To skip or not to skip:
The influence of ad characteristics on online video skipping
Master thesis, MSc, Marketing Intelligence
University of Groningen, Faculty of Economics and Business
June 22, 2020 CAROLINE KICHLER Studentnumber: 4088646 Nassaulaan 57A 9717 CH Groningen tel.: +43 (0)660-3738313 e-mail: c.kichler@rug.nl Supervisor
Prof. Dr. T.H.A. Bijmolt
Second Supervisor Prof. Dr. L.M. Sloot
Acknowledgements:
I would like to thank my supervisor Prof. Dr. T.H.A. Bijmolt for his support, especially in these current times. Moreover, I express my gratitude to DVJ Insights for providing the data used in this study with a special thanks to Mark. Last but not least, I thank my family, friends (and dog) for supporting me.
Abstract
Table of Content
I. List of Tables ... 1
II. List of Figures ... 2
1 INTRODUCTION ... 3
1.1 Online Advertisement Avoidance Behavior ... 4
1.2 Research Question ... 5
1.3 Relevance of the study ... 5
1.4 Structure ... 6
2 THEORETICAL BACKGROUND ... 7
2.1 Ad Avoidance ... 7
2.1.1 Skipping ... 8
2.1.2 Differences between offline and online ad avoidance ... 8
2.1.2.1 Attention-getting Techniques ... 9
2.1.2.2 Detect the ad’s content ... 9
2.1.2.3 Intrusiveness ... 9
2.2 Theoretical Framework ... 10
2.2.1 Celebrity ... 11
2.2.2 Real People in Real Settings ... 11
2.2.3 Slice of Life ... 12 2.2.4 Sensual Appeals ... 12 2.2.4.1 Visual Cues ... 13 2.2.4.2 Audio Cues ... 14 2.2.5 Brand Presence ... 14 2.2.5.1 Logo ... 15 2.2.5.2 Brand Mentions ... 16 2.2.6 Content of the Ad ... 16
2.3 Hypotheses Overview ... 17 3 METHODOLOGY ... 19 3.1 Research Design ... 19 3.2 Sample Description ... 19 3.3 Procedure ... 20 3.3.1 Individual Coding ... 21 3.4 Variables ... 21 3.4.1 Dependent Variables ... 21 3.4.1.1 Skipping Rate ... 22 3.4.1.2 Viewing Time ... 22 3.4.2 Independent Variables ... 22 3.4.2.1 Celebrity ... 22
3.4.2.2 Real People in Real Settings ... 22
3.4.2.3 Slice of Life ... 23
3.4.2.4 Sensual Images ... 23
3.4.2.5 Audio Cue: Music Only ... 23
3.4.2.6 Brand’s Logo ... 23
3.4.2.7 Brand Mentions ... 23
3.4.2.8 Content of the ad ... 23
3.4.2.9 Cognitive Complex Ad Characteristics... 24
3.4.3 Covariates ... 24
3.4.3.1 Respondent Individual Characteristic ... 24
3.4.3.2 Attitude toward Ads ... 25
3.4.3.3 Duration of the Ad ... 26
3.5 Missings and Outliers ... 28
3.6 Multicollinearity ... 28
3.7 Methods of Analysis ... 28
3.7.2 Hazard Model ... 29
3.7.2.1 Proportionality Assumption ... 30
4 RESULTS ... 32
4.1 Descriptive Analyses ... 32
4.2 Binary Logit Model ... 33
4.2.1 Multicollinearity ... 33
4.2.2 Model Fit ... 33
4.2.3 Findings ... 34
4.3 Hazard Model ... 36
4.3.1 Kaplan-Meier Survival Function ... 36
4.3.2 Proportionality Assumption Testing ... 38
4.3.3 Violation of the Assumption ... 39
1 INTRODUCTION
In our lives, we are constantly exposed to different forms of advertising. Whether it is a billboard that we pass on the way to work, an online advertisement before watching a workout video on YouTube or a TV commercial while waiting for the daily news. The evolution of advertisements (hereafter, ads) has grown from year to year and companies try to attract the attention of the consumer in multiple ways, resulting in an enormity of ad clutter (Chatterjee, 2008). But over the years the purpose of ads has changed. In the early days of advertising, the goal was to inform consumers; for example, about a new product innovation or specific price discounts, which were afterwards used to make informed purchase decisions (Teixeira, 2014). However, ads are aimed at eliciting positive feelings rather than providing the consumer with new information (Biswas, Olsen & Carlet, 1992). Due to the intensive use of the internet, people can easily find the required information by using a search engine and retrieving any information regarding the brand. Instead of relying on the informativeness of advertisements, it became easy for a consumer to get the requested information on the web within seconds (Teixeira, 2014). The emergence of new media channels led to an increase in the exposure of ads. New challenges arose as the more ads there are out there the more difficult it becomes to stand out of the mass and reach the intended target group. Big players like Facebook, Google and YouTube are all displaying advertised content and it has become their main source of income (Belanche, Flavián & Pérez-Rueda, 2017). The former two dominate the market, accounting for 46.4 percent of digital ad spending (Handley, 2017) while the latter is the most visited website globally, with it covering 85 percent of the top global brands during YouTube videos (Belanche, Flavián & Pérez-Rueda, 2019).
While video ads were predominantly left to the TV ad section, platforms such as Facebook, Twitter and LinkedIn are lately shifting towards video ads features as well (Dukes, Liu & Shuai. 2020). In 2018, total expenditure for video ads accounted for over $90 billion dollars (Vidyard Benchmark Report, 2019). But managers can also reap benefits from online ads. A study from 2013 showed that they score better on recall and likeability than TV ads (IAB, 2013).
Online Advertisement Avoidance Behavior
There are a number of factors that affect the perception and effectiveness of ads. Advertisements are rarely perceived well by its audience. People are neither enjoying commercial breaks during movies nor like to watch an ad before a YouTube video. As a result, ads are suffering from the perception of being annoying and intrusive (Hegner, Kusse & Pruyn, 2016). Nowadays, viewers have many options to dodge ads. Whether it is through installing internet ad blockers or switching the channel on TV when a commercial comes up (Edwards, Li & Lee, 2002).This creates a challenge for advertisers who are trying to create compelling and less intrusive ads (Aaker, 2010; Campbell, Mattison Thompson, Grimm & Robson, 2017). Despite the booming trend of online video advertising, they are also not spared from ad avoidance. This can be explained as consumers are in a goal-directed state when using the internet (Seyedghorban, Tahernejad & Matanda, 2016). Consider a scenario in which a person wants to listen to a song on YouTube. The person types in a song title and after clicking on the video an ad pops up, thereby interfering with its initial intention. Due to this goal-directed state, internet ads are perceived as more intrusive than ads in traditional media which increases the viewer’s urge to avoid the ad (Edwards et al., 2002). A feature in pre-roll ads which is able to counteract the feeling of intrusiveness is the option to skip an ad (McCoy, Everard, Polar & Galletta, 2008). In these kinds of ads viewers have the possibility to decide whether to continue watching the ad or to go directly to the content they wanted in the first place.
Measuring the skipping rate can serve as a proxy for the effectiveness of an ad. These skippable pre-roll videos must still consider the goal-directed state of the viewer and endeavor to keep the perceived intrusiveness of the ad as low as possible. Previous research found out that approaches that work with traditional advertising may lead to increased avoidance of an online video ad (Baek & Morimoto, 2012; Campbell et al., 2017).
advertised product or brand when making buying decisions (Keller, 2007; Crommelin, Gerber & Terblanche-Smit, 2014). Hence looking at the skipping rate is an important measure for advertisers and content providers.
Research Question
The resultant research question in this context is:
What are the predictors that determine online video ad skipping behavior and how do the appeals and characteristics used in TV ads influence the skipping of online video ads?
The emergence of online video ads rose questions about the effects of characteristics in these ads. Previous research even found out that the elements that usually work in a TV setting can lead to opposing effects when used in pre-roll ads (Teixeira, 2014). According to Campbell et al. (2007) especially gentler, less brash and less overt advertising - features that are not advisable in TV advertising- reduce pre-roll skipping while basic emotions like happiness increase skipping.
But what are the concrete differences that advertisers should care about and consider when creating an online ad compared to a TV ad? What appeals help to boost the viewer’s intention to view the ad; For example, does the presence of a celebrity help to make the viewer curious to see the whole ad or does it have the opposite effect and rather deter viewers from watching it?
The central focus on this work is to describe the underlying factors that determine the duration of watching an ad before it is being skipped. Therefore, we consider ad watching as an indicator of effectiveness in terms of the viewer’s interest in the ad’s content and the willingness to spend time to watch it. More precisely, this work focuses on two indicators: ad viewing time - the time the viewer deliberately watches an ad - and ad skipping - whether a viewer skips an ad or not. The aim of this study is to give an insight on which features are increasing the viewing time of skippable pre-roll video ads, by testing features and appeals that are usually used in TV ads.
Relevance of the study
and recall (Percy & Elliott, 2009). The outcome of this thesis can be useful for managers as it provides insights on which ad characteristics decrease online ad skipping and therefore are recommended for application in an online video ad setting.
From a scientific perspective, this research is relevant because it extends the research field by looking at the factors that are previously used to predict TV ad avoidance. While for traditional advertisement a lot of research has focused on this topic, there seems to be a lack of research investigating skippable pre-roll ads.
Structure
2 THEORETICAL BACKGROUND
Ad Avoidance
Researchers have studied ad avoidance since the introduction of TV commercials in the 1950s (Joo, 2014). Advertising avoidance is defined as “all actions made by media users that differentially reduce their exposure to ad content” (Speck & Elliott, 1997: 61). Nowadays, viewers have different options to dodge ads. Whether it is through installing ad blockers to avoid online ads or through changing the channel during commercial breaks in order to not be forced to listen or watch it regardless of the brand and content (Gustafson & Siddarth, 2007). Viewers have developed avoidance strategies that reduce the chance for advertisers to attract new customers through the advertising message (Heeter & Greenberg, 1985; Greene, 1988). From an advertiser perspective, an advertising form is only effective if the viewer is exposed to the advertising message whereas avoiding an ad reduces the chance of exposure (Heeter & Greenberg, 1985; Joo, 2016). Speck and Elliot (1997) identified three different strategies that people use to avoid ads; physical, cognitive and mechanical avoidance, while this thesis focuses on the latter one. Mechanical avoidance includes all acts of ad avoidance that uses technological means such as a remote control in the case of TV, a dial in the case of radio (de Gregorio, Jung & Sung, 2017) or a skip button in the case of online video ads (Joo, 2016). Mechanical avoidance is different to the other avoidance strategies because it requires an active action by the viewer (Joo, 2016). Three actions emerge from mechanical avoidance; zapping, zipping and skipping. Zapping refers to switching television channels when a commercial comes up, zipping means fast forwarding ads and skipping is the act of stopping an ad through clicking on a button (Fransen, Verlegh, Kirmani & Smit., 2015). As this study focuses on online video ads, the investigated mechanical avoidance will be skipping.
does not require a shift between states. As a result, mechanical avoidance seems to be very relevant to online video ads due to the ease of using the skip button (Joo, 2016).
2.1.1 Skipping
Online videos can appear before, during or after watching the accessed online video - so called pre-, mid- or post-roll video-ads (Teixeira, Wedel & Pieters, 2012). In skippable pre-roll ads, the viewer is entitled to close the ad with a click after a certain time. In the aforementioned TrueView in-stream ads this is possible after 5 seconds. By clicking on the skip button, the viewer can directly jump to the desired content. A conclusion of this is that only if the value of the ad is higher than the desired content, the viewer will decide to continue watching the ad (Dukes et al., 2020).
Skippable pre-roll video ads are used by advertisers due to its effectiveness in comparison to other ad formats such as non-skippable or display ads (Guttmann, 2020). For non-skippable ads people must watch the whole ad before they can continue to see their desired content. In contrast, the format of skippable ad seems to be more preferred by viewers as they have the power to decide whether to watch the ad (Belanche et al., 2017).
While in the traditional context turning away from an ad implies to avoid an ad, we have to be careful when interpreting the implications for skipping. Skipping is a less extreme form of ad avoidance and should not be equated with ad abandonment (Campbell et al., 2017). Ad abandonment in the context of skippable pre-roll video ads would be when a viewer refuses to watch the ad during the time where he/she is forced to watch the ad, hence during the first 5 seconds (Goodrich, Schiller & Galletta, 2015).
2.1.2 Differences between offline and online ad avoidance
2.1.2.1 Attention-getting Techniques
The active state of the user can be seen either as an advantage or disadvantage. Proponents say that due to the heightened attention of the consumer it is easier to attract the attention in an online ad than in traditional media (McCoy et al., 2008). On the other hand, advertisers need to consider that techniques that work in the traditional media setting might have a different effect when the viewer is already in an active state. Usually, advertisers make use of attention-getting tactics which aim to attract the viewer’s attention. Attention in the media is important because for consumers to be affected by advertising messages, they need to first pay attention to the ad (Teixeira, 2014). In traditional advertising, viewers can be easily distracted, therefore attention-getting tactics are effective to make the viewer focus their attention to the ad. However, this does not necessarily have to hold for online video advertising.
This thesis even goes one step further and proposes that attention-getting techniques increases the skipping rate of pre-roll video ads. Nowadays, people have got used to attention-getting characteristics in ads which might even backfire because viewers get more easily aware that they are watching an ad instead of the desired content. For example, while using a celebrity in an ad was once a valuable approach to attract the user’s attention, it became rather conventional these days (Campbell et al.,2017; Till & Shimp, 1998).
2.1.2.2 Detect the ad’s content
One negative consequence can be that due to certain ad characteristics it is easier for viewers to recognize that they are watching an ad instead of the requested video. Viewers may at first be unaware that an ad appears and mistake it with the content they initially wanted to watch. If an ad with attention-getting characteristics is shown in the pre-roll setting, it may increase the recognizability of ads.
This recognition can have negative consequences. Research has found that consumers are spending less time when they know they are viewing content of an advertisement leading to an avoidance of the ad (Kim, Pasadeos, & Barban; 2001; van Reijmersdal, Neijens, & Smit, 2005).
2.1.2.3 Intrusiveness
Therefore, pre-roll ads are more likely to evoke perceptions of irritation and intrusion than ads in traditional media (Li, Edwards & Lee, 2002, Campbell et al., 2017).
Theoretical Framework
Advertisements typically contain different kinds of features while some of these features will be evaluated in this thesis. A specific attribute’s presence usually aims at grabbing the viewer’s attention while these characteristics might backfire in online video ads, due to its level of intrusiveness and the ease to detect an ad.
2.2.1 Celebrity
A typical appeal that can often be seen in advertisements is the presence of celebrities in which a well-known person acts as a brand ambassador. Advertisers found that due to the attraction of celebrities, these ads are able to break through the cluttered advertising world (Atkin & Block, 1983; MacInnis, Moorman & Jaworski, 1991). A study by Kneale that was undertaken in 1988 has shown that during the Grammy music awards the average zapping rate was decreased from 10% to 2% just by featuring the singer Michael Jackson. Using celebrities as brand advocates is a promising strategy in traditional media because typically positive outcomes such as ad recall, likeability and ad recognition are expected (Campbell et al., 2017; Friedman & Friedman, 1979). On the contrary, the positive effect on ad recognition can be a major disadvantage for online video ads because the easier and faster the ad can be recognized, the more likely it will be skipped (Campbell et al., 2017).
Hence it is expected that ads with celebrities are increasing ad recognition which makes the viewer want to continue to the desired content. For this reason, we expect the following hypothesis:
Hypothesis 1. Online video ads showing celebrities increase skipping compared to ads
without this feature.
2.2.2 Real People in Real Settings
While celebrity endorsement seems to remain a common strategy for advertisers, it also entails its negative effects. Women are comparing themselves to attractive models in the ad, which eventually leads to negative effects on their self-esteem (Bower & Landreth, 2001). This should never be the intention of advertisements while at the same time there is an increasing demand for authenticity and truth in ads. Consumers not only expect basic information about the product but rather emotions and values with which they can identify themselves (Zatwarnicka-Madura & Nowacki, 2018).
While this appeal has a positive effect on the trustworthiness towards the ad, it is uncertain whether using real people in real setting affects the skipping rate. Due to the consumers likeability of authentic and forthright stories, the following hypothesis is expected:
Hypothesis 2. Online video ads that show real people in real settings decrease
skipping as compared to ads without this feature.
2.2.3 Slice of Life
Another executional cue which is frequently used in advertising is the slice-of-life appeal. This can be expressed by showing a happy situation in an advertisement which is elicited by the usage of the product (Belch & Belch, 2009). Often a real-life situation is displayed, for example a woman complaining about brittle hair or the deposition of lime in a washing machine, after which the advertised product is presented as the answer to the problem. Therefore, the goal of the advertising is to show how the use of the product or service is enhancing the situation. Advertisers are implementing this kind of cue because it has been shown to be effective in terms of brand recall and brand loyalty (Saxena Arora & Arora, 2017). Consumers who are facing a similar situation in real life, will remember the product or service as the problem solver. Advertisers also use this type to counteract prior negative attitudes with the product or service (Saxena Arora & Arora, 2017).
From a managerial perspective, this appeal is an effective application for ads, while there still is uncertainty if the proposed measure is advisable in regards to online video ads. One could argue, that due to the wide implementation of this appeal it is easier to detect the ad as an ad which is a disadvantage in the online video ad environment. Therefore, the following hypothesis will be tested:
Hypothesis 3. The slice-of-life appeal in online video ads increases skipping compared
to ads without this appeal.
2.2.4 Sensual Appeals
engages the consumers' senses and affects their perception, judgment, and behavior” (Krishna, 2012: 332).
Humans interact with the world through their senses and therefore the prevalence of sensory cues in an ad influence the viewers thoughts, feelings and subsequently purchase decisions (Krishna, 2012). While scent can be important for product use and purchase decisions, sensual attributes in commercials can only happen through sight and sound (Meyers-Levy, Bublitz & Peracchio, 2009). A study by the international research institute Millward Brown (2013) revealed that 99 percent of all brand communication focuses on these two senses. Advertisements therefore mainly appeal to our eyes and ears, with the aim to emotionally connect the viewer.
2.2.4.1 Visual Cues
Sensory images effectively appeal to consumer’s senses and evoke a memorable experience (Krishna & Schwarz, 2014), be it through showing a fresh cooked meal or the visualization of the scent of a cookie. An ad with sensual images includes sensory attributes such as stunning images, strong colors or visual imagery (Haase, Wiedmann, Bettels & Labenz, 2018; Teixeira, 2014). The focus is more on strong images and visual stimulation as compared to ads that focus on the benefits of the product or on engaging the viewer through an emotional storyline (Teixeira, 2014). Another sensory cue which is widespread in advertisements is the use of sexual images (MacInnis et al., 1991; Wirtz, Sparks & Zimbres, 2018). These are ads that show nudity or sexual gestures with the aim to attract the attention and evoke a feeling of arousal in the viewer (Wirtz et al., 2018).
On one hand, advertisers argue that using sensory stimulation generally creates liking and a more positive perception towards the ad, others believe that sensory stimulations were overused during the recent years, and hence are less effective than primarily thought (Haase et al., 2018). Researchers like Lindstrom (2005) call consumers to be visually sophisticated, knowing that what they see in the ad is not always what they get in reality. Even though, there is uncertainty whether and by which size it affects the skipping rate, the following hypothesis is assumed:
Hypothesis 4. Online video ads with sensory images increase skipping compared to ads
2.2.4.2 Audio Cues
Another sensual cue that is often used in ads is music which aims to connect the viewer through the ears. Most people enjoy listening to music because it helps bringing them in a different state of emotion. Our daily lives are accompanied by music either voluntarily or involuntarily (Sloboda, O’Neill & Ivaldi, 2001; North, Hargreaves & Hargreaves, 2004). Music is constantly appearing in ads and its influence on people has created a lot of attention (Gorn, 1982). This is because music attracts and holds the attention of the viewer and maintains its interest with different music styles (Brooker & Wheatley, 1994). Already in 1984, Hecker considered music to be “the single most stimulating component of advertising” (Hecker, 1984: 3).Music appears in ads since the beginning of radio shows (Brooker & Wheatley, 1994; Kellaris, Cox & Cox, 1993) and numerous advertisers rely more on music than on words (Dunbar, 1990).
While in the traditional media setting this is an important element for getting noticed in a cluttered ad environment, it might not be as useful in an online setting. Some users might feel disturbed by the music or feel that it is unsuitable for the current situation. For example, if a person wants to listen to a song on YouTube while suddenly a different song appears in the ad web users will most likely feel irritated resulting in a feeling of intrusiveness.
Furthermore, ads that overuse sensory cues have showed to cause feelings of irritation leaving consumers to be negatively affected by the ad (Aaker & Bruzzone, 1985). Both reasons lead to the following two assumptions:
Hypothesis 5. Online video ads in which only music is played and no speaking occurs
increase skipping compared to ads in which there is speaking.
Hypothesis 6. Online video ads with sensual images increase skipping when combined
with music and vice versa.
2.2.5 Brand Presence
can only show its positive effects if the viewer recognizes and subsequently remembers the brand in the ad (Teixeira et al. 2012; Zigmond, Dorai-Raj, Interian & Naverniouk, 2009). While brand presence has showed its dependence on how likely the viewer zaps a TV ad (Teixeira, Wedel & Pieters, 2010) the same is assumed for skipping rates of online video ads. Both verbal (brand mentions) and non-verbal (brand logo) brand presence characteristics are discussed.
2.2.5.1 Logo
The swoosh sign, the bitten-off apple or the capitalized yellow “M” are just three examples of the multiple logos that can be easily associated with the corresponding brands. Companies invest large amounts of money in the development of the logo, which represents the personality and values of the company (Bernstein, 1986) and further aims to affect consumers perception and recognition of the brand (Girard, Anitsal & Anitsal, 2013).
Logos are presented to the costumer in multiple occasions, be it on the product package, the corporate’s website or in advertisements (Bottomley & Doyle, 2006). Researchers found that the prevalence of logos are an effective tool in ads, because it captures the viewer’s attention effectively (Cian, Krishna, & Elder, 2014; Pieters & Wedel, 2004) Besides grabbing attention consumers easily recognize the brand in the ad from remembering the logo (Pham, Pallares-Venegas & Teich, 2012).
Logos can be visual at different moments in the commercial. Some commercials even show the logo throughout the whole time. When showing the logo early or more frequently, a study by Teixeira et al. (2010) affirmed that it improves recall and recognition but brand presence generally increases ad avoidance. A significant negative effect of Logos was also found by Baltas (2003) who tested the clicking rate for banner ads with the reasoning that showing the logo together with the ad decreases the interest of the viewer because it is already revealed what the ad is about. Only people who were generally interested in the advertised brand in the first place showed a higher clicking rate (Baltas, 2003). Given these findings from previous studies the following hypothesis can be postulated:
Hypothesis 7. Logos which are presented during the whole time of the online video ad
2.2.5.2 Brand Mentions
What marketers also can decide on is the frequency of the brand name mentioned in the ad. Mentioning the brand is important because it ensures that the consumers reaction to the advertising is associated with the advertised company. Eventually, consumers remember the content of ads but fail to remember the advertised brand name (Crommelin et al., 214) or even worse, consumers associate the ad message with a different brand (Baker, Honea & Russell, 2003). As opposed to other ad elements, brand names have the power to efficiently store the ad message in the memory of the viewer (Alba et al., 1991).
Another fact which might impact ad effectiveness is the repetition of the brand name which strengthens the association between the ad content and the advertised ad (Alba et al., 1991). While repetition is positively affecting subsequent responses of the consumer, the effect on the skipping rate still has to be examined because to date no research has investigated whether, and by how much, the number of brand mentions impacts advertising effectiveness in terms of the skipping rate. Crommelin et al. (2004) affirmed that the longer the brand is present in an advertisement, the easier it is for the viewer to recognize the brand. From these results, mentioning the brand is advisable. However, considering the skipping rate it is expected that the repetition of the brand name increases skipping because once the viewer is aware of the ad and the advertised brand the viewer will lose interest and wants to click it away. Hence the following hypothesis can be stipulated:
Hypothesis 8. The more often the brand is mentioned in an online video ad, the higher
the rate of skipping.
2.2.6 Content of the Ad
advertising and especially the time before our daily lives were dependent on the internet, consumers were relying on ads to provide them with informational content. However, nowadays, consumers are not requiring this need for product information anymore (Teixeira, 2014). Furthermore, given the goal-directed state, the user might already be looking for a specific information while surfing the web. If suddenly an ad appears, the user might feel distracted and irritated from the initial goal. Therefore, the following hypothesis is derived:
Hypothesis 9. Online video ads that provide entertaining rather than informative value
decrease skipping as compared to ads with informative value.
2.2.7 Cognitive Complex Ad Characteristics
Another technique can be to include cognitive complex ad characteristics. These are characteristics that trigger certain emotions which require cognitive capacity in order to process them. In total there are seven different emotions - exhilaration, warmth, fun, humor, nostalgia, relaxing, and shock – that cause cognitive load (Campbell et al., 2017). Four of the characteristics show a positive valence, while nostalgia, relaxing and shock can be either neutral, positive or negatively associated. All these characteristics are complex in nature, meaning they demand high cognitive resources (Schmidt 2002; Strick, Holland, Van Baaren &Van Knippenberg, 2009).
Watching an ad with these characteristics requires the cognitive capacity of the viewer, while at the same time the viewer cannot use its cognitive resources to experience intrusiveness. Therefore, it is more difficult for the viewer to recognize the ad as an ad (Campbell et al., 2017). Furthermore, Goodrich et al. (2015) affirmed that some of the emotions are decreasing the feeling of intrusiveness which leads to the following hypothesis:
Hypothesis 10. Online video ads that trigger cognitive complex emotions decrease
skipping as compared to ads that do not trigger those emotions.
Hypotheses Overview
3 METHODOLOGY
Research Design
The data that was used for estimating and validating the model was collected through large-scale copy tests in three different countries, namely the Netherlands, Germany and the United Kingdom. The company responsible for the data collection was the marketing research firm DVJ Insights which provided the data through an internal portal. The whole sample size consisted of 9.838 evaluations, while in total 96 distinct ads were evaluated. Each respondent watched four to five commercials, while a maximum of two of them were in a skippable format. On average, each individual ad was evaluated by around 100 respondents. The time period over which the data was collected was approximately one month. The initial data set included only ads in which the respondent was allowed to skip it after five seconds.
The ads represented a wide range of product categories, from life insurances to food and beverages. In total, 96 distinct ads appeared in the study comprising of 34 German, 30 Dutch and 32 English advertisements. An almost equal number of evaluations came from the Dutch and English dataset, 3.152 and 3.181 respectively, while 3.505 evaluations was included within the German dataset.
Sample Description
The length of the ads varied between each advert. The shortest length of a full online ad was 10 seconds while the longest lasted for 21 seconds with a standard deviation of 3.47 seconds. The mean of all ads was 16 seconds, which is slightly above the recommended minimum length of 12 seconds (Google Support Display Specs Help, 2020). 26% of the 96 videos were around 15 seconds long, while the majority (37 ads) had a length of 20 seconds (figure 4). The researcher from DVJ Insights had already excluded ads that lasted less than 6 seconds.
Procedure
participant was forwarded to a news article or to a YouTube video when suddenly an ad appeared at the place where he/she thought to see the requested content. This is a major benefit of this test since pre-roll video ads are extremely common in the online world and respondents were using their own electronic device which further contributed to create a real setting. The ads were presented in random order and each participant watched several advertisements. Depending on the country where the study appeared the respective national language was used in the ad, hence in this case either Dutch, English or German. After looking at the ads, participants were asked questions about their general attitude towards ads. Subsequently, other metrics like brand impact and associations were measured. Due to their irrelevance to this study, this data was not made available. The provided data file was in the SPSS (Statistical Package for Social Sciences) format.
3.3.1 Individual Coding
To answer the research question and test the hypotheses, more metrics than the ones asked in the questionnaire needed to be included. Before the disclosure of the SPSS files, three excel files were sent out by DVJ Insights with a clickable link directly to the ad video itself. Each ad was watched multiple times and coded accordingly to the features explained in the variables part. As an example, the celebrity variable was coded as 1 when a celebrity appeared in the ad and 0 when no celebrity appeared. Due to the Dutch language barrier, subtitles were used to translate the content of the ad (if available). After coding all 96 ads, the additional variables were included as dummies in the matching SPSS file.
Variables
In total, nine independent variables and five covariates were selected to parameterize the model. A summary of the variables is given in table 3.
3.4.1 Dependent Variables
3.4.1.1 Skipping Rate
For each evaluation per respondent, the data indicates whether the viewer skipped the ad or not. The variable is binary, with two levels. A zero indicates that the ad was watched for its whole length, and a 1 if the ad was skipped. For example, a 15-second commercial that was watched for 13 seconds would be qualified as being skipped and hence indicated with the level 1. Looking at the determinants that affect the skipping rate serves as a first guidance for the subsequent analysis due to the assumption that an increase in the skipping probability, leads to a decrease in the viewing time. Skipping is an appropriate proxy for the viewer’s motivation to watch an ad (Campbell et al., 2017). Hence, explanatory variables that have a specific effect on the skipping probability are expected to have the same effect when accounting for the time up until an ad was skipped (Siddarth & Chattopadhyay, 1998).
3.4.1.2 Viewing Time
While the first dependent variable did not distinguish whether it was skipped at second 8 or at second 12, the second dependent variable signals the time until an ad was skipped or, in the case where no skipping occurred, the seconds of the full length of the ad. The viewing time for each evaluation represents the time elapsed until the ad is being skipped, or watched to completion. The initial value was subtracted by five seconds, in order to account for the fact that respondents were not able to skip an ad in the first five seconds.
3.4.2 Independent Variables
The executional cues that were used to measure advertising effectiveness were based on theory and consisted of nine different variables: celebrity, real people in real settings, slice of life, sensual images and audio cue, a brand’s logo, the number of brand mentions, the value of the ad and the number of cognitive complex characteristics.
3.4.2.1 Celebrity
An ad can appear with different kinds of endorsers, whether it is an expert, a regular consumer or celebrity endorser. The variable indicates a 1 if a famous person (related to the relevant country) appeared in the ad.
3.4.2.2 Real People in Real Settings
customers using their products and talking about their product experience. Whenever the ad made use of this appeal the variable indicates a 1.
3.4.2.3 Slice of Life
The slice of life variable usually implemented in ads for home care products, however this can also be seen in other categories. Whenever the ad showed a slice of life situation, the variable was coded as 1.
3.4.2.4 Sensual Images
Ads can use visual cues in form of strong images, visual imaginary and other visual stimulations. The variable indicates a 1 whenever an ad used sensual images.
3.4.2.5 Audio Cue: Music Only
The variable MusicOnly indicates whether the ad only used music in the foreground, as compared to talking in the ad. The model also included the interaction effect of an audio and visual cue. Although the investigation of the interaction is not the main object of the study, it is believed that it might provide additionally valuable insights.
3.4.2.6 Brand’s Logo
Some advertisers choose to show the logo of the brand continuously in their ad in order to burn-in the brand burn-in the mburn-inds of the customers (Keller, 2007; Teixeira et al., 2012). Whenever the ad showed the logo throughout the whole duration of the ad, the variable indicates a 1, whereas a 0 indicates when a logo was briefly shown or not at all.
3.4.2.7 Brand Mentions
This variable indicates the number of times the brand was mentioned during the ad, regardless if it was through a voice over or an on-camera character. The highest number of brand mentions was 5 times, while some of the ads did not mention the brand name once.
3.4.2.8 Content of the ad
3.4.2.9 Cognitive Complex Ad Characteristics
When investigating the seven emotions that are considered as cognitive complex, each emotion was coded individually. After looking at the emotions separately, a sum of all emotions was calculated. Hence, if an ad showed emotions of relaxation and nostalgia, the variable indicates a 2. The reason for creating a numerical variable is that the underlying assumption is the more cognitive complex emotions the ad elicits, the more difficult it is to focus on something else, such as skipping. The following emotions were investigated:
• Exhilaration: ads that created a thrill (Campbell et al., 2017)
• Fun/Entertaining: amusing, pleasurable ads with a feel-good atmosphere (Ducoffe, 1995; Edwards et al., 2002)
• Humor: ads with a punchline or inherent irony (Goel & Dolan, 2001)
• Nostalgia: ads reflecting on the past (Muehling & Sprott, 2004), which can be bittersweet and often creates conflicting emotions such as happiness and sadness (Davis, 1979; Holak & Havlena, 1998)
• Relaxing: ads that highlight safety or serenity (Campbell et al., 2017) • Shock: ads with a surprising storyline (Campbell et al., 2017)
• Warmth: an ad that elicits “a positive, mild, volatile emotion involving physiological arousal and precipitated by experiencing directly or vicariously a love, family, or friendship relationship” (Aaker, Stayman & Hagerty, 1986: 366)
3.4.3 Covariates
Besides the independent variables mentioned in the section above, researchers found additional factors that seem to affect the skipping rate. On the basis of prior research, the following variables will be additionally controlled: gender, gender consistency with the target audience, age, the respondent’s attitude towards ads and the length of the ad.
3.4.3.1 Respondent Individual Characteristic
The included customer characteristics are of interest due to their effect on skipping behavior based on previous research. Findings for ad avoidance for traditional media concluded that women are more likely to avoid ads because they are more active during commercial breaks as compared to men who kept looking at the TV screen (Manrai & Manrai, 1995).
On the other hand, Belanche et al.’s (2015) study revealed that women are watching skippable ads for a longer time than men. While other researchers (Siddarth & Chattopadhyay, 1998; de Gregorio et al., 2017) report no significant differences in ad avoidance between men or women. Considering the world today and the personnel connections that people form with their electronic devices, gender might not play as a significant role as it used to be in advertising literature (Joo, 2016).
Moreover, the consistency between the proposed target audience of the ad and the gender of the respondent is considered, with 0 indicating that the gender did not match and a 1 when the gender fit with the target audience. Showing a commercial aimed at the right target group was found to reduce switching within TV commercial breaks (van Meurs, 1999). Whenever the target audience was not clear or the ad was suitable for both male and females, the variable indicates a 1 as well.
Besides gender, the age of the respondent might also have a significant impact on the skipping probability (Speck & Elliot, 1997). Research on both zapping and skipping data affirmed that younger people tend to avoid ads more than older people, or in other words, watch a lower duration of the ads than older people (Cronin & Menelly, 1992; Danaher, 1995; Joo, 2016). Belanche et al. (2017) confirmed that older viewers tend to watch online video ads longer than younger viewers, resulting in a longer viewing time for older people.
3.4.3.2 Attitude toward Ads
analysis was performed. First, a principal component analysis was conducted, resulting in two dimensions with eigenvalues above 1.0. Using one factor explains 46.36% of the variance while using two increased the explained variance to 71.73%. However, because the second dimension only contained one item, it will be excluded from the study. After proceeding with one factor which contained the items ad avoidance, ad quality and ad likeability, requirements in terms of KMO scores (0.664>0.6) and the Bartlett’s test of sphericity (α = 0.05) were satisfied. Furthermore, reliabilities were tested within the factor. The reliability index (Cronbach’s alpha coefficient) was higher than the minimum acceptable level (0.671 > 0.6) and deleting an item would have decreased the coefficient (Malhotra, 2010).
3.4.3.3 Duration of the Ad
Investigating the skipping probability, one must consider the length of the ad. Usually there is a time screen indicating the seconds that are still left until the video is finished or can be skipped (Hegner et al., 2016). As this survey only focused on skippable ads, the length of an ad might not play as an important role because people could simply click away the ad after five seconds. However, duration is still important, because the longer an ad, the longer the time span in which a person is able skip. This would lead to the assumption that a 14 second ad is about two times more likely to be skipped than an ad with a 7 second ad format (Gustafson & Siddarth, 2007). Due to the reason that viewers were only allowed to skip once the first five seconds elapsed the variable was adjusted for five seconds.
TABLE 3
Missings and Outliers
Before proceeding with the data analysis, missings were detected and cleaned from the dataset. Ads that lasted for less than 6 seconds as well as ads that could not be skipped during its whole length wear already excluded by the researchers. Due to the thorough data cleaning by DVJ Insights, minimal data-collection errors were found. Out of the 9,899 cases, 133 values were missing for three variables. Any missing responses were treated by casewise deletion, where each response with a missing value is discarded from the analysis. This resulted in 9,854 remaining cases. Due to the very low percentage of missings and a still sufficiently large data sample no bias to the results is expected with casewise deletion (Malhotra, 2010).
Furthermore, 16 outliers were detected. The boxplot of all observations that were skipped, revealed that there were a few observations in which respondents skipped even before the 5 second mark (see appendix I). Due to the reason that it was not possible to skip within the first 5 seconds, it is assumed that respondents abruptly exited the questionnaire. Therefore, these evaluations were removed from the data set. All of the outliers were found within the Dutch data and it affected 13 distinct advertisements.
Multicollinearity
Before testing the hypotheses and research question, the Pearson correlations were calculated to show the relationship among variables. First a correlation matrix was built which can be found in detail in the appendix (see appendix I). Four variables were detected with correlations above 0.5; Content significantly positively correlated with CognCompl (0.551) and
TimesBrandname significantly negatively correlated with MusicOnly (-0.607). To test for
multicollinearity, the correlating variables were regressed against each other. The highest VIF score resulted from regressing TimesBrandname against MusicOnly with a score of 1.459 which is within the tolerance level of below four (Malhotra, 2007). Hence, we can conclude that multicollinearity is not a problem within the dataset.
Methods of Analysis
The unit of analysis is each individual evaluation of the ads by a respondent. Thus, it studies the skipping at a disaggregate level which is similar to the analyses of established advertising researchers such as Stewart & Furse (1986), Olney et al. (1991) and Siddarth & Chattopadhyay (1998).
The viewer’s decision to skip an online video ad is decomposed into two components – first, whether or not a skipping occurred and subsequently, the time a skip took place. To assess the influence of the explanatory variables on whether or not an ad is skipped it is modeled with a binary logit model. Whereas the second component – the timing of a skip – is tested through a proportional hazard model. The hazard model adds value to this thesis because unlike the binary logit model it has the possibility to incorporate the information of the time to an event (Mills, 2011). The approach of first modeling skipping in a binary logit and then continue with a duration model is based on the paper by Siddarth and Chattopadhyay (1998) who did a similar approach with TV ads zapping behavior.
3.7.1 Binary Logistic Model
A binary logistic model is useful to test whether a relationship exists between the predictor variables and the dependent variable. In a binary model the dependent variable can only have two values, either a 0 for an event not taking place and a 1 for an event taking place. The outcome is the probability that a viewer will skip an online video ad. In order to predict probabilities that lie within 0 and 1, OLS regression cannot be applied (Leeflang, Wieringa, Bijmolt & Pauwels, 2016). Logit Regression on the other hand creates values within the 0 and 1 range (Leeflang et al., 2016).
3.7.2 Hazard Model
The second analysis is the hazard model, also called duration model or survival analysis, which is a statistical method that focuses on the timing and duration of a specific event. Survival analysis has been used for different managerial reasons (e.g. firm survival, customer lifetime) (Boehm, 2008). The event in this case is the decision to skip an ad and the outcome is the hazard rate which represents the conditional probability that a skip occurs at a particular time interval (t).
from 0 to 16 seconds. We can derive the survival rate S(t) = 1 – F(t), indicating the probability that a skip has not occurred yet. The hazard function F(t) focuses on “skipping”, while the survival function S(t) focuses on watching the ad for its full length (Mills, 2011).
Knowing F(t) and S(t), the following hazard rate can be computed:
h(t) = 𝑓(𝑡)
𝑆(𝑡)) (1)
which can be interpreted as the conditional probability that a skip occurs at t, given that is has not occurred yet (Leeflang et al, 2016).
To account for the explanatory variables Cox proportional hazard model was applied (Cox, 1972), which is written as follows:
h(t|X) = h0(t) exp(X,β) (2)
3.7.2.1 Proportionality Assumption
One important assumption of the Cox regression is that the hazard rate for each individual variable is constant over time, hence proportional. This is tested prior to the estimation of the Cox regression and can be examined through different tests. When the test proves that hazard rates are constant over time, the researcher can continue with the Cox model (Mills, 2012). However, if the proportionality assumption is violated by an explanatory variable then the model needs to encompass variables that change over time (Therneau, Crowson & Atkinson, 2020). There are several methods that enable the Cox model to account for these time-varying effects.
Two methods are proposed in this thesis first including time-dependent covariates through the interaction with the logarithm of time and second through using specific intervals of time (Therneau et al., 2020). The first is usually the more recommended solution (Quantin, et al., 1996).
3.7.2.1.1 Stratification
3.7.2.1.2 Log Interaction with Time
The second method is to recode variables violating the proportionality assumption into time-dependent covariates (Therneau et al., 2020). Those variables that did not satisfy the proportionality assumption were included in the model with its main effect and the log interaction of time (Allison, 1995). This will incorporate the fact that the variable’s coefficient changes over time resulting in a non-constant hazard rate (Borucka, 2014).
For each variable violating the assumption the model additionally includes the interaction effect of time. The resulting formula is written as follows:
hi(t|Y1,Y2) =h0(t)exp(β1Xi1 + β2Xi2 + + β3Xi3 + β4Xi4 + β5Xi5 + β6Xi6
+ β7Xi7 + β8Xi8 + β9Xi9 +β10Xi10 + β11Xi11 + β12Xi12 + β13Xi13 + β14Xi14 +
β15Xi4*ln(t)+ β16Xi5*ln(t) + β17X i7 *ln(t) + β18X i12 *ln(t) - ln(t)) (3)
i the advertisement of a specific brand
Y1 the adapted viewing time of the respondent of advertisement i, Y2 the event of skipping,
ho the baseline function,
X1 the use of a celebrity in advertisement i,
X2 the use of real people in real settings in advertisement i, X3 the use of a slice of life appeal in advertisement i, X4 the use of sensual images in advertisement i, X5 the solely use of music in advertisement i, X6 the continuously use of a Logo in advertisement i,
X7 the numbers of times the advertised brand is mentioned in advertisement i, X8 the informative vs. entertaining content in advertisement i,
X9 the scale of cognitive complex characteristics in advertisement i, X10 the gender of the respondent of the evaluation of advertisement i, X11 the consistency with the target audience of advertisement i, X12 the age of the respondent of the evaluation of advertisement i,
4 RESULTS
Descriptive AnalysesThe descriptive analysis of the 9,838 evaluations show that in 41.94% (4,126) the respondents watched the full length of the video while in 58.06% of these evaluations the ad was skipped. In addition to examining the skipping rate, the viewing time is also investigated in this thesis. Figue 5 shows the cumulative number of skipped ads plotted against the viewing time. It visualizes that skipping starts right at the beginning and that after 3.5 seconds the number of skipped ads stagnates. Around 40% of all ads are skipped within the first three seconds and the highest number of ads is skipped between second 1 and 2 (1.866 ads). This is consistent with the analysis by Campbell et al. (2017) who also reported a peak in skipping after the “Skip” button appears.
Binary Logit Model
For testing the determinants on the probability to skip an online video ad, a binary logistic regression was performed. The binary decision variable equals 1 if the ad was skipped and 0 if an ad was watched for its whole length. The results are shown hereafter.
4.2.1 Multicollinearity
To test whether the intercorrelations among the predictor variables are high, the data was tested for multicollinearity. The correlation matrix in chapter 3 already anticipated tolerable correlations between the predictors. Likewise, the VIF scores of the model are all below the critical value of 5 (see appendix II) (Leeflang et al, 2016). As expected, multicollinearity is not an issue in the model.
4.2.2 Model Fit
To determine the fit of the model the Pseudo R², Hit Rate and likelihood ratio test was inspected. The explanatory power is shown by the percentage of the total variance explained by the model. Opposed to a linear model the R² is not meaningful in a binary logit model. Instead, Pseudo R² were computed which lead to similar interpretations (Leeflang et al., 2016). The three most popular Pseudo R² McFadden, Cox& Snell and Nagelkerke were computed which all compare the log-likelihood of a null model with the log-likelihood of the estimated model (Leeflang et al., 2016). The values of each measure respectively are 9.95%, 13.39% and 7.71%. Making comparisons with other studies, e.g. Bellmann, Schweda & Varan. (2010) who reported a R²
FIGURE 6
Skipping rate per Country
FIGURE 7
for logistic regression of 25%, and van Meurs (1998) of 31.7% the values are comparably low indicating that there are other factors that explain the variance in the dependent variable. The predictive validity of the model is assessed through the hit rate (Leeflang et al., 2016). The model has correctly classified 45.15% of true negatives (Skipped = 0), and 77.96% of true positives (Skipped = 1). This indicates the power of the model to predict if an ad gets skipped, however it lacks power when predicting if an ad is watched for its full length. Nevertheless, the total number on right classifications is 64.20% which is higher than the benchmark rates of the maximum chance criterion (58.06%) and the hit rate of the naïve model or so-called proportional chance criterion (51.30%) (Morrison, 1969). Thus, it can be concluded that the proposed model is valid.
Another indicator is the model’s predictive ability to a new dataset. The data was split into two different sets after the rows were brought in a randomized order. The estimation set formed 7.378 evaluations which is 75% of the evaluation while the remaining 25% with 3.460 evaluations in total were used to test the model. This allows to detect how the model performs on new data that was not used for the estimation of the model. The estimation sample data consists of 41.54% not skipped ads and 58.46% skipped ads, which is almost a similar resemblance of the real dataset (41.94% and 58.06%, respectively). The hit rate of the holdout sample is 64.51%, which is an adequate number.
The likelihood ratio test shows difference between the null model and the estimated model (Leeflang et al. 2016). The results show that the model with the explanatory variables is significantly better (p<0.001) (see appendix II).
Lastly, the top-decile-lift and Gini coefficient were measured which indicates the predictive validity of the model (Leeflang et al., 2016). The TDL shows a value of 1.39, meaning that the model is 1.39 times better at predicting the top decile than a null model (see appendix II).
4.2.3 Findings
No significant relationship was found for the interaction effect of SensImages and MusicOnly,
Logo, TimesBrandname, and the covariate GenderCons.
The results of the analysis address the impact of the ad’s characteristics on the likelihood of skipping. The odds ratio shows the change in the odds of the event (skipping) when changing the independent variable by one unit. In the case for the categorical variables the value indicates the odds of skipping an ad whether the cue is present or absent. As already discussed only the predictors with at least marginal significance will be interpreted.
The findings from the analysis are as follows. If the ad has a celebrity endorser in the ad versus having not a celebrity in the ad, the probability of skipping decreases by a factor of 0.8140 or 18.60%. In other words, having no celebrity in the ad increases the likelihood of skipping holding the effect of other explanatory variables constant. Another significant variable is
RealLife which indicates that online video ads with real life settings decrease the probability of
being skipped by 22.83% as compared to ads without this element, keeping all other variables constant. Regarding the content of the ad, the outcome affirms that entertaining ads have a 13.54% higher probability of being skipped than ads with informative value. The last significant variable is the number of cognitive complex characteristics. With every additional unit increase the probability of skipping decreases by 14.48%.
Hazard Model
The second analysis is the hazard model, which tries to determine the timing and duration of a skip. The model consists of two dependent variables, first the event variable that indicates whether a skip happened or not (Skipped) and second the time variable that shows for how long the ad has been watched (SecofSkip). Since respondents were able to skip an ad after watching it for at least 5 seconds, the time variable was subtracted by the value of 5 seconds. The same variables as in the binary logit model are used as predictors.
4.3.1 Kaplan-Meier Survival Function
any of the independent variables. In table 5, each row corresponds to a different second, beginning at the second 0, and ending at second 15.037. After second 0.054, 31 ads had already been skipped. The corresponding plot 8 visualizes the survival probability of an ad, regardless of its characteristics. It is visible that the survival probability drops substantially in the first few seconds, reconfirming that most skipping happens right after the viewer is able to skip an ad.
4.3.2 Proportionality Assumption Testing
In order to estimate the Cox model, a key assumption is the proportionality of the hazards. In other words, the different levels of covariates need to be constant over time. The proportionality assumption should be tested prior to the estimation (Mills, 2012). The variables can be statistically tested for its proportionality by looking at the scaled Schoenfeld residuals (Mills, 2011). The fulfilling of the assumption is a condition of the (proportional) Cox model because it estimates an average hazard rate over the whole time period (Therneau et al., 2020). If any variable violates the assumption, those parameter estimates will be either over- or underestimated. The output of the Schoenfeld residuals reveals whether the proportionality assumption holds or not (Broström, 2012; Dessai & Patil, 2019). If the p-value of the global test is lower than 0.05, the assumption does not hold (Broström, 2012). The outcome of the test reveals that for five variables proportionality is not fulfilled excluding the interaction effect of
This can be also shown by plotting the coefficient of each variable against time. Two examples are given, one graph illustrating one variable not violating the assumption (Celeb) (figure 10) and one violating the assumption (Age) (figure 11). The value of the average hazard over time is shown by the dashed horizontal line. A covariate is measured at the baseline. Whenever the time-varying hazard significantly deviates from the average hazard over time it violates the proportionality assumption (Therneau et al., 2020). While the black horizontal line shows that the effect of Celeb is constant over time, the black line of the covariate Age increases over time, indicating non-proportionality (Kalbfleisch & Prentice, 2002).
4.3.3 Violation of the Assumption
Hosmer & Lemeshow (1999) stress that violations of the assumption need to be taken into account to circumvent biased estimates and a loss in the power of the model. There exist different solutions to make the model proportional. For example, modifications of the variables by accounting for time-varying effects can help to avoid non-proportionality (Zhang, Reinikainen, Adeleke, Pieterse & Groothuis-Oudshoorn, 2018). In this thesis, two different methods are tested. First, a model with log interactions of time and secondly, stratified model with a step function.
For all the variables violating the proportionality assumption the data was split into specific time intervals. The interaction effect of SensImage and MusicOnly was excluded from both models. The outcome of the stratified model result in different coefficients for the prespecified intervals (Therneau et al., 2020). Due to the reason that the interaction model achieved a lower BIC we will continue with the interpretation of that model. For comparisons of the methods the exact estimation of the stratified model is available in appendix III.
4.3.4 Model Fit
The table 6 reports some measures signaling the overall fit of the model. The information criteria (AIC & BIC) have been already used to pick the model with the highest fit. The explanatory power of the Cox model is reported through the calculation of the Pseudo R² which show the following percentages; McFadden of 18.74%, Cox &Snell and Nagelkerke of around 85%. These numbers indicate an appropriate fit of the model.
the data, this means the model is concordant. The model predicts 95% correctly, which is considerably larger than the benchmark of 50%.
Further, the likelihood ratio test leads to the rejection of the null hypothesis implying that the fit of the model is better than the model without interactions as well as the null model (see appendix III).
4.3.5 Findings
Similar to the binary logit model, the odds ratio – in this case the hazard ratio – will be interpreted (Mills, 2012). The hazard and the odds ratio have the identical form, however, when interpreting Cox models, it is referred to the ratio of rates rather than of odds (Hosmer & Lemeshow, 1999). Compared to the odds ratio it measures the survival rate of the entire time period instead of looking at the endpoint of the time (Hosmer & Lemeshow, 1999). From the variables that were used to answer the research question, six out of nine are significant and therefore their effects will be interpreted.
An ad with a real people in real setting feature has a negative effect on the skipping rate, with a probability of being 30% skipped less than an ad without this feature. The variable SliceLife has a negative effect (β=-0.18) and the probability of skipping an ad decreases by 17% when a slice of life appeal is applied as compared to when it is not applied. The significant estimate for
Logo equals to -0.10, with a hazard ratio of 0.90 which means that ads with a Logo are
approximately 10% less likely to be skipped as compared to ads that do not continuously show the brand’s logo.
Due to the fact that the proportionality assumption is violated for five of the variables in the model the interaction effects need to be take into account to examine the varying hazard rate throughout the time period. Figure 12 visualizes the hazard rate for the variables that depend on time.
Ads with sensual images and ads that play mainly music in the ad are on average more likely to be skipped than when these characteristics are reversed. However, when accounting for the time effect, the difference between an ad with and without these characteristics depend on the second of the ad. The hazard ratio exceeds 1 in the early phase of an ad, but gets lower than 1 after 2.5 seconds indicating that ads with these characteristics are less likely skipped as the ad continues.
Interpreting the effect of the variable TimesBrandname is not entirely clear because it entails that the longer the viewer watches the ad the higher the number of times the brand is mentioned. Therefore, in this case we consider the coefficient of the main effect of the variable which constitutes the “average” effect of the variable (Allison, 1995). Each additional increase in the time the brand is mentioned increases the probability of skipping by 40%.
5 CONCLUSION
To conclude, this thesis has tested several factors that showed to have an impact on advertising effectiveness in previous papers. The outcome of the study reveals that some of the tested characteristics indeed led to a change in the skipping rate. Therefore, we can conclude that some techniques persuade the viewer to watch the whole content of the ad while others cause the viewer to prematurely skip the advertisement.
The aim of the research was to find an answer to the proposed research question by investigating ten hypotheses that formed the research model of this thesis. The explanatory variables that have an effect on the avoidance of TV ads were used for this research. The assumption was that successful strategies and characteristics used for TV ads does not correlate for online video ads and may even lead to contrary effects. As stated by Baek & Morimoto (2012), approaches that work with traditional advertising may have undesirable outcomes when used in online ads.
Hypotheses Testing
Table 7 gives an overview of the prior expected effects of the variables together with the results of both analyses. All variables used to answer the hypotheses were significant in at least one of the models. Interestingly, for almost all variables the effects that were significant for one model were not significant for the other model. Only H2 is approved by both models.
Derived from the research of Campbell et al. (2017) the hypotheses were assumed based on the following assumptions; First, attention-getting techniques are ineffective because the viewer is already in a heightened active state. Second, specific characteristics increase the already high level of intrusiveness. Third, appeals that are often used in advertisements help the viewer to recognize the occurrence of the ad. However, the results reveal that this is not entirely true due to having different signs of the effects than expected.
Considering the outcome of the first model, we conclude that the attraction of a celebrity is larger than the viewer’s wish to continue to the desired video which contradicts with the assumption of H1. This outcome confirms the theory by Atkin & Block (1983) and MacInnis et al. (1991), who showed that people enjoy watching ads containing celebrities.
viewers like to see an authentic and real message (Zatwarnicka-Madura, 2018) or alternatively because this appeal makes it more difficult to detect the ad as an ad.
Contrary to our expectation using a slice of life appeal is advisable in an online setting. This might be because this appeal brings value to the ad by demonstrating the product as the solution to a problem.
