Social Television by Robin Papa

1

Master Thesis Marketing:

The Effects of Mentioning Social Media

During Television Broadcasts

On Social Media Behaviour

University of Groningen

Faculty of Economics and Business

The Effects of Mentioning Social Media

During Television Broadcasts

On Social Media Behaviour

ABSTRACT

This paper is a case study of a broadcast by NOS on social television techniques and its effect on television ratings and number of tweets. It provides a framework for time series analysis for social television research and examines how the above two metrics are influenced by emotional moments during the broadcast and short intermissions in which social media activity is presented during a broadcast. Furthermore, we establish an introductory chapter on the concept of social television, to establish new grounds for research in this upcoming field of marketing. Previous research on social television failed to address the trending nature of the data, which this paper does by using a Vector Auto-Regression model and Impulse Response Functions. We examined data of a day-long broadcast of the Dutch coronation of April 30th _{2013, during which}

three-minute blocks of social media coverage was piloted. We find that spoken encouragement to tweet during such social media blocks has greater impact on number of tweets sent than counterparts without encouragement and that any effect wears off quickly after the social media block. We also find that television ratings in this case study do not affect the amount of tweets sent. Finally, we see that social media blocks have a greater effect on number of tweets sent than emotional moments. The managerial implications derived from these results have great impact for television producers planning to implement social television techniques.

Keywords: Social Television, Twitter, Television, Social Media, Hashtags, Emotion, VAR JEL Code: M31, L82

ACKNOWLEDGEMENTS

2

0. CONTENTS

1.3 Managerial and Scientific Relevance ... 5

1.4 Concepts and Definitions ... 6

1.4.1 Social Television ... 6

1.4.2 Buzz ... 6

1.4.3 Twitter ... 7

1.4.4 Hashtags ... 7

1.5 Structure of this Paper ... 7

2. SOCIAL TELEVISION ... 9

2.1 Channels of Communication ... 9

2.2 Social Television as Earned Media ... 9

2.3 Social Television as Means of Engagement ... 9

2.4 Twitter as Measure for the Television Industry ... 10

2.4.1 Tweeting About Television or “The Input behind Television” ... 10

2.4.2 Tweeting For Television or “The Input to Add to Television”... 10

2.4.3 Twitter-Enhanced Television or “The Input to Change Television” ... 10

2.4.4 The Dimensions Above as Television Audience Research ... 11

3. THEORETICAL BACKGROUND & CONCEPTUAL MODEL ... 12

3.1 Theoretical Foundation in Related Fields of Research ... 12

3.2 Related Effects of Television Ratings and Tweets ... 13

3.3 Effects of Social Media Blocks on Television Ratings and Tweets ... 14

3.4 Effects of Emotional Moments on Tweets ... 15

3.5 Effects of Age on Tweets ... 15

3.6 Conceptual Model ... 16

4. METHODOLOGY ... 17

4.1 Initial Approach ... 17

4.2 Data Description ... 17

4.2.1 Tweets per Minute & Television Ratings ... 18

4.2.2 Social Media Blocks ... 18

3

4.3 Vector Auto Regression Preparations ... 23

4.3.1 Procedure ... 23

4.3.2 Unit Root Test for Stationarity ... 23

4.3.3 Cointegration... 24

4.4 Lag Length ... 24

4.5 VAR(X) Model Specification ... 25

4.6 Impulse Response Functions ... 25

5. RESULTS ... 26

5.1 Model Fit... 26

5.2 VAR(X) Model Estimation ... 26

5.2.1 The Effects of TV Ratings ... 27

5.2.2 The Effects of Tweets per Minute ... 27

5.2.3 The Effects of Social Media Blocks ... 28

5.2.4 The Effects of Emotional Moments ... 29

5.2.5 The Effects of Age ... 29

5.3 Impulse Response Function Analysis ... 30

5.3.1 IRF Analysis for Television Ratings and Tweets per Minute ... 30

5.3.2 Manual IRF Analysis for Social Media Blocks and Emotional Moments ... 31

6. DISCUSSION & MANAGERIAL IMPLICATIONS ... 33

6.1 Television Ratings & Tweets ... 33

6.2 Social Media Blocks ... 34

6.3 Emotional Moments ... 35

6.4 Age ... 36

7. LIMITATIONS & FUTURE RESEARCH ... 37

7.1 Limitations ... 37

7.1.1 Ability to Generalize Results ... 37

7.1.2 Not Accounting for Retweets ... 37

7.13. Sample Size of Social Media Blocks ... 37

7.2 Future Research ... 37

7.2.1 Emotional Stock ... 37

7.2.2 Older Demographic ... 38

4

1. INTRODUCTION

The ways of watching television have been intensively changing over the past couple of years, as social media usage is increasingly integrated in people’s daily lives. Currently almost every new television series makes applications available for tablets and phones and usage of these applications while watching television is rising steadily. The creators of television shows establish conversation with viewers by maintaining social networking channels such as Facebook pages and Twitter feeds. Sometimes during live broadcasts of certain television programmes the audience is asked to comment on the show via Twitter. These examples show that social television is a field of social media that is becoming a trend for television networks.

1.1 Problem Specification

Social television is a field that attempts to combine social media networks with television productions and gains popularity under creators of television programs. We find Twitter cooperating with talent show The Voice USA (Edelsburg, 2011), see hashtags (explained in section 1.4) displayed during many primetime television series and Nielsen acquiring social television metrics startups (Nielsen, 2012). The author of this paper observes a trend in these developments and asks the question why social television is of increasing interest for television producers and if that interest is justified.

Social television is a concept that entails the use of so-called second-screen devices, such as smartphones or tablets to share thoughts about what they are viewing on social networks, during the process of watching the first-screen, the television set. This trend is interesting for marketers because a rapidly expanding real-time focus group starts to exist. Currently, social television is believed to be a source of added value for both audiences and producers of television shows in four dimensions: (1) Social TV is believed to increase ratings, (2) to improve the viewing experiencing of the audience, (3) to be a means of engagement with shows and brands beyond the broadcast window and (4) to provide a new marketing and advertising platform (Ad Age, 2012).

It should be noted that knowledge regarding these four dimensions is coming from sources in the commercial social television industry and may very well be biased towards the notion that social television is of positive value. For example, television rating market leader Nielsen, which recently took over social television analytics company SocialGuide, is the author of one well-cited report on the effects of social television buzz on television ratings (Nielsen, 2012). Furthermore, a connection between buzz and hype or ratings for television shows is commonly expected (Proulx & Shepatin, 2011), yet we find little scientific proof for any causality. We therefore conduct a case study in which social television techniques (by making use of social media content) are applied by the producers during a broadcast of the Dutch coronation (more on this in section 1.2). We hope to derive practical insights on the effects of social television on television ratings and the number of tweets sent.

Because the field of social television consists of many types of application and sometimes unique terminology, we introduce some of its underlying concepts in section 1.5 and elaborate on these concepts in chapter 2. The next section elaborates on the approach of this case study.

1.2 Research Approach

5

place on the 30th of April 2013, during which Queen Beatrix abdicated and her son, Prince Willem-Alexander, became king of The Netherlands.

To account for second-screen activity, we use Twitter as source of buzz. More specifically, we focus on the number of tweets being sent at any given minute during the broadcast. Concepts regarding Twitter will be explained more thoroughly in section 1.5: concepts and definitions.

To examine the effects of social media coverage during television broadcasts on the number of tweets, we state the following research question:

What is the effect of television events on the number of tweets sent during and after these events?

In this research question, television events are defined as either staged social media intermissions during the broadcast in which social media activity is discussed, or emotional moments that occur live through the day. An elaboration on this categorization of events can be found in section 3.3 and 3.4.

1.3 Managerial and Scientific Relevance

Social television techniques might be able to recover to some extent the loss of live viewers in various scenarios. For example, social media may be used to bridge the gap between seasons1_{, measure sentiment}

about certain shows and generate feedback on how to improve formats. Social television may also engage live show viewers if it offers ways of engagement with the viewer during these shows. Such practices are already seen during talent shows, such as “The Voice”. Furthermore, by adding value by using social television practices as described above, the rising trend in piracy of television series can be contained to some extent (Business Insider, 2013)

Scientifically, no frameworks on measuring the effectiveness of social media techniques in the television landscape have been established, nor has sufficient research been done regarding the social television field. In the field of social sciences, research on television is often conducted, though not from a marketing perspective. The uniqueness of social television lies in the possibilities it offers for marketers to increase engagement and measure show effectiveness. At the same time, social television opens up a new advertising possibility, which makes the techniques surrounding it even more important to understand for television network’s marketing departments. A specific subject of research would regard the field of advertising, which, through digitalization and socialization of television systems, may shift from measuring effectiveness through mere awareness to measuring effectiveness through engagement (Griner, 2012).

The theoretical foundation for this case study will be based on research in related fields, as because of the novelty of this research subject, little to no previous research has been done. Therefore chapter 3 will show current insights regarding direct response marketing, word-of-mouth communication, and emotional effects in advertising, to form a foundation for the hypotheses tested in this case study.

1 _{This paper uses the North-American definition of seasons. Episodes of television programs are usually broadcasted in sets of 20 to 26 in a}

6

1.4 Concepts and Definitions

In the upcoming section we clarify some of the terminology of importance for this research. We discuss the concept of social television, buzz and hashtags. We introduce several measurements of these concepts, which will be used and more thoroughly explained in later chapters.

1.4.1 Social Television

Social Television is a term used to describe current integration of social media interaction with television programming (Hill et al., 2012). Previous research in the field of social television has defined the concept in a broad sense, consisting of all technology supporting social practices while watching television (Nathan, 2008; Cesar, et al., 2008), also known as Connected TV (Cesar & Geerts, 2011). However, specific developments in the field of online social media platforms, such as Twitter, require us to adapt a more precise definition. We therefore suggest to define Social TV as a concept where social media platforms support practices that connect people watching television and accomplish the creation of a communal experience of group viewing while not physically together, where conversation can be established among and between creators, publishers (broadcast networks) and recipients of television content. A definition as such offers the possibility to gain insights on social TV from a television production perspective, in addition to discussing merely the underlying principles of twitter user behaviour (for example: Java, et al., 2007; Deller, 2011). An increase in use of Twitter and other social media platforms regarding a specific television show is currently a measure of marketing buzz for television creators (Hill, 2012).

1.4.2 Buzz

Buzz is a term originating from the field of word-of-mouth marketing, but somewhat problematic to define. Word-of-mouth marketing is about engaging the customer through purposive components other than product distribution and advertising. Instead, it focuses on creating buzz, which is seen as the commercialization of increasing or starting word-of-mouth conversations, which “tend to be random and spontaneous in nature, occurring in a natural, unpredictable pattern of communication” (Ahuja et al., 2007, 152). The television industry shows tendencies to stimulate these word-of-mouth conversations under its viewers, as seen in section 2.2. Viewers of a television show may very well create buzz, as according to Mangold & Faulds (2009: 362), “people are more likely to communicate through both word-of-mouth and social media when they are engaged with the product, service or idea”. In this paper we are primarily interested in such an effect of buzz on ratings, but also measure for causality in the opposite direction, because buzz can also be seen as an influencer of ratings, in addition to television ratings being an influencer of buzz.

7

It therefore seems that correlation between television ratings and buzz exists, but the direction of causality is still a point of discussion. This paper therefore accounts, as stated, for both directions of causality.

1.4.3 Twitter

The dataset used primarily consists of messages on Twitter, which are called tweets. Twitter makes access to these messages possible, which creates the opportunity of storing messages that fulfil certain criteria. Storing specific messages is possible through Twitter’s application programming interface (API). The API was used to collect tweets containing a certain keyword or tag that is mentioned in the message, preceded by a hash sign (#), which makes searching and collecting messages on a certain subject easier. These keywords preceded by a hash sign are called hashtags.

Another concept regarding Twitter is the process of retweeting, which is a form of quoting someone else’s message. In this case study, retweeting would consist of reading messages on Twitter about the coronation and then deciding to repost this (usually accompanying a short comment of the “retweeter”). A retweet can be compared to the forwarding of an email.

1.4.4 Hashtags

Hashtags are keywords that make the collection of messages about a specific subject possible on the social network Twitter, a function that is useful for users, organizations and researchers. On television, the broadcaster usually shows a hashtag on-screen to encourage sending tweets containing that specific hashtag. For example, during the broadcast of the Dutch coronation, NOS showed the hashtag #troon (which translates to throne) during the broadcast and let NOS reporters use this hashtag in their tweets early on, to try and establish it as the conventional one. Sometimes a small amount of Twitter users do not comply with the convention of using the preferred hashtag.

The advantage of hashtags is that they make messages containing important keywords easier to find. For example, when searching for the word ‘Lost’, one would find messages such as “I was lost”, while searching for #Lost often shows messages related to the TV series Lost. Using hashtags enables users to create specific threads of conversation. The following message is an example of a tweet that has been sent out the coronation, using hashtag #NOS and hashtag #troon:

@Ma*en: Wat een mooi feest! Doet NL goed. Complimenten en respect voor verslaglegging door #NOS en idem voor scriptschrijvers van vandaag #troon

(The above tweet roughly translates to “What a beautiful festivity! The Dutch are doing great! My compliments and respect for the reports by #NOS and also for the scriptwriters of today #throne”.)

In chapter 2, current insights on the concepts regarding Social Television will be further explained and the current state of affairs regarding these concepts will be highlighted.

1.5 Structure of this Paper

8

9

2. SOCIAL TELEVISION

In this chapter several concepts regarding social television are examined and past research on these concepts, though currently in its initial stage, will be discussed in order to establish a theoretical framework. This chapter serves as elaboration on the social television landscape and will provide ground for research in this case study.

2.1 Channels of Communication

Why is it important for television broadcasters that their audience tweets about what they see on the screen? When considering types of promotional efforts, a distinction can be made between paid, owned and earned media (Burcher, 2012). Paid media are channels of communication for which an organization pays, for example advertisements for a television show in a newspaper. Owned media are channels that an organization controls, such as the channels on which NOS broadcasts. Earned media is another type of communication channel, such as conversations on social media platforms and the concept of word-of-mouth, which cannot be bought, hence the term earned.

2.2 Social Television as Earned Media

The advantage of broadcasters is that they can use owned media to induce earned media with relatively small investments when compared to paid media, as social media elements can be produced in-house. Organizations in other industries usually try to establish a degree of earned media by using both owned and paid media as a spark (Burcher, 2012).

Initiatives to measure and establish conversations in the form of earned media are appearing in the television industry. We see an increase in application development for tablets and phones, adapted to television shows, which enable interaction between users and show producers. One example is The Voice USA application, which lets people vote for their favourite performance in this live talent show (Bergman, 2012). The year before, The Voice USA cooperated with Twitter in order to let audiences vote using tweets (Edelsburg, 2011). Furthermore, increasingly more shows place hashtags on-screen during television broadcasts to stimulate conversations about the show on the Twitter platform. Nielsen, a renowned organization for the provision of television ratings, recently acquired social television analytics company SocialGuide (Nielsen, 2012). Another sign that social television endeavours are gaining support is the fact that Twitter hired Fred Graver, television executive, as their head of television. These recent developments beg the question why earned media is seemingly of interest for the television industry and, more importantly, is this interest justified?

2.3 Social Television as Means of Engagement

One of the goals stated by the NOS is to increase engagement with the audience. This is currently achieved by asking reporters, correspondents, show hosts and editors to establish rapport with the audience, for example by using personal or company-owned Twitter accounts. During broadcasts of events such as the elections, NOS tries to elaborate on what is being said on social media platforms such as Twitter.

10

2011: 237), in terms of creating the most buzz. These findings make Twitter a valuable tool in measuring behaviour regarding word-of-mouth and the effect of television has on this behaviour.

2.4 Twitter as Measure for the Television Industry

Harrington et al. (2012) distinguish four dimensions of interaction and intersection of an old medium, television, and a new medium, Twitter. In order to sketch an image of the current state of affairs regarding the use of Twitter by both audiences and producers of television shows, we will highlight these dimensions below.

2.4.1 Tweeting About Television or “The Input behind Television”

Twitter is often used as backchannel or “water cooler”-platform for audiences of television programmes. The social media platform sustains an aspect of behaviour that exists for decades, the difference being that discussions about television shows at the water cooler are replaced by having these interactions during the live broadcast with other viewers. Shared viewing and receiving or sending peer recommendations are the goals that people have for using Twitter while watching television (Montpetit et al., 2009).

2.4.2 Tweeting For Television or “The Input to Add to Television”

Many television networks incorporate social media as a structured way to engage audiences. Instead of a backchannel, Twitter is to some extent used as a part of the show itself. An example of this phenomenon can be seen in talent shows such as The Voice of Holland, where certain tweets of audiences are shown on-screen. In this case study a similar approach is used by NOS in its coronation broadcast, as certain tweets are highlighted in so-called social media blocks during the broadcast. A broadcaster that asks its audience to use a specific hashtag is another form of influencing tweeting about television. This practice is more commonly used, as it offers a non-intrusive way to let the audience connect on Twitter in shows that do not ask for audience participation on-screen, such as Dutch television programmes Nieuwsuur, a broadcast on current affairs and news, or Studio Voetbal, a football discussion programme. Activities like these can transform the monologue that television entails into a dialogical interaction between producers and audience, which in various cases may add value to a broadcast. This paper focuses on this specific type of social television.

2.4.3 Twitter-Enhanced Television or “The Input to Change Television”

Many possibilities for inter-media communications between social media and television have not yet been explored. Many networks therefore experiment with twitter-enhanced television programming. CBS series Hawaii-Five-O stated it would let its Twitter audience choose the ending of an episode, so that one of the suspects portrayed would indeed be chosen as the person who committed the crime. This did not affect the ratings substantially, nor the number of tweets, as the number of viewers did not notably rise and the specific episode did not attract more tweets than another episode of the same series (Proulx, 2013). This example shows one of many possibilities where Twitter is a platform that lets television audiences engage with television shows.

11

inadvisable to implement. This research will contribute to the strategic backdrop of social television by offering knowledge that increases effective decision-making about social media in the television landscape.

2.4.4 The Dimensions Above as Television Audience Research

All dimensions mentioned above (tweeting about television, tweeting for television and twitter-enhanced television) can be used as an input of television research. The rich stream of data that tweets contain offers a rich form of feedback for television makers. The nature of this data, qualitative as well as quantitative, can provide grounds for empirical research in the field of television.

12

3. THEORETICAL BACKGROUND & CONCEPTUAL MODEL

This chapter introduces the research design of this paper, contains the conceptual model used for analysis of the case study and derives hypotheses from this model. We distinguish trigger events, such as emotional moments and social media blocks. Furthermore, we add television ratings and number of tweets as relevant variables.

3.1 Theoretical Foundation in Related Fields of Research

This paper is an exploratory research regarding questions that need answering as the techniques that social television provides are progressing and more widely used. This chapter shows what variables will be examined and what types of relation are expected. A summary of all hypotheses can be found in section 3.6 in the form of a conceptual model.

As concluded in the previous chapter, scientific research on social television is non-existent. In order to provide theoretical foundation for this case study, we therefore resort to similar fields of research. We divide this study into multiple sections of interest and find related research for each of them. These fields are (1) the related effects of television ratings and tweets, (2) the effects of social media blocks on television ratings and tweets, (3) the effects of emotional moments on tweets and (4) the effects of age on tweets. For each section any relevant previous research is discussed.

When examining (1) the related effects of television and tweets per minute with the expectations by the industry as described in chapter 2 in mind, we see that Twitter is used by viewers of television shows as a communication channel about that show. Television producers tend to stimulate this form of communication, as they believe that this type of word-of-mouth communication (see also section 1.5 and 2.4.1) increases the show’s publicity and television ratings. From this perspective, we can find a related field of research in word-of-mouth as an influencer of satisfaction. The possibility exists that word-word-of-mouth about a television show may lead to a form of satisfaction about this show among non-viewers, which in turn might turn this potential audience into viewers.

We find that (2) social media blocks, as an effect on television ratings and tweets per minute, are a form of triggers. We also see the use of triggers, albeit in another form, in the field of direct response marketing, where triggers such as certain types of characteristics in for example mailings, television commercials or radio commercials are tested for size of effect. Most often these effects are measured in propensity to buy or actual purchases. In our case, we examine how a trigger (the social media block) results in a similar type of action (switching channels or sending tweets). Furthermore, we find theoretical founding on triggers in word-of-mouth research, regarding the manner in which word-of-word-of-mouth communication can be influenced or provoked, as these social media blocks function as a trigger for word-of-mouth communication via Twitter. Similar to the research fields mentioned above, we find overlap between (3) emotional moments and the emotional aspects affecting word-of-mouth or propensity to take action. For example, we use research regarding emotional appeals in direct advertising, affective communication and its relation to consumer behaviour and emotion in advertising to establish predictions on the effect of emotional moments broadcasted on desired behaviour or actions (tweeting).

13

3.2 Related Effects of Television Ratings and Tweets

As described in section 1.4.2, social media buzz is seen as a predictor for television ratings, but television ratings are also seen as a predictor for social media buzz. Social media is encouraged on-screen, which would result in more buzz as an effect of more viewers. On the other hand, before a new television series launches, buzz is seen as a measure for the initial success, measured in television ratings. The direction of causality, therefore, is not very clear: are changes in buzz affected by television ratings or are television ratings affected by buzz?

We therefore test for both directions of causality between television ratings and changes in tweets per minute (buzz). The causal effect of television ratings on tweets per minute is based on the notion that when more people watch the broadcast, the more tweets should be sent. The research by Nielsen (2011) that was introduced earlier showed that a correlation exists between television ratings and social media activity. An article from The Nation (2013) describes the state of social television decisively, saying “the local TV industry has not yet figured out whether the number of tweets and additional exposure on social media like Facebook

has anything thing to do with ratings.”

We previously mentioned Mangold and Faulds (2009: 362), who stated that “people are more likely to communicate through both word-of-mouth and social media when they are engaged with the product, service or idea”. Based on this notion we assume that viewers who watch the broadcast are in the broadest sense engaged with this broadcast and are therefore more likely to communicate through word-of-mouth and social media. We therefore expect that a change in television ratings affects the number of tweets positively.

H1: A change in Television Ratings positively affects a change in Tweets per Minute.

As said, we also test for effects in the opposite direction, from tweets per minute to television ratings. No previous research exists regarding this direction causality. We can however find theoretical foundations in word-of-mouth effects research. Word-of-mouth marketing has very large carry-over effects when compared to media messages or promotional events and is effective in the process of referring new customers (Trusov et al., 2009). We can compare word-of-mouth communication to tweeting in the sense that Twitter is an often-used platform for brand expressions. For example, Jansen et al. (2009) show that 19% of tweets include mention of brands.

Gruen et al. (2005) suggest that electronic word-of-mouth (eWOM) in the form of discussion board conversations increase likelihood to recommend a product but does not influence repurchase intentions. This outcome already provides a critical note on the effectiveness of gaining viewers through social media while the broadcast is airing, as recommending a product can be compared with communicating in the form of tweeting or retweeting (see section 1.5.3) but repurchase intentions can be seen as a clear action that has to be undertaken, comparable to switching on the television set, which in turn increases television ratings. Actions as such are seemingly not clearly affected by eWOM.

14

Thus, contrary to popular belief (discussed in chapter 2), we assume no causality of tweets per minute on television ratings, as stated in hypothesis 2.

H2: A change in Tweets per Minute does not positively affect a change in Television Ratings.

3.3 Effects of Social Media Blocks on Television Ratings and Tweets

During the broadcast of the coronation, seven social media blocks have been scheduled, in which tweets of the audience and the hashtag #troon have been shown on-screen (more on this in section 4.2.2). In some cases the host of these blocks encourages its audience to tweet. The latter version is expected to result in more tweets. Such a form of encouragement as described can be seen as a call to action or trigger, which is a technique used in direct response marketing. A call to action promotes change in behaviour, for example from passively watching television to sending a tweet. Nevertheless, also merely showing information, like the social media blocks in this case study, can lead to behaviour change (Smith, 2006).

Research on Twitter behaviour during a broadcast of The Voice of Germany, a German talent show, shows an 18.8% increase in tweets when a hashtag was shown on-screen, comparing the number of tweets three minutes before and three minutes after showing the hashtag (Hill, Benton, 2012). In all social media blocks that were shown during the coronation broadcast a hashtag was shown.

This study asks if a social media block is indeed effective in increasing the number of tweets that are sent during and after that block and test for this relation.

H3: The occurrence of a Social Media Block results in an increase of Number of Tweets.

As stated, we also want to know if a vocal encouragement to tweet leads to a greater increase in number of tweets when compared to its encouragement-lacking counterpart. To illustrate, one of the spoken encouragements was: “keep on twittering, we like it a lot”. If we compare an encouragement during a social media block with a direct mailing, we see that both entities wish for action by the recipient. Phelps et al. (2004) discuss that a messages with strong emotional elements, such as humor, fear, sadness or inspiration are more likely to be acted upon. If we classify a social media block with an encouragement as a form of inspiration, then such blocks should lead to more tweets.

Furthermore, in direct response marketing, a distinction between relational (retailer-image enhancement) and promotional (call-to-action) messaging can be made, where promotional messages seem to work well on short-term but not on long-short-term basis (Abad et al., 2011). When we relate this to vocal encouragement, we should see a short-term rise in tweets and no term effects. Note however, that the above research defines long-term in multiple days, where the examined broadcast in this paper per definition does not exceed a day. Regardless, we measure for increased effect on the number of tweets by social media blocks with vocal encouragement, as stated in hypothesis 4.

H4: Social Media Blocks with Vocal Encouragement result in a larger increase of Number of Tweets when compared to Social Media Blocks without Vocal Encouragement.

15

be exceeded when using social media blocks (NOS, 2013)2_{. Research shows that when sending personalized}

messages, customers react negatively when this ideal level of communication is exceeded (Godfrey et al., 2011). In said paper the ideal level of communication is measured as the number of times any contact with a customer was made, which is comparable to a social media block in which engagement with the viewer is sought. Furthermore, it was shown that when a message is more difficult to ignore such as unwanted email, it is seen as intrusive, which evokes negative reactance, which in our case could emerge in switching channels or turning off the television. Then again, when the receiver has significant control over message processing, which is the case when watching television, a message rarely evokes negative reactance (Nordhielm, 2002).

Based on the above research, we believe that on one hand a lack of fit between users of Twitter and recipients of Twitter-related broadcasting might be influencing television ratings negatively, as a result of age, level of communication and impossibility to ignore. On the other hand, when stating that message processing when watching television is controllable by the viewer, a social media block that is received by a viewer that lacks fit with the message may result in ignorance by this viewer. The television ratings should therefore not be affected by the occurrence of social media blocks, as non-users of Twitter will ignore the message or at least not negatively experience the message. In this light we state a fifth hypothesis, which entails the prediction that there is no effect in terms of loss of viewers.

H5: The Occurrence of Social Media Blocks Does Not Affect Television Ratings.

3.4 Effects of Emotional Moments on Tweets

Besides staged social media blocks, we also state that spontaneous moments of emotional relevance to the viewer are an indication of a change in tweets per minute. For example, we believe that during certain events, such as the balcony scene with the Royal family or the King’s speech, people are more inclined to tweet. To find theoretical grounds for this assumption, we may look at research regarding direct mailing. Such research shows that impact of extraordinary design categories affect opening rates of these mailings. For example, when a mailing is of extraordinary format and colours, it is opened more often (Feld et al., 2012). This outcome shows that differentiation in messaging can turn into action. In this specific case, this might mean that an extraordinary moment or event, which we call an emotional moment, results in more tweets.

Furthermore, Mehta and Purvis (2006) show that highly emotional advertising enhances recall of the advertisement, as opposed to what was believed by brain theorists before said paper was published. The same research points out that recall shows moderately high correlation with attention. Attention, along with motivation and ability, might result in behaviour. In this light we recite Phelps et al. (2004), stating that messages containing strong emotional elements are more likely to be acted upon. More specifically, emotional moments can result in higher attention, which then may result in behaviour such as sending a tweet, as stated in the sixth hypothesis.

H6: The Occurrence of Emotional Moments Affects the Number of Tweets at the time of occurrence.

3.5 Effects of Age on Tweets

Finally, we test if a rise in younger audiences results in more tweets per minute, keeping the goal of NOS in mind, which is targeting a younger audience through social media integration in its programming. We use an

16

age category of 6-19 years, as our data consists of additional television ratings for this group specifically. Twitter users aged 15 to 25 consist of 73.7% of the users of the social media network (Beevolve, 2012). Furthermore, 73% of American teens used social networking sites like Twitter, whereas only 47% of adults use these sites (Lenhart et al., 2010). We therefore expect that a rise in young people watching will result in a rise in number of tweets, as stated in hypothesis 7.

H7: When the number of people between 6 and 19 years old viewing the broadcast increases, the number of tweets increases.

3.6 Conceptual Model

To summarize, this paper’s model consists of 6 variables, of which the variables TV ratings and tweets per minute are initially expected to be of autoregressive nature. We therefore include lagged effects of ratings and tweets in the model, as shown by the curved arrows in figure 1. We test for effects on TV ratings and tweets per minute as a result of trigger events. These trigger events consist of social media blocks and any emotional moments that were shown on-screen. We make a distinction between social media blocks with spoken encouragement to tweet and blocks without this encouragement. More on this distinction will be discussed in chapter 4 on methodology. Finally, we add age as a variable to test for the effect on tweets by the 6 to 19 year old age category.

In measuring tweets per minute, we do not make a distinction between tweets and retweets. This is due to the manner of acquisition of data for this case study. This lack of distinction in the data set may affect the results and will be discussed in chapter 7 on limitations.

FIGURE 1

17

4. METHODOLOGY

This chapter provides a description of data used and the preparations taken to establish the final model. We will explain the steps taken to establish the VAR(X) model, test for stationarity and cointegration, show the equation of the model and finally discuss the use of Impulse Response Functions (IRFs).

4.1 Initial Approach

To understand how television ratings are influenced by use of social implementation as described in section 2.4.3, we focus on the coronation broadcast on the 30th _{of April 2013 by NOS, which was the sole broadcaster of}

the event on the Dutch television system. The broadcast lasted 14 hours and provide coverage of the abdication of Queen Beatrix and the succession by her son, Prince Willem-Alexander, who then became king of The Netherlands. The television ratings were provided by NOS and a database of tweets that were sent during the day was created. Furthermore, a timeline was made of the occurrence of social media blocks by NOS. Social media blocks are three-minute interruptions of the broadcast in which a reporter highlights several tweets and discusses remarkable content that was shared using social media, limited to tweets that were sent using the hashtag #troon. More information on social media blocks is stated in section 3.3.2.

The nature of this case study creates the necessity to account for historical values of variables, because when hypothetically a shock impacts the number of viewers, then the latter is not only influenced by that shock over time, but also by its previous values. To elaborate: when we see a rise in television ratings after a certain event, a pitfall would be to simply allocate the difference in viewers to the happening of that event. Nevertheless, this would not account for trends that were already present in the number of viewers before that event. It might very well be the case that the number of viewers was already rising before the event, which would mean that the event’s effect does not account for the whole rise, but just a part of it.

To account for trends such as mentioned, time series analysis is preferred. A value that is explained by its previous values is called autoregressive. In our case we have two time series, the number of viewers and the number of tweets, which results in the need to create a multivariate time series model, a vector autoregression model (VAR). Furthermore, to test for the effects of trigger events, we add exogenous variables, which results in a VAR(X) model.

By using a VAR(X) model we can measure the true effect a trigger event (a social media block or emotional moment) has on both time series data, while accounting for the trends of these time series (Pauwels et al., 2002; Steenkamp et al., 2005).

4.2 Data Description

18

The initial collection of tweets consisted of 187.000 entries, which contained tweets sent with hashtag #troon during April 30th _{2013 from the start of the broadcast until the end of the broadcast, which lasted from 8:59 till}

22:59. We then calculated the number of tweets per minute, which resulted in 840 (14 hours * 60 minutes) observations. We also gathered television ratings per minute from 8:59 till 22:59. Descriptive statistics of these time series data is shown in table 1. A graphical presentation of these time series can be seen in figure 2.

TABLE 1

Descriptive Statistics (Per-Minute Data)

Mean Median Min. Max. S.D.

Television Ratings* _{4158.4 3953 1200 6607 1343.0}

Tweets per Minute 184.9 155 23 1284 148.1

* in thousands

At some moments in time, television ratings data was missing, for example during intervening news broadcasts. Missing data for television ratings has been replaced using the linear interpolation.

FIGURE 2

Number of Tweets per Minute (solid) and TV Ratings in 10,000s (dashed) Timescale in minutes where 0 = 8:59 and 840 = 22:59

Figure 2 shows peaks in tweets per minute that are very sharp. Furthermore we see that television ratings and tweets per minute show some form of correlation, as the trend of both time series looks similar.

4.2.2 Social Media Blocks

During the eight hours of broadcasting there were seven social media blocks. During these blocks, that had an average duration of three minutes, a NOS presenter discussed several messages that were shared on social

19

media networks, mainly tweets using the hashtag “throne” (#troon). A specification of these social media blocks and their timestamps can be found in table 2. These events could not be controlled and the spoken encouragements or lack thereof are of random nature. Neither the time of the event nor the sequence of the blocks could be scheduled by the researcher beforehand, because the NOS needed flexibility in the scheduling of broadcasting certain events.

As mentioned in section 3.3, some blocks would contain an encouragement to tweet, while others would not. We therefore work with a division of social media block types as shown in table 2, where four blocks do not contain a spoken form of encouragement, while three blocks do.

TABLE 2

Specification of Social Media Blocks

Timestamp Contents of event Spoken Encouragement

09:27AM-09:30AM four tweets highlighted No

11:11AM-11:13AM four tweets highlighted Yes

12:47PM-12:50PM five celebrity tweets highlighted No 15:41PM-15:43PM tweets of foreign newspapers and television Yes 16:46PM-16:49PM tweets with photo’s showing pets + prize contest No

19:21PM-19:23PM three tweets highlighted Yes

22:26PM-22:29PM three tweets highlighted No

In the next part we will highlight the Social Media Block of 11:11AM with encouragement and the 15:41PM Social Media Block without encouragement to provide an initial overview of the data present in order to give an indication of the type of data.

As seen in figure 3, the 11:11AM social media block seems to only provide a rise in change of the number of tweets during the three minutes the block is broadcasted, which was during minute 133, 134 and 135 (or 11:11-11:13). We see the strongest rise in tweets during the 133th minute of the broadcast and see the effect wear off. After the block has ended at the 135th _{minute, a decline in the amount of tweets sent sets in.}

Table 3 shows the values seen in figure 3; the rise in tweets are respectively 632, 347 and 153 in the 133rd_,

134th _{and 135}th _{minute into the broadcast. The column “change in tweets sent at T=x” shows the values}

depicted in figure 3.

FIGURE 3

20

We also see that a new equilibrium establishes around 420 tweets per minute (tweets sent amount to 418, 410 and 424 around the 140th_{, 141}st _{and 142}nd _{minute). Nevertheless, the number of tweets per minute lowers}

after 11:20 and at 11:40 it returns to values beneath the pre-block value of 152 tweets per minute, 29 minutes after the social media block, which is not shown in the table below.

TABLE 3

Changes in Tweets at Social Media Block 11:11-11:13

If we compare the graph and table above with the social media block at 15:41PM depicted in figure 4, we see similar patterns in tweets per minute.

FIGURE 4

Tweets per Minute in Differences & Social Media Block 15:41-15:43PM

Time (T) Tweets sent at T=x Cumulative tweets sent at _T=x Change in Tweets sent at _T=x Cumulative Change in _{Tweets sent at T=x}

21

At 15:40 (T=401), a minute before the social media block was visible, the number of tweets sent was 75. In the first minute of the social media block, 412 more tweets were sent, as seen in column “change in tweets sent at T=x” in table 4.

TABLE 4

Changes in Tweets at Social Media Block 15:41-15:43PM

In this block we see a lower cumulative number of tweets that’s being sent during and after the block when compared to the social media block at 11:11, but the percentage increase is higher for the 15:41 block. Table 4, column “Tweets sent at T=x”, shows that this social media block resulted in a faster return to pre-block values. The tweets per minute amount to 77 at 16:01, 20 minutes after the social media block, as opposed to 29 minutes after the 11:11-11:13 block. We initially find a preliminary solid increase in the number of tweets sent during and after the broadcast of a social media block. This effect is immediate and quite large but wears off quickly.

The examined examples show short-lived large effects on the number of tweets during a social media block. After the social media block has ended, the effects of the block without encouragement wears off faster than that of the block with encouragement. We also see that the first minute of social media block creates the largest effect on the number of tweets.

Time = T Tweets sent at T=x Cumulative tweets sent at _T=x Change in Tweets sent at _T=x Cumulative Change in _{Tweets sent at T=x}

22

Besides the effects of social media blocks, we also considered that some moments on television which we classified as emotional are of emotional value to the audience. We define these moments as emotional moments and for classification we look at live blogs by prominent websites of high journalistic quality and collect a sample of moments that these blogs classify as newsworthy. We defined an event as newsworthy when at least two out of three news sources made a note of the event and elaborated on the content of the event. When multiple news entries of one source reported the same event, we used the timestamp of the first entry about the event. When times of the entry between sources were different, the earliest time was taken as timestamp of the event.

We then assumed that these moments are “emotional” to the extent that they are functioning as trigger to tweet about the happening shown on television. In other words, these events are seen as substantially important and provide content for the audience to possibly talk about. We only covered events that were directly related to the Royal House, which excludes coverage of interviews with attending audience, because we believe that these events do not add substantially to any engagement or emotion the audience may feel. The studied broadcast is meant as coverage of the experiences of the Royal House and we may assume that the audience watches the broadcast for the reason of watching those experiences and that these result in more probable triggers to tweet. Furthermore, facts about shown events, e.g. information about clothing manufacturers and information about foreign royal guests, are disregarded for the sole reason that the timestamps of the news entries stating these facts cannot accurately display the time when something or someone was first visible on-screen, thereby not providing enough accuracy to measure an effect on, for example, tweets per minute. Nevertheless, it might very well be possible that for example the dress of Queen Maxima was a provocation of emotional reactions for the viewing audience, but this research does not account for it for the reason stated before.

The above rules result in the following classification of emotional moments during the broadcast on April 30th 2013, as shown in table 5.

TABLE 5

Specification of Emotional Moments According to Liveblogs*

Timestamp Description of Emotional Moment

10:06 Queen Beatrix signs abdication.

10:31 Balcony Scene Princess Beatrix, King W-A, Queen Maxima & princesses. 13:50 Royal couple enters church.

14:14 Long round of applause for Princess Beatrix. 14:21 King Willem-Alexander’s oath.

14:51 Princess Amalia yawns and this results in laughter. 15:01 Beatrix and little princesses leave the church.

23

20:57 “Unexpected”** disembarking of King and Queen during boat tour.

* Liveblogs used are Dutch news sites Reformatorisch Dagblad (2013), De Telegraaf (2013), Nu.nl (2013) ** Stated as “unexpected” because the event was unknown by broadcaster NOS, which caused fuss.

The result is a classification of ten emotional moments which were of importance during the day of the broadcast.

4.3 Vector Auto Regression Preparations

This empirical analysis begins by testing for stationarity in time series of television ratings and tweets per minute. Augmented Dickey-Fuller Unit-root tests indicate occurrence of evolution for tweets per minute but indicate stability for television ratings.

4.3.1 Procedure

To correctly establish the model, we follow the steps as shown in table 6. We will first test for Unit Roots, then test for cointegration and finally derive impulse-response models.

TABLE 6

Modelling Procedure in Steps

1 Test for unit roots

2 If no unit roots are present, build VAR model on the level, then go to step 5. If unit roots are present, go to step 3.

3 Test for cointegration

4 If no cointegration is present, build VAR model on the differences, then go to step 5. If cointegration is present, build vector error-correction model, then go to step 5.

5 Derive impulse-response models and associated persistence levels. 4.3.2 Unit Root Test for Stationarity

It is important to know whether a time series is stationary or evolving for modelling purposes, by which stationary means that changes in data do not have a permanent impact on the data later in time and by which evolving means that changes in data are affecting data later in time, for example by establishing a new equilibrium. The latter implies that any effects are permanent. A unit root therefore implies that a portion of a shock in tweets per minute is persistent and will influence its long-run behaviour (Dekimpe & Hanssens, 1995).

We use an Augmented Dickey-Fuller Unit-root test for each variable, testing for stationarity in the level of the variable (which is the time series as such) and included an intercept in the equation. We then used a Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test to confirm the Dickey-Fuller test results.

24

concludes that Tweets per Minute is evolutionary and Television Ratings is stationary. The KPSS test, shown in table 8, confirms these results for at confidence level of 95%.

After differencing the variable of Tweets per Minute, we did find stationarity. This means that television ratings is an I(0) variable and tweets per minute is an I(1) variable. This means we need to account for stationarity issues when modelling the VAR model, so we follow the procedure of Toda & Yamamoto (1995).

TABLE 7

Augmented Dickey Fuller Test Results (Trend and Intercept)

Variable Dickey-Fuller

t-Statistic t-Statistic (1%) t-Statistic (5%) t-Statistic (10%) Probability Root Unit onary Stati-Television Ratings -2.102 -3.969 -3.415 -3.130 0,5435 No Yes Tweets per Minute -6.971 -3.968 -2.865 -2.568 0,0000 Yes No TABLE 8 KPSS Test Results

Variable _Statistic Test Critical Value _(1%) Critical Value _(5%) Critical Value _(10%) Reject h0 at Confidence

Level of 95% Stationary? Television Ratings 0.1758 0.7390 0.4630 0.3470 No Yes Tweets per Minute 0.4731 0.7390 0.4630 0.3470 Yes No 4.3.3 Cointegration

Since the data consists of time series that are integrated of different orders, a I(0) and a I(1) variable, we cannot test for cointegration, as series should be integrated of the same order (Giles, 2013). We therefore do not estimate a Vector Error Correction model but a Vector Autoregression model, according to the steps described in section 4.3.1. The model will be depicted in section 4.4.

4.4 Lag Length

Tweets per Minute is an I(1) variable, as shown in the previous section, so we take first differences of this time series to establish stationarity: this results in the variable Tweets per Minute. We then test for optimal lag length. Using the Schwarz Information Criterion and Hannan-Quinn Information Criterion, we found an optimal length of p = 6. The LR, FPE and AIC criterions advise a lag length of 14. The lag length test results are shown in table 11.

TABLE 11

VAR Lag Order Selection Criteria

Criterion LR FPE AIC SC HQ

25

As stated by Liew (2004, p.6), the Hannan-Quinn Information Criterion “is found to outdo the rest in correctly identifying the true lag length” with a “relatively large sample (120 or more observations)”. As we have 840 observations, we will choose the lag length proposed by the HQ criterion, which value is 6.

4.5 VAR(X) Model Specification

The vector of variables Tweets per Minute and Television Ratings are assumed endogenous because both graphs seem to have similar patterns. These two variables are explained by their own past and the past of the other variable. As shown in section 4.4, we will use a lag length of 6, so that we measure the lagged effects of Tweets per Minute and Television Ratings up to and including 6 minutes into the past. The VAR(X) model also includes a dummy for social media blocks with spoken encouragement to tweet, a dummy for social media blocks without this encouragement and a dummy for emotional moments. A summary of all variables in the model can be found in table 9.

TABLE 9

Variables (Per-Minute Data)

Conceptual Variable VAR Variable Endogeneity

Television Ratings TVRit Endogenous

Tweets per Minute TPM it Endogenous

Social Media Block without Encouragement SMB0 it Exogenous

Social Media Block with Encouragement SMB1 it Exogenous

Emotional Moment EMO it Exogenous

The resulting VAR(X) specification is therefore given by the following equation. [ ] [ ] [ ] [ ] [ ] [ ] [ ]

For which is the coefficient of past television ratings on the current ratings and is time in minutes.

Coefficient accounts for the effect of a difference in tweets on television ratings. Coefficient represents

the effect of past television ratings on the difference in tweets per minute and coefficient accounts for the

effect of a difference in tweets on itself. Coefficient and are the constants, and are the

coefficients for social media blocks without encouragement, and for those with encouragement. The

coefficients and account for the effect of emotional moments. 4.6 Impulse Response Functions

26

5. RESULTS

This chapter will discuss the results from the VAR model established in chapter 4. To this end, we examine VAR(X) coefficients (section 5.2), and Impulse Response Functions (section 5.3). The methodological assumptions behind these techniques are described in the previous chapter.

5.1 Model Fit

We modelled the VAR, as depicted by the equation in section 4.4, using 6 lags and found a model fit of adjusted for Tweets per Minute adjusted for Television Ratings. An value of 0.99 might point to spurious regression. However, if the Durbin Watson statistic is close to a value of 2, spurious regression is probably non-existent, according to Baumohl and Lyocsa (2009, p. 11). The Durbin Watson statistic for this VAR(X) model has a value of 1.99, so the high is not explained by spurious regression and the model is reliable.

Furthermore, Valle (2011) shows that trending time series regressions usually have high values and do not provide a good measure of model explanation. If we detrend television ratings by taking first differences I(1), we find an of 0.46, which shows a good and probable fit. We still have to use the I(0) version of the variable for VAR analysis according to the VAR procedure in section 4.3.1

The Tweets per Minute model fit of can be explained as this model only explains relationships between tweeting behaviour and watching television, while Twitter is a medium for various types of messages and hashtag #troon would also be used by people not watching television. Therefore, not all differences are explained by occurrences of social media blocks, emotional moments, lagged effects of tweets, or changes in television ratings.

5.2 VAR(X) Model Estimation

To measure the effect of emotional moments and social media blocks, with and without encouragement, we discuss the estimates of coefficients that the VAR model provides. These coefficients are shown in table 12. Significant results are marked bold.

TABLE 12

VAR Coefficients (lags p = 6)

Television Ratings (t-statistic) Tweets per Minute (t-statistic) Television Ratings at t=-1 1.546 (44.6869)* _{-0.012 (-0.267)} Television Ratings at t=-2 -0.558 (-8.684)* _{0.070 (0.825)} Television Ratings at t=-3 0.045 (0.661) -0.082 (-0.932) Television Ratings at t=-4 0.028 (0.412) 0.085 (0.961) Television Ratings at t=-5 0.101 (1.581) -0.014 (-0.167) Television Ratings at t=-6 -0.163 (-4.752) * _{-0.045 (-1.006)}

27

Tweets per Minute at t=-2 0.022 (0.869) -0.181 (-5.481) *

Tweets per Minute at t=-3 -0.066 (-2.750) -0.064 (-.2051) *

Tweets per Minute at t=-4 -0.015 (-0.635) -0.161 (-5.101) *

Tweets per Minute at t=-5 0.009 (0.365) -0.077 (-2.432) *

Tweets per Minute at t=-6 -0.002 (-0.096) -0.095 (-3.001) *

Constant 10.170 (2.246) * _{-7.700 (-1.292)}

Social Media Block without Encouragement -20.191 (-1.844) † _{77.080 (5.354)}*

Social Media Block with Encouragement 3.680 (0.251) 257.570 (13.405)*

Emotional Moment -10.780 (-0.857) 36.260 (2.192)*

* Significant at confidence level of 95% † Significant at a confidence level of 90%.

Furthermore, table 12 shows the VAR coefficients for the three dummy variables as effect on tweets per minute and television ratings, which show immediate effects. We will discuss these results in the following sections.

5.2.1 The Effects of TV Ratings

We tested for effects between the variables Television Ratings and Tweets per Minute in first differences. We anticipated that a change in television ratings did not lead to a change in tweets per minute and also anticipate the reverse, where a change in tweets per minute does not lead to a change in television ratings. We did, however, assume an effect of social media blocks on both ratings and tweets.

We find that the variable Television Ratings mainly influences itself, the first minute positively, then negatively in the second and sixth minute. The results demonstrate that television ratings are of trending nature and past data especially impacts new data in the two minutes before, along with a significant correlation at six minutes before a measured television rating.

We also find that television ratings are no significant prediction for a difference in the number of tweets, as the coefficients of all lags are insignificant in explaining tweets.

5.2.2 The Effects of Tweets per Minute

The difference in number of tweets per minute is seemingly influenced continuously over time by lagged values, though the trend is negative. This means that impulses are needed to create a propensity to tweet, other than itself.

28

There exists a significant correlation between the number of tweets sent and the occurrence of social media blocks (SMBs), as seen in table 13. The coefficients calculated in the VAR model show the short-term effect of SMBs on the difference in tweets per minute. SMBs with and without spoken encouragement to tweet respectively have a coefficient of 257.57 and 77.08 when relating to the difference in number of tweets. Table 13 shows the coefficients of individual social media blocks. One outlier can be seen at 11:11AM, where the shown social media block is shown and has a direct effect of 544.19 more tweets the first minute. Nevertheless, even when accounting for this outlier, we find that a social media block with vocal encouragement has a larger effect on change in tweets per minute than the blocks without vocal encouragement. We use a Z-test to compare the coefficients of a social media block without encouragement (77.08) and a social media block with encouragement (257.57), shown in table 12. The estimates of the standard errors are respectively 14.398 and 19.215.

We use the following equation, where h0: and h1 for :

Z =8.664, then p=0.000 which means we reject h0, so we conclude that the coefficients differ significantly.

TABLE 13

Individual SMB Coefficients

* Significant at confidence level 95% † Significant at confidence level 90%

A curious value that was found is that a social media block without encouragement to tweet has a significant coefficient of -20.19 on the number of viewers. A preliminary conclusion would state that when such a social media block is shown on-screen, the number of viewers lowers by 20,190 the minute after the start of the block. Nevertheless, because only SMB’s without encouragement have a significant effect on TV ratings and SMB’s with encouragement do not, we look at individual cases to find an explanation for this result. The initial possible explanation for the drop in viewers would be that this part of the audience dislikes twitter-related messages. When we look at table 14, however, we find a different explanation.

Timestamp Spoken Encouragement VAR Coefficient on Tweets _{per Minute in Differences}

29

TABLE 14

Individual SMB Coefficients

* Significant at confidence level 95% † Significant at confidence level 90%

We see that most individual blocks have no significant relationship to the number of viewers, except for the last social media block of the day, without encouragement, at 22:26PM. We expect that at the end of the broadcast, a social media block can be a trigger for the audience to turn off the TV, for example when going to bed, or switching channels

5.2.4 The Effects of Emotional Moments

In our model we find a significant coefficient of 36.26 regarding the direct impact of an emotional moment on a difference in tweets per minute*.3

A possible explanation for this small effect when compared to social media block effects may be that a change in the difference in tweets already provides the cause for a following change, which is the autoregressive effect of tweets. A change in emotion would then not explain any change in tweets. Furthermore, emotion might simply not be a large factor in explaining if people tweet. We will elaborate on these statements in chapter 6.

5.2.5 The Effects of Age

When testing for the effects of television ratings of 6-19-year-olds on the number of tweets, we found no significant results, comparable to television ratings of the whole population.

3_{Robustness check: to verify if the dummy variables were correctly coded, we checked if peaks in tweets occurred just outside the}

specified dummy duration, but this did not seem to be the case. We also coded different dummy variables regarding duration and starting times but found no reason to believe that the initial coding was incorrectly timed.

Timestamp Spoken Encouragement VAR Coefficient on _{TV Ratings}