Using Twitter content to predict brand popularity
Amanda Pieper
2
Master’s Thesis Marketing Intelligence
Supervisor: Dr. H. Risselada
3
Using Twitter content to predict brand popularity
Amanda Pieper
Abstract
Communication platform Twitter is growing. This implies an increase in user-generated content (UGC). As this UGC is often brand-related, firms might be able to use this UGC to monitor brand-related topics discussed by consumers on Twitter. Furthermore, firms might be able to use this topic information to predict and perhaps influence their brand popularity. However, little is known about the effects of online topic chatter on brand popularity. This study therefore attempted to examine to what extent topics that dominate in Twitter content of consumers and firms can predict brand popularity. Latent topics were extracted from consumer Tweets and brand Tweets. Time series of topic intensities and relative search volume were examined, with relative search volume serving as a proxy for brand popularity. Topics were extracted from consumer and brand tweets with Latent Dirichlet Allocation (LDA). Topic words occurrences were converted to topic intensities, which were analyzed by using Vector Auto Regression (VAR) and Impulse Response Functions (IRF). In this study, we did partly find evidence that topics dominating in Twitter content can predict brand popularity. Results with regard to brand popularity prediction indicated that sustainability and product innovation chatter lead to significant negative effects on brand popularity, whereas other topics did not influence brand popularity. Another finding with regard to topic extraction is that brands mostly tweet about topics relating to product quality and service quality, whereas consumers tweet about more diverse topics, like product innovation, celebrity endorsement, sustainability and consumer satisfaction. This paper can give insight in how managers can monitor online chatter related to their brands. It can also guide managers in determining which topics have to be managed to influence brand popularity.
4
Contents
1 Introduction 5 2 Theoretical background 7 3 Data 14 4 Method 17 5 Results 216 Discussion and conclusion 33
7 References 36
5
1 Introduction
The user base of microblogging service Twitter is growing; where users sent more than 350.000 Tweets every minute in 2016 (Internet live statistics 2016), nowadays users send more than 520.000 Tweets every minute (Internet live statistics 2019). One reason for this growth is that Tweets can be quickly and easily sent (Murthy 2018). Because of this ease, Twitter is growing in importance; Twitter plays a prominent role in activism, disaster recovery and elections (Murthy 2015).
Tweets are a form of user-generated content (UGC). UGC is a type of word-of-mouth (WOM). WOM in commercial situations involves consumers sharing attitudes and opinions about businesses and products with other people (Jansen et al. 2009). Twitter is mainly used as a communication platform; “Twitter’s users are ultimately trying to say something to each other” (Murthy 2012, p. 1071). Users are motivated to contribute to Twitter because they receive intrinsic utility from the act of posting content, which means that simply posting content is satisfying (Ryan and Deci 2000; Toubia and Stephen 2013). Users also contribute to Twitter because of image-related utility; users are motivated by the perception of others (Toubia and Stephen 2013). Users seek status and want to be socially accepted (Lampel and Bhalla 2007; Toubia and Stephen 2013). Social media posts therefore serve as an important tool for self-affirmation, which arises from the need for identity confirmation. Users use Twitter as a tool to say ‘I exist’ (Murthy 2012). In addition to this, Twitter is also mainly analysed as a communication platform; social researchers have found Twitter useful for understanding complex social processes due to its easily accessible data and brief messages (Murthy et al. 2015; Sloan et al. 2015). Research has therefore mainly focused on influence and information diffusion on the Twitter network (Cha et al. 2010; Weng et al. 2010; Bakshy et al. 2011; Wu et al. 2011). Twitter is also used for text mining purposes, to get insight into consumer perceptions and behaviours (Fader and Winer 2012; Culotta and Cutler 2016).
6
In Twitter communication, two key aspects can be distinguished; consumer-to-consumer (C2C) and firm-to-consumer (F2C) communication. C2C communication consists of content which is not generated by firms, but generated through social interactions by consumers (Stephen and Galak 2012; Colicev et al. 2018). F2C communication consists of brands sending messages to consumers via Tweets. Online marketing in the form of Tweets posted by brands themselves is one of the best practices used to establish a brand in the online world and to increase its popularity (Aggrawal et al. 2017). The reason for this is that marketing online can be used in a more effective way; showcasing the brand’s products, maximizing brand reach and getting feedback (Aggrawal et al. 2017). This broader brand reach for firms and the increased production of brand-related UGC on a microblogging platform like Twitter might induce brand-related chatter, as the content generated by brands can change customer attitudes toward the brand, influence future purchase, or affect whether consumers will talk about the brand with their friends (Berger et al. 2019). Brands can thus gain exposure and strengthen relationships with customers through social media marketing (Kim and Ko 2012), which could affect the popularity of a brand. User-generated contents like comments or reviews on social networking sites have already been used for social media popularity analysis (Jamali and Rangwala 2009; Siersdorfer et al. 2014; Arora et al. 2016; Bansal et al. 2016; Aggrawal et al. 2017; Garroppo et al. 2018; Antoniadis et al. 2019). However, they mainly focus on predicting the popularity and acceptance of the social media comments itself, rather than the popularity of brands. In addition to this, the focus is often on content generated by consumers only (C2C) (Garroppo et al. 2018).
In this article we want to predict the popularity of a brand out of social media content, using both content generated by consumers (C2C communication) and firms (F2C communication). The aim of this study is therefore to examine whether topics that dominate in Tweets, expressed in UGC by consumers and online marketing by brands, can serve as a measurement tool for brand popularity. This leads to the following research question; To what extent can topics that dominate in Twitter content of consumers and firms predict brand popularity?
To answer the research question, we will search for latent topics emerging from brand Twitter data and examine to which extent topic intensities can predict brand popularity. The latent topics emerging from Twitter data will be linked to topics which can be found in literature. According to literature, topics where consumers and firms could talk about on Twitter that might influence brand popularity are product quality, consumer satisfaction, service quality, sustainability, product innovation and celebrity endorsement. Latent topics will therefore be linked to the aforementioned topics. These concepts will be elaborated further on in the theoretical section.
The goal of this research is firstly to find the most prevalent latent topics in Tweets by applying Latent Dirichlet Allocation (LDA). Then, word groups emerging from LDA will be grouped together into general topics that might influence brand popularity. Topic intensities resulting from LDA will serve as input for Vector Auto Regression (VAR), to examine to which extent these topics can predict brand popularity. Tweets by consumers and Tweets by brands will be separated from each other during LDA, to be able to discover differences in occurring topics between Tweets of consumers and brands and to examine whether there are different effects on brand popularity.
7
whereas consumers tweet about more diverse topics, like product innovation, celebrity endorsement, sustainability and consumer satisfaction.
Addressing the research question is relevant for several reasons. This study combines topic models with sentiment. Topic extraction has already been used to analyze data. However, the novel part in this study is that we do not only extract topics for the purpose of data summarization, we also examine the intensity of these topics. That is, we want to know how ‘intense’ chatter with regard to a certain topic is. Furthermore, this study relates topic intensities to brand popularity. This study could be relevant for management in several ways. If the Twitter discussion affects brand popularity, brands have the possibility to measure their brand popularity by monitoring brand chatter on Twitter. Consequently, brands can manage their brand popularity by managing their Twitter content. Brands might be able to manage their Twitter content in such a way that it becomes more likely for the brand to grow in popularity.
This article is structured as follows. Firstly, we will elaborate on the concept of brand popularity. Secondly, theory will be discussed of topics that might influence brand popularity. Then data characteristics and methods of analysis are presented, followed by the results and discussion section.
2 Theoretical background
Brand popularity
According to Yohn (2014, p.3), a brand is “a bundle of values and attributes that define the value you deliver to people through the entire customer experience, and the unique way of doing business that forms the basis of your company’s relationships with all of its stakeholders … Your brand is what your company does, your brand is not what you say you are - it is what you do.” This unique way of doing business can influence how the brand is doing; to achieve an increase in margin, revenue and market value, the firm must match their brand with consumer needs and wants (Chasser and Wolfe 2010). If firms succeed in this matching, it can increase brand performance and thus brand popularity. Brand popularity has already been defined as the accumulation of market acceptance over time (Kim and Chung 1997). Brands can score high or low on brand popularity during a certain time period. This overall score on popularity is the accumulation of market acceptance, the sum of market acceptance at a certain point in time.
Previous research on the concept of brand popularity is limited (Whang et al. 2015). Brand popularity has mainly been used to predict brand loyalty (Raj 1985), sales (Kim and Chung 1997), and differences in consumer attitudes and responses (Dean 1999; Lee and Kim 2008; Kim and Park 2013; Whang et al. 2015). This means that research has mainly focused on the influence of brand popularity on other concepts, instead of a focus on determinants for brand popularity itself. As mentioned earlier, there is also little research on how social media content can predict the popularity of the brand, as it is mostly about popularity of the social media content itself.
8
Topics that could influence brand popularity will now be discussed. The topics are expected to influence brand popularity for the following reason. Brand image, which consists of associations and beliefs, is formed partly by social media chatter (MacInnis et al. 2015). The topics that will be discussed below are topics which are discussed by consumers and brands on social media (Zeng et al. 2010; Mount and Martinez 2014; Andersson and Öhman 2017; Jiang et al 2017; Khamis et al. 2017; He et al. 2018). Chatter about these topics on social media can create associations and beliefs and could thus influence brand image. This brand image created by topic chatter could in turn affect brand popularity. The following topics will be discussed: product quality, consumer (dis)satisfaction, service quality, sustainability, product innovation and celebrity endorsement.
Product quality
Quality constitutes an inner feature of a product, that interacts directly with customers’ perceptions and can be recognized through experience (Cannatelli et al. 2017). Consumers tend to express their perceptions and opinions about product quality, product defects and product-related ideas in social media content (Jiang et al. 2017). Perceived product quality is a relevant determinant of firm performance (Cannatelli et al. 2017), which causes perceived product quality to have a strong impact on business success (Olbrich et al. 2017). If the quality of a product is perceived as poor, it loses attractiveness and its market share will decrease (Olbrich et al. 2017). Information expressing how consumers perceive product quality is consumed by an extensive audience on social media. This means that this information easily attracts public’s attention and has great social influence (Wang et al. 2017). Information on social media expressing how other consumers perceive product quality makes consumers able to better judge product quality, which could affect consumers’ own quality perceptions and their intentions to buy the product (Moussa and Touzani 2008). This effect of product quality related chatter on social media causing a change in intention to buy, can be regarded as a change in market acceptance of the product. That is, the brand becomes more or less widely sought after and is more or less frequently purchased by consumers. Therefore, if chatter about product quality perceptions changes consumers’ intentions to buy, this chatter could affect their acceptance of the product and thus brand popularity. Consumer (dis)satisfaction
9
One third of the social media users follow brands on social media (Edison Research 2014) and major social media channels like Facebook and Twitter are used extensively for customer complaints (Dekay 2012; Einwiller and Steilen 2015). This public nature of complaining online can increase the negative effects for firms if they do not handle the complaints well; when a company ignores online complaints, it leads to dissatisfaction and it reduces consumers’ intention to repurchase (Mattila and Mount 2003; Van Noort and Willemsen 2012). Because of the public nature and social influence online, other consumers will be exposed to the complaint, which could affect their satisfaction level as well. This indicates that the expression of satisfaction or dissatisfaction in a social media environment can affect market acceptance, which implies that chatter with regard to (dis)satisfaction on social media could have an effect on brand popularity.
Service quality
Service quality can be defined as a level of achievement in customer service, which refers to customers’ evaluation and perceptions of organizations’ service offering that could be favorable or not (Collins 2017). Due to a highly competitive market place, service quality has to be ensured to encourage customers’ return intentions (Lewison 1997). Higher service quality is associated with higher customer satisfaction (Ramanathan and Karpuzcu 2011) and positive WOM (Carrillat et al. 2007), which makes it an essential element of business’ survival (Collins 2017). How consumers experience service quality depends largely on their expectations. Their experience leads to a subjective evaluation and this perception is expressed (Collins 2017). Kornberger (2011) showed in a banking context that the concept of service quality is appealing and relevant for most consumers: 40 percent of potential customers were willing to switch to ING if ING was easier than its competitors. Easier meant easy contact with the firm, transparency and a fast and efficient firm, providing advice when the customer needs it, which comprises the concept of service quality.
Just like the expression of (dis)satisfaction, perceived service quality can be expressed on social media (Pan and Crotts 2012). Service quality has already been measured by studying the online opinions of consumers (He et al. 2018). Where perceived service quality explicitly refers to opinions and experiences with regard to the quality of the service that has been offered, the concept of satisfaction comprises opinions about all aspects of a brand and/or product. More and more customers are posting their service experiences on social media, using review websites, blogs, Facebook and Twitter (He et al. 2018). This service quality chatter affects sales performance (Sonnier et al. 2011), by guiding public opinion and forming consumers’ expectations. Therefore, service quality related chatter on social media can influence consumers’ return intentions (Zeng et al. 2010) by guiding consumers’ expectations with regard to the service quality of the brand. These service quality expectations can be reflected in increased or decreased market acceptance of the brand, which implies that brand popularity can be affected by service quality social media chatter.
Sustainability
10
Sustainability information on social media also spreads awareness relating to production conditions (Saeed et al. 2019).
Consumers are more favorable to companies that can provide products or services that satisfy their environmental needs (Chen et al. 2006). When consumers perceive firms as social responsible, this can compensate for the absence of a strong brand or a small advertising budget (Van Doorn et al. 2017). In addition to this, consumers have lower purchase intentions for products made with unethical processes when they have some control over production of these products than when they have no role in production (Paharia 2019). Direct responsibility causes feelings of guilt by consumers, which indicates that consumers value ethical and sustainable production (Paharia 2019). This matches with the finding that activation of self-accountability (the desire to live up to your own self-standards) leads to more positive reactions to ethical appeals, as consumers want to avoid anticipated guilt (Peloza et al. 2013). This valuation of ethical and sustainable production has led to an increase in sales of sustainable products and an increased pressure on firms to address these sustainability-related issues in their production process (Saeed et al. 2019). Examples of industries addressing these issues are the fashion industry (Kozlowski et al. 2012), the car industry (Vynakov et al. 2016; Koppelaar and Middelkoop 2017) and the (organic) food industry (Bezawada and Pauwels 2013). Research shows that addressing sustainability-related issues by means of corporate social responsibility (CSR) can positively affect customer satisfaction and this in turn can positively affect the market value of the firm (Luo and Bhattacharya 2006).
Consumers have an inclination to educate each other about a company’s sustainable practices (Saeed et al. 2019). This positive and negative sustainability-related information on social media significantly influences consumers’ intention to purchase sustainable products (Saeed et al. 2019). Satisfied customers can act as brand ambassadors and spread positive WOM, and unsatisfied customers can turn against a product or brand and spread negative WOM (Saeed et al. 2019). Thus negative social impacts can lead to negative publicity and positive social impacts can lead to positive publicity on social media. This publicity influences consumers’ intention to purchase these sustainable products and the acceptance of the product. Therefore, expressing sustainability-related information on social media could affect the popularity of the brand.
Product innovation
To grow, firms have to bring new products to the market (Langerak and Hultink 2006). According to Von Stamm (2009), innovation is the major driving force in organizations, caused by intensifying competition for consumers, employees and other critical resources. It is essential for firms to have the ability to continuously develop successful innovative products, services and processes (Von Stamm 2008). The reason for this is that firms can attract demand with new products and can maintain their position in competition with other firms (Christensen and Lundvall 2004). Product innovation is therefore crucial to attract customers from competitors (Christensen and Lundvall 2004).
11
channels. When consumers discuss innovation of a particular brand on social media, consumers might perceive that the firm is more aware of new preferences and trends, which could lead to an increased likelihood that their needs will be considered in new products of the brand (Prabhu et al. 2005). This could affect the popularity of the brand, as expectations towards the brand and market acceptance may change.
Celebrity endorsement
A celebrity endorser is defined as “anyone who enjoys public recognition and who uses this recognition on behalf of a consumer good by appearing with it in an advertisement (McCraken 1989, p.310). The use of celebrities has been a common marketing practice for firms to support or endorse their brand quality and corporate image (Prentice and Zhang 2017). A celebrity can attract new users and revive a product that has lost market share by creating new interest from consumers (Atkin and Block 1983). The use of a celebrity can also enhance brand equity (Till 1998), build brand credibility (Petty and Lindsey-Mulkin 2006) and influence purchase intentions (Spry et al. 2011). The reason for this is that a celebrity can attract a large audience (Khamis et al 2017). When celebrities lend their names to major brands, the celebrity can expose his or her own audience to that brand (Khamis et al. 2017). In the social media environment, consumers are saturated with so much to choose from. Competing for attention has therefore become important. Using celebrities to endorse a brand online is one way of getting this attention (Khamis et al. 2017).
Consumers are often low involved when processing advertisements (Petty et al. 1983). This is likely to be similar to the process of scrolling through posts on social media. Consumers then use the celebrity as a heuristic cue: the positive affect associated with a celebrity spills over to the product, which increases liking for the product (Petty et al. 1983). Beliefs about the brand have become consistent with beliefs about the celebrity (Fennis and Stroebe 2016). The use of celebrities to endorse products works because: “…using celebrities alongside products affects consumer behavior because of transference, attractiveness and congruence” (Lea-Greenwood 2013, p.77). This means that consumers believe that some of the skills of the celebrity might rub off on them if they purchase the brand (transference) or the attractiveness of the celebrity might rub off on them (attractiveness) (Lea-Greenwood 2013). There also must be a fit between the brand and the celebrity to make it credible to the consumer that the celebrity will wear the brand, which makes it more implicit that the celebrity is paid by the brand to wear the brand (Lea-Greenwood 2013).
12
Chung and Cho 2017). Consumers tend to like people who disclose information to them (Greene et al. 2006) and consumers are consequently more likely to accept the celebrity’s claims in an ad, which in turn lead to increased brand credibility (Chung and Cho 2017). This brand credibility serves as an indication of brand quality (Erdem and Swait 1998), which in turn leads to brand loyalty (Sweeney and Swait 2008) and purchase intentions (Wang and Yang 2010).
However, in this research we do not focus explicitly on celebrities endorsing a brand on social media. The focus is on chatter related to celebrities. That is, consumers who talk about celebrities that endorse a brand, it is not the celebrity itself that is posting the content. Nonetheless, we argue that the direct celebrity endorsement effects mentioned earlier might also hold for celebrity endorsement chatter. As consumers are low involved when scrolling through social media posts, simply the occurrence of the name of the celebrity might be able to produce the same effects as direct celebrity endorsement, even the post is not produced by the celebrity but by another consumer. The positive affect associated with the celebrity can still be activated when seeing the name of the celebrity, and spill over to the brand. Therefore, social media chatter with regard to celebrity endorsement may have a positive effect on brand popularity, due to increased brand credibility, brand loyalty and purchase intentions.
Sentiment
A sentiment is the emotional response of an individual toward an external stimulus (Lyu and Kim 2016). Capturing the sentiment of words from opinions and attitudes expressed in Tweets is important for monitoring brand reputation (Hassan 2017). About 30 percent of all words used in communication on social media signify a sentiment (Lyu and Kim 2016). This sentiment can be analyzed with sentiment analysis which is defined as the task of identifying positive and negative opinions, attitudes and emotions towards some subject or the overall polarity of a document (Pang and Lee 2008). Sentiment is for example reflected in text through affect; opinionated words like ‘great’, reflecting positive sentiment and ‘sad’, reflecting negative sentiment (Hassan 2017). However, the sentiment of words can also differ according to their context (Hassan 2017). It can be derived from this that all the topics where consumers talk about on social media mentioned previously, can be expressed with both negative or positive sentiment, which determines the relation with brand popularity. That is, consumers can be positive but also negative about, for example, product quality. Social media is often used to point out disputable business activities and dissatisfaction to a broad public, which has economic consequences for firms, such as a damaged public image (Stiglbauer et al. 2014). Derived from this, we argue that social media chatter about a topic, characterized by negative sentiment, could have a negative effect on brand popularity, whereas social media chatter about a topic, characterized by positive sentiment, could have a positive effect on brand popularity. The effect of topic chatter thus depends on the sentiment of the chatter.
Interdependencies between social media chatter and brand popularity
13
(Chung 1995). This implies that increased brand popularity leads to an increase in sales. It can be argued that when more consumers experience the use of a product due to increased sales, they might spread more WOM about the product, which could lead to increased social media chatter about the brand or an aspect of the brand, caused by increased brand popularity.
One example of this interdependency between brand popularity and brand chatter is the relation between product quality chatter and brand popularity. Previously we mentioned that social media chatter about product quality could increase brand popularity. However, brand popularity can also be used as an external cue for product quality, which influences consumers’ evaluation of alternatives (Chung 1995; Kim and Chung 1997; Dean 1999). Thus, when a consumer knows that a brand is a top-selling brand, he might evaluate the quality of that product more positively, which could increase chatter about product quality on social media. This implies that the relation between social media chatter about product quality and brand popularity can go both ways. The same holds for the concept of consumer satisfaction; when a consumer knows that a brand is a popular brand, this can affect the experience of the brand, which leads to a higher level of satisfaction (Chung 1995; Kim and Chung 1997). This could lead to more satisfied online chatter about the brand, which can in turn increase brand popularity even more. This means that brand popularity can increase consumer satisfaction, but consumer satisfaction can also increase brand popularity. Therefore, if brand popularity provides value to customers, then customers tend to return this value to firms by enhancing brand loyalty and ultimately providing more sales (Scherer and Ross 1990). This indicates that brand popularity might not only be the result of brand chatter, but it could also be a determinant for brand chatter. This phenomenon of Twitter topics influencing brand popularity and brand popularity influencing Twitter topics will be addressed in this research with a VAR model, which allows for these interdependencies.
In this research a brand popularity research framework has been developed to explore social media chatter concepts that affect brand popularity and to increase our understanding of how brand popularity can affect social media chatter. The research framework is shown in figure 1. The method used in this research allows for interactions between all concepts in the research framework.
Figure 1. Brand Popularity Research Framework1.
1 The research framework illustrates the relations that are examined with a VAR model. However, not all interdependencies are represented with arrows in the model.
14
3 Data
Dataset characteristics
This study has been designed to determine to which extent social media chatter about product quality, consumer satisfaction, service quality, sustainability, product innovation and celebrity endorsement influences brand popularity. This will be investigated by using two different data sources. Twitter data is used for the extraction of latent topics, which will result in topic intensities. Relative search volume data from Google Trends is used as a measure for brand popularity.
Twitter data characteristics – Topic intensities
The Twitter data has been provided by Squaremoon, a company specialized in analyzing the social media culture and bringing brands and the internet closer to each other (Squaremoon 2019). The Twitter data consist of four datasets from three brands; Fashion Nova, Louis Vuitton and Tesla. We employ three brands for comparison purposes. The first dataset contains consumer (non-brand) Tweets for all three brands. These are Tweets posted by consumers. This dataset consist of a 10 percent sample of all consumer Tweets. A sample has been taken to make the data amount manageable, as there is a vast amount of consumer Tweets about these three brands. The other three datasets each contain brand Tweets for one of the three brands. These are Tweets posted by the brands themselves. No sampling has been applied to the brand Tweets, because there are relatively few brand Tweets compared to consumer Tweets. Therefore, all data from these three datasets have been used. The first dataset, containing consumer Tweets from all three brands has been separated into three datasets, each containing consumer Tweets for one brand. This resulted in six datasets.
The six datasets originally contained the variables id, date and time of the Tweet, and the text of the Tweet itself. In addition to this the number of replies, retweets and likes have been included. The included language variable indicated the language of the posted Tweet. We have chosen to only select English Tweets, because text input consisting of one single language is more suitable for applying LDA. In addition to this, brand Twitter accounts often have brand accounts in multiple languages. From this, we can expect that the same topics will occur from the different accounts, regardless of their language. For each brand, all Tweets posted between Sunday 3 September 2017 and Saturday 31 August 2019 have been collected, thereby covering two years of online Twitter data. The date variable has been used to select the Tweets within this period. The modified six datasets now only contain the observations for this two-year period between Sunday 3 September 2017 and Saturday 31 August 2019, each containing English brand or consumer Tweets for one of the three brands.
15 Brand Brand or consumer (non-brand) Tweets Number of English observations/Tweets Average number of Tweets per week Number of weeks with posted Tweets Number of weeks with no LDA input Fashion Nova Fashion Nova Louis Vuitton Louis Vuitton Tesla Tesla Consumer Brand Consumer Brand Consumer Brand 65.699 14.957 47.382 1.541 46.764 961 632 144 456 15 450 9 100 95 80 104 98 95 4 9 24 0 6 9 Table 1. Brand Tweets and consumer Tweets data characteristics.
The three brands mentioned are chosen for several reasons. Firstly, these brands show relatively large fluctuations in search volume compared to other brands (Google Trends 2019a; 2019b; 2019c). The fluctuations are shown in figure 2a, 2b and 2c. We are interested in these fluctuations, because we are interested in fluctuations in brand popularity, which is measured by fluctuations in relative search volume. Secondly, these brands are characterized by a high volume of Tweets (almost on a daily basis). This high volume of Tweets is necessary for performing LDA and especially for calculating topic intensities, as we need enough Tweets for each week to be able to calculate these intensities on a weekly level. Lastly, these brands are among the most searched-for brands (Twitter 2019a; 2019b; 2019c). This means that consumers talk a lot about these brands on Twitter. For example, Fashion Nova and Louis Vuitton ranked first and second in the top 10 of most searched-for fashion brands in 2018, respectively (Google Trends 2018). Because of these relatively large fluctuations in relative search volume and vast amounts of Twitter data, these brands will be suitable to discover latent topics in Tweets and to examine to which extent these topics can predict popularity of brands in general.
The process will be as follows for each of the six datasets separately; The text of the Tweets will be used to extract latent topics. LDA will be applied on each dataset. This means that six LDA analyses will be performed; one for each brand, consumer Tweets and brand Tweets separated. The number of Tweets for each brand (consumer or brand Tweets) will serve as the number of observations for each LDA process. The output of each LDA will result in word groups, which will be interpreted as topics. The topics extracted will consequently be used to calculate topic intensities. Topic intensities will be calculated on a weekly level, by using a week variable (starting at Sunday because the relative search volume per week in Google trends also starts at Sunday) that assigns Tweets to a particular week. The process of calculating topic intensities is explained in the method section. This leads to 104 topic intensities within the two-year period, thereby resulting in enough observations for the VAR model, as the topic intensities are time-varying.
Google Trends data characteristics – Brand popularity measure
16
In addition to this, relative search volume from Google Trends is found to be able to predict stock returns (Bijl et al. 2016) and Leitch and Sherif (2017) found that Twitter sentiment can forecast stock returns. A great, popular brand is often characterized by high returns (Yohn, 2014). If relative search volume can predict returns, it might also be suitable to measure brand popularity. Moreover, if Twitter content can be used to predict returns, Twitter content might also be able to predict brand popularity. It can be concluded that relative search volume data from Google Trends is a suitable proxy for brand popularity. The relative brand search volume can be operationalized as the search volume at time t, relative to the highest search volume of time period t-104 until t-1, with t in weeks. The span of the analysis covers two years. The relative search volume reflects how often the brand is relatively searched for on Google, compared to other weeks within the given time span. A value of 100 is the peak popularity for the search term (see figure 2a, 2b and 2c). A value of 50 means that the search term is half as popular as the peak popularity (Google Trends, 2019d). Topic intensities in Tweets, resulting from LDA, will then be used to examine to which extent these topics can predict the relative search volume on Google at week t, and thus the popularity of the brand. The relative search volumes from Google Trends will be used in the VAR models.
Figure 2a. Search interest for Fashion Nova between 3 September 2017 and 31 August 2019.
Figure 2b. Search interest for Louis Vuitton between 3 September 2017 and 31 August 2019.
17
4 Method
In this research, Latent Dirichlet Allocation (LDA), Vector Auto Regression (VAR) and Impulse Response Functions (IRF) will be used to serve the research purpose. These methods will be explained below.
Latent Dirichlet Allocation
Latent Dirichlet Allocation (LDA) is a common tool for topic extraction in text analysis, and can help to answer research questions with the purpose of summarizing texts or identifying consumer and market trends (Berger et al. 2019). In this research, the goal of performing LDA is to extract topic intensities from Twitter data, for each week, covering a period of two years. This will result in two topic intensity numbers for each week for each of the three brands: one for the Tweets posted by the brand for each of the three brands, and one for the Tweets posted by consumers for each of the three brands. These Tweets by brands and consumers will be separated, because these two types of Tweets might be conceptually different. That is, brands might talk about different topics on Twitter than consumers do, with different intentions, which could lead to different effects on brand popularity. This supposition will be examined by this separation.
Latent topics that are not readily observed in text in online consumer content have already been analysed by using LDA in multiple previous researches (Tirunillai and Tellis 2014; Büschken and Allenby 2018; Netzer et al. 2019). Tirunillai and Tellis (2014) used LDA to extract dimensions of quality and their importance from product reviews. Büschken and Allenby (2018) used LDA to extract topics discussed in online hotel and restaurant reviews and to examine whether these topics could predict hotel and restaurant ratings. Netzer et al. (2019) applied LDA to loan application texts to extract topics discussed in loan requests, to relate these topics to default probabilities. This previous research shows that extracting topics from text with LDA can serve as an important tool for generating marketing insights, as LDA text analysis allows for quantifying information contained in textual data (Berger et al. 2019). In this study LDA will be applied to datasets containing Tweets of three brands, covering two years, to extract topics discussed in these Tweets. We will only select the data which contain English Tweets, to be better able to perform LDA. The model description below is an adapted version of the model description by Büschken and Allenby (2018).
18
dependent on the specific topic. Words with a high probability characterize the topics, which indicates that for each topic there will be other words with a high probability of occurrence.
In other words, the 𝑛𝑡ℎ word that appears in review 𝑑, 𝑤𝑑𝑛, is generated by the process of first choosing a topic 𝑧𝑑𝑛 out of the latent topic pool which has a multinomial distribution (𝜃𝑑). Then, given this chosen
topic, a word 𝑤𝑑𝑛 is chosen (Büschken and Allenby 2018). This word has a certain probability of
occurring, given the topic. This means that a word is chosen from 𝑝(𝑤𝑑𝑛∣ 𝑧𝑑𝑛, Φ), where Φ is a matrix of word-topic probabilities {𝜙𝑚,𝑡} for word m and topic t, which indicates a different probability of
occurrence for all words, for each topic (Büschken and Allenby 2018). This leads to the following formula for the LDA model:
𝑝(𝑤𝑑𝑛= 𝑚 ∣ 𝑧𝑑𝑛= 𝑡, Φ ) = 𝑝(𝑤𝑑𝑛= 𝑚 ∣ 𝜙𝑡),
(4.1) with the probability of choosing word m, given the chosen topic t, out of the matrix of word-topic probabilities, which is equal to the probability of choosing word m out of the specific row with word probabilities belonging to that specific chosen topic. More specific, 𝜙𝑡 indicates the vector of word probabilities for the specific topic t (Büschken and Allenby 2018).
The topics 𝑧𝑑𝑛 and words 𝑤𝑑𝑛 are seen as discrete random variables. Because there are more than two topics and more than two words, a multinomial, discrete distribution is used, which is the dirichlet distribution;
𝑝(𝜃𝑑) ~ 𝐷𝑖𝑟𝑖𝑐ℎ𝑙𝑒𝑡(𝛼),
𝑝(𝜙𝑡) ~ 𝐷𝑖𝑟𝑖𝑐ℎ𝑙𝑒𝑡(𝛽),
(4.2) with the vector of topic probabilities 𝜃𝑑 for a given document and the vector of word probabilities 𝜙𝑡 for a given topic both characterized by a dirichlet distribution.
Co-occurring words that appear within a document indicate the presence of a latent topic (Büschken and Allenby 2018). It is assumed that the latent topics 𝑧𝑑𝑛 can vary from word to word. In this research, one
document consists of all Tweets posted within one specific week t, either posted by the brand, or posted by consumers. From each of these groups of Tweets, several groups of words will appear as output from the LDA. These groups of words will each be interpreted, that is, a latent topic will be allocated to each group of words. It is likely that this latent topic can be linked to one of the topics expected to influence brand popularity, mentioned in the theoretical section. This results in each Tweet week being characterized by its own set of latent topics, for both the Tweets posted by the brand and the Tweets posted by consumers, separately.
19 Vector Auto Regression
To examine changes in the occurrence of topics in Tweets over time, a Vector Auto Regression (VAR) model will be used. A VAR model is a flexible model which is suitable for analyzing time series data (Leeflang et al. 2017). The VAR model is useful when the goal is to explain more than one variable (Leeflang et al. 2017). These variables might be explained by their own historical patterns or by other variables. That is, a vector of time series variables will be regressed on lagged vectors of these variables. In our research, we have multiple topic intensity variables for each brand and Tweet type (consumer or brand Tweet), and a relative search volume variable.
A VAR model can be written in a reduced-form model, where all right-hand-side variables are predetermined at time t (Leeflang et al. 2017). The advantage of this is that the system can be estimated without imposing restrictions (Leeflang et al. 2017). This is useful for calculating impulse response functions. This leads to the following VAR formula:
𝑦𝑡 = 𝑐 + 𝐵1𝑦𝑡−1+ 𝐵2𝑦𝑡−1+ ⋯ + 𝐵𝑝𝑦𝑡−𝑝+ 𝑒𝑡
(4.3)
The 𝑛 × 1 vector of 𝑛 endogenous variables 𝑦 is regressed on constant terms, which may include a deterministic time trend and seasonality terms, which are exogenous variables. The endogenous variables 𝑦 are also regressed on their own past. 𝑝 is the number of lags and 𝐵 the 𝑛 × 𝑛 coefficient matrix of a given lag (Leeflang et al. 2017). Thus, 𝐵 is a matrix with coefficients associated to lag i. 𝑦 implies multiple variables, and each of these variables has several lags. A lag is a past value of a variable. This leads to a model where all variables are both independent and dependent variables: the contemporaneous value of a variable constitutes the dependent variable and the lagged values of all variables in the model constitute the independent variables.
For example, when we run a VAR model with the topic intensities from the two topics that emerged from the consumer Tesla Tweets, sustainability and product innovation, a model with two lags could be represented as a VAR system consisting of the two following equations:
20
research, we have several endogenous variables: the topic intensities per topic and the relative search volume variables (as a proxy for brand popularity). These are endogenous variables, because their values can be determined by other variables within the model. A search peak dummy will serve as exogenous variable, which captures outside forces in our model. This dummy will be further explained in the results section.
Before estimating a VAR model, we have to check for Granger causality, to examine whether the interrelationships between the variables that serve as input for the VAR model are indeed present (Gijsenberg and Verhoef 2019). For each variable it will be checked whether the variable is affected by the history of the other variables in the VAR model. That is, to which extent have the past values of variable X predictive value for variable Y, after correction for the own history of variable Y. To perform the test, the lag order 𝑝 has to be determined. 𝑝 is the number of lags that is taken into account in the computation of the test. More lags implies that a larger history of the variables is taken into account. A significant Granger causality implies that the past 𝑝 values of variable X indeed help in predicting variable Y. The test results are shown in the results section.
We also have to check whether the data is stationary or evolving by performing a Kwiatkowski–Phillips– Schmidt–Shin (KPSS) test (Kwiatkowski et al. 1992). Stationarity is the tendency of a time series to revert back to its deterministic components, such as a fixed mean or a mean and trend (Leeflang et al. 2017). This means that variables have constant statistical properties over time (Gijsenberg and Verhoef 2019). A significant KPSS test implies that the variable is non-stationary or evolving. When a variable is evolving, first-differencing must be applied. The differencing process removes a trend from the data (Leeflang et al. 2017). First-differencing thus implies that the differenced first value of a variable is the outcome of observation 2 minus observation 1. After first-differencing, non-stationarity is removed and these variables become stationary. Tests for stationarity are shown in the results section.
21 Impulse Response Functions
The coefficients resulting from the VAR-model will be interpreted by means of Impulse Response Functions (IRF). Direct interpretation is infeasible, as the number of coefficients is very large. We are interested in the net result of the modeled actions and reactions over time (Leeflang et al. 2017). “IRF simulate the over-time impact of a change (over its baseline) to one variable on the full dynamic system” (Leeflang et al. 2017, p.135). This means that each endogenous variable (in our research the topic intensity variables and the relative search volume variable) is explained by a weighted average of current and past errors, or shocks, to the variable itself and to the other endogenous variables (Leeflang et al. 2017). This shock affects the whole system; a change to a variable can be regarded as a shock to the whole series of the variable. That is, a shock in variable 𝑦1 can affect the future value of 𝑦1, but it can also affect variable 𝑦2 and its future values, or 𝑦3 and its future values, up to variable 𝑦𝑡 and its future values.
An IRF examines the impact of the shock to each other variable in the system, during the shock (period 0) and for each period after the shock (Leeflang et al. 2017). That is, the IRF shows the impact of an one-unit positive change in the impulse variable on the response variable, over the next periods. Immediate effects (effects during the shock), cumulative effects and permanent effects (lasting effects after the shock) can be shown with IRF. The IRF gives the expected level of the shock. This shock level is surrounded by a 95% confidence interval, which implies a low and a high estimate. When the lower bound of the 95% confidence interval is below zero, and the upper bound is above zero in a particular period, this means that there is no significant effect of the impulse variable on the response variable for that period. When the lower bound and upper bound of the 95% confidence interval are both below or both above zero, this means that there is a significant negative or positive effect.
5 Results
LDA AnalysisThe analysis of the Tweets starts with LDA for each of the six datasets (three brands consisting of consumer and brand Tweets). Before performing LDA, the Tweets were preprocessed using several steps. The following steps were applied to all six datasets:
1. Deleting numbers and punctuation;
2. Removing a set of standard English stop words (Feinerer and Hornik 2019); 3. Substituting capital letters with lower-case letters;
4. Stemming, to transform the plural words into words with the same stem;
5. Removing a set of unnecessary words: they, this, just, even, these, will, don’t, much, give, first, also, think, make, now, want, still, never, lot, got, thought, sure, without, whenever, unlike, somehow, yes, tend, today, the, have, day, not, but, and, the, for, from, out;
6. Removing words that are not very informative for the specific brand. This word set differs per dataset (table 2). These words are either the brand name itself, or they are that general that they appear in every topic. These words are therefore not informative for the content of the different topics.
22
like ‘collect’ and ‘collection’. Both results with and without stemming have been examined, but these results did not show this possible disadvantage at a large scale. However, multiple words with the same stem did occur in the LDA results. Therefore, we have chosen to apply stemming.
In addition to these general steps which have been applied to all six datasets, some datasets required additional preprocessing. The three datasets containing brand Tweets for the three brands contained non-ASCII characters and had to be re-imported by specifying UTF-8 encoding. The dataset containing Fashion Nova brand Tweets required additional steps, as these characters appeared multiple times in every Tweet. This led to non-informative codes appearing in the lists of words belonging with high probability to topics. The coding parts were removed from the Tweets. The Fashion Nova brand Tweets dataset also contained many similar web links. These have also been removed from the Tweets.
Brand Brand or consumer
(non-brand) Tweets Removed words Fashion Nova Fashion Nova Louis Vuitton Louis Vuitton Tesla Tesla Consumer Brand Consumer Brand Consumer Brand
fashionnova, @fashionnova, nova,, fashion, nova fashionnova, @fashionnova, nova,, fashion, nova, shop louisvuitton, @louisvuitton, louis, vuitton, loui
louisvuitton, @louisvuitton, louis, vuitton, lvfw, lvmenfw, louisvuitton',, louisvuitton',
car, Tesla, @tesla, tesla, tsla, tsla,, teslamodel, model2, elon, musk, elonmusk
car, Tesla, @tesla, tesla, model
Table 2. Non-informative brand specific words that are removed from the analysis.
Topic number selection
LDA might be complex, as “the interpretation of topics can be challenging” (Berger et al. 2019, p.10) and there is “no clear guidance on the selection of the number of topics” (Berger et al. 2019, p.10) to extract. Therefore, a two-sided approach has been applied for the selection of the number of topics to extract for each dataset. First, a graph with four metrics to determine the optimal topic number has been created, by using the FindTopicsNumber function from the ldatuning package in R (Murzintcev 2019). However, for all six datasets, only one of the four metrics seemed to be informative. That is, the metrics are informative when they either show a maximum or minimum. The topic number where the metric reaches a minimum or maximum is the optimal number of topics. The CaoJuan2009 measure (Juan et al. 2009) was the only measure displaying a minimum. Secondly, with the result of this metric in mind, common sense has been used to interpret which number of topics led to the most distinctive topics. The four metrics used are derived from Griffiths and Steyvers (2004), Juan et al. (2009), Arun et al. (2010) and Deveaud et al. (2014), and are labeled as Griffiths2004, CaoJuan2009, Arun2010 and Deveaud2014 respectively. The optimal number of topics can be found where the CaoJuan2009 (Juan et al. 2009) and Arun2010 are at their minimum (Arun et al. 2010), and where the Griffiths2004 (Griffiths and Steyvers 2004) and Deveaud2014 are at their maximum (Deveaud et al. 2014). For further explanation of these metrics we refer to Griffiths and Steyvers (2004), Juan et al. (2009), Arun et al. (2010) and Deveaud et al. (2014).
23
The resulting graphs for each of the six datasets are shown in figure 3a, 3b, 3c, 3d, 3e and 3f. For all six datasets, Griffiths2004, Arun2010 and Deveaud2014 are not informative, as they do not show a clear minimum or maximum. Therefore, we only inspect the CaoJuan2009 measure. The CaoJuan2009 measure selects the best LDA model based on density. Juan et al. (2009) showed that “the LDA model performs best when the average cosine distance of topics reaches the minimum” (Juan et al. 2009, p. 1780). This explains why the optimal number of topics can be found where the CaoJuan2009 measure is at its minimum. Their method is based on statistics of the whole corpus. This seems suitable for our datasets, as the analysis of topic number determination also applies to a whole corpus in our analysis, where no new documents will appear.
Figure 3a. Topic number measures for Figure 3b. Topic number measures for
consumer Tweets Fashion Nova. consumer Tweets Louis Vuitton.
Figure 3c. Topic number measures for Figure 3d. Topic number measures for
consumer Tweets Tesla. brand Tweets Fashion Nova.
Figure 3e. Topic number measures for Figure 3f. Topic number measures for
24
When applied to our datasets, the CaoJuan2009 measure recommends to extract 2, 3, 4 or 7 topics from the consumer Fashion Nova data, with the minimum lying at 3 topics. For the consumer Louis Vuitton data, 2, 4 or 6 topics are recommended, with the minimum lying at 2 topics. For the consumer Tesla data, 3 or 7 topics are recommended, with the minimum lying at 3 topics. For the brand Fashion Nova data, 3 or 4 topics are recommended, with the minimum lying at 3 topics. For the brand Louis Vuitton data, 7, 11 or 12 topics are recommended, with the minimum lying at 12 topics. Lastly, for the brand Tesla data, 3, 4 or 6 topics are recommended, with the minimum lying at 4 topics. However, because only one of the four measures seems to be informative, and no confirmation for the right topic number from other measures can be found, we will also use common sense to interpret the topic number that gives the most distinctive topics.
The results of this interpretation process are shown in table 3, 4, 5, 6, 7 and 8. The process has been executed by analyzing each topic number between two and eight topics, or more if the CaoJuan2009 measure indicated that more than eight topics might be preferred. For these topic numbers, we have searched for distinctive words, as these really characterize a topic. A certain number of topics was chosen when this topic number led to topics that were clearly distinctive from each other. For each of the six cases, except for the Louis Vuitton brand Tweets, one of the topic numbers indicated by the CaoJuan2009 measure as optimal topic number has been chosen, indicating that the CaoJuan measure was indeed helpful in determining the optimal topic number for each dataset. For the Louis Vuitton brand Tweets, five topics have been chosen, because these number of topics seemed to produce the most distinctive topics. The 7, 11 or 12 topics indicated by CaoJuan2009 produced to many non-distinctive topics. In each table, the top column gives the latent topic assigned to each word group by means of interpretation. Each latent topic belongs to an overall topic. These overall topics are written in brackets. Topic interpretation
For all six datasets (Fashion Nova, Louis Vuitton and Tesla, with brand or consumer Tweets) we performed an LDA analysis. The interpretation of the results from the consumer Fashion Nova Tweets is explained below. In a similar vein we did this for the other five datasets (see appendix A). The resulting subtopics and the corresponding overall topics of all six LDA analyses are shown in table 4. Table 3 displays the top 10 words associated with three topics for the consumer Fashion Nova Tweets. Topic 1 is a description of “Product”, as shown by the use of words such as “dress”, “model”, “jeans”, “design” and “outfit”. Topic 2 describes the sharing aspect of the product and is concerned with words like “Poshmark”, “ship” and “share”. Poshmark is an online platform where clothes can be sold and bought. Topic 3 describes the wearing aspect of the product, displayed by the words “fit”, “girl”, “ass” and “size”. In general, topic 1 “product” and topic 3 “wearing” can be grouped under the broader topic of product quality. Topic 2 ‘sharing’ can be grouped under “service quality”, as Poshmark is a service offered to share clothes. It can be concluded from the analysis that consumers of Fashion Nova mostly talk about aspects of products of Fashion Nova and the service associated with it.
25
Rank Topic 1 “Product”
(product quality) Topic 2 “Sharing” (service quality) Topic 3 “Wearing” (product quality) 1 2 3 4 5 6 7 8 9 10 Love Dress Buy Item Model Jeans Design Outfit Video Fit Cloth Buy Style Time Poshmark Love Jeans Ship Dress Share Ship Check Dress Jeans Love Cloth Fit Girl Ass Size
Table 3. Top 10 words from LDA analysis for consumer (non-brand) Fashion Nova Tweets.
Brand Brand or consumer
(non-brand) Tweets
Overall topic Sub topic within overall topic
Fashion Nova Fashion Nova Louis Vuitton Louis Vuitton Tesla Tesla Consumer Brand Consumer Brand Consumer Brand Product quality Service quality Product quality Service quality Product innovation Celebrity endorsement Service quality Consumer satisfaction Product innovation Sustainability Product quality Celebrity endorsement Sustainability Product innovation Product quality Service quality Product Wearing Sharing Sporty image Women experience Lovely image Win action Distinctive bag Fashion art show Celebrity fashion
Male celebrity endorsement Win action
Online luxury bag promotion
Spring summer collection Charity Bag Exploring image Perfume campaign Electric driving Innovation New car model
Performance Power Contact
26 Topic intensities
After grouping the word groups into topics, topic intensities are calculated. These topic intensities are calculated per week, by using the week variable in the datasets. As mentioned before, topic intensities consist of the number of weekly occurrence of top 10 words belonging to one topic, divided by the total number of words occurring in Tweets of that specific week. This will be done for each brand, with brand posts and consumer posts separated. With one topic, one overall topic is mentioned. This means that the words belonging to subtopics with the same overall topic, are grouped together to calculate one overall topic intensity, which could lead to more than 10 words belonging to one topic. To illustrate this, the subtopic “product” and “wearing” from consumer Fashion Nova Tweets both belong to the overall topic “product quality”. Therefore, all words of these two topics will be used together to calculate the topic intensity of product quality. Words that belong to more than one subtopic within a single overall topic are only counted once for the topic intensity. To illustrate this, subtopics “product” and “wearing” both contain the word “love”, but the count of “love” is not doubled. That is, duplicate words are removed. The total number of words occurring in one Tweet week consists of all words, that is, duplicate words are not removed from this word total. The data used for topic intensity calculation, that is, for counting the topic words and the total number of words, is the original, uncleaned data, that has not been prepared already for LDA. The topic intensities for each week are calculated for each dataset by using a for loop. The topic intensities and relative search volume variables for the same brand (three brands in total, consumer and brand Tweets separated) are combined into one dataset, which will serve as input for the VAR model.
VAR Analysis
After calculating topic intensities using the topics from the LDA analysis, the topic intensities have been used to estimate multiple VAR models. The VARS package in R has been used to estimate these models (Pfaff 2008).
First, data preparation has been applied to make the data suitable for VAR analysis. First, topic intensities of topics belonging to the same brand and Tweet type (for example all four topics from consumer Louis Vuitton Tweets) have been combined into one dataset. Rows have been added in the datasets for the weeks with missing topic intensities. For these weeks, a row is added in the data with zero values. This means that for these weeks, the topic intensities are set to zero. This seems logical, because having no Tweets implies no chatter about the topic, resulting in a topic intensity of zero for that week. This way, all topic intensity sets now contain 104 weeks, with a zero value for weeks without Tweets. The extension of the topic intensity data with zero values is necessary, as the relative search volume data from Google Trends with 104 weeks of data must be attached to the topic intensity datasets. This way, the topic intensities in a particular week will match with the relative search volume of that week.
Granger causality tests
27
Vuitton data, some of the topic intensity variable combinations (lag order 5) and some of the Louis Vuitton relative search volume combinations (lag order 1) showed a significant Granger causality. With regard to the brand Louis Vuitton data, some topic intensity variable combinations showed a significant Granger causality (with multiple lag orders). The Louis Vuitton relative search volume variable combinations did not show a significant Granger causality. With regard to the consumer Tesla data and the brand Tesla data, both topic intensity variable combinations and Tesla relative search volume variable combinations did show a significant Granger causality for some of the combinations. It can be concluded from this that for five out of six models, estimating a VAR model seems suitable. Only the brand Fashion Nova data did not show any Granger causality. If there are indeed no interrelations, as this test proposes for this dataset, we could expect that the brand Fashion Nova VAR model will probably not lead to significant results.
Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test for stationarity
To test whether the time series of variables in our datasets are stationary, we performed multiple KPSS tests. The null hypotheses is that the time series are stationary. The alternative hypothesis is that the time series are not stationary and thus evolving. Table 5 shows for each dataset which variables are stationary and which are non-stationary at the 10% level (each dataset will lead to one VAR model).
Dataset Variable p-value Stationary or evolving
Consumer Fashion Nova Tweets
Brand Fashion Nova Tweets
Consumer Louis Vuitton Tweets
Brand Louis Vuitton Tweets
Consumer Tesla Tweets
Brand Tesla Tweets
Relative search volume Topic intensity product quality Topic intensity service quality
Relative search volume Topic intensity product quality Topic intensity service quality
Relative search volume
Topic intensity product innovation Topic intensity celebrity endorsement Topic intensity service quality Topic intensity consumer satisfaction
Relative search volume
Topic intensity product innovation Topic intensity sustainability Topic intensity product quality Topic intensity celebrity endorsement
Relative search volume Topic intensity sustainability Topic product innovation
Relative search volume Topic intensity product quality Topic intensity service quality
p < 0.01* p > 0.1 p > 0.1 p < 0.01* p > 0.1 p > 0.1 p > 0.1 p < 0.01* p = 0.082. p < 0.01* p < 0.01* p > 0.1 p > 0.1 p > 0.1 p > 0.1 p > 0.1 p = 0.029. p = 0.017. p > 0.1 p = 0.029. p = 0.054. p < 0.01* Evolving Stationary Stationary Evolving Stationary Stationary Stationary Evolving Evolving Evolving Evolving Stationary Stationary Stationary Stationary Stationary Evolving Evolving Stationary Evolving Evolving Evolving
28
11 out of 22 variables are evolving at the 10% level. In reality, 9 variables are evolving, as the relative search volume variables for both Fashion Nova datasets and both Tesla datasets are the same variables. These non-stationary variables had to be made stationary, as non-stationary time series could lead to incorrect estimates in the VAR model. We therefore applied first-differencing to these 9 time series to make them stationary. This implies that all models which contain one or more first-differenced variables lose the first observation. As this is the case for all models, except the brand Louis Vuitton Tweets, all models are left with 103 observations per variable. The brand Louis Vuitton variables still have 104 observations per variable. After applying first-differencing, the KPSS tests have been performed again to check whether the variables are now indeed stationary. The KPSS test confirmed stationarity. After first-differencing the non-stationary variables, we determined whether we should only include a constant in our VAR model or a constant and a trend. These two options have been combined with a dataset without a dummy for peaks in relative search volume and with a dataset with a dummy for peaks in relative search volume. We chose to include an intercept (constant) in the model as the independent variables (topic intensities of lagged variables) could become zero. Therefore, the intercept has an intrinsic meaning: the value of the dependent variable when one of the independent variables is equal to zero. The option to use a dummy might be useful, as we want to explain variation in the ‘regular’ data pattern, and not in the peaks. These search peaks might distort our results. With regard to the search peak dummy, figure 2c shows that in week 11 (12 November 2017), 12 (19 November 2017) and 23 (4 February 2018) there are peaks in Tesla search volume. For these three peaks, a dummy is included in the VAR model for the consumer Tesla Tweets. To test which of the four combinations gave us the best model, we specified the four models and chose the model which returned the lowest BIC. The results showed that the model specification that included a constant and a search peak dummy variable resulted in the lowest BIC value. Therefore, the six VAR models have been estimated by including the search peak dummy and the constant only, resulting in VARX models. That is, the models now contain an exogenous variable: the search peak dummy which does affect the model, but is not affected by the model. An example of the examination of BIC values for the four model specifications for the consumer Tesla Tweets is given in table 6.
Model specification Search peak dummy included BIC
Constant
Constant and Trend
Constant
Constant and Trend
No No Yes Yes -676.8876 -664.3146 -840.8221 -827.9243
Table 6. BIC values for four different model specifications (lowest BIC in bold).
29
collaboration between Fashion Nova and celebrity Cardi B started. With regard to Louis Vuitton, figure 2b shows search peaks in week 15 (10 December 2017), 16 (17 December 2017), 17 (24 December 2017), 68 (16 December 2018), 69 (23 December 2018) and 74 (27 January 2019). For these six peaks, a dummy is included. The search peaks in December are probably due to the Christmas season. In January 2019, Louis Vuitton launched a new collection, in which Michael Jackson was honored. All these aforementioned special events and seasonal effects are captured in the search peak dummy. VARX estimations
First, for each brand (Fashion Nova, Louis Vuitton and Tesla) and Tweet type (consumer and brand Tweets) the optimal number of lags has been determined by using the BIC. This is necessary for determining the optimal VARX model. The maximum lag length for which the BIC has been calculated has been set to 10, as this is already a lag length which results in a model with too many parameters, making the model inefficient. This way, VAR models with 1 until 10 lags have been estimated and R determined the most favorable model by comparing their BIC values. For all six models the BIC showed the lowest value for a model with a lag number of 1. This optimal lag number of 1 is efficient, as it results in less parameters to estimate and thus more degrees of freedom. Namely, a downside of a VAR model is the risk of over parameterization. Each included lag adds multiple parameters that have to be estimated. This increases when more endogenous variables are included (Gijsenberg and Verhoef 2019). Therefore, all VARX models will be estimated with 1 lag, resulting in multiple parameter estimates. IRF interpretations
To interpret the parameter values, we will use IRF. That is, we do not interpret all parameters individually. The first reason for this is that there is a vast amount of parameter values. The second reason is that all variables in a VAR model are interdependent. A single parameter therefore provides little information on the reaction of the model on a shock. An IRF gives the reaction of a variable to a shock. The variable that causes the shock is the impulse variable, and the reacting variable is the response variable. The reaction of the response variable will be shown in a plot. As giving a shock to all variables in all models is not feasible, we only give a shock to the topic intensities of each topic, and check how the relative search volume variable of the corresponding brand responds (which is the measure for brand popularity). That means that we do not examine impulse-response relations between multiple topics, only between topics and search volume. As we have 16 topics in total, this leads to 16 IRFs. In the IRFs, we set ‘ortho’ to true. This makes it possible for the response function to not start at zero in period 0, indicating that a direct effect of the shock is possible. The resulting IRFs are shown in figure 4a, 4b, 4c, 4d, 4e, 4f, 4g, 4h, 4i, 4j, 4k, 4l, 4m, 4n, 4o and 4p. The periods are represented on the X-axis. The amount of change in relative search volume (response variable) caused by a unit impulse/shock from the impulse variable is represented on the Y-axis.
