Online video advertising (skippable and non-skippable) versus television advertising

Online video advertising (skippable and non-skippable)

versus television advertising

A comparison of its effectiveness in influencing consumers’ purchase

behaviour

By

Online video advertising (skippable and non-skippable)

versus television advertising

A comparison of its effectiveness in influencing consumers’ purchase

behaviour

By

Tom Blanke

Faculty of Business and Economics Department of Marketing

MSc Marketing Intelligence and MSc Marketing Management

Author: 1st Supervisor: 2nd Supervisor:

Tom Blanke P.S. van Eck Dr. Ir. M.J. Gijsenberg

t.blanke@student.rug.nl p.s.van.eck@rug.nl m.j.gijsenberg@rug.nl

Parelstraat 56 Nettelbosje 2 Nettelbosje 2

DUI326 DUI332

PREFACE

Before you lies my master thesis, as a result of my research on the effectiveness of online video advertising and television advertising from February to June 2017. It has been written to complete the Marketing Intelligence and Marketing Management Master programs at the University of Groningen, and to extend my knowledge on advertising mediums which benefits me in my future career as a marketing consultant.

I am satisfied with the end result of this thesis and the great learning experience that conducting this research was for me. I would like to thank my supervisor Peter van Eck for his support and guidance throughout the whole process of writing this master thesis and for his useful feedback to further improve and revise my work. Furthermore, I am grateful that GfK provided me the comprehensive and rare dataset for this research, which I could not have collected on my own. I hope you will enjoy reading my thesis.

Tom Blanke

ABSTRACT

TABLE OF CONTENT

1. Introduction 8

2. Literature review 12

2.1 Television advertising 12

2.2 Online video advertising 13

2.3 Skippable and non-skippable online video advertising 14

2.4 Consumers’ purchase behaviour 15

2.5 The direct effect of advertising on consumers’ purchase behaviour 16

2.5.1 Television advertising 16

2.5.2 Online video advertising 17

2.5.3 Television advertising vs online video advertising 18 2.5.4 Skippable vs non-skippable online video advertising 19 2.6 The retention effect of advertising on consumers’ purchase behaviour 19 2.7 The synergy effect of advertising on consumers’ purchase behaviour 21

2.8 Conceptual model 22 3. Methodology 23 3.1 Data collection 23 3.2 Data sample 24 3.2.1 Data preparation 24 3.2.2 Data description 25 3.3 Variables 26 3.4 Analysis plan 31 3.4.1 Specification 32 3.4.2 Estimation 33 3.4.3 Validation 35 4. Results 38 4.1 Specification 38

4.1.1 Specification Logistic Regression 38

4.1.2 Specification count model 40

4.2 Estimation, validation, and use 42

4.2.1 Preliminary checks Logistic Regression 42 4.2.2 Estimation, validation, and use Logistic Regression 43 4.2.3 Hypothesis testing Logistic Regression 55

4.2.4 Preliminary checks count 56

5. Discussion 71

5.1 Conclusion 71

5.2 Scientific contribution 74

5.3 Managerial contribution 76

5.4 Limitations and opportunities for future research 77

References 79

Appendix I: Multicollinearity table 85

Appendix II: Retention rates, per week, per medium: Logistic Regression 86 Appendix III: Comparison best retention variables with direct variables: Logistic

Regression 89

Appendix IV: Online retention variables: Logistic Regression 90 Appendix V: Retention rates, per week, per medium: Negative Binomial Regression 91 Appendix VI: Comparison best retention variables with direct variables:

Negative Binomial Regression 94

1. INTRODUCTION

Advertising has changed significantly over the past years which might have an impact on the effectiveness of advertising. In 2008, 801 million Euros was spent on internet advertising and 855 million Euros was spent on television advertising in the Netherlands (Statista, 2013). In 2013, the advertising expenditures increased to 1.255 billion Euros spent on internet advertising and 933 million spent on television advertising in the Netherlands (Statista, 2013). Worldwide, advertisers still prefer to spend their advertising budget on television advertisements rather than on internet advertisements (Draganska et al., 2014). In 2016, television advertising received the largest share of the advertising expenditure, namely 41.2 percent (Statista, 2016). However, the global growth of advertising spending in internet advertising in 2016 was 14.6 percent, whereas the growth of advertising spending in television advertising was 2.8 percent (Statista, 2017). The research by Kumar (2015) shows that there is not only an increase in money spent on internet advertising, but consumers are also switching from traditional media to interactive media such as internet, mobile, and interactive television. These significant changes in advertising expenditures and advertisement consumption make it interesting to take another look at the work of Sethuraman, Tellis and Briesch (2011) on the effectiveness of advertising. They show that television advertising has a significantly positive effect on sales, and that the effectiveness of television advertising has decreased over this period.

The objective of this thesis is to renew parts of the meta-analysis of Sethuraman, Tellis, and Briesch (2011) regarding the effectiveness of advertising, and extend it by not only analysing the effectiveness of traditional advertising, but also the effectiveness of online advertising. The main research question of this research is: “How effective is online video advertising (skippable and

non-skippable) relative to television advertising, regarding their influence on consumers’ purchase behaviour?”. The effectiveness of online video advertising (skippable and non-skippable) and

television advertising will be evaluated based on three types of effects: their direct effect, their retention effect, and their synergy effect.

Therefore, consumers are less tolerant of internet advertising than television advertising (Ducoffe 1996 from Belanche et al., 2017). This suggests that television advertising would be more effective than online video advertising. In addition, to compare the direct effect of television advertising and online video advertising, this study makes the comparison between the direct effects of skippable and non-skippable advertisements. The researchers Belanche et al. (2017: 75) describe skippable advertisements as follows: “with skippable advertisements consumers are not required to sit through the entire commercial information before they may access the content they seek; instead, they decide whether to watch the advertising videos voluntarily”. Earlier research shows that the level of interest and tolerance of consumers in a commercial depends on the amount of control they perceive to have (Huang, Lurie, and Mitra 2009; Logan 2013; Raney et al., 2003). This indicates that skippable advertisements are more effective than non-skippable advertisements (Yoo and Stout 2001 from Belanche et al., 2017).

Besides the direct effect of advertising, this study will also investigate the retention effect of advertising. The retention effects of the advertising will be investigated based on: 1) how large the influence of the advertisement is on consumer’s purchase behaviour over time, and 2) how long the advertisement influences the purchase decision of the consumers. This is important because when advertising managers assume the effect of advertising will last one year, while the duration in reality is only a few weeks, poor decisions will be made (Clarke, 1976).

The three types of effects will be evaluated in relation to consumers’ purchase behaviour. Consumers’ purchase behaviour will be assessed with respect to the following two elements: purchase decision (a purchase probability) and purchase quantity decision (the number of products bought). These two elements are selected based on prior research that shows that consumers’ brand purchase behaviour results from the consumers considering two questions: which brand should I buy? and how much should I buy? (Bucklin and Lattin, 1991; Chintagunta, 1993; Gupta, 1988; Krishna, 1994; Krishnamurthi and Raj, 1988; Neslin, Henderson, and Quelch, 1985).

In addition, to extending the work of Sethuraman, Tellis, and Briesch (2011), which is focused on the effectiveness of advertising at an aggregate level, on overall sales, with online video advertising. This research also revises their analysis by analysing the effectiveness of advertising mediums on individual consumer’s purchase behaviour. Therefore, this research will make an academic contribution by investigating whether the influences of the different advertising channels at an aggregate level also hold on an individual household level. Furthermore, this research provides clarification on whether advertising is effective for a product that is in its mature stage of the product life cycle. The findings of Sethuraman, Tellis, and Briesch (2011) suggest that companies should focus on advertising during the early stages of the product lifecycle and on price promotions during the subsequent stages. The findings of Mela, Gupta, and Lehmann (1997), however, show that price promotions have negative effects in the long run, while advertising has positive effects on consumers’ brand choice behaviour.

This study reacts to the contradiction in prior research about the effectiveness of advertising for a product that is in the mature stage of the product life cycle. Therefore, a soft drink brand that is in its mature stage is the focus of this study. To investigate the effectiveness of advertising for this brand, the scope of this research consists of Dutch households who purchased soft drinks, more specifically, in regards to their brand purchases and not to their category purchases. This restriction is made to investigate the direct effect of a firm’s own brand advertisements on the demand for their own brand.

2. LITERATURE REVIEW

This research explores the effectiveness of different advertising mediums regarding their ability to influence consumers’ purchase behaviour. This chapter starts with describing the different advertising mediums and the dependent variable, consumers’ purchase behaviour, which will be evaluated in this research. Subsequently, the three types of effects, direct effect, retention effect, and synergy effect, on which the advertising channels will be evaluated regarding their ability to influence consumers’ behaviour will be discussed. Finally, based on prior research regarding these types of effects, hypotheses will be formulated and graphically presented in a conceptual model. 2.1 Television advertising

Television advertising is “the sending of promotional messages or media content to one or more potential program viewers” (IPTV Dictionary, 2005). In the Netherlands in 2014, 7.4 of the 7.6 million households owned a television, and in 2015 the average daily time spent watching television in the Netherlands was 190 minutes (Statista, 2015; CBS, 2017). These figures indicate an important advantage of television advertising, as it enables firms to reach a high number of consumers. Furthermore, Keller (2013: 222) stated that television is a powerful advertising medium because it is effective in vividly (in sight, sound, and motion) showing a product’s attributes and it informs consumers about the benefits in a compelling manner. Another advantage of television advertising is that television advertising can be used in a persuasive manner by illustrating user and usage imagery, brand personality, emotions, and other intangible aspects. Keller (2013: 223) also discusses a few disadvantages of television advertising. Consumers may overlook the actual product related message and the brand itself due to fact that the advertisements are very short and often possess distracting creative elements. The advertising clutter on television may also cause consumers to ignore or forget the television advertisements. Another disadvantage is that consumers may not get exposed to the advertisements because they have the opportunity to skip commercials. In addition, the production and placement cost of television advertisements are high.

The explicit memory, “the part of the memory that consumers can consciously bring to mind” (Fennis and Stroebe, 2016: 49), and the implicit memory, “the part of the memory that is mentally available but not consciously accessed” (Fennis and Stroebe, 2016: 49). The influence of advertising on the explicit memory can largely be addressed by measurements of recall. However, the influence of advertisements on implicit memory cannot be measured by addressing what consumers recall with a survey. Therefore, a great advantage of this study is that it is based on data of the actual exposure to television advertising. This is also an adjustment to the research of Sethuraman, Tellis, and Briesch (2013) who measured television advertisement in terms of a firm’s advertising expenditures.

2.2 Online video advertising

Online video advertisements are also ignored or forgotten by consumers because of the advertising clutter. Finally, firms may lose the control over the content, because everyone has access to it.

Online video advertisements can appear in four ways: as an in-stream video, on the right hand side of the video watch page, in search results, and on the home page (Pashkevich et al. 2012). The advertisements in this research were shown as an in-stream video advertisement, which is the prevalent online video format. The researchers (Pashkevich et al. 2012: 65) defined in-stream video advertisement as “a short video – much like a television commercial – that is played prior to the user-selected video content”. An advantage of the in-stream video advertisement compared to other online form of advertising (e.g. buttons and banners) is that they are less intrusive (Winer, 2009), which is something that is valued by consumers.

2.3 Skippable and non-skippable online video advertising

They make use of skippable online video advertisements, which are advertisements “consumers are not required to sit through the entire commercial information before they may access the content they seek; instead, they decide whether to watch the advertising videos voluntarily” (Belanche et al., 2017: 75). The form of skippable advertisements that is the focus of this research is the by YouTube in 2010 introduced TrueView-in-stream advertisements. These online video advertisements give viewers the possibility to watch the advertisement entirely or to go to the desired content after 30 seconds of viewing, thereby enhancing the viewers’ active role (Pashkevich et al., 2012). With the TrueView-in-stream advertisements YouTube reacted to the viewers’ experience of the placement of video advertisements before the viewed content as intrusive (Pashkevich et al., 2012). The main advantage of skippable online videos is that they increase the amount of control that consumers perceive to have, which increases consumers’ level of interest and acceptance of the advertisement (Huang, Lurie, and Mitra 2009; Logan 2013; Raney et al. 2003; Fortin and Dholakia, 2005). In addition to the increase in control for advertisement viewers, it also increases their motivation and loyalty, and thereby increasing firms’ chances of achieving their advertising objectives (Raney et al. 2003). Another advantage of skippable advertisements for advertisers is that they only have to pay for the advertisements if consumers watch the advertisements for at least 30 seconds (Pashkevich et al., 2012. Therefore, the skippable advertisements improve the viewers’ advertisements experience without reducing the advertising value for advertisers (Pashkevich et al., 2012). This research analyses the effectiveness of the skippable advertisements that are viewed completely or for at least for 30 seconds. The advantages of skippable advertisements are noticed by advertisers. In 2012, 70 percent of the in-stream advertisements on YouTube were skippable (Pashkevich et al., 2012), and in 2014 all top 100 global brands used skippable advertisements (Connolly, 2015 from Belance, Flavián, Pérez-Rueda, 2017).

2.4 Consumers’ purchase behaviour

Therefore consumers’ purchase behaviour in this research is assessed regarding the following two elements: 1: purchase decision (a purchase probability), and 2: purchase quantity decision (the number of products bought). In addition, the researchers Lewis and Reiley (2014) show that advertising increases both the probability of the purchase and the average purchase quantity. Therefore, the results regarding the two elements of consumer’s purchase behaviour are expected to behave into the same direction (e.g. both positive or both negative).

To the best of the researcher’s knowledge, this is the first research that makes the comparison between television advertising and online video advertising, regarding their effectiveness of influencing individual consumer’s purchase behaviour. Earlier studies that analysed the effectiveness of these mediums focused on its effects on overall sales.

2.5 The direct effect of advertising on consumers’ purchase behaviour 2.5.1 Television advertising

Despite the changes in effectiveness over time, the effect of television advertising on sales was positive (Hu, Lodish, and Krieger, 2007; Sethuraman, Tellis, and Briesch, 2011). In addition, other researchers have shown that television advertising has a positive effect on sales (Danaher and Dagger, 2013; Keller, 2013). Therefore, this research will investigate whether television advertising also has a positive effect on consumer’s purchase behaviour. Thus, the following hypothesis will be tested in this research:

H1: Television advertising has a positive direct effect on (a) consumers’ purchase decision and (b) consumers’ purchase quantity decision.

2.5.2 Online video advertising

2.5.3 Television advertising vs online video advertising

This study answers the call from the researchers Draganska, Hartmann, and Stanglein (2014) to compare television and internet advertisements regarding their impact on sales, by assessing its impact on consumers’ purchase behaviour. The researchers Draganska, Hartmann, and Stanglein (2014) show that there is no significant difference between the effectiveness of television advertising and internet advertisements, regarding brand recall. Recall can be seen as an indicator of purchase behaviour, therefore these results suggest that there should be no difference in the effectiveness of television advertising and online video advertising (Keller, 2013: 74). However, the research conducted by Logan (2013) shows that consumers regard television advertising as less intrusive than online video advertising. Therefore, consumers are less tolerant of internet advertising than television advertising (Ducoffe 1996 from Belanche et al., 2017). Consumers do not accept the role of advertising as a means to subsidize access to internet content because they believe online content should be free, whereas they accept it for television advertising. Furthermore, fewer respondents of the survey conducted by Schlosser, Shavitt, and Kanfer (1999) had a positive perception of internet advertisements than they had of the general advertisements, amongst which television advertisements. Earlier research shows that consumers’ perceptions and attitudes towards advertisements are a good indicator of their influence on purchase decisions (Katz and Stotland, 1959; Rosenberg and Hovland, 1960: both from Schlosser, Shavitt, and Kanfer, 1999). Therefore, the fact that the respondents’ attitudes towards television advertisements were more favourable than toward internet advertisements indicates that television advertising is expected to have a greater impact on consumers’ purchase decisions (Schlosser, Shavitt, and Kanfer, 1999). Based on the great support from literature, the following hypothesis will be investigated in this research:

2.5.4 Skippable vs non-skippable online video advertising

As mentioned earlier, providing viewers of online videos with the choice to view the advertisement or not, increases viewers interest for and tolerance of the advertisement (Pashkevich et al., 2012; Huang, Lurie, and Mitra 2009; Logan 2013; Raney et al., 2003). Therefore, enabling consumers to skip an advertisement improves their purchase intentions (Yoo and Stout 2001 from Belanche et al., 2017). Consequently, this paper will investigate the following hypotheses:

H4: Skippable advertising has a greater direct impact on (a) consumers’ purchase decision and (b) consumers’ purchase quantity decision than non-skippable advertising.

2.6 The retention effect of advertising on consumers’ purchase behaviour

Siegel et al. (2016) state that advertising has a cumulative effect, where current advertisements have a greater impact than prior advertisements. Thus, to assess the individual effect of advertising channels, not only the direct effect should be considered but also its retention effect. This research defines retention effect as the period after initial exposure that advertisements keep having an effect on consumers’ purchase behaviour due to consumers’ memory of the advertisement (GfK). The researchers Aravindakshan and Naik (2011) show that consumers possess memory of advertisements, that the created brand awareness by advertisements does not disappear directly after consumers have seen the advertisement. In addition, they state that the retention rate differs per brand, based on different brand awareness. Therefore, this research investigates not only the period that an advertisement keeps having an effect on a consumer’s purchase behaviour, but also whether the retention rates differ per advertising channels. Retention rate is defined in this research as the percentage of the advertisements that keeps having an effect on consumers’ purchase behaviour over time.

Therefore, this research is not focused on the long-term effects (effect in years) of advertising. In addition, the researchers Lewis and Reiley (2014) show that the effects of online advertising on offline sales stopped three weeks after the advertising campaign.

To the best of the researcher’s knowledge, this is the first study that compares television advertising with online video advertising (skippable and non-skippable advertising) regarding their retention effects. Because there is no prior research, the following hypotheses are derived from the discussed literature and the hypotheses for the direct effects of the different advertising channels. H5: Television advertising has a longer impact on (a) consumers’ purchase decision and (b) consumers’ purchase quantity decision than online video advertising (both non- and skippable ads).

H6: Television advertising has a greater impact on (a) consumers’ purchase decision and (b) consumers’ purchase quantity decision over time than online video advertising (both non- and skippable ads).

H7: Skippable advertising has a longer impact on (a) consumers’ purchase decision and (b) consumers’ purchase quantity decision than non-skippable advertising.

2.7 The synergy effect of advertising on consumers’ purchase behaviour

Synergy is defined as “the combined effect of multiple activities exceeds the sum of their individual effects” (Belch and Belch, 1998: 11 from Naik and Raman 2003), and the researchers Naik and Raman (2003: 385) stated that the synergy effect “is the added value of one medium as a result of the presence of another medium, causing the combined effect of the mediums to exceed the sum of their individual effects”. An important concept regarding synergy is the integrated marketing communication (IMC), which emphasizes the advantages of creating synergy across multiple advertising channels, the advantage that each medium enlarges the contributions of all other channels. The results of the work of Naik and Raman (2003) confirms the IMC concept, by showing that the impact of the multimedia activities, television, print, radio, internet, direct response, sales promotion, and public relations taken together, can be a lot larger than the sum of their individual

effects. Their findings show that when synergy increases, the advertiser needs to invest a smaller

proportion of their media budget in the more effective activity, and a larger proportion of their media budget in the less effective activity. Because in case of multiple advertising channels with synergy, the optimal investment in one advertising channel depends not only on the effectiveness of that advertising channel, but also on the budget spent on the other advertising channels and their effectiveness. Therefore, overall the effectiveness of the different advertising channels, increases faster when a larger proportion of the advertising budget is spend on the least effective channel. This makes it interesting for marketers to get more insights in both the direct and synergy effects. Although the concept of IMC has already been widely accepted by advertisers, the amount of research on it is scarce (Naik and Roman, 2003), which makes the synergy effect an interesting part of this study. Next to the work of Naik and Roman (2003), other earlier research also confirms the existence of the synergy effect between multiple advertising channels (Naik and Peters, 2009; Danaher and Dagger, 2013; Dinner, van Heerde, and Neslin, 2014; Verhoef, Neslin, and Vroomen, 2007; Dijkstra et al., 2005; Rutz and Bucklin, 2011). In this research the synergy effect between multiple advertising channels will be investigated regarding their effect on consumer’s purchase behaviour. Therefore the following hypothesis will be tested in this research:

2.8 Conceptual Model

Figure 1 provides a graphical representation of the expected relationships between the concepts of this research, based on the literature review.

3. METHODOLOGY

This chapter consists of the following elements: data collection, data sample, variables, and analysis plan. In the data collection, the research design through which the data is collected is discussed. The second part of this chapter is about the data preparation steps that were conducted, which resulted in the data sample of this research, and provides a description of the resulted data sample. The third part discusses how the independent variables, the dependent variables, and the effects of the independent variables on the dependent variables are measured. In addition, the covariates for which its effects are controlled in this research are reviewed. Lastly, the analysis plan consists of the analysis steps through which the results are collected and the hypotheses are assessed.

3.1 Data collection

To overcome this drawback, GfK, the provider of the panel data that is used in this research, ensured that the panel was a good representation of the population, with the use of a stratified sampling method. With stratified sampling, the entire population is divided into smaller groups based on common attributes or characteristics. Afterwards, from each group a representative random proportion is taken, a comparable proportion that the groups represent in the population. These samples are then pooled into a random sample of the entire population (Investopedia, n.d.). This way, the panel is a good representation of the Dutch consumers of the soft drink category, which is the population that this research wants to describe. Another drawback that Malhotra (2010: 112) discusses is response bias. A few causes for response bias are boredom of the panel members and an incomplete diary. GfK managed to keep the response bias as minimal as possible, by closely managing the data provided by the panel members. In case of any strange results (e.g. fewer data) the panel members were contacted to check whether they filled in everything in the right way and whether they were still participating. To assess the effectiveness of the different advertising channels, this study needed information about what consumers purchase and about their exposure to advertisements. Therefore, the panel data for this research is provided through both a purchase panel and a media panel. A purchase panel is “a data-gathering technique in which respondents record their purchase online in a diary” (Malhotra, 2010: 150), while a media panel is a data gathering technique that automatically records respondents’ exposure to advertisements, to supplement the purchase information (Malhotra, 2010: 151).

3.2 Data sample 3.2.1 Data preparation

Subsequently, the households who did not participated in both the online and television panel were excluded from the data sample, to ensure that the same people were evaluated for the direct, retention, and synergy effect of the advertising channels on the households’ purchase behaviour. After this step still 780 households and 70200 observations remain in the dataset.

The researchers Sethuraman, Tellis, and Briesch (2011) state that the effects of advertising channels differ significantly per data interval. Therefore, it is important to select the appropriate data interval to estimate advertising effectiveness. The conventional view for determining the correct data interval is the inter-purchase time for the product category (Sethuraman, Tellis, and Briesch, 2011). The inter-purchase time is the time between each transaction of a customer and the previous transaction by the same customer (Kuo and Chen, 2004). As can be seen in table 1, the households bought the brand more often than once a week in only 1.1% of the cases. Therefore, the prevalent inter-purchase time in this dataset equals a week, and the data was aggregated from day level to week level. This resulted in the final dataset of 780 households and 10140 observations. Before starting to analyse this dataset, the dataset was examined on duplicates and outliers. In this examination one observation was excluded from the dataset, because one observation from the same household on the same day existed twice in the dataset.

Table 1

Frequency table of the brand purchases per week Weekly purchases Frequency Percentage

0.00 8913 87.9 1.00 1114 11.0 2.00 94 0.9 3.00 16 0.2 4.00 3 0.0 Total 10140 100.0 3.2.2 Data description

These observations were collected from 780 households over a period of thirteen weeks in 2014 (January, February, and March in 2014). These households purchased the soft drink brand that is the focus of this research 1227 times. In those purchases the amount of products bought by the households ranges from zero to fifty-four with an average of .57. These products were bought by the households at a mean price per unit of €1.31. The households bought 2033 units of competitor brands. Most households in the panel were based on their prior consumption classified as light (38.3%) and medium consumers (27.7%) of the soft drink brand. During the three months all the households together were exposed to 2975 advertisements through television, 129 skippable advertisements through YouTube, and 36 non-skippable advertisements through the RTL website. The descriptive statistics of the socio-demographical variables in this research showed that the most common age amongst the households lays between 55-65, namely 18.2%, and even 75.5% of the households have an average age of 40 or higher. The average household size is 1.48 persons, and most frequently the households existed out of two persons (35.1%). Most households completed another education after high school as 34% finished MBO and 27.2% finished HBO/WO. The income of the households is spread across the different levels, with the most frequent incomes between €1700 - €1900 (10.8%) and €2100 -€2300 (10.3%). They were also spread throughout the Netherlands, where the largest number of households lived in the districts rest west (28.2%) and south (21.1%).

3.3 Variables

Television advertising. The first advertising channel of which the effects on consumers’ purchase

Online video advertising. The second advertising channel of which the effects on

consumers’ purchase behaviour will be measured is online video advertising. The exposure of households to online video advertisements is measured through a browser plugin, which is installed on all the devices through which the households can be exposed to online video advertisements. This way, all the URL calls and search queries of the panel members are registered. Based on this, GfK determines to which online video advertisements the households are exposed. Consequentially, the knowledge of the internet usage of the households and their exposure to online video advertisements provides insight in whether the households were exposed to skippable or

non-skippable online video advertisements. The exposure of households to non-skippable online video

advertisements is measured as the contact moments of households with advertisements on YouTube, which they watched for at least 30 seconds. The exposure of households to non-skippable online video advertisements is measured as the contact moment of households to online video advertisements on the RTL website that they have to watch entirely before they can watch their desired content. Online video advertising is included as a nominal variable in the dataset, representing whether households where exposed to skippable and/or non-skippable advertisements. Additionally, the exposures to skippable and non-skippable advertisements are also added to the dataset as nominal variables, in which 1 represented exposure to the advertisement and 0 no exposure to the advertisement.

Consumer’s purchase behaviour. In this research, the dependent variable is consumers’

Direct effect, synergy effect, and retention effect

The direct effect, synergy effect, and retention effect of the different advertising channels will be assessed through the parameters that are obtained in the Logistic Regression and the count model. Regarding the parameters for the different advertising channels in the Logistic Regression, their effectiveness will be evaluated in terms of their effect on the probability that a consumer will buy the brand or not. As for the parameters in the count model, the parameters of the different advertising channels, and, therefore, their effectiveness, will be evaluated based on their multiplicative effect on the amount of products that the consumer buys.

Retention effect

The retention effect of the different advertising channels will be analysed with the use of adstock variables with a short half-life. The researchers Broadbent and Fry (1995) state that adstock modelling with a short half-life is a well-established way to assess the retention effect of advertisements over a few week, in which a short half-life is ten weeks or less (Broadbent, 2000). Because this research focuses on the short-term and medium-term retention effects, the adstock variables with a short half-life fit perfectly. Adstock with a long half-life, six months to two years, however, would have a better fit in case the focus was on the long-term effect (Broadbent and Fry, 1995). Adstock is “a measure of cumulative advertising exposure that sums current advertising exposure levels with discounted levels of prior exposure” (Siegel et al., 2016: 6.). The following equation to calculate the adstock variables is derived from earlier research that incorporated adstock variables in their brand choice models on comparable data (Terui and Ban, 2008; Ban, Terui, and Abe, 2011; Aravindakshan and Naik, 2011):

𝐴𝑡 = 𝑋𝑡 + 𝑟𝑒𝑡𝑒𝑛𝑡𝑖𝑜𝑛 𝑟𝑎𝑡𝑒 ∗ 𝐴𝑡 − 1

Table 2

Example of the made adstock calculations Household Wk. TV_contact TV_25_w3

1723 1 1 1.00

1723 2 0 .25

1723 3 0 .0625

1723 4 0 0.00

Table 2 shows an example of the calculation of an adstock variable for television advertising with a retention rate of 25% for three weeks. As can be seen, the household is exposed to one television advertisement in week one, and this exposure is assumed to retain 25% of its influence in week two. Therefore, despite the fact that the household is not exposed to a television advertisement in week two, the adstock variable: TV_25_w3 is .25 in week two for household 1723. The exposure to the advertisement in week one is assumed to still have an effect in week three, but it holds only 25% of its effect that it had in week two, which results in .0625. Because the television advertisement in this example was assumed to have an effect for three weeks, the effect of the exposure to the television advertisement in week 1 stops after week 3. Therefore, the adstock variable: TV_25_w3 is .00 in week four for household 1723.

Synergy effect

For the analyses of the synergy effect between the advertising mediums, synergy variables will be created. These variables will be created by multiplying the adstock variables or the original variables of the advertising mediums with each other. This is a common way of creating the synergy variables and it has been used by many earlier researchers (Naik and Peters, 2009; Naik and Raman, 2003; Danaher and Dagger, 2013). If the analyses on the retention effects have as its results that the advertising mediums have a retention effect, then the adstock variables will be used. Otherwise the original variables will be used.

Covariates

Price. Earlier research has shown that the price that consumers have to pay for a product

influences their purchase probability, and, thereby, their purchase behaviour (Lichtenstein, Ridgway, and Netemeyer, 1993; Sethuraman, Tellis, and Briesch, 2011). Price has a negative influence on consumers’ purchase behaviour, meaning that higher prices have a negative effect on consumers’ purchase behaviour. Therefore, the covariate price, the price per unit, is included in this study as an interval variable. The by GfK provided dataset included only the paid prices for the soft drink brand. The price for the weeks in which the household did not make a purchase, therefore, is unknown. The missing prices were replaced by the average price per week, which was derived from the prices that other households paid for the brand during that week.

Competitor purchases. The research of Brown et al. (2012) shows that in a competitive

market, where consumers buy multiple brands, it is more likely that consumers make use of the brand information provided in advertisements. Therefore, advertisements may have a greater impact on brand performance and result in more brand purchases (Brown et al., 2012). For that reason, competitor purchases are included as a covariate in this study. Competitor purchases are included as a nominal variable, in which 1 represents that the household bought a competitive brand and 0 that the household did not bought a competitive brand.

Prior Category purchases. This covariate indicates how frequently a household bought the

brand in the past. Earlier research shows that previous purchases of soft drinks in the supermarket by consumers are an indication of their current preferences (Dubé, 2004). In other words, past purchases of consumers influence their current purchases. Therefore, prior category purchases are included as a covariate in this study. Prior category purchases are included in the dataset as a nominal variable, where 0 represents not a consumer, 1 a light consumer, 2 a medium consumer, and 3 a heavy consumer.

District. The researchers Sethuraman, Tellis, and Briesch (2011) show that the advertising

Age. Research by William and Page (2011) shows that consumers of different age categories

have different preferences regarding advertising channels. Especially younger people, for example, have lessened their time watching television or stopped watching television altogether, while increasing their use of digital television (Oblinger, 2008). Therefore, the variable age can influence the relationship between advertising channels and consumers’ purchase behaviour, and is included as a covariate in this research. Age was included in the dataset as an interval variable.

The researcher Dubé (2004) shows that both Income and Size of the Household have an impact on consumers’ purchase behaviour. His study shows that a larger household results in more purchases. Additionally, income and education influence consumers’ brand choices, as people with higher incomes and higher education levels are more inclined to buy high-end quality brands (Dubé, 2004; Verrecken et al., 2005). Thus, income, size of the household, and education are included as covariates in this research. These three covariates are included in the dataset as interval variables. 3.4 Analysis plan

The analysis plan for this research consists of the model building steps: specification, estimation, validation, and use (Leeflang et al., 2015: 19). The fourth step, “use”, is not extensively discussed in this analysis plan. The resulting models from this research will be used to accept or reject the set hypotheses based on the coefficients in the models.

3.4.1 Specification

In the specification step of the model building process, the variables that will be included in the model are specified (Leeflang et al., 2015: 20). Regarding the model specification, the in paragraph 3.3 discussed independent variables were selected based on theory. Next to the literature based part of the specification, cross tabs, independent sample t-tests, one-way ANOVA’s, and bivariate correlation matrixes are used as a secondary variable selection tool, to see if the variable selection based on the data is in line with the selection based on theory. Besides the selection of the variables, these analyses, will also be used to get a first impression of the relationships between the independent variables and the dependent variables. Because the two elements of consumers’ purchase behaviour are different types of variables, the first impression for both elements cannot be obtained with the use of the same analyses.

For the second element, consumers’ purchase quantity decision, the first impression of the influence of the interval independent variables will be created with the use of a bivariate correlation matrix. The bivariate correlation matrix can assess the strength of the association between one interval variable and another (Malhotra, 2010: 562). Therefore, it gives a good first impression of the relation between the interval independent variables and the interval second element of the dependent variable, purchase quantity decision. As the second element of the dependent variable is an interval variable, the independent sample t-test can again be used for the nominal independent variables with two categories. However, the independent sample t-test cannot be used for the nominal independent variables with more than two categories. A first impression of their influence on the second element of the dependent variable is created with the use of a one-way ANOVA. It assesses whether the nominal variable has an influence on the number of products bought by the household by analysing whether the average amount of products bought by the household differs per category of the independent variable (Malhotra, 2010: 531).

3.4.2 Estimation

Leeflang et al. (2015: 20) define the estimation step as “the determination of parameter estimates for a model”. Before the parameters of the models are actually derived from the models, the below stated preliminary requirements for the Logistic Regression and the count model were checked and met.

Assumptions Logistic Regression (StatisticsSolutions, n.d.): 1. A binary dependent variable.

2. Factor level 1 of the dependent variable should indicate the desired outcome and factor level 0 the undesired outcome.

3. The Logistic Regression should have the optimal model fit and should consist of merely the meaningful variables.

4. The independent variables need to be linearly related to the log odds. 5. The observations should be independent from each other.

Assumptions Poisson Regression/Negative Binomial Regression (Statistics Laerd, n.d.): 1. A dependent variable that exist out of count data.

2. The model should consist out of at least one independent variable, measured on a continuous, ordinal, or nominal scale.

3. The observations should be independent from each other.

4. The dependent variable follows a Poisson distribution (or for the Negative Binomial Regression a Negative Binomial distribution).

5. The mean equals the variance of the model (or for the Negative Binomial Regression, the variance should exceed the mean; there should be over-dispersion).

Estimation Logistic Regression

The aim of the Logistic Regression in this research is to determine how the probability that a consumer buys the brand is influenced by the independent variables (Leeflang et al., 2015: 264). Therefore, the effectiveness of the advertising mediums will be evaluated regarding their influences on the probability that the consumer is going to buy the product. In other words, the parameters of the Logistic Regression are interpreted in relation to the probability of the consumers buying the brand. The parameters of the independent variables are acquired through the use of maximum likelihood estimation, and these parameters will be interpreted by taking the exponents of the parameters, the odd ratios. The odd ratios of the different advertising channels will indicate the effect of a consumer being exposed to an advertisement relative to not being exposed to it on the probability that the consumer buys the brand, given that the other independent variables are held constant. By comparing these effects of the different advertising channels this study compares the effectiveness of the different advertising channels on consumer’s purchase decision.

Estimation Poisson Regression/Negative Binomial Regression

The exponents of the parameters indicate the factor with which the amount of products that the consumer is expected to buy is multiplied in case the independent variables changes by one unit. Thus, for the different advertising channels, the exponents of their parameters is the number by which the consumer’s amount of products bought is multiplied when they are exposed to an advertisement relative to not exposed, given that the other independent variables are held constant. By comparing the exponents of the parameters of the different advertising channels, this study compares the influence of the different advertising channels on consumer’s purchase quantity. 3.4.3 Validation

In the validation stage, an assessment is made of how well the created model fits the data and, therefore, if the estimated parameters fit well with the data. This assessment is (except for the Hit Rate) based on a comparison between the created model with all its variables and a model with only an intercept. The created model is only valid when it is significantly better than the intercept only model (Leeflang et al., 2015: 156).

Validation Logistic Regression

To assess the validity of the Logistic Regression, the following validation measures will be used to evaluate whether the Logistic Regression indeed has a significantly better fit with the data than the intercept only model.

1. The Hit Rate “indicates the percentage of correctly classified observations” (Leeflang et al., 2015: 269). It assesses in what percentage of observations the model correctly predicts whether the consumer will buy the product or not, and thereby it assesses how well the model performs. 2. Nagelkerke R2 is a pseudo r-square statistic, which is calculated with the following equation

(Leeflang et al., 2015: 269): 𝑅2₌1 − ( 𝐿𝑛𝑢𝑙𝑙 𝐿𝑘 ) 2 𝑛 1 − (𝐿𝑛𝑢𝑙𝑙)2𝑛

3. McFadden R2 is also a pseudo r-square statistic, which is calculated with the following equation (Leeflang et al., 2015: 269):

𝑅2 = 1 − ( 𝐿𝐿𝑘 𝐿𝐿𝑛𝑢𝑙𝑙)

LLk stands for the log likelihood of the complete model and LLnull for the log likelihood of the null model. Again, the higher the pseudo r-square the better the model fits the data.

4. Akaike’s Information Criterion (AIC) is an information criterion, where counts, the lower the AIC, the better the model fits (Leeflang et al., 2015: 158). AIC is calculated through the following formula:

𝐴𝐼𝐶 = (−2𝐿𝐿 + 2𝐾)/𝑁

K indicates the number of parameters, N the number of observations, and LL the log likelihood of the model.

5. The likelihood ratio test “is used to investigate two models that are nested, and it chooses that model that has the highest likelihood, given the observed data” (Leeflang et al., 2015: 197). Nested means that the larger model is an extended version of the smaller model, which is the case here, since the intercept only model is compared with an extended version. The likelihood ratio is calculated with the following formula:

Validation Poisson Regression/Negative Binomial Regression

The validity of the selected count model will be assessed with the use of the following model fit measurements:

1. Akaike’s Information Criterion (AIC) 2. Likelihood ratio test

In addition to the already discussed Akaike’s information criteria and likelihood ratio test, the model fit of the count model as well will be assessed through the following information criteria:

3. Bayesian Information Criterion (BIC):

𝐵𝐼𝐶 = −2 ln(𝐿) + ln(𝑁) 𝑘 4. Consistent Akaike’s Information Criterion (CAIC):

𝐶𝐴𝐼𝐶 = −2 ln(𝐿) + (ln(𝑁) + 1) 𝑘

4. RESULTS

This chapter reports the results of the analyses conducted in this research. It starts with the specification of the Logistic Regression and the count model based on the data sample. Through this it checks the model specification based on literature, and it provides an initial indication of the relationships between the advertising channels and the two elements of consumers’ purchase behaviour. Secondly, the estimation, validation, and use of the models are discussed. This starts with checking the model assumptions. After that the Logistic Regression and count model are estimated, validated, and interpreted. This model building process start with only the dependent variable, the covariates, and the direct effects of the advertising mediums included. The next step is assessing whether the retention variables of the advertising variables have a better fit with the data than the direct variables. Finally, the process ends with analysing whether the synergy effects are significant, and if the model benefits from adding synergy variables to the model or not. This way the models start relatively simple and become more complex. The models only become more complex when the new variables (the retention and synergy variables) improve the models, resulting in a better model fit. Based on the resulting final models the hypotheses will be accepted or rejected.

4.1 Specification

4.1.1 Specification Logistic Regression

In this section, cross tabs and an independent sample t-test are performed to see whether the model specification based on theory is in line with the variable selection based on the data sample. These will also provide an initial benchmark of the relationships between the independent variables and the dependent variable that will be part of the Logistic Regression.

Table 3

Chi-square tests for the nominal independent variables Variable Pearson

Chi-Square value Sig. TV_contact 5.197 .023 YT_contact 2.578 .108 RTL_contact .033 .855 Competitor Purchase 97.762 .000 Prior category purchases 155.356 .000 District 31.891 .000

The cross tabs contradict the literature, and lead to the initial impression that the different advertising channels do not explain much of the variance in consumers’ brand purchase decision. There is only a significant difference in the brand purchase decision between the consumers who were exposed to advertisements through television and those who were not. In addition, all nominal covariates seem to result in different brand purchase decisions, which confirms the inclusion of them in the Logistic Regression.

For the rest of the covariates that were selected by literature, the interval variables, an independent sample t-test was conducted. The independent sample t-test shows whether the averages of the interval independent variables are significantly different for the people who bought the brand compared to those who did not buy. Therefore, it provides an initial impression of the relationships between these covariates and the first element of the dependent variable: purchase decision.

Table 4

Independent sample t-test for the interval independent variables Variable t Sig. (2-tailed) Price -8.477 .000 Age -2.137 .033 Income 4.455 .000 Household size .825 .026 Education 1.340 .180

4.1.2 Specification count model

As for the Logistic regression, multiple analyses are performed to see if the model specification for the count model based on theory is in line with the variable selection based on the data sample. These analyses also provide a first impression of the relationships between the independent variables and the dependent variables within the count model. However, because the count model will be performed on an interval dependent variable, the first impression will be created through different analyses. The impression of the two level nominal independent variables will be created by an independent sample t-test. A one-way ANOVA will be used for the more than two level nominal independent variables. Lastly, a bivariate correlation matrix will be applied for the interval independent variables.

Table 5

Independent sample t-test for the two level nominal independent variables Variable t Sig. (2-tailed) TV_contact 1.013 .311 YT_contact 1.143 .253 RTL_contact .496 .620 Competitor Purchases -6.932 .000

As was the case for the purchase decision of consumers, the independent sample t-test also contradicts literature regarding the influence of the advertising mediums on consumers’ purchase quantity decision. Table 5 shows that the average amount of products bought of the brand does not differ significantly between the people who were exposed to advertisements through the different advertising mediums, and those who were not. Consequently, this t-test suggests that it is not expected that the advertising channels will have an effect on the purchase quantity decision of consumers. In addition, table 5 confirms the inclusion of competitor purchases in the count model (p = .000).

Table 6

One-way ANOVA for more than two level nominal independent variables

Variable F Sig.

Prior category purchases

51.020 .000

District 7.315 .000

Secondly, a one-way ANOVA was performed to see whether the average amount of products bought by the households was significantly different per district and per category of the variable prior category purchases. Table 6 shows that the amount of products bought by the households differs significantly per district and per category of the variable prior category purchases. Consequently, it indicates that it is likely that these covariates have an influence on the purchase quantity decision of consumers, and it confirms the inclusion of them in the count model.

Table 7

Bivariate correlation matrix for the interval independent variables Variable Pearson Correlation Sig. (2-tailed) Price -.383 .000 Age .006 .568 Income .022 .027 Household size .059 .045 Education .014 .146

4.2 Estimation, validation, and use

4.2.1 Preliminary checks Logistic Regression

4.2.2. Estimation, validation, and use Logistic Regression

Direct effect

The process of building the best performing Logistic Regression starts with analysing the direct effects of the advertising mediums on the first element of the dependent variable, purchase decision. The direct variables are combined in the Logistic Regression model with the in the methodology discussed covariates. Table 8 shows the resulted complete Logistic Regression.

Table 8

Logistic Regression with all the variables included

Variable Exp(B) (SE) Sig.

TV_contact 1.150* (.076) .065 YT_contact 1.502 (.261) .119 RTL_contact 1.103 (.547) .858 Price .266*** (.139) .000 Competitor Purchases .258*** (.118) .000 PCP (0) not a consumer .000 (1) Light consumer 1.454*** (.137) .006 (2) Medium consumer 2.561*** (.139) .000 (3) Heavy consumer 5.019*** (.143) .000 District (0) North .000 (1) East 1.167 (.135) .255 (2) South 1.229 (.129) .111 (3) 3 Largest cities 1.548*** (.148) .003 (4) Rest West .886 (.134) .363 Age .966* (.018) .055 Income 1.025*** (.009) .006 Household size .880*** (.034) .000 Education .994 (.030) .848 *** p < .001, ** p < .05, * p <.1

Table 9

Significance level if variable removed Model if removed

Variable Sig. of the change

Step 1: YT_contact .135 RTL_contact .860 Education .848 Step 2: YT_contact .132 Education .848 Step 3: YT_contact .132

Table 9 shows that the variables, exposure to advertisements on the RTL website, education, and exposure to advertisements on YouTube, do not significantly contribute to the Logistic Regression, and that they can be removed from the model without changing the model fit significantly. Table 10 confirms this, as it points out that the model fit changes barely.

Table 10

Measurements of model fit of the Logistic Regression -2LL Cox & Snell

R2

Nagelkerke R2

AIC

Complete model 5858.043 .148 .284 .5816610

Model without RTL_contact 5858.074 .148 .284 .5814669

Model without RTL_contact and education

5858.110 .148 .284 .5812732

Model without RTL_contact, education, and YT_contact

5860.377 .148 .283 .5812995

Table 11

Logistic Regression without YT_contact, RTL_contact, and education

Variable Exp(B) (SE) Sig.

TV_contact 1.149* (.076) .067 Price .268*** (.139) .000 Competitor Purchases .258*** (.118) .000 PCP (0) not a consumer .000 (1) Light consumer 1.459*** (.137) .006 (2) Medium consumer 2.563*** (.139) .000 (3) Heavy consumer 5.069*** (.143) .000 District (0) North .000 (1) East 1.171 (.135) .243 (2) South 1.234 (.129) .103 (3) 3 Largest cities 1.546*** (.148) .003 (4) Rest West .887 (.133) .368 Age .967** (.017) .048 Income 1.024*** (.009) .006 Household size .884*** (.034) .000 *** p < .001, ** p < .05, * p <.1

Table 11 suggests that exposure to television advertisements has a marginally significant positive direct effect on consumers’ brand purchase decision (Exp(B) = 1.149, p = .065). Furthermore, in line with the results from the performed cross tabs, the exclusion of the direct effects of exposure to online video advertisements on the RTL website (non-skippable) and on YouTube (skippable) from the Logistic Regression, indicates that they have no significant effect on whether or not consumers buy the brand. Table 11 also indicates that, confirm the results from the conducted cross tabs and the independent sample t-test, except from education, all the covariates appear to have a significant influence on consumers’ purchase decision.

Retention effect

The first part is analysed by comparing the model fit and the significance level of the different retention rates (25%, 50%, or 75%) for each advertising medium per week. The comparison for week 10 is depicted in table 12, while the comparisons of the other weeks can be found in appendix II.

Table 12

Comparison retention variables week 10 10 weeks

-2LL Sig. Exp(B) Sig. (whole model) Hit rate Nagel -kerke R2 Cox & Snell R2 McFadden R2 AIC LR TV_25_w10 5863.128 0.452 1.027 .000 90.7% .148 .283 0.2163466 0.5815708 1618.659*** TV_50_w10 5863.355 0.564 1.018 .000 90.7% .148 .283 0.2163162 0.5815932 1618.432*** TV_75_w10 5863.641 0.835 1.005 .000 90.7% .147 .283 0.2162780 0.5816214 1618.146*** YT_25_w10 5858.150 0.118 1.393 .000 90.7% .148 .284 0.2170119 0.5812771 1623.637*** YT_50_w10 5858.283 0.130 1.313 .000 90.7% .148 .284 0.2169941 0.5812902 1623.504*** YT_75_w10 5858.756 0.184 1.210 .000 90.7% .148 .283 0.2169309 0.5813369 1623.031*** RTL_25_w10 5860.033 0.545 1.357 .000 90.7% .148 .283 0.2167602 0.5814628 1621.754*** RTL_50_w10 5859.602 0.359 1.519 .000 90.7% .148 .283 0.2168179 0.5814203 1622.185*** RTL_75_w10 5859.160 0.249 1.568 .000 90.7% .148 .283 0.2168769 0.5813767 1622.627*** *** p < .001, ** p<.05, *** p<.05

These comparisons resulted in the best performing retention variables per week and per advertising medium. The next step is comparing the retention variables for the ten weeks with each other and with the direct variables (see appendix III). The comparison with the direct variables is made to see whether or not there is a retention effect for the different advertising mediums. If the retention variables for the advertising channels are not performing better than the direct effect variables, there appears to be no retention effect. The result of this comparison is shown in table 13. In both table 12 and table 13 only the retention variables are depicted. However, the retention variables are analysed in combination with the other variables within the developed Logistic Regression, during the analyses on the direct effects.

Table 13

Best performing retention variables in the Logistic Regression -2LL Sig. Exp(B) Sig.

The variable for the direct effect of television advertising has a better fit with the data and a higher significant level than all the retention variables for television advertising. This indicates that television advertising has no retention effect, but only a positive direct effect on consumers’ brand purchase decision (Exp(B) = 1.149, p = 0.067). Just like the direct effect, the retention effect for online video advertisements on the RTL website seems to be insignificant (Exp(B) = 1.873, p = 0.149). The results, on the other hand, do indicate that online video advertisements on YouTube have a retention effect (see table 13). The results suggest that the skippable advertisements to which consumers are exposed to through YouTube have a marginal significant positive effect (Exp(B) = 1.380, p = 0.052) on consumers’ brand purchase decisions for three weeks and with a retention rate of 75%. In addition, because the results suggest that YouTube advertisements have a marginally significant positive effect all along the three weeks after exposure, the results indicate that they must also have a positive direct effect. The results from the retention variables overrule the results from the direct variables of YouTube, because the retention variables have a better fit with the data, and are, therefore, better able to explain the relationships within the data sample.

Because the results indicate that exposure to RTL advertisements does not have a significant retention effect, but exposure to YouTube advertisements has a marginally significant effect, it is not clear whether online video advertisements have a positive direct effect and a positive effect over time on consumers’ purchase decision. To get clear answers, further research is conducted to see whether online video advertising (Skippable and/or non-skippable advertisements) has a significant retention effect (see appendix IV).

Table 14

Model fit best performing online retention variable and online variable -2LL Sig. Exp(B) Sig.

(whole model) Hit Rate Nagel -kerke R2 Cox & Snell R2 McFadden R2 AIC LR Online_75_w3 5855.864 .026 1.457 .000 90.7% .148 .284 0.2173175 0.5810517 1625.923*** Online_contact 5858.653 .175 1.386 .000 90.7% .148 .283 0.2169447 0.5813267 1623.134*** *** p < .001, ** p<.05, *** p<.05

Furthermore, because the results suggest that online video advertisements have a significantly positive effect during the three weeks after exposure, the results also indicate that they must have a significantly positive direct effect after the exposure on consumers’ brand purchase decision. In both table 14 and in appendix IV, only the retention variables are depicted. However, the retention variables have been analysed in combination with the other variables within the developed Logistic Regression, during the analyses on the direct effects.

Synergy effect

The last step in the process of building the Logistic Regression is to analyse whether the synergy effects are significant, and to determine if the model fit increases from adding synergy variables to the model. Because the analyses of the retention effects shows that for YouTube advertisements and Online video advertisements there exist a retention effect, the analyses of the synergy effect with these advertising mediums are conducted with the use of the retention variables. To be able to test the synergies between the advertising mediums, the direct or retention variables (based on which performed best) need to be included in the model as well. Therefore, the insignificant RTL_75_w2 retention variable is added where needed.

The following five synergies will be tested: 1. TV_contact X YT_75_w3

2. TV_contact X RTL_75_w2 3. YT_75_w3 X RTL_75_w2