CUSTOMER RESPONSE TO PUSH ADVERTISING COMMERCIALS: A LATENT PROFILE ANALYSIS

(1)

Master Thesis

CUSTOMER RESPONSE TO PUSH ADVERTISING

COMMERCIALS: A LATENT PROFILE ANALYSIS

By

Sara van Es

26

th

_{June 2017}

University of Groningen

Faculty of Economics & Business

MSc Marketing Intelligence & Marketing Management

1

st

_{supervisor: dr. ir. M. (Maarten) Gijsenberg}

2

nd

_{supervisor: P. S. (Peter) van Eck}

Nieuwe Blekerstraat 84

(2)

2

Preface

In front of you is my final project completed for the MSc Marketing Intelligence & Marketing Management at the University of Groningen,

my master thesis called "Customer Response to Push Advertising Commercials: A Latent Profile Analysis".

The reason why I chose the subject of (online) media channels is because it is a relatively new topic that has not been studied in depth already. Intuitively, people consider using online media channels as a great strategy

for organisations to attract new customers or to increase their spendings. I was interested in finding out whether this effect on sales could be observed

in reality and what kind of psychological processes are behind this behavioural change of customers.

It would have been very difficult for me to write this thesis without the feedback from my first supervisor Maarten Gijsenberg. I want to thank him for always clearly pointing out the improvements that could be made

in every chapter and also for finding the time to answer my questions about issues with the dataset and the methods that I used in my thesis. I also want to thank my second supervisor Peter van Eck, who provided

the option to work with a dataset from GfK panel. Without the great opportunity to work with this dataset consisting of a large number of households, it would not have been possible to present an answer to my research question. Two other people who helped me while I was writing my thesis by providing feedback are Stephanie Becker and Margriet van Weperen. I also want to thank both of them for their advice during the

entire process.

The general message of this thesis is that as long as organisations spend time on trying to understand who their customers are, it is likely that they

will succeed in presenting suitable media strategies to these individuals. I think that this idea not only applies to the processes described in my

thesis, but also to the process behind writing my thesis.

When I started writing my thesis, I never expected that I would be able to apply certain types of analysis that I had never used - or even heard of - before. A valuable thing that I learned was that if you put enough time into

trying to understand new things, there is a big chance that you will actually succeed in what you are trying to do.

Sara van Es, Groningen, 26th

(3)

3

Abstract

In this study, the Elaboration Likelihood Model is applied to the process of push advertising

activities to argue that these are particularly effective methods in a fast-moving consumer good

setting, in which customers often make buying decisions based on heuristics. A panel dataset

consisting of 7.742 households was used to test the effect of the exposure to three types of push

advertising activities (i.e. television commercials, in-stream commercials on a media streaming

website and in-stream commercials on YouTube) on the amount of euros spent by the respondents

on a soft-drink brand. A Latent Profile Analysis was applied to place similar respondents into

discrete homogeneous segments. The results indicated that a model dividing the respondents into

three segments had the best fit with the data. The identified segments are the frozen non-buyers

consisting of the majority of the respondents who seldom buy products from the brand and are not

affected by any type of push advertising; the delicate potentials who are profitable (potential)

buyers of the brand but who are strongly affected by the exposure to online push advertising in a

negative way; and finally the fickle fans who are not strongly affected by any type of push

advertising but are perceived as the most profitable buyers of the brand.

Key words: push advertising, segments, fast-moving consumer goods, in-stream commercials,

(4)

4

communication technologies. Where organisations used to have access to consumers through only a few traditional media channels (e.g. television, radio), new communication technologies created opportunities for organisations to reach consumers through an increasing number of media channels (Franz, 2000; Mareck, 2014). However, while the number of media channels expanded, in many organisations the marketing budget remained on the same level as a few decades ago (Danaher and Dagger, 2013).

Nowadays, the allocation of this marketing budget is often based on past experiences or intuitive feelings of managers instead of on insights based on data about the effectiveness of the marketing activities (Wiesel et al. 2011; Zhao and Zhu, 2010). In order to stimulate an optimal allocation of the marketing budget, it is becoming increasingly relevant to study the effectiveness of the new online media channels in the academic field of marketing.

In this study, the existing knowledge about the effectiveness of the media channels will be updated by focusing specifically on the effectiveness of push advertising channels. Push advertising channels can be characterised in the sense that they interrupt the activities of the user while serving on the internet, for example in the form of in-stream video commercials or banner advertising (Spilker-Attig and Brettel, 2010). In contrast, on-demand channels are actively sought out by consumers, typically by visiting a price-comparison website or by using a search engine to find more information about certain products. The major difference between push channels and on-demand channels is that users often perceive push channels to be intrusive, whereas this is not the case with on-demand channels (Unni and Harmon, 2007; Belance et al., 2017). Furthermore, previous studies give different conclusions about the effects of push advertising channels, with Manchanda et al. (2006) showing an effect of push advertising on sales, whereas in other studies no effect was found on the amount of orders (e.g. Spilker-Attig and Brettel, 2010) or the buying intentions of customers (e.g. Unni and Harmon, 2007).

While contradicting results are found regarding the effectiveness of push advertising channels, the previous studies did find an indication of heterogeneity across consumers in their response towards push advertising (e.g. Cho, 2003; Manchanda et al., 2006; Möller and Eisend, 2010). In these studies, researchers highlighted the need to further explore the diverse customer responses to push advertising activities. Also, while much attention has been given to banner advertising in earlier studies on push advertising activities, less

consideration has been given to the effect of in-stream video commercials as a type of push advertising (“push advertising commercials”). The current study will address both gaps by looking at the effect of three types of push advertising commercials on the buying behaviour of customers and by identifying

(6)

6

Furthermore, demographic information about the households was made available to find commonalities among the customer segments. Where previous studies on push advertising briefly touched upon the heterogeneity in customer response (e.g. Manchanda et al., 2006; Cho, 2003), these studies did not explore any commonalities among customer segments based on demographics. A practical contribution of this study is that it allows organisations to identify the segments of consumers who will actually change their buying behaviour after coming into contact with certain types of push advertising commercials. Therefore, organisations will be able to deliver an integrated marketing mix in which they tailor the advertising message to the specific needs of the customer segments that are reached through the specific push advertising channel. In this sense, the study provides new insights into stimulating an optimal budget allocation of marketing departments in fast-moving consumer good (FMCG) industries.

(7)

7 THEORETICAL REVIEW

In this chapter, an interdisciplinary approach is used to create an understanding of the process behind push advertising and the ways it can influence customers, building on knowledge from the fields of psychology and marketing. Next, the unique characteristics of a FMCG setting are explored, together with the possible consequences for the effectiveness of push advertising. In the end of this chapter, an overview is given of the three types of push advertising commercials that are explored in this study and their possible effects on the purchase behaviour of customers.

Push advertising channels

The field of psychology laid the foundation for influential theories on the process of advertising. One of these theories resulted in the Elaboration Likelihood Model (ELM), which explores the processes underlying the effectiveness of communication (Petty and Cacioppo, 1986). In particular, this model is suitable for understanding the processes behind communications that aim to change the behaviour of consumers. In short, this model states that individuals who feel a high level of motivation, use the central

route to thoroughly process the information they receive, which results in a change in attitude that is

relatively persistent (Cacioppo and Petty, 1984). However, individuals who do not feel a high level of motivation use the peripheral route to process the received information only in a superficial way, which will at best result in a temporary attitude shift. The level of motivation is determined by the degree of personal relevance the person experiences when receiving information on the subject and could therefore be different in each situation (Petty, Cacioppo and Goldman, 1981).

In order to adapt the ELM to the field of online advertising, Cho (1999) argued that users often have the option to voluntary click on links before receive any advertising message. This leads to a situation in which consumers feel they are in control of the type of advertisements they receive. When users click on the link, they will perceive this information to be personally relevant for them, which - together with their active role in deciding which advertisements to view - automatically results in a higher level of involvement (Cho, 2003; Pashkevisch, Sundar, Kellar and Zigmond, 2012). In their classification of the different online media channels, Spilker-Attig and Brettel (2010) based their definition of on-demand channels on this active role that users experience when coming into contact with search engine advertising or price-comparison websites. The active role leads to central processing of the information, making consumers focus on the product information in the advertisement (Petty and Cacioppo, 1981; Cho, 2003).

(8)

8

(Möller and Eisend, 2010; Pashkevisch et al., 2012). This passive role and the lack of motivation to process the advertisement results in a lower level of involvement. In turn, this leads to peripheral processing of the advertisement, making individuals focus on the peripheral cues in the ad as opposed to in-depth product information (Cho, 2003).

The previous findings have important implications for the difference in effectiveness between push advertising channels and on-demand channels. In previous studies it was often concluded that a push approach is less effective than an on-demand approach in influencing buying decisions, while resulting in a higher concern for privacy and feelings of intrusiveness (Unni and Harmon, 2007; Spilker-Attig and Brettel, 2010; Pashkevitsch et al., 2012). However, the existing literature suggests that peripheral cues such as the size of the advertisement, the use of colour and dynamic effects are more influential in push advertising (Cho, 2003) than in on-demand advertising. In other words, if push advertising only results in peripheral processing, which leads to a temporary attitude shift in consumers then this type of advertising could be more effective in situations in which consumers make decisions based on superficial cues. For instance, Unni and Harmon (2007) found that push advertising was effective when the consumer already showed a preference for buying a certain product and only needed a last push to actually purchase the product. The next section explores a specific setting in which push advertising could be an effective method for influencing the buying behaviour of consumers.

Fast-moving consumer goods

The FMCG industry is characterised by selling supermarket goods to consumers at a relatively low price level on a regular basis (KPMG, 2016). Examples of the products that are commonly sold in this industry are meat, vegetables, personal hygiene and soft drinks. Although the markets for fast-moving consumer goods are often mature and saturated, they account for a large part of the consumer's budget (Stuart and Sumner, 2012). Consequently, the FMCG industry is made up out of a large number of brands competing for the customer's attention (Haddad, 2016). The FMCG industry could be one of the settings in which push advertising is an effective technique for influencing buying behaviour. The reason for this is that the FMCG industry is characterised by an unlimited amount of choices, which stimulates consumers to make choices based on heuristics and simple cues.

In the last few decades, a popular method to reach consumers through this wide range of offerings has been to provide price promotions (e.g. Rao, 2009; Selwood, 2016). However, as the profit margins on

convenience goods are already quite small and the price perceptions of consumers are influenced by price promotions, organisations have started to look for other ways to differentiate their products in the FMCG industry (Roderick, 2016). Since consumers often perceive the products from different brands in the FMCG industry as being relatively similar to each other (e.g. Thøgersen, Jørgensen and Sandager, 2012),

(9)

9

When buying convenience goods, consumers use heuristics to decrease the amount of time and effort they spend making choices in the shopping area (e.g. Dickson and Sawyer, 1990; Thøgersen, Jørgensen and Sandager, 2012). Therefore, it can be argued that consumers will also feel a low level of motivation to process the information in advertisements related to these convenience goods. Instead, they might focus on peripheral cues to decide on an attitude towards the brand or product. In support of this view, Reinartz and Saffert (2013) found that the buying decisions of consumers are influenced twice as much by creative ads in the FMCG industry, as opposed to ads describing the product attributes or benefits of the product. Because consumers in the FMCG industry are more affected by these peripheral cues as opposed to in-depth product information (Cho, 2003), push advertising might be particularly useful to influence buying behaviour in this setting.

Heterogeneity in customer response

In studying the effectiveness of push advertising in a FMCG industry, it is important to take into account possible heterogeneity in customer response to this type of advertising. This way, organisations can focus on the segments of consumers that are most influenced by this push advertising. As explained in the introduction, earlier studies touched upon this subject by stating that different segment of consumers can be identified based on their response to push advertising strategies.

At the lowest level, this heterogeneity in customer response is perceived to be caused by differences in the level of specific product involvement, with consumers who show a high level of involvement with the product category having a higher click-through-rate when viewing a banner advertisement (Cho, 2003). In addition, Manchanda et al. (2006) perceive the heterogeneity to be the result of the level of sensitivity from consumers to the amount of advertising they are exposed to and their sensitivity to the number of channels in which the advertisement is shown.

Figure 1. The three types of push

(10)

10

In this study, the heterogeneity is further explored by dividing consumers into segments who react

differently to three types of push advertising channels. The three types of push advertising channels can be placed on a continuum from traditional to modern, which is determined by the level of personal relevance (i.e. to what extent can the commercials be adapted to the interests of the viewer) and the level of

accessibility (i.e. on how many devices has the viewer access to the content in which the commercials are presented). A visualisation of the differences between the three push advertising channels is presented in

figure 1.

By observing whether changes in the three push advertising variables lead to different purchase behaviours of consumers, it will be possible to find segments of consumers who alter their buying behaviour after coming into contract with certain types of push advertising (see push advertising variables in figure 2). After finding segments of consumers who react differently to the push advertising variables, commonalities in demographics among segments can be identified based on the demographic variables (see demographic

variables in figure 2). In the next section, the characteristics of the three types of push advertising

commercials displayed in the conceptual model along with their possible influences on the buying behaviour of consumers are further explored.

Figure 2. Conceptual model displaying the three push advertising variables as affecting the purchase behaviour of

consumers while the segment to which the customer belongs performs as a moderator

Traditional television commercials

The use of television channels to broadcast commercials is the type of push advertising channel that can be considered the most “traditional” of the three push advertising channels in this study, which is also

(11)

11

previous years. In contrast, the age group between 18 – 24 years only watched 77 hours of television per month, which is a downward trend from the years before (Nielsen, 2016).

On average, the three oldest age categories show an upward trend in the number of hours spent watching television compared to the previous years, while a downward trend can be observed for the two youngest age categories. Instead of watching television, the age groups 18 – 24 years and 25 – 34 years have started to spend more and more time watching online content via their electronic devices. In fact, every age group shows an upward trend in terms of hours spent on viewing videos via laptops or tablets, indicating that people in general are getting more interested in the option to watch content online.

Figure 3. Hours of monthly television watching and video watching on the computer per age group

in the second quarter of 2016 (Nielsen, 2016).

In previous audience reports, a link was found between the level of education and the amount of television watching a person reports. Individuals with a higher level of education watch less television on average (Bureau of Labour Statistics, 2011), especially early in the morning and later in the day (Friedman, 2013). Child Trends Databank (2014) also found a linear relationship between the education of the parents and the amount of television watching of their children. In eighth grade, children spent on average 26 hours per week watching television when their parents had only completed high school. In contrast, the children of parents who had completed graduate school only spent 14 hours per week watching television.

Other findings show that the household category of a potential customer can be linked to the number of times they are exposed to television advertisements. For instance, the Bureau of Labour Statistics (2011) found that in households with children below 6 years old, the adults spent an average of 2.09 hours watching television. When the youngest children were between 6 and 17 years old, the respondents spent 2.31 hours watching television. Adults watched on average 3.13 hours television per day when the household did not consist of any children younger than 18 years old. This indicates that individuals have more time to spend on watching television when there are no young children in their household. The effect of traditional television advertising on the purchase behaviour of customers has been already demonstrated in previous studies (e.g. Zhou, Zhou & Ouyang, 2003; Liaukonye, Teixeira & Wilbur, 2015).

(12)

12

When comparing traditional advertising with online advertising, Danaher and Dagger (2013) found that traditional media channels are more effective than online media channels in influencing purchase

behaviour. The authors also found that while the traditional media channels were more influential in terms of sales, the online media channels increased the amount of traffic to the websites of the companies. This indicates that the strength of the online media channels could be in creating awareness for the brand instead of directly affecting the amount of purchases.

In-stream commercials on media streaming websites

When the revenues of television advertising started to decrease, more and more TV channels and radio channels decided to start broadcasting their shows on the internet (Jones, 2007). By making the content of their radio or tv-shows available online, these channels tried to gain additional advertising revenues by presenting viewers with clickable in-stream video commercials. Advertising via online television streaming shares similarities with traditional television advertising in that it allows TV channels to play non-skippable video commercials to target audiences.

However, it offers an advantage in the sense that younger consumers who are difficult to target through traditional television advertising are particularly attracted by the option to watch shows online (Logan, 2011; Nielsen, 2016). A reason for this could be that young adults prefer the option to decide where, when and how they want to watch the TV or radio show with the accompanying advertisements. In a recent survey by Pew Research Center (2013) it was also found that the percentage of American adults who said to watch online videos on a regular basis was higher for those with a college degree (84%), compared to the individuals who had only finished high school (71%) or been in college (80%). This indicates that

individuals with a higher level of education are more likely to come into contact with push advertising via media streaming websites than people with a lower level of education.

(13)

13

the brands associated with the ads and the messages of the ads when being exposed to online television streaming. The viewers also rated these ads higher in terms of likeability than television commercials. However, a drawback of this study is that the authors used an aggregate approach to find these differences, instead of comparing the effectiveness of each commercial in online television streaming with the same traditional television commercial.

A major difference with in-stream commercials on YouTube is that the commercials on online media streaming websites are often the exact versions of the commercials that were presented to viewers of the traditional TV or radio show (Logan, 2013). Therefore, these in-stream commercials are not specifically matched to the interests of the individual consumer who is watching the content. Instead, the in-stream commercials will be relevant to the average viewer of a certain TV or radio show.

In-stream commercials on YouTube

The popular platform YouTube attracts more than a billon users every day from countries all over the world. This streaming platform reaches users from every age category and is very popular with men and women between 18 – 34 years, who often prefer using this media channel over watching television

(YouTube, 2016). Specifically, 72% of the individuals between 18 – 34 years old use YouTube on a weekly basis. Of the age group between 35 – 50 years old, 58% uses YouTube. Between 51 – 69 years old, 43% of the individuals is a weekly YouTube user (Milward Brown Digital, 2015). On average, a video on YouTube has a duration of 4,2 minutes (Lella, 2014). More than half of the views come from users who watch YouTube videos on their mobile devices. This mobile option provides viewers with the opportunity to watch videos from every possible location where they have internet.

The process of attracting customers via YouTube commercials seems to be particularly effective when the brand already has a high level of awareness among customers. Kononova and Juan (2015) found that brand recognition increased the likelihood that the ad was recognised by viewers later on, indicating that strong brands gain more awareness through in-stream commercials on YouTube. The authors also found that users recalled YouTube advertisements better when they were already primed with a certain theme. As the platform has the advantage of being owned by Google, it has the opportunity to target commercials specifically to the interests of its audience, which increases the chance that the viewer is interested in the message of the advertisement.

(14)

14

of privacy as viewers notice that the ads are based on their browsing behaviour. In turn, this violation of privacy weakens the effect of personal relevance on the buying intentions of viewers.

Taking everything from this chapter together, table 1 provides an overview of the characteristics of the three push advertising channels. In table 2, an overview is provided of the demographic characteristics of the customers who are likely to be exposed to the push advertising commercials.

Table 1. Overview of the differences between traditional television commercials, in-stream commercials viewed during online

television streaming and in-stream commercials on YouTube based on the available literature.

Television Media streaming website YouTube Age of users Depends on the audience of the TV

show. The older age groups watch on average more television than the

younger age groups

Depends on the audience of the TV show, particularly attractive to the younger age

groups.

The younger age groups watch YouTube videos more

often than the older age groups, especially between

18 – 34 years old.

Level of education A higher level of education of parents is linked to a lower duration of television watching in the family

A higher percentage of people with a college degree said to watch online videos on a regular basis in comparison to

individuals who only completed high school

Household category Parents with young children watch less television than parents with

older or no children < 18 years

- No information available -

Table 2. Overview of the demographic characteristics of the audience of television channels, media streaming websites and

the YouTube platform based on previous audience reports.

Television Media streaming website YouTube Accessibility Only available on devices with

television streaming option

Available on many devices with internet connection

Available on nearly all devices with internet connection

Duration of videos Duration of the tv or radio show, which is usually > 20

minutes

Duration of the tv show, which is usually > 20 minutes

Depends on the video, with the average YouTube video being

4,2 minutes and the average YouTube session 40 minutes

Option to skip ads Not available Not available Available for certain ads

Option to zap Available Not available Not available

Length of commercial break

A couple of minutes, which makes it possible to perform other activities during the break

The length of one or two TV commercials; often too short to perform another activity

Only a few seconds; too short to perform another activity

Personal relevance Targeted to the general audience of the TV show

Targeted to the general audience of the show; often

the same commercials as during the original TV show

Linked to queries and demographics of users; companies only pay when

consumers view the ad

Intrusiveness Relatively low, as people traditionally associate the production of TV shows with

the need for commercials to cover the production costs

Higher than traditional TV advertising, as people associate online streaming to

a lesser extent with the need for commercials to cover

production costs

High when ads are not skippable as people do not traditionally associate a streaming platform with the need for commercials

Privacy Lower personal relevance means that ads are being perceived as less intrusive in terms of violations of privacy

Lower personal relevance means that ads are being perceived as less intrusive in terms of violations of privacy

Higher personal relevance means that ads are being perceived as

(15)

15 METHODS

Data collection

The dataset in this study originated from GfK panel, an international market research organisation. The organisation provided access to panel data of 10.703 Dutch households in the period December 30th, 2013 - March 29th, 2014. The variables described the experiences of customers with a specific brand in the soft-drink category on each day and were measured by means of self-reports and plug-ins on electronic devices. Three types of variables were accounted for in the dataset, with the first type of variables indicating the push advertising channels the households had been exposed to on each day. The second type of variables was related to the purchase behaviour of the household on each day and the third type of variables described their demographic characteristics.

Push advertising channels

The GfK panel dataset took into account the exposure of customers to three types of push advertising channels. First, the dataset reported the number of times an individual household was exposed to traditional television commercials (T). The second type of push advertising channels were in-stream commercials on online media streaming websites (R), which are characterised by the broadcasting of television shows. The third type of push advertising variables consisted of the exposure to in-stream commercials on the media platform YouTube (Y). In table 3, an overview is provided of the push advertising variables and their operationalisation in the dataset.

The active and passive variables for the exposure to TV commercials were combined into a new variable describing the number of times the household was exposed to a TV commercial of the brand (‘TV contact’). Respondents who did not participate in both the online panel and the TV panel (either passive or active) were removed from the dataset, resulting in 7.742 households of whom enough information was available to establish whether they had been exposed to each type of push advertising channels or not.

Variables Operationalisation

TV contact (T) The number of times the household members reported coming into contact with an ad on TV (passive, measured by self-reports) or the number of times that the household members should have come into contact with the ad given their TV viewing behaviour (active, measured by a recording device). A household either participated in the active or passive measurement.

RTL contact (R) The number of times that the household members came into contact with an online ad from the brand on the streaming website of Dutch broadcasting channel RTL, as measured by a plug-in on their devices.

YouTube contact (Y) The number of times that the household members came into contact with an online ad on from the brand on YouTube, as measured by a plug-in on their devices.

(16)

16

The respondents indicated in 6,2% of the daily observations that they had been exposed to a television commercial of the brand (see table 4). As indicated in figure 4, only 0,7% of the respondents were exposed to more than one television commercial on a day. The maximum number of times a respondent was exposed to a commercial of the brand on television was four times on one day. Interestingly, the respondents did not often report coming into contact with an ad of the brand via online channels. In respectively only 0,05% and 0,1% of the daily observations did the respondents report to have been exposed to a commercial on either RTL or YouTube.

Frequency (N) Percentage (%)

Variable No Yes % No % Yes

YouTube contact 696.057 723 99,9 0,1

RTL contact 696.425 355 99,95 0,05

TV (passive) 110.025 7.335 93,8 6,2

Table 4. The frequency and percentage of observations in which respondents were exposed to the push advertising and

marketing mix variables.

Figure 4. The percentage of daily observations in which the respondent reported to have been exposed to television

commercials of the brand.

Purchase behaviour

(17)

17

Purchase Do the household members report purchasing any products from the brand (1) or not (0)

Purchase value (S) The amount of euros the household members reported spending on products from the brand

Price per litre The price that the household members paid per litre of the product (in €)

Table 5. The purchase variables in the GfK dataset together with their operationalisations.

The distribution of the purchase value was first computed by including every respondent in the dataset, which resulted in an average purchase value of €0,06 per daily observation (see table 6). The average purchase value was €4,35 if only the observations were included in which a purchase was made (indicated by *). The relatively high standard deviation of 4,07 indicated large differences between the households in the price that they paid for the products. The maximum amount of euros paid for a purchase was €49,14. At first sight there appeared to be several outliers in the price per litre that the respondents had paid

according to the dataset. After a closer observation of these individual observations, it became clear that the price per litre was sometimes calculated incorrectly in the dataset. Therefore, these values were computed again using the volume that the customer reported to have bought and the price they paid for their purchase.

Variable N Min Mean Std. Dev. Max

Purchase value 696.780 0,00 0,06 0,69 49,14 Purchase value* 9.289 0,23 4,35 4,07 49,14 Price per litre* 9.289 0,53 1,28 0,42 4,75

Table 6. Descriptives of the purchase variables in the dataset

In figure 5, a boxplot of the newly computed price per litre is displayed. As expected, respondents paid on average a lower price per litre when the product was purchased during a promotional period. The lowest and highest values in this boxplot were not treated as outliers, because the lowest prices were paid by customers who each purchased between 18 and 36 units of the brand on one day. This indicates that they received some form of stack discount from the store. On the other hand, the highest prices were paid by customers who bought their product in small volumes (e.g. 200ml or 330ml), which resulted in relatively high prices per litre.

(18)

18

Demographic characteristics

When identifying customer segments based on the push advertising variables and purchase variables, it is relevant to explore the characteristics of each segment. The respondents in GfK panel were asked to fill out questions about their demographic characteristics, as described in table 7.

Buying intensity The intensity of the household in terms of buying behaviour, computed by using the total purchase value of the household in the entire period. with 1= never (€0,00), 2= light (< €7,90), 3= medium (< €23,50), 4= heavy (> €23,50).

Age of housewife The age of the housewife in number of years.

Household cycle The stage of the household cycle, indicated as 1 = single, 2 = family, youngest child is <6 years old, 3 = family, youngest child is between 6 – 12 years old, 4 = family, youngest child is between 13 - 17 years old, 5 = couple without any children younger than 18 years.

Level of education The highest level of education the household members completed (1 = primary school, 2 = LBO/MAVO, 3 = MBO, 4 = HAVO.VWO, 5= HBO/WO bachelor’s degree, 6 = HBO/WO master’s degree).

Table 7. Demographic variables in the GfK dataset, along with their operationalisations.

Approximately 30% of the respondents never bought the brand during the observational period (see figure

6). The majority of the respondents spent less than €7,90 in the three months they were providing

self-reports. Another part of 21,3% of the respondents indicated that they had paid between €7,90 and €23,50 for products of the brand. Finally, 14,2% of the respondents labeled the ‘heavy users’ spent an amount higher than €23,50 on the brand during the observational period.

Figure 6. Pie chart of the buying intensity of the respondents

(19)

19

displays, the age categories 12 – 19 years and 20 – 24 years were underrepresented by the dataset, with less than 2% of the respondents belonging to one of these categories. This could become an issue in further analyses, as the literature review revealed that the youngest age groups are most likely to watch videos on RTL or YouTube.

Figure 7. Percentage of respondents belonging to each age category

The majority of the respondents (46,4%) reported being in a relationship while not having any children younger than 18 years (see figure 8). This household category includes the subcategories of younger couples who might expect to have children in the future, older couples who already have children older than 18 years and older couples who do not have any children. However, in the current dataset there is not enough information available to determine who belongs to which subcategory.

Almost 30% of the respondents indicated that they were living in a 2+ person household (‘family’) with their children. Within this household category, there was information available about the age of the youngest child in the family. Finally, 23,6% of the respondents indicated being single without having any children. Again, this household

category includes the subcategories of both younger singles and older singles who might have children older than 18 years.

Figure 8. The percentage of respondents

(20)

20

In terms of education (see figure 9), only a minor part of the respondents (1,3%) reported primary school as their highest completed degree. A quarter of the respondents indicated to have completed LBO/MAVO and 4,9% reported having finished HAVO or VWO. Approximately one third of the respondents finished an education at MBO. More than 37% of the respondents received either a bachelor’s degree or a master’s degree at HBO or WO. The distribution of education in the dataset can be viewed as more or less representative for the level of education of the Dutch population (Central Bureau of Statistics, 2013).

Model specification

When looking at the purchases that individual customers make after coming into contact with push advertising, it is important to take into account possible wear-in and wear-out effects. The term wear-in describes the fact that advertising may not immediately affect purchase behaviour after the campaign is launched, but instead takes a longer time to start affecting the purchase behaviour of consumers. In contrast to this popular belief, Tellis (2009) concluded after an extensive literature review that wear-in takes place relatively quickly, especially when consumers are ‘forced’ to come into contact with the advertisements. To be more precise, Wiesel et al. (2011) studied the number of days in which different types of marketing activities (e.g. flyers, banners) affected the purchase behaviour of consumers. They found that wear-in takes place very shortly after launching the campaign, with a couple of days being the maximum wear-in time and the exact duration varying with the type of marketing communication. Their conclusions were taken into account in this study by assuming that consumers have the opportunity to change their purchase behaviour from the moment they come into contact with in-stream commercials.

On the other hand, the term advertising adstock (Broadbent, 1979) describes the lagged effect of advertising on customer behaviour, suggesting that the effects of advertising on customer behaviour can be visible for a long time. In addition, Tellis (2009) stated that the wear-out of advertising campaigns occurs slowly, especially when the delivered message is difficult to decode due to ambiguity, complexity or emotional Figure 9. The percentage of respondents

(21)

21

context. Wiesel et al. (2011) found that 90% of the effect of marketing communications was reached within seven days, with complete wear-out taking place within a maximum of nine days. Instead of including an unlimited amount of lagged effects from previous weeks in the model, it is therefore assumed that in-stream commercials are only affecting the purchase behaviour of consumers within that week (effects of

advertising in week t) and to a minor extent in the week thereafter (effects of advertising in week t-1). Based on these findings, a model accounting for the effect of push advertising variables on the purchase behaviour of customers can be described (see data collection for the explanation of each variable):

𝑺_𝒊𝒕 = 𝛽0 + 𝛽1(𝑇_𝑖𝑡) + 𝛽2(𝑇_𝑖𝑡−1) + 𝛽3(𝑅_𝑖𝑡) + 𝛽4(𝑅_𝑖𝑡−1) + 𝛽5(𝑌_𝑖𝑡) + 𝛽6(𝑌_𝑖𝑡−1) + 𝜀𝑖𝑡 Ordinary Least Squares

An Ordinary Least Squares (OLS) regression can be used to estimate the effect of the push advertising variables on the purchase value of customers. However, an issue with performing an OLS is that a loop of causality between the independent variables and the dependent variable can be identified in the model (Wooldridge, 2009). That is, the three push advertising variables not only influence the amount of sales, but the amount of sales also influences the levels of the independent variables. That is, if someone expects to purchase a certain product in the next few days, they could be more interested in viewing commercials about this product in that same week. The assumption is made that customers are not planning their purchases more than a few days in advance, as this study specifically focuses on a FMCG setting.

When applying an OLS, this loop of causality increases the amount of unexplained variance caught by the three endogenous variables, resulting in inflated estimations of the regression coefficients (Petrin & Train, 2010). Therefore, a two-step estimator Control Function Approach is applied to control for the endogeneity in the model. An assumption for this type of analysis is that the instrumental variables in the model are over-identified compared to the endogenous variables in the model (Yamano, 2010). It is possible to identify three endogenous variables in the current model, which means that a minimum of four instrumental variables have to be found in order to apply the two-step estimator.

Another requirement for the instrumental variables is that they are correlated to the endogenous variables, but not to the dependent variable in the model (i.e. the amount of sales) in order to separate the part of the endogenous variable that is correlated to the error term. The instrumental variables are created manually by using the lagged versions (t-2 and t-3) of each endogenous variable in the model. Again, the assumption is made that the effect of advertising and price promotions disappear relatively quickly and are not related to the sales after a period of longer than a week .

First, in order to check whether endogeneity is indeed an issue, each of the endogenous variables is regressed on the eight instrument variables in the model, following this example:

(22)

22

The residuals(𝛔̂ , 𝒊𝑻 𝛔̂ 𝒊𝑹and 𝛔̂ ) 𝒊𝒀 are computed for each regression line. After that, they are included as

additional regressors in the original OLS model to check whether they are significantly different from zero, indicating that the corresponding push advertising variables are endogenous (Wooldridge, 2009). If they are significant, the estimates of the regression coefficients should no longer be affected by endogeneity when performing an OLS on the model that includes the residuals as additional regressors:

𝑺𝒊𝒕 = 𝛽0 + 𝛽1(𝑇𝑖𝑡) + 𝛽2(𝑇𝑖𝑡−1) + 𝛽3(𝑅𝑖𝑡) + 𝛽4(𝑅𝑖𝑡−1) + 𝛽5(𝑌𝑖𝑡) + 𝛽6(𝑌𝑖𝑡−1) + 𝛽7(̂σ𝑖𝑇) + 𝛽8(̂ + 𝛽9(σ_𝑖𝑅) ̂ + σ_𝑖𝑌) 𝜀_𝑖𝑡

Latent Profile Analysis

After solving the potential endogeneity issue in the model and computing the OLS, a second technique is applied to divide consumers into a number of homogeneous segments based on their reaction to the push advertising variables. Latent Profile Analysis uses a set of independent variables to determine the probability that a respondent belongs to a certain cluster (“latent class”). In this study, the observed score (X) of a respondent on a certain push advertising variable (u) from the vector of six push advertising variables ‘P’ (specified in the previous section) is used to determine the probability (η) that a respondent belongs to any cluster (j) from a number of ‘K’ clusters:

𝑓(𝑋𝑢) = ∑ 𝜂𝑗 𝐾 𝑗=1 ∏ 1 √2𝜋𝜎 _𝑖𝑗2 𝑃 𝑢=1 exp (−(𝑥𝑖− 𝜇𝑖𝑗) 2 𝜎 _𝑖𝑗2 )

We are interested in finding out the values of 𝜼_𝒋 for each individual respondent, as this determines to which

cluster the respondent is assigned during the analysis. It is possible to estimate the probability that a customer belongs to a certain segment because of the assumption that the clusters can each be described by a normal distribution (Templin, 2006). To compute this, we use the mean for push advertising variable u for respondents from cluster j (given by 𝝁_𝒊𝒋) and the variance for push advertising variable u for respondents from cluster j (given by 𝝈_𝒊𝒋). Each respondent is placed in the cluster that results in the highest probability, with∑𝐾_𝑗=1𝜂𝑗 always adding up to zero.

The new regression formula taking into account the different segments can then be described as follows:

𝑺_𝒊𝑲𝒕 = 𝛽1_𝐾(𝑇_𝑖𝑡) + 𝛽2_𝐾(𝑇_𝑖𝑡−1) + 𝛽3_𝐾(𝑅_𝑖𝑡) + 𝛽4_𝐾(𝑅_𝑖𝑡−1) + 𝛽5_𝐾(𝑌_𝑖𝑡) + 𝛽6_𝐾(𝑌_𝑖𝑡−1) + 𝛽7_𝐾(̂σ𝑖𝑇) + 𝛽8_𝐾(̂ + 𝛽9σ𝑖𝑅) 𝐾(̂ + σ𝑖𝑌) 𝜀𝑖𝐾𝑡

(23)

23

similarity between them and to place observations in clusters together (Vermunt & Magdison, 2002). Moreover, it is relatively easy to incorporate different scale types in the analysis. This is particularly valuable in the current study as a combination of numeric variables and nominal variables are used to describe the purchase behaviour of customers. In contrast to when performing a traditional regression analysis, in a Latent Profile Analysis the data do not have to meet any criteria related to the distribution (Magidson & Vermunt, 2002). Consequently, any lack of support for the assumptions related to the distribution does not change the suitability of this type of analysis.

(24)

24 RESULTS

This chapter is divided into two parts, with the first part describing the results of the Ordinary Least Squares combined with the two-step estimator Control Function Approach which computes each regression

coefficient one time for the entire group of respondents. The second part describes the outcomes of the Latent Profile Analysis, which computes each regression coefficient one time for each segment.

Ordinary Least Squares

Endogeneity

In the first step of this analysis, a check for endogeneity was performed using the Wu-Hausman Test. The results of this test were not significant with F(3, 100633)=0,66 and p=0,58, indicating that the three push advertising variables are not endogenous. To confirm this finding, the residuals𝛔̂ , _𝒊𝑻 𝛔̂ _𝒊𝑹and 𝛔̂_𝒊𝒀 were included in the model. None of these residuals were significant (respectively t=0,77 and p=0,437; t=0,07 and p=0,944; t=0,395 and p=0,693), which confirms that a regular OLS model can be applied to explain any variance in purchase value.

Model estimation

In the second step, the model without any corrections for endogeneity was estimated. This OLS model was significant with F(6, 100638)=3,54 and p<0,01. However, the R2_{of the model was extremely low (<0,01)}

which indicates that the model only explains a small part of the variance in the purchase value. When looking at the individual parameters (see table 8), it can be observed that the variables YouTube contact (t) and TV contact (t-1) are significant, with respectively t=2,41 (p= 0,016) and t=2,84 (p<0,01). In other words, the amount of YouTube commercials that a customer is exposed to in week t increased the amount of euros they spent on products from the brand in that week. Also, the amount of TV commercials that the customer viewed in the previous week increased their purchase value in week t. However, the value of the regression coefficients reveals that the effect of the variables is negligible, with values close to zero. The other independent variables are not significant, which indicates that they do not explain a significant amount of variance in the purchase value of respondents.

Variable Coefficient Std. Error T-value P-value

Constant 0,38 0,01 50,42 <0,01** TV contact (t) 0,01 0,01 1,40 0,16 TV contact (t-1) 0,02 0,01 -0,32 <0,01** RTL contact (t) -0,03 0,10 2,41 0,75 RTL contact (t-1) 0,09 0,10 2,84 0,34 YouTube contact (t) 0,16 0,07 0,96 0,02** YouTube contact (t-1) 0,05 0,07 0,70 0,48

Table 8. The estimates of the regression coefficients and their p-values

Goodness of fit

As indicated before, the R2_{of the model is extremely low. This means that the dataset does not adequately}

(25)

25

not every assumption for an OLS is met in the dataset.

For instance, the plot of the residuals versus the fitted values shows that the relationship between the independent and dependent variables in the model does not follow a linear pattern (see figure 10a). Instead, the relationship appears to follow the pattern of a log distribution with a large inflation of zero’s because of the large number of households who did not purchase anything. However, when performing the OLS model again after log-transforming the purchase values, the adjusted R2_{did not become higher and the p-values of}

the individual parameters only slightly decreased (see figure 10b).

Secondly, the Kolmogorov-Smirnoff Test indicated that the pattern of the residuals deviated from a normal distribution with D=0,495 and p<0,001. As can be observed in figure 11a the histogram of purchase value is skewed to the right, describing the large number of zero’s in the dataset. When only looking at the purchase values > 0, the pattern did not resemble a normal distribution either (see figure 11b). According to the Central Limit Theorem this violation of the normality assumption (see figure 12a and figure 12b) should not be an issue as the dataset is larger than a few hundreds of observations (Ghasemi & Zahediasl, 2012).

Figure 10a. Plot of the residuals versus the fitted values with

purchase value as dependent variable.

Figure 10b. Plot of the residuals versus the fitted values with

the log transformation of purchase value as dependent variable.

Figure 11a. Histogram of purchase value when

including all observations

Figure 11b. Histogram of purchase value when only

(26)

26

Figure 13. Plot of the average value of the residuals per week The multicollinearity assumption describes that the independent

variables in the model should not be highly correlated with each other. A correlation matrix did not reveal any correlations higher than 0,25 between the independent variables. The VIF scores were computed to confirm that there were no unusual high correlations (Leeflang, Wieringa, Bijmolt & Pauwels, 2015). They were each < 4 which indicates that there are no extreme cases of multicollinearity in the dataset (see table 9). A plot of the residuals over time revealed a level of dependency between the residuals over time (see figure

13). The plot shows that a low value is more often

followed by a high value in the week thereafter, whereas a high value is more often followed by a low value in the week thereafter. In order to test for autocorrelation, the Durbin-Watson Test was performed. However, the DW-value of 1,54 indicated that there is no high level of autocorrelation in the dataset, being relatively close to the “no autocorrelation” indicator of 2 and relatively far away from the “negative autocorrelation” indicator of 4.

Table 9. VIF-scores of the independent variables

Variable VIF-score TV contact (t) 1,07 TV contact (t-1) 1,07 RTL contact (t) 1,01 RTL contact (t-1) 1,01 YouTube contact (t) 1,05 YouTube contact (t-1) 1,05

Figure 12b. QQ-plot of purchase value when only including

observations higher than zero

Figure 12a. QQ-plot of purchase value when including all

(27)

27 Latent Profile Analysis

In the Latent Profile Analysis a number of homogeneous segments were determined based on the reaction of respondents to the push advertising variables. In the previous section, it was reported that the push advertising did not result in endogeneity in the model. Therefore, in this analysis the push advertising variables did not have to be corrected for endogeneity either.

Number of clusters

Seven models were initially identified, containing one to seven clusters to divide the respondents into homogeneous groups. While the first cluster remained approximately the same in size when adding new clusters (91,6%), the size of the second cluster decreased with every new cluster that was being added to the model (see table 10). The Log-Likelihood (LL) values and BIC values of each of the models are presented in figure 14.

This scree plot reveals that a model going from one to two clusters results in the highest increase in LL-value. After that, The LL-value only increases slightly when adding more clusters, suggesting that the fit of the model with the data does not get significantly better when adding more clusters. The BIC scores quickly decrease between one and two clusters too, after which they only decrease slightly when adding more clusters. The AIC3 and CAIC scores follow a similar pattern. Thus, in terms of LL and the other information criteria, a model with two clusters would be preferred.

Figure 14. Scree plot

containing the Log-Likelihood values (LL) and BIC, with the x-as indicating the number of clusters included in the model

Table 10. Proportion of

respondents in each cluster when comparing models with two to five clusters

Number of clusters Two Three Four Five

(28)

28

However, when looking at the scree plot of R2_{it can be observed that a major increase in R}2_{takes place}

between two and three clusters (see figure 15), indicating that a model with two clusters explains

significantly less variance of the purchase value than a model containing three clusters. A Log-Likelihood Test confirmed that the model with three clusters fitted significantly better (X2₌_1.082.864_{; p<0,01) than a}

model with only two clusters. Taking together the scores of the LL-values, information criteria, R2_values

and class sizes, together with the level of interpretability resulted in the final decision to use three clusters.

Model estimation

In the model with three clusters, the majority of the respondents (91,6%) were placed in the first cluster, which had a mean purchase value of €0,00 and R2_{value smaller than 0,001. The second cluster was smaller}

(6,9% of respondents) and had a slightly higher purchase value of €3,37 and R2_{value of 0,043. While the}

third cluster only captures a minor proportion of the respondents (1,5%), it is also the cluster with the highest purchase value of €11,43 and R2_{of 0,013.}

When looking at the estimates of the individual parameters (see table 11), it becomes clear that cluster 3 has the highest intercept, indicating that when these respondents are not exposed to any form of push

advertising, they still choose to buy products from the brand. In contrast, when respondents from cluster 1 and cluster 2 are not exposed to push advertising, their purchase value is close to zero. However, while the purchase value of the respondents of cluster 2 is affected by the push advertising variables (revealed by the relatively high percentage of explained variance and number of significant parameters), this is not the case for respondents from cluster 1.

The variables TV contact (t) and TV contact (t-1) are close to zero in the three clusters. This indicates that the purchase value of the respondents is not significantly affected by these variables, which is confirmed by the p-values of respectively 0,58 and 0,20. Interestingly, TV contact (t-1) was significant in the null model, which does not divide the respondents into clusters. The variable was also significant in the regression model specified in the previous section.

The other parameters in the model do significantly affect the purchase value of the respondents in the three clusters. The reaction to the variable RTL contact was described by the same pattern for all three clusters (Wald= 2,24; p= 0,69). When respondents were exposed to zero or one commercial(s), this resulted in a

0 0,2 0,4 0,6 0,8 1 1 2 3 4 5 6 7 R²

Figure 15. Scree plot displaying the R2

(29)

29

positive effect on purchase value, while seeing two commercials affected the purchase value of respondents in a negative way.

In contrast, the variable YouTube contact has a significant effect on the purchase value of respondents, but the effect is different between the three clusters (Wald= 178,47; p< 0,001). The post-hoc analysis (see table

12) revealed that cluster 1 and cluster 2 were significantly different in their reaction to the exposure to

YouTube commercials. In figure 16, the individual coefficients of the three clusters are plotted. It can be observed that the exposure to zero or only one

YouTube commercial(s) has a slightly negative effect on the purchase value in cluster 1, while being exposed to two commercials results in a positive effect. In contrast, seeing zero or one commercial(s) has a slightly positive effect on the purchase value in cluster 2 and cluster 3, whereas being exposed to two commercials affects the purchase value negatively.

Table 11.

Regression coefficients of the intercept and push advertising variables with the bold values indicating that the purchase value is significantly affected by the specific parameter. The nominal variables are displayed using effect coding. * = significant < 0,05 ** = significant < 0,01 *** = significant <0,001

Cluster 1 Cluster 2 Cluster 3 Wald P-value

Intercept (B0) -0,07 -0,63 2,91 3,23 0,36 TV contact (B1) 0,00 -0,02 0,36 1,94 0,58 RTL contact (B2) zero commercials one commercial two commercials 1,12 1,12 -2,25 1,07 0,95 -2,02 1,79 7,15 -8,94 26,76 <0,001*** YouTube contact (B3) 21615755,68 <0,001*** zero commercials one commercial two commercials -0,76 -0,76 1,52 1,08 0,82 -1,91 0,74 3,17 -3,91 TV contact t-1 (B4) 0,00 0,08 0,01 4,65 0,20 RTL contact t-1 (B5) zero commercials one commercial two commercials -0,73 -0,73 1,46 0,85 1,05 -1,90 4,69 3,18 -7,87 10972048,61 <0,001*** YouTube contact t-1 (B6) zero commercials one commercial two commercials 0,44 0,44 -0,88 0,98 0,93 -1,91 1,08 3,12 -4,20 85,03 <0,001*** -5 -4 -3 -2 -1 0 1 2 3 4

zero one two

Cluster 1 Cluster 2 Cluster 3

Figure 16. The reaction of the three clusters when being exposed to

(30)

30

The variable RTL contact (t-1) also has a significant effect on the purchase value of respondents (p< 0,001). As can be observed in the post-hoc analysis, the reactions of the three clusters do not follow the same pattern, with cluster 1 and cluster 2 in particular experiencing a different reaction to the commercials (p = 0,03). Specifically, cluster 1 reacts slightly negative when coming into contact with zero or one commercial(s) and has a slightly positive reaction when being exposed to two commercials (see figure 17). On the other hand, respondents from

cluster 2 react slightly positive to being exposed to zero or only one commercial(s) and experience a slightly negative reaction when having been exposed to two commercials on RTL at time t-1.

Finally, the variable YouTube contact (t-1) has a significant effect on the purchase value of respondents. The three clusters reacted in a similar way to being exposed to zero, one or two YouTube commercials in the previous week. When they were exposed to zero or only one commercial(s), there was a positive effect on the purchase value of the respondents. However, when the respondents came into contact with two commercials in the week before, this negatively affected the purchase value.

Demographic profiles

In the previous part the individual parameters of the model were estimated, together with their effect on the purchase value of respondents in the three clusters. In this section, the demographic profiles of the three clusters are described in terms of buying intensity, the age of the housewife, the level of education and the household cycle categories.

Wald P-value Intercept (B0) 0,78 0,68 TV contact (B1) 1,94 0,38 RTL contact (B2) 2,24 0,69 YouTube contact (B3) cluster 1 vs cluster 2 cluster 1 vs cluster 3 cluster 2 vs cluster 3 178,47 175,90 0,78 1,53 <0,001*** <0,001*** 0,68 0,47 TV contact t-1 (B4) 4,65 0,1 RTL contact t-1 (B5) cluster 1 vs cluster 2 cluster 1 vs cluster 3 cluster 2 vs cluster 3 73,52 34,38 0,77 1,53 <0,001*** 0,03* 0,68 0,47 YouTube contact t-1 (B6) 1,15 0,89

Table 12. Post-hoc analysis with the results of the paired

comparison test if the difference between the clusters was significant. -10 -8 -6 -4 -2 0 2 4 6

zero one two

Cluster 1 Cluster 2 Cluster 3

Figure 17. The reaction of the three clusters when being exposed to

(31)

31

The buying intensity of the three clusters is displayed in figure 18. This histogram reveals that cluster 1 has a large proportion of respondents in the “never” and “light” group, which means that the customers in this cluster are not regular buyers of the brand. On the other hand, cluster 2 consists of respondents from all categories, but mostly from the “medium” and “heavy” group. The buying intensity of this group can therefore be described as relatively high. Finally, cluster 3 has an extremely large proportion of respondents in the “heavy” group and almost no respondents in the “never” and “light” category. The respondents from this cluster buy products from the brand on a regular basis and are highly profitable for the organisation.

Figure 18. Histogram of the proportion of the buying intensity in the three clusters

The age of the housewife of the three clusters is relatively similar (see figure 19), with the first cluster consisting of a higher proportion of respondents from a higher age category (55- 75 years) than the other two clusters. Both cluster 2 and cluster 3 have a slightly higher proportion of younger respondents than cluster 1. However, in general there are no large differences between the three clusters in terms of age.

Figure 19. Histogram of the proportion of age of the housewife in the three clusters

0 0,1 0,2 0,3 0,4 0,5 0,6 0,7

Never Light Medium Heavy

(32)

32

The respondents of the three clusters are also very similar in their level of education, as can be observed in

figure 20. The bars in the histogram are equally sized for cluster 1, cluster 2 and cluster 3 for each of the six

education options, indicating that there are no large differences between the clusters in terms of education.

Figure 20. Histogram of the proportion of respondents of the three clusters in each level of education

Finally, the proportion respondents in the different household cycles is approximately the same for the three identified clusters (see figure 21). The lines in this plot more or less follow the same pattern, which means that there are no large differences between the clusters in terms of the household category either.

Figure 21. Plot of the proportion of respondents of the three clusters in each household cycle

In conclusion, the largest differences between the demographics of the three clusters are based on the buying intensity of the respondents. The other demographic variables follow the same pattern across the different clusters and can therefore not be linked to specific buying behaviours of the respondent.

0 0,2 0,4 0,6 0,8 1 1,2 Primary school

LBO/MAvO MBO HAVO/VWO HBO/WO

(33)

33 SUMMARY & DISCUSSION

In this chapter, the results of the Ordinary Least Squares and Latent Profile Analysis are summarised and further interpreted, after which recommendations are made to individuals who are interested in exploring the use of push advertising strategies. Then, the limitations of the current study are described and suggestions for further research on the topic of push advertising commercials are presented.

Key findings

The main purpose of this study was to find out whether push advertising commercials have an effect on the buying behaviour of different segments of customers in a FMCG setting. Using observations from a panel consisting of 7.742 households, it was possible to estimate the effect of three types of push advertising commercials on the amount of euros spent by the respondents.

Before placing the respondents into clusters of individuals who have a similar reaction to certain push advertising activities, an OLS with all respondents together was estimated. This analysis revealed that when individuals had seen a television commercial of the brand in the week before, this slightly increased the amount of euros they spent on the brand in the week thereafter. Similarly, when respondents were exposed to a YouTube commercial in a certain week, the amount of euros spent on products from the brand in that same week increased slightly. However, the effect of the two variables on the purchase value of customers was negligible and the model only explained less than 0,1% of the variance in purchase value. The other push advertising variables did not explain a significant part of the purchase value of customers. As the assumptions of the distribution (e.g. normality, linearity) were not met, it was predicted that another type of analysis would be more suitable for describing the effect of the push advertising variables on the purchase value of customers.

The Latent Profile Analysis was performed to check whether dividing the respondents into three

homogeneous segments would increase the fit of the model to the data. An advantage of this analysis is that no assumptions are made about the distribution of the dataset. Indeed, when dividing the respondents into three segments, the fit of the model increased significantly. In the model with all respondents together (see previous section), the explained variance of purchase value was less than 0,1% whereas in the model with three segments this increased to approximately 5,6% explained variance of the purchase value.

CUSTOMER RESPONSE TO PUSH ADVERTISING COMMERCIALS: A LATENT PROFILE ANALYSIS

Master Thesis