To what extent do user comments explain
the effect of brand post characteristics on CTR?
An analysis of Facebook posts of six NGOs
MASTER THESIS
MSc Marketing
Intelligence & Management
To what extent do user comments explain
the effect of brand post characteristics on CTR?
An analysis of Facebook posts of six NGOs
Siri de Ruiter
MANAGEMENT SUMMARY
Although NGOs and other organizations widely use social media marketing and evaluate the effectiveness of their efforts in terms of click-through and user engagement, little is known about the drivers of these metrics and how they relate. Therefore, this study aims to investigate how certain Facebook brand post characteristics affect the click-through rate (CTR) and whether this effect can be explained by user engagement on that post. More specifically, the effects of entertaining features, vividness and interactivity in brand posts are analyzed. Concerning engagement, this study focuses on user comments since these, as opposed to other forms of engagement, contain content that can be analyzed in more detail. In particular, the volume of comments and the extent to which users ‘tag’ each other have been studied.
Brand post-level data from Facebook pages of 6 non-governmental organizations (NGOs) was used for the analyses. The results of the log-linear regression models reveal that no indirect effects have been found, which means that user comments do not explain the effects of brand post characteristics on CTR. However, comment volume and the presence of tags have a small positive effect on CTR. Furthermore, several direct effects of brand post characteristics have been found. Vividness has a negative effect on CTR, whereas it positively affects comment volume and tagging. Including a contest in a brand post positively contributes to CTR and comment volume, but hashtags, person or organization tags and questions decrease CTR. Entertaining features have no effect on either user comments and CTR. Moreover, it is demonstrated that boosting a post is effective for driving CTR. A remarkable finding is that NGOs that address health causes show lower CTR than environmental NGOs. However, the cause of this effect needs to be investigated in further research.
PREFACE
The paper in front of you happens to be my masterpiece, literally. I have been working on it from September 2017 till January 2018 in order to fulfill the graduation requirements of the Master Marketing Intelligence & Management at the University of Groningen. Although I am very glad that this project is finished now, (most of the time) I enjoyed working on it. My thesis enabled me to apply and extend the skills and knowledge that I acquired during the master, which was challenging and interesting at the same time.
There are some people who were important during the process of writing this thesis and deserve to be thanked for that. First of all, I would like to thank my supervisor, dr. Hans Risselada for his guidance during this project and for connecting me to the company that was involved in this project. Obviously, I would like to thank this company, Pieter-Paul in
particular, and its clients for their participation in this project and for allowing me to work with their data. Nick, better known as ‘my data nerd’, also deserves special thanks, as he was always willing to help me when I encountered any struggles with the data (and there were quite some…). Lastly, I would like to thank my friends and family for their support during this project.
I hope you will enjoy reading my thesis!
Siri de Ruiter
TABLE OF CONTENTS
1. Introduction 5
2. Literature review 9
2.1 Background of the conceptual model 9
2.2 Click-through rate (CTR) 12
2.3 Brand post characteristics 13
2.3.1 Entertainment 14 2.3.2 Vividness 14 2.3.3 Interactivity 15 2.4 User comments 16 2.4.1 Volume 17 2.4.2 Tagging 19
2.5 User comments and CTR 21
2.6 Control variables 22
2.6.1 Environmental vs. Health NGOs 22 2.6.2 Paid vs. Unpaid posts 23
2.7 Overview of hypotheses 24 3. Research design 25 3.1 Data 25 3.2 Operationalization of variables 26 3.3 Descriptive statistics 30 3.4 Method 32 3.4.1 Mediation 32 3.4.2 Model specification 34
3.4.3 Validation of the log-linear regression model 35
4. Results 36 4.1 Model 1 36 4.2 Model 2 36 4.3 Model 3 37 4.4 Model 4 37 4.5 Validation 40
4.5.1 Validation of the mediation effects 40
4.5.2 Robustness checks 41
5. Discussion 42
5.1 Findings and theoretical implications 42
5.2 Managerial implications 44
5.2.1 Enhancing click-through 45 5.2.2 Enhancing user comments 45
5.3 Limitations and Future research 46
1. INTRODUCTION
Marketers have placed increased emphasis on social media marketing over the years. Given their unrestricted reach and opportunities for targeting specific consumers, social media platforms have largely been adopted by organizations as part of their marketing strategy. In 2013, over 15 million brands were already present on the most popular social media platform, which is Facebook (Koetsier, 2013). This is not surprising, as Facebook reported to have 1.71 billion active users in 2016 (Tynan, 2016). Furthermore, it is suggested that Facebook influences over half of consumers’ online and offline purchases (McCarthy, 2015). Accordingly, creating effective social media strategies is of major importance for organizations. In particular, Krueger and Haytko (2015) demonstrate that social media platforms can be extremely powerful for NGOs. Social media have especially offer new opportunities for creating awareness for social causes and reaching potential contributors (Krueger & Haytko, 2015; Mano, 2014). Since NGOs are generally constrained by limited marketing budgets, they can benefit heavily from applying effective social media strategies.
A specific way in which organizations are active on social media is via establishing a brand pages on Facebook through which they can post brand messages and acquire fans. Since these posts reach a specific audience, they can be considered as a form of advertising (De Vries, Gensler & Leeflang, 2012). In a more recent study, De Vries, Gensler & Leeflang (2017) demonstrate that impressions generated by brand posts are effective means for brand building and customer acquisition purposes, which suggests that it is important for organizations that a post reaches as many relevant users as possible. An important feature of Facebook that contributes to enhancing post reach is that brand fans are entitled to engage with posts. Gensler et al. (2013) argue that user engagement contributes to brand building. Whenever brand fans engage with a post, their friends receive a notification about this action in their news feed, which further amplifies the spread of the post throughout the network. De Vries et al., (2017) find that social messages initiated by organizations, like brand posts, foster interactions among consumers and in turn influence sales.
it increases the probability that this content is relevant to other users. As a result, understanding the factors that affect post popularity is necessary to generate effective social media strategies.
To date, research on this topic has mainly studied the popularity of brand posts in terms of the amount of user engagement that the brand post has generated, which is measured by the number of likes, shares and comments (Cvijikj & Michahelles, 2013; De Vries et al., 2012; Lee, Hosanagar & Nair, 2015; Luarn, Lin & Chiu, 2015; Stephen et al., 2015; Tafesse, 2015; Trefzger, Baccarella & Voigt, 2015; Wallace et al., 2014). It was found among others that brand posts that contain entertaining features lead to more user engagement. Furthermore, the amount of user engagement varies for different levels of vividness and also interactive features affect the number of comments on a post (Cvijikj & Michahelles, 2013; De Vries et al., 2012;Luarn et al., 2015; Tafesse, 2015). These findings show that marketers can exert influence on what information about their brands is being shared among social media users by generating relevant and appealing brand posts that foster engagement (Muñiz & Schau, 2011; Parent et al., 2011).
Although likes and shares are only measurable in numbers, comments contain actual content that can be analyzed. However, the content of user comments remains an underinvestigated topic in the literature on social media marketing (De Vries et al., 2017). Since this content could contain very relevant information for brands, it is important to further investigate the role of user comments in explaining social media effectiveness. Therefore, this study addresses both the volume and the content of user comments. More specifically, Facebook users heavily utilize the ‘tagging’ feature, which enables them to share certain - assumingly relevant - content with friends by inserting their name in a comment. Although this form of engagement is often used in practice, little is known about when and why users tag each other and more importantly, what effect it has. This makes it interesting to study this type of commenting.
more useful to consider click-through as a measure of brand post success than user engagement.
In particular, Kumar et al. (2016) argue that whether a post leads to purchase behavior depends on the sentiment of the post and the user’s response to it. Therefore, placing appealing brand posts is crucial for organizations to generate click-through and engagement among social network users. Moreover, Colicev et al. (2017) find an indirect effect of owned-social media, like brand posts, on brand awareness and purchase intent through user engagement. This indicates that both the characteristics of the brand post as well as the way in which users engage with the post determine its success. Since the way in which users engage with a post is determined by its characteristics, a mediation effect is proposed. As a result, this study aims to investigate how certain brand post characteristics affect CTR and to what extent this effect can be explained by user comments on that post. In addition, the study accounts for differences between health and environmental NGOs and whether the organization paid for ‘boosting’ a post.
This is relevant to investigate, since it connects brand post characteristics to user engagement and organizational performance at the same time. Moreover, social media presence has been found to be of major importance for the success of NGOs (Krueger & Haytko, 2015), which demonstrates that gaining more insights on how to optimize their social media strategies is relevant. It would be valuable to find out whether results are similar to other types of organizations. Concerning user comments, this study contributes to understanding the effect of the number of comments by addressing it relative to the number of impressions on a brand post. Furthermore, this paper aims to close a gap in literature by addressing a recent call for further research by De Vries et al. (2017) to examine the content of user comments on brand posts to gain more insights on the underlying processes that explain the effectiveness of brand posts. This study is the first to extend literature on social media marketing by investigating tagging behavior.
classify whether user comments included tags, Graph API was used. A log-linear regression was performed on the data and bootstrapping helped to determine the significance of the indirect effects.
The results reveal that entertaining elements do not contribute to either CTR or user comments. Vividness has a negative effect on CTR, whereas positive effects were found on user comments. Moreover, limited effects for interactivity were found. Including a contest in a brand post contributes to CTR, but negative effects were found for hashtags, person or organization tags and questions. Furthermore, click-through is lower for NGOs that address health causes and it is demonstrated that boosting is effective for driving CTR. The number of comments and the presence of tags have a small positive effect on CTR. However, no indirect effects have been found, which means that user comments do not explain the relationship between brand post characteristics and CTR.
This study contributes to extant literature in several ways. First, it considers social media content, engagement behavior and organizational performance measures simultaneously, which provides new insights on how these constructs affect each other. Secondly, it extends existing literature by addressing the volume of comments relative to the number of impressions that a brand post generates, which enables researchers to draw more meaningful insights. Furthermore, this study is the first to consider the drivers and effects of user tagging behavior. The insights contribute to the successful development of social media marketing strategies for NGOs and specifically on whether and how brand post characteristics can be used to foster comments and increase CTR. Since CTR is widely used as a performance measure, marketers can benefit from understanding how to optimize this metric. Moreover, if marketers understand the effects of their social media content on user responses, they can exert influence on what is said about their brands in the online environment. This would be relevant for optimizing acquisition and brand building efforts.
2. LITERATURE REVIEW
This chapter contains an extensive discussion of existing literature on the topic of this study. First of all, relevant research findings will be presented as a means to provide background information to the readers and to introducethe conceptual model. Subsequently, each variable included in the conceptual model will be reviewed. Accordingly, this review serves as the foundation for the development of the hypotheses that will be tested later.
2.1 Background of the conceptual model
Researchers have increasingly started to investigate how organizations apply social media for marketing purposes (Swani et al., 2014) and how social media users respond to these efforts (Jahn and Kunz, 2012; Kim and Ko, 2012). Nowadays, the challenge is to understand how these marketing efforts translate to increased organizational performance. Existing literature on social media marketing merely considers businesses as the unit of analysis. However, Krueger & Haytko (2015) find that appropriate use of social media can have an enormous impact on NGO performance as well, which is why the focus of this study is on gathering insights on social media effectiveness for this specific type of organization. NGOs need to create brand awareness, brand loyalty and establish trustworthiness to receive funds and support from individuals. Therefore, a good marketing strategy is of major importance to them. One of the reasons for organizations to integrate social media into their marketing strategy is that it enables them to reach large but targeted audiences at relatively low costs. Since NGOs are often constrained by limited marketing budgets, proper targeting is crucial for them to find potential supporters. Social media are effective platforms for building community support, as people who share similar interests are connected through for instance friendships or brand pages (Saxton & Waters, 2014). This offers NGOs opportunities for targeting on a more advanced level than geographic or demographic characteristics (Van Beekum, 2017). Hence, NGOs can benefit from implementing effective social media strategies.
users can be transferred to another website. This action is called click-through. As this click expresses immediate interest in the advertised brand (Briggs & Hollis, 1997), the click-through rate (CTR) is the most commonly used measure for advertising success (Lohtia et al., 2013).
Apart from clicking on a brand post, Facebook enables users to engage with brand posts by liking, sharing, or commenting on them (Wallace et al., 2014). As such, users can immediately respond to the content that they are exposed to and interact with the organization or with other users. Van Doorn et al. (2010) define customer engagement behavior as “going beyond transactions, and may be specifically defined as a customer’s behavioral manifestations that have a brand or firm focus, beyond purchase, resulting from motivational drivers.” In the context of social media, engagement is conceptualized as “a multidimensional construct that reflects users’ brand-related cognitive, affective and behavioral actions with dimensions indicating various levels of intensity of users’ interactions with brands” (Hollebeek et al., 2014; Vivek et al. 2012). Motivations for users to engage with brand-related content are expressing their identity, having social interactions with either consumers or brands, obtain or disseminate information and entertainment (Leung, 2009; Muntinga et al., 2011; Shao, 2009). Stephen et al. (2015) demonstrate that users engage when the content of the brand post is relevant to the user’s personal goals. Furthermore, Bitter (2014) shows that user engagement is affected by a user’s self-brand relationship and interactions with friends.
fans. Hence, Facebook enables a brand to reach a large audience with only one brand post, which is one of the major benefits of social media marketing.
Another reason for creating popular content is that Facebook only wants to show relevant content in user’s news feed. As the network continues to grow, Facebook has developed an algorithm to determine whether a post is relevant to a user. Part of this algorithm entails the popularity of the content, which is determined by the amount of user engagement a post generates (Facebook, 2017). The higher the predicted relevance of the post, the more people will be exposed to the post. Hence, it is valuable for organizations to create engaging content to enhance the number of impressions. As a consequence, marketers attempt to popularize brand content by implementing features that encourage users to engage with a post.
Whereas engagement is currently measured in terms of volume, comments can contain much more information. Brand post characteristics have been found to affect comment volume, but De Vries et al. (2017) suggest that it would also be worthwhile to investigate the content of comments to obtain insights on what consumers talk about and how engaging brand posts are. Kumar et al. (2016) argue that the effect of content created by organizations depends on the sentiment of the message and the user’s response to it. This indicates that both the characteristics of the brand post as well as the way in which users engage with the post determine its success. Since the way in which users interact with a post is determined by the characteristics of that brand post, this research aims to investigate how certain brand post characteristics affect CTR and whether this effect can be explained by the way in which users comment on that post. Thus, it will be examined whether user comments have a mediating effect on the relationship between post characteristics and CTR. Consequently, the conceptual model is proposed as follows in Figure 1. Each component included in the model will
be discussed individually in the remainder of this chapter. Based on that, hypotheses will be formulated.
2.2 Click-through rate (CTR)
The most commonly used metric to measure online advertising success is CTR (Lohtia et al., 2003). A click refers to “a user-initiated action of clicking on an ad element, causing a redirect to another web location” (PriceWaterHouseCoopers, 2001). Often this ‘other web location’ is the homepage of the advertised brand or a page that relates to the content of the ad. A link does not necessarily have to be a hyperlink, it could also be a clickable element integrated in the post. CTR has been defined by Lohtia et al. (2013) as “the ratio of the number of times an advertisement is clicked relative to the number of advertisement impressions.” This metric is widely used since it reflects a user’s behavioral response to the advertising message and indicates immediate interest in the advertised brand. Also, it is easy to observe, measure and interpret (Briggs & Hollis, 1997). Pavlou and Stewart (2000) consider CTR as a predictor of brand interest and potential purchase or contribution behavior, as a user deliberately chooses to leave the website he or she is browsing in order to collect more information. If the goal of an NGO’s social media strategy is to increase interest and funding, CTR can be considered as a better measure of brand post success than the amount of engagement a brand post generates (Dorrington, 2016). Therefore, this study uses CTR as a measure of brand post success.
To date, research has mainly considered CTR in the context of search engine advertising and banner advertisement, whereas it is also very applicable to brand posts (De Vries et al., 2012). Briggs and Hollis (1997) have identified that CTR can be explained by audience-related and advertising-related factors. Audience-related factors entail the immediate relevance of the message to the audience and the involvement that the ad creates. These comprise interest in the category, attitude towards the context in which the ad appears and familiarity with the brand. It was found that people who are more involved with the product category or have a more favorable attitude are more likely to click on the ad (Cho, 2003). This suggests that proper targeting improves CTR (Briggs & Hollis, 1997; Chandon et al., 2003; Chatterjee et al., 2003). Also, familiar brands receive double the click-through as opposed to unfamiliar brands (Dahlén, 2001).
2.3 Brand post characteristics
As social media offer both organizations and its customers new ways of interacting with and between each other, Mangold and Faulds (2009) argue that social media communication generated by organizations is an essential part of their promotion mix. A variety of research has discovered that brand posts affect brand awareness and consideration, brand loyalty and acquisition (De Vries et al., 2017; Kumar & Mirchandani, 2012; Rapp et al., 2013). Furthermore, Fulgoni (2016) finds that these messages can complement traditional marketing efforts if they spread through the social network.
Compared to offline advertising, the online context offers a higher degree of variety regarding advertising content, presentation forms (vividness) and interactivity. These aspects have been classified as design features and it is argued that these features might elicit different responses from the people exposed to the ad (Tang et al., 2015). Online ads have been found to affect users’ cognitive, affective and behavioral responses (Eroglu et al., 2001). According to Mansfield (2014), NGOs need “compelling content in order to evoke empathy and trigger the impulse in donors and supporters to take action.” However, a clear view of what compelling content entails, is lacking. Since limited research has been conducted on the effects that brand post characteristics have on CTR so far, literature on the effectiveness of banner advertising is used to develop hypotheses. Although banner ads and brand post are not the same, De Vries et al. (2012) argue that there are similarities between two which suggests that literature on banner advertising applies to brand posts in relation to CTR as well. More specifically, both are small online advertisements that people can click on. Because of their size, a major challenge is to attract attention from the audience that is exposed to the ad. Therefore, special characteristics or features need to be integrated in the message to capture users’ attention (De Vries et al., 2012). A key difference between the two is that banner ads show up involuntarily whereas brand posts appear in the news feed of brand fans who indicated their interest in the brand. However, this notion does not hold for users who are exposed to the brand post through engagement by friends or when the organization has paid to spread the brand post.
suggests that more research is needed. In the remainder of this section, each of these characteristics will be discussed individually.
2.3.1 Entertainment
One of the reasons why users are active on social networks is for its entertainment value (Lin & Lu, 2011; Park et al., 2009). Entertainment provides people with an opportunity to fulfill user needs like distracting and diverting themselves, aesthetic enjoyment and emotional release (Bronstein, 2013; Muntinga et al., 2011). Entertaining brand posts include content that does not refer to the brand or a particular product, rather they include features like humorous videos, teasers, slogans or wordplay (Cvijikj & Michahelles, 2013). Raney et al. (2003) argue that information that is enriched with entertaining elements is usually more positively evaluated by users and it leads to higher website revisit intentions compared to information without entertaining elements. Previous studies (Cvijikj & Michahelles, 2013; Lin & Lu, 2011; Park et al., 2009) demonstrate that entertainment is the most important factor in explaining behavior of social network users and engagement with a brand post. Hence, it is expected that users who are exposed to an entertaining post are more inclined to click on it, as it is congruent with their motivations for being online.
H1a: Brand posts containing entertaining features generate higher CTR. 2.3.2 Vividness
Secondly, the brand post characteristic vividness will be discussed. Nisbett and Ross (1980) define vivid stimuli as 1) emotionally interesting, 2) concrete and image provoking and 3) proximate in a sensory temporal or spatial way. In the context of brand posts, vividness refers to the extent to which the post stimulates different senses (Steuer, 1992) or the extent of media richness (Cvijikj & Michahelles, 2013). Shrum (2002) argues that vividness “captures the ability of content to depict a situation in ways that approximate reality.” Hence, brand posts including a video are considered to be more vivid than a photo, since it stimulates sight and hearing and it seems closer to reality. In turn, photos are more vivid than post containing just text. Therefore, different brand posts can have different levels of vividness.
the content of the video. However, vividness has been found to positively affect audience response on websites (Coyle & Thorson, 2001) and brand posts (Sabate et al., 2014). The authors attribute this effect to the ability of vivid content to represent a virtual brand experience that is close to reality. Cho (1999) argues that highly vivid banners increase users’ intention to click as it captures attention from the rest of the screen. Similarly, Lohtia et al. (2003; 2007) and Chua and Banerjee (2015) find that highly vivid banners promote CTR. Based on these argumentats, the following hypothesis is proposed:
H1b: The higher the level of vividness in a brand post, the higher the CTR. 2.3.3 Interactivity
The third brand post characteristic is interactivity, which can be defined as “the degree to which two or more communication parties can act on each other, on the communication medium and the message, and the degree to which such actions are synchronized” (Liu and Shrum, 2002). Steuer (1992) expresses interactivity as "the extent to which users can participate in modifying the form and content of a mediated environment in real time." Interactive brand posts thus include elements that encourage consumers to interact with the brand and with each other. For instance, brand posts containing a question or a call to action are more interactive than brand posts which only contain text, as they encourage consumers to answer the question or perform an action. Other interactive elements that have been defined are open spaces, hashtags, page or person tags, contests, voting options and links (De Vries et al., 2012; Tafesse, 2015).
Mand (1998) found that interactivity of banner ads has a substantial effect on CTR. Whereas interactive features are meant to trigger a behavioral response from users who are exposed to an ad, it is argued that redundant interactivity leads to distraction and should therefore be avoided. Similarly, Lohtia et al. (2003) find that higher degrees of interactivity lower CTR. Tang et al. (2015) strengthen this finding as they demonstrate that ads containing interactive elements lead to avoidance behavior. An explanation for this negative effect would be that the goal of including interactive features is not to drive users to the website, but instead to encourage users to engage with that ad. Therefore, including interactive features is expected to be more important for creating interaction with and among users and driving virality of a post than for generating clicks. Hence, it is expected that:
2.4 User comments
According to Muñiz and Schau (2011), marketers need to induce users to engage with brand posts by providing relevant content and information. This is valuable for organizations, since the level of engagement with brand pages in terms of likes, shares and comments popularizes brand content and spreads the content through users’ network of friends (Lipsman et al., 2012). Various researchers have identified brand post characteristics that contribute to explaining the amount of engagement a brand posts provokes. In these studies, engagement is used as a measure of brand post popularity and success (Cvijikj & Michahelles, 2013; De Vries et al. 2012; Lee, Hosanagar & Nair, 2013; Luarn, Lin & Chiu, 2015). Moreover, psychological motivations for users to engage with content have been identified. These internal triggers include among others the need to share information and to express self-identity, expertise, feelings, emotions, excitement or concern for others (Swani et al., 2017)
Whereas likes and shares are only ‘clicks’ on a button which involve relatively little effort, placing a comment on a post requires deeper levels of engagement as the user actively responds to the content of the brand post (Sabate et al. 2014; Swani et al., 2017). Also, commenting allows users to publicly voice their opinion. Research has shown that brand fans who are involved with creating user-generated content (UGC), such as comments on Facebook, are likely to be brand advocates. As a consequence, user comments can be regarded similar to electronic Word-Of-Mouth (eWOM) communication (Cvijikj & Michahelles, 2013; Daugherty, Eastin & Bright 2008; De Vries et al., 2012). Furthermore, it is likely that these expressions of engagement induce Facebook friends of engagers to engage with the content as well, which further amplifies the spread of a post (Babic Rosario et al., 2016; Kumar et al., 2013; Schivinski et al., 2016). Comments are classified as ‘earned’ social media (ESM), since it generates exposure for the brand that is not directly initiated by the brand itself (Fulgoni, 2016). According to Colicev et al. (2017), ESM can be influenced by owned social media (OSM), which refers to the brand post.
To date, several studies have included the volume of comments as a predictor of brand post popularity (Cvijikj & Michahelles, 2013; De Vries et al., 2012; Stephen et al., 2015; Trefzger et al., 2015) and De Vries et al. (2012) also consider valence of comments in their study. From these papers, it appears that user comments on brand posts are driven by the content and design of the brand post (Berger & Milkman, 2012; Lohtia et al., 2003, Wallace, 2014). Fulgoni (2016) and De Vries et al. (2017) address the need for metrics that enable measuring user comments on social networks as a direction for further research. Subsequently, this study elaborates on commenting activity by not only extending literature on comment volume, but also by making a first step in addressing its content. A popular action that Facebook users take to share certain posts with friends is tagging behavior. Although this form of engagement is often used in practice, little academic research has been conducted on it in the field of marketing. This makes it interesting to investigate the drivers and the effects of tagging behavior and the volume of comments in this study.
2.4.1 Volume
The volume of comments reflects the total number of user comments a brand post generates. Various studies have used the volume of comments as a measure of brand post popularity (Cvijikj & Michahelles, 2013, De Vries et al., 2012; Stephen et al., 2015), since brand post popularity has been found to play a role in causing behavior. However, the absolute comment volume does not say much without any context. The number of comments is for instance dependent on the number of brand fans and the number of people that are reached by a post. Facebook pages with 200 fans obviously reach fewer users than Facebook pages with 20,000 fans. Therefore, Dawley (2016) argues that it is more valuable to assess the number of comments relative to another measure. Therefore, this study measures the number of comments relative to the number of impressions, which is similar to click-through. Due to the limited availability of research that considers this share of comments, literature on comment volume will be reviewed to develop hypotheses.
posts more appealing to engage with than other types of posts. This could be explained by the fact that they evoke feelings and emotions, which are motivations for users to engage with content. Akpinar and Berger (2017) confirm this idea by finding that the use of emotional appeals, such as entertainment, motivates users to place more comments, since it induces deeper engagement. Furthermore, entertaining content contributes to fulfilling user needs for hedonistic pleasure and emotional release, which makes the content relevant to the user and therefore he or she might be more inclined to actively process the post and engage with it (Muñiz & Schau, 2011). This leads to the following expectation:
H2a: Brand posts containing entertaining features lead to a higher share of comments.
Vivid elements might lead to engagement among users, because of its impact on various senses (Coyle & Thorson, 2001).Luarn et al. (2015) find that people are more likely to engage with brand posts containing a medium level of vividness (photo and text) than posts with low or high levels of vividness. Cvijikj & Michahelles (2013) show in their study that content with lower levels of vividness (statuses and photos) are perceived to be more attractive than content containing higher levels of vividness (videos and links). Moreover, they find that photos received fewer comments than status posts. In contrast, Chua and Banerjee (2015) find that higher levels of vividness triggered commenting activity. Stephen et al. (2015) find no effect of including photos or videos on the number of comments that a brand post generates. Hence, there is no consensus with regard to the effect of vividness on the number of comments. However, the variation in findings can partly be attributed to differences in measurement. Combining the results of the previously discussed studies, it is proposed that medium levels of vividness (photos) have the largest impact on the number of comments. An explanation for this is that they contribute more to attracting attention from consumers than brand posts with low vividness (status updates) and that they do not distract attention as much as brand posts with high levels of vividness (videos). Moreover, it is expected that posts with videos generate more comments than status posts, since they attract more attention from the audience. Hence:
Since brand posts can differ in their degree of interactivity, it is likely that brand posts containing higher levels of interactivity, such as questions or contests, perform better in terms of generating engagement among the audience. This intuition is confirmed by research from Luarn et al. (2015) as they show that brand posts containing high levels of interactivity lead to higher levels of engagement than brand posts with medium or low interactivity. Moreover, they find that interactivity is especially effective in eliciting comments. De Vries et al. (2012) demonstrate similar results, brand posts containing high levels of interactivity are positively related to the number of comments. Similarly, Lee et al. (2013) find that including blanks or a question in a brand post increases comments. According to Stephen et al. (2015), asking questions in a post has a positive effect on comments as well. However, Cvijikj and Michahelles (2013) find a negative effect between interactivity and engagement and explain this by the notion that Facebook is mostly used in short sessions, and therefore engagement with highly interactive posts would require too much time for users. However, this opposing effect might be attributed to the difference in measurement compared to the other studies that have been discussed. The majority of research suggests that higher levels of interactivity lead to a higher number of comments, which makes sense since the objective of this brand post characteristics is to trigger a response from consumers. Especially highly interactive features such as questions or contests are often included in posts to elicit commenting activity. It is therefore expected that:
H2c: Brand posts containing higher levels of interactivity lead to a higher share of comments. 2.4.2 Tagging
Tafesse (2015) argues that sharing behavior contributes to the reach and impact of brand posts. As the information in the brand post is shared with friends who receive a notification about this, the likelihood that they will see the content increases due to curiosity (Oeldorf-Hirsch & Sundar, 2015). Moreover, Hess and Ring (2016) argue that eWOM is perceived more credible and thus more influential when they originate from personal sources. This might be similar for comments that include tags.
Reasons for users to share brand posts comprise self-presentation purposes, sharing useful information with others and deepening social connections (Berger & Milkman, 2012; Wojnicki & Godes, 2008). Oeldorf-Hirsch and Sundar (2015) find that users who tagged their friends experienced a greater sense of community than those users who did not, suggesting that tagging is used to involve other people in what is perceived to be relevant content for them. Consequently, tagging can be considered as a form of targeting. Whereas organizations usually target prospects themselves, tagging enables users to target other users. Tagging behavior is therefore considered to be highly valuable for marketers. Moreover, when a person is tagged, this action is shared in the news feed of friends of the tagged person, which might generate a snowball effect among potentially interested users.
With regard to entertaining features, Luarn, Lin and Chiu (2015) find that these exhibit a higher number of shares than other content types. This finding is consistent with Cvijikj and Michahelles (2013) who propose that brand fans find entertaining posts more appealing to share with friends than other types of post. Various studies have shown that content that triggers emotion tends to be shared more (Araujo et al., 2015). More specifically, Yuki (2015) finds that happiness-evoking content has most impact on sharing behavior. This corresponds to the notion that people use social media to enjoy themselves. As entertaining content contributes to feeling happy and positive emotions, it is likely that entertaining content leads to more tagging. Stephen et al. (2015) confirm this idea as they find a positive effect of humorous content on the number of shares. Furthermore, Berger & Milkman (2012) argue that sharing entertaining content contributes to entertaining others, which can help to deepen social connections. Therefore, it is posited that:
as the level of vividness increases, it is expected that the content becomes more interesting and attention drawing. This increases the likelihood that users will share the content with friends. It is proposed that this holds for tagging as well:
H3b: The higher the level of vividness, the more positive the effect on tagging activity.
With regard to interactivity, literature on brand post sharing finds that people are more likely to share highly interactive content than content with medium or low levels of interactivity (Luarn et al., 2015). More specifically, they find increased chances of sharing when a link is included in a post. Similarly, Stephen et al. (2015) find that including URLs in a brand post have a positive effect on the number of brand post shares. However, Tafesse (2015) demonstrates a negative effect of interactivity on the number of shares. An explanation for this might be that interactive features increase the complexity of a post, leading to lower audience response. However, also the measurement in this study is different than in the previously discussed studies, since tallying is used to determine the degree of interactivity. As the general notion of including interactive features in a post is positive, as long as it is not overwhelming, it is proposed that:
H3c: Brand posts including interactive features have a positive effect on tagging activity. 2.5 User comments and CTR
Stephen and Galak (2012) find positive effects of earned social media on sales. Moreover, Colicev et al. (2017) report that earned social media positively affects user mind-set metrics and firm value. According to Stephen et al. (2015), placing a comment requires active processing of the brand post and consequently it increases the effectiveness of the message in affecting behavior. As discussed earlier, comments can be considered as a form of eWOM. Pauwels et al. (2016) find that eWOM is more important than paid marketing in driving online store traffic, which suggests that user comments indeed affect CTR. Liu et al. (2017) state that eWOM affects purchase decision-making. Therefore, commenting activity might play a partially mediating role in the relationship between brand post characteristics and CTR. This effect is expected to be partial, since comments are not the only factors that explain the relationship between brand post characteristics and CTR. Other factors that have been found to influence this relationship are for instance other forms of user engagement such as likes and shares and audience-related factors (Berger & Milkman, 2012).
enhances reach of a brand post (De Vries et al., 2012; Sabate et al., 2014). Since friends of a fan are likely to be somewhat similar, it could be argued that they share interests. Therefore, friends might be more likely to click. Furthermore, Colicev et al. (2017) find that earned social media volume, positively affects user interest in the brand and purchase intent. Given that CTR serves as a predictor for these two constructs, it is suggested that comment volume positively affects CTR as well. Moreover, Pauwels et al. (2016) argue that users might develop perceptions about brands based on the popularity of posts. If many users comment on a post, another user who sees the post might be more inclined to click on it based on the expectation that other users perceived the content to be interesting. Hence, the following hypothesis is proposed:
H4a: The share of comments has a positive effect on CTR.
De Vries et al. (2012) find that customer-to-customer social messages are effective in increasing preference and acquisitions. Since tagging entails that a user involves a specific other user in the online content, it is likely that the content he or she is exposed to is relevant to them. Furthermore, eWOM from a personal source is perceived as more credible and thus more influential in determining user behavior (Brown & Reingen, 1987; Duhan et al., 1997). Hence, it is expected that tagging behavior positively affects CTR and therefore:
H4b: The share of tags has a positive effect on CTR. 2.6 Control variables
Finally, the conceptual model takes into account that two variables that might affect the proposed relationships. As these variables have only been added to the model to control for their potential effects, no hypotheses have been developed.
2.6.1 Environmental vs. Health NGOs
for the way in which they operate and market themselves. NGOs want to induce consumers to give a donation, which implies that they need to convince them to do so. As opposed to businesses, consumers do not immediately receive something in return for their donation, which is the case if consumers buy a product. Therefore, it is argued that consumer motivations for supporting an NGO differ from motivations to involve themselves with businesses. Since a major challenge for NGOs is to find users who are interested in supporting them, it is necessary that they possess insights on how to reach and attract those users.
This study considers data from environmental and health NGOs. People might hold different attitudes towards the environment than to problems that are related to one’s health. A specification between these two types of NGOs is therefore included in the model in order to control for possible differences between how users respond to social media marketing efforts of the type of NGO. It is not necessarily expected that there will be a difference, but if there appears to be a difference in user response this would be relevant for the NGOs that participate in this study. This specification is included in the model as a control variable, as there is no theoretical ground for estimating separate effects and this is not a focus of the study. Hence, the reason to include this variable is to found whether there are any effects
2.6.2 Paid vs. Unpaid post
Facebook decomposes brand post reach into organic, viral and paid reach. Initially, organizations place an ‘organic’ brand post on Facebook, which means that the organization does not pay for placing the brand post. The organic post can generate ‘earned’ or ‘viral’ impressions when users engage with the content (Fulgoni, 2016). In that case, the post spreads through the network and reaches other users without the organization paying for it.
Facebook also offers the opportunity to promote brand posts. This action is called ‘boosting’ and the impressions generated by this promotion are classified as ‘paid’ instead of organic. When an organization pays to boost a post, the post still looks the same except a label is added to it in order to indicate that the post is ‘sponsored’. Organizations can use this feature to reach an audience that is likely to be interested in the brand but not a fan of the brand page yet. The organization can determine how the post should be boosted based on target audience characteristics like age, gender, interests or geographic location. Moreover, a budget and a time slot for boosting can be specified (Facebook, 2017).
this have implications for organizations’ marketing strategies. Especially in the case of NGOs, which generally have limited marketing budgets, it is interesting to know whether boosting brand posts is effective or not. For instance, if a larger audience is reached with a boosted post, but it does not lead to higher engagement or CTRs, there is no incentive to boost the post and the organization could in turn allocate their budget more efficiently. Furthermore, it could indicate that viral reach might be more effective in attracting interesting prospects to the brand, since these users are likely to share similar interests with brand fans. This would emphasize the need for creating content that fosters engagement.
2.7 Overview of hypotheses
Many variables and their expected effects have been described in this chapter. Therefore, an overview of all the hypotheses that have been developed is provided below.
H1a: Brand posts containing entertaining features generate higher CTR. H1b: The higher the level of vividness in a brand post, the higher the CTR. H1c: Brand posts containing higher levels of interactivity lead to lower CTR.
H2a: Brand posts containing entertaining features lead to a higher share of comments.
H2b: Brand posts containing medium levels of vividness generate a higher share of comments than brand posts with high levels of vividness and both generate a higher share of comments than brand posts with low levels of vividness.
H2c: Brand posts containing higher levels of interactivity lead to a higher share of comments. H3a: Brand post containing entertaining features have a positive effect on tagging activity. H3b: The higher the level of vividness, the more positive the effect on tagging activity. H3c: Brand posts including interactive features have a positive effect on tagging activity. H4a: The share of comments has a positive effect on CTR.
3. RESEARCH DESIGN
This chapter elaborates on how the research has been designed. First, the data will be discussed, followed by the operationalization of the variables. Subsequently, descriptive statistics of the data will be provided. Lastly, the method that was used to test the hypotheses will be presented.
3.1 Data
Facebook brand post data of six non-governmental organizations (NGOs) was empirically investigated to test the hypotheses. Data from this platform was used in this study since it is the most widely used social network among consumers and organizations. Furthermore, previous studies have shown the suitability of Facebook as the subject of study when investigating user engagement (Cvijikj & Michahelles, 2013; De Vries et al., 2012; Stephen et al., 2015; Swani et al., 2017; Tafesse, 2015). The data was provided, with permission of the NGOs, by a Dutch growth marketing agency that cooperates with these organizations. The dataset includes data on individual brand posts that were posted between 1 October 2016 and 30 September 2017. Together, the NGOs placed a total number of 2,083 posts on their brand pages during this period.As this study aims to explain which factors drive behavior of users who are exposed to a brand post, only brand posts that include a link were used for further analysis. It is assumed that the goal of brand posts that include a link is to generate traffic to a website, which corresponds to triggering a certain behavior from users. Brand posts that do not contain a link are assumed to be more focused on creating awareness or building relationships with users. The most common metric to measure potential behavior and interest in a brand is CTR (Pavlou & Stewart, 2000). In order to be able to acquire this metrics, a link is needed, since a brand post cannot generate clicks without a link. From the total number of posts in the dataset, 1,003 posts included a link to an external website. Facebook Insights automatically collects brand post data including metrics like the number of comments, impressions and link clicks that a post generates.
method is limited to a maximum of 1,000 comments per brand post. Of the 1,003 brand posts that are included in the dataset, only six brand posts generated more than 1,000 comments in total. Another explanation is that this method only extracts comments that are a direct response towards a brand post. Hence, responses to other users’ comments are excluded from this dataset (see Appendix I). At first sight, this might be considered as a limitation of the study. However, since the aim is to measure the effect of the brand post characteristics on tagging behavior, it can be beneficial that only the direct responses to the brand posts are measured. This means that the comments included in the analysis are primarily driven by the brand post characteristics and not by other users’ comments. It could however occur that users have responded to other users’ comments without using the ‘reply-to’ option, but the nature of the data does not allow to account for this. (including responses to other comments). Important to be aware of is that these excluded comments could have an impact on CTR for which the model does not account.
3.2 Operationalization of variables
This study addresses success of a brand post in terms of its click-through rate (CTR). This variable is composed by dividing the number of link clicks by the total number of post impressions. As this results in relatively small numbers, percentages are used for convenience in further reporting. Facebook posts include a link when they contain a clickable element. This element could either be a hyperlink in the textual part of the post or it could be a separate clickable element integrated in the design of the brand post. Hence, a link is not necessarily clickable text in a brand post, it could also be an image or video. Furthermore, the two types of links can be combined within one post (see Appendix II). Facebook Insights only reports the total number of link clicks that a post generated, meaning that there is no specification on the number of link clicks that different clickable elements within one post generated. This is for instance the case when a post contains multiple hyperlinks or a combination of a clickable element and one or multiple hyperlinks. Only 58 (5,8%) brand posts in the dataset contained both a hyperlink and a clickable element and therefore this is not assumed to be problematic.
The model further accounts for the share of paid impressions by including a control variable for boosted posts.
Apart from the brand post data that can be retrieved from Facebook Insights, all the brand posts were manually coded based on whether they contained entertaining, vivid and interactive features. This was done manually, since the default data that Facebook reports does not provide sufficient information to classify posts. Furthermore, the content of photos and videos is not visible in the dataset and the detection of certain brand post elements like wordplay are often context based which is hard to detect in an automated manner. In order to classify posts, the provided URL that directs to the brand post was opened and accordingly the brand post was evaluated. The researcher executed this procedure twice for the whole data set and then compared the classification results to account for the subjectivity of the researcher. Inconsistencies could occur due to misinterpretation or typing errors. In total, 56 inconsistencies were detected and these posts were re-evaluated and recoded.
A brand post is regarded as entertaining when it includes features like humorous videos or images, anecdotes, teasers, slogans or wordplay (Cvijikj & Michahelles, 2013; Luarn et al., 2015). As such, this variable is binary coded based on the presence of any of these features.
For vividness, three different levels (low, medium, high) have been determined based on reviewing and combining previous research. Vividness is low when a post only includes a status update, this corresponds to a post with just text. The medium level of vividness refers to the presence of a photo or image in a post and high vividness means that the post contains a video (Cvijikj & Michahelles, 2013; De Vries et al., 2012; Luarn et al., 2015; Tafesse, 2015). Furthermore, Facebook classifies posts with separate clickable elements as ‘links’ by default, regardless of whether there is a photo or video integrated in this element. Therefore, all brand posts that were classified by Facebook as a link needed to be checked and recoded if necessary.
mentioned. Examples of calls to act are ‘Read more: ...’, ‘Donate now via: ...’ or ‘More information can be found here...’. Examples of questions directed towards the audience are ‘What is your reason to…?’, ‘Do you want to be…?’ or ‘Are you…?’ Since these three interactive features directly serve to enhance user engagement and links clicks, they are considered to be highly interactive.
For both vividness and interactivity, the low level is used as the base category in the analyses. An overview of the operationalizations can be found in Table 1. Important to note is that previous studies also include links as an interactive feature (Cvijikj & Michahelles, 2013; De Vries et al., 2012; Luarn et al., 2015). However, this feature becomes redundant to take into account in this study, since only brand posts that contain a link are considered. In addition, it has been argued that including hyperlinks in a post might be obsolete because it is so widely used that people have become blind for this form of interactiveness (Liu et al., 2017). Moreover, Lei et al. (2017) find no relationship between the presence of a link and the number of comments in their study.
With regard to user comments, the volume of comments entails the total number of comments that a post generated and this metric is provided by Facebook Insights. This metric includes all comments that a post generated, so also responses that users gave to other users’ comments on a post. As it was found that the number of comments is positively correlated to the number of impressions (r = 0.58, p < 0.001), this study considers the effect of comment volume relative to the number of impressions. To account for this, a share of comments is calculated for each brand post by dividing the number of comments by the total number of impressions. This variable is called ‘Share of Comments’.
number of comments including a tag by the total number of comments included in this dataset. Accordingly, the shares were integrated into the total dataset. Ideally, one would account for names that are mentioned in the content of a brand post, so that these are not classified as tags when they appear in comments. For instance, if people respond to a brand post that promotes a blog post from person X, those comments that include name X should not be classified as a tag. Due to time constraints, this was not possible, but follow up studies might consider this. When referring to this variable, the name ‘Share of Tags’ will be used from now on.
3.3 Descriptive statistics
The average number of brand posts per NGO is 167.17. However, the highest number of post placed during the observation period by one NGO is 521, whereas the lowest number of post is 10, which demonstrates that the NGOs in the study differ in how active they are placing posts with links on Facebook. The other four NGOs all placed around 100 posts including a link. Furthermore, there were no missing values detected in the total dataset.
From the total number of posts that were used for analysis, 152 (15.15%) posts contained entertaining features such as wordplay and humorous pictures. With regard to vividness, the majority of posts, 848 (84.55%), included a photo. Calls to act (66.90%) and questions (45.76%) were the most used interactive features. Since multiple interactive features can be present within one brand post, the sum of the percentages in Table 2 which are related to interactivity exceeds 100%.
Table 2: Descriptive statistics of the independent variables
The average number of comments that an individual post generated is 51.53 (SD = 144.06), but the maximum number of comments on a post equals 2,182. Furthermore, the average percentage of users commenting on a post is 0.06% of the number of impressions, whereas the maximum value is 0.48%. Hence, there is large variation in the number of comments that a post evokes.
Considering tagging behavior, it appears that 66.5% of the brand posts included comments with tags, which demonstrates that this feature is indeed popular among Facebook users. This supports the relevance of investigating this variable. The share of tagging per brand post fluctuates heavily(Appendix IV), but if the brand post generated tags, on average 39.2% of the user comments included a tag.
(7.48%) are boosted and of these posts, the average share of impressions that is acquired by boosting the post is 0.64 and the maximum share is 0.98.
Since the dependent variable CTR is calculated based on link clicks and impressions, a scatter plot of these variables is presented in Appendix III for more insight into the composition of this variable. The majority of posts achieved between 0 and 250.000 views and between 0 and 2500 link clicks. Moreover, the average CTR is 0.46% (SD = 0.53), but the highest CTR is 4.68%. The boxplot in Figure 2shows that there are few relatively high CTR values. It considers these values as outliers, but it might be that there are reasons for why these values are so high. Since it is unlikely that these outliers can be attributed to errors in the dataset, they are further inspected. Table 3 provides an overview of the descriptives for the mediators and the dependent variable.
Table 3: Descriptive statistics of the mediators and the dependent variable
When taking a closer look at the 20 posts that generated more than 2% CTR, it strikes that 19 of the posts include a separate clickable element and 19 posts originate from environmental organizations. Furthermore, it appears that 18 of the posts included a photo or video and 13 contain high levels of interactivity. Five posts were boosted. However, no clear pattern across these posts can be detected immediately. Therefore, these values are not adjusted or removed from the dataset, as the analysis might reveal why these posts score high on CTR. For an overview of the descriptive statistics of all variables, please refer to Appendix V.
1936). Therefore, all correlations were calculated with Pearson’s R and integrated in Appendix VI. The reason behind this is that the correlations are merely used for understanding of the associations between variables, they are not considered to be crucial for the analyses.
Weak correlation is detected between share of comments and CTR (r = .19, p < .001). Also, share of tags and CTR correlate weakly (r = .13, p < .001). Furthermore, it appears that there is weak, but significant correlation between the two mediators (r = .22, p < .001). If they would strongly correlate, modifications like merging the two variables would be necessary to measure their effects. Since the variables are both related to user comments, this makes some correlation between the two plausible and therefore the mediators will be considered as individual constructs in the model.
An association that is worth mentioning is the type of NGO and CTR (r = -.34, p < .001), as it indicates that the two might be causally related as well. Also remarkable is the absence of correlations between the share of comments and other variables. It would be expected that the independent variables would be at least to some extent associated with this variable as existing research has consistently found effects of brand post characteristics on comments. In general, it appears that correlations between the variables are weak. Appendix VIprovides a correlation matrix of all variables included in the analysis.
3.4 Method 3.4.1 Mediation
conditions that should be met by the correlation between the mediators and the outcome variable. The first condition refers to directionality, which addresses the need for other than statistical evidence for the direction of the relationship between mediators and outcome variables. Theoretical ground to support the statistic model that is tested for mediation is therefore important. The model in this study is based on existing literature, which means that this condition is met. The second condition entails reliability, meaning that measurement errors in the mediators and outcome are ignorable. Since most of the data is retrieved directly from Facebook, it is unlikely that errors exist in the number of comments and CTR. The classification of tagging is subject to some measurement errors, which means that reliability is not optimal. The unconfoundedness condition specifies that variables omitted from the model have ignorable effects. Since the model measures a limited amount of brand post characteristics and focuses only on explaining user comments as a form of engagement, this might be an issue. However, a RESET-test (see Section 3.4.3) reveals that this is not the case. Furthermore, sensitivity analysis is a performed to deal with this condition. Whereas these first three conditions address the unbiasedness of the indirect effect, the remaining three conditions cope with the size of the indirect effect. Distinctiveness refers to whether the mediators and outcome are distinct constructs. This can be achieved through high reliabilities of the variables, small true correlations and large sample sizes. Fifth, the power condition states that larger effects and sample sizes contribute to increasing the statistical power of finding true non-null effects. This study meets these conditions as the sample size is large (n = 1003) and correlations between the variables are small (Appendix VI). However, small correlations might also indicate problems with regard to the mediation effect, since it implies that associations between variables are small. Therefore, the effect sizes might turn out to be low. Lastly, the mediation condition implies that the effect of the independent variables on the outcome is transmitted via the mediator(s). This is reasonable when the previous five conditions are met and when the indirect effect is significant. Section 5.4.1 will elaborate on how to determine when an indirect effect is significant.
3.4.2 Model specification
In order to test the hypotheses that have been developed and whether mediation is present, log-linear regression analyses have been performed. The reason for this is that the dependent variable and the mediators have been integrated in the dataset as shares that range from 0 to 1 and therefore they are proportional variables. Since only a small proportion of the total audience that is reached by a post clicks or comments on a post, these variables show probability masses close to 0. This means that the variables follow a positively skewed distribution, and therefore deviate substantially from the normality line (see Appendix III and IV). To account for this nonnormality, the variables CTR, share of comments and the share of tags have been log transformed. Before the transformation, the 0 values in these variables have been changed to 0.0001. This change is made because the original variables can take a value of 0, but taking the log from 0 is not possible, which would lead to missing values (Benoit, 2011).
The model that explains the direct effect of the brand post characteristics on CTR and the mediators can be expressed as:
where
yij dependent variable i (CTR, share of comments or share of tags respectively) for brand post j
Entertainmentj dummy variable indicating whether the entertaining features are present in brand post j (baseline is no entertaining characteristic) Vividnessgj dummy variables indicating whether the vivid element g at brand
post j is present or not (baseline category is no vividness)
Interactivityhj dummy variables indicating whether the interactive characteristic h at brand post j is present or not (baseline category is no interactivity) Boostj indicates the share of impressions of brand post j that the NGO has
paid for
NGOj dummy variable indicating whether brand post j is placed by an NGO that addresses an environmental or health problem (baseline category is environmental)
The model that includes all variables can be expressed as:
where
Share of Commentsnj indicates the number of comments relative to the number of impressions n for brand post j
Share of Tagspj indicates the share of the number of comments that includes a tag p for brand post j
3.4.3 Validation of the log-linear regression model
Before discussing the results, an initial model including all variables was estimated to check whether the model complies with the assumption of linear models and to make alterations in case violations were detected. No evidence of misspecification of the variables (RESET = .236, p = 1) or heteroskedasticity (Breusch-Pagan = 19.886, p = .069) was found. When testing for multicollinearity, all VIF values ranged between 1 and 2, which indicates that this is not an issue.
Furthermore, the significance of the Shapiro-Wilk test of the residuals (p < .001) suggests that the error terms in the model are not normally distributed. Accounting for this by Box-Cox transformation did not remove this issue from the model. Although Leeflang et al. (2015) note that increases in sample sizes more often lead to rejections of the normality assumption, the residuals were further inspected. No correlations between the predictors and the error terms were found, which means that the variables can be considered exogenous. A reason for non-normality of the error terms might be the existence of outliers in the residuals. Based on Figure 3, a filtered dataset containing only the 32 observations (3,19%) for which the residuals are outliers was made. From this dataset it appeared that all of these brand posts have in common that their CTR values are extremely low,
