Optimizing marketing strategies MASTER THESIS

(1)

Optimizing marketing strategies

Measuring the effectiveness of marketing channels for generating sales

and assessing the moderating role of price promotions

MASTER THESIS

MSc Marketing Management | Marketing Intelligence

(2)

Optimizing marketing strategies | Master thesis by Stephanie Becker

2 Optimizing marketing strategies

Measuring the effectiveness of marketing channels for generating sales

and assessing the moderating role of price promotions

MASTER THESIS

MSc Marketing Management | Marketing Intelligence

June 26, 2017

Author

1

st

_Supervisor

₂

nd

_Supervisor

Stephanie Becker (S2917920) dr. ir. Maarten Gijsenberg dr. Peter Van Eck stephanie.becker0309@gmx.de m.j.gijsenberg@rug.nl p.s.van.eck@rug.nl Josef-Knettel-Straße 24 Nettelbosje 2 Nettelbosje 2 55411 Bingen 9747 AE Groningen 9747 AE Groningen Germany The Netherlands The Netherlands

(3)

3 Abstract

A new marketing environment shaped by technology developments creates big challenges for today’s marketing managers. An efficient investment of the marketing budget is one prerequisite to succeed in this competitive environment. To do so, marketers must know which advertising channels are most effective for generating brand sales.

This thesis has the goal of developing new insights about the impact of different offline and online advertisement channels in generating sales. Furthermore, the moderating role of a price promotion is examined. More specifically, an investigation has been made to determine whether a price promotion can increase the effectiveness of an advertising channel for generating sales.

Based on previous research it is assumed that the three researched types of advertising have a significant positive effect on sales. Additionally, it is also assumed that a price promotion enhances the effectiveness of a marketing channel on generating sales.

For this investigation, data concerning consumers’ media- as well as purchase behavior towards a multinational soft drink brand is used. The data is provided by the research institute GfK and reflects represents households over an observation period of three months.

The data analysis shows that there is only a significant positive effect of advertisement that is shown on the website of the Dutch television channel RTL on generating more sales. The other investigated online channel, namely Youtube, has a significant negative effect on the sales value. For television advertising, no significant positive effect could be found. For the moderating role of price promotions, significant positive effects could only be found for the two kinds of online advertisement and the purchase value.

These results give marketing managers important insights about which advertising channels can convince consumers to spend more money on a certain brand. However, marketing managers should not take the positive effects of advertising for granted. They need to track which channels are effective for generating sales and for these channels, marketers must develop creative advertising messages to convince the consumer to buy their brand.

(4)

4 Preface

Prior to starting my bachelor studies in BA Communications & Marketing at the International School of Management in Frankfurt, my interest had already been spurred on the subject of marketing. My further work and research into this topic has served to only deepen my interest for marketing. Fortunately, various internships in different countries helped me to turn my theoretical knowledge into practical experience. While working in different companies, my passion for marketing grew and I found out that I am especially interested in the marketing of consumer goods. Consequently, in 2015 I started my master studies MSc Marketing at the University of Groningen. Studying at a very prestigious university in a foreign country with a different educational system was not easy for me in the beginning, but it gave me the opportunity to gain new insights and perspectives while increasing my knowledge base. I challenged myself even more when I decided to do the double track MSc with specializations in both Marketing Management and Marketing Intelligence.

This master thesis represents the last component of my master studies. During the progress of this paper, I especially learned how to work independently and how to handle multiple difficulties. I would like to show my gratitude to some people who have supported me during the creation of this master thesis. Most importantly, I want to thank my family and my friends for comforting me when I experienced difficult times. Furthermore, I am very thankful for the valuable feedback and guidance of my first thesis supervisor dr. ir. Maarten Gijsenberg. Moreover, I want to thank my second supervisor dr. Peter Van Eck for putting time and effort into the assessment of this thesis. A special thank you goes to my two fellow students Sara Van Es and Louise Nyman with whom I regularly exchanged feedback to get the most out of our master theses. The conceptualization of this master thesis has been a great experience for me since I could not only apply the knowledge and methods I learned during my studies but also extend my expertise about the effectiveness of marketing channels. Overall, I found that it complemented perfectly the outstanding education I received in the University of Groningen.

Groningen, June 2017

(5)

5

6 1 Introduction

‘Half of the money I spend on advertising is wasted; the trouble is, I don’t know which half.’

This quote by John Wanamaker from the 1990s reflects the big challenges that marketing managers are facing when it comes to measuring the effectiveness of marketing channels. Today’s marketing environment shows an increase in competition as a result of the rapid development of the internet and the effects of globalization (Deshwal 2015). This new media landscape, which developed from a solely offline model to a diverse offline/online environment with plenty of different channels, leads to changing consumer needs that reflect complexities for marketers (Keller 2014; Kumar & Venkatesan 2005). In this media setting, marketing practitioners are forced to spend their marketing budget carefully to secure the optimal use of traditional and new media channels. To do so, marketers need to track which channels work best for different consumers and they must develop robust marketing programs that work as well as possible for each individual consumer and his shopping preferences (Keller 2014). These are very important tasks, as the goal of marketing is to attract and retain profitable customers to the business (Doyle & Saunders 1990). To reach this goal, advertisements and the channels on which they are delivered to the consumers, need to be effective. Consequently, the effectiveness of advertisement channels is an intermediate marketing goal. According to Lavidge and Steiner (1961), an important determinant for this effectiveness is whether a consumer purchases a product or not.

This thesis investigates the effects of different offline and online channels on the purchase behavior of consumers. More specifically, the effectiveness of two online channels and one offline channel are studied in this research. Furthermore, it is investigated whether there is a moderating role of price promotions on the relationship between seeing an advertisement on a certain channel and the value of the purchase. It could well be that price promotions enhance the effectiveness of an advertisement channel because they can give consumers an additional reason to buy a certain product (Jang et al. 2012).

The main research question of this thesis is the following:

(7)

7

To answer this main research question, underlying research questions are developed: 1. What effect does a contact with an online advertisement have on brand sales? 2. What effect does a contact with a TV advertisement have on brand sales?

3. Which of the three marketing channels has the highest positive impact on brand sales? 4. What is the effect of a price promotion on marketing channel effectiveness?

With the support of the main research question as well as the sub questions, the aim of this thesis is to contribute new insights about the effectiveness of different offline and online media channels for increasing the purchase value of consumers. The second goal is to generate new findings about the moderating role of price promotions on the relationship between seeing an advertisement on a certain marketing channel and purchasing the advertised product for a certain value.

With these insights, marketing managers can estimate more precisely the impact of different marketing channels and consequently they will be able to spend their marketing budget more efficiently. This will lead marketers one step closer to marketing’s ultimate goal of attracting and retaining customers to the company.

For academics, this thesis has the goal of expanding literature about the investigated offline and online channels and their effectiveness for generating sales. Furthermore, this thesis shall broaden researcher’s knowledge about the moderating role of price promotions.

The two investigated online advertisements are shown as pre-roll advertisements1_{placed on} YouTube and on the website of the Dutch television channel RTL. The offline advertisement is a spot that is shown during commercial breaks on Dutch television. Each of the advertisements is investigated separately from each other. It is also analyzed which of these advertisements has the highest positive impact on brand sales and therefore is the most effective one. This thesis is based on a dataset that represents a low-involvement product of a fast-moving consumer goods (FMCG) company. The data is provided by the marketing research firm GfK and consists of daily data of 10703 households covering 90 days from December 2013 to March 2014.

(8)

8

(9)

9 2 Theoretical Framework

Based on the academic literature that is reviewed in this chapter, a conceptual model as well as hypotheses are developed. The chapter starts off by illustrating the conceptual model, followed by an investigation of the meaning of effectiveness. It will then continue with an elaboration on the effectiveness of online and offline advertisement. Lastly, an argument is made for the inclusion of the moderator variable and the control variables.

2.1 Conceptual Model

Figure 1 provides a graphical illustration of the conceptual model and the hypotheses. The model shows how online and offline advertisement is expected to affect offline sales, influenced by price promotions.

2.2 Literature Review and Hypotheses

The following paragraphs contain a literature review about related academic studies. They are fundamental for the hypotheses development.

2.2.1 Effectiveness based on the ‘Information Processing Model of Communications’

A well-thought through mix of different marketing channels that a firm can choose to implement to influence consumers and to convey a certain message, is crucial for generating

(10)

10

sales. It is highly important to consider that consumers respond in different ways to advertising and that these responses can be either affective (feelings), cognitive (thoughts) or behavioral (actions) (Fennis & Stroebe 2010). In this study, the behavioral response is observed and a channel is perceived effective when the action of a purchase is made.

According to Keller’s (1993) ‘Information Processing Model of Communications’ six different steps can be identified when examining how consumers process communications. Only with the realization of each step, an advertisement on a certain channel is effective. Exposure is the first component of the model. This step represents the number of brand-related experiences that have been accumulated by a consumer. The higher the number of exposures to a certain brand, the greater is the consumer’s ability to recognize and to recall this brand (Keller 1993). Attention is the second step, which identifies whether a consumer notices a certain communication. One way to increase the likelihood of attention in a commercial spot is to delay brand identification until the end of the spot (Keller 1993). Nowadays, in a time in which consumers are targeted with thousands of advertisements per day, a communication must stand out to get the attention from a consumer. As a third step, a consumer must process the information in such a way that the message, which the firm wants to convey, is understood by the consumer. This goal is called Comprehension and might be difficult to achieve, especially when a message is ambivalent. Based on the message that a consumer just received and hopefully understood, a company is then aiming for a favorable response, called Yielding. This favorable response means that the consumer sees the benefit of the advertised brand. The next step is called Intentions. Here, the company wants the consumer to have the intention to buy a specific brand. The Behavior of buying that brand is the last component of the model and only when it comes to this last step, an advertisement on a certain channel can be considered as effective (Keller 1993).

Figure 2: Components of Effectiveness based on the ‘Information Processing Model of Communications’

2.2.2 Online advertisement effectiveness

(11)

11

challenges. On the practical side, marketers need to take steps against the lack of accountability that has undermined marketing’s credibility and threatens the standing of the marketing department in firms (Verhoef & Leeflang 2009). For marketing managers, the optimal allocation of the marketing budget is a crucial determinant for the success of their department. However, also for marketing researchers, which represent the theoretical side, it is nowadays highly important and interesting to investigate which of the many different online touchpoints work best (Goldfarb 2013).

In comparison to the academic literature about own-channel advertising effects between offline advertising and offline sales, literature about the cross effects of online advertisement on offline sales, is limited (Dinner et al. 2014). Baxendale et al. (2015) study the impact of different online and offline touchpoints on brand consideration, but not on actual purchase behavior. They find that in-store communications are most influential, followed by peer observation and brand advertising, then word-of-mouth and retailer advertising. In contradiction to the research by Baxendale et al., there are some other studies having sales as a dependent variable. Wiesel et al. (2010) investigate the impact of marketing communication activities on the online and offline purchase funnel and ultimately sales with the example used being a European office furniture supplier. The authors observe that there is a large impact of Google AdWords2_{and a much lower impact of flyers and faxes on sales.} Very similar to Google Adwords, which is investigated by Wiesel et al., is advertisement that is placed on Yahoo, which is researched by Lewis and Reiley (2014). These two authors measure the effectiveness of online advertising on offline sales based on a controlled experiment on Yahoo. The results show ‘positive, sizable and persistent effects of online advertisement on sales’ (Lewis and Reiley 2014, p. 45). The researchers also find that online advertisements increase both the likelihood of purchasing and the average purchase amount. Dinner et al.’s (2014) research also takes brand sales as a dependent variable. In their study, the authors focus on the presence, magnitude and carryover of cross-channel effects for online advertising and traditional media. The paper investigates how advertising expenditures translate into sales. In comparison to this thesis, the research of Dinner et al. elaborates on cross effects in both directions, online to offline and offline to online. For a high-end clothing and apparel retailer, the researchers observe that online advertisement is highly effective and has strong cross effects on offline sales (Dinner et al. 2014).

2_{Google AdWords is an online system that links advertisements with searched terms. This enables accurate presentation of}

(12)

12

In summary, there is some research which measures the effectiveness of different online touchpoints. Literature confirms the positive effect of online advertisement on sales in general. This study investigates whether these findings are also appropriate for a FMCG brand.

The following hypothesis is developed:

H1: A contact with an online advertisement has a significant positive effect on brand sales.

2.2.3 Offline advertisement effectiveness

Since offline advertising, such as TV-, radio- and print ads, has a much longer history than online advertisement, more literature exists for offline than for online marketing. However, the literature, that is interesting for this thesis, often compares the effectiveness of online and offline channels, but does not focus exclusively on offline touchpoints. Yet, there is some literature elaborating on offline channels and therefore it can contribute to the research of this thesis.

The study by Sethuraman et al. (2011) builds on the first empirical generalizations of advertising elasticity from 1984. The authors give insights in how the marketing environment has changed from 1984 to 2011. The researchers focus on advertising from print and television and conclude that the decrease in the effectiveness of offline advertising is based on changes such as increased competition, globalization and the advent of the internet.

(13)

13

on how television, radio, newspaper and internet advertisement affects consumers’ choice of media channels. They ascertain that consumers with high attention for computers use less offline media such as radio, print, and television. In contrast, consumers with high attention for these offline media options have higher utility for using both forms of media, online and offline (Lin et al. 2013).

Another interesting finding is the one of Wilbur (2008). The author found that most television viewers are not interested in television commercials. This disinterest results in aversive behavior which can lead to channel switching as soon as commercial breaks begin (Wilbur 2008).

In summary, prior research regarding the effectiveness of offline media compares several media channels with one another in terms of their effect on attitudes, awareness and purchase behavior. However, no study compares television advertising as a traditional media format to different online media formats and their effectiveness for generating sales. Thus, this study fills this gap by measuring household-level exposure to television advertisements and connecting it to purchase data for these consumers.

As literature states a strong impact of offline advertisement on positive brand attitudes, the following is hypothesized:

H2: A contact with a TV advertisement has a significant positive effect on brand sales.

2.2.4 The role of price promotions

(14)

14

(15)

15

In comparison to literature about the direct effect of price promotions on purchase intentions, literature about the moderating effect of price promotions on the relationship between seeing an advertisement on a certain channel and having certain purchase intentions, is more limited. However, there are some studies that use other independent variables and/or another dependent variable than is used in this thesis, but that also investigate the moderating role of price promotions. One example is the research of Luk and Yip (2008) that elaborates on the moderating role of monetary sales promotion on the relationship between brand trust and purchase behavior. The authors find that monetary promotions can boost sales volume but at the expense of profit margin, since over-reliance on monetary sales promotions can destroy the trust that consumers have in a brand. Ultimately this can lead to brand switching behavior as well as increased price sensitivity. This insight fits very well to the above-mentioned finding by Selvakumar & Vikkraman (2011) who state that the direct effect of a price promotion on brand equity is negative in the long-run.

(16)

16

or to not purchase the product. This whole psychological path is also investigated by Jang et al. (2012). The authors find that both advertisement (in the consideration stage) and price information (in the choice stage) might influence the considerations and choices of consumers (Jang et al. 2012).

Based on these findings of the existing literature, it is highly probable that price promotions have a positive short-term effect on sales. Again, it is important to keep in mind that this thesis investigates sales in the short-term, not in the long-term. Therefore, for this study, the results that researchers find for the short-term are important for the hypothesis development of the moderating role of price promotions. Furthermore, they do not only affect the generation of sales directly (Jang et al. 2012) but are also expected to influence the relation between advertisement and sales during the whole purchase decision process (Jang et al. 2012). Therefore, the variable ‘Price Promotions’ is the moderator in this conceptual model. It can be expected that consumers are confronted with the different forms of advertisement which would increase their purchase intentions, but the purchase value would be even higher in case of a price reduction. Based on this perspective, the following hypotheses are developed: H3a: The effect of a contact with an online advertisement on brand sales is stronger when there is a price promotion.

H3b: The effect of a contact with a TV advertisement on brand sales is stronger when there is a price promotion.

2.2.5 Income and Education

In this research, income and education are incorporated as control variables.

(17)

17

healthier consumption (Hiza et al. 2013). Consequently, it is expected that families with a lower level of education consume a higher amount of soft drinks than families with a higher education.

(18)

18 3 Methodology

After reviewing literature extensively and developing a conceptual model as well as hypotheses, this chapter focuses on the data collection and the data analysis of this research. First, the focus is on the data collection and the data description. Thereafter, results of a first analysis of the demographic and independent variables are given. Then, the model that shall be used for explaining brand sales, is defined. Finally, it is explained what kind of regression is used for the analysis.

3.1 Data Collection and Data Description

Since the data is collected from the research institute GfK, which is an external source, the importance of elaborating on whether the data is good data, is increased compared to if it was self-collected data. According to Leeflang et al. (2015) four different criteria are crucial for this investigation, namely availability, quality, variability and quantity.

Availability of the data refers to a company’s ability to collect data, but also to construct or extract data from company data bases (Leeflang et al. 2015). In this case, the availability of the data is given since it is retrieved from GfK and is made available for the use of this research. Before investigating if the data fulfills the other three criteria, it is important to understand which information the data reflects. Therefore, some insights on the data are given first: The data encompasses two datasets that cover information from 10703 households.

The first dataset represents 23 variables that record media behavior as well as purchase behavior over a period of 90 days (December 2013 - March 2014). Regarding the households, this dataset is time-variable since the media and purchase behavior can change from day to day.

The media behavior is reflected by variables that report which household got reached with what kind of advertisement. Three different kinds of advertising are recorded, namely two online advertisements and one offline advertisement. The two online advertisements are pre-roll advertisement videos placed on YouTube and on the website of the Dutch television channel RTL. The offline advertisement is a spot that is shown during commercial breaks on Dutch television. For those cases in which an advertisement of the soft drink brand is seen by a household, the data also contains information about how often it was seen on a particular day by a certain household.

(19)

19

credibility. In other words, the used methods of the research should measure what they are supposed to measure (Leeflang et al. 2015). Reliability assesses the degree to which a measure is subject to random error (Leeflang et al. 2015). In this research, the media behavior is investigated with passive measurement techniques. For this, electronic trackers were installed in the computers and TVs of each household. Passive measurements have the advantage of giving an accurate depiction of reality (Blumberg et al. 2008). Hence, validity and reliability of the data is given.

The purchase behavior is reflected by variables that report whether, when and how much a household bought of the soft drink brand. A purchase is measured in value (monetary) as well as volume (liters). Even though the soft drink brand is advertised both online and offline, it is only distributed offline. Hence, the dataset only includes offline purchase data.

The second dataset is time-invariable. It encompasses 18 variables that report not only geographic and sociodemographic- but also psychographic characteristics.

Geographic and sociodemographic information are given by variables like Education, Age of housewife, Income, Size of town and Postcode.

Psychographic variables display how many households participated in which measurement panels. 9380 households participated in the measurement panel for online behavior, whereas only 1443 households took part in the measurement panel for offline behavior. 1304 households participated in both the online measurement as well as the offline measurement. For this research, only those respondents are included who participated in both measurement panels. In this way, reliability can be increased. Furthermore, to increase reliability, it is crucial to include only those households for which daily data is available. In total, 1176 households participated continuously (every day) during the 3-months observation period in both the online as well as the offline measurement panel.

(20)

20

Regarding the fourth prerequisite, namely variability, Leeflang et al. (2015) state that it is impossible to measure the impact on the dependent variable if the independent variable does not show any variation. Consequently, the larger the variation within an independent variable, the better the estimation of the effect of an independent variable on the dependent variable (Leeflang et al. 2015). The variation of each variable can be seen in table 1. This table shows the descriptive statistics (N, Minimum, Maximum, Mean and Standard Deviation) of the analyzed variables of this research. Here, it is important to mention that the moderating variable Price Promotion is not included in the table since it is coded as a Dummy variable, with the values 0 and 1, indicating whether a price promotion takes place during a purchase or not. Furthermore, it is crucial to explain that the three independent variables OnlineYouTube, OnlineRTL and OfflineTV are mostly 0, which means that most consumers are not reached by any of the advertisements during the 3-months observation period. This is indicated by the mean value, which is very low for these variables. The deviation value for TV advertisement is higher than the ones for online advertisement since some people – even though only a small few – even get reached four times a day with a TV advertisement. The two control variables Income and Education are recoded from string variables into numeric variables. For income, 1 is indicating an income of less than 700, - per month and 19 stands for an income of more than 3901, - per month. The mean value of 9.06 indicates that the income of the participating households is distributed equally. Education is grouped into seven different categories, ranging from ‘Bo Basisonderwijs’ (lowest) to ‘Hw WO doktoraal’ (highest). The mean value of 4.55 shows that the respondents tend to have a higher education. Consequently, enough variation is given in most variables, however, the amount of variation is low for both online advertising variables. This should be considered when analyzing the impact of the parameter estimates of both online advertisements.

Table 1: Descriptive Statistics

Descriptive Statistics

N Minimum Maximum Mean Std. Deviation

OnlineYouTube 105840 0 2 .0016 .03981

OnlineRTL 105840 0 1 .0005 .02195

OfflineTV 105840 0 4 .07 .286

Income 105840 1 19 9.0612 4.59238

(21)

21 3.2 Analysis of the Sociodemographic and Independent Variables

This chapter consists of two parts. The first part focuses on the data. Here, preliminary insights about the most important sociodemographic as well as the independent variables are given. It is examined to what extent the values of these variables are representative for the values of the population in the Netherlands. The second part of the chapter defines a regression model that is suggested to be used for the analysis of the hypotheses.

As illustrated in chapter 3.1, the original sample of this research contains 10703 households and is reduced to 1173 households due to some inconsistency in the dataset. The most important characteristics are summarized in table 2.

Table 2: Sociodemographic Analysis

Answer Category Frequency Percentage

Characteristic Age of Housewife

< 24 6 0,5% 25 - 34 132 11,2% 35 - 44 227 19,3% 45 - 54 325 27,7% 55 - 74 441 37,5% > 75 45 3,8% Characteristic Education Low education 216 18,4% Middle education 563 47,9% High education 397 33,7%

Characteristic Income Per Month

(22)

22

In most of the participating households, the housewife is between 55 and 74 years old (37,5%). The youngest housewife is 20 years old, whereas the oldest one is over 75 years old. According to Centraal Bureau voor de Statistiek (CBS) (2017), most women in the Netherlands are 47 years old. It can therefore be concluded, that the age of most women in this data is higher than the age of most women in the Netherlands. Furthermore, the CBS (2017) concludes that 23% of the population is younger than 20 years. Hence, this sample exists of relatively more elderly people and relatively less young people compared to the age distribution in the Netherlands. Moreover, 47,9% of all households have a middle education, followed by 33,7% with a high education. These statistics that are generated from all 1173 households match with the data generated by the CBS. Their findings conclude that 60% of all people in the Netherlands had a middle education and 29% had a higher education in 2014 (CBS 2014). The income characteristic which reaches from 700, - to over 3900, - per month is divided equally into ten different groups. Originally, this variable was divided into 19 different groups (see chapter 3.1), however, to provide a clearer structure, it is reduced from 19 to ten groups. Here, the biggest group (19,6%) is represented by respondents who earn between 1501, - and 1900, - per month. According to the CBS, this income bracket can be seen as representative for the Netherlands since most Dutch people earn approximately this amount of money each month (CBS 2017). To conclude, the sample of this research is viewed as a proper representation of the population in the Netherlands.

Since the purpose of this thesis is the measurement of the effectiveness of advertisements for generating sales, it is crucial to get a feeling for both the media behavior and the buying behavior of the respondents. When analyzing the data, it is conspicuous that many households are mostly not confronted with any of the media channels. This phenomenon is represented in the following paragraphs and figures.

3.2.1 YouTube (Online)

Per day, 99,84% of the 1173 households do not see the advertisement of the soft drink brand on YouTube. 0,15% see it once per day and no household sees it more than one time. According to Van Manen (2016), who conducted a research about the use of social media in the Netherlands, 48% of ‘Generation X’ (36 – 55 years old) and 37% of ‘Babyboomers’ (56+ years old) are actively using YouTube. Furthermore, Osterveer (2015) investigates that YouTube is with 77% the most popular online channel in the

(23)

23

Netherlands, followed by the channel ‘Uitzending gemist’ with 50%. These results seem controversial to the given media behavior. However, here it is crucial to consider that Van Manen’s research investigates how many people of different segments use YouTube in general. However, this thesis analyses whether respondents of this research are confronted with one precise brand advertisement on YouTube. Even though many people might be confronted with a particular marketing channel, the likelihood of seeing one special advertisement is small since they get targeted with many advertisements each time they use this channel.

3.2.2 RTL (Online)

When analyzing the respondent’s media behavior towards the other online advertisement on the RTL website, the overlapping tendencies with YouTube are obvious. 99,95% of respondents are not reached by this channel per day and only 0,05% see the video advertisement on this internet site once per day. The research by Osterveer (2015) also gives interesting insights regarding the use of the website of the Dutch television channel RTL. According to his study, this website is ranked as number 6 with 13%

in the Netherlands. Again, even though 13% represents a considerable number

of users, the likelihood of seeing one special advertisement on the RTL website is a lot smaller since the audience gets reached by many advertisements each time it uses this channel.

3.2.3 TV (Offline)

93,7% of the investigated households is not reached by the TV advertisement of this brand, however at least 6% see the advertisement twice per day. According to a study by the Stiching Kijkonderzoek (SKO) (2016), nowadays Dutch people watch approximately three hours of TV per day on average. However, here as well it needs to be considered that even though Dutch people might spend some hours per day watching TV, the likelihood is small that they get to see the advertisement of the soft drink brand. This is because firstly there are thousands of TV commercials every day,

Figure 4: Contact with RTL advertisement per household per day

(24)

24

and secondly many people watch movies, series or TV shows but are busy on other devices, such as their mobile, as soon as the advertisement break begins (Wilbur 2008).

3.2.4 Purchase of the soft drink brand

The households are divided into four different groups based on the monetary value they spend on the brand during the observation period. The majority of the respondents, namely 37,1%, are ‘light’ shoppers which mean that they spend less than 7,90€ for products of the soft drink brand. 30% of the households never buy the brand during the observational period. 19,4% of the respondents pay between €7,90 and €23,50 for products of the brand. Finally, 14,2% of the respondents are marked as ‘heavy users’ since they spend an amount higher than €23,50 on

the brand during the observational period. It would be highly interesting to investigate possible reasons for this ratio since the numbers that are found by the Dutch Association for Soft Drinks, Waters and Juices in 2013 are not really corresponding with these findings. According to their research the consumption of soft drinks was on average 99,4 liters in 2013 (Dutch Association for Soft Drinks, Waters and Juices 2013). All results from 2009 until 2013 for different types of drinks are represented in table 3. Considering the high number of 99,4 (table 3) it could be expected that a lot more of the investigated households would spend more money on soft drinks within the period of three months.

Table 3: Consumption per capita in liters

Consumption per capita in liters

2009 2010 2011 2012 2013 Soft drinks 102,7 101,8 102,6 102 99,4 Non-processed waters 21,6 21 21,5 21,2 21,7 Juices 17,3 16,9 17 17,3 16,4 Nectars 10,4 11 11,4 11,9 11,8 Syrups 15,1 15 14,9 15 14,9

(25)

25 3.2.5 Purchase during a price promotion

Of all households that made a purchase during the observation period, 22,5% did so while a price promotion took place. A reason for this ratio might be that most consumers, who feel the need to have a soft drink, buy their favorite brand regardless of whether the price is decreased or not (Ramanathan & Muyldermans 2010).

(26)

26 3.3 Model

Specification

The aim of this thesis is to measure the effect of three different marketing channels for generating sales and to assess the moderating role of price promotions. To achieve this goal, a model needs to be developed which can measure these effects. Here, the first step is to decide which dependent variable and which independent variables to use. The second step is to carefully consider the mathematical form of the model (Leeflang et al. 2015). Two of the most popular mathematical forms are the linear additive model and the multiplicative model (Leeflang et al. 2015). The linear additive model is linear in parameters and in variables. This model expects that the overall effect of the independent variables is equal to the sum of their individual effects (Leeflang et al. 2015). The main shortcoming of restricting the estimate to be additive is that potential interactions between the independent variables are missed (Leeflang et al. 2015). The multiplicative model is non-linear in the parameters. It assumes that each variable interacts with each other. To make a multiplicative model linear, logarithms of the dependent variable and of the independent variables need to be taken (Leeflang et al. 2015)

Since the focus of this research is to measure the separate effects of each independent variable on the dependent variable and no interactions between the independent variables are expected, a linear additive model is developed.

The variable Brand Sales is used as the dependent variable. It would not be sufficient to only measure by which household on which day the brand was purchased or not purchased. The brand can be bought for different prices and in different sizes, hence, it is interesting to investigate for which value consumers purchase the brand depending on the advertising channel they get reached by. The value of a purchase is calculated by multiplying the price per liter (in cents) with volume (in liters). Consequently, Brand Sales is represented by the value in cents.3_{OnlineYouTube, OnlineRTL, OfflineTV as well as the moderator variable Price} Promotion and its interaction effect with these three variables are used as independent variables. Furthermore, two control variables, namely Education and Income are included. Moreover, Brand Sales of the previous day is part of the model. This variable is added to the model since it is very likely that a purchase of the soft drink brand on the previous day influences whether it is again bought on the next day. It is assumed that many consumers associate soft drinks with a kind of refreshment that they purchase in moments they want to do themselves some good and therefore they might not buy this kind of product every day.

3_{The reason for indicating brand sales in cents, not in euros, is that this transformation enables a more meaningful interpretation}

(27)

27

𝑆𝑎𝑙𝑒𝑠&' =

𝛽*+ 𝛽- 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'+ 𝛽6𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'+ 𝛽: 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'+ 𝛽=𝑃𝑃'+ 𝛽?𝑃𝑃'𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'+

𝛽@𝑃𝑃'𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'+ 𝛽A𝑃𝑃'𝑂𝑓𝑓𝑙𝑖𝑛𝑒&'+ 𝛽B 𝐸𝑑𝑢 + 𝛽E 𝐼𝑛𝑐 + 𝛽-* 𝑆𝑎𝑙𝑒𝑠&'H- + 𝜀&'

Where: 𝑆𝑎𝑙𝑒𝑠_&' = Purchase of household i on day t 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒_&' = Contact with YouTube ad of household i on day t 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&' = Contact with RTL ad of household i on day t 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&' = Contact with TV ad of household i on day t 𝑃𝑃' = Price Promotion on day t 𝑃𝑃_'𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒_&' = Interaction effect between Price Promotion and Contact with YouTube ad of household i on day t 𝑃𝑃'𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&' = Interaction effect between Price Promotion and Contact with RTL ad of household i on day t 𝑃𝑃_'𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉_&' = Interaction effect between Price Promotion and Contact with TV Ad of household i on day t 𝐸𝑑𝑢& = Education of household i 𝐼𝑛𝑐_& = Income of household i 𝑆𝑎𝑙𝑒𝑠_&'H- = Lagged purchased of household i on day t-1 𝜀_&' = Error term of household i on day t

The next paragraph explains the procedure of the model estimation. It is elaborated on the decision to use a two-stage least squares (2SLS) regression instead of an ordinary least squares (OLS) regression to explain brand sales.

3.4 Possible Issue of Endogeneity

Before estimating the model and interpreting the results in the next parts of this thesis, this chapter focuses on the possible issue of endogeneity. In academic literature, many researchers do not draw intense attention to this characteristic, but they only mention at the end of their paper that possible endogeneity is one limitation of their study. This constraint arises since simple statistical tests such as the ordinary least squares regression (OLS), which is often used to investigate the effectiveness of marketing channels (Nannestad 2008), cannot account for the endogeneity problem.

(28)

28

independent variables are dependent on the dependent variable and hence the error term 𝜀. Possible causes for endogeneity are missing variables, common-method variance and measurement errors (Antonakis et al. 2014, Leeflang et al. 2015). In the event of endogeneity, the obtained coefficients will be biased (Leeflang et al. 2015).

The most common way to test for endogeneity is to conduct a 2SLS regression, also called instrument variable estimation. To use this approach, it is crucial to identify possible endogenous variables as well as instrument variables. Instrument variables must fulfill three conditions: First, more instrument variables than endogenous variables need to be identified. Second and third, instrument variables must be relevant and exogenous. Relevance refers to the degree to which the instruments correspond with the endogenous variable. The research of Stock et al. (2002) suggests that the relevance of an instrument variable can be measured by the F-statistics of first-stage regressions. According to these authors, an instrument variable is more relevant when it has a higher F-statistic (Stock et al. 2002). Exogeneity refers to the degree to which an instrument is uncorrelated with the error term in the second stage. According to Bascle (2008), the chance of replacing one endogenous variable with another can be reduced when testing for instrument exogeneity. Due to these prerequisites, it can be very difficult to identify suitable instruments. In practice, it is problematic to find variables that have a strong correlation with the endogenous variable but not with the error term in the second stage. Instrument relevance and exogeneity often work against one another. As instrument strength increases, i.e. the instrument becomes more like the endogenous variable, it can well be that it may be related to the error term in the same way as the endogenous variable (Semadeni et al. 2014). Furthermore, there are other prerequisites, not only related to the instrument variables that need to be confirmed before testing for endogeneity with the 2SLS regression. Firstly, it is important that both the dependent variable as well as the independent variables are quantitative. Secondly, the distribution of the dependent variable must be normal for each value of the independent variables. Furthermore, the variance of the distribution of the dependent variable should be constant for all values of the independent variables. Finally, the relationship between the dependent variable and each independent variable should be linear. Since the data used for this research satisfies these conditions, the coefficient for brand sales can be estimated using the instrument estimator two-stage least squares.

(29)

29

variables as well as instrument variables are chosen since this research investigates a low-involvement soft drink product and hence, it can be expected that a consumer is influenced by an advertisement on day t, however advertisements from the previous day (t-1) and from two days before (t-2) do not influence the purchase behavior of day t. This statement is based on the assumption that consumers do not plan their purchase of a FMCG product a few days in advance. Yet, seeing an advertisement on day t-1 or t-2 might have an effect on seeing that same advertisement on day t, because consumers are already familiar with that advertisement and brand and hence, they might be more likely to like or dislike it (Ansari & Joloudar 2011). In this thesis, the 2SLS regression is conducted and the issue of endogeneity is considered from two different perspectives:

First, a control-function approach is used to investigate the issue of endogeneity in the model. In this approach, each of the three possibly endogenous variables is regressed on the six lagged versions, i.e. the instrument variables. Then, the residuals are computed for each regression line. They are then additionally included in the model to investigate whether the residuals have a significant p-value, indicating that the corresponding advertising channel variables are endogenous (Wooldridge 2009). However, if they are not significant, endogeneity is not an issue. The following three formulas illustrate how each of the three possibly endogenous variables is regressed on the six instrument variables:

𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&' = 𝛼* + 𝛼- 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'H-+ 𝛼6 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'H- + 𝛼: 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'H-

+ 𝛼= 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'H6+ 𝛼? 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'H6 + 𝛼@ 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'H6

𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿_&' = 𝛼_*+ 𝛼_- 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿_&'H-+ 𝛼₆ 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒_&'H- + 𝛼_: 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉_&'H- + 𝛼= 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'H6+ 𝛼? 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'H6 + 𝛼@ 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'H6

𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&' = 𝛼* + 𝛼- 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'H-+ 𝛼6 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&'H- + 𝛼: 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'H-

+ 𝛼₌ 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉_&'H6+ 𝛼_? 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒_&'H6 + 𝛼_@ 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿_&'H6

When conducting these three regressions, their residuals (error terms) are calculated. This calculation works in a way that the predicted value is subtracted from the observed value. In theory, the residual calculation works with the following formula:

e = y – ŷ

(30)

30

𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒𝐸𝑟𝑟𝑜𝑟&' = 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&' − 𝑂𝑛𝑙𝚤𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒q'

𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿𝐸𝑟𝑟𝑜𝑟_&' = 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿_&' − 𝑂𝑛𝑙𝚤𝑛𝑒𝑅𝑇𝐿_q'

𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉𝐸𝑟𝑟𝑜𝑟&' = 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&' − 𝑂𝑓𝑓𝑙𝚤𝑛𝑒𝑇𝑉q'

After calculating the residuals, they are plugged into the main model and an OLS regression is conducted:

𝑆𝑎𝑙𝑒𝑠&' =

𝛽* + 𝛽- 𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&' + 𝛽6𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&' + 𝛽: 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'+ 𝛽= 𝑃𝑃'+ 𝛽?𝑃𝑃'𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒&' +

𝛽@ 𝑃𝑃'𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿&'+ 𝛽A 𝑃𝑃'𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉&'+ 𝛽B 𝐸𝑑𝑢&+ 𝛽E 𝐼𝑛𝑐& + 𝛽-* 𝑆𝑎𝑙𝑒𝑠&'H- +

𝑂𝑛𝑙𝑖𝑛𝑒𝑌𝑜𝑢𝑡𝑢𝑏𝑒𝐸𝑟𝑟𝑜𝑟&' + 𝑂𝑛𝑙𝑖𝑛𝑒𝑅𝑇𝐿𝐸𝑟𝑟𝑜𝑟&'+ 𝑂𝑓𝑓𝑙𝑖𝑛𝑒𝑇𝑉𝐸𝑟𝑟𝑜𝑟&' + 𝜀&'

Only if the p-values of the residuals are significant, the characteristic of endogeneity is indeed present in the model.

Second, to confirm the results of the control-function approach regarding the issue of endogeneity, the Wu-Hausman-test is used. For this, a normal OLS regression is executed and afterwards a 2SLS regression is conducted. Also here, two stages must be specified. The first stage involves using instrumental variables to determine the possibly endogenous independent variable. Then, the second stage uses the predicted value from the first stage as an independent variable in the second stage. The intuition behind this regression is that the first stage partials out any common variance between the endogenous independent variable and the instrumental variable. In this way, the predicted value does not share any variance that is related to the error term in the second stage. Finally, the Wu-Hausman-test is used to test whether endogeneity is an issue in the model. In this test, the null hypothesis is the following:

Ho: Variables are exogenous

(31)

31

(32)

32 4 Results

The results section of this thesis first provides the findings of the two approaches, namely control-function approach and Wu-Hausman-test, that are conducted to test whether endogeneity is an issue in the presented model. Afterwards, the goodness of fit of the specified model is investigated. Thereafter, the model is explored for inaccurate estimates in the variance of the parameters. Here, three disturbance term assumptions, namely autocorrelation, heteroscedasticity and nonnormality are examined. Furthermore, the issue of multicollinearity, which is a specification problem, is elaborated on. As a last step, the final model is estimated and the output is interpreted.

It is important to know that the significance of all investigated variables is judged on the 5% significance level. If a variable has a p-value of under 5% (0.05), it is significant, however if the p-value is higher than 5%, the variable does not have a significant effect.

4.1 Testing for Endogeneity

After testing for endogeneity with the 2SLS regression, more specifically the control-function approach, there is evidence that two of the possibly endogenous variables are exogenous, however, the variable OnlineYouTube is endogenous. This insight can be generated since computing the three residuals and including them in the model, do not evoke significant p-values for the variables OnlineRTL and OfflineTV (p-value Residual RTL = 0.1104; p-value Residual TV = 0.7198, see table 5). This indicates that these two advertising channel variables are not endogenous. The residual of the variable OnlineYouTube however has a significant p-value of 0.0062 (see table 5). Consequently, endogeneity is present in the model for this variable.

(33)

33

endogeneity is expected to be a problem, therefore the 2SLS regression needs to be performed.

Having a closer look at the results of the first possibly endogenous variable OnlineYouTube, it is conspicuous that the coefficient estimate of this variable between the OLS regression and the 2SLS regression changes a lot. In the OLS regression, the coefficient estimate of OnlineYouTube is 4.45 (see table 7 in appendix A) and in the 2SLS regression it is -320.36 (see table 8 in appendix A). This huge value change implies that there is probably an endogeneity bias in the OLS regression. However, whether there is a bias or not needs to be investigated further. First, it is important to investigate the strength of the chosen instrument variables, which is done by looking at the first stage regression. Here, the significant p-value (p = 0.000, see table 8 in appendix A) shows that the first-stage regression explains the variable OnlineYouTube very well. Second, to finally test whether the results of 2SLS are needed or whether the results of the OLS regression are sufficient, the Wu-Hausman-test is conducted. Since the p-value of this test is significant (p-value = 0.0315, see table 9 in appendix A), the null hypothesis can be rejected. This means that endogeneity is a problem for the variable OnlineYouTube. These different steps of the 2SLS regression are also conducted for the two other possibly endogenous variables OnlineRTL and OfflineTV. However, for these variables, the p-values are insignificant, namely p-value = 0.1637 for OnlineRTL and p = 0.2817 for OfflineTV (see table 10 and 11 in appendix A).

The next paragraph investigates the goodness of fit of the specified model. Thereafter, the model is explored for inaccurate estimates in the variance of the parameters. The three disturbance term assumptions autocorrelation, heteroscedasticity and nonnormality are examined.

4.2 Goodness of Fit

(34)

34

represents that the predictor variables do not explain the dependent variable very well. The lack of explained variance in the model could be attributed to the fact that there are many zeros in the dataset for both the media behavior and the purchase behavior (since most households did not see an advertisement during the observation period and did not buy the brand).

4.3 Autocorrelation

According to Leeflang et al. (2015), autocorrelation (also called serial correlation) is present if the residuals show some systematic pattern over time. This means that there is some relationship over time in the residuals which leads to the fact that the estimates for the parameters in the model are not as efficient as they could be. Consequently, one would get an inaccurate idea of the variance of the effects if autocorrelation is present (Leeflang et al. 2015). Therefore, it is highly important to detect autocorrelation. In most studies, this is done with the aid of the Durbin Watson test. Durbin-Watson is a measurement statistic of autocorrelation or serial correlation in the residuals of regression analysis (Michalakis et al. 2008). However, for the model used in this research, the Durbin-Watson test cannot be used since one of the independent variables is the lagged version of the dependent variable (Lagged Brand Sales). Therefore, the Breusch-Godfrey test is used to investigate whether autocorrelation is present. The p-value of this test does not show a significant value (0.1327, see table 13 in appendix B), which means that the null hypotheses of no autocorrelation cannot be rejected. Hence, no serial correlation is present in the model.

It is also checked to find out whether there was autocorrelation if the lagged variable was not included. It would be very typical if autocorrelation would be an issue in that situation, because the most typical approach to solve for autocorrelation is to add a lagged version of the dependent variable as an independent variable (and this is done in the model). Indeed, if the lagged version was not included, the p-value would be significant (0.0060, see table 12 in appendix B).

However, the model used in this study has an independent variable which is the lagged version of the dependent variable, and in this model, no serial correlation is present.

4.4 Heteroscedasticity

(35)

35

efficiency of the parameter estimates. However, the presence of heteroscedasticity factually means that the standard errors cannot be trusted in the same manner as they could if the residuals were homoscedastic. To detect for heteroscedasticity in the proposed model, the Breusch-Pagan test is executed. This test calculates the F-statistic with the R-squared from the regression (𝑅_r6s), the number of predictor variables (K) and the number of observations

(T). The following formula is used in the Breusch-Pagan test to calculate the F-statistic: 𝐹 = uv ss /x

(-Hu_v ss )/({HxH-) (Leeflang et al. 2015). This test produces a chi squared test statistic with 13

degrees of freedom (since there are 13 parameters in the model) when the null hypothesis of no heteroscedasticity is satisfied. That test statistic in this case is 3.61 (see table 14 in appendix C) and the p-value is 0.000 (see table 14 in appendix C). Consequently, the null hypothesis of homoscedasticity can be rejected. This means that the variance of the residuals increases as a function of at least one of the independent variables, but potentially more than one. Hence, there is the problem of heteroscedasticity in this dataset. This problem can also be reviewed in figure 8, 9 and 10. In these plots, the unstandardized residuals are plotted against the predicted residuals which show a pattern. Plots 9 and 10 show that the variance of the independent variables differ a lot, meaning that the data is heteroscedastic.

(36)

36 4.5 Nonnormality

To be able to judge the significance values (p-values), there needs to be some distribution. The distribution that is used to calculate the p-values is the normal distribution, hence, the normality of the residuals is needed in order to calculate the p-values. Consequently, the normality assumption is important for trusting the significance values (Leeflang et al. 2015). There are several opportunities for testing normality. In this case, figure 9 and figure 10 show that the normality of the residuals is not given in this data, indicating the large number of zeros in the dataset.

4.6 Multicollinearity

The last assumption that needs to be tested is the presence of multicollinearity. Multicollinearity is given when there is a very high correlation between independent variables (Malhotra 2010). This can cause unreliable estimations of the partial regression coefficients. Furthermore, when independent variables are highly correlated, a meaningful interpretation of the regression model becomes very difficult since their signs and their magnitude cause wide confidence intervals (Malhotra 2010; Leeflang et al. 2015). The most common way to detect multicollinearity is to first examine the Pearson Correlation Matrix to get a first insight in whether there are some values that are highly correlated with each other (Malhotra 2010). Each value in this matrix (see table 4) should be between -0.5 and 0.5. If there are some critical values that are lower or higher than this, the Variance Inflation Factor (VIF) scores should be inspected (Leeflang et al. 2015). They can be used for the final decision in determining whether multicollinearity is present in the model (Malhotra 2010). However, there is some disagreement between different researchers regarding the critical value of the VIF

(37)

37

score that determines whether multicollinearity is present or not. According to Leeflang et al. (2015), a VIF value below 5 indicates that no multicollinearity is present. Malhotra (2010) suggests a VIF value of 4 as a cut-off point.

The Pearson Correlation Matrix for the model used in this research shows that the independent variables do not highly correlate with each other since their values are between -0.5 and 0.5. Consequently, it is not necessary to check whether the VIF scores are below 4. It can be concluded that no multicollinearity issues can be detected in the model used for this research.

Pearson Correlation Matrix Brand

Sales Contact with YouTube Ad Contact with RTL Ad Contact with TV Ad PP PP * YouTube Ad PP

* RTL Ad PP * TV Ad Edu. Inc. Lag Brand Sales Brand Sales - Contact with YouTube Ad .008 - Contact with RTL Ad .001 .010 - Contact with TV Ad .002 -.009 -.002 - PP .973 .007 .066 -.001 - PP * YouTube Ad .039 .113 .059 -.001 .038 - PP * RTL Ad .011 .077 .000 -.002 .013 .469 - PP * TV Ad .250 .000 -.001 -.035 .255 -.006 -.004 - Edu. .004 .002 -.003 -.035 .003 -.002 .001 -.002 - Inc. .018 .002 .001 .001 .016 .000 .000 .006 .265 - Lag Brand Sales .011 .006 .335 .246 .009 .005 -.001 .002 .004 .018 -

4.7 Model Correction

(38)

38

(39)

39 4.8 Estimation Results

Table 5 shows the 2SLS regression results of the final model that is estimated with the control function approach. As mentioned before, the adjusted R2_{of the model is moderately low} (0.521) which indicates that the model explains 52,1% of variance in the purchase value of the observed households.

Dependent Variable: Brand Sales Method: Least Squares

Sample: 1 105840

Included observations: 105840

HAC standard errors & covariance (Bartlett kernel, Newey-West)

Table 5: Regression Output

Variable Estimate Std. Error t-statistic Prob.

Contact with YouTube Ad -822.6108 302.6899 -2.717669 0.0066

Contact with RTL Ad 1984.902 1154.617 1.719100 0.0856

Contact with TV Ad 1.896975 3.951006 0.480125 0.6311

Price Promotion 223.5858 6.606004 33.84585 0.0000

Price Promotion * YouTube Ad 1.746257 0.476123 3.667661 0.0002

Price Promotion * RTL Ad 5.938983 0.261019 22.75310 0.0000

Price Promotion * TV Ad 4.238652 5.809715 0.729580 0.4656

Education 0.039987 0.097069 0.411940 0.6804

Income 0.116555 0.031681 3.678972 0.0002

Lagged Brand Sales 0.005123 0.002898 1.767698 0.0771

Residual YouTube 827.2467 302.5256 2.734469 0.0062

Residual RTL -1955.072 1154.609 -1.693275 0.1104

Residual TV -1.442157 4.020396 -0.358710 0.7198

Constant -0.599021 0.595708 -1.005561 0.3146

R-squared 0.521031 Mean dependent var 5.607039

Adjusted R-squared 0.520972 S.D. dependent var 65.90473

S.E. of regression 45.61386 Akaike info criterion 10.47843

Sum squared resid 2.20E+08 Schwarz criterion 10.47970

Log likelihood -554504.7 Hannan-Quinn criter. 10.47882

F-statistic 8855.349 Durbin-Watson stat 1.993330

Prob (F-statistic) 0.00000 Wald F-statistic 6.56E+08

Prob (Wald F-statistic) 0.00000

4.8.1 Online advertisement effectiveness

(40)

40

increases, instead of decreases sales (𝛽 = 1984.902). The parameter estimate indicates that a one-unit increase in seeing a RTL advertisement increases sales by 1984,902 cents (19,84€). Consequently, hypothesis 1 is only partly supported since both online advertisements have a statistically significant impact on brand sales. However, OnlineYouTube correlates negatively with brand sales, but OnlineRTL correlates positively with brand sales.

4.8.2 Offline advertisement effectiveness

Offline advertising does not have a significant impact on brand sales. This is reflected by the p-value of the variable OfflineTV, which is insignificant (p-value = 0.6311). However, the coefficient estimate of this variable is positive (𝛽 = 1.896975). Therefore, it can be concluded that seeing a TV advertisement of the soft drink brand has a positive effect on brand sales of the brand, but this effect is not statistically significant. Therefore, hypothesis 2 is rejected.

Optimizing marketing strategies MASTER THESIS