the market is opened to online sales, the prices are reduced. Their model predicts that e-commerce introduction should shrink the firms with higher costs, and sometimes put them out of business. At the same time, the market shares will shift to firms with lower costs. The model’s predictions were tested in three different industries; travel agencies, bookstores, and new auto dealers. The predictions are consistent in all these industries. As e-commerce is growing, the number of firms with higher costs drops. The effect on firms with lower costs is not significant or even positive.

1_{Note that Goldmanis et al. (2010) are using size of the firm as a proxy for costs. Earlier studies show that higher}

(14)

9

3. Methodology

In this section, the methodology used in the study will be presented and described. It will contain a description of the data used, empirical models and relevant diagnostic tests. In addition to this, eventual problems that could affect the models' results will be discussed and dealt with.

3.1 Data

The data used for the empirical study is gathered from the official database Eurostat and put together in a panel on a regional level. More information on the variables can be found later in this section. The data spans over the year from 2009 to 2016 and contains 8 regions from Sweden, 5 from Denmark and 7 from Norway (NUTS 2). This gives a total of 176 observations. The data do not have any dropouts; hence the panel is strongly balanced.

Eurostat is using the NACE classification code when describing different industries. NACE is divided into four different levels ofhierarchy. Level one contains 21 sections and is identified with letters between A and U. Level two and three are identified with numerical codes on a two- and three-digit level (Eurostat). This study will work with data on a three-digit level from section G, which is wholesale and retail sales.

NUTS is a system which is used to identify different regions within the European Union. Like the NACE classification, NUTS come with different levels of hierarchy; NUTS 1, NUTS 2 and NUTS 3 (Eurostat). This study will work only with regions on a NUTS 2 level. See appendix for a list of all regions included.

(15)

10 3.1.1 Variable description

The data set contains two dependent variables. The first dependent variable represents the total number of enterprises in the G47 class and its subclasses on a three-digit level, per region. The class contains 9 subclasses, G47.1 – G47.9. Each subclass represents a different activity of an enterprise. Since the study is focusing on the effect of e-commerce on the ‘offline’ retail industry, the two subclasses in the G47 class that are covering online sales and sales not in stores are excluded from the dataset. This study will focus on the enterprises in class G47 that work with specialized retail sales in stores. That is, subclass G47.2 – G47.7. However, since the data for G47.2 and G47.3 are not complete, these are unfortunately not included in the study. The names of the variables for each class and subclass are enterprises47 and enterprises474-

enterprises477. In table 3.1 below, the included classes and their descriptions are presented.

Table 3.1

Subclass Description

G47 Retail trade, except motor vehicles and motorcycles

G47.4 Retail sale of information and communication equipment in specialized stores G47.5 Retail sale of other household equipment in specialized stores

G47.6 Retail sale of cultural and recreation goods in specialized stores G47.7 Retail sale of other goods in specialized stores

Source: Eurostat

The other dependent variable represents the number of employees per enterprise in the G47 class and its subclasses on a three-digit level, per region. The variable was generated by dividing the number of people employed in each class by the number of enterprises in the same class. The names of the variables are emp_ent47 and emp_ent474-emp_ent477.

The main independent variable, that is used in all models is internet_purchase. This variable represents the total number of individuals who ordered goods or services over the internet in the last 12 months, presented as a proportion of the total population in each region. This variable can be found in some earlier literature (see for example Goldmanis et al, 2010).

(16)

11

dividing the total employment in the age span 15 to 74, with the total population in the same age span. It is used to control for economic patterns.

3.1.2 Descriptive statistics Table 3.2 Variable Number of observations Mean Standard deviation Min Max enterprises47 ₁₆₀ _6897.96 _3751.96 ₂₆₁₃ _{16 848} enterprises474 ₁₆₀ _247.77 _164.5 ₅₀ ₇₂₈ enterprises475 ₁₆₀ _954.08 _492.76 ₃₄₂ ₂₂₅₃ enterprises476 ₁₆₀ _559.19 _303.81 ₂₀₂ ₁₃₁₆ enterprises477 ₁₆₀ _2411.38 _1253.77 ₇₂₇ ₅₃₉₀ emp_ent47 160 5.44 1 3.99 8.81 emp_ent474 160 4.36 1.1 2.35 7.94 emp_ent475 160 6.03 1.15 3.35 9.8 emp_ent476 160 4.4 1.24 2.52 8.32 emp_ent477 160 4.05 0.67 2.92 6.06 totemp_ratio (%) 160 66.06 3.21 60.41 73.69 internet_purchase (%) 160 71.82 6.13 56 84

In table 3.2 above descriptive statistics of the data set are presented. In the table, one can see that it does not exist any extreme values for any variable. There are 160 observations for each variable, which means there are no dropouts in the data. In many cases, however, the mean value is much further away from the highest value than the lowest. This will be discussed more in 3.1.3 below.

3.1.3 Skewness

(17)

12

This is resulting in a data set that is skewed to the right. By taking the logarithm of each variable some of the skewness in the data disappears.

3.2 Model specification

The models that are estimated in this study are quite like those in the study by Goldmanis et al. (2010). Two effects are to be modelled to answer the research question; the effect of e-commerce on the number of enterprises in retail (1), and the effect of e-e-commerce on the number of employees per enterprise in retail (2). This will be done for each retail class included in the study. In total there will be six different models – three models to answer research question 1 and three models to answer research question 2. More detailed information on each model can be found in section 3.2.2. Model of type (1) and (2) with the different dependent variables are presented in general form below.

The general form for the model to answer research question 1

𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑠𝑖𝑡= 𝑓(𝑖𝑛𝑡𝑒𝑟𝑛𝑒𝑡_𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑖𝑡, 𝑡𝑜𝑡𝑒𝑚𝑝_𝑟𝑎𝑡𝑖𝑜𝑖𝑡)

where i = 1, 2, …, 20 and t = 2009, 2010, …, 2016.

__________________________________________________________________________

The general form for the model to answer research question 2

𝑒𝑚𝑝𝑙𝑜𝑦𝑚𝑒𝑒𝑠/𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑖𝑡= 𝑓(𝑖𝑛𝑡𝑒𝑟𝑛𝑒𝑡_𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑖𝑡, 𝑡𝑜𝑡𝑒𝑚𝑝_𝑟𝑎𝑡𝑖𝑜𝑖𝑡)

where i = 1, 2, …, 20 and t = 2009, 2010, …, 2016.

___________________________________________________________________________ The first model describes the relationship between the growth of e-commerce and the number of offline enterprises in the retail industry. The index i describes each panel, which in this case are regions, and t describes the time period, which goes from 2009 through 2016. The variable

enterprises represent the number of enterprises in the different retail classes that are included

(18)

13 3.2.1 Fixed effects vs Random effects

In the data set that is used for this study, there are 20 heterogeneous regions. If all heterogeneity could be controlled for, a simple multiple regression would probably be the best-suited model to answer the research question. However, all characteristics that differ between regions cannot possibly be included in the model and hence a different model must be used. One model that controls for these omitted variables is the fixed effect regression model. This model control for all heterogeneity between the regions. The characteristics must, however, be constant over time. Otherwise, the model is inconsistent (H. Greene, 2012)

An alternative model to the fixed effect model is the random effect regression model. Unlike the model with fixed effects, this model is considering all the unobserved heterogeneity between the regions to be random. This model assumes that the characteristics that are time invariant to be strictly uncorrelated with the independent variables that are included in the model2_(H.

Greene, 2012).

To decide which model is best suited for the given data set, the Hausman specification test was used. This test is basically comparing the estimated coefficients from the fixed effect model with those from the random effect model. The results of the Hausman test differs across the different classes. In most models, random effects are to be used, but in two models, fixed effect is to be used. (See appendix for more details about the tests.)

In addition to the models with regional fixed and random effect, models with time fixed effects will be estimated. This will be done by adding the dummy variables for each year into the models. There will also be models where an interaction term between the country dummies and the internet_purchase variable is included. With this interaction term included, country-specific effects of e-commerce can be displayed. A Hausman test was used for all these models as well. The results differ here as well. With the year fixed effect included, random effects are to be used in most models (with two exceptions). With the interaction term included, fixed effects are to be used in most models (with one exception). All models are presented in section 3.2.2 below.

(19)

14 3.2.2 Empirical models

Models with random effects

(20)

15 2010 = 1 𝑖𝑓 2010, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 ⋮ 2016 = 1 𝑖𝑓 2016, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝐷𝐸𝑁𝑖𝑛𝑡 = 𝑐𝑜𝑢𝑛𝑡𝑟𝑦 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑒𝑓𝑓𝑒𝑐𝑡 (𝐷𝑒𝑛𝑚𝑎𝑟𝑘) 𝑁𝑂𝑅𝑖𝑛𝑡 = 𝑐𝑜𝑢𝑛𝑡𝑟𝑦 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑒𝑓𝑓𝑒𝑐𝑡 (𝑁𝑜𝑟𝑤𝑎𝑦)

Models with fixed effects

(21)

16 where 𝑌 = 𝑙𝑜𝑔𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑠47 𝑙𝑜𝑔𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑠474 𝑙𝑜𝑔𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑠476 𝑙𝑜𝑔𝑒𝑛𝑡𝑒𝑟𝑝𝑟𝑖𝑠𝑒𝑠477 2010 = 1 𝑖𝑓 2010, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 ⋮ 2016 = 1 𝑖𝑓 2016, 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝐷𝐸𝑁𝑖𝑛𝑡 = 𝑐𝑜𝑢𝑛𝑡𝑟𝑦 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑒𝑓𝑓𝑒𝑐𝑡 (𝐷𝑒𝑛𝑚𝑎𝑟𝑘) 𝑁𝑂𝑅𝑖𝑛𝑡 = 𝑐𝑜𝑢𝑛𝑡𝑟𝑦 𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑒𝑓𝑓𝑒𝑐𝑡 (𝑁𝑜𝑟𝑤𝑎𝑦)

In the two tables above, the complete models that are used for the empirical study are presented. The estimated coefficient 𝛽1 is, given that the other variables in the model are held constant,

the effect of internet_purchase on Y. 𝛽₀ and 𝛼_𝑖 are two constants. 𝛽₀ are the same for all regions while 𝛼_𝑖 differs between regions. Note that it is not indexed by time, since the heterogeneity among regions that are not included in the models are considered time invariant. Also, note that there is no 𝛼𝑖 in the models with random effects. In model 1c and 2c, country specific effects

are modelled. The reference country is Sweden, and the effect of internet_purchase in Swedish regions is represented by is 𝛽₁. 𝛽₁₀ and 𝛽₁₁ represents how the effect in Danish and Norwegian regions differs from the Swedish regions. If the coefficient is positive, the effect of

internet_purchase in that country is larger, and vice versa. 𝑢_𝑖𝑡 and 𝑣_𝑖𝑡 are random terms that captures all the factors, except those included in the models, that explains the variance in Y. 𝑢𝑖𝑡

is in the models with fixed effects and 𝑣_𝑖𝑡 is in the models with random effects3. Since all the variables are in their logarithmic form, coefficients can be interpreted as elasticities, meaning if the independent variable x increases with one per cent, Y will increase/decrease with 𝛽_𝑥 per cent. In this case however, one must be a little bit careful with the interpretation. This is because the variable internet_purchase is already written in percentage form. The coefficient for this variable must therefore be interpreted as follows; If internet_purchase increases with 1 per cent, for example from 60 to 60.6, Y will change with 𝛽₁ percent. One thing that should be noted

(22)

17

however, is that it is not the size of the coefficients that are interesting, but rather the sign before them.

3.2.3 Assumption

All models with regional fixed effects are using OLS (ordinary least squares) when estimating the coefficients. For these estimates to be valid there are some assumptions that must be met. One assumption is that the error term is normally distributed with a constant variance over all the values of the independent variables. If the variance is not constant heteroskedasticity exists in the data set. The consequences of this are that the standard errors for the estimated OLS coefficients are wrong which are leading to biased test results. To eliminate any risk of this, robust standard error is used in the models. These standard errors are consistent with both homoskedasticity and heteroskedasticity (Stock & Watson, 2015). Another assumption is that the independent variables are uncorrelated with the error term. If a correlation exists both between the independent variables and the dependent variable, omitted variable bias (OVB) exists in the model (Stock & Watson, 2015). A simple solution to this is to add the omitted variable to the model, but since not everything can be measured this may sometimes not be a possibility. This means that there will most likely always be some variables causing problems to the model. One thing that may cause OVB, is region-characteristics that differs between the panels in the data set. By using a fixed effect model, these differences are controlled for and some of the risks of OVB is therefore removed. Note that these differences are not allowed to vary over time. If they do, they should preferably be included in the model. But as mentioned earlier in this section, it is not possible to include and control for all heterogeneity that exists.

(23)

18

at the literature that is discussing why people buy goods online, for example, Lieber and Syverson (2011), the number of enterprises in a region or town is not highlighted as a reason for people buying their goods online. It is rather the convenience of finding information, reviews, and prices all in the same place that makes customers buying their goods online. One can, therefore, argue that neither the number of enterprises or the size of enterprises does affect the internet purchase habits, and hence no endogeneity should exist in the model.

Two other assumptions that must be met is that no large outliers exist in the data set and that there is no perfect multicollinearity between any of the variables. Both assumptions are met.

(24)

19

4. Results

In this section, the empirical results will be presented. The results will be displayed in tables and briefly described in words. A more comprehensive discussion will be given in section 5.

Table 4.1

Model 1a Variables enterprises47 (RE) enterprises474 (RE) enterprises475 (RE) enterprises476 (RE) enterprises477 (RE) internet_purchase -0.105** (0.045) -1.108*** (0.114) -0.432*** (0.054) -0.426*** (0.054) -0.268*** (0.025) totemp_ratio 0.411 (0.151) -0.386 (0.717) -0.242 (0.243) -0.230 (0.466) 0.346 (0.244) Observations 160 160 160 160 160 R2 0.20 0.46 0.48 0.28 0.3

Model 1b, with year fixed effects

Variables enterprises47 (RE) enterprises474 (RE) enterprises475 (RE) enterprises476 (RE) enterprises477 (RE) internet_purchase -0.074 (0.059) -0.204* (0.117) -0.130** (0.054) 0.101 (0.066) 0.004 (0.035) totemp_ratio 0.458** (0.186) -0.292 (0.373) -0.256 (0.182) -0.245 (0.250) 0.412*** (0.116) Observations 160 160 160 160 160 R2 _0.30 _0.84 _0.76 _0.77 _0.82

Model 1c, with year fixed effects and country-specific effects

Variables enterprises47 (FE) enterprises474 (FE) enterprises475 (FE) enterprises476 (FE) enterprises477 (FE) internet_purchase 0.102** (0.053) -0.277*** (0.091) -0.079 (0.073) 0.002 (0.072) 0.046 (0.043) totemp_ratio 0.125 (0.164) 0.236 (0.356) -0.390** (0.168) 0.330** (0.152) 0.335** (0.136) DENint -0.356*** 0.069 -0.187 (0.131) -0.235** (0.094) 0.330** (0.152) -0.065 (0.049) NORint -0.176** (0.066) 0.489*** (0.101) 0.119 (0.087) -0.080 (0.073) -0.064 0.059 Observations 160 160 160 160 160 R2 _0.45 _0.86 _0.81 _0.81 _0.82

Notes: All variables are logarithmic. Clustered robust standard errors are presented in the parenthesis. All values are rounded down to 3 decimals. FE = Fixed effects. RE= Random effects. *, **, ***, significant on a 10%, 5% and 1% significance level

(25)

20

4.1 Model 1a

The results of model 1a are displayed in the first section in Table 4.1. The results are in line with the earlier studies presented in section 2, that is e-commerce has a negative effect on the number of enterprises in the retail sector. In all sub-sectors, the effect is significantly negative on a 1% significance level, while in the entire sector, it is significantly negative on a 5% significance level. A significant effect of totemp_ratio on a 5% significance level can only be found in class G47. The r-squared values are not too high in any of the classes, but not too small either. The coefficients are to be interpreted as elasticises. For example, in the entire retail sector (G47), if internet_purchase increases with 1%, ceteris paribus, the number of enterprises will decrease on average with 0.105%. However, as mentioned earlier, the size of the effect will not be deeply analysed in this paper. What is more interesting is the sign before the coefficients to see whether the effects of e-commerce on the outcome variables are positive or negative.

4.2 Model 1b

The results of model 1b are displayed in the second section in Table 4.1. In this model, year fixed effects are used by adding the dummy variables for each year. The coefficients for these dummies are however not displayed in the table. What can mainly be noted in the results for these models are the loss of significance for the effect by internet_purchase. Only in subclass G474 and G475 is the effect of e-commerce significantly negative on a 10% significance level (in G475 on a 5% significance level also). In the other classes, the effect could be non-existent with year fixed effects. The possible reasons for this are discussed in section 5. The coefficients for totemp_ratio are still low in significance, but with lower standard errors. In subclass G477, the effect went up from non-significant to significant on a 1% significance level.

4.3 Model 1c

The results of model 1c are displayed in the last section in Table 4.1. In this model, country-specific effects are included. The r-squared values went up a little in almost all models with the interaction terms included. In sub-sector G474 it went up as high as 0.86. Because of the interaction term, the interpretation of the coefficients is different from the models above. The coefficient for internet_purchase represents the effect of e-commerce on the number of enterprises in the reference country. In this case, it is Sweden. The coefficients for DENint and

NORint shows how the effects in Denmark and Norway differs from the effect in Sweden. In

(26)

21

coefficients for Denmark and Norway is however significantly negative, which means the effect is smaller or even negative in these countries. The results differ between all classes and so does the effect between countries.

Table 4.2

Model 2a Variables emp_ent47 (RE) emp_ent474 (RE) emp_ent475 (RE) emp_ent476 (FE) emp_ent477 (FE) internet_purchase 0.130** (0.063) 0.118 (0.182) 0.475*** (0.127) 0.310** (0.140) 0.215*** (0.060) totemp_ratio 1.222** (0.471) 0.580 (0.661) 0.271 (0.578) 1.215 (0.863) 0.411 (0.524) Observations 160 160 160 160 160 R2 _0.10 _0.03 _0.15 _0.10 _0.08

Model 2b, with year fixed effects

Variables emp_ent47 (FE) emp_ent474 (RE) emp_ent475 (RE) emp_ent476 (FE) emp_ent477 (FE) internet_purchase -0.067 (0.076) -0.162 (0.200) 0.072 (0.100) -0.363** (0.158) -0.100 (0.065) totemp_ratio 1.013** (0.407) 0.769 (0.736) 0.167 (0.498) 1.127* (0.586) 0.150 (0.295) Observations 160 160 160 160 160 R2 _0.41 _0.09 _0.36 _0.49 _0.62

Model 2c, with year fixed effects and country-specific effects

Variables emp_ent47 (FE) emp_ent474 (FE) emp_ent475 (RE) emp_ent476 (FE) emp_ent477 (FE) internet_purchase -0.129 (0.127) 0.350** (0.160) 0.008 (0.098) 0.058 (0.097) -0.153 (0.092) totemp_ratio 1.071 (0.420) -0.060 (0.640) 0.238 (0.502) 0.240 (0.382) 0.186 (0.287) DENint -0.140 (0.140) -1.067*** (0.336) 0.078*** (0.024) -1.310*** (0.144) -0.174 (0.110) NORint 0.397** (0.139) -0.406 (0.258) 0.065*** (0.020) 0.202** (0.088) 0.405*** (0.109) Observations 160 160 160 160 160 R2 _0.45 _0.21 _0.36 _0.73 _0.69

Notes: All variables are logarithmic. Clustered robust standard errors are presented in the parenthesis. All values are rounded down to 3 decimals. FE = Fixed effects. RE= Random effects.

*, **, ***, significant on a 10%, 5% and 1% significance level

(27)

22

4.4 Model 2a

The results of model 2a are presented in the first section in Table 4.2. As the results of model 1a, these are also in line with the results of the earlier similar studies. Only in subclass G474, no significant effect of internet_purchase on the number of employees per enterprise can be shown. In the retail sector as a whole, internet_purchase has a positive effect on a 10% significant level. In its subsectors, except G474, internet_purchase has a positive effect on a 1% significance level. Only in class G47, a significant effect on a 10% significance level can be shown for the variable totemp_ratio. The r-squared values are quite low in all classes. It varies between 0.08 and 0.16 in the classes where a significant effect of e-commerce is displayed. The coefficients are to be interpreted as elasticities. In the entire retail sector (G47), the number of employees per enterprise will on average increase with 0.115% if

internet_purchase increase with 1%, ceteris paribus.

4.5 Model 2b

The results of model 2b are presented in the second section in Table 4.2. In this model, year fixed effects are included. Just like in model 1b, the effect of internet_purchase on the number of employees per enterprise loses significance with year fixed effects. Similar results were found by Goldmanis et al. (2010). Only in class G476 a significant effect on a 5% significant level is shown. It is however negative, which is not what is expected. This is also discussed more in section 5. Most of the coefficients for totemp_ratio is still low in significance, but with smaller standard error than in the model with no year fixed effects.

4.6 Model 2c

(28)

23

5. Discussion

In this section, the results presented in section 4 are discussed. In addition to this, a discussion regarding the choice of the empirical models and their limitations will be given.

5.1 Discussion

In this thesis, the effect of the growth in e-commerce on the retail sector in Scandinavia has been empirically studied and presented. Like shown in section 4, the results differ between the models. In the model with only regional fixed (or random) effects, a significant effect of e-commerce can be displayed in all classes, while with year fixed effects it cannot. By using year fixed effects, every year is isolated which means that an individual regression is created for each year. The coefficients of the independent variables are however still the same for all regions and should be interpreted as the effect given that everything else is held constant, including the year. This means that only the effect that is not due to any trends across the years are displayed. As mentioned in the results, the significance of the control variable (total employment ratio for the entire region), went up with year fixed effects. What this implies is that there is something that is changing in the economy over the years. What it may be is however not certain, but it may explain some of the effects of e-commerce, hence the lost significance in that variable. Adding dummy variables to a model does in many cases improve the model by controlling for possible omitted variables. The results must, however, regardless be dealt with caution. The coefficients for the dummy variables are not presented in the results, but they are nevertheless worth discussing. By looking at the coefficients it is quite easy to distinguish the trend over the years. In the model where the effect of e-commerce on the number of enterprises are modelled, the coefficients are decreasing on average over the years. In the model that is modelling the effect of e-commerce on the number of employees per enterprise, the coefficients are increasing on average over the years. This means that the effect of e-commerce on the two dependent variables is already partly explained by which year it is. Even though the trend may be explained by e-commerce, the model will not display it. This brings up the question of whether the model with region fixed (or random) effects only is, in this case, the better one of choice.

(29)

24

determine a common effect. This is tested on a country level in model 1c and 2c by using the interaction terms between country dummies and the e-commerce variable. The results vary across the countries and classes. In some classes, a significant difference is displayed, while in some it is not. It is however quite certain that there are some differences across the countries and the classes, but these differences must be studied more thoroughly before any other statements are made.

When comparing the results of this thesis to the results of the similar earlier studies it is evident that there are some similarities. As my estimated models are close to the ones of Goldmanis et al. (2010), their results are those of interest when comparing. Like mentioned in section 2, they studied three markets, bookstores, travel agencies, and car dealers. Year fixed effects were used for all markets, but for travel agencies, only regional fixed effects were included. Like in most cases in this thesis, the effect of e-commerce lost significance on the market for travel agencies when using dummies for each year. The results for car dealers are not interesting to compare with since this is excluded in the data set used in this thesis. The results for bookstores are however more interesting. They suggest that e-commerce has a negative effect on the number of enterprises as a whole and that it has a negative effect on smaller enterprises and a non-existent or even positive effect on larger enterprises. These results are in line with the results of this thesis (from the models without any year fixed effects, and in some cases from the models with year fixed effects).

5.2 Limitations

(30)

25

(31)

26

6. Conclusions

In this section, conclusions about the results are made. In addition to this, some proposals regarding future research are given.

6.1 Conclusions

As the results of this study partly disprove each other in some cases, it is not easy to make a straight conclusion about them. The results for model 1a and 2a confirm the expectations made before the study was started. That is, if more people go online to purchase their goods, the number of “offline” retail enterprises will decrease. The firms that will survive are the bigger ones that can afford a lower marginal cost without being forced to exit the market, hence the average number of employees per enterprise will increase. Whether the model is good or bad is however left unsaid. It still needs some improvements. With year fixed effects included in the models, e-commerce does not show a significant effect in most cases. It is, however, unlikely that the effect is non-existent. Only in subclass G474 (Retail sale of information and communication equipment in specialized stores) and G475 (Retail sale of other household equipment in specialized stores), e-commerce has a negative effect on the number of enterprises with year fixed effects. Regarding the other classes, one cannot say whether the effect is negative, positive, or non-existent. The conclusion for these models may, therefore, be that improvements must be made, by for example adding more variables to control for heterogeneity across the regions.

6.2 Further research

(32)

27

7. References

Baye, Michael & Morgan, John & Scholten, Patrick. (2007). Information, Search, and Price Dispersion, Handbook on Economics and Information Systems, vol. 1, (T. Hendershott, Ed), Amsterdam and Boston: Elsvier, 2007.

Brynjolfsson, E. & Smith, D. M. (2000). Frictionless Commerce? A Comparison of Internet and Conventional Retailers, Management Science, 46, (4), 563-585.

Clay, K., Krishnan, R. & Wolff, E., (2001). Prices and Price Dispersion on the Web: Evidence from the Online Book Industry, Journal of Industrial Economics, Blackwell Publishing, vol. 49(4), pages 521-39.

Goldmanis, M., Hortaçsu, A., Syverson, C., Emre Ö. (2010). E-Commerce and the Market Structure of Retail Industries, Economic Journal, Royal Economic Society, vol. 120(545), pages 651-682, 06

Greene, W. H. (2012). Econometric analysis, 7th edition. Upper Saddle River, N.J., Prentice Hall

Lieber, E. & Syverson, C. (2011). Online vs. Offline Competition, Oxford Handbook of the

Digital Economy, 2012.

PostNord, 2010. Distanshandeln I Norden 2010.

https://www.postnord.com/globalassets/global/sverige/dokument/publikationer/2010/distansh andeln-norden-2010.pdf?amp;epslanguage=sv [Retrieved 2019-04-04]

PostNord, 2017. E-handeln i Norden 2017.

https://www.postnord.com/globalassets/global/sverige/dokument/publikationer/2017/ehandeln -i-norden-2017_se_lowres.pdf [Retrieved 2019-04-04]

(33)

28

Resnick, P., Zeckhauser, R., Swanson, J. & Lockwood, K. (2006). The value of reputation on eBay: A controlled experiment. Exp Econ, 9: 79

Scott Morton, F., Zettelmeyer, F. & Silva-Risso, J. (2001). Internet Car Retailing, The Journal

of Industrial Economics, vol. 49, No. 4, Symposium on E-Commerce (Dec., 2001), pages.

501-519.

Sengupta, A. & Wiggins, S. N. (2014), Airline Pricing, Price Dispersion, and Ticket Characteristics on and off the Internet, American Economic Journal: Economic Policy, 6, issue 1, pages 272-307

Stock, J.H. & Watson, M.W. (2015). Introduction to Econometrics. Updated 3rd edition. London: Pearson Education Limited.

Data

• Eurostat. Regional ICT Statistics (isoc_reg). [Retrieved 2019-04-20]

• Eurostat. SBS – regional data – all activities (sbs_r). [Retrieved 2019-04-20]

(34)

29

8. Appendix

List of regions included in the study

(35)

30

Hausman tests, for the entire retail sector (class G47)

If prob>chi2 < 0.05, fixed effects are to be used.

Model 1a

Test: Ho: difference in coefficients not systematic chi2(2) = (b-B)'[(V_b-V_B)^(-1)](b-B) = 1.75

Prob>chi2 = 0.4161

Model 1b

Prob>chi2 = 0.9983

Model 1c

Prob>chi2 = 0.0001

(V_b-V_B is not positive definite)

Model 2a

(36)

31

Model 2b

Prob>chi2 = 0.0346

(V_b-V_B is not positive definite)

Model 2c

Prob>chi2 = 0.2609

E-commerce: the end of offline retail?