Let’s get engaged : predicting facebook post engagement within social automotive advertising

Let’s Get Engaged: Predicting Facebook

Post Engagement within Social Automotive

Advertising

Master’s thesis

Communication Science: Persuasive Communication Word count: 7391

Abstract

In the recent years, Facebook has earned its spot as an integral part of the digital experience of internet users. Such popularity has been widely recognised by businesses that have incorporated branded fan pages into their promotional strategies. This phenomenon had turned the social networking site into a cluttered space in which brands are doing their utmost to have their voice heard and engage their potential customers. The following research

presents a framework that focuses on exploring which communication characteristics make that user engagement more likely to occur. Random forests and regression analysis were utilized in an aim to determine which post characteristics are most predictive of user

engagement. Results indicate that the type of media (image or video) within a Facebook post is the characteristic that is most closely correlated to user engagement rate, partially followed by social content with an educational theme. Based on the results, several implications for future research and business practices were discussed.

Keywords: Facebook fan pages, post characteristics, engagement rate, random forests, regression analysis, car advertising

Table of contents

1. Introduction ……….4

2. Theoretical framework ………....6

2.1. Identification (Type of posting) ………...7

2.2. Content format (Type of media) ………..8

2.3. Categorisation (Content appeals) ……….9

2.3.1. Emotional content appeals ………..…………10

2.3.2. Functional content appeals ………..…10

2.3.3. Educational content appeals ……….11

2.3.4. Community content appeals ………11

2.4. Models for predicting user engagement………..12

3. Methodology .………14 3.1. Sample ………14 3.2. Coding categories ………...……16 3.3. Coding process ………...17 3.4. Dependent variables ………...19 3.5. Coding reliability………19 4. Results ………...20 4.1. Descriptives ………20 4.2. Random forests ………..21 4.3. Multiple regression ………24 5. Conclusion ……….26 6. Discussion. ………29 7. Reference list ………31 8. Appendices……… 36 8.1. Appendix A……… 36 8.2. Appendix B ………39 8.3. Appendix C ………40

Introduction

The creative advertising industry is notoriously competitive and is characterized by a high employee turnover rate, while being under constant threat by emerging technologies, automation, and general marketing budget cuts (Richards, 2016). It comes at no surprise, therefore, that every possible advantage is seized by advertisers and marketing professionals in their struggle to gain a competitive edge. A development, that a large part of commercial professionals has recognized as an opportunity, is the emergence of social media into the mainstream. Social networking sites (SNS) have stimulated new ways in which people communicate, make decisions, socialize, collaborate, learn, entertain themselves, interact with each other or even do their shopping (Sabate, Berbegal-Mirabent, Cañabate, & Lebherz, 2014).

Currently, social media are responsible for 30% of the time an average consumer spends online (Mander, 2017). Perhaps unsurprisingly, social networking sites therefore have become an increasingly more sought-after promotion channel by marketing professionals. Roughly half of social networking users follow brands on platforms such as Facebook and Instagram (Mander, 2017). Also, respectively 44% and 53% of the active user bases of Facebook and Instagram have reportedly ‘liked’ their favourite brands on these social media. According to a study from Econsultancy and Adobe, which surveyed international marketers on how their spending on various digital marketing channels would change in 2017, social media marketing remains a top priority for brand professionals (Richter, 2017).

Essentially, the investments in social media are made with the expected return of being able to interact with brand followers and foster customer relationships. A common strategy for achieving that is by cultivating brand communities in the form of ‘brand fan pages’ where customers have the option to interact with a company by engaging with the content provided by a brand (McAlexander, Schouten, & Koenig, 2002; Muñiz & O'Guinn,

2001). Social media brand followers tend to not only be loyal and committed to the brand, but also receptive to being informed about company-related activities (Bagozzi & Dholakia, 2006). According to Dholakia and Durham (2010) social media followers have been shown to be more frequent physical store visitors, to generate more word-of-mouth, and to be more emotionally connected to the brand of question than non-brand followers. From the customer perspective, motivations to follow brands on social media include the ability to monitor promotions and discounts, find out product information, reach customer service, consume entertainment content, or simply offer feedback (Ali, 2015). In other words, brand fan pages provide a platform for customers’ relationship with the brand in question and a channel that could benefit both brand followers and businesses. The ease of feedback provision in the form of customer response and engagement options such as liking or sharing make measuring individual brand posts’ performance an industry-accepted indicator for brand health. In this paper, the aforementioned customer feedback to posts in a branded social media page has been defined as ‘user engagement’. The present study tests the idea that certain

characteristics of social media posts can influence user engagement either positively or negatively.

When it comes to the specific industry of choice, a review of the literature showed that branded social content has been explored in the context of a number of industries,

including automotive (Tafesse, 2015), yet further knowledge on the latter could be derived by combining social content exploratory frameworks. The present study focuses on the

automotive industry, as the field still lacks scientific direction on what type of content engages most. Furthermore, automotive brands on social media and particularly Facebook provide enough available and, most importantly, fluid content to explore engagement

patterns. These expected patterns are further explained below. Thereby, the research question is defined as follows:

RQ: Which post characteristics in Facebook posts of automotive brands can predict user engagement?

Theoretical framework

The concept of social media engagement is often discussed in the context of brand post popularity. The focus in exploring the drivers of popularity has been on scrutinizing certain social post characteristics and their likelihood to correlate with customer engagements. For instance, De Vries, Gensler and Leeflang (2012) have attempted to identify the most

prominent predictors of brand popularity by using theory from the field of banner and word-of-mouth (WoM) literature. In their paper, the authors criticized the lack of theoretical foundation of management-oriented studies about post popularity, and, to be more precise, the lack of formal testing of which commercial activities enhance brand post popularity. Despite the continuous aspiration on behalf of the academic world to better understand the trends behind the drivers of social engagement, studies on the topic have mostly been such of a descriptive nature (De Vries et al., 2012). The present paper aims to take a step further and presents a predictive model that identifies a number of social post characteristics, and

measures their capabilities to predict follower engagement. Essentially, follower engagement provides a metric for measuring the impact of a post, which is constituted of the sum of the total amount of Facebook ‘likes’, ‘reactions’, ‘shares’ and ‘comments’ a social post has.

This study employs a slightly modified version of the overarching framework related to brand post characteristics suggested by Moro, Rita, and Vala (2016), namely dividing the post characteristics in three subgroups:

• Identification — features that allow identifying each individual post (‘time of posting’); • Content format — the type of social content (‘photo’ or ‘video’);

Past scientific work around the proposed post characteristics is described for the remainder of the review section.

Identification: Type of posting

Considering the nature of Facebook when it comes to a heightened information overload, as part of the user experience, it is essential to consider posting time as a variable that could determine post performance. For example, research from Golder, Wilkinson, and Huberman (2007) provides evidence that most of the user activities on Facebook happen on weekdays, whereby brand activity on Facebook is distributed nearly exclusively between Mondays and Fridays (Buddy Media Inc., 2011). In support of this argument is the

benchmark for engagement, as research by Rutz and Bucklin (2011) found that click-through rates on online advertisement decrease significantly over the weekend. Consequently, day of publication is known to influence post performance (Sabate et al., 2014). Identifying peak patterns of user activity can further add to the importance of the variable. Golder et al. (2007) claim that Facebook activity is in its pinnacle in evenings, while engagement rates during this period exceeds the daily average by 20%. Related to this, however, is the idea that trying to communicate a branded message in a less cluttered digital space would be more beneficial for the post’s popularity, as it would appear on top of users’ news feeds. The paper by Sabate et al. (2014) provides further evidence of the claim that publication during business hours tends to result in increased engagement. Considering the aforementioned claims, the following hypothesis regarding the ‘time of publication’ variable is deducted:

H1a. Branded Facebook posts posted during office hours predict user engagement. H1b. Branded Facebook posts posted on working days predict user engagement.

Content format: Type of media

The type of content (video, photo or a link) is another post characteristic that has been proven impactful in predicting social media performance. In the paper by Moro, Rita, and Vala (2016), the authors found content type to be the most relevant feature for determining likelihood of user engagement. The type of social content has most often been discussed in relation to its vividness (Pletikosa Cvijikj & Michahelles, 2013). Luarn, Lin, and Chiu (2015) found that medium levels of vividness (such as a post involving a text caption and an image) result in most engagements due to the simplistic and effortless capabilities to attract user attention, as compared to videos which typically require more processing effort. Brookes (2011) support such evidence with claims that images receive 22% more engagements than video posts, making those prevailingly more impactful on followers’ proactive attitude towards the post in question. Recent trends however, hint towards video domination on social media (Grosman, 2017). With the latest technological developments around video viewing (such as the ‘auto-play feature’) in the social networking platforms themselves, videos have been competing successfully with images for most engagements. Making use of users’ sensory processing system by utilizing attention grabbing techniques such as dynamic animations and contrasting colours of still images is another reason for brands to choose interactive imagery for their promotional materials on SNS (Sabate et al., 2014). This paper follows the reasoning identified by Sabate et al. (2014) and distinguishes content type (video and image), rather than relying on vividness. This eliminates subjective bias and disregards potential personal perceptions in users’ eyes. Considering the variance between scientific and practical evidence on which content type engages most, for this paper it is assumed that both videos and images could predict engagements equally.

H2a. Branded Facebook posts that include a video predict user engagement. H2b. Branded Facebook posts that include a photo predict user engagement.

Categorisation: Content appeals

The present paper also employs a qualitative categorization of appeals. A common framework in academic literature in the past has been the utilization of uses and gratifications theory as an approach to understand the goals and motivations of individuals to engage with content. Furthermore, evidence by Chauhan and Pillai (2013), De Vries et al. (2012), and Pletikosa Cvijikj & Michahelles (2013) suggests that content type has a vast potential to drive consumer engagement. Drawing on uses and gratification theory, first developed by Katz (1959), studies have shown that social media has mostly been used with the purpose of social interaction, identity formation, information acquisition, entertainment and economic reward (Tafesse, 2015). Past studies have observed the engagement behaviour of brand followers on social media in line with those motivations. Hence, social media content is often designed with entertainment, relational, informational, and transactional themes in mind (De Vries et al., 2012; Sinclaire & Vogus, 2011). The present paper suggests a division that is based on a study by Tafesse and Wien (2017), which, in turn, outlines 12 exhaustive and mutually exclusive categories of brand posts. The research focuses on 4 of those that fit the automotive industry, the communication around it and the motivations outlined in this paragraph. These are ‘emotional brand posts’, ‘functional brand posts’, ‘educational brand posts’, and ‘brand community’. Entertaining social media content, which in this case of the present paper entails emotional themes, is deemed a vital factor for online community participation (Park, Kee, and Valenzuela, 2009). For instance, Cvijikj & Michahelles (2013) found higher response effect for entertaining brand posts. The same is valid for informational content, which could be described as content around functional and educational brand traits (Dholakia, Bagozzi, &

Pearo, 2004). Muntinga, Moorman, and Smit (2011) further discover that while information and entertainment are the primary sources of motivation to engage with a brand online, remuneration content also presents a frequently mentioned stimuli for online community contribution. The latter theme often manifests in community-centred posts in the form of sweepstakes, user-generated content or any way the user might directly benefit from participating in a branded activity. The chosen content appeals are described in more detail below.

Emotional content appeals

The first category entails emotional brand posts, or in other words, posts that

constitute emotion-laden language, inspiring stories or humor in order to stimulate affective response by consumers (Tafesse & Wien, 2017). This enables the brand to create a

connection with social followers on an emotional level (Davis et al., 2014). Integrating emotion in social advertising has been acknowledged by academia as a fruitful strategy and, perhaps unsurprisingly, receives popularity in the field. In a study by Tafesse and Wien (2017), emotional posts accounted for the largest category sample across the sample in number of industries - hinting to the importance of this e-post characteristic. Furthermore, emotional posts are shown to be capable of communicating additional message cues such as functional, educational and brand image (Davis, Piven, & Breazeale, 2014).

Functional content appeals

On the other hand, functional brand posts promote specific practical traits and

attributes of a product or service (Tafesse & Wien, 2017). These tend to focus on the benefits of the brand, including communicating qualities around competitive advantage, affordability, efficiency, style criteria, and performance. Tafasse and Wien identified functional posts as a

category that most benefits consumers’ abilities to make informed purchase decisions by providing them with meaningful product or service information. The latter was believed to involve also product features or technical specifications. Highlighting newly developed products and solutions, functional attributes of brand posts also reward consumer’s novelty-seeking behaviour (Tafesse, 2015). Tafesse and Wien (2017), however, also identified a possible obstacle that could potentially negatively impact engagement success with functional posts - namely, the fact that they are often seen as a form of a sales pitch, thus potentially raising resistance from consumers.

Educational content appeals

Educational brand posts were another strongly pronounced category in the study by Tafesse and Wien (2017). These concern primarily the task of educating customers on ways to acquire new skills when it comes to the application and use of brand-related products or services. The authors outline the idea that educational posts tend to often discuss broader industry trends and potential issues that could be relevant to the consumer - with those being often presented in conjunction with a persuasive message aimed to raise brand benefits. Tafesse (2015) further elaborates on the effects that educational posts have on consumers, namely, their ability to empower brand followers to solve product-related issues or discover creative ways to apply those in their day-to-day lives.

Brand community appeals

The final category that was explored in this paper is brand community posts. While the category accounted for only 7% of all branded posts in the benchmark study by Tafesse and Wien (2017), the trend in the car industry may differ. The field of automotive is

BMW Stories underline the importance of cultivating a strong brand community. Zaglia (2013) further identified the importance of brand community posts and their ability to counter the perception that brand pages generally lack a communal spirit or that consumers fail to develop a strong feeling of belongingness just by following a business on social media.

Based on previously outlined research on the post characteristics category, the present paper assumes that the suggested themes will have an impact on the amount of user

engagement, as outlined in the hypotheses below.

H3a. Branded Facebook posts that include emotional content predict follower engagement. H3b. Branded Facebook posts that include functional content predict follower engagement. H3c. Branded Facebook posts that include educational content predict follower engagement. H3d. Branded Facebook posts that include community-stimulating content predict follower engagement.

Models for predicting user engagement

Fundamentally, previous scientific research has tested the influence of Facebook brand-page posts on engagement in a number of ways. More specifically, scientific works have explored content themes and forms that correlate to a number of metrics of engagement on the SNS. A common theme when developing the model across a number of these articles has been to group the variables that are supposed to predict engagement. For instance, Luarn et al. (2015) differentiate between media (such as vividness and interactivity) and content type (content themes) of posts when trying to predict user engagement. Pletikosa Cvijikj and Michahelles (2013) take a similar approach by including a category regarding posting time. Post characteristics such as position of post on page; valence of comments; post reach and clicks; amount of text are among have also been explored to possess a correlation to

engagements (Trefzger, Baccarella, & Voigt, 2015; Tafesse, 2015; De Vries et al., 2012; Moro et al., 2016). Sabate et al. (2014) followed the same reasoning in their model

development, but they factored the number of followers as a moderating variable. Unlike the rest of the aforementioned studies that solely use Facebook likes comments and shares, the present paper factors the number of followers at the time of the posting as an important novelty aspect. In this case, the target variable is identified as engagement rate, which is explained in detail in the methodology section and calculated to factor page popularity, thus creating a fair comparison between posts. In that sense, the study by Pletikosa Cvijikj and Michahelles had most similarity with the present one, as the authors factored the number of impressions when determining engagement.

In line with that way of thinking, the conceptual framework (Figure 1) is developed as such:

Figure 1. Hypothesised conceptual framework for relations between post characteristics and online engagement. Positive effect is assumed.

Another common pattern with studies which employ models that are meant to explore the relationship between post characteristics and user engagement was doing so by evaluating the performance of each individual post characteristic. For instance, Sabate et al., de Vries et al., and Tafesse employ regression analyses to explore any existing correlations between their variables. Authors such as Luarn et al., and Trefzger et al., on the contrary, compared the effects of post characteristics on post popularity by comparing the means, or in other words, by conducting analyses of variance (ANOVAs). This paper utilizes a model, which has presently not been used for comparing Facebook post characteristics, namely random forests. The output was then benchmarked to a multiple regression, as the said analysis being the more common approach in academia.

Methodology Sample

The current research involved a content analysis of user engagement strategies by automotive brands on Facebook. Ten car brands were chosen on the basis of a report by Brandwatch (2016), which ranked the social media presence of all automotive companies. The social media monitoring company ranked the brands based on five criteria, namely social visibility, general visibility, net sentiment, reach growth and social engagement. Considering all factors, the top ten car marques with the highest score were used for the current paper - namely Audi, BMW, Dodge, Honda, Jeep, Lexus, Mercedes-Benz, Porsche, Volkswagen, and Citroen. Table 1 provides further information on the brands, including the Brandwatch social

composite score and number of followers of the brand’s Facebook-page per December 2017. In order to get a grasp of the type of social content that is unrelated to local market specifics, the Facebook pages of all automotive brands were chosen with the region set to ‘global’. Audi was the only Facebook brand page for which the US region was used, due to

the brand having separate pages for each region around the world. Choosing either the global or US options also made it possible to have a dataset in English, which eased the coding process.

The time frame involved 30 days of content (01/11/2017 until 30/11/2017), thus data from the period were collected from each of the 10 Facebook pages during this period. This resulted in a sample size of 500 Facebook posts (observations) in total. Each of the brands produced a different amount of posts, as indicated in Table 2 above, with Mercedes-Benz being the most prolific with 117 posts, while Audi’s Facebook presence was the least active with 11 posts for the period.

Coding categories

The codebook and definitions were developed based on the previously outlined theories around Facebook post characteristics. Four coding categories were outlined, namely descriptives, form characteristics, content characteristics and user engagement.

Descriptives. Descriptives handle characteristics from the Facebook posts that are unique to the post itself and are used to identify it. Those include the post hyperlink, date and hour of the publishing, and the name of the automotive brand the post belongs to.

Form characteristics. Form characteristics describe the specific content form that was posted - In the context of this paper, specifically the presence of images and videos in the posts was coded.

Content characteristics. The third coding category, content characteristics, was

developed to distinguish specific themes within the Facebook post content. Such themes were previously identified as ‘emotional’, ‘functional’, ‘educational’, and ‘community-centric’ posts, and a separate category was assigned to each of the posts. It was possible to assign more than one content characteristics to a single post, which was often the case for posts including ‘functional’ information about fuel consumption next to a text-based emotional appeal.

Engagement signals. The last coding category involved marking the total number of signals of engagement with the posts by users, which was the sum of likes, shares, comments and reactions. Based on the hypotheses, a separate nominal-dichotomic variable for each of the characteristics was created and consequently coded with either ‘1’ or ‘0’, with ‘1’ stating the post contains the characteristic of question. In comparison, the dependent variable was defined as numerical. How the specific variables within these coding categories were defined, is described in more detail below.

Coding process

For the data-collection process, the social analytics and statistics tool Fanpage Karma was used. The tool automatically identifies a number of variables per post that fit within the previously outlined coding categories.

Descriptives. For the descriptives, Fanpage Karma exports a ‘Post link’ which

presents an URL link to the individual post. Furthermore, the tool extracts the specific ‘Time of day’ and ‘Date’ the Facebook post is published. In the context of this study, posts that were published during office hours (9am until 6pm) were manually coded as such, while the same was done for posts that were posted during the 22 working days in November.

In total, 401 (80.2%) Facebook posts were coded as being published during a weekday, while the number for those that were posted between 9am and 6pm was 262 (52.4%).

Format characteristics. The relevant form characteristics were automatically identified by Fanpage Karma as well. The tool defined whether a certain post involved an image (photo or picture) or a video (animation or a video uploaded on the Facebook

platform), and those were coded accordingly. The coding procedure resulted in 127 videos, which is 25.4 per cent of all posts, whereas 360 of the observations (72%) in the sample included images.

Content characteristics. Furthermore, separate columns were also created for each of the relevant content categories based on the themes that the branded post communicates, namely ‘Emotional’, ‘Functional’, ‘Educational’, ‘Community’. Here, manual coding was conducted and the ‘Post Caption’ was used as a source of information to determine how to code the relevant themes. The post caption is the written text that accompanies each video or image in a Facebook post. Coders assigned one or more themes to each of the posts

Appendix A1 with an extensive guideline on how to determine the relevant content themes. A summary of the coding categories is given below.

Posts containing an emotional appeal such as joy, humor, affection, sadness,

adventure, excitement or nostalgia were coded as 'emotional'. Furthermore, the release of new advertisement was also coded as emotional content due to the nature of feelings these are developed to incite.

The 'functional' theme was defined with content involving rational statements. In other words, any “practical information about the cars that communicate some sort of

competitive advantage were coded within the aforementioned theme. Any sort of information that provides technical feature of the car brands was also considered within ‘functional’ category.

The 'educational' theme was defined to handle post captions that aim to teach social media users not only about best practices around service usability, but also content that was aimed at informing users of a new model, model edition or external car service releases. A strict guideline was developed so that coders could differentiate between 'functional' and 'educational' posts.

Lastly, posts that was directed to the brands' offline or online community bases were coded under the variable 'community.’ Examples included reposts of user-generated content such as amateur photographs by car enthusiasts, organized car show events, Facebook raffles, and content around external sponsorships.

In total, 165 emotional (33%), 121 functional (24.2%), 66 educational (13.2%) and 175 community-based (35%) posts were coded as such.

Dependent variables

In order to conduct the analysis towards the target variable, the present paper looked at the ‘Total number of engagements’, which entailed the sum of the variables ‘Number of Reactions’, ‘Number of Likes’, ‘Number of Shares’ and ‘Number of Comments’.

Furthermore, the idea that the size of the Facebook brand page directly influences the total number of engagements was factored in this paper by introducing the variable ‘Engagement Rate’. This is an industry-accepted measure in which engagement volume is divided by the number of users or events that could have triggered the action (Yamaguchi, 2017). In the context of this study, it was measured what percentage of the followers at the time of the publication engaged with the post. This resulted in the creation of a new variable,

‘Engagement rate’, which is a division between previously outlined ‘Total number of engagements’ and ‘Number of fans at the day of post’. The number was then turned into a percentage. The mean value of the variable engagement rate was reported at 0.078, meaning that, on average, approximately 0.08% of the car brand followers engaged with the content provided by company pages. The range differed for the variable with 0 (meaning no user engagement for a single post) on the minimum end and 2.214 being the maximum value. Standard deviation equaled 0.142.

Coding reliability

Both the first coder (the author) and the second coder (also a student from the Master’s programme in Communication Science at the University of Amsterdam) were provided with a coding manual and coding sheet with instructions on determining the relevant posting time, type of media and content category (see Appendix A2).

Inter-coder reliability was evaluated on 10% (50 posts) of social media posts with representation of all brands from the sample (Neuendorf, 2016). This was done with the aim

of evaluating coding’s accuracy and reliability, and identifying disagreements between coders (Fico, Lacy, & Riffe, 2008). Krippendorff’s alpha value was calculated at 0.91, which is notably above the cutting point of 0.80, as identified by de Sweert (2012), and gives a strong reason to consider the coding procedure as reliable within content analysis. The following section looks at the results of the analysis.

Results

0.08% of the Facebook page followers engaged with the branded posts on average (M = .078, SD = 0.14). Meanwhile, for the predictors of engagement rate, three of the independent variables were a majority within the dataset. Namely, 72% of Facebook posts contained images (n = 360, M = .720, SD = 0.45), 80% of the entire sample (N = 500) was published during a weekday (n = 401, M = .802, SD = 0.40), and 52% during working hours (n = 262, M = .524, SD = 0.50). Respectively, 25% of the sample included a video (n = 127, M = .254, SD = 0.45). From the content characteristics, 33% of the branded car posts had an emotional appeal (n = 165, M = .330, SD = 0.47), 24% of the sample communicated a brand or service function (n = 121, M = .242, SD = 0.43), while 13% aimed to educate about the brand (n = 66, M = .132, SD = 0.34). Lastly, exactly 35% of the dataset engaged the community in a way (n = 175, M = .350, SD = 0.48).

Assuming all the independent variables have an effect on the engagement rate of automotive brands’ Facebook post, the aim of this study is to evaluate the different degrees of importance of each of the predictors. Therefore, a framework that tests the relationship

between post engagement rate and media type, time of posting, or content appeal was considered for this paper. The approach to testing the hypotheses was to use a machine learning model that would provide an indication of the importance of each of the post

features and later test those towards an established statistical method for modelling a relationship between variables.

Random forests

Hence, random forests modelling was utilized as a method for supervised machine learning due to its ability to construct a multitude of decision trees (Breiman, 2001). Scikit-learn software identifies random forests as “a meta estimator that fits a number of classifying decision trees on various sub-samples of the dataset and use averaging to improve the

predictive accuracy and control overfitting” (Scikit-learn.org, n.d.). Decision trees essentially perform the process of going from observations about an item (the independent variables) to conclusions about the item's target value (represented by the variable engagement rate), and thereby ranking predictors in importance (Kim, Lee, Shaw, Chang, & Nelson, 2011).

The algorithm for running random forests was input in Jupyter Notebooks and relied on the utilization of the RandomForestRegressor Python tool for data mining and analysis from the Scikit-learn library. As the algorithm was tweaked to fit the dataset, the decision trees estimated so-called ‘feature importance values’ for each of the predictor variables. Feature importance’ essentially rates the predictors in order from lowest to highest, with the latter being the more important in predicting engagement rate. The highest feature

information gain will have the highest linear coefficient in a multiple regression (Scikit-learn.org, n.d.).

Figure 2 below renders the importance of each feature (or that case, independent variable) that random forest assigns to them, and in turn, their predictive accuracy on engagement rate. The model returns a score, whereas features that produce large values for this score are ranked as more important than features which produce small values (Breiman, 2001). As can be seen, branded Facebook posts that had a video attached to them were the

most valid predictor of engagement rate, featuring an ‘feature importance’ rating at 0.663. In context, the said score ranges between 0 and 1, with 1 meaning that the feature of question fully predicts the target variable. Here, it is important to note that interpreting the importance of correlated variables within the random forest model presents a difficulty, which applies to most model-based feature selection methods according to Datadive.net (2014). Hence, computing the feature importances within random forest provides information on the importance of each variable in the context of the model alone. In other words, images in the car brand posts came second as the strongest feature importance variable with 0.271, which, interpreted, entails that images are roughly 2.5 times less important for predicting

engagement rate than videos. Whether the post involved the community theme was the third strongest predictor of engagement and scored 0.033. Figure 2 also shows the feature

importance scores of the remainder of the independent variables. Whether the content of the post involved educational themes such as the release of a new car model, model edition or external car service releases was 3 times less important than content that was specifically referred to the community. Posts published during weekdays were slightly more important (0.007) for predicting engagement rate than those that were posted during working hours (0.005). Functional statements that presented some sort of informative appeals, on the other hand, performed approximately 1.5 times better than content that is primarily emotionally-laden.

Figure 2. Ranking of the post characteristics’ according to their feature importance for the Random Forests model.

In sum, this analysis suggested that the inclusion of a video or image were much more important for promoting consumers’ engagement with car brands’ Facebook posts than any of the other features measured in the study. Community-centred brand posts were the most important post characteristic from the selected content themes, with educational content following closely. Lastly, random forests computed limited differences between the two factors measuring time of posting namely, publishing during office hours and weekdays. The following section will test the validity of the findings of random forests by conducting a more common relationship analysis – a linear regression.

Multiple regression

Essentially, the output from the random forests model served as an indicator of what to expect when compared to another widely used method in social science research - multiple regression analysis. The analysis explored the assumed linear relationship between the

predictors and engagement rate, and was developed to build upon the assumptions of variable importance outlined by random forests. Consequently, the approach tested whether the variables that were rated of highest predictor importance in the Random forest model also performed well in the context of a linear relationship. As previously mentioned, the clear dominance of video, image and community posts was expected to result in a high correlation coefficient in a regression. A multiple regression analysis was run using SPSS with the previously outlined post characteristics (‘image’, ‘video’, ‘weekday’, ‘working hours’, ‘emotional’, ‘functional’, ‘educational’, ‘community’) as independent variables, while ‘engagement rate’ was input as the dependent variable in the setup of the analysis.

Due to the nature of the test, the adjusted R Square is of importance. A value of 0.014 indicates that 1.4% of the total variability in engagement rate is explained by the model. The low percentage of explained variance suggests that the model explained very little of the variability in engagement with car brands’ Facebook posts. The conducted ANOVA test provided evidence that there is a lack of significance of the independent variables on

engagement rate at the p<.05 level for the conditions F(8, 491) = 1.901, p = 0.058. Rather, the null hypothesis (being that the model has no interpretative capabilities) could be rejected with 10% significance at p<.1.

Looking at the coefficients (Table 2), and more specifically the p-value of each, it becomes apparent that those Facebook posts that contained an ‘image’ could significantly predict user engagement, with significance level at 5% (p = .051). Also, educational content (t = 1.856, p = .064) returned a p-value of 0.064 indicating that posts that included new

information about the car brands marginally increased engagement rates. Whether the post was posted during a ‘weekday’ (t = 0.209, p = .835), or during ‘working hours’ (t = 0.775, p = .439) were proven to not be significant predictors in the model. Surprisingly considering the output from random forests videos (t = 1.309, p = .191) were also not found to be significant. No statistically significant interaction was found in the rest of the dependent variables, namely Facebook posts with ‘emotional’ (t = 0.538, p = .591), ‘functional’ (t = -1.276, p = .439) and ‘community’ (t = 0.824, p = .410) themes.

In order to further explore the relationship between the predictors and engagement rate, the multiple regression analysis was run at higher aggregation levels. Thereby, new variables with composite scores were composed between the independent variables with similar characteristics. In other words, the predictive score of Facebook posts that included an

image was paired with the posts that contained a video, resulting in the newly compound variable ‘media type’. The same procedure was done for the cases in which posts were during a weekday or during working hours, putting the variables under the umbrella term ‘business hours’. In order to determine how to aggregate the four themes within the content

characteristics, a principal component analysis was run. The test resulted in 2 components, as factor loadings paired Facebook posts that include functional and educational appeals

together (β2 = .621, β3 = .690), while emotional appeals were matched with community-centred content (β1 = .731, β4 = -.937). The multiple regression was then run with the newly created predictors, namely ‘media type’, ‘business hours’, ‘EmotionalOrCommunity’, ‘FunctionalOrEducational’, as independent variables, and ‘engagement rate’ as the target variable once again. Results of the newly-ran multiple linear regression indicated that there was not a collective significant effect between ‘media type’, ‘business hours’,

‘EmotionalOrCommunity’, ‘FunctionalOrEducational’ and engagement rate, (F(4, 495) = 1.217, p = .303, R2 = .002). The individual predictors were examined further and indicated that, similar to random forests and the regression which was ran at higher aggregation levels, ‘media type’ came closest to being significant (t = 1.896, p = .059). Whether the publication was posted during ‘Business hours’ (t = 0.736, p = .462), or had an

‘EmotionalOrCommunity’ (t = 0.630, p = .529) or ‘FunctionalOrEducational’ (t = -0.006, p = .995) content themes, were proven to not be significant predictors in the model.

Conclusion and Discussion

Deriving from the performed statistical tests, it is evident that the strongest predictors of engagement rate with the posts of automotive brands on Facebook are the types of media included in the branded Facebook post. The model based on random forest, decision trees

indicated that the inclusion of a video in a car-branded Facebook post was the strongest predictor of engagement - twice more important than the inclusion of an image. The regression analysis showed a similar outcome in support of the hypothesis that the use of a video in a social media post is correlated to the amount of user engagements it receives. The use of images in as Facebook post of automotive brands were also found to be predictive of engagement rate with statistical significance. Hence, the results suggest that H2a (the

assumption that social media posts including a video predict engagement) is supported, while H2b which claims that images are correlated to user engagements on Facebook is also

supported.

The findings are in line with the results of a number of past studies. For instance, post vividness as a characteristic of a social media publication has been found to significantly affect engagement with social media posts in earlier studies by Luarn et al. (2015) and De Vries et al. (2012) The same was concluded by Sabate et al. (2014), but described as post richness in their research. The predictive performance of both image and video post

characteristics comes hardly as a surprise considering the vivid nature of communication of car brands in general. Considering that out of 500 posts from the sample only 11 included neither a video or an image, this speaks for the car brands’ common reliance of visual appeal.

Neither having the branded Facebook post published on a weekday, nor during working hours, proved to be predictive of engagement rate, meaning that both H1a and H2b must be rejected by the present study. These findings contradict earlier findings from studies by Sabate et al. (2014) in regards to posting during business hours and by Pletikosa Cvijikj and Michahelles (2013) when it comes to differences between the days of the week and engagements. A possible explanation for the lack of a significant correlation between the time that a post appeared and engagement rates could be the global nature of the pages. Choosing the international Facebook business pages of the brands made sense for practical

reasons, such as language benefits and being able to provide a good overview of a brand’s communication efforts on social media. However, the global nature of the pages meant that there is a high chance of discrepancy between the time zones of the intended target groups, origin of branded page and, lastly, the coders of the present paper. This limitation is

especially valid for posts that were deemed to have been posted during working hours. Thus, future research might focus on a sample of only regional pages of car brands when comparing predictors of Facebook engagements related to time when certain content was posted.

With regard to content characteristics of the car branded Facebook posts, the results from the compared statistical models differ. Albeit with a fairly lower so-called ‘feature importance value’ than content involving an image or a video, the random forest model determined community-centered posts on Facebook to be more than three times more

indicative of engagement rate than the remainder of the observed content themes. A possible explanation for this result could be the fact that often the sole purpose of posts that

specifically address a community is to instigate a certain action or response in the form of user engagement. Unfortunately, this assumption was not supported by the regression analysis, as community posts were not deemed significantly predictive within the context of this model. This finding is in line with a paper by Pletikosa Cvijikj and Michahelles (2013) which found that remuneration content had no effect on sharing ratio. Contrary to that claim, De Vries et al. (2012) provided evidence that the number of comments on a post could be enhanced by an interactive post characteristic or a question directed towards the social page user base. All in all, H3d was not supported.

The regression analysis showed that branded car posts that involved an educational element in their content came closest to being a significant predictor. This once again creates evidence in support of a finding by Pletikosa Cvijikj and Michahelles, namely that branded informative content increases the level of engagement through liking and commenting.

Thereby, as the predictor significant only at 10% significance level, H3c is only partially supported. Branded posts that included emotionally-laden or functional content were not found significant in terms of predicting engagement rate in any of the observed models, thus neither H3a and H3b were supported. Further exploration of different content themes would be worthwhile.

In conclusion, the research question of the present paper sought to explore which of the formerly-mentioned characteristics of a Facebook post are predictive of follower engagement in the context of automotive social media branding. Results provide evidence that social posts that include an image or a video of car-related content are considered to be the characteristics that are most closely related to user engagement rate, followed narrowly by an educational theme as a post feature and with less affirmation of significance in the model.

Discussion

A common pattern in past literature on social media content is the testing of potential relationships between the post characteristics and the different types of user engagements, namely likes, shares and comments (De Vries et al., 2012; Trefzger, Baccarella, & Voigt, 2016; Tafesse, 2015; Pletikosa Cvijikj & Michahelles, 2013). Focusing on the engagement rate allows a fair comparison between branded Facebook pages regardless of their respective sizes, yet, the measure disregards possible correlations between the exact types of Facebook engagements and the predictors. This presents an opportunity for future research in which the relationships between post characteristics and the types of user engagement are explored in more in depth, for instance by utilizing the decision trees model performed in this paper.

The present study has some limitations that have the potential to tamper its results’ generalizability. Firstly, the comparison between the time of the publication of global pages

is dependent on the research authors’ locations. While the day of the week in this case could still be considered an accurate measure for both global and regional marketers, the result around time of the publication can not be generalized to regional social media strategies. Comparing branded Facebook pages within a single country market would have made the results more applicable to business professionals, as this reflects user engagement behaviour during the day more accurately. Secondly, another limitation of this study is the similarity in choice of content theme categories of Facebook posts and their observed recurring

dependence on the brand’s general tone of voice. To illustrate that, functional and educational content was often hard to distinguish and could have been combined within a single theme, while Dodge having primarily community-oriented posts was indicative of the uniformity in content that the dataset had. Related, sponsoring certain posts as paid media is another factor that could have had an effect on the number of engagements a single post had received. This is again largely related to the promotional strategy which social media marketers behind each individual car brand have utilized. Future research can tackle the aforementioned limitation by exploring the relationship between the publisher by brand name and the number of engagements.

In conclusion, the present paper presents a framework that could be utilized by social media marketing professionals within the automotive industry in their pursuit of tailoring an effective Facebook strategy. It is important to note that the studied SNS content was very specific to the automotive industry, be it by relying on vivid imagery or by communicating a specific car-related theme. That being said, the results from the present paper are

generalizable to communication channels within the digital ecosystem of a brand, such as other SNSs like Instagram and Twitter, or the official corporate blog. On an industry-level, the evidence from this research could be valid for a number of industries that sell not only a

product, but also an experience, as in travel, or a status symbol, as is the case for a large number of tech and luxury brands.

Literature list

Ali, A. (2015, May 26). Why Do We Follow Brands on Social Media? [Infographic]. Retrieved January 30, 2018, from https://www.socialmediatoday.com/social-business/asadali/2015-05-24/business-social-media-infographic

Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32.

Bagozzi, R. P., & Dholakia, U. M. (2006). Antecedents and purchase consequences of customer participation in small group brand communities. International Journal of

research in Marketing, 23(1), 45-61.

https://doi-org.proxy.uba.uva.nl:2443/10.1016/j.ijresmar.2006.01.005

Brookes, E. J. (2011). The Anatomy of a Facebook Post: Study on Post Performance by Type, Day of the Week, and Time of Day. Vitrue Inc. Retrieved.

Brandwatch. (2016). The Automotive Industry (Rep.). Brandwatch.

Cvijikj, I. P., & Michahelles, F. (2013). Online engagement factors on Facebook brand pages.

Social Network Analysis and Mining, 3(4), 843-861. https://doi-org.proxy.uba.uva.nl:2443/10.1007/s13278-013-0098-8

Chauhan, K., & Pillai, A. (2013). Role of content strategy in social media brand

communities: a case of higher education institutes in India. Journal of Product & Brand

Management, 22(1), 40-51.

https://doi-org.proxy.uba.uva.nl:2443/10.1108/10610421311298687

De Swert, K. (2012). Calculating inter-coder reliability in media content analysis using Krippendorff’s Alpha. Center for Politics and Communication, 1-15.

Davis, R., Piven, I., & Breazeale, M. (2014). Conceptualizing the brand in social media community: The five sources model. Journal of retailing and consumer services, 21(4), 468-481. https://doi-org.proxy.uba.uva.nl:2443/10.1016/j.jretconser.2014.03.006 Datadive.net. (n.d.). Selecting good features – Part III: random forests. Retrieved January 30,

2018, from http://blog.datadive.net/selecting-good-features-part-iii-random-forests/ Dholakia, U. M., & Durham, E. (2010). How effective is Facebook marketing?. Harvard

business review, 88(3), 26.

De Vries, L., Gensler, S., & Leeflang, P. S. (2012). Popularity of brand posts on brand fan pages: An investigation of the effects of social media marketing. Journal of interactive marketing, 26(2), 83-91.

Fico, F. G., Lacy, S., & Riffe, D. (2008). A content analysis guide for media economics scholars. Journal of Media Economics, 21(2), 114-130.

https://doi-org.proxy.uba.uva.nl:2443/10.1080/08997760802069994

Grosman, L. (2017, November 07). Video Marketing: The New King Of Content. Retrieved January 30, 2018, from

https://www.forbes.com/sites/forbescommunicationscouncil/2017/11/07/video-marketing-the-new-king-of-content/#77e7c1937634

Golder, S. A., Wilkinson, D. M., & Huberman, B. A. (2007). Rhythms of social interaction: Messaging within a massive online network. In Communities and technologies 2007 (pp. 41-66). Springer, London. https://doi-org.proxy.uba.uva.nl:2443/10.1007/978-1-84628-905-7_3

Katz, E. (1959). Mass communications research and the study of popular culture: An editorial note on a possible future for this journal. Departmental Papers (ASC), 165.

Kim, J. W., Lee, B. H., Shaw, M. J., Chang, H. L., & Nelson, M. (2001). Application of decision-tree induction techniques to personalized advertisements on internet storefronts. International Journal of Electronic Commerce, 5(3), 45-62.

http://www.jstor.org.proxy.uba.uva.nl:2048/stable/27750981

Luarn, P., Lin, Y. F., & Chiu, Y. P. (2015). Influence of Facebook brand-page posts on online engagement. Online Information Review, 39(4), 505-519.

Mander, J. (2017, May 16). Social Media Usage Rises To 2 Hours Per Day | GlobalWebIndex. Retrieved January 30, 2018, from

https://blog.globalwebindex.net/chart-of-the-day/daily-time-spent-on-social-networks/ McAlexander, J., Schouten, J., & Koenig, H. (2002). Building Brand Community. Journal of

Marketing, 66(1), 38-54. http://www.jstor.org.proxy.uba.uva.nl:2048/stable/3203368

Muniz, Jr., A., & O’Guinn, T. (2001). Brand Community. Journal of Consumer Research, 27(4), 412-432. http://www.jstor.org.proxy.uba.uva.nl:2048/stable/10.1086/319618

Muntinga, D. G., Moorman, M., & Smit, E. G. (2011). Introducing COBRAs: Exploring motivations for brand-related social media use. International Journal of advertising,

30(1), 13-46. https://doi-org.proxy.uba.uva.nl:2443/10.2501/IJA-30-1-013-046

Moro, S., Rita, P., & Vala, B. (2016). Predicting social media performance metrics and evaluation of the impact on brand building: A data mining approach. Journal of Business Research, 69(9), 3341-3351.

https://doi-org.proxy.uba.uva.nl:2443/10.1016/j.jbusres.2016.02.010 Neuendorf, K. A. (2016). The content analysis guidebook. Sage.

Park, N., Kee, K. F., & Valenzuela, S. (2009). Being immersed in social networking environment: Facebook groups, uses and gratifications, and social outcomes.

CyberPsychology & Behavior, 12(6), 729-733. https://doi-org.proxy.uba.uva.nl:2443/10.1089/cpb.2009.0003

Rutz, O., & Bucklin, R. (2011). From Generic to Branded: A Model of Spillover in Paid Search Advertising. Journal of Marketing Research, 48(1), 87-102.

http://www.jstor.org.proxy.uba.uva.nl:2048/stable/25764566

Richards, K. (2016, March 22). Why So Many People Leave Advertising. Retrieved January 12, 2018, from http://www.adweek.com/brand-marketing/why-so-many-people-leave-advertising-170337/#/

Sabate, F., Berbegal-Mirabent, J., Cañabate, A., & Lebherz, P. R. (2014). Factors influencing popularity of branded content in Facebook fan pages. European Management Journal, 32(6), 1001-1011. https://doi-org.proxy.uba.uva.nl:2443/10.1016/j.emj.2014.05.001 Scikit-learn. (n.d.). 3.2.4.3.2. sklearn.ensemble.RandomForestRegressor. Retrieved January

30, 2018, from

http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html Sinclaire, J. K., & Vogus, C. E. (2011). Adoption of social networking sites: an exploratory

adaptive structuration perspective for global organizations. Information Technology

and Management, 12(4), 293-314.

Tafesse, W. (2015). Content strategies and audience response on Facebook brand pages.

Marketing Intelligence & Planning, 33(6), 927-943. https://doi-org.proxy.uba.uva.nl:2443/10.1108/MIP-07-2014-0135

Tafesse, W., & Wien, A. (2017). A framework for categorizing social media posts. Cogent

Business & Management, 4(1), 1284390.

https://doi.org/10.1080/23311975.2017.1284390

Trefzger, T., Baccarella, C., & Voigt, K. I. (2015). Antecedents of brand post popularity in Facebook: the influence of images, videos, and text.

Yamaguchi. (2017, January 12). How to Calculate Engagement Rate for Social Media Marketing. Retrieved January 30, 2018, from

https://www.origamilogic.com/blog/calculate-engagement-rate-social-media-marketing/

Zaglia, M. E. (2013). Brand communities embedded in social networks. Journal of business

research, 66(2), 216-223.

https://doi-org.proxy.uba.uva.nl:2443/10.1016/j.jbusres.2012.07.015

Appendix A. Codebook

This codebook is designed to help the coders in the process of coding post characteristics derived from 10 automotive Facebook brand pages. The coders should refer only these definitions and instructions.

Definitions

Unit of Analysis: Each individual component which is coded. Specifically, the unit is each mined Facebook post.

Post link: A unique web link that identifies the specific brand post. Click on the hyperlink to open the Facebook post on your browser.

Engagement rate: A variable consisting of the sum total between the likes, shares,

comments and reactions of a single Facebook branded post, divided by the number of page followers at the time of the publication, and the value turned into a percentage.

Descriptives

Weekday: Mark the posts that have been published during a weekend (all 4th, 5th, 11th, 12th, 18th, 19th, 25th & 26th of November, 2017) with a ‘0’ and the ones posted during rest of the days during November with ‘1’.

Working hours: Mark the posts that have been published between 9am and 6pm with ‘1’, while the rest with a ‘0’.

Format characteristics

Video: Posts that include a video attachment (often paired with a caption) should be marked with ‘1’, while the rest are coded with ‘0’.

Image: Posts that include an image attachment (often paired with a caption) should be marked with ‘1’, while the rest are coded with ‘0’.

Content characteristics

The purpose or intention of the branded post, which is typically communicated within the post caption. Whenever a post caption is absent, the coder must derive meaning from the visual content (in this case an image or a video). For this category of characteristics, it is possible for a Facebook post to have more than one theme.

 Emotional post: Brand posts that employ emotion-laden language to trigger affective responses should be marked with ‘1’. The theme includes language that expresses joy, humor, affection, sadness, adventure, excitement, nostalgia, pride, or inspiration. Commercials are also considered emotional content due to the perceived emotional effect those are meant to have on the consumer.

 Functional post: Brand posts that highlight the functional/beneficial attributes of products and services should be marked with ‘1’. The theme includes any practical information that communicates competitive advantage, like performance parameters and emissions metrics.

 Educational post: Posts aimed to educate followers about the brand itself and new skills when using it’s products or services should be marked with ‘1’. The theme incorporates any information teaches the customer more about the ca brand. This includes model releases, special edition product launches, the promotion of a campaign, or teasers.

 Community-oriented post: Posts that promote active participation among existing members and stimulate brand followers to take a certain action should be marked with ‘1’. Also, community-centered posts should address the customers directly, or

specifically communicate an event that community member can attend (such as a car show). Raffles and contests are also considered within the theme. So as

Appendix B. Coding sheet

1. Video  ‘1’ for Yes  ‘0’ for No 2. Image  ‘1’ for Yes  ‘0’ for No 3. Weekday  ‘1’ for Yes  ‘0’ for No 4. Working hours  ‘1’ for Yes  ‘0’ for No 5. Emotional post  ‘1’ for Yes  ‘0’ for No 6. Functional post  ‘1’ for Yes  ‘0’ for No 7. Educational post  ‘1’ for Yes  ‘0’ for No 8. Community-centered post  ‘1’ for Yes

 ‘0’ for No