Do Firms Deal with Customer Generated Branded Content on Twitter in the right way?

(1)

Do Firms Deal with Customer Generated

Branded Content on Twitter in the right way?

A case study about to what extent the way in which firms respond to customer generated branded content (CGBC) affects customer engagement (CE) and customer engagement

(2)

1

Do Firms Deal with Customer Generated

Branded Content on Twitter in the right way?

A case study about to what extent the way in which firms respond to customer generated branded content (CGBC) affects customer engagement (CE) and customer engagement

sentiment (CES).

By

Jasmijn Staal

Master Thesis Marketing

S2324660

Nassaulaan 29a

9717 CG Groningen

06 – 48 79 31 52

j.j.h.staal@student.rug.nl

University of Groningen

Faculty of Economics & Business

Department of Marketing

PO Box 800, 9700 AV Groningen

First supervisor: Dr. J.C. Hoekstra

(3)

2

ABSTRACT

Nowadays, it is increasingly important for firms to deal with user generated content (UGC) in the right way. Inadequate response to UGC can be fatal, because it drives product awareness and influences consumers’ purchase behavior. This explorative study is part of a larger study and aims to provide first insights into how a firm’s response towards customer generated branded content (CGBC) influences customer engagement (CE) and customer sentiment (CES). CGBC provides valuable insights for marketers in managing unstructured big data; however, this is found to be extremely challenging. Previous research mainly focuses on the study of structured data (e.g. online ratings) and the importance of a firm’s response towards customers. However, which firm response characteristic works best in terms of CE and CES when responding to CGBC remains untapped. In this study, a conceptual model for exploring these relationships is set up based on current literature. This model may serve as a basis for a code book to be used in content analysis. This code book was used to code a total of 951 tweets that were collected from four large Dutch companies on Twitter. Results show in the first place that firm response characteristics change the focus or the characteristics in a customer tweet that are important in predicting CE. Furthermore, the way a firm responds affects CES by creating less negative and intense tweets. These results could provide managers with tools for choosing which characteristics to use when responding to CGBC, and increase insight into the separate effects of CGBC, firm response characteristics and subsequent customer response on CE and CES. Future researchers can use the recommendations and could test the relationships on a larger scale.

(4)

3

ACKNOWLEDGEMENT

This thesis represents the final part of the Master Marketing Management. My goal for this master thesis process was to learn a lot and get the most out of this interesting thesis topic, but more importantly; enjoy it! So that is what I did and I think I have never been so dedicated to a specific topic for such a long time. I even enjoyed diving into the analysis parts of my thesis, where, upfront, I was most afraid of. Of course I experienced periods where I did not know how to deal with the overload of information and how to analyze it. Nevertheless, I think reached my goal: I enjoyed the thesis process and learned a lot in this short period of time.

I would like to thank the people who supported me during this thesis process. First of all, I would like to express my gratitude to Janny Hoekstra and Bianca Harms for providing me with constructive and detailed feedback, being flexible and being enthusiastic about this study. Thanks to both of you, I was able to enjoy every part of this thesis process, push my limits in doing research and overall I was able to learn a lot. Moreover, I would thank my parents for supporting me during my whole ‘study-career’. They show respect for every choice I make and really motivate me to make the most out of it, which I am very grateful for. Furthermore, I would like to thank Lotte de Boer for listening to my frustrations, having breaks (incl. too much coffee) in a very strict schedule, sending me the funniest memes, going to the gym to relax and all the giggling in the university library when we were supposed to do thesis stuff. I also like to thank Koen Schuurman for calling me when I had questions about the methodology and statistical analysis. Further, I like to thank my roommates (the pinkpanters) who cooked me dinner when I came home late and cheered me up when I was stressed about deadlines. I also like to thank Marin Kelava for putting tons of critical, but very legit comments and suggestions in my thesis, and showing me that late night sessions in the university library can be fun. Last but not least, I like to thanks Rosalie Hofwegen and Willie Offringa for helping me with checking my thesis on grammar.

(5)

4

1. INTRODUCTION

In 2008, Dave Carroll created the song ‘United Breaks Guitars’ (#UnitedBreaksGuitars) about how United Airlines broke his $3.500 Guitar and the reaction he received: the claim was rejected because it was submitted after the standard 24-hour timeframe. This video was a hit on YouTube with over a 5 million views in two days. Within four weeks media speculated about a decline in stock prices by 10% for United Airlines, costing the stockholders about $180 million (Ayres, 2009). Furthermore, brand damage was undeniable (McCarthy, 2009). Nine years later the hashtag is still used on Twitter and expanded by the hashtag #UnitedBreaksPassengers, because of a new social media crisis. April 2017, a passenger of United Airlines was violently and forcibly ejected from flight 3411 and another passenger taped the situation (Victor and Stevens, 2017). Again, this content went viral; the way United Airlines reacted in the first place was largely criticized (McCann, 2017). Such a social media crisis is archived on the Internet and remains visible for a long time. So user-generated content (UGC) can really impact the firm, and these examples show the importance of an appropriate reaction by the firm to prevent or reduce brand damage. Inadequate response to UGC can be fatal because UGC shapes perceptions about what large masses of consumers think about the brand, thus driving product awareness and influencing purchase behavior (Gensler, Völckner, Liu-Thompkins and Wiertz, 2013). Hence, it is important for firms to deal with UGC in the right way.

UGC refers to ‘situations whereby customers freely choose to create and share information of

value’ (Campbell, Pitt, Parent and Berthon, 2011:87) and is found to play a more influential

role in driving purchases than marketer generated content (Goh, Heng and Lin, 2013). Nowadays, customers actively take part in the marketing communications process, which results in firms losing control in creating and communicating their brand strategies (Fernando, Suganthi and Sivakumaran, 2014). A substantial amount of UGC is brand-related and can be used to form customers’ brand perceptions (Smith, Fischer and Yongjian, 2012). When looking at the example of #UnitedBreaksGuitars, customer brand perceptions can be easily influenced by customers creating UGC, and thereby have a large impact on the firm. Therefore, this study focuses on customer generated branded content (CGBC), which is defined as ‘‘customers’

generating content about a brand, regardless of positive or negative connotation or intent’

(Christodoulides, Jevons and Bonhomme, 2012: 2) that ‘arises predominantly due to the

internal motivational state of customers, irrespective of explicit intended f actions’ (e.g.

(8)

7 words, CGBC is about brand-related content that is created by the customer independently from the firm. Current research studies focus on structured data (e.g. online ratings by providing mean and variance), while Kannan and Lee (2017, p. 28) show the need and importance of researching the content of posts (unstructured data) because it contains valuable and direct information expressed by consumers. Hence, this research dives into the unstructured data of CGBC.

CGBC becomes more visible when other customers like, share (retweet) and/or comment on this CGBC, and is called customer engagement. Customer engagement (CE) entails ‘the

mechanics of a customer’s value addition to the firm through indirect contribution’ (Pansari

and Kumar, 2017, p. 295), where indirect customer value addition arises from customer likes, shares and/or comments. Dessart, Veloutsou and Morgan (2015) states that creating engaged customers is becoming one of the key objectives of many marketing professionals, where the positive implications of customer engagement are driving academic and practical interest (Vivek, Beatty, Dalela and Morgan, 2014). Moreover, CE has been viewed as a new key metric for gaging firm performance (Hollebeek, Glynn and Brodie, 2014; Kumar and Panasari, 2016). Research has been done on the financial consequences of customer engagement behavior (CEB) resulting from firm-initiated CE. However, the effect of customer-initiated CE is assumed to be different and remain untapped (Beckers et al. 2017). Hence, looking at the effect of CGBC on CE is aimed to expand current literature.

(9)

8 and returns) (e.g. Baer & Hill, 1994; Fan and Niu, 2016; Liu, Burns and Hou, 2017; Sparks et al., 2016; Rauschnabel et al., 2016; Van Laer and De Ruyter, 2010), research in this field is rather limited (Gensler et al. 2013; Kannan and Li, 2017). Hence, combined with CE as a new key metric for gaging brand performance, this study aims to explore the relationship between the way in which a firm responds towards CGBC influences CE.

Customers involved in UGC are more likely to be brand advocates, share information and opinions (Shivinski and Dabrowski, 2016). Kannen and Li (2017) state in their research agenda that sentiment analysis can be used as indicator for how well actions of the firm are viewed by the customer. Hence, this research also incorporates customer engagement sentiment (CES) that refers to the general message valence and total message intensity represented in the

customer tweet (Makarem and Jae, 2016; Mishra and Satish, 2016; Villaroel Ordenes,

Ludgwig, Ruyter, Grewal and Wetzels, 2017). By incorporating CES, this study aims to explore the opinions of customers about the way in which firms respond.

Building on existing literature in the fields of SSCT, service recovery theory and digital marketing, the growing interest in research about the effects of user-generated social media communication and the limited research on firm response to UGC (Goh et al., 2013; Schivinski and Dabrowski, 2016), this exploratory study aims to describe the relationship between CGBC, the way in which a firm responds, and CE and CES, thereby extending existing literature. UGC may provide helpful insights for marketers to manage and monitor this extremely challenging unstructured big data (Kannan and Li, 2017; Liu et al., 2017). Moreover, implications on how social media can be used to enhance customer services can be derived from this study (Kannan and Li, 2017). In sum, this research aims to uncover to what extent the way in which firms

respond to customer generated branded content (CGBC) affects customer engagement (CE) and customer engagement sentiment (CES).

To explore this relationship the following sub-questions are formulated:

1. How do the characteristics of CGBC affect CE?

2. How do the characteristics of a firm’s response affect CE?

3. How do the characteristics of a customer response towards the firm affect CE?

4. How do firm’s response characteristics change the characteristics in customer tweets?

(10)

9 This explorative study uses existing literature as a basis for generating a fairly complete code book to collect as much relevant data as possible that will provide possible insight into the relationship between the firm’s response characteristics towards CGBC and their effect on CE and CES. The focus of this study is on the platform Twitter, since it is, for the top fortune 100 firms, found to be (with 87% usage) more popular than Facebook. Moreover, Twitter is an established and effective customer communication platform that can be used for listening and handling CGBC (Liu et al., 2017). Based on Veldwijk (2018), the availability of GCBC and firms responding to it, four large Dutch companies (KLM, KPN, NS and Bol.com) were selected to collect data from. To analyze the unstructured data that CGBC contains and account for more complex characteristics conveyed in tweets, such as humor and sarcasm, this research makes use of quantitative content analysis (QCA) using human coders instead of automated text analysis.

(11)

10

2. LITERATURE REVIEW

In the review that follows, section 2.1 provides a visual representation of the relevant characteristics per model component and how these characteristics relate to each other. Section 2.2 discusses the types of customer engagement and customer engagement sentiment and in section 2.3 the characteristics of CGBC are discussed. Section 2.4 discusses the firm’s response characteristics.

2.1 Conceptual model

The aim of this chapter is not to generate hypotheses, but to create an overview of relevant variables per model component to serve as basis for generating a fairly complete code book. Characteristics of CGBC are assumed to influence the CE types (replies, mentions, retweets and likes). A number of characteristics of firm response will be discussed: content format,

denial, apology and/or compensation, direct message, customer name, understanding, compliment, voice of response, signature, related humor, firm liking, firm retweeting and speed of response. They are assumed to be influenced by the characteristics of CGBC (content format, general message valence, total message intensity and hashtags). They are proposed to

influence subsequent reactions of customers (CES: general message valence and total message

intensity). Furthermore, we expect that the firm’s response mediates the relationship between

CGBC characteristics and CE types. Figure 1 visually represents the proposed relationships. In the following sections (2.2, 2.3 and 2.4) the relevant variables per model component are discussed.

(12)

11 2.2 Customer engagement and customer engagement sentiment

In this study, customer engagement (CE) refers to ‘‘the customers’ value addition to the firm

through indirect contribution’ (Pansari and Kumar, 2017, p. 295), where indirect customer

value addition occurs in the form of engagement in terms of replies, mentions, retweets and likes and expressed in counts (Cho, Schweichart & Haase, 2014; Peters, Chan, Kaplan,

Ognibeni, and Pauwels, 2013; Wonderwesen Fafesse, 2016).The ripple effect of customer

brand discussion on social media is found to indirectly impact firm performance and is therefore important to take into account (Hogan, Lemon and Libai, 2003; Kumar, 2013). Moreover, CGBC contains valuable information about the brand and the firm’s performance, and has significant potential business value in CE, which has been viewed as a new key metric for gaging brand performance (Hollebeek, Glynn and Brodie, 2014; Liu et al., 2017; Kumar and Panasari, 2016; Tirunillia and Tellis, 2012). Hence, CE is included in this study as key metric.

Customer engagement sentiment (CES) refers to the general message valence and total

message intensity represented in the customer tweet (Makarem and Jae, 2016; Mishra and

Satish, 2016; Villaroel Ordenes, Ludgwig, Ruyter, Grewal and Wetzels, 2017). The goal of sentiment analysis is to identify customer opinions (positive, negative or neutral) and to detect message intensity (strong vs. weak) in a message (Pang and Lee, 2008). Liu et al. (2017, p. 241) assume that each tweet on Twitter explicitly expresses the writer’s opinion on aspects of the firm, such as brands, products and services. To uncover these opinions, they incorporate valence in their study. Moreover, Makarem and Jae (2016) incorporate both valence and intensity to uncover motivations in boycott behavior, which results from actions that consumers take based on their opinions about a firm. Valence and intensity often appear jointly and are important in understanding heterogeneous emotional experiences (Makarem and Jae, 2016; Presi et al., 2014). Hence, both valence and intensity are incorporated in CES to uncover these opinions and heterogeneous emotional experiences.

(13)

12 2.3 Customer generated branded content

In the context of this study, customer generated branded content (CGBC) is about brand-related content that is created by the customer independently from the firm and contains the following characteristics: content format, general message valence, total message intensity and hashtags which will be elaborated in section 2.3.1, 2.3.2, 2.3.3 and 2.3.41_{(Mishra and Satish, 2016;}

Villaroel Ordenes, Ludgwig, Ruyter, Grewal and Wetzels, 2017; Meijer and Homburg, 2015).

2.3.1 Content formats

On Twitter, the following content formats can be used: text, URLs, photos, animations, videos and polls. Kane, Alavi, Labianca and Borgatti (2014) found that different content formats (text, multimedia and hypermedia) affect how content flows across social media platforms. Findings from Chang, Yu and Lu (2015) in social media marketing (based on ELM factors) show that post popularity plays an important role in persuasion for users in both central and peripheral route processing (high and low elaboration). Since post information contains both responses to the content and popularity numbers, it influences central and peripheral processing of customers. Content formats, such as animations, may enhance post popularity which is found to directly influences behavior intentions (Chang et al., 2015). Moreover, Van De Velde et al. (2015) found that the use of a URL significantly relates to retweet probabilities, which enhances post popularity. Abitbol and Lee (2017) show that posts on facebook including photos, graphics and/or URLs generate higher CE in terms of number of likes, shares and comments in comparison to text-only messages. Hence, the same may be the case for Twitter, so photos, graphics and URLs may increase CE, and text may decrease CE. In sum, as in Abitbol and Lee (2017), different content formats are proposed to have different effects on CE.

2.3.2 General message valence

General message valence refers in this study to whether the overall message is positive,

negative or neutral and is found to generate different effects on i.e. e-WOM or engagement

(Liu et al., 2017; Tirunillai & Tellis, 2011; Van Doorn et al., 2010). Sentiment on Twitter often occurs in information (neutral), complaints (negative) and opinions (potentially positive) (Smith et al., 2012). Chang et al. (2015) and Batra and Keller (2016) suggest that what a message conveys (e.g. valence of e-WOM) is more important than the volume (e.g. number comments/replies) of e-WOM conversations. Valence is a popular measure in brand sentiment

(14)

13 analysis and is used to measure success of social media initiatives (Smith et al., 2012). Sentiment analysis of Liu et al. (2017) based on1.7 million tweets from 20 brands across five industries, shows that brand-related tweets generally express customer emotions, where negative (47,8%) sentiment is almost three times higher than positive sentiment (16,9%). This shows that unhappy customers are more likely to tweet about their bad experiences than happy customers (Liu et al., 2017). Critical CGBC (negative valence) can really hurt the brand, because it is found to have larger effects on returns than positive CGBC (Tirunillia and Tellis, 2012). However, CGBC can still benefit the firm in terms of positive e-WOM (Gregoire et al., 2015; Poch and Martin, 2015). Based on findings of these studies, it could be possible that negative general message valence could influence CE more than positive or neutral messages.

2.3.3 Total message intensity

For this research, message intensity is defined as ‘the degree of positivity or negativity, or

alternatively the neutrality of a message’ (Floh, Koller and Zauner, 2013, p. 650). In other

(15)

14 2.3.4 Hashtags

Hashtags are ‘keywords prefixed with # symbol in tweets’ (Ma, Sun and Cong, 2013, p. 1399). Hashtags, also called social tags, are created to efficiently categorize the enormous amount online content that users generate, but also to facilitate users in the content discovery process (Nam and Kannan, 2014). Users classify tweets around specific themes by using hashtags (Marres & Weltevrede, 2013; Page, 2012). Hashtags are a powerful resource for promoting visibility among a Twitter update (Page, 2012). Because tweets are classified around a specific theme and directly directed towards the firm (which is the case for replies), it costs extra time and expenses for the firm to identify and track the messages in hashtags. However, hashtags offer a great opportunity to positively surprise customers and therefore may be worth spending extra money and time tracking hashtags (Abney et al., 2017).

Hashtags develop associative structures around brand-related content and thereby become a valuable source in deriving thoughts and perceptions about the brand, product and firm (Nam and Kannan, 2014). As in the example of #UnitedBreaksGuitars and #UnitedBreaksPassengers, customers spreading messages using these hashtags generate negative associations and perceptions with the brand and therefore negatively affect the brand image. Moreover, Van De Velde et al. (2015) found that the use of hashtags significantly relates to retweet probability. So, hashtags can be used to obtain volume (e.g. number of likes) and valence (positive/negative/neutral) of tags associated with the brand. Therefore, hashtags are proposed to have an effect on CE and CES.

2.4 Firm response characteristics

(16)

15 2.4.1 Importance of the way a firm responds

Effectively responding to social media complaints can benefit the firm in terms of both complaining customers and other social media users (Abney et al., 2017). This is in line with service recovery literature, where service recovery refers to ‘all actions that an organization

may take to rectify a failure’ (Presi et al., 2014, p. 1606). When recovery is successful the

customer’s post-failure satisfaction may exceed pre-failure satisfaction, also called the service recovery paradox (SRP) (Matos et al., 2007; Presi et al., 2014). Building on social crisis communication theory (SCCT) and the framework for collaborative brand attacks (CBAs), which can be seen as collective negative e-WOM, increased traffic on online and mobile platforms enhances CBAs/online firestorms, which may result in large quantities of negative GCBC (e.g. due to a bad service experience) (Rauschnabel et al., 2016). Hence, recovery is important, because negative e-WOM is found to have a larger impact on returns than positive e-WOM (Tirunillia and Tellis, 2012). Firm response characteristics may work as the SRP and are found to be effective in protecting the firm’s reputation, but only under the condition of matching the right firm response characteristic with the right situation (Coombs and Holladay, 1996; Coombs, 2006). Hence, based on current literature, several firm response characteristics are formulated and discussed in section 2.4.2, 2.4.3 and 2.4.4.

2.4.2 Firm response: What

This section discussed the firm response characteristics that represents what content is used in firm response. The following characteristics are incorporated in this section and will be further elaborated on: content format, denial, apology and/or compensation, direct message, customer

name, understanding and compliment. Content format

As for CGBC, as described in section 2.2.3, various content formats in the firm response are proposed to have different effects on engagement (Kane et al., 2014; Abitbol and Lee, 2017). Hence, this study argues the same effect for the firm’s response on CE.

Denial

(17)

16 require a fight (e.g. rumors). Therefore the effect of denial may be positive or negative on CE and CES, depending on the situation.

Apology and/or compensation

Wirtz and Mattila (2003) found that recovery outcomes (e.g. compensation), procedure (e.g. speed of recovery) and interactional treatment (e.g. apology) have a joint effect on post-recovery satisfaction. Compensation is found to be a poor substitute for a bad post-recovery process (delay in response without apology), while on the other hand compensation was effective in mixed recovery situations (delay in recovery with apology or immediate response without apology) (Wirtz and Mattila, 2003). Jung and Seock (2017) accounted for the separate effects of apology and compensation, but also for the joint effect of combining apology with compensation. They found that providing the customer with an apology, with or without a tangible compensation, improves customer satisfaction and thereby provides an increase of positive WOM intentions. In a content analysis about online reactions to an apology, Coombs and Holladay (2011) found that most posts indicated acceptance and show positive purchase intentions, hereby confirming the effectiveness of the SCCT. Therefore, both compensation and apology are proposed to affect CE and CES.

Direct message

Direct messages (DMs) are used for keeping a conversation private and confidential (Gunarathne, Rui, Seidmann, 2017). Well-known companies use DMs to generate a form of private complaining (Gregoire et al., 2015). Asking customers to send a DM instead of replying in public prevents other users following the conversation and thereby the firm creates a private conversation. Therefore it can be argued that, when a firm asks a customer to send a DM and the customers agrees by sending the firm a DM, the public conversation ends. Logically, CE will be lower than for a conversation without the firm asking the customer to send a DM.

Customer name

(18)

17 predicting overall service quality and customer satisfaction, which they think is surprising. In sum, contradicting findings for using the customer name arise so it could be possible that this has an effect on CE and CES either negatively or positively.

Understanding

According SCCT, understanding is used by the firm to express concern to the customer/victim (Coombs, 2006). Magnini, Ford and Markowski (2007) propose their study about service recovery in offline context (tourism) that using understanding/empathy may create paradoxical satisfaction after a service failure. The study of Wirtz and Mattila (2004) suggests that understanding/empathy may increase customer perception of fairness (Wirtz and Mattila, 2004). However, their study only incorporates apology and compensation and leaves understanding as future research topic. Hence, understanding is included in this study and is proposed to have an effect on CE and CES.

Compliment

Coombs (2006, p. 248) defines compliment as crisis response, where the crisis manager praises stakeholders and reminds them of good things in the past. Hence, the firm thanking or praising the customer for their help or attentiveness can be described as giving the customer a compliment. Dai, Shin, Kashian, Jang and Walther (2016) found that people like to be appreciated by people they know, but also by people they do not know. They also found that compliments are effective at increasing liking, because people who receive compliments are found to feel obligated to like the complimenter (Dai et al., 2016). From this it could be argued that giving the customer a compliment will increase liking, which may possibly result in engagement.

2.4.3 Firm response: How

This section discusses the firm response characteristics that represents how content is conveyed in firm response. The following characteristics are included in this section: voice of response,

signature use, related humor and like and retweet customer tweets. Voice of response and signature use

(19)

18 Weinberger, 2000). Sparks et al. (2016) found that different voices of response have different effects on customer concern inferences. When using a human voice, response customer concern inferences scores were more favorable compared to a professional voice of response. Therefore this research incorporates human and professional as different voices of response, where a human voice is proposed to affect CE and CES positively.

To make conversations more personal, some companies (e.g. NS) may use a signature of the customer service agency in the tweet. Back in the days Goldsmith (1999) already suggested that personalization should be the basis in relationship marketing. Keller (2013) confirms this and shows the must for marketers to create personalized experiences for customers that creates awareness and cultivate loyalty. Therefore the use of a signature and possible effect on CE and CES is also taken into account when responding to the customer.

Related humor

(20)

19

Like and retweet customer tweets

Spreading good words and positive publicity about how well a firm resolves a complaint is called boasting (Gregoire et al., 2015). Crisis management and service recovery literature are focused on solving complaints to prevent bad e-WOM, while firms also receive compliments from customers. The easiest way to share compliments is by retweeting them. By liking a tweet, the tweet appears on the firm’s profile page in the list of likes and the creator of the tweet is notified. In that way a firm shows the user that they appreciate their tweet and spreads positive e-WOM, which could have an effect on CE.

2.4.4 Firms response: When

(21)

20

3. METHODOLOGY

Section 3.1 describes the method for data collection, the platform on which the data will be collected from and how the data is collected. In section 3.2 the operationalization of the variables is discussed, followed by intercoder reliability in section 3.3, data cleaning in section 3.4 and statistical validity in section 3.5. Lastly, in section 3.6 method of analysis is presented.

3.1 Method for data collection

Content analysis has proven to be useful in gaining a better understanding of print ads, television commercials, product placement, outdoor advertising and websites (Ashley & Tuten, 2015). This study uses quantitative content analysis (QCA) with human coders to discover relationships between CGBC, firm response and CE (Abitbol & Lee, 2017). QCA is ‘a research

technique for the systematic, objective, and quantitative description of the manifest content of communication’ (Berelson, 1952, p.18). QCA is the most appropriate technique for this

research, because it allows researchers to use messages that the customer (CGBC) and the firm (firm response) have actually shared and the actual engagement (CE) on these messages (Abitbol & Lee, 2017). While other studies containing automated text analysis make use of dictionary-based analysis such as Linguistic Inquiry and Word Count (LIWC) and other advanced packages such as R and Python, this is not sufficient for this research. Compared to automated text analysis, human coders provide a richer and deeper understanding of the context where life was lived in (Humphrey and Wang, 2017; Kassarjian, 1977). Hence, human coders were used to uncover finer shades of meaning such as humor and sarcasm since this is not possible with automated text analysis (Humphrey and Wang, 2017).

This research contains a case study, to explore relationships on a smaller scale to pre-test feasibility, improve quality and efficiency, provide valuable insights and make suggestions for a larger scale study (Nadler, 1967; Van Teijlingen and Hundley, 2001). Moreover, Sekaran and Bougie (2016) state that a case study is the documented history of noteworthy events that have taken place in a given institution, where tweets can be seen as documented history of noteworthy events and the selected firms as the given institution.

Twitter

(22)

21 with followers on the Twitter chronological ‘public timeline’, the users’ individual Twitter page (‘micro blog’) or in private (only followers from the users feed are able to see the messages) (Honeycutt and Herring, 2009; Williams, Inversini, Buhalis and Ferdinand, 2015). Twitter is the most popular microblogging platform and is known for its variety of users that express their opinions about all kinds of topics (Pak and Paroubek, 2010). Hence, Pak and Paroubek (2010) argue that Twitter serves as a valuable source for real-time opinions of people. Presi et al. (2014) found that Twitter (compared to Facebook) contains the most emotionally motivated CGBC (i.e. tweets that contain vengeance), highlighting the importance of a good service recovery for this platform to prevent negative e-WOM. Twitter is an established and effective customer communication platform that can be used for listening and handling CGBC (Liu et al., 2017).

A comparison of Twitter and Facebook showed that there was no significant difference in the proportion of subscribed users who shared content (Chiu, 2013). Chiu (2013), which suggests that quality of the content is the main factor for sharing content and not the platform. Further, among the Fortune 100 companies, Twitter is (with 87% usage) more popular than Facebook (Narayanan, Asur, Nair, Rao, Kaushik, Mehta, Athalye, Malhotra, Almeida and Lalwani, 2012). Therefore, this research makes use of the platform Twitter. An overview of the relevant conversation aspects of Twitter is provided in table 1.

Table 1:

Twitter conversation (Williams et al., 2015)

Twitter conversation Description

CGBC The first tweet in the conversation created by the customer directed towards the firm. Replies ‘Replies’ start with @username in tweets and are replies to earlier generated tweets.

Mentions ‘Mentions’ contain @account in the message, but not at the beginning of the message as in a ‘reply’

Retweet A ‘Retweet’ is sharing another user’s tweet to the accounts that follow your account Hashtags (#) Hashtags (#) organize content on Twitter. Users can follow a # or monitor a # to see what is

shared under the specific # (even if they don’t follow the user that generates the tweet) (e.g. #UnitedBreaksGuitars)

Likes Users can ‘like’ tweets when they like/agree with the tweet Direct message (DM) ‘Direct message’ (DM) are private messages sent to individuals

Data collection

Sample

(23)

22 Veldwijk (2018). Furthermore, the sample of companies was, where possible, matched with the sample of Veldwijk (2018). The sample of this case study includes four large Dutch companies from different industries. An overview of the companies and their Twitter accounts is presented in table 2.

Table 2:

Companies in the case study

KLM KPN NS Bol.com

Twitter account @KLM @KPNWebcare @NS_online @Bol_com Joined Twitter on July 2009 August 2009 March 2010 January 2009 Number of followers of the firm 2.352.000 68.800 210.000 56.900 Number of followings by the firm 67.200 27.400 10.800 18.900 Total number of tweets by the firm

947.000 662.000 1.187.000 170.000 Time response

promise 60 minutes + live updated banner on Twitter

30 minutes + live updated banner on Twitter

Within 15

minutes Within 60 minutes between 08:00-23:00

Note: numbers collected on November 11 2017 (rounded weights)

Tweet selection

Per firm 60 CGBC tweets (referred to as type 1, see table 3) were coded and labeled as a case, followed by the whole conversation between the customer and the firm on Twitter. These subsequent tweets were labeled as the firm response (type 2) and the customer response to the firm (type 3) (see table 3).

Table 3:

Coding type of tweet

Type 1 CGBC. The first tweet that the customer sends to the firm

Type 2 Firm response to the customer. The response of the firm towards type 1 and type 3 tweets. Type 3 The customer response towards the firm. The response of the customer towards a type 2 tweet.

(24)

23 Second, to answer the research and sub-questions, only CGBC containing a firm response can provide data for exploring this relationship. For CGBC without a firm’s response it is impossible to measure the firm’s response’s characteristics and the effect on CE.

Furthermore Type 1 tweets were selected so that they provide a representative image with as little as possible external influence, such as bank holidays, day of the week, weather, product fails, etc. Thus, the cases were selected based on; first, language, only Dutch type 1 tweets were selected. Second, as in i.e. research from Pak and Paroubek (2010) an even distribution of positive, negative and neutral (valence) type 1 tweets per firm is handled. Nowadays, in estimating effects, is agreed upon at least 20 observations (type 1 tweets). Therefore 20 type 1 tweets for each valance type are selected to provide variance in the tweets. Third, distribution of tweets in time: the type 1 tweets were selected on different time of the day, days, weeks and months. Lastly, all tweets were collected before November 8, 2017, so that only tweets with 140 characters were part of the sample.

Tweet collection

The tweets were coded in a separate Excel file for each firm. After finishing the coding, the Excel files were merged into one SPSS file. The structure of the data set, including variable naming, can be found in appendix C.

3.2 Operationalization of the variables

Similar to Abitbol and Lee’s work (2017), a code book was developed in three steps to operationalize the relevant variables. First, based on literature as described in chapter 2 the initial code book was developed. Second, this initial code book was tested by pre-coding 10

cases (followed by the whole conversation) to see if other relevant variables were missing.

(25)

24 For the customer tweets the firm variables were coded as missing (99) and the other way around. An overview of the different variables and their coding can be found in table 4.

Table 4:

Overview of the variables and their coding

Variables Description Coding

Customer engagement

Replies How many replies are there on the tweet? Count the number of replies

Mentions How many mentions are there on the tweet? Count the number of mentions

Retweets How many retweets are there on the tweet? Count the number of retweets

Likes How many likes are there on the tweet? Count the number of likes

Content format

Text Does the tweet contain text? Present (1) or absent (0)

Photo Does the tweet contain a photo? Present (1) or absent (0)

Video Does the tweet contain a video? Present (1) or absent (0)

URL Does the tweet contain an URL? Present (1) or absent (0)

Animation Does the tweet contain text an animation? Present (1) or absent (0)

Poll Does the tweet contain a poll? Present (1) or absent (0)

Type 1 and 3 tweets

General message valence Is the customer tweet overall positive, negative or neutral in

valence? Positive (1), negative (-1) or neutral (0)

Total message intensity What is the total message intensity of the customer tweet? Based on:

General message intensity Length of the tweet Expressive punctuation Capital letter words

General intensity of emoticons Profanity & insult

Sarcasm

Sum of all intensity points. The formative scale on which intensity is measured can be found in appendix A.

Intensity of Emoticons How many high/moderate/low intensity emoticons were used

in the customer tweet? Count the number of high/moderate/low intensity emoticons in a tweet

Hashtags How many hashtags were used in a tweet? And how many positive/negative/neutral hashtags were used in the customer tweet?

Count the total number of hashtags and count the number of

positive/negative/neutral hashtags in the tweet

Type 2 tweets

Response time of the firm How many minutes did it take the firm to respond to the

customer tweet?

(26)

25

Apology Has the firm offered an apology in the response tweet? Present (1) or absent (0)

Compensation Has the form offered a compensation in the response tweet? Present (1) or absent (0)

Related humor Has the firm used related humor in the response tweet? Present (1) or absent (0)

Voice of response What voice of response has the firm used in the response

tweet? Professional (1) or Human (2)

Direct message (DM) Has the firm asked the customer to send a DM in the response

tweet? Present (1) or absent (0)

Firm liking Has the firm liked the customer tweet? Present (1) or absent (0)

Firm retweeting Has the firm retweeted the customer tweet? Present (1) or absent (0)

Signature use Has the firm used a signature in the response tweet? Present (1) or absent (0)

Denial Has the firm used denial in the response tweet? Present (1) or absent (0)

Customer name Has the firm used the customer name in the response tweet? Present (1) or absent (0)

Compliment Has the firm used a compliment in the response tweet? Present (1) or absent (0)

Understanding Has the firm used understanding in the response tweet? Present (1) or absent (0)

Control variables

Followers How many followers has the Twitter account (customer or

firm) that created the tweet? Count the number of followers of the Twitter account

Followings How many followings has the Twitter account (customer or

firm) that created the tweet? Count the number of followings of the Twitter account

Total number of tweets How many tweets has the Twitter account (customer or firm)

created in its Twitter account lifetime? Count the total number of tweets created by the Twitter account

Source of reply Is someone else then the customer or firm inferring the

conversation? Someone else is inferring (1) or not (0)

Competition Does the tweet contains information about the competitor

and/or the competitor is mentioned in the tweet? A competitor is mentioned/competitor information in the tweet yes (1) or no (0)

Type of tweet What is the type of tweet in the conversation? CGBC (1; type 1), the firm response to the customer (2; type 2) or the customer response towards the firm (3; type 3)

Tweet exchange What is the length (number of tweets in conversation) of the Twitter conversation?

Count the total number of tweets in the conversation and assign every tweet in that conversation this number (e.g. when the conversation contained 6 tweets, every tweet in that conversation received an 6 on tweet exchange)

Order What is the order of this specific tweet in the conversation? If the tweet is the second tweet in the conversation this tweet receives a 2

(27)

26

Firm To which firm is the tweet related? KLM (1), KPN (2), NS (3) or Bol.com

(4)

Response time promise What is the number of minutes that the firm handles as response promise towards a customer tweet?

Number of minutes

Time At what time was the tweet sent? The time in mm:ss that the tweet was

sent (e.g. 15:04)

Date At what date was the tweet sent? The date in dd-mm-yyyy that the tweet

was sent

3.3 Intercoder reliability

To ensure coder validity and reliability for the actual coding, next to the author, a second coder familiar with Twitter was used to pre-code 10 cases. The second coder sat with the author to walk through the aim of the research, the conceptual model, the variables, the thoughts behind the code book and the coding worksheet. Both coders received the code book containing the specified variables and a coding worksheet. Hema, another Dutch firm that also uses Twitter as customer service channel, was selected to practice the coding worksheet. Both coders filled out the worksheet and, where needed, differences were addressed via discussion. Hereafter, both coders coded 10 cases (approximately 40 tweets) for an intercoder reliability check. Krippendorff’s alpha (α) is often used in content analysis, because it ‘defines a large family of

reliability coefficients and embraces several known onces’ (i.e. scott’s pi) (Hayes and

Krippendorff, 2007, p.82). Alpha checks agreement between coders, is applicable on data containing missings and is suitable for different measurement scales, which is necessary for this research (Hayes and Krippendorff, 2007; Krippendorff, 1980). Only variables were disagreement in coding is suspected were used in conducting the Krippendorff’s α that should be above 0,80 (Krippendorff, 1980). All type 1 and 3 characteristics and all type 2 characteristics except ‘response time of the firm’, ‘firm liking’and ‘firm retweeting’ were included (see table 4). The DVs and control variables were not included. For the variables that received an α > 0,80, also the Cohen’s Kappa (κ) is computed to see if it still meets the criteria of the κ (Cohen’s κ range: 0,550–0,865) (Altman, 1999; Landis and Koch, 1977; Vilnai-Yavetz and Tifferet, 2013).

Applying the SPSS syntax KALPHA macro of Hayes and Krippendorff (2007) show mostly satisfactory intercoder reliability results (α > 0,80). There was no coder disagreement (α = 1,0) about the following variables: ‘apology’, ‘related humor’, ‘voice of response’, ‘signature use’,

‘direct message’, ‘denial’, ‘customer name’, ‘compliment’, ‘intensity of emoticons’ and

(28)

27 was present for ‘general message valence’ (α = 0,94, 95%CI 0,80 to 1,0) and ‘total message

intensity’ (α = 0,99, 95%CI, 0,98 to 0,10). Only ‘understanding’ received a higher

disagreement score (α = 0,63, 95%CI 0,27 to 1,0). Hence, for this variable also the Cohen’s Kappa is conducted which indicates, based on guidelines form Altman (1999) and Landis & Koch (1977), that there is substantial agreement, κ = 0,62 (95%CI 0,013 to 0,017), p = ,005. Based on these outcomes, the coding for ‘understanding’ was discussed in more depth and specified/modified in the code book.

3.4 Data cleaning

The separate excel files in which all tweets were collected were imported in SPSS and merged into one file. IBM SPSS Statistics 23 software was used to explore and clean the data, checking the model assumptions for analyzing and interpreting results, and in conducting the different analyses. By checking frequencies and descriptive statistics, oddities and missings were indicated. Some coding errors were solved (e.g. a missing value where a value should be present). ‘Content format “poll”’ as well as the firm response ‘denial’ were excluded from analysis, because they were constant (zero occurrence). Extreme values in CE were not seen as outliers, because these extreme values are controlled by including control variables such as the ‘number of followers’, ‘number of following’, ‘total number of tweets’ and ‘source’. Based on the frequencies (low occurrence) table of type 2 tweets (see table 14) and lack of theoretical foundation for the relationship between these variables and CE ‘compensation’, ‘firm liking’ and ‘firm retweeting’ were not included in the multiple regression analysis.

Customer tweets (type 1 and 3) contain missings for the firm tweets (type 2), where firm tweets contain missings for customer related variables. So the misings in the dataset are there, because they are collected and coded this way. In the process of conducting multinomial logistic, logistic and multiple regression analysis to explore the relationships, this seems to cause two restrictions in the current dataset (appendix C).

The first restriction occurs because of the missing values in the current dataset. By performing a multiple regression analysis it regresses IVs on a DV containing missing values, which results in errors. This is also the case for the ‘response time of the firm’, ‘firm liking’ and ‘firm

retweeting’. Replacing these missing values by variable means or creating dummies for the

(29)

28 The second restriction was caused by the coding according the ‘type of tweet’ and again also by the missing values. By coding all firm responses as 2 in the current dataset it was not possible to make a distinction between the firm response to CGBC (type 1) and the firm response to the customer response (type 3). The variable ‘Order’ could not fix this coding restriction, because some CGBC starts with two tweets as the customer needed more characters than one tweet allowed. Also, the firm often replies in three tweets. So it was not possible to say something about the separate relationships of a firm response (type 2) on CGBC (type 1) and on customer response (type 3) and the effect on CE, because type 2 contains responses towards both. Splitting the dataset by the ‘type of tweet’ variable to organize output by these separate groups followed by a multiple regression analysis also caused errors due to the first restriction; the missing values.

Both restrictions can be solved by adapting the original datasets, e.g. by adding a new variable to indicate the difference in firm response on CGBC (type 1) and the customer response to the firm (type 3) so that the separate relationships can be explored as well. Moreover, creating separate data sets or by restructuring the data set so that structural missings are eliminated can solve the missing restriction. Due to a lack of time, this is not included in this research and remains for further research.

3.5 Statistical validity

When estimating parameters in ordinary least squares’ (OLS) regression the following four assumptions need to be satisfied, in order to ensure and uncover the degree of statistical validity for interpreting the results: autocorrelation, heteroscedasticity, multicollinearity and normality (leeflang, Wieringa, Bijmolt and Pauwels, 2015). After testing the assumptions and adjustments/transformations in the data where needed, the different multiple regressions were conducted.

Autocorrelation

(30)

29 (lowest DW = 1,614 , highest DW = 2,238). So, no autocorrelation in the residual is present, meaning that significance levels in the regression results can be interpret.

Heteroscedasticity

Heteroscedasticity reduces the efficiency of the parameter estimates and is not desirable (Leeflang et al. 2015). To check whether the variances of the error terms are constant (homoscedasticity), residual scatter plots were computed when performing the regression analysis to indicate heteroscedasticity. The residual plots indicate that error terms of the DVs are heteroscedastic since the dots are all clustered and not spread in a cloud, as with homoscedastic data. Different transformations of the DVs (square root, log, Box-Cox) were used to reduce heteroscedasticity and generate homogeneity, however none of these transformations solved this problem. Furthermore, opportunities for using the macro of Hayes and Cai (2007) (HCREG) were explored, which can be used in transforming the errors on continuous basis and thereby produce more reliable regression results. Using this macro requires additional programming language and is in this research not further explored due to difficulties that arise from looking at the separate effects of CE for type 1 and 3 tweets. However, when the other assumptions are met, significance tests are unaffected and OLS estimations can be used, even when heteroscedasticity is detected (Richard Williams, 2015). In sum, when the other assumptions hold, significance tests will be unaffected, however violation of this assumption should be kept in mind.

Multicollinearity

Multicollinearity checks the correlations between two or more IVs. When multicollinearity is high (>10) parameter estimates become unreliable (Leeflang et al. 2015). The Variance Inflation Factor (VIF) scores were used to check the correlations between the IVs. Extreme

multicollinearity occurred (VIF values up to 427.312) when regressing the control variables on

the DVs. Mean centering and standardization of DVs gave the same results. Moreover, creating

dummy variables with the biggest category as reference category did not improve the values. However, when selecting the control variables with p values < 0,05 for the specific DVs (see table 15), the multicollinearity assumption was, for the regression analysis concerning customer response characteristics, not violated anymore (VIF < 10)(Curto and Pinto, 2011).

(31)

30 ,888, p < ,001), which implies that both are very related towards each other. Upon leaving

‘followings’ out of the analysis, only a small change in R2 occurred and all multicollinearity

problems disappeared (VIF < 10). Activities on a Twitter account (i.e. creating a tweet, engaging with other tweets) are shared on the timelines of the ‘followers’ of this Twitter account and not on the timelines of the ‘followings’ of the Twitter account. Altogether, leaving ‘followings’ out of this analysis resulted in the least changes in predicting variance in CE. Now all VIF values are < 10 and almost all of them < 5 and thereby the VIF values do not exceed the cut-off value of 10 (Curto and Pinto, 2011). Clearly, no correlation between two or more independent variables in all the models exists. All VIF values are presented in the multiple regression results tables (see table 15, 16, 18 and 20).

Normality

(32)

31 (Field, 2009). Green (1991) suggests a sample size for testing multiple correlations of N > 50 + 8m (where m is the number of IVs). In this research, in the case the maximum number of IVs is 22, a sample size of N > 50 + 8*22 = 226 is required. This holds true since the total sample size is 951 and the minimum sample size in multiple regression is 235.

In conclusion, all assumptions are met except the heteroscedasticity assumption, so results should be interpreted keeping these outcomes in mind.

3.6 Method of analysis

To explore the relationships of the customer- and firm characteristics on CE, OLS regression is used. The DVs used in the regression analysis were transformed by means of the Box-Cox transformation (see section 3.5). Because of the restrictions in the data set (discussed in section 3.4) a revised version of the conceptual model, that visually represents the multiple regression analysis, is provided in figure 2.

Figure 2 - Revised conceptual model

Effect of control variables on CE

Prior to testing the overall conceptual model, a multiple regression analysis is performed including all 21 control variables regressing them on the four CE variables. The statistical equations is as follows:

(1) CExi = β0 + β1* CFSi+ β2 * CFG + β3 * CTTi + β4 *CSi + β5 * CCi + β6 * CDKPNi + β7 * CDNSi + β8 * CDBOLi +

β9 * CTEi + β10 *CDTMi + β11 * CDTAi + β12 * CDTEi + β13 * CDWKNDi + β14 * CDJi + β15 * CDJLi + β16 * CDAi + β17 *

CDSi + β18 *CDOi + β19 *CDNi + β20 *CDPGi + β21 *CRPi + ε Where:

CExi = Customer engagement (Norm and Z) ( x = Replies, Mentions, Retweets and Likes; i = 1…950)

CFSi = Number of followers of the twitter account (Z) (i = 1…950)

(33)

32

CTTi = Total number of tweets by the twitter account (Z) (i = 1...950)

CSi = Source; others than customer and firm infer the conversation[JH1] (i = 1...950)

CCi = Competition is mentioned in the tweet (i = 1...950)

CDKPNi = Dummy variable for the firm KPN (1) and other (0) (i = 1…950)

CDNSi = Dummy variable for the firm NS (1) and other (0) (i = 1...950)

CDBOLi = Dummy variable for the firm Bol.com (1) and other (0) (i = 1...950)

CTEi = Total tweet exchange in conversation (Z) (i = 1…950)

CDTMi = Dummy variable for tweet sent in the morning (1) and other (0) (i = 1…950)

CDTAi = Dummy variable for tweet sent in the afternoon (1) and other (0) (i = 1…950)

CDTEi = Dummy variable for tweet sent in the evening (1) and other (0) (i = 1…950 )

CDWKNDi = Dummy variable for tweet sent in the weekend (1) and other (0) (i = 1…950)

CDJi = Dummy variable for tweet sent in June (1) and other (0) (i = 1…950)

CDJLi = Dummy variable for tweet sent in July (1) and other (0) (i = 1…950)

CDAi = Dummy variable for tweet sent in August (1) and other (0) (i = 1…950)

CDSi = Dummy variable for tweet sent in September (1) and other (0) (i = 1…950)

CDOi = Dummy variable for tweet sent in October (1) and other (0) (i = 1…950)

CDNi = Dummy variable for tweet sent in November (1) and other (0) (i = 1…950)

CDPGi = Dummy variable for product type ‘good’ (1) and other (0) (i = 1…950)

CRPi = Time response promise of the firm towards the customer (Z) (i = 1…950)

As explained in section 3.5, regressing all control variables on the DVs caused multicollinearity problems. Therefore, only the control variables with a p < 0,05 for the specific DV (replies, mentions, retweets and likes) were selected as control variables in further regression analysis. In addition, this is the source of the differences in equations (2, 3, 4, 5,6 and 7). Moreover, the large amount of dummies with more than two categories resulted in an ambiguous and hard to interpret constant coefficient, because it represents several reference categories in one constant. Lastly, an insignificant constant occurred for all the models including the control variables.

Effects of CGBC and customer response characteristics on CE

An overview of the statistical equations for regressing CGBC (type 1) on CE and the effect of the response of the customer to the firm’s response (type 3) on CE are presented as follows:

(2) CERi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * TIi+ β7 * DGVPii + β8 *

DGVNi + β9 * POSHi + β10 *NEGHi + β11 * NEUHi + β12 * HIEi + β13 * MIEi + β14 * LIEi + β15 * CSi + β16 *

CFSi + β17 * CFGi + β18 *CTTi + β19 * CTEi + β20 * CDKPNi + ε

(3) CEMi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * TIi+ β7 * DGVPi + β8 *

DGVNi + β9 *POSHi + β10 *NEGHi + β11 * NEUHi + β12 * HIEi + β13 * MIEi + β14 * LIEi + β15 * CSi + β16 *

CFSi + β17 * CFGi + β18 * CDJi + ε

(4) CERTi,Mi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * TIi+ β7 * DGVPii + β8 *

DGVNi + β9 * POSHi + β10 *NEGHi + β11 * NEUHi + β12 * HIEi + β13 * MIEi + β14 * LIEi + β15 * CSi + β16 *

CCi + β17 * CDJLi + ε

Where:

CExi = Customer engagement (Norm and Z) ( x = Replies, Mentions, Retweets and Likes; i = 1…246 (type 1); i = 1…235 (type

3))

CFTi = Content format text present (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CFPi = Content format photo present (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CFVi = Content format video present (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CFUi = Content format URL present (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CFAi = Content format animation present (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

TIi = Total intensity of the tweet (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

DGVPi = Dummy variable for positive tweets (1) and other (0) (i = 1…246 (type 1); i = 1…235 (type 3))

DGVNi = Dummy variable for negative tweets (1) and other (0) (i = 1…246 (type 1); i = 1…235 (type 3))

POSHi = Number of positive hashtags used in the tweet (Z) ((i = 1…246 (type 1); i = 1…235 (type 3))

NEGHi = Number of negative hashtags used in the tweet (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

(34)

33

HIEi = Number of high intensity emoticons used in the tweet (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

MIEi = Number of moderate intensity emoticons used in the tweet (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

LIEi = Number of low intensity emoticons used in the tweet (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

CSi = Source; others than customer and the firm infer the conversation (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CFSi = Number of followers of the twitter account (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

CFGi = Number of followings by the twitter account (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

CTTi = Total number of tweets by the twitter account (Z) (i = 1…246 (type 1); i = 1…235 (type 3))

CTEi = Total tweet exchange in conversation (Z)

CDKPNi = Dummy variable for the firm KPN (1) and other (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CDJi = Dummy variable for tweet send in June (1) and other (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CCi = Competition is mentioned in the tweet (1) or not (0) (i = 1…246 (type 1); i = 1…235 (type 3))

CDJLi = Dummy variable for tweet send in July (1) and other (0) (i = 1…246 (type 1); i = 1…235 (type 3))

ε = error term that follows a normal distribution with a mean of 0

When performing the regression analyses for model 2, 3 and 4, a negative adjusted R2 occurred for mentions in type 1 (one negative adjusted R2) tweets and type 3 (all negative adjusted R2) tweets. To track changes in (adjusted) R2 and see when a negative (adjusted) R2 occured, forward selection procedure was applied (Halinski and Feldt, 1970). Forward selection is based on the statistical characteristics of the IVs and control variables. In other words, this study contains nominal scaled variables and scale variables as specified in SPSS. First all nominal scaled IVs were added, followed by the scaled IVs and the same was done for the control variables. Forward selection can be used to uncover confounds, detect reasons for a negative R2 and to see how the different models perform. The regression models performed based on forward selection can be find in Furthermore, some constants remain insignificant.

Effects of firm response characteristics on CE

As referred to in section 3.4, selecting the control variables for de DVs with p values < 0,05 did not solve multicollinearity for the firm’s response’s characteristics. Therefore, ‘Following’ was left out to solve this problem. In order to test the effects of the firm response (type 2) on CE, the following statistical equations are presented:

(5) CERi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * DVRi + β7 * APi + β8 * RHi + β9 *

DMi + β10 * SIGi + β11 * CNi + β12 * UNDii + β13 * COMi + β14 * CSi+ β15 * CFSi + β16 *CTTi+ β17 * CTEi + β19 * CDKPNi

+ ε

(6) CEMi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * DVRi + β7 * APi + β8 * RHi + β9

*DMi + β10 * SIGi + β11 * CNi + β12 * UNDii + β13 * COMi + β14 * CSi+ β15 * CFSi + β16 * CDJi + ε

(7) CERTi,Mi = β0 + β1* CFTi + β2 * CFPi + β3 * CFVi + β4 * CFUi + β5 * CFAi + β6 * DVRi + β7 * APi + β8 * RHi + β9

*DMi + β10 * SIGi + β11 * CNi + β12 * UNDii + β13 * COMi + β14 * CSi + β15 * CCi + β16 * CDJLi + ε

Where:

CExi = Customer engagement (Norm and Z) ( x = Replies, Mentions, Retweets and Likes; i

CFTi = Content format text present (1) or not (0) (i = 1…469)

CFPi = Content format photo present (1) or not (0) (i = 1…469)

CFVi = Content format video present (1) or not (0) (i = 1…469)

CFUi = Content format URL present (1) or not (0) (i = 1…469)

CFAi = Content format animation present (1) or not (0) (i = 1…469)

DVRi = Dummy variable for voice of response where (1) is professional and (0) is human (i = 1…469)

(35)

34

RHi = The firm used related humor in the tweet (1) or not (0) (i = 1…469)

DMi = The firm mentioned direct message in the tweet (1) or not (0) (i = 1…469)

SIGi = The firm used a signature in the tweet (1) or not (0) (i = 1…469)

CNi = The firm used the customer name in the tweet (1) or not (0) (i = 1…469)

UNDi = The firm used understanding in the tweet (1) or not (0) (i = 1…469)

COMi = The firm used a compliment in the tweet (1) or not (0) (i = 1…469)

CSi = Source; others than customer and the firm infer the conversation (1) or not (0) (i = 1…469)

CFSi = Number of followers of the twitter account (Z) (i = 1…469)

CTTi = Total number of tweets by the twitter account (Z) (i = 1…469)

CTEi = Total tweet exchange in conversation (Z) (i = 1…469)

CDKPNi = Dummy variable for the firm KPN (1) and other (0) (i = 1…469)

CDJi = Dummy variable for tweet send in June (1) and other (0) (i = 1…469)

CCi = Competition is mentioned in the tweet (1) or not (0) (i = 1…469)

CDJLi = Dummy variable for tweet send in July (1) and other (0) (i = 1…469)

ßk = kth population regression coefficient (k = 1…19)

ε = error term that follows a normal distribution with a mean of 0

Firm response characteristics and changes in customer tweet characteristics

A restriction in the current data set (explained in section 3.5) makes testing the relationship by regression analysis between CGBC, the firm response, CE and CES for this research not possible. However, to still say something about the relationship between CGBC, the firm response and CE a t-test for unequal sample sizes to compare significant coefficients from type

1 and type 3. This comparison is made to uncover how the firm response changes characteristics

in the customer tweet. Moreover, to test the relationship between the CGBC, the firm response and CES a Kruskal-Wallis test and independent t-test were conducted to see if the ‘general

message valence’ and ‘total message intensity’ of type 1 and type 3 tweets differ (see figure 3).

Hence, looking at these differences, may imply that the characteristics of the firm’s response influence ‘general message valence’ and ‘total message intensity’ of the customer tweet.

Do Firms Deal with Customer Generated Branded Content on Twitter in the right way?

Do Firms Deal with Customer Generated

Branded Content on Twitter in the right way?

Do Firms Deal with Customer Generated

Branded Content on Twitter in the right way?

By

Jasmijn Staal

Master Thesis Marketing

S2324660

Nassaulaan 29a

9717 CG Groningen

06 – 48 79 31 52

j.j.h.staal@student.rug.nl

University of Groningen

Faculty of Economics & Business

Department of Marketing

PO Box 800, 9700 AV Groningen

First supervisor: Dr. J.C. Hoekstra

ABSTRACT

ACKNOWLEDGEMENT

TABLE OF CONTENTS

1. INTRODUCTION

2. LITERATURE REVIEW

3. METHODOLOGY