UNIVERSITY OF AMSTERDAM AMSTERDAM BUSINESS SCHOOL
THE IMPACT OF FREQUENCY AND SPACING OF
BRAND-GENERATED CONTENT ON THE DYNAMICS OF CONSUMER
INTERACTION ON FACEBOOK PAGES
Master Thesis
MSc in Business Administration – Marketing Track
Author Evaldas Jankauskas 11603097 Supervisor Dr. Abhishek Nayak submitted 25-01-2018
2
Statement of originality
This document is written by Student Evaldas Jankauskas who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
3 Index
Abstract ... 4
1. Introduction ... 5
2. Literature review ... 9
2.1 Value of consumer engagement on social media ... 9
2.2. Drivers of consumer interaction on social media ... 10
2.3. Effects of the advertising repetition on the effectiveness of the ad ... 13
2.4. Effects of the message sentiment on the consumer interaction ... 16
2.5. Spacing effects of the brand-generated content ... 17
2.6. Effects of the generation rate of new messages on the growth rate of interaction ... 18
2.7. Conceptual model ... 20
3. Method ... 21
3.1. Data ... 21
3.2. Measurement of variables ... 23
3.2.1. Measurement of variables - model 1 ... 24
3.2.2. Measurement of variables - model 2 ... 27
3.2.3. Measurement of variables – model 3 ... 28
3.2.4. Descriptive statistics ... 29
3.3. Statistical procedure ... 30
3.3.1. Statistical procedure - model 1 ... 30
3.3.2. Statistical procedure - model 2 ... 33
3.3.3. Statistical procedure - model 3 ... 34
4. Results ... 35
4.1. Correlation analysis ... 35
4.2. Hypotheses testing – model 1 ... 36
4.3. Hypotheses testing – model 2 ... 39
4.4. Hypotheses testing – model 3 ... 41
5. Discussion ... 44
5.1. General discussion ... 44
5.2. Theoretical implications ... 46
5.3. Managerial implications ... 48
5.4. Limitations and future research ... 49
6. References ... 51
4 Abstract
This study examines the influence of frequency and spacing of brand-generated content on the dynamics of consumer interaction on social media. Dynamics of consumer interaction was operationalized in two ways: as the level of interaction and the growth rate of interaction of a post. Additionally, the moderating role of sentiment on the relationship between the frequency and the level of interaction was examined. Also, this paper investigated the role of posting and spacing between messages on the number of brand followers. Two data sets were collected via Facebook’s API consisting of 6,471 and 932 brand posts respectively. Results showed that frequency of brand-generated content and the level of consumer interaction has an inverted u-shape relationship and that the level of consumer interaction is positively influenced by the time period (or space) between the posts. Furthermore, findings of this study suggest that posting on social media (higher audience reach specifically) is positively associated with unfollowing by followers and that the growth rate of interaction of the post depends on the rate of new message generation by the same brand. Based on these findings, marketing managers are advised to employ a more moderate posting behavior in terms of posting frequency in order to maximize consumer interaction.
Key words: frequency, spacing, social media, brand-generated content, engagement,
5
1. Introduction
Social media marketing is an attractive marketing method for fostering relationships with customers. About 30% of social media users find social networking sites important when searching for information about brands as well as showing their support towards them (Nielsen, 2017). This engagement with brands on social media is one of the factors driving company outcomes. For example, consumer engagement in social media brand communities is found to have a positive impact on purchase spending (Goh, Heng, and Lin, 2013), brand equity (Christodoulides and Jevons, 2012), and brand attitude (Schivinski and Dabrowski 2016). Consumer engagement involves both consumer interaction and co-creation of the content (Smith and Gallicano, 2015). In order to enhance engagement with brand content, marketers must persuade consumers to interact with those messages by sharing, commenting or liking them (Chang, Yu, Lu, 2014). Hence, interaction is the crucial step towards improving consumer engagement.
While marketers rely on experimenting in order to find elements that drive consumer interaction, researchers use vast social media data in order to examine relationships between brand message characteristics and consumer interaction with those messages. For example, Vries, Gensler, and Leeflang (2012) studied the impact of post’s vividness, interactivity, content, position of a post and valence of comments on brand post popularity as represented by number of likes and comments. Wang et al (2016) examined the impact of topic, tone and the length of post on social media engagement defined not only by the number of likes and shares, but also by the likability of characters featured in the post. Chang, Yu, and Lu (2014) studied how argument quality, post popularity, and post attractiveness can lead to consumer engagement. Similarly, Lee and Hong (2016) investigated the impact of emotional appeal, informativeness and creativity of a message on positive consumer behavior towards brand message. However, little is known so far about the effects of the frequency and spacing of
6
brand-generated content on the dynamics of consumer interaction. Advertising research shows that advertising frequency has an impact on various consumer behavior and attitude outcomes and suggests that there is an optimum level of exposure to advertising that yields greatest results (Schmidt and Eisend, 2015; Broussard, 2000; Malaviya, 2007). Homburg, Ehm, and Artz (2015) note that consumers show diminishing returns to active firm engagement. Moreover, research on advertising repetition in traditional channels suggests an inverted u-shape relationship between ad repetition and message effectiveness. This happens because at a certain number of exposures negative factors, such as boredom and irritation (Heflin and Haygood, 1985), kick in and overweigh positive ones. As a result, the effectiveness of an ad starts diminishing. This effect is also known as the wear-out effect. On the other hand, Lee, Ahn, and Park (2015) suggest that inverted U-shape relationship between repetition and attitude towards the brand does not hold true in online environments. This is the case because users can control their exposure to advertising, therefore they do not expose themselves to the ad to the extent that they feel adverse toward it. As firm-generated brand content on social media is a form of advertising, it is interesting to examine, whether wear-out effect occurs in the context of social media and user interaction.
In addition, the effect of advertising repetition is found to depend on the time period, or space, between ad exposures (Janiszewski, Noel, and Saywer, 2003). Spacing between exposures affect learning (Sawyer, Noel, and Janiszewski, 2009), attitude towards the brand (Schmidt and Eisend, 2015), purchase spending (Sahni, 2015), attrition rate and customer response (Dreze and Bonfrer, 2008). Moreover, recent study by Wang, Greenwood, and Pavlou (2017), who investigated the influence of posting on the propensity to unfollow the brand on the largest social media in China WeChat, found that posting leads to higher likelihood of unfollowing the brand, which in turn has a negative effect on the long term sales. However, WeChat may be considered to be more intrusive than Facebook because of the differences in
7
how followers get notified about new brand posts. Therefore, it is interesting to examine whether the same effect of posting holds true on Facebook. Finally, viral marketing research suggests that the growth rate of interaction with the content depends on the rate of creation of other messages (Karnik, Saroop, and Borkar, 2013). Based on the findings from previous studies, it is evident that frequency and spacing may have a significant influence on the level and growth rate of user interaction.
Furthermore, two-sided advertising research suggests that inclusion of negative information in product related messages can yield better results in terms of persuasive power than if no negative information is included (Eisend 2006). In addition, political communication researchers found that sentiment-carrying Twitter messages tend to be retweeted more often and more quickly (Stieglitz and Dang-Xuan, 2013). Therefore, it is suggested that the effect of message frequency and spacing on the level of consumer interaction is moderated by the sentiment of the message. In other words, the optimal level of message frequency is expected to be higher for emotionally-charged firm-generated brand messages as compared to neutral ones. Hence, the following research questions are proposed:
RQ1: How does frequency and spacing of brand-generated content affect the dynamics of consumer interaction on social media and how is this effect moderated by the sentiment of
the content?
RQ2: How does posting on social media affect the unfollowing by brand followers? RQ2a: Does the spacing between messages help reduce the negative effect of posting on the unfollowing by brand followers (if such effect is present)?
In order to answer these questions two data sets were gathered via Facebook’s API consisting of 6,471 and 932 brand posts respectively. Two separate data sets were needed to examine the frequency effects on the overall level of consumer interaction as well as on the growth rate of interaction. Therefore, post and page data for 7 international brands from 5
8
different product categories for the period of 2 years were collected to examine the frequency effects. To investigate the effect of posting on the growth rate of consumer interaction, 11 brands were tracked for the 7-week period in order to capture the development of the interaction. In addition, the impact of posting on the propensity to unfollow the brand was examined. Consequently, three separate regression models were built to test the hypotheses. Results showed that frequency of posting and the level of consumer interaction has an inverted u-shape relationship and that the level of consumer interaction is positively influenced by the space between the posts. Further, findings suggest that posting on social media is positively associated with unfollowing by followers and that the growth rate of interaction of the post depends on the rate of new message generation by the same brand.
The study has few theoretical and practical contributions. Answering to the call for research (Vries, Gensler, and Leeflang, 2012) to include the dynamic aspects of interaction, this study contributes to the social media literature by examining the effects of the rate of new message generation on the growth rate of interaction of the post. In addition, this study adds to the stream of research on the wear-out effects in online environments by including higher number of exposures and by testing the type of firm communication (social media communication) that previously has not been studied. Finally, this study contributed to the recent research (Wang, Greenwood, and Pavlou, 2017) by examining the effect of posting on the unfollowing by brand followers. As for practical contributions, findings of this study have implications for marketing managers with respect to the frequency and spacing of posting. This paper provides evidence for a more moderate posting strategies in terms of frequency.
The remainder of this paper is structured as follows. The next chapter gives an overview of the latest developments in regard to the drivers of consumer interaction on social media. Afterwards, chapter three describes the data collection procedure and research method. Results of the analyses are discussed in chapter four. Finally, conclusions and implications of the
9
results of this study are discussed in chapter five, together with the limitations and suggestions for possible future research directions.
2. Literature review
In this chapter relevant theoretical concepts and the most relevant findings from current literature about the drivers of interaction on social media are discussed. Subsequently, the chapter continues with an overview of the literature on the ad repetition and spacing effects on the effectiveness of the ad. Further, overview continues with the role of sentiment in sharing behavior and the impact posting has on the following behavior of social media users. The chapter ends with a conceptual model, which graphically illustrates proposed hypotheses.
2.1 Value of consumer engagement on social media
Social media is defined as an online platform that “enables consumers to create, interact, collaborate and share in the process of creating as well as consuming content” (Obar
and Wildman, 2015, p.746). Large portion of social media users find social networking sites important when searching for information as well as showing their support towards brands (Nielsen, 2017). Marketers, in turn, tapped into this opportunity and employed social media to generate content to engage consumers in order to increase information sharing (Goh, Heng, and Lin, 2013). Consumer engagement with brands can be defined as progression from one-way reception and interaction with the messages to cognitive involvement in creating, responding to, and distributing information (Smith and Gallicano, 2015). Therefore, consumer engagement is an ongoing process with consumer interaction as an initial step towards engagement.
The growth of social media also led to investigations of the value of consumer engagement in social media for companies. Studies have shown that consumer perceptions of
10
co-creation and community have a positive impact on consumer engagement that, in turn, positively affects consumer-based brand equity (Christodoulides, Jevons, and Bonhomme, 2012). Other studies found that engagement in social media brand communities leads to an increase in purchase expenditures (Goh, Heng and Lin, 2013). They suggest that this effect is achieved through embedded information and persuasion of the social media contents. However, social media content is not limited to functional information. Consumers have different motivations for engaging with brand-related content. While some users search for product-related information, others seek inspiration and entertainment (Homburg, Ehm, and Artz, 2015). In addition to the influence of embedded information and persuasion on purchase behavior, consumers are more willing to purchase a good or a service featured in a social media content when they have a strong intention to engage in viral behavior (Lee and Hong, 2016). This suggests that focus on increasing the number of likes and shares may prove to be an effective brand communication strategy (Lee and Hong, 2016). Liking, sharing and commenting on a brand content is similar to word of mouth communication and is the indication of consumer interaction behavior (Vries, Gensler, and Leeflang, 2012).
2.2. Drivers of consumer interaction on social media
As consumer interaction with brand content is shown to have positive effects on various brand outcomes, both marketers and researchers have turned to studying possible drivers of consumer interaction behavior. For example, Vries, Gensler, and Leeflang (2012) studied relationship between brand post characteristics, content of the brand post, position of the brand post, the valence of comments and brand post popularity as represented by number of likes and comments. They found that post popularity can be enhanced by creating a vivid and interactive brand post and positioning the brand post on top of the brand fan page. Interestingly, Chang, Yu, and Lu (2014) used post popularity (previously used as dependent variable) as variable
11
affecting like and share intentions via perceived usefulness and preferences. Additionally, they included argument quality and perceived post attractiveness as antecedents of like and share intentions. They suggest that marketing managers can strengthen content quality by having famous individuals with connections promote their marketing content. Moreover, vivid images may attract lower level fans, while useful content may appeal higher level fans to respond and share. Wang et al (2016) added to the previous research by examining the impact of topic and tone of post on social media engagement defined not only by the number of likes and shares, but also the likability of characters featured in the post and found that specific topics, tone and length of a post are associated with social media engagement. Similarly, Lee and Hong (2016) studied impact of emotional appeal, informativeness and creativity of brand message on the intention to express empathy operationalized as intention to click like for the brand content and found that informativeness and creativity of the content are the key drivers of favorable behavioral responses. Wagner, Baccarella, and Voigt (2017) extended message appeal research by distinguishing between emotional and functional appeals and found that some appeals have positive and others have negative impact on consumer interaction. Interestingly, they found that some appeals with negative impact on user interaction are used extensively, while appeals with positive impact are used rarely. A summary of independent and response variables as well as findings of previous research are presented in table 1.
12
Table 1. Summary of research on the drivers of consumer engagement on social media
Antecedents Dependent variables Findings Author(s)
• Post appeals • User interaction • “Some post appeals have positive and others have negative impact on user interaction.” (Wagner,
Baccarella, and Voigt, 2017, p. 1).
• Some appeals with positive impact are rarely used, while others with negative impact are used quite often
Wagner, Baccarella, and Voigt (2017) • Emotional appeal • Informativeness • Advertising creativity
• Purchase intention • Informativeness and advertising creativity are the key drivers of favorable behavioral responses to an SNS ad and that intention to engage in favorable user responses is positively associated with purchase intention
Lee and Hong (2016)
• Content category • Page total likes • Type of content (link,
photo, status, video) • Post month
• Post hour • Post weekday • Paid (yes/no)
• Engagement • Type of the content is considered to be the most relevant input feature for the model
• Status type posts are likely to result in twice the impact of the remaining types.
• Publications related to special contests and offers are likely to generate posts with greater impact than product and other non-explicit brand related contents
Moro, Rita, and Vala (2016)
• Topic of the post • Tone of the post • Length of the post
• Character’s likability • Number of likes • Number of shares
• The level of social media engagement varies by topic • People are more engaged to the
narratives associated with dreams, education, and romantic relationship • Tone of the post is positively
related to the number of likes and shares as well as the likability of the character • Tone of the narrative can
influence how people make evaluations Wang et al (2016) • Argument quality • Post popularity • Post attractiveness • Like intention • Share intention
• Post popularity, argument quality and attractiveness reinforce preference and usefulness
• Like intention of brand followers is the essential factor in their sharing intention • Different levels of user
expertise affect the willingness to like and share
Chang, Yu, and Lu (2014)
13
• Sentiment of a tweet • Sharing volume • Sharing speed
• Sentiment-carrying Twitter messages tend to be retweeted more often and more quickly compared to the neutral messages Stieglitz and Dang-Xuan (2013) • Post vividness • Post interactivity • Informational content • Entertaining content • Position of a post • Valence of comments • Day of the week (c) • Message length of brand
post (c)
• Product category (c)
• Number of likes • Number of comments
• Positioning the brand post on top of the page improves post popularity
• Vivid and interactive brand posts enhance the number of likes
• Share of positive comments on a brand post is positively related to the number of likes • Number of comments can be
increased by the interactive post characteristic (e.g. a question) • Both negative and positive
comments are positively related to the number of comments
Vries, Gensler, and Leeflang (2012)
2.3. Effects of the advertising repetition on the effectiveness of the ad
The effects of advertising repetition have been extensively studied in traditional channels, such as television and print. Previous research that studied the impact of advertising repetition usually examined single ad repetition strategy. As noted by Schmidt and Eisend (2015), majority of the findings point to the same conclusion: initial exposures to an ad first increase the effectiveness of the ad because of learning and habituation (also known as wear-in effect) (Schmidt and Eisend, 2015), but at the certawear-in number of exposures receiver becomes bored and additional exposures lead to lower attention or even irritation (Heflin and Haygood, 1985). Therefore, the effectiveness of the ad diminishes (wear-out effect). Wear-out effect can be found in relation to consumer attention toward the ad (attention wear-out), brand recall (learning wear-out), and attitude and purchase intention (acceptance wear-out) (Lee, Ahn, and Park, 2015). Wear-in and wear-out effects can be explained by Berlyne’s two-factor theory (Berlyne, 1970). It suggests that positive and negative factors interact with one another and have an impact on the effect of repetition on attitude (Schmidt and Eisend, 2015). Positive factors, such as learning and habituation, result in positive thoughts, while negative factors,
14
including redundancy and boredom, result in negative thoughts (Schmidt and Eisend, 2015). Attitude towards the ad increases with repetition until learning is saturated. Then, each additional exposure leads to boredom and negative thoughts that exceed positive ones. At this point, attitude decreases and the effect of repetition becomes negative (Schmidt and Eisend, 2015). Two-factor theory suggests an inverted U-shape relationship between the number of exposures and the impact of the message (Berlyne, 1970; Pechmann and Stewart, 1988; Schmidt and Eisend, 2015).
Advertising repetition research in traditional channels suggests an inverted u-shape relationship between message repetition and effectiveness of the message. However, advertising research in online environments present mixed results. For example, Manchanda et al. (2006) studied the effect of banner ads on internet purchases and found that advertising repetition is positively correlated with revenues, but that the size of return on each exposure declines with each additional exposure. This implies the diminishing effect of ad repetition. On the other hand, Lee, Ahn, and Park (2015) suggest that inverted u-shape relation between advertising repetition and effectiveness does not fully hold true in the context of online advertising. They studied the effects of repetition on attention, memory, and attitude, and found that wear-out only occurs in regards with attention, but not with memory or attitude. However, the high-repetition condition consisted of 8 exposures, which, according to the authors, may not be sufficient to achieve wear-out effect, and, thus, needs to be studied further. Having in mind high volume of brand posts on social media, the likelihood of wear-out to occur increases. Although brand-generated content on social media does not carry all the features of traditional advertising definition, it can be considered a form of advertising. According to a working definition by Daglen and Rosengren (2016), advertising can be described as a “brand-initiated communication intent on impacting people”. Definition has shifted from such terms
15
changes in advertising formats over the last 20 years (Dahlen and Rosengren, 2016). Based on this definition, it is evident that brand-generated content on social media can be considered to be a form of advertising. Brand-generated content is the communication effort initiated by a brand and it is intended to attract people’s attention and consequently induce them to engage (anticipated impact) (Vries, Gensler, and Leeflang, 2012). Therefore, based on the literature of advertising frequency and the link between brand-generated content and advertising, the following hypothesis regarding the frequency of brand-generated content and the level of consumer interaction is proposed.
H1: Frequency of brand-generated content on social media and the level of consumer interaction have an inverted u-shape relationship
Furthermore, Wang, Greendwood, and Pavlou (2017) found that social media posts are positively associated with unfollowing by the brand followers. Using a unique dataset from a Chinese fashion retailer present on social media platform WeChat, authors examined how a firm’s social media posts affect the propensity of its followers to unfollow the firm. They found that posting on social media increases the likelihood that existing followers of a brand page will unfollow the brand. Authors suggest that this effect is driven by interruption induced annoyance and concern for privacy. However, they did not study whether posting on social media helps attract new followers. In addition, the way Facebook notifies its users about new brand posts is different compared to WeChat. Users who follow brands on WeChat receive notifications about new brand posts in the same way as messages from friends (Wang, Greenwood, and Pavlou, 2017). On the other hand, users on Facebook encounter brand posts when exploring newsfeed on their own choice. In other words, brand followers on Facebook chooses when to be interrupted. As a result, Facebook may be considered to be less intrusive. Having in mind these differences between Facebook and WeChat, it is interesting to examine, whether the same brand following dynamics hold true on Facebook. Although intrusion may
16
not be the main driver of the propensity to unfollow the brand on Facebook, wear-out effect of advertising repetition suggests that additional exposures may lead to boredom and negative thoughts (Schmidt and Eisend, 2015). Assuming that majority of brand followers are not new followers and that they have been previously exposed to brand posts, each additional post may lead to negative thoughts and, therefore, may increase the likelihood of unfollowing the brand. Moreover, the larger the audience is reached by a post, the more significant losses in terms of the number of followers would be. Therefore, the following hypothesis is proposed:
H2: Higher audience reach of a single post is positively associated with unfollowing by followers
2.4. Effects of the message sentiment on the consumer interaction
Two-sided advertising research suggests that inclusion of negative information in product related messages can yield better results in terms of persuasive power (Eisend 2006), which, in turn, have a positive impact on the like and share intentions via perceived usefulness (Chang, Yu, and Lu, 2014). Furthermore, political communication researchers found that emotionally-charged Twitter messages tend to be retweeted more often and more quickly (Stieglitz and Dang-Xuan, 2013). Previous research on written communication has shown that emotional words may evoke cognitive processes such as attention (Bayer, Sommer and Schacht, 2012; Kissler et al., 2007). Consequently, a higher level of cognitive involvement may increase the likelihood of behavior response in terms of sharing (Luminet et al., 2000; Peters, Kashima, and Clark, 2009). In addition, Huffaker (2010) who studied impact of emotions in computer-mediated communication showed that, in discussion forums, users who use emotion-rich language in their messages get more feedback compared to those who do not. This is applicable to both positive and negative emotions. Similarly, Joyce and Krauta (2006) studied newsgroups and found that positive emotions in messages encourages participation by
17
reinforcing a sense of community. Finally, content that arouse positive or negative emotions is more viral compared to the emotionally neutral content (Stieglitz and Dang-Xuan, 2013). Thus, the presence of emotional words in social media brand-generated content may get more attention and evoke more arousal, which in turn may positively influence sharing behavior and, therefore, interaction in general. Hence, it is proposed that the effect of post frequency on the level of consumer interaction is moderated by the sentiment of the message. In other words, the optimal level of message frequency is expected to be higher for sentiment-carrying brand messages as compared to the neutral ones.
H3: Relationship between the frequency of brand-generated content and the level of consumer interaction is moderated by the sentiment of the message
2.5. Spacing effects of the brand-generated content
The effects of advertising repetition on memory formation depend on the time period, or space, between ad exposures (Janiszewski, Noel, and Sawyer, 2003). Phenomenon is known as spacing and it could be defined as the time period between exposures to a stimulus. In advertising, spacing effect suggests that longer periods of time between exposures lead to better learning than shorter periods (Sawyer, Noel, and Janiszewski, 2009). It is found to have an impact on various consumer response measures. For example, Schmidt and Eisend (2015) studied the optimal level of exposures that maximize consumer response to an ad, and found that spaced exposures strengthen repetition effects on attitude toward the brand. Similarly, Sahni (2015) studied the impact of advertising spacing on the likelihood of a purchase in online environments, and also found evidence for the spacing effect. The study suggests that likelihood of a purchase increases when the ads are spread, even if it requires moving some ads away from the purchase moment (Sahni, 2015). Finally, Dreze and Bonfrer (2008) examined
18
the impact of communication frequency on the customer retention and spending, and found that message spacing has an impact on both attrition and the customer response.
Spacing effect is related to the two-factor theory (Berlyne, 1970) and, more specifically, to the wear-out effect. In case of longer space between two exposures, it takes more time for the receiver to familiarize himself/herself with the stimulus before boredom and/or irritation develops, and exceed the positive effect of learning (Schmidt and Eisend, 2015). Moreover, larger periods of time between exposures let recipients process information, form associations with information stored in memory, and, therefore, improve recall (Schmidt and Eisend, 2015). Finally, spacing effect has greater impact on the novel stimuli than on familiar ones (Sawyer, Noel, and Janiszewski, 2009). This is the case because it takes more effort to process more sophisticated advertising content, and, as a result, it may not “wear” as well as simpler advertising content (Sawyer, Noel, and Janiszewski, 2009). As spacing reduces the risk of irritation, it may also diminish the propensity to unfollow the brand, which, according to Wang, Greenwood, and Pavlou (2017) is driven by the irritation induced annoyance and concern for privacy. Hence, the following hypotheses regarding the spacing of brand-generated content and the level of interaction and propensity to unfollow the brand are proposed.
H4: Time period (or space) between brand-generated content positively affects the level of consumer interaction on social media
H5: Time period (or space) between the brand-generated content has a positive impact on the number of brand page followers
2.6. Effects of the generation rate of new messages on the growth rate of interaction Consumer interaction with brand-generated content is a mechanism for information diffusion. High rate of user interaction could make a message highly popular or viral. To
19
understand the dynamics of the growth rate of user interaction, viral marketing research is used. Number of studies have proposed models of information diffusion in order to measure diffusion structure and likelihood and define what makes content viral. For example, contagion or epidemic models base their assumptions on the mechanism of social contagion – active word of mouth or passive observation of the actions of others (Chandrashekaran, Grewal, and Mehta, 2010). Similarly, contagion is also the basis for cascade models, which “define an iterative process of diffusion in a network” (Bourigault, Lamprier, and Gallinari, 2016) and characterize
information dissemination as a branching process (Nematzadeh et al., 2014). Another large group of models are threshold models, that “incorporate the idea of “social reinforcement” by assuming that each adoption requires a certain number of exposures” (Nematzadeh et al.,
2014).
Structure of the network is the foundation for modeling information diffusion (Nematzadeh et al., 2014). For example, Bakshy et al. (2012) studied the role of social networks in online information diffusion and found that users that are exposed to signals about friends’ information sharing behavior, are more likely to share information, and do so earlier in the process. Karnik, Saroop, and Borkar (2013) examined how and why some content on social media becomes viral. They suggest that a message becomes viral after it surpasses a certain response threshold. More interestingly, study suggests that this threshold depends on the stream of competing messages: the higher the rate of new message generation, the higher the threshold will be. This suggests that the growth rate of interaction with the content depends on the rate of generation of other messages (Karnik, Saroop, and Borkar, 2013). Therefore, the following hypothesis regarding the growth rate of interaction is proposed.
H6: Growth rate of interaction with brand-generated content depends on the rate of new message generation by the same brand
20
2.7. Conceptual model
In previous sections six hypotheses were established. Three variables represent the frequency dimension of the model and two variables represent consumer interaction. The graphical illustration of conceptual model is represented in figure 1 below.
Figure 1. Conceptual model of the effect of frequency and spacing of the brand-generated content on the dynamics of consumer engagement
Audience reach of a post Frequency of brand-generated content Level of consumer interaction Spacing of brand-generated content Growth rate of consumer interaction Number of brand page followers Sentiment of a message Frequency Interaction H1 H3 Rate of new message generation H4 H6 H2 H5
21 3. Method
This chapter represents the empirical part of the study. First, data collection process and final sample is outlined and data is described. Afterwards, operationalization and measurement of variables included in the analyses are discussed. Finally, a description is provided for the statistical approach and procedure that was used in order to test the expected relationships as discussed in the literature review.
3.1. Data
To answer the research question, post and page data from multiple brands present on Facebook were collected. Facebook is the largest social networking site in terms of the number
of active users (“Social media statistics”, 2017), which makes it a representative source for social media and, therefore, appropriate source for this study. Furthermore, the use of Facebook in previous research on user interaction reinforces the suitability of this social networking platform (Vries et al., 2012).
The data was collected via Facebook’s API using statistical software RStudio and its package RFacebook. The nature of dependent variables required different datasets. In order to test post frequency and post spacing effects on the level of consumer interaction (H1, H3, and H4), post, interaction and page data needed to be collected. However, testing the effects of audience reach and post spacing on the propensity to follow or unfollow the brand (H2 and H5), and testing the impact of the rate of new message generation on the growth rate of interaction with a post (H6), required data that reflected development of user interaction over time. Therefore, two separate datasets were gathered. Data collection procedure and descriptive statistics of the data are outlined next.
First, in order to test the frequency and spacing effects on the level of consumer interaction, post, interaction, and page data were collected. Facebook allows to extract data no
22
older than 2 years, therefore, data for the period of 96 weeks (2015/12/07 – 2017/10/08) was collected. Moreover, data was extracted on the 8th and 9th of November to allow for a time lag, which is important in order to capture final level of consumer interaction. Social networking sites are extremely dynamic so the content posted for more than 30 days is unlikely to get more interaction (Sabate et al, 2014). Finally, Facebook limits the availability of historic page data in regards to the number of page followers (or page likes). It allows to extract the number of page followers by country, however, the number of countries is limited, therefore, the total number of followers collected in this way is not equal to the actual total number of followers as seen on the Facebook page of a brand. In other words, the number of page followers extracted via API usually accounted for less than 98% of actual total number of followers, depending on the brand. As a result, 7 international brands that were actively posting content on Facebook with more than 85% of total followers available via API, were investigated. The brands were from five different categories: online retail, city marketing, sports goods, cars, and mobile phones. The following information was collected: 1) text message of a post, 2) type of the post (event, link, status, photo, or video), 3) number of brand page followers, 4) number of likes, comments, and shares of a post, and 5) the date and time the post was created. The total of 6.471 brand posts were collected.
Second, in order to examine how audience reach and post spacing affect the number of page followers, and to test how the rate of new message generation impacts the growth rate of interaction with a post, a second dataset was produced and was used in 2 separate models. One was used to test hypothesis H2 and H5 (the effects of audience reach and post spacing on the number of page followers), and the other was used to test hypothesis H6 (the effect of the rate of new message generation on the growth rate of interaction). To capture the changing dynamics of interaction with a post, 11 brands were tracked in the period between 2017/09/25 and 2017/11/14. During this period, post and interaction data was scraped every 2 hours during
23
the day. In order to improve generalizability of the findings brands were selected based on the type of product or service they produce. Hence, brands with differing type of product or service were given priority. Moreover, number of brands turned out to be relatively passive on social media with the annual number of posts less than 100, therefore, the level of activity on the platform was additional criteria for brand selection. The total of 932 posts were collected.
To test the relationship between the audience reach and the number of page followers, 822 posts were taken into consideration. Some posts were removed from the analysis due to the operationalization of interaction variable. It was operationalized as the total number of interaction with the post during the first 20 hours since the post was created. Such time period was selected after examining descriptive statistics of the dataset. The average time period until the growth rate of interaction hits plateau (growth rate does not exceed 3% for consecutive 5 days) is 64 hours. In addition, around 75% of final level of interaction occurs during the first 20 hours since the creation of the post. Therefore, the level of interaction was cut to the first 20 hours of post’s life. As mentioned earlier, some posts needed to be removed because there was no interaction data for the first 20 hours since the post was created. Some posts only had interaction data for the first 30 or 40 hours since the posting. This occurred because the time period between some of the data extracts were longer than 2 hours. Therefore, such posts were removed in order to keep analysis consistent.
3.2. Measurement of variables
Data analysis section consists of 3 separate models, therefore, variables of each model are discussed separately. First, variables that were used to test the relationship between the frequency and spacing and the level of consumer interaction are outlined. Afterwards, variables that were used to test the effects of post reach and post spacing on the number of followers are
24
discussed. Finally, variables of model 3, which is designed to test the relationship between the rate of new message generation and the growth rate of consumer interaction, are described.
3.2.1. Measurement of variables - model 1
Dependent variable
Response variable in model 1 is the level of consumer interaction with brand-generated content. In order to examine the frequency effects on the level of interaction, a period of seven days was used. The variance in the number of daily posts is relatively high. Some brands post multiple messages per day while others may only post few times a week. In addition, the average number of daily posts is relatively low compared to the number of weekly posts, which may not be sufficient for curvilinear relationship to occur. Hence, the level of consumer interaction was operationalized as the total number of likes, shares and comments per week.
Independent variables
The main independent variable, frequency of brand-generated content is operationalized as the total number of posts per week. Spacing of the brand-generated content was calculated as the average time period expressed in hours between the posts in each week and was grouped into 2 categories based on the 50th percentile. 2 categories were used instead of 3 as the model with 2 categories better explains the variance in dependent variable (R2 = 27.7% with 2 categories compare to R2 = 24.8% with 3 categories). In addition, dummy variables with categories were used instead of continuous variable in order to reduce multicollinearity, which is caused by the correlation between the number of weekly posts and the average space between posts.
Sentiment of a message was defined as the average sentiment score of all messages in
25
tool has been proven effective in classifying emotions in short informal messages (Stieglitz and Dang-Xuan, 2013). Sentistrength is based on a human-designed dictionary of emotional terms with an additional collection for negations (e.g. “not satisfied”), booster words (e.g. “very good”), amplifications (e.g. “saaaad”), emoticons, spelling corrections, and other factors (Stieglitz and Dang-Xuan, 2013). SentiStrength has a scale of 1 (neutral) to 5 (strongly positive) for a positive sentiment and a scale of -1 (neutral) to -5 (strongly negative) for a negative sentiment. Every text message is given both a negative and a positive sentiment score. SentiStrength has been shown to be superior compared to standard machine learning methods
in terms of accuracy rate for positive sentiment strength (Thelwall et al., 2010). In general, sentiment can be operationalized in terms of polarity (positive, negative, or neutral) and in terms of total amount of sentiments. This is important when a message has both positive and negative sentiment words. In these cases, polarity of a message fails to capture the level of emotionality of the message because the positive and negative sentiment scores cancel each other out. For example, message could have polarity=0, although the message, in fact, is heavily emotional and not neutral as polarity would indicate. Therefore, the total amount of sentiment instead of polarity was used. The measure of the total amount of sentiment is based on Stieglitz and Dang-Xuan (2013):
sentiment = (positive – negative) - 2
Control variables
• Message length. Advertising research suggests that message length may have an impact on click-through rates (Robinson, Wysocka, and Hand, 2007). Therefore, message length is included as control variable. It is operationalized as the average number of characters of all brand messages in a given week.
• Post type. Vries, Gensler, and Leeflang (2012) showed that vividness of a post operationalized as post type (photo, video, text etc.) has an effect on brand post popularity.
26
Therefore, post type is included as control variable and is based on the classification used by Facebook. There are 5 types of posts: event, status, link, photo, and video. Dummy variables for photography and video posts were created, and link, status, and event messages were used together as a baseline due to their low level of occurrence. Dummy variables video posts and photography posts were calculated based on the number of posts on a given week. Dummy variable video posts was assigned 1 for weeks where the number of video posts was higher compared to other types of posts. For example, if there were 6 video posts and 5 photo posts in week 34, dummy variable video posts would be assigned value 1 and dummy variable photo posts would be assigned value 0. In weeks where the number of photography and video posts was equal (e.g. 4 photo posts and 4 video posts), both variables were assigned value 0.
• Level of page following. Studies have shown that the number of followers has an influence on the level of interaction with the message (Stieglitz and Dang-Xuan, 2013). Therefore, number of brand followers was taken into consideration as well. However, due to the fact that Facebook gives limited access to the historic data, variable number of page followers was operationalized as the level of page following. The number of page followers extracted via Facebook’s API usually accounts for less than 98% of actual total number of followers as seen on the brand page. Hence, the changes in the number of page followers extracted via API may not reflect the actual changes in the number of page followers. Therefore, the levels of page following were used instead of numerical values. Different levels were
determined based on the 33rd and 66th percentiles of the number of followers. Brands with number of page followers below 33rd percentile have low level of following, brands between 33rd and 66th percentiles have moderate level and brands above 66th percentile have a high level of following. Hence, 2 dummy variables for the level of page following were created: page following moderate and page following high.
27
3.2.2. Measurement of variables - model 2
Dependent variable
Dependent variable, number of brand page followers, was operationalized as the change in the number of page followers. To test how a new post affects the propensity to follow or unfollow the brand page, the change in the number of page followers must be captured. To do that, the difference between the number of page followers 1 day after the post was created and the number of page followers at the day of the posting was calculated.
Independent variables
The main independent variable affecting the propensity to follow or unfollow the brand, is the audience reach of the post. Simply put, the more users post reaches the more people would follow or unfollow the brand. The reach of the post is only available to the administrator of the brand page, therefore, the total number of page likes, comments, and shares 20 hours after the post was created, was used as a proxy for the post reach. Such time period was selected after examining descriptive statistics of the dataset. Around 75% of the final level of interaction occurs during the first 20 hours since the creation of a post. Therefore, the largest effect on the number of page followers is expected to occur within this time frame. Finally, spacing of the brand-generated content was operationalized as the time period since the previous post
expressed in hours.
Control variables
Similar to model 1, message length, sentiment of a message and post type were used as control variables. Message length was operationalized as the number of characters of the brand message. Sentiment of the message was defined using the same approach as in model 1. In
28
regards with post type, 2 dummy variables representing photo and video posts were created with the link and status posts representing a baseline.
3.2.3. Measurement of variables – model 3
Dependent variable
Growth rate of a post was operationalized as the number of new likes, shares, and comments per hour. The variable was calculated based on the records made during each data extract. Data was extracted every 2 hours during the day, therefore, the growth rate was calculated by dividing the total number of new likes, shares, and comments by the time period between the current and previous extract or the time the post was created. The time period may also be other than 2 hours. For example, if the extract was made at 10AM and the post was created at 4AM, the time period used in calculations of the growth rate would be 6 hours, because there were no extracts made during the night.
Independent variable
Rate of new message generation was operationalized as the number of new posts on the
last extract. For example, if the post in question is analyzed at a certain point in time (e.g. at 10AM, the time of the first extract of the day) the number of new messages would be calculated as follows: the total number of posts at 10PM (the time of the previous extract) would be subtracted from the total number of posts at 8PM.
Control variables
Additional factors, such as time period since the creation of the post, the number of followers and the growth rate of the previous post may also have an effect on the growth rate of a post. For example, the level of interaction is often higher at the beginning of post’s life
29
compared to later stages. Further, brands with more followers would naturally receive more interaction and the previous post that gets a lot of interaction may bring traffic to the brand page and in turn positively affect the level of interaction of the latest post. Therefore, time period since the post was created, the number of page followers and the growth rate of previous post were used as control variables. Time period since the post was created was expressed in the number of hours, number of page followers was operationalized as the total number of followers at the day of data extract, and the growth rate of the previous post was defined as the number of new likes, comments, and shares per hour.
3.2.4. Descriptive statistics
The average number of brand followers was 13,227,367 per brand (SD = 10,950,435); the number of brand posts taken into consideration in this study was, on average, 924.34 per brand (SD = 206.41), meaning that a brand, on average, created 9.63 posts per week. The average number of interactions (likes, comments, and shares) per brand post was 5,317 (SD = 20,947).
Companies used various techniques to stimulate brand followers to interact with the content (see table 2). On average, about 53% of the brand posts were photographs and 32% were videos. The least frequent type of the content were status posts and event posts, which, on average, accounted for 0.1% and 0.9% of total posts respectively. Furthermore, the relative shares of positive, neutral, and negative brand posts were 0.47, 0.44, and 0.09 respectively. The average time period between two posts was 17.07 hours (SD = 18.09) and the average text length of a brand post was 140.99 characters (SD = 102.71).
30
Table 2. Brand post characteristics model 1
Post type Relative frequency (%) Min (%) Max (%)
Event 0.9% 0% 3.7%
Link 14.7% 0.3% 52.3%
Photo 52.8% 31.4% 81.4%
Status 0.1% 0% 0.3%
Video 31.5% 8.5% 68.3%
In model 2 (hypotheses H2 and H5), the average number of brand followers was 28,228,748 per brand (SD = 16,115,583) and each brand created 1.49 posts per day on average. The average number of interaction (likes, comments, and shares) per brand post during the first 20 hours since the post was created was 3,788 (SD = 17,422) and the average number of new followers after 1 day of the posting was 2,332 (SD = 3,768). In model 3 all 932 posts of the second data set were taken into account, which is equal to 84.72 posts per brand and 1.69 posts per day. The average number of brand followers was 27,904,930 per brand (SD = 15,828,111) and the average number of new interaction per hour per brand post was 358 (SD = 4,907). Finally, the average growth rate of the previous post was 189.99 likes per hour (SD = 520.56).
3.3. Statistical procedure
3.3.1. Statistical procedure - model 1
To perform statistical analyses, the SPSS was used. Before running regression analyses, descriptive statistics, skewness, kurtosis, and normality tests were computed.
In order to select appropriate analysis method, the nature of dependent variables was identified. The level of user interaction is count variable with a Poisson distribution (Cameron and Trivedi, 2005) and is highly skewed with skewness statistic of 5.268 and kurtosis statistic of 40.313, which indicates that distribution has heavy tails. To normalize the distribution, and, therefore, fulfill the normality assumption of linear regression, multiple transformations have
31
been tested. Log-transformation has shown the best results, with skewness statistic of -.249 and kurtosis statistic of -.127, close to that of normal distribution. To test the curvilinear relationship between the number of posts per week and total weekly interaction, quadratic term for posts per week was introduced. It allowed to identify whether there is a curve in the regression line and what is the nature of the curve (a u-shape or an inverted u-shape). Variable total posts was mean-centered and its quadratic term was calculated in order to avoid multicollinearity problem, which could distort coefficients of the model.
Direct relationships were examined by the use of hierarchical regression. In step 1, the control variables message length, post type photo, post type video, page following moderate, and page following high, were entered into the equation. Total posts centered, and total posts centered squared were entered into the model in step 2. In step 3, spacing of brand-generated
content was entered into the equation. In order to test the moderating role of sentiment, an SPSS
macro of Hayes (2012) was used.
The model to explain the level of consumer interaction can be expressed in the equation below. Description of each variable is provided in table 4.
log $ = &'+ &)*+,- + &.*+,-/0 + &123/-,- + &4/5236,7 + &89:;_1 + &>9:;_2 + &@A-BC-: + &D*+E$*ℎ + &G*+E$H9 + &)'*7I+A+ + &))*7I+69 + K
Table 4. Description of variables for model 1
Acronym Name of variable Description
y Level of consumer interaction Total number of weekly likes, comments, and shares
PoCe Posts centered Total number of brand posts per week. Variable has been mean-centered in order to avoid high level of correlation with its squared term PoCeSq
32
PoCeSq Posts centered squared Squared term of the mean-centered total number of brand posts per week. This variable is introduced to test whether there is a curvilinear relationship between the frequency of brand-generated content and the level of consumer interaction. AvSeCe Average sentiment centered Average sentiment score per week. Variable has been
mean-centered in order to avoid correlation with the interaction term int_1, which represents interaction between mean-centered average sentiment and the mean-centered total number of posts to test the moderating role of the sentiment.
SpAvHCa Spacing category Dummy variable representing a category of spacing. Variable was grouped into 2 categories (low and high) based on the 50th percentile. Category low level of spacing is used as a baseline. int_1 Interaction term 1 Variable represents the interaction between the mean-centered
average sentiment and the mean-centered and squared total number of posts. The term is introduced to test the moderating role of sentiment on the curvilinear relationship between the frequency of brand-generated content and the level of user interaction.
int_2 Interaction term 2 Variable represents the interaction between the mean-centered average sentiment and the mean-centered total number of posts. The term is introduced to test the moderating role of sentiment.
MesLen Message length Average number of characters of a message. PoTyPh Post type photo Dummy variable indicating whether the number of
photography posts was highest in a given week.
PoTyVi Post type video Dummy variable indicating whether the number of video posts was highest in a given week.
PaFoMo Page following moderate Dummy variable indicating whether the level of page following for a given brand is moderate.
PaFoHi Page following high Dummy variable indicating whether the level of page following for a given brand is high.
33
3.3.2. Statistical procedure - model 2
Dependent variable of model 2, operationalized as the change in the number of page followers, is count variable with skewed and kurtotic distribution with skewness statistic of
1.477 and kurtosis statistic of 2.905. To normalize the distribution, and, therefore, fulfill the normality assumption of linear regression, multiple transformations have been tested. Square root transformation has shown the best results, with skewness statistic of .207, close to that of normal distribution. Direct relationships were examined by the use of hierarchical regression. In step 1, the control variables message length, message sentiment, post type video, and post type photo, were entered into the equation. Space between the posts was entered into the model
in step 2 and the level of interaction was entered into the equation in step 3.
The model to explain the level of consumer interaction can be expressed in the equation below. Description of each variable is provided in table 5.
sqrt $ = &'+ &)P:Q + &.E5/** + &1RC-: + &4;/-:;9 + &8RE$5* + &>RE$5H + K
Table 5. Description of variables for model 2
Acronym Name of variable Description
y Change in followers Change in the total number of page followers 1 day after the post was created.
TpSPP Time since previous post Time period since the previous post, expressed in hours. mLen Message length Total number of characters in a message.
tSenti Total sentiment score Total amount of sentiment in a message.
mTypP Message type photo Dummy variable representing whether a brand post is a photography post or not.
mTypV Message type video Dummy variable representing whether a brand post is a video post or not.
34
3.3.3. Statistical procedure - model 3
Regression was applied to test hypothesis H6, that is, to examine whether the growth rate of interaction is affected by the rate of new message generation. Because the dependent variable growth rate of interaction represents true-event count data (i.e. integer based and nonnegative) and its standard deviation is larger than its mean, the analysis has to be adjusted for overdispersion. As none of the transformations helped to normalize the distribution, negative binomial regression model was applied, assuming that the dependent variable has the negative binomial distribution (Cameron and Trivedi, 1998). Negative binomial regression is based on the log-transformation of the conditional expectation of the response variable and needs an exponential transformation of the estimated coefficients for assessing the effect sizes (Stieglitz and Dang-Xuan, 2013).
The model to explain the level of consumer interaction can be expressed in the equation below. Description of each variable is provided in table 6.
log P $ ∗ = &)log(U-V*+B;B) + &.log(E9R-/9:,X) + &1log *7Q-I+ +
&4log(*X-*+YXZ) + K
Table 6. Description of variables for model 3
Acronym Name of variable Description
y Growth rate of interaction The number of new likes, shares, and comments per hour. NewPosts Number of new posts The number of new posts on the previous extract.
TimeSinCr Time since created Time period since the post was created, expressed in hours. PageFo Page followers The number of page followers at the day the data was extracted. PrePoGrR Previous post growth rate Hourly growth rate of interaction of the previous post.
35 4. Results
In this chapter, the results of the data analyses are presented. First, correlation matrices are discussed. Further, a description of the data preparation will be given, followed by the actual hypotheses testing based on the results of regression analyses. Direct relationships between independent variables and response variables are discussed.
4.1. Correlation analysis
An overview of descriptive statistics and correlations are presented in appendices 1-3. A first observation derived from the tables is that none of the variables are highly correlated. In regards to model 1, variable post type photo is moderately correlated with post type video (r = -.568, p < .01) and weakly correlated with page following moderate (r = -.452, p < .01). In addition, average sentiment score has a moderate positive correlation with the average message length (r = .477, p < .01). Correlation makes sense as the likelihood of a message to
contain emotional words increases with the length of a message. Furthermore, total posts centered is moderately correlated with total posts centered squared (r = .603, p < .01) and
spacing category (r = -.649, p < .01). Finally, variable page following moderate has a moderate negative correlation with page following high (r = -.546). Although there are no highly correlated variables in model 1, variance inflation factor (VIF) was calculated later together with model estimates to test for multicollinearity.
In model 2, most of the variables are either weakly correlated or not correlated at all. Only post type photo is highly correlated with post type video (-.703, p < .01). It can be explained by the fact that the baseline post type (link posts and status posts) has relatively low frequency. Therefore, variable post type photo has the value of 1 and variable post type video has the value of 0 or vice versa, which makes the two variables highly correlated. As a results,
36
variance inflation factor (VIF) was calculated later in the process in order to test for multicollinearity.
In model 3, all variables were weakly correlated. New interaction per hour and growth rate of interaction of the previous post display the strongest positive correlation of all the
variables (r = .217, p < .01) and new interaction per hour and time since the post was created exhibit the strongest negative correlation (r = -.140, p > .01). The later can be explained by the diminishing rate of growth of interaction. The more time passes since the creation of the post, the less interaction the post gets. As for the positive relationship between the new interaction per hour and the growth rate of interaction of the previous post, it could be argued that higher
interaction with the previous post drives traffic to the Facebook page of a brand which leads to an increase in interaction with other brand-generated content. To conclude, correlation coefficients provide an early indication of the fitness of variables for the regression model.
4.2. Hypotheses testing – model 1
Hierarchical regression was used to isolate the effects of control variables and examine the effects of post frequency on log-transformed total interaction. First, 5 control variables were entered into equation: message length, post type photo, post type video, page following moderate, and page following high. Model was statistically significant in step 1 with F(5, 659)
= 32.377; p < .001, and explained 19.7% of variance in interaction.
Number of weekly posts centered, and number of weekly posts centered squared were
entered into the model in step 2. Model explained 24.5% of total variance and was statistically significant with F(7, 657) = 30.515; p < .001. Number of total posts and its squared term explained additional 4.8% in variance, after controlling for message length, post types, and the level of page following (R2 change = .048; F(2, 657) = 20.958; p < .001).
37
In step 3, spacing of brand-generated content was entered into the equation. Model explained 26.9% of variance and was statistically significant with F(8, 656) = 31.473; p <.001. Spacing of brand-generated content explained additional 3.2% of variance and the change in
R2 was statistically significant with F(1, 659) = 29.058; p< .001.
Finally, in step 4, interaction terms for average sentiment amount and total posts (both centered and centered squared) were entered into equation. Final model explained 28.7% of
variance and was statistically significant with F(11, 653) = 23.861; p < .001. Interaction terms explained additional 0.1% of variance, but the change in R2 was statistically insignificant with F(1, 653) = .437; p > .05. In addition, none of the interaction terms were statistically significant for a confidence level of 95%. Therefore, it can be concluded that moderating effect of sentiment on the relationship between posting frequency and the level of consumer interaction is not present (hypothesis H3 rejected).
Furthermore, in the final model, six out of eleven variables were statistically significant, with the page following moderate recording highest positive Beta (β = .510, p < .001) and total
number of posts centered squared recording highest negative Beta (β = -.212, p < .001), which indicates an inverted u-shape regression curve and provides support for hypothesis H1. In addition, spacing between posts was positively related to the level of consumer interaction (β = .269, p < .01). This supports the notion that time period (or space) between the posts reduces the wear-out effect of frequency on the effectiveness of the post and, thus, positively affects the level of consumer interaction (H4 accepted). Finally, post type photo (β = .116, p < .05),
page following high (β = .396, p < .001), and total posts centered (β = .506, p < .001) were positively related to the log-transformed total weekly interaction. Finally, each variable has an acceptable level of variance inflation factor (VIF) as it does not exceed the recommended maximum value of 10 (Hair et al., 1995). Therefore, regression model does not have a multicollinearity problem. Model estimates and other statistics are presented in table 7.