Is silence golden? The influence of corporate
socio-political activism on eWOM and firm performance: A
Black Lives Matter event study.
Is silence golden? The influence of corporate
socio-political activism on eWOM and firm performance: A
Black Lives Matter event study.
Master’s Thesis Marketing EBM867B20
MSc Marketing Intelligence & Marketing Management University of Groningen
Faculty of Economics and Business
10th of January, 2021 Sophie Bode Oosterhamrikkade 33F 9713 KA Groningen s.j.bode@student.rug.nl S3774341
ABSTRACT
Driven by highly competitive marketplaces and proliferating socio-political issues in the 21st century, consumers have been increasingly demanding firms to express their values by becoming advocates for one side of a partisan socio-political topic – also conceptualized as Corporate Socio-political Activism (CSA). The Black Lives Matter (BLM) movement is recognized as one of the largest social movements in recent history, causing polarization in society in many ways. Hence, as one can assume, brands attesting their support for this movement has the potential to jeopardize or enforce consumer relationships, thus making its repercussions for firm’s financial performance uncertain. Limited research has been devoted to the consequences of CSA, and especially the question of how brand’s activistic initiatives propagate on social media has remained unanswered. Consequently, the aim of current research is to identify the impact of BLM-related CSA on online evaluations (electronic word-of-mouth) via Twitter, and subsequently on investor responses and stock performance. In doing so, several moderators that can influence online responses toward BLM-related CSA are examined, based on the Social Identity and Self-Congruity theories. Using an event study methodology and sentiment analysis to uncover the polarity of online evaluations on Twitter, I analysed 42 BLM-related CSA events initiated by 20 firms over a period of two months. Through a series of panel and logistic regressions, I find that, on average, BLM-related CSA elicits negative online response. However, the negative valence can be mitigated and even turned around into positive valence dependent on the ethnicity of the brand’s following. The relationship is also contingent on whether or not the topic of BLM is more focal due to any events related to it. Moreover, I find that the volume of tweets increases due to a brand’s CSA and that the opinions become more polarized. eWOM generally educes favourable investor responses, yet become unfavourable when opinions become increasingly disunited. Interestingly, while investors initially act in advantage of the firm, this beneficial effect is diminished when BLM-related CSA is repeated over time. These findings contribute to the current body of literature on CSA and provide valuable insights in how CSA can influence online chatter and investor decision making.
PREFACE & ACKNOWLEDGEMENTS
Before you lies the thesis “Is silence golden? The influence of corporate socio-political activism on eWOM and firm performance: A Black Lives Matter event study”, a captivating study I have been working on from September 2020 until January 2021. As an advocate for this social movement and what it stands for, I was inspired by the responses BLM brought forth from many brands globally. I therefore greatly appreciate the opportunity to explore the underlying dynamics of brands as activists, especially in support of this movement.
During these few months I have gained an incredible amount of knowledge in terms of CSA theory, event study methodologies, performing (sentiment) analyses on unstructured data and using the Python language. I am hugely grateful for this journey and the concepts I can now call part of my knowledge. That said, it would not have been such a thought-provoking process (although definitely challenging at times) without the assistance of a few invaluable people.
Firstly, I want to express my gratitude for my supervisor, Dr. Evert de Haan. Without his knowledge, trust, time and calmness I would not have been able to carry out this research
successfully. Every Google Meet videocall with him left me energised and at peace at the same time, which showcases Evert’s supervisor skills. Secondly, I gratefully thank Dr. Mehrad Moeini Jazani for his counsel and for giving me the opportunity to attend an MSI webinar, which
deepened my understanding of the practical application of my topic and elevated my eagerness to study brands as activists even more. I also wish to gratefully thank my fellow master students Kelly, Justina, Annebeth & Permata for keeping me level-headed. Their constant encouragement did not go unnoticed and motivated me to stay the course, even when I felt the data was
analysing me instead of the other way around – so to speak. Last but definitely not least, I have an inestimable amount of respect for my dear family and friends, for enduring my (at times off the charts) stress-level and all the moods that come along with it. They have especially created an environment of ongoing support and pride, which means more to me than I can express in words.
CONTENTS
1 INTRODUCTION ... 1
2 LITERATURE REVIEW ... 5
2.1 Corporate Socio-political Activism ... 5
2.1.1 Antecedents of CSA ...5
2.1.2 CSA motives ...6
2.2 Brand evaluations ... 7
2.2.1 Brand evaluations and congruency...8
2.3 Online brand evaluations and its response to CSA ... 10
2.4 BLM and CSA credibility ... 11
2.5 Model and hypotheses development ... 12
2.5.1 Online brand evaluations ... 13 2.5.2 Online brand evaluation on firm performance ... 14 2.5.3 CSA on firm performance ... 16 2.6 Moderating variables ... 16 2.6.1 Events related to BLM ... 16 2.6.2 Background of target audience ... 17 2.6.3 Leadership’s ethnic diversity ... 19 3 DATA COLLECTION ... 20
3.1 Conditions for the dataset... 20
3.2 Dependent variable ... 21 3.3 Independent variables ... 23 3.3.1 BLM-related CSA ... 23 3.3.2 eWOM ... 24 3.3.3 Background variables ... 26 3.3.4 Events related to BLM ... 27 3.3.5 Ethnic diversity of leadership roles ... 28 4 METHODOLOGY ... 28 4.1 Event study ... 29
4.2 Model specification for panel regression ... 29
4.3 Data preparation ... 30
4.4 Model specification and assumptions... 32
4.4.1 eWOM (volume and valence) ... 32
4.4.2 Financial performance (AR, CAR, CAAR)... 34
4.4.3 Direct effects ... 36
5 RESULTS ... 38
5.1 Descriptive statistics and model-free evidence ... 38
5.2 Hypotheses tests ... 43 5.2.1 Effect of BLM-related CSA events on eWOM ... 43 5.2.2 The effect of eWOM on financial performance ... 48 5.2.3 Alternative direct effects ... 51 6 DISCUSSION AND IMPLICATIONS ... 53 6.1 Theoretical implications ... 54 6.2 Managerial implications ... 55
7 LIMITATIONS & FURTHER RESEARCH ... 56
8 APPENDIX ... 58
8.1 Appendix A... 58
8.2 Appendix B... 62
1
1 INTRODUCTION
“We must dismantle white supremacy. Silence is not an option.”
- Ben & Jerry’s, 2020
On the 2nd of June 2020, Ben & Jerry’s campaigned this statement alongside a four-point call to action to annihilate police brutality and racial injustice. Strictly ideological brands such as Ben & Jerry’s are known to speak up about socio-political topics, but more often than not, brands remain reticent and focus on their performance characteristics rather than their pursuit of societal change (Eilert & Nappier Cherup, 2020; Vredenburg et al., 2020; Kotler & Sarkar, 2017). McKinsey’s view on the silence of firms on socio-political topics is attributed to “short-term financial pressures, a lack of familiarity with the issues, and the sense that specialists in the public-affairs and legal departments handle this sort of thing” (Bonini et al., 2006, p. 21). However, recent events concerning police brutality and racial injustice have shown us brands have come a long way since McKinsey’s statement in 2006.
2 As much as there has been (online) support for the movement, a recent study by Pew Research Center (2020) among ten thousand U.S. adults shows that the numbers of supporting and opposing respondents are close, with 29% strongly supporting the movement and 30% strongly opposing. The polarized opinions from the public are further illustrated by the pungent reactions towards brand conducts. For example, Starbucks faced backlash after their company policy disallowed employees to wear apparel that supports the BLM movement, as it could induce feelings of “divisiveness” (Murphy, 2020, para. 2). Likewise, Adidas’s human resources procedures and the lack of promoting diversity thereof, resulted in the resignation of the chief of HR, as a reaction to employee rage (Loh, 2020). Conversely, a campaign between Nike and Black Lives Matter-activist and Football player Colin Kaepernick resulted in antagonists to launch the adversary hashtag #BoycottNike (Kelner, 2018).
These partite responses towards the issue that the firm addresses are crucial in delineating firm’s engagement in corporate social and political activities from Corporate Socio-political Activism. The key characteristic of Corporate Socio-political Activism is the dichotomized nature of the social issues firms actively voice their concerns or opinions about, as is the case with the BLM-movement. As one can assume, voicing an opinion that is in favour of one side of a partisan societal or political topic, brings the risk of losing consumer groups that have opposing views. Therefore, only recently firms have been taking active stances on such topics, driven by the progressive concern of consumers for firm’s contribution to society as a whole (Bhagwat et al., 2020; Bhattacharya et al., 2020; Hambrick & Wowak, 2019; Klein & Dawar, 2004; Stanaland et al., 2011; Eilert & Nappier Cherup, 2020).
3 et al., 2009). In turn, these positive consumer perspectives lead to enhanced brand awareness, brand image, brand engagement and ultimately brand equity (Torres et al., 2012). However, literature also cautions that the effects of CSR initiatives largely depend on the companies’ values. Vredenburg et al. (2020) assert that when the consumers’ trust is discredited, commonly by the act of ‘woke washing’, consumer hate arises following into anti-brand activism. Woke washing entails the act of spreading an activism-motivated message while not having the value-driven social corporate practices to support this message (Romani et al., 2015). In such terms, corporate brand activism can influence consumers’ evaluation of said brand in an opposing fashion when the firm does not have the corporate values that support their advocate belief for the realms they take a stand on.
4
What is the effect of firm’s engagement in corporate socio-political activism with regards to the Black Lives Matter movement on online brand evaluations and subsequently on firm
performance?
Overall, current research centers four main questions: 1) How does corporate socio-political activism influence consumers’ brand evaluations (i.e. how do consumers respond to BLM-related CSA) in terms of valence and volume? 2) How do events related to BLM, ethnicity, political preference, ethnic diversity in leadership roles and the transparency of brands modify these brand evaluations? 3) How do brand evaluations affect firm performance? and 4) Is there a direct effect of socio-political activism on firm performance? While answering these questions, current research aspires to advise brand managers on the importance of brand-cause fit, what the consequences of CSA might be and how to effectively respond to partisan societal topics. Hence, current research will shed light on several characteristics than can moderate the relationship between social issue messages and the consumer’s evaluation.
5
2 LITERATURE REVIEW
In this section the definitions, relationships and existing literature on Corporate Socio-political Activism (hereafter mentioned as “CSA”) and (online) brand evaluations will be provided. The theoretical framework is set up as follows. Firstly, the concept of CSA is presented, followed by how (online) brand evaluations are formed and how these concepts relate to each other. Thereafter, the connection with the BLM movement is described. Subsequently, these relationships are visualized in a conceptual model. Following, the study’s hypotheses for the relationship between CSA, (online) brand evaluations and firm performance with moderating variables are brought forward.
2.1 Corporate Socio-political Activism
Corporate Socio-political Activism is a rather new term with many corruptions. Scherer et al. (2016) use the term Political Corporate Social Responsibility in their research. Nalick et al. (2016) speak of Corporate Socio-political Involvement, whereas Rim et al. (2020) draw on Corporate Social Advocacy. Scholars also mention the broad term ‘corporate activism’ (Eilert & Nappier Cherup, 2020; Sethi, 1982; Vredenburg et al., 2020). Although corrupting terms have been used, research on CSA as characterized by Bhagwat and colleagues (2020) is scarce. Popular definitions do not mention actual actions of the firm, but rather limit socio-political activism to ‘statements’ that are communicated (e.g. Hambrick & Wowak, 2019). However, Bhagwat et al. (2020) found evidence that 40% of the acts of activism in their sample were accompanied by a form of action. Therefore they have defined CSA as “a firm’s public demonstration (statements and/or actions) of support for or opposition to one side of a partisan socio-political issue” (Bhagwat et al., 2020, p. 1). Socio-political issues are characterized by their divisive, unsettled, emotionally-charged and contested nature (Nalick et al., 2016). Examples are contentions regarding same-sex marriage, LGBTQ+ rights, immigration or gun control but also racial injustice (Bhagwat et al., 2020; Nalick et al., 2016).
2.1.1 Antecedents of CSA
6 law” (Bhagwat et al., 2020, p. 3; Bhattacharya et al., 2020, p. 2050; McWilliams & Siegel, 2001, p. 117). In practice, CSR resources are for example devoted to sustainability, improving labor policies, environmental causes or human rights (Sen & Bhattacharya, 2001). Similarly to CSA, firms that engage in CSR are doing so to address issues that do not only affect the firm but society as a whole (A. Bhattacharya et al., 2020; Nalick et al., 2016). However, as Bhagwat et al., (2020) depict, initiatives of CSR oftentimes give attention to issues where opinions are neutral instead of polarized, unlike with CSA (i.e. partisanship is low, and societal consensus is high). For CSA practices, there is no agreement on how to respond to socio-political issues, which can engender heated discussions across societal groups (Nalick et al., 2016).
Firms can, beyond societal topics, also pursue to form government policies in favour of the firm – commonly known as CPA (Hillman & Hitt, 1999; Lawton et al., 2013). Alike CPA, Hambrick & Wowak (2019) point out that the primary aim is of CSA to influence political change in an espoused direction. Nonetheless, the difference between CPA and CSA is that CPA has an underlying motive that furthers financial payoff rather than social payoff (Bhagwat et al., 2020). Thus, in case of CSA, there is a weaker link to the business activities (Nalick et al., 2016). Another key difference is the degree of which the actions are public. While with CSA firms take a public stand, CPA is performed quietly (Bhagwat et al., 2020; Nalick et al., 2016). On these grounds, CSA initiatives bring more risk and are more controversial than CSR or CPA initiatives and therefore require a different approach.
2.1.2 CSA motives
7 mechanism of firm-consumer relationships is the attachment consumer feel towards brands that act on goodwill – by creating consumer-company identification.
The marketing perspective contends that firms are using CSA as a tactic to stand out in an increasingly competitive marketplace (Vredenburg et al., 2020) by leveraging the associations that consumers have of them to gain sustained advantage over competitive firms (Sen & Bhattacharya, 2001). This means, while consumers might share certain associations with competing brands, a brand’s unique positioning as “activistic” can be a compelling reason to buy from that particular brand (K. L. Keller, 1993).
Taken as a whole, it appears that an important aspect of a brand’s motive to engage in corporate social behavior is the positive effects it has on brand evaluations. Following this train of thought, in the next subsections, I will discuss how brand evaluations are formed, the relevancy of congruency for (positive) evaluations and how these constructs relate to CSA, especially in the online environment.
2.2 Brand evaluations
Considering that a brand is a vital resource for any business not only from a marketing but also from a financial perspective (K. L. Keller, 1993; Lassar et al., 1995; Moisescu, 2007), it is of importance to understand how consumers form their opinions about them. Brands are defined as “a name, term, sign, symbol, or design, or combination of them which is intended to identify the goods and services of one seller or group of sellers and to differentiate them from those of competitor” (Keller, 1993 p. 2; Kotler, 1991, p. 442). Keller (1993) asserts that a firm’s most valuable asset is the knowledge of how the brand is perceived in the consumers’ mind, as it is a driver of consumer behavior (e.g. purchase intent). He furthermore argued that favourable evaluations are formed when consumers believe that the brand has attributes that satisfy their needs.
8 firms that engage in CSA, as the act of taking responsibility for societal and political topics humanizes the firm (C. B. Bhattacharya & Sen, 2003; Olsen et al., 2014).
Bhagwat et al. (2020) propose that within the framework of socio-political issues, it is of importance to find congruency between the personality traits a brand consists of and the consumer’s perspective of the brand for increased financial performance. As Korschun et al. (2017, p. 21) explain: “When consumers are exposed to a political issue, they evaluate the company’s behavior in large part based on how the company’s actions and statements align with their central and enduring values”. Therefore, the forming of brand evaluations and the importance of congruency will be examined hereafter.
2.2.1 Brand evaluations and congruency
Fishbein & Ajzen (1975) propose that brand evaluations are formed in a two-fold manner: 1) through the salient beliefs a consumer has about the brand (i.e. whether or not a brand has certain (beneficial) attributes) and 2) the judgement of those beliefs (i.e. whether or not the consumer thinks it is good or bad that the brand has these attributes or benefits). These attitudes can be based on the consumer’s perspective on product characteristics, but also about non-product related characteristics such as the actions of companies selling those products (C. B. Bhattacharya & Sen, 2003; K. L. Keller, 1993).
Research has implicated that consumers form their beliefs about a brand based on their self-concept. Namely, through the notion that the image the consumer has of him- or herself impacts how a brand’s attributes (as formed in their mind) are evaluated (Aaker, 1997; C. B. Bhattacharya & Sen, 2003; K. L. Keller, 1993). This phenomenon is the basis of Social Identity Theory (Tajfel, 1974) and Self-Congruity Theory (Sirgy, 1982), both of which posit that consumers generally feel a match with brands that share common characteristics with them, like views and morals (C. B. Bhattacharya & Sen, 2003; Hur et al., 2020; Sen & Bhattacharya, 2001).
9 & Sen, 2003). This process can be explained through the Self-Congruity Theory developed by Sirgy (1982), which illustrates that people are most comfortable with buying products from organizations that are consistent with the social identities that define their self-concept (Kleine et al., 1993). In other words, this self-categorization process advances when the self-concept is in concordance with the firm’s identity. Several authors propose that strong congruence and strong incongruence correlate with consumer purchase behavior in opposing ways (Landon, Jr., 1974; Sirgy, 1982). For instance, Sirgy (1982) introduced the following forms of congruency and its impact on purchase motivation: 1) Positive self-congruency: when there is a positive self-image and a positive product-image. Here the motivation to purchase will be high, affirmed by Landon Jr. (1974), in several product domains. The high motivation is based on the premises that consumers are inclined to purchase a product that is positively valued to maintain an optimistic self-image and thus uphold their self-esteem (Sirgy, 1982); 2) Positive self-incongruency: when there is a negative self-image and a positive product-image. Here there will be a conflict regarding purchase motivation. On the one hand, the utilization of a product that holds a favourable image to the buyer will lead to a more favourable self-image, for example through positive reinforcement by advertisements or the buyer’s relevant reference group (Reed, 2002). On the other hand, from a self-consistency perspective, consumers tend to avoid buying products that are not congruent with their self-image belief (Sirgy, 1982); 3) Negative self-congruency: when there is a negative self-image and a negative product-image. Here there will be a conflict regarding purchase motivation. The conflict arises from two motives: either to protect their esteem or to be self-consistent (Sirgy, 1982). With the former motivation, consumers might avoid buying products that progressively lowers their negative self-image. Hence to protect their self-esteem, they will avert making a purchase. With the latter motivation, consumers might approach products that are consistent with their self-image, albeit negative; 4) Negative self-incongruency: when there is a positive self-image and a negative product-image. In this case there will be avoidance behavior (i.e. no purchase motivation). The discrepancy between the consumers self-concept and the brand’s identity creates dissonance, which threatens consumers’ sense of self (Swann et al., 1992). This results in restraint from buying the product (Sirgy, 1982).
10 consumers that feel congruent (incongruent) with a brand have preferential (unfavourable) attitudes towards the brand (Loureiro et al., 2017), which are both strong motives for eWOM (Fatma et al., 2020; Hennig-Thurau et al., 2004). Especially in CSR settings, as found by Fatma et al. (2020), considering the anthropomorphical effects CSR has on the brand – which enhances consumers’ willingness to evaluate online. Moreover, Xun & Guo (2017) argue that investigating eWOM as proxy for brand evaluations is effective as it quantifies to what extent the online evaluation impacts firm performance and for how long, in contrast to regular WOM – since traditional evaluations only impact a limited number of peers and with a much lower speed. They further argue that the decontextualized virtual environment facilitates honest opinions about firm’s actions since the tie to the company itself is weaker, due to the anonymity of user accounts. As a result, brand evaluations are conceptualized as eWOM in the present research. In the next paragraph, eWOM and its relationship with CSA will be discussed.
2.3 Online brand evaluations and its response to CSA
11 media increases the quality of the CSR message through its two-way communication propensities which fosters consumers’ demand for authenticity. When the quality of the message increases, proposingly, the arguments in the message are perceived as stronger. Leong et al. (2019) found that stronger arguments tend to persuade the target audience to engage in information processing, adopting and sharing, as per the Information Adoption Model (Sussman & Siegal, 2003). Notably, Eberle et al. (2013) found that the valence of the responses of other consumers on online CSR messages also matter. Positive evaluations by others seem to have the greatest effect on identification with the firm, which in turn stimulates positive eWOM. Contrarily, negative evaluations by others have a significant negative effect on message credibility, which stimulates negative eWOM. These findings suggest that CSA activities can foster and increase online evaluations in the form of eWOM, especially when influenced through peers – which is stimulated by social media through its sharing, commenting and evaluating attributes (Babić Rosario et al., 2020; Chu et al., 2020).
2.4 BLM and CSA credibility
It is of importance to recognize that there are extraneous components that can either reinforce or jeopardize the positive effects of customer-firm congruence in controversial settings. Especially since research clarifies that consumers who identify with brands’ CSR initiatives engage in positive consumer behavior (Hur et al., 2020), and negative consumer behavior when those initiatives endanger the credibility of the brand (Alcañiz et al., 2010). This might be the case when consumers feel that a brand’s actions are driven by selfish motivations. Swaminathan et al. (2020) argue that integrity and credibility have to be prominent when firms take a clear stance in controversial topics, as failing them can foster identity conflict among consumers that use the brand to communicate important values. Integrity means that the brand’s intentions and values are relevant for the realms they take a stand on (Swaminathan et al., 2020). Credibility has to do with the truthfulness and believability of the message (Eberle et al., 2013).
12 wrong-doings of a firm (Romani et al., 2015), which one can expect is the case with BLM-movement, seeing the amount of public support.
Moreover, the sincerity of the action has been seen as a key indicator of whether or not a company is perceived as credible (i.e. value-driven and not profit-driven) in their social responsibility intentions in empirical research by Alcañiz et al. (2010). Similarly, experimental research by Korschun et al. (2017) shows that when consumers agree with a political stand of a company, purchase intent increases – moderated by corporate hypocrisy. They conceptualize corporate hypocrisy as a mismatch between the values of the company and the companies’ actions. When corporate hypocrisy increases, purchase intent mitigates significantly. In other terms, consumers can be congruent with the firm in terms of political ideologies, however, if they feel the firm or brand is hypocritical they will less likely purchase their products or services. This outcome is the strongest for result-oriented companies, which are companies that focus on performance more than beliefs and values. Arguably, this indicates that opposing consumers have more compelling responses than supporting consumers. Taking these findings into account, one can conclude that in order to encourage positive consumer behavior, firms have to pursue credibility and integrity when communicating their opinion on controversial topics.
2.5 Model and hypotheses development
13 Figure 1: Conceptual model of the influence of BLM related socio-political activism on abnormal stock returns,
mediated by online brand evaluations. The relationship between the IV (X) and mediator (C) is moderated by events related to BLM (M), political preferences and ethnical background of the followers of the brand (M) and the
diversity among employees of the brand (M).
2.5.1 Online brand evaluations
14 on a brand’s involvement with this movement – as it modifies the level of identification (C. B. Bhattacharya & Sen, 2003; K. Keller & Aaker, 1990). Due to the polarized nature of these CSA activities, this will expectedly cause a dispersion in the evaluation of a brand (Bhagwat et al., 2020). A dispersion in evaluation is furthermore assumed seeing that people that voice their opinions online are often very satisfied or very dissatisfied (de Haan, 2020). Hence, I propose that the evaluations of a brand will get increasingly branched out when brands voice their point of view on the BLM discussion. Moreover, seeing the idiosyncratic characteristics of social media, it is predicted that a large audience is reached with these BLM-related campaigns. For these characteristics, consumers are stimulated to interact from to-consumer and consumer-to-brand, even more so due the high involvement with the topic. In turn, this increases the influence on several other consumers, thereby increasing the volume of eWOM. Therefore, I propose the following hypotheses:
H1a: The volume of eWOM will increase due to the firm’s engagement with BLM related CSA. H1b: eWOM will be more polarized (i.e. have a higher standard deviation in terms of sentiment)
due to the firm’s engagement with BLM-related CSA.
2.5.2 Online brand evaluation on firm performance
The evaluation of a brand is a vital component of the consumer’s decision-making process. Namely, because it is salient information coming to mind when a consumer wants to purchase a product or service (K. L. Keller, 1993). Accordingly, evaluations (also in the form of eWOM) are a good predictor of future purchase behavior, because it forms the basis for the preference for one brand over the other (K. L. Keller, 1993; Xun & Guo, 2017).
15 is bound to be harmed (Kudeshia & Kumar, 2017). For this, investing poses more risks as there is potential for increasing backlash by the public (Bhagwat et al., 2020). Positive valence, however, signals the secure brand-customer relationship (e.g. through increased customer loyalty). Moreover, positive evaluations can result in the recruitment of new customers (Peng et al., 2015). Both encourage financial performance (Babić et al., 2015; Luo et al., 2010). Positively as well as negatively valenced brand evaluations are monitored and evaluated by investors to forecast future cash flow prospects (Bhagwat et al., 2020; Luo et al., 2010), and therefore expectedly increase or decrease the firms stock performance based on investor decision making. Furthermore, the volume of the eWOM reflects the magnitude of the topic at hand, which enhances the positive or negative effects of the valence of eWOM. Consequently, I propose:
H2a: Positive (negative) eWOM regarding a firm positively (negatively) associates with the
firm’s abnormal stock return.
H2b: The volume of eWOM moderates the relationship between eWOM valence and
financial performance in a way that an increased amount of volume strengthens the positive (negative) effects of positive (negative) eWOM valence.
16 H2c: Increased negative sentiment of eWOM impacts the abnormal stock return more
profoundly compared to increased positive sentiment of eWOM.
2.5.3 CSA on firm performance
While I expect a mediated relationship between corporate social activism and stock return via eWOM, literature suggests that this mediation might be partial. Bhagwat and colleagues (2020) found evidence for other confounding variables influencing the relationship between CSA and stock return. Their overall negative proposition of firm’s CSA on stock performance is confirmed when the CSA initiative deviates from legislator values as well as shareholder values. They confirmed that this divergence from values increases perceived uncertainty for investors, as the responses from key stakeholders are then rather unpredictable (Bhagwat et al., 2020). Nalick et al. (2016) furthermore argue that investors are less prone to buy a stake in companies that engage in CSA as they deem that the firm’s time, resources and attention is allocated to CSA, which decreases their dedication for other, likely more profitable, activities. Therefore, I hypothesize:
H3: Firm’s engagement in BLM-related CSA has a negative effect on abnormal stock
returns.
2.6 Moderating variables 2.6.1 Events related to BLM
17 accompanied by social sharing and the seeking of social support (Pennebaker et al., 1997), which in the 21st century is commonly via online communication tools (Glasgow et al., 2016). Hence, expectedly, the more events happen, the more the community will voice their concerns with or for the movement. This will result in progressive awareness on the topic, especially when voiced via the online sphere (Babić et al., 2015; De Choudhury et al., 2016). Furthermore, the more events happen, the more the supporting public will feel that there is an unwavering lack of accountability for responsible police officers – meaning the battle for racial equality and police humaneness is far afield (Carney, 2016). In consequence, I expect that the increasing awareness of the movement and the topics related to it as well as the disappointment with the progress of the movement will contribute to consumers’ increasing demand for firm’s CSA on the matter. This expectation is based on the notion that present day consumers expect brands to help solve societal and political issues in the world (Kotler & Sarkar, 2017). The CSA of brands will then become more focal, thus, increasing the dispersion of opinions. Therefore, I hypothesize:
H4a: Firms engagement in BLM-related CSA increases when events related to BLM occur. H4b: Events related to BLM moderate the relationship between BLM-related CSA and
eWOM in a way that events increase eWOM volume and increases the polarized opinions.
2.6.2 Background of target audience
18 consumer’s self-concept (i.e. political and ethnical background) à the higher the identification and/or tendency to strengthen this self-concept through consumption à the higher the likelihood of forming favourable attitudes towards the brand. Research affirms that consumers ascribe even more positive associations to a brand if the attributes of the brand are seen as important to them (Fishbein & Ajzen, 1975; K. L. Keller, 1993), which further strengthens the hypotheses that political and ethical background moderate the CSA and brand evaluation relationship.
19 H5a: The congruency (incongruency) between BLM-related CSA and political ideology of the
target audience moderates online brand evaluations towards BLM-related CSA such that eWOM is more positive (negative) when the congruence decreases (increases).
H5b: The congruency (incongruency) between BLM-related CSA and ethnical background of the
target audience moderates online brand evaluations towards BLM-related CSA such that eWOM is more positive (negative) when the congruence decreases (increases).
2.6.3 Leadership’s ethnic diversity
Several cases show that brands that have taken a stance in favour of the BLM-movement have faced backlash if their HRM policies, marketing activities or corporate values did not match their message of support. Exemplary, L’Oréal Paris drew criticism from a large community when their statement of solidarity expressed that “speaking out on the topic is worth it”, while in the past they have dropped a model from their campaign who spoke out about racism and white supremacy (Young, 2020, para. 3). Pepsi, in attempt to open conversation about racial injustice and police brutality via an ad starring Caucasian reality-tv star Kendall Jenner, received intense denunciation. Consumers felt Pepsi used imagery from real protests to sell their product, instead of addressing the underlying issue (Victor, 2017). In this context, however, it seems that the most adverse reactions are given towards firms that support the movement while diversity is still a subject of discussion within their firm (Stengel, 2020; Harper, 2020). As such, one can assume that brands that declare their support for racial justice while not having a diverse employee portfolio can be seen as deceptive and dishonest. These impact the brand’s integrity and credibility, which is linked to negative consumer behavior (Alcañiz et al., 2010). Moreover, it has been known that companies attract prospective organizational members that match with the companies ideologies, also known as stakeholder alignment, even in the context where these ideologies are distinct (Hambrick & Wowak, 2019). This would imply that brands that stand firmly on their support or opposition of a partisan topic would also attract employees that agree with their ideology. In turn this can influence the credibility and integrity of BLM-supporting messages. Taking this into account the following hypotheses are proposed:
H6a: The transparency in ethnical diversity of leadership roles within the brand
20
publishing (not publishing) a diversity and/or inclusion report positively (negatively) influences eWOM.
H6b: Ethnical diversity of leadership roles within the brand moderates the relationship
between BLM-related CSA and brand evaluations, in a way that high (low) ethnic diversity positively (negatively) influences eWOM.
3 DATA COLLECTION
The following sections will detail the collection of data. First, the conditions for the dataset are explained. Then, the different sources of data for the study are described, starting with the dependent variable and followed by the independent and moderating variables. To analyse the hypothesized relationships, I have collected data on 42 BLM-related CSA events, initiated by 20 publicly traded firms from different industries between the 25th of April and the 25th of June of 2020.
3.1 Conditions for the dataset
21 data to analyse eWOM. Lastly, the firm should be part of the list on The Social Listening
(www.social-listening.org) that provides information on the political ideology of brand’s Twitter
following (further explained in section 3.3.3.1). A full list of the brands in the sample can be found in Table 9 in Appendix A.
3.2 Dependent variable
To measure the reactions of the stock market towards BLM related CSA, a formula was used to equate abnormal stock returns. This is the underlying mechanism of an event study methodology, which is explained in section 4.1.
Abnormal stock returns measure whether the rate of return deviates from the expected return of the overall market in a period of time, which could be due to any changes in the market or even marketing capabilities (Angulo-Ruiz et al., 2018). Financial theory proposes that there is an advantage to measuring firm performance via abnormal stock returns, as stocks have a forward-looking focus and can have a distinguished link to specific events opposed to other performance measures such as return on sales, return on assets or return on equity (Geyskens et al., 2002). Moreover, stock prices can change in real time. When any new information becomes public, investors can react quickly by buying or selling a stock based on their expectations of the news’ impact on future cash flows (Geyskens et al., 2002). Hence, abnormal returns can be positive or negative, depending on the underperformance or overperformance of the stock related to the expected rate of return – which is related to investor responses to new information.
Data on stock performance was yielded via Yahoo! Finance, who provide daily stock data. The abnormal returns were calculated via equation (1).
!"#$ = "#$− (("#$) (1)
Where Rit is the daily return on day t of the stock per firm i, E(Rit) is the expected return of the stock on day t per firm i. This abnormal return, or unexpected change in the stock price, can then be attributed to a specific event that took place at time t (Geyskens et al., 2002).
To calculate E("#$), I used the Capital Asset Pricing Model (Sharpe, 1964). The Capital Asset Pricing Model accounts for the time value of money that could influence a firm’s return and is based on the following equation (2):
22 Where E(Rit) is the expected return of the stock on day t per firm i. ", is the risk-free rate of interest, .#$ is the beta of investment and ((("/) − ",) is the market risk premium. Daily data on the risk-free rate of return and the market risk premium is extracted from the Fama-French library (2020). Similar to Geyskens et al. (2002) and Beckers et al. (2017), the beta of investment is an ordinary least squares parameter estimate, obtained by regressing the market risk premium on the expected returns over an estimation period of 250 days prior to the event period. The abnormal returns are then calculated as in equation (1). To assess whether the returns are indeed abnormal, they have to be significantly different from zero. By means of a t-test, results show the returns are indeed abnormal (p < 0.05, M = 0.22%).
In line with related studies (e.g. Bhagwat et al., 2020), I also compute the cumulative abnormal returns (CAR) for several windows around the BLM related CSA events, via equation (3).
0!"#(1′, 1) = ∑ !"$ #$
$5 (3)
In doing so, I can capture any spill over effects in the abnormal returns which can for example be caused by a delayed response of investors or information leakage before the event day. Namely, because in practice information is not disseminated during the event day only, but also days prior and post the event day (Geyskens et al., 2002), especially when user-generated content is considered (Tirunillai & Tellis, 2012). The most common event window in event studies is three days, starting at t1 = -1 and ending at t2 = 1, but recent research also used a window of five days (e.g. Bhagwat et al., 2020). Hence, CAR is computed for alternative t’ and t ∈ {-2, -1, 0, 1, 2}. The most significant window is chosen via a t-test, as described by Brown & Warner (1985), which is in line with the procedures of previous studies (e.g. Beckers et al., 2017; Bhagwat et al., 2020; Geyskens et al., 2002). The five-day window is the most significant window and will therefore be used in further analyses (p < 0.01, M = 1.21%). Succeeding, because the event study is conducted over 42 (N) events, I averaged the CAR into a cumulative average abnormal return (CAAR), via equation (4) (see e.g. Geyskens et al., 2002; Beckers, 2017). Where N is the number of CSA events being studied.
0!!"(1′, 1) = ∑9 0!"#(15, 1)/8
23 Each of these distinct abnormal returns calculations above will be analysed as a dependent variable, explained in section 4.4.2, because their characteristics reveal whether the effects are immediate or subjected to spillover effects. Further, analysing each individual abnormal return variable separately serves as a robustness check of the model.
3.3 Independent variables 3.3.1 BLM-related CSA
The independent variable is the CSA events that twenty brands have initiated in the beforementioned time period. Multiple BLM-related CSA events i are analysed per brand J to ensure enough observations are included in the dataset. Moreover, analysing multiple events gives the opportunity to test the robustness of the research concerning whether or not the CSA effects are stronger for the first CSA announcement or statement versus later actions.
Hence, there are a few independent variables to consider. Firstly, I compute a dummy variable indicative of whether or not a firm had initiated an act of CSA on a particular day. Secondly, three count dummy variables are classified to indicate the ascending events (i.e. the dummies represent whether it was the first, second, or third CSA event).
24 Table 1. Examples of BLM-related CSA in the sample
Brand BLM-related CSA Date Political
stance
Average polarity
CAR
Nike Nike initiates the “For once, Don’t Do It” campaign to bring attention to racism and asking their followers not to turn their backs on the problems in America.
30th of May 0.14 0.08 1.49%
EA Electronic Arts (EA) added an in-game message in their popular soccer game FIFA20 about their support for the BLM movement and vouched to combat racism in their games.
8th of June 0.10 0.001 1.98%
Domino’s Domino’s announced to donate to supporting organizations such as the Black Girls Code and the National Urban League as well as investing $1 million into their newly found “Black Franchisee Opportunity Fund”.
12th of June 0.12 0.03 0.99%
Twitter Twitter initiated a series of billboards with handpicked statements on the Black Lives Matter movement to be broadcasted in different US cities to motivate protesters.
19th of June 0.22 -0.01 1.62%
Political stance (-1 = Extremely Conservative & 1 = Extremely Liberal) Average polarity (<0 = Negative & >0 = Positive)
3.3.2 eWOM
25 19% of Tweets being related to a brand (Jansen et al., 2009). Twitter is chosen as a source of eWOM for several reasons: 1) Twitter is one of the most used forms of social media with 340 million daily users who send out 500 million Tweets per day (Sehl, 2020), which enables customer’s voice to be heard more easily, and information to be spread faster and with a wider community (Xun & Guo, 2017); 2) Users can write up to 280 characters (used to be 140 characters until 2017; Twitter, 2020), which increases the volume of information and aids data collection; 3) Tweets can be easily scraped via Python software; 4) By default, Tweets are publicly visible, which benefits scraping methods (Jansen et al., 2009); 5) Many brands are active and use Twitter for their CSR messages (K. Lee et al., 2013); 6) Twitter is increasingly used during emergencies or disasters (Jansen et al., 2009), such as the events related to BLM, which advances user data or opinions on brand related matters on the topic; 7) Tweets remain available forever unless deleted by the user or Twitter itself (de Haan, 2020).
The tweets scraped for this study are written in English, and it is clear that the tweets are indeed about the brand in question. I used “snscrape”, a Python package developed by “JustAnotherArchivist” in 2018, to scrape all tweets where brand i was mentioned between the 25th of April and the 25th of June. To scrape the tweets, only the twitter handle of the brand (e.g. @Twitter) is searched for rather than the brand name combined with #BlackLivesMatter or #AllLivesMatter, to get a general view of users’ evaluation. Also, the package does not allow for this specification due to recent changes in Twitter’s policy. Furthermore, by confirming that the Twitter handle of the brand is mentioned in the tweet, it is ensured that the user meant to reach the brand which enhances the evaluative disposition of the tweet and guarantees that the user is talking about a particular brand (e.g. not ‘a target’ instead of Target). In total, I scraped a little over 3.5 million tweets, which averages at around 2.800 tweets per day per brand.
3.3.2.1 Sentiment analysis
26 events (Liu & Zhai, 2012). With the ‘qdap’ package, one can uncover polarity scores, which indicate how positive or negative a text is. The analysis is sectioned as follows. First, for each tweet, the algorithm of the package tags positive and negative polarized words which are a product of the sentiment lexicon of Hu & Liu (2004). Next, the algorithm derives information from the context of polarized words by taking a cluster of four words before and two words after each polarized words, which is labelled as the “context cluster” (de Haan, 2020). This context cluster (<=> ) can unveil whether the polarized words in the tweets are shifted in valence due to its context, and thus are classified in four types (see e.g. de Haan, 2020 and Jockers, 2017):
1) Neutral words, which do no shift the valence of the polarized word;
2) Negating words, which reverse the valence of the polarized word (e.g. “I do not like it.”); 3) Amplifying words, which increases the impact of the polarized word (e.g. “I really like
it.”);
4) De-amplifying words, which reduce the impact of the polarized word (e.g. “I hardly like it.”);
Subsequently, the sentiment score per tweet is calculated by summing up the polarity score per context cluster of a tweet, and divide them by the square root of the number of words (de Haan, 2020). Neutral/mixed tweets equal a polarity score of 0, negative tweets equal a polarity score < 0 and positive tweets equal a polarity score > 0. Finally, the polarity scores per tweet are aggregated as an average score per day per brand.
3.3.3 Background variables 3.3.3.1 Political background
To operationalize the political background of the consumers, I measured the political stance as a ratio of Liberal to Conservative followers. The website The Social Listening (
http://www.social-listening.org) reveals the political associations of 604 different brands on Twitter. I identified the
ratio of individual brands in the dataset in terms of Liberal versus Conservative followers for brand
i via equation (5).
(?@$AB CDEFDG$AHD @, I#JDEAB ,@BB@K#GHL)M(?@$AB CDEFDG$AHD @, N@GODEPA$#PD ,@BB@K#GHL)
27 The ratio equals +1 if all followers are Liberal and -1 if all followers are Conservative.
3.3.3.2 Ethnic background
To account for the ethnical background of the following of the several brands, data was collected from Google Trends (https://www.trends.google.com) and the US Census bureau (2019) to calculate a diversity score per brand. I used Google Trends to see from which American state the distinct brands were most searched within the same time period as the abnormal stock return calculation. To illustrate, the term ‘Nike’ was most searched within this period from the state New York. Google Trends accounts for the size of the different states when providing the state that most searched a specific term, and therefore, also smaller states can be brought forward. This data is then matched with data of the US Census bureau, who present the states’ population by ethnicity. The data classifies the ethnicities of every state as: Hispanic, Non-Hispanic White, Non-Hispanic Black, Non-Hispanic Asian and Non-Hispanic American Indian. The US Census bureau (2019) identifies every citizen that is Non-White as ethnic diverse. However, different proxies for ethnic diversity are generated. For example, the variable ‘Black_Proxy’ is created using solely the population of Non-Hispanic Blacks in relation to the total population for the diversity score. In this way, the study will reveal whether the BLM affected population (i.e. the Non-Hispanic Blacks) significantly identifies with the matters at hand, or whether (a combination of) other ethnic groups unveil similar results. The diversity score depicts the ratio between the proxies for several ethnic diverse groups in relation to the total population. Thus, the diversity score of ethnicity for each brand i is calculated by equation (6).
?@$AB CDEFDG$AHD @, D$QG#F R#PDEOD F#$#SDGO @, O$A$D
?@$AB C@CTBA$#@G @, O$A$D
(6)
Where the ‘ethnic diverse citizens’ are proxied by the several Non-White ethnicities mentioned before. The ratio equals 1 if all followers are part of the ethnic diverse proxies and 0 if all followers are Non-Hispanic White.
3.3.4 Events related to BLM
28 event happening and the value 0 indicated no event happening per day t. This timeline of events is constructed based on information from Pew Research Center (2020) combined with announcements of major US news outlets. Pew Research Center analysed the hashtag #BlackLivesMatter after crucial incidents related to the BLM movement, and therefore composed an outline the movement overtime. Events include but are not limited to fatal arrests, such as the arrest of George Floyd on May 25th, but also a series of protests against racial injustice initiated in 50 states of America as well as countries across the world (Nardini et al., 2020). The list of events can be found in Table 10 in Appendix A.
3.3.5 Ethnic diversity of leadership roles
I determine the ethnic diversity of the leadership roles of brand i based on publicly available data sources (e.g. Wall Street Journal) and corporate documents (e.g. diversity and/or inclusion reports). First, a binary variable is created to indicate whether or not the firm at hand published a Diversity & Inclusion report since 2019. Here, a 0 indicates no report was published and a value of 1 indicates a report was published. This binary variable will proxy the transparency of the firm in terms of their employee diversity. Thereafter, the diversity in ethnicity of each firm’s leadership is rated as an index of the total directory representation. To illustrate, Nike reports that 4.8% of their leadership positions are filled by Black employees, which would result in an observation of 4.8 for ‘Leadership Black’ versus an observation of 95.2 ‘Leadership Non-Black’ for Nike. In light of the current research, only the representation of Black leaders is considered. Namely, because former research has shown that integrity of the brand in CSA initiatives relate to the relevancy for the realms that brands take a stance on (Swaminathan et al., 2020). Firms that engage in BLM related CSA are initiating that Black lives matter, hence, only the Black representation of the leadership roles are included.
4 METHODOLOGY
29 4.1 Event study
As aforementioned, this study measures abnormal stock returns as an evaluation of firm performance, in relation to certain BLM-related CSA events. Measuring abnormal stock returns in accordance with an ‘event’ is the underlying mechanism of an event study methodology. Event studies originate from financial research, but have been broadly used in marketing studies in recent years (see e.g. Sorescu et al., 2017) as well as in studies of corporate social responsibility (see e.g. McWilliams et al., 1999). The objective of event studies is to investigate the impact of an event through changes in stock prices (Delattre, 2007). Events can be announcements of mergers, changes in organizational structures – or more marketing related – the introduction of new products. Recent advances by Bhagwhat et al. (2020) concluded that this methodology is also applicable for announcements linked to CSA.
Event studies have an underlying theory of market efficiency (Fama, 1970), which makes two assumptions: 1) in an efficient market, stock prices reflect all publicly available information; and 2) in an efficient market stock prices change immediately when new information becomes available (Sorescu et al., 2017). Under these assumptions, investors react to announcements that carry new information to alter their expectation of future cash flows of firms (Sorescu et al., 2017). Hence, if the changes in abnormal stock returns are statistically different from zero, one can infer that the event holds important information for investors, seeing they revised their expectation of the market based on this event, either by buying or selling stocks (Delattre, 2007). As mentioned in section 3.2, that is the case for current research.
4.2 Model specification for panel regression
In the current research, data is collected on twenty brands observed over multiple T time periods (25th of April until the 25th of June in 2020). Hence, the data is classified as panel data. This signifies that the data have both cross-sectional dimensions (i.e. multiple brands rather than a single entity) as well as time-dimensions (i.e. multiple periods in time). Panel data are considered as balanced when all cross-sections have the same number of observations (Wooldridge, 2002). In current research, the brands are all observed within the same time period, hence, the dataset is balanced (Ti = T for all i). A statistical test with the function ‘punbalancedness’ from the R package
30 First in panel regression methodology is the determination of whether or not the data suits a panel regression model significantly better than a regular OLS regression model. The OLS regression differs from the panel regression, as it does not consider heterogeneity across groups or time (Torres-Reyna, 2010). This determination can be established with a Breusch-Pagan Lagrange Multiplier test (LM test). Here, the LM test shows significance for all the specified models (p < 0.05), indicating that a panel regression model suits the data significantly better than the OLS model.
Next, I determine which panel model fits the panel data best. Panel data are generally associated with three types of models: the pooled model, the random effects model and the fixed effects model. The pooled model is the most restrictive panel data model, and is therefore not much used in literature (Katchova, 2013). The random effects model assumes that individual-specific effects ai are distributed independently from the regressors x. In other terms, the intercept is treated as a
‘random’ variable and therefore is included in the error term (Wooldridge, 2002). The fixed effects model allows for individual-specific effects ai to be correlated with the regressors x. This implies
that the individual-specific effects are treated as parameters to be estimated for each cross-section observation for i. Hence, ai is included as an intercept, which means that every brand i has a
different intercept, but the same slope parameters (Katchova, 2013). Ergo, whether or not the data allows for a random effects or fixed effects model is dependent on whether the individual-specific effects ai are correlated with the regressors. This correlation is detected with the Hausman test.
The null-hypothesis assumes there is no correlation, which prefers the random effects model. The alternative hypothesis assumes correlation, for which the fixed effect model is the appropriate model. For every specified model, as explained in the subsequent sections, the Hausman test was performed. On all occasions, the random model was preferred (p > 0.05), which indicates there is no significant correlation between the unique errors and the regressors as the individual-specific effects move independently from the regressors (Katchova, 2013; Torres-Reyna, 2010).
4.3 Data preparation
31 (abnormal) returns and consequently the CAR and CAAR are therefore not measured on 19 days. This results in a panel data model with N = 20 brands and T = 42 days is n = 840 observations. Further, the variable that indicates the percentage of Black leaders within the firms have some missing observations. These missing observations are a result of firms not publishing a diversity- and/or inclusion report, hence, the data on leadership representation is not available. However, as this is the case for five out of twenty brands, a lot of vital observations are missing. Therefore, the missing data for these five brands is imputed with the average observation of the remaining fifteen brands. Nevertheless, the fact the firms did not publish a diversity nor an inclusion report holds crucial information for the hypotheses. This information is accounted for by creating the dummy variable ‘Diversity Report Dummy’, indicative of whether or not the firm has published a diversity report.
32 polarity, negative polarity and volume further have VIF-values > 5. The VIF-scores are reduced by changing the positive and negative polarity scores to percentages of the total number of tweets rather than a sum of positive or negative tweets per day and are therefore included in the models.
4.4 Model specification and assumptions
In the subsections below, I will introduce the model specification and assumptions per individual hypothesis. For each of the models, a random effects panel regression analysis is performed by means of the R-package plm. In the first section of the analysis, the characteristics of eWOM are measured in relation to the IV and the moderators. Therefore, tweet volume, positive polarity, negative polarity, average polarity and the partisanship of the polarity (i.e. the standard deviation) are the dependent variables. In the subsequent analysis, the relationship between eWOM and financial performance is considered. Hence, the financial performance indicators (abnormal returns, cumulative abnormal returns and cumulative average abnormal returns) are the dependent variables. Additionally, analyses are performed to measure any direct effects. Along with it, the panel regression assumptions are presented for every individual model, as they all have distinct dependent variables and therefore relate to the assumptions in different ways. An overview of all the variables used in the panel regressions can be found in Table 11 in Appendix A.
4.4.1 eWOM (volume and valence)
To test the hypotheses of the effects of BLM-related CSA on eWOM (whether or not moderated by background characteristics, transparency and authenticity of the firm and BLM-related events), ten models are specified. These models differ in their dependent variable, but have the same independent variables and moderators. The direct effects and moderator effects are modelled separately, so that the moderators do not influence the direct effect. Thereafter the moderators are included as interaction terms. As such, the following models 1 – 5 (equation 7) are specified for the direct effects and models 6 – 10 (equation 8) are specified for the moderated effects:
UVWXJ$ = Y + Z ∗ 0\! U]U^1J$+ _J$ (7)
UVWXJ$ = Y + Z ∗ 0\! U]U^1J$+ ` ∗ 0\! U]U^1J$∗ .aX U]U^1$+ b ∗ 0\! U]U^1J$∗
cde=1=fge =hUdedijJ$+ k ∗ 0\! U]U^1J$∗ .egfl eUghUmnℎ=pJ$+ q ∗ 0\! U]U^1J$∗
33
where eWOMbt : (1) positive polarity, (2) negative polarity, (3) average polarity, (4)
standard deviation or (5) log(volume);
t: day 1 … day T; b: brand;
ε: error term
For interpretation purposes, the volume variable is log transformed during analyses, as then one unit increase in x explains a percentage increase in the volume. This holds more interpretational power than a unit increase in volume, which means one tweet extra was posted on a particular day
t. Also, as the volume variable is highly skewed, the log-transformation results in a more normally
distributed variable.
34 models 1 – 10. After, the Durbin Watson test was performed again, resulting in p-values > 0.05 (Tables 14 and 15 in Appendix B). With these p-values, I fail to reject the null-hypotheses that the residual errors are correlated over time, which means the assumption of independent residual errors is now met.
Next, the assumption of equal variances across the residuals is analysed, otherwise known as the assumption of heteroskedasticity. When the regression does not have minimum variance, the efficiency of the parameter estimates is reduced, which can lead to biased outcomes (Leeflang et al., 2015). For panel data regressions, the package “lmtest” includes a heteroskedasticity test with the Breusch-Pagan-Godfrey test statistic. The null-hypothesis of this test is that the variances in the residuals are equal. The alternate hypothesis is that the variance in the residuals are not equal, which calls for heteroskedasticity rather than homoskedasticity. For models 1, 3 and 6 – 10, the p-values were below 0.05 (see Tables 16 and 17 in Appendix B), resulting in a rejection of the null-hypothesis and the conclusion that the variances in the residuals are heteroskedastic. For panel models, the function vcovHC provides robust covariance matrix estimates which presents heteroskedasticity-consistent assessment of the variables’ effect. Resultingly, the estimates are reliable again.
Lastly, the assumption of a normal distribution of the error term is analysed. The error term has to have a normal distribution for the test statistic for hypothesis testing to be effective (Leeflang et al., 2015). As explained in Leeflang et al. (2015), one can examine an approximate normal distribution by plotting histograms of the residuals. For models 1 – 10, the histograms shown approve the approximate normality assumptions, which can be found in Appendix B (Figure 5). Note that I also performed tests to determine whether the data has any cross-sectional dependencies by means of the Pesaran’s CD test. For all the models in the study that had significant cross-sectional dependencies, the GLS-transformations to solve the violation of autocorrelation also lifted the problem of residuals that are correlated across entities.
4.4.2 Financial performance (AR, CAR, CAAR)
35 effect with volume is added. The financial performance measures are not log transformed for two reasons: they can become negative, which holds important interpretational value and second, they are already measured in percentages.
t=^g^f=ge pUmudmvg^fUJ$= Y + Z ∗ pdn=1=]U pdegm=1jJ$+ w ∗ ^Uig1=]U pdegm=1jJ$+ ` ∗
n1g^hgmh hU]=g1=d^J$ + b ∗ log (]de{vU)J$ + _J$ (9)
t=^g^f=ge pUmudmvg^fU#$= Y + Z ∗ pdn=1=]U pdegm=1jJ$+ w ∗ ^Uig1=]U pdegm=1jJ$+ ` ∗ n1g^hgmh hU]=g1=d^J$ + b ∗ ]de{vUJ$ + q ∗ ]de{vUJ$ ∗ pdn=1=]U pdegm=1jJ$+ | ∗
]de{vUJ$ ∗ ^Uig1=]U pdegm=1jJ$+ } ∗ ]de{vUJ$ ∗ n1g^hgmh hU]=g1=d^J$ + _J$ (10)
where Financial performancebt: (1) Abnormal Returns, (2) Cumulative Abnormal
Returns, (3) Cumulative Average Abnormal Returns;
t: day 1 … day T; b: brand;
ε: error term
36 Next, the presence of heteroskedasticity is tested. The Breusch-Pagan-Godfrey tests are also significant for models 11, 14, 13, 15 and 16. The same solution was applied for these models as in section 4.4.1. Lastly, the normality assumption was examined and approved in line with section 4.4.1 (Figure 6 in Appendix B).
4.4.3 Direct effects
There are seven models to be estimated for direct effects. First, the models to determine whether or not BLM-related CSA events and the count dummies have a direct effect on the firm’s financial performance (models 17 – 19). Second, I model the direct effect of BLM-related events on the probability of a BLM related CSA-event to happen (model 20). Third, I model the direct effect of BLM-related events on the dummies that represent the first, second or third CSA approach and their probability to happen (models 21 – 23). Models 17 – 19 (equations 11) will be estimated by a Panel regression. Model 20 and 21 – 23 (equations 13 and 14) will be estimated by a Logit regression. The models are specified as follows:
t=^g^f=ge pUmudmvg^fUJ$= Y + Z ∗ 0\! U]U^1J$ + Z ∗ 0d{^1R/~; J$ + w ∗ 0d{^1R/~J$ + b ∗
0d{^1R/~ÄJ$ + _J$ (11)
where Financial performancebt: (1) Abnormal Returns, (2) Cumulative Abnormal
Returns, (3) Cumulative Average Abnormal Returns;
t: day 1 … day T; b: brand;
ε: error term
c(0\! U]U^1J$ = 1) = Y + Z ∗ .aX U]U^1$+ _J (13)
c(0d{^1_h{vvjJ$= 1) = Y + Z ∗ .aX U]U^1$+ _J$ (14)
where Count_dummy = (1) dummy for the first CSA-event, (2) dummy for the second
CSA-event and (3) dummy for the third CSA event.
