The Role of Online WoM Sentiment and Search Volume for Brands

The Role of Online WoM Sentiment and

Search Volume for Brands

The Impact on Monetary Brand Equity

Master Thesis

Department of Marketing University of Groningen

Laura Schum

– S3953424 –

Phone: +49 172 7939266 Email: l.schum@student.rug.nl

Supervisor: Dr. Evert de Haan Second Supervisor: Dr. Arnd Vomberg

Abstract

With the rise of user-generated content on social media platforms, big data is becoming increasingly qualitative in nature: Users share actual outspoken opinions on platforms such as Twitter – ranging from both positive, negative, and neutral experiences to criticism about topics, products, firms, and brands. As a ‘hot topic’ nowadays, the impact of online word-of-mouth on brand equity has been investigated with a primary focus on customer-based brand equity. The impact on monetary brand value – as published yearly by renowned rankings such as Interbrand and Forbes – has been less of a subject. Therefore, this study aimed at identifying the impact of online word-of-mouth, i.e., positive and negative Twitter sentiment as well as brand popularity in terms of search volume, on financially based brand equity. Further, differences of effects are investigated for company types behind the brand, i.e., B2C, B2B, or a mix of both. Tweets were extracted for 53 brands across a period of ten years. Further, brand popularity data using Google Trends as well as brand value data for all considered brands across four well-established rankings were collected. With sentiment analysis, tweets were analyzed for polarities, after which panel regressions were conducted. I found that the presence and strength of both the sentiment and popularity effects differ to a certain extent across rankings. Further, the impacts are distinct across company types as well as depending on whether absolute or relative brand value is of interest. Positive online WoM appears to impact absolute brand value mainly, while negative online WoM does not have a substantial impact in most cases – possibly implying that strong brands are protected from negative chatter by a strong name as well as brand supporters already. The lacking effect of sentiment for relative brand value may be due to spillover effects within industries – thus not necessarily causing differences in brand value. Search volume impacts both absolute and relative brand values similarly to sentiment. Interestingly, most effects are in favor of B2B and mix brands. The findings extend the existing body of research as to the impact of online WoM on brand value depending on the company type and bring implications for marketing managers in two ways: attracting investors and improving its department’s internal standing in the organization.

Table of Contents

1 Introduction ... 1

2 Theoretical Framework ... 3

2.1 Literature Review ... 4

2.2 Conceptual Model and Hypotheses ... 8

3 Research Design ... 12

3.1 Data ... 12

3.2 Methodology ... 17

3.3 Panel Regression Model Specification ... 18

3.4 Descriptive Statistics and Model-Free Evidence ... 20

4 Results of the Analysis ... 26

4.1 Panel Model Diagnostics ... 26

4.2 Results of the Panel Regression on Absolute Brand Value ... 29

4.2.1 Main Effects ... 29

4.2.2 Interaction Effects of Sentiment and Search Volume ... 32

4.3 Results of the Panel Regression on Relative Brand Value... 35

4.3.1 Main Effects ... 35

4.3.2 Interaction Effects of Sentiment and Search Volume ... 37

5 Discussion ... 40

5.1 Theoretical Implications and Further Research ... 40

5.2 Managerial Implications and Recommendations ... 46

5.3 Limitations ... 49

References ... 51

1

1 INTRODUCTION

‘Information is the oil of the 21st century, and analytics is the combustion engine.’ (Peter Sondergaard – Senior Vice President, Gartner)

With the rapid growth of big data in the last decades, increasingly more attention is drawn to how to transform these into information that is relevant and valuable for decision-making processes. Since big data growingly consist of both internal and external unstructured data (Verhoef, Kooge, & Walk 2016), it is even more crucial for companies to make use of sophisticated analytical skills and techniques to derive insights that support in creating a sustainable competitive advantage. Considered as online word-of-mouth (WoM) with currently around 3.8 billion active users (Cooper, 2020), one external source of unstructured data is social media: A channel used by both consumers and companies that enables firms to reach, engage and communicate with (potential) customers (Cooper, 2020) while eliminating communication constraints about time, place and channels (Kim & Ko, 2012). It further allows companies to easily extract actual outspoken consumer opinions about them, their products, and brands (De Haan, 2020).

3

1. How do online WoM sentiment and search volume influence monetary brand equity? 2. To what extent does this influence differ across company types?

For the purpose of this study, brand value data of 53 brands in 15 industries for the period 2010 - 2019 were collected. With the existence of several brand value ranking providers and their distinct methodologies, the effects were investigated for a selection, i.e., for only those brands present in all rankings. Further, 7.6 million tweets in total, as well as Google Trends data for each of the brands for the respective period were extracted and matched with the brand value data. In my empirical application, I analyzed the extent to which monetary brand value is influenced by online WoM’s sentiment and (search) volume employing panel regression analyses. Further, effects across company types were compared. I show that the ranking’s methodologies differ to a certain extent as to the impact of both online WoM sentiment and search volume on brand value. Further, the results deviate depending on whether absolute or relative brand value is considered, thus focusing on either individual or competitive settings. Against expectations, it is mostly positive online WoM sentiment that has an impact on absolute brand value, while negative WoM seems not to affect brand value too much. Relative brand value, on the other hand, is only weakly affected by negative social media chatter. Instead, search volume clearly determines a brand’s value in nearly all cases. Further, online WoM sentiment and search volume do not always go hand in hand but differ depending on a brand’s business orientation, i.e., being a B2C, B2B, or mix brand. The rest of this thesis is structured as follows: First, the literature review provides background information on the importance of unstructured data, in particular online WoM, and insights into the existing body of research regarding online WoM and brand equity. Based on this, the conceptual model of this study is elaborated on. Next, the research design, data sources and collection, as well as methodology are discussed. Afterward, panel regression results are explained, followed by discussing findings, academic and managerial implications and limitations.

2 THEORETICAL FRAMEWORK

4 combined construct of the effect of online WoM on monetary brand equity exists (Lim et al., 2020), both research streams are discussed independently. After, the conceptual model used is presented.

2.1 Literature Review

The potential of unstructured customer insights

In the last decades, companies realized that generating customer insights is a crucial aspect for decision-making in marketing in order to drive profitable growth and outperform the competition in the long run (Gupta & Zeithaml, 2006; Osborne & Ballantyne, 2012). Data usage, as part of organized collection and analysis processes, is referred to as database marketing (Blattberg, Kim & Neslin, 2008). However, databases contain increasingly more unstructured data, which are projected to make up 80% of business data (Schneider, 2016). Such data are strongly linked to the

3V model, which describes the nature of big data by the large volume, velocity, and variety.

Potential sources of unstructured data are customer reviews, surveys, complaints, customer feedback, as well as online WoM. In fact, an IDC Digital Universe study in 2012 revealed that only 3% of potentially useful data worldwide is tagged while even less (1%) is analyzed (Burn-Murdoch, 2012). This shows that experts view novel challenges when it comes to investigating unstructured data while ascribing immense untapped potential (Chen, Mao, & Liu, 2014). Research in many fields already proved the benefits of using unstructured data: Netzer, Lemaire, and Herzenstein (2019) demonstrated how text information improves model performance for loan default prediction. However, since this data type is qualitative in nature and not directly suited for analyses, it requires structuring and pre-processing before existing analysis techniques can be used (Verhoef et al., 2016). This effort requires firms to develop skills for effective data management and analysis to gain potentially valuable knowledge generated from beyond traditional data. The role of online word-of-mouth for companies

networking-5 based communication channel among consumers used to exchange their experiences with products and services (Griffin & Hauser, 1993; Van den Bulte & Lilien, 2001). Hennig-Thurau, Gwinner, Walsh and Gremler (2004) showed that online compared to offline WoM is more powerful due to the immediacy, reach, and high accessibility. Cheung and Lee (2012) investigated that consumers with a high level of satisfaction are more prone to share their positive experiences online compared to traditional WoM, while the same applies in case of negative experiences and dissatisfaction (Lee & Cude, 2012). As a large part of online WoM, social media is used both as a tool to market goods and as a means to communicate with consumers. It has become an integral part of the marketing mix in the online context (Cheung, Lee & Rabjohn, 2008). Next to review websites and blogs, consumers increasingly use Facebook and Twitter to share their experiences (He, Tian, Hung, Akula & Zhang, 2018). Further, Edison Research (2014) revealed that one-third of all social media users follow brands, while increasingly using Facebook and Twitter for complaints (Goodman & Dekay, 2012; Einwiller & Steilen, 2015).

6 Based on this, it is reasonable to assume that customer satisfaction and advocacy expressed as WoM have a positive impact on financial metrics as well. In terms of customer behavior, Menichelli and De Haan (2020) investigated that customer feedback metrics (e.g., written and verbal opinions) enhance the predictive value of customer churn behavior beyond traditional data gathered from customer databases. De Haan, Verhoef, and Wiesel (2015) further demonstrated that mindset metrics allow for precise prediction of customer retention compared to using data on customer relationship length and demographics only. Findings based on online WoM are thus increasingly used to predict both future firm performance and customer behavior (De Haan, 2020). Online word-of-mouth’s influence on brands

7 towards brands (Chatterjee, 2001). Research further showed that sentiment analysis is useful to assess brand popularity (Ghiassi, Skinner & Zimbra, 2013; Mostafa, 2013), and using Twitter is a best practice approach for brand establishment and popularity increase (Aggrawal, Ahluwalia, Khurana & Arora, 2017). Notably, brand equity is improved using a celebrity endorser (Till, 1998), which is especially relevant now in times of influencer marketing (Chung & Cho, 2017).

The potential of online WoM for brand equity

Brand equity, compared to other brand indicators, is increasingly in the spotlight since several studies proved it to foster dual-value creation, for both customers and firms (García-Osma, Villaseñor, & Yagüe, 2015; Keller, 2013). This highlights the importance of long-term strategic thinking as part of brand management (Wood, 2000). For instance, brand equity is positively related to firm performance, profitability, and return on investment (Aaker & Jacobson, 2001; Wang & Sengupta, 2016; Felício, Duarte, Caldeirinha & Rodrigues, 2014). Hence, it is not only a strategic outcome of branding (Wang & Sengupta, 2016) but also deemed to be vital as it allows for long-term performance measurement, which is especially supportive of the marketing function. This is the case as marketing activities increasingly aim at strategically developing and maintaining brand image, reputation, and market share (Motameni & Shahrokhi, 1998) while being exposed to the consistent need to justify expenses (O’Sullivan & Abela, 2007). External perspectives, on the other hand, need to be considered as well: Brand equity is also of high importance to investors and shareholders who are interested in its economic significance, i.e., the effect on the company’s balance sheet (Cobb-Walgren, Ruble & Donthu, 1995). For instance, Murphy (1990) early identified that brand value is of great importance when it comes to company takeovers. Moreover, since marketing decisions affect a company’s stock prices (Simon & Sullivan, 1993), it is assumed that the publication of brand value rankings affects a company’s stock performance as well. Overall, brand equity as a metric is considered to be of interest as it enables the quantification of the brand’s intangible value, thus allowing stakeholders, such as brand managers and investors, to assess and compare brands and improve decision-making (Millward Brown, 2019).

8 Shahrokhi 1998): the marketing perspective and the financial perspective. The marketing perspective views brand equity as ‘a set of brand assets and liabilities linked to a brand, its name, and symbol, that adds to or detracts from the value provided by a product or service to a firm and/or to the firm’s customers’ (Aaker 1991, p. 16), thus influencing the consumer response to the brand’s marketing (Keller, 2013). The financial perspective, on the other hand, considers the brand’s monetary value by regarding the incremental discounted future cash flows resulting from branded product revenue as compared to non-brand revenue (Simon & Sullivan, 1993).

Currently, there is only limited research available when it comes to the connection of (online) WoM and brand equity: Previous studies revealed an impact of social media on determining a company’s brand equity (Baehr & Alex-Brown, 2010; Chang & Chuang, 2011). However, the existing body of research shows a strong focus on CBBE. For instance, Bambauer-Sachse and Mangold (2011) investigated that negative online WoM has a detrimental effect on CBBE. Severi et al. (2014) revealed an interdependency of online WoM and CBBE. Moreover, Schivinski and Dabrowski (2015) identified effects of social media chatter on different elements of brand equity, e.g., brand awareness, loyalty, and perceived quality. Menichelli and De Haan (2020) further pointed out that online WoM can be used to identify customer attitudes over time, hence leading to a better understanding of how consumers view a brand. To the best of my knowledge, there is only one study published by Lim et al. (2020) that investigated the impact of social media on brand equity from the financial point of view. In particular, the authors examined the extent to which activities of several social media types, e.g., the number of followers, comments, and likes, affect brand equity as well as how these differ across industry types. However, these activities are primarily firm-driven, thus not representing regular online WoM, which is initiated by consumers (Chatterjee, 2001). Since online WoM is increasingly being used by popular brand value ranking providers (see 3.1 Data), insights generated through social media chatter may support companies in understanding how their brand equity is derived and determine the extent to which it can be used to manage online WoM actively.

2.2 Conceptual Model and Hypotheses

Despite the previously mentioned studies, the influence of social media chatters’ elements

9 Further, previous studies primarily took CBBE into account, which is based on consumer attitudes and opinions but does not provide a quantifiable measurement in itself. Further, these studies have considered either specific brands or types of industries without providing detailed implications for company types. These gaps should be addressed in this study.

With this theoretical basis, the conceptual model of this thesis’ study is presented in Figure 1. The model consists of one dependent (DV), one independent (IV), and two moderator variables. As discussed, the impact of online WoM on brand equity is investigated. For this purpose, online WoM is split into the critical components sentiment and (search) volume, while sentiment represents the IV, which is expected to be moderated by both (search) volume and company type. The impact of these factors is investigated for monetary brand equity, which represents the DV.

Figure 1: Conceptual model – online WoM and brand equity

Dependent variable: Monetary brand equity

10 Independent variable: Online WoM sentiment

Several studies revealed that consumers’ expressed feelings, attitudes, and perceptions have an impact on brand indicators. Saif (2017) identified sentiment analysis of comments expressed in tweets as an integral part of monitoring a brand’s reputation. Aaker and Jacobson (2001) explained that a brand that can create favorable associations could improve its brand equity, while Sen and Lerman (2007) proved that the valence of online reviews significantly impacts consumers’ attitudes towards a product. Positive feelings towards a brand do not only encourage consumers to stay loyal but also to engage in WoM (Kotler & Armstrong, 2010; Keller, 2013). On the other hand, Zarantonello, Romani, Grappi and Bagozzi (2016) found that once consumers develop a feeling of hatred towards a brand, its value decreases. Chatterjee (2001) stated that negative online reviews have a more negative effect on perceived reliability and purchasing intentions for less known compared to well-known retailers. This can also be applied to strong and weak brands. Further, Bambauer-Sachse and Mangold (2011) demonstrated that negative online WoM has a brand dilution effect. Lastly, De Haan (2020) investigated that changes in electronic WoM sentiment, especially negative sentiment, are significant predictors of firm performance. Based on these findings, it is assumed that online WoM sentiment affects brand equity as well. With the asymmetry of effects, past research revealed that dissatisfied customers engage more heavily in WoM than satisfied customers (Anderson, 1998). Thus, brand equity could be more heavily affected by negative as compared to positive WoM. The hypotheses are as follows:

H₁: Positive online WoM has a positive effect on monetary brand equity. H₂: Negative online WoM has a negative effect on monetary brand equity.

H₃: The effect of negative online WoM is stronger than the effect of positive online WoM on monetary brand equity.

Moderator: Online WoM (search) volume

11 online WoM volume is associated with brand awareness: The more consumers are confronted with WoM on or are searching for a specific brand online, the more aware and the more likely they are to recognize and recall a brand (Keller, 2013). Yang, Kim, Amblee and Jeong (2012) further point out the credibility aspect of WoM volume: Increased WoM conveys information about the number of people who connect certain emotions or experiences with a brand, thus reducing uncertainty about attitude formation of others. Nevertheless, the volume itself does not prove any valence: In the case of negative online WoM, a high volume may negatively affect a brand’s reputation, and hence equity, e.g., in case of a publicity crisis. On the other hand, an increased volume of positive online WoM may strengthen the brand’s equity, e.g., in case of a successful campaign. Thus, it is hypothesized that online WoM volume moderates the relationship of sentiment and brand equity:

H₄: The effect of online WoM sentiment on monetary brand equity is stronger with a higher online WoM volume.

Moderator: Company type

12 which end-consumers are less expressive. Nevertheless, the expected lower level of online WoM for B2B brands might also imply a greater potential and sensitivity for these brands. Therefore, there is no clear direction expected. The hypotheses are:

H₅: The effect of online WoM sentiment on monetary brand equity is moderated by the company type. H₆: The effect of online WoM volume on monetary brand equity is moderated by the company type.

I aim at addressing marketers by offering insights on how online WoM affects a brand’s monetary brand equity across company types. Furthermore, this information serves as the basis for how to leverage brands and community management on social media in order to manipulate brand value. The results should further advise to which extent online WoM information should be incorporated to guide investment strategies for brand management.

3 RESEARCH DESIGN

The study at hand was designed to determine the effect of online WoM sentiment on monetary brand equity, which is assumed to be moderated by both online WoM volume and the company type. For this purpose, a quantitative approach was used to manage qualitative data: Online WoM sentiment was quantified utilizing text analytics. In the following, the three data sources, brand

rankings, social media posts, and Google Trends are described. Afterward, the data collection

process of Twitter scraping and the sentiment analysis is elaborated on, followed by the explanation of the panel regression analyses to determine the relationship between the IVs and brand equity. Eventually, a descriptive analysis provides initial model-free evidence.

3.1 Data

Brand value rankings

For the measurement of monetary brand equity, publicly available data provided by several institutions specialized in brand valuation are used. Obtaining data from www.rankingthebrands.com, the following rankings are considered to be the most popular: The

13 research by Lim et al. (2020), I decided against choosing only one brand ranking provider due to the popularity of numerous rankings nowadays: Some companies and investors might be more likely to consider one ranking over another due to different firms, methodologies, and distinct data sources. Further, neither a combination nor a merge of rankings is used due to substantially different methodologies and thus potentially high-value ranges, which might lead to a distorted picture of a brand’s value. A summary of the brand rankings is shown in Appendix A, Table 11.

Forbes’s methodology is based on an evaluation of more than 200 brands, while primarily focusing

on a brand’s financial strength (Badenhausen, 2018). After determining earnings before interest and taxes for each brand and averaging these over the previous three years, a charge of 8% of the capital employed for the specific brand is subtracted, i.e., assuming that a brand should be able to generate at least 8% on earnings from that capital. Next, the maximum corporate tax rate in the parent company’s domestic country is applied while allocating a percentage of earnings to the brand depending on how important brands are considered for this industry. However, Forbes requires brands to have a strong presence in the U.S., which is also incorporated in other ranking methodologies, yet rather one-sided as it neglects the economic power of other worldwide regions. Another factor to consider is the fact that Forbes calculates an index by weighting the financial value of the brand more than consumer perceptions. This may be considered a biased approach as brand equity is not only based on financial success but especially on consumer attitudes.

Brand Finance, on the other hand, computes the brand value based on the so-called ‘Royalty Relief

14 to its competitors, it shows a strong emphasis on the royalty rate which is most interesting to tax authorities and courts. Also, in this case, the consumer’s perspective is rather minimally integrated. Comparing the rankings of Millward Brown (BrandZ) and Interbrand, there are many parallels visible: Both approaches consider a holistic view of the brand through assessing globally, evaluating their present and future potential, and incorporating both the financial aspects as well as consumer perceptions. Millward Brown calculates the brand value by adhering to both financial value and brand contribution (Millward Brown, 2019). Using corporate earnings, the proportion of earnings derived from the firm’s intangible assets is determined and applied by the metric ‘intangible ratio’. Further, an attribution rate is applied to the intangible earnings to identify the proportion of overall earnings attributable to the brand. Lastly, considering the projected earnings of the brand, the branded intangible earnings are multiplied by the metric ‘brand multiple’ which provides the brand’s financial value. The brand contribution then quantifies the brand’s relative strength compared to the competition by eliminating the financial contribution. For this, a brand equity model is used to quantify current demand, price premium, and future demand and price.

Interbrand, on the other hand, considers three critical factors for the brand valuation (Interbrand

15 Choice of brands

Each brand that is present in at least one of the four rankings was considered. Brand value information is available for all brands in all years (2010 – 2019) for the different rankings, apart from Forbes for which data is available for the period 2013 – 2019 only. Due to the nature of being popular brands, the criterion is met that all companies are active on social media. This ensures a high volume of Twitter posts. I made use of the predetermined industry categories by Interbrand, but moved brands of the industries alcohol, beverages, and sporting goods into other industry categories due to a very small number of observations. Further, I added telecommunications as an additional separate industry initially due to the high number of brands operating in this sector. However, none of these brands were part of all rankings, thus, eliminating this new industry category completely. With that, I ended up with 14 industries totally: apparel, automotive, business

services, diversified, electronics, energy, financial services, FMCG, logistics, luxury, media, restaurants, retail, and technology. These industries were applied to brands of other rankings as

well based on the respective ranking’s industry classification (if available) and subjective evaluation. Further, each brand was classified into B2C, B2B, and mix brands based on Wikipedia’s industry and product categorization and own judgment. It is to be noted that in cases in which brands changed their name in the considered period, only the most recent brand name was chosen and applied consistently to the years and rankings. This was necessary for Citibank (previously Citigroup/Citi), Dell (previously Dell Technologies/Dell Enterprises/Dell EMC), HP (previously Hewlett Packard Enterprises) as well as T-Mobile (previously Deutsche Telekom). Thus, a total of 53 brands are considered. Appendix A, Table 12 shows the full list of brands. For the operationalization of brand equity, transformations are applied to obtain more detailed information on the effects of online WoM on different perspectives of monetary brand equity:

1. Absolute brand value: actual monetary value of brand x (in million $) in year t

2. Brand value change: %-change in monetary brand x value in year t compared to year t-1 3. Relative brand value: monetary value of brand x in year t compared to the sum of the

monetary values of other brands in the same industry in year t Online WoM / social media data

16 However, with the rise of social media, this study focuses on social media mainly. Nevertheless, social media data are not unrestrictedly available due to data and consumer privacy concerns. Therefore, I made use of Twitter data, also called ‘tweets’, which are either consumer- or firm-generated. These tweets are posts of up to 280 characters long, which can be sent to one’s online network (Russel & Klassen, 2018). Twitter is becoming increasingly important due to its prominent role in activism, disaster recovery and elections (Murthy, 2015), and enjoys a user base of currently around 330 million monthly active users worldwide (Clement, 2020). Compared to other social media platforms, tweets are infinitely available – as long as both the platform and the author do not delete them -, can be easily extracted using scraping tools and show a high activity rate of companies (De Haan, 2020). Further, Twitter is known as a suitable tool for text mining in order to obtain insights on consumer opinions, attitudes, and behavior (Fader & Winer, 2012; Culotta & Cutler, 2016; Berger et al., 2019). Online content can impact a customer’s attitude towards the brand (Berger et al., 2019). Thus, tweets are used to investigate online WoM sentiment. Google Trends

Due to tweet extraction restrictions using Twitterscraper (see 3.2 Methodology), tweet volume per brand cannot unconditionally be regarded as a proxy for popularity. Hence, Google Trends is used as an alternative as it collects immense amounts of data on Internet searches globally and is even able to predict a company’s stock returns (Bijl, Kringhaug, Molnár & Sandvik, 2016). Showing relative search volume about brands on Google, it is a proxy of popularity: A high search volume indicates high popularity. Since the analysis covers ten years (2010 – 2019), the Google Trend score at year t is viewed in relation to the highest search volume for the brand within the total period. A score of 100 reflects the peak popularity, while it is to be noted that Google Trends provides figures per month, thus necessitating to take the average to result in a yearly figure. Figure

2 shows an example of the global Google Trend popularity of ‘Amazon’ in the period 2010 – 2019.

17 3.2 Methodology

Twitter scraping

Based on the programming language Python, the tool TwitterScraper is used to extract tweets (available at https://github.com/taspinar/twitterscraper). Compared to previous text mining packages, TwitterScraper does not create barriers regarding the availability of past tweets. With this, it is possible to collect tweets per brand and year. It is to be noted, however, that a maximum of approx. 14,500 tweets per request can be obtained. Therefore, numerous requests are run per brand for each year, resulting in a total of 530 requests and approx. 7.6 million tweets. Only English tweets are considered due to the simplicity of analysis. Nevertheless, TwitterScraper does not cover retweets, i.e., tweets which have been shared by other users.

Sentiment analysis

18 Panel regression analysis

For the investigation of the actual extent to which online WoM sentiment and (search) volume influence monetary brand equity over time, a panel regression analysis is used. Panel models combine cross-section and time-series data, where ‘a fixed sample of population elements is measured repeatedly on the same variables’ (Malhotra, 2010, p. 110). In this study, the same cross-sectional units, i.e., brands, are observed over ten years concerning multiple variables, covering yearly brand value data from several ranking providers, sentiment scores, and online search volume. A complete list of variables is found in Appendix B, Table 13. By considering panel data, changes on the individual brand level are detected, and a large amount of data can be collected, which allows for high variability and improved statistical efficiency (Kennedy, 2003). Further, panel analysis is an approach that may account for heterogeneity among the cross-sections, which in this case represents the fixed effects model. In this study, comparisons are made on the company type-level. Thus, company-type-specific intercepts through a dummy variable can be included, which act as regressors and enable the control of unobservable effects on the DV. This approach considers company type as a time-invariant regressor, compared to sentiment and search volume, which differ over time. On the other hand, the random effects model assumes that individual-specific effects are correlated with regressors. Using the Hausman test, it is possible to statistically assess which model is better suited based on differences between coefficient estimates obtained by both alternatives. The R package plm was used. Since several brand rankings were used, distinct panel models were created and compared. In particular, the models were contrasted based on model fit and parameter effects for online WoM sentiment, (search) volume, and company type. Furthermore, the effects per company type were tested for significant differences using a Z-test. 3.3 Panel Regression Model Specification

A panel regression was conducted for each transformation of the DV, as discussed in section 2.2

Conceptual Model and Hypotheses. Further, the panel regression was applied to each brand

ranking provider separately, to 1) not mix up different methodologies used, and 2) compare the outcomes of the models eventually. In the following, the model specifications are elaborated. It is to be noted that due to multicollinearity issues, the sentiment variables polarity_mean and

polarity_sd were excluded. Further, industry and type show a high correlation, while the inclusion

19 interactions between industries. Therefore, only the company type was used to allow for comparisons. Appendix C, Table 14 - Table 17 show the correlation matrices for each case. Important to consider is that for the purpose of this study, the Hausman test was used to determine which model fits the data better: If the p-value < 0.05, the fixed effects model is more suitable than the random effects model since there is a systematic difference in coefficients. Initially, the absence of heteroscedasticity was tested to allow for the Hausman test. In this section, modeling issues are discussed per DV separately.

Dependent variable: Log of absolute brand value

Initially, absolute brand value was used as DV. After a first estimation, however, extremely low R²s result, i.e., sentiment and search volume do not well explain the DV alone. Alternatively, the log of absolute brand value was used as DV while controlling for its lag as an IV. This improved the model’s performance and showed significance for some IVs. Since only 53 brands were observed, and thus a limited amount of observations was available, an evolutionary approach was used to investigate the effects. Model (1) only features main effects, while model (2) includes interactions of sentiment and search volume. Initially, also the change in absolute brand value as a separate DV should be investigated but using the log of absolute brand value made this DV redundant due to the same interpretation, hence using it as validation (Appendix C, Figure 25). (1) Model for main effects

𝐿𝑜𝑔(𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒)𝑏𝑟𝑡= 𝛼 + 𝛽 ∗ 𝑙𝑎𝑔(𝑙𝑜𝑔 (𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒)𝑏𝑟𝑡) + 𝛾 ∗ 𝑝𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑏𝑡+ 𝛿 ∗

𝑝𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑏𝑡+ 𝜖 ∗ 𝐺𝑜𝑜𝑔𝑙𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦𝑏𝑡+ 𝜃 ∗ 𝑡𝑦𝑝𝑒𝑏+ 𝜀

20 Dependent variable: Relative brand value

In the case of relative brand value as DV, a log-transformation was not necessary since it is shown as a percentage already. The same evolutionary approach as for the absolute brand value was applied, with the model (1) showing the main effects, and model (2) including interaction effects. (1) Model for main effects

𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒𝑏𝑟𝑡= 𝛼 + 𝛽 ∗ 𝑙𝑎𝑔(𝑟𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒𝑏𝑟𝑡) + 𝛾 ∗ 𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑏𝑡+ 𝛿 ∗

𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑏𝑡+ 𝜖 ∗ 𝐺𝑜𝑜𝑔𝑙𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦𝑏𝑡+ 𝜃 ∗ 𝑇𝑦𝑝𝑒𝑏+ 𝜀

(2) Model for interaction effects of sentiment and search volume 𝑅𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒𝑏𝑟𝑡 = 𝛼 + 𝛽 ∗ 𝑙𝑎𝑔(𝑟𝑒𝑙𝑎𝑡𝑖𝑣𝑒 𝑏𝑟𝑎𝑛𝑑 𝑣𝑎𝑙𝑢𝑒𝑏𝑟𝑡) + 𝛾 ∗ 𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑏𝑡+ 𝛿 ∗ 𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑏𝑡+ 𝜖 ∗ 𝐺𝑜𝑜𝑔𝑙𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦𝑏𝑡+ 𝜃 ∗ 𝑇𝑦𝑝𝑒𝑏+ 𝜗 ∗ 𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑏𝑡 ∗ 𝐺𝑜𝑜𝑔𝑙𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦𝑏𝑡+ 𝜇 ∗ 𝑃𝑜𝑙𝑎𝑟𝑖𝑡𝑦 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑏𝑡∗ 𝐺𝑜𝑜𝑔𝑙𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦𝑏𝑡+ 𝜀 where b: brand r: ranking provider t: year ε: error term; 3.4 Descriptive Statistics and Model-Free Evidence

Before the data could initially be investigated, data cleaning and preparation was necessary as part of the data transformation process (Verhoef et al., 2016), concerning standardization, matching, and consolidation. After brand names and industries have been standardized across years and rankings, the brand value data could be matched with the extracted tweet sentiment and Google popularity scores and consolidated into a complete data frame.

Table 1: Descriptive statistics of IVs

Mean Median Std. Dev. Min Max

21 Brand _Finance Brand_value_mio 29569 21379 24754 7994 17003 Polarity_pos_mean 20.52 20.54 6.32 4.35 16.98 Polarity_neg_mean 12.21 11.09 6.82 3.12 8.62 Google_popularity 66.21 68.92 19.68 8.92 55.46 Forbes Brand_value_mio 26577 17000 28177 5700 10500 Polarity_pos_mean 21.13 21.16 6.42 4.70 17.39 Polarity_neg_mean 12.30 10.82 7.77 3.12 8.46 Google_popularity 63.83 65.83 20.02 9.08 49.25

Table 1 shows an overview of descriptive statistics of the considered IVs. The rankings differ to

some extent when it comes to brand value, showing a large spread, i.e., standard deviation, for BrandZ in particular. In case of the remaining IVs, the figures are somewhat similar which can be ascribed to the fact that only brands have been considered which are present in all rankings. Thus, mainly yearly variations cause differences in the figures. Further, Figure 15 (Appendix C) features that the brand selection showcases a high dominance of three industries – technology, financial

services, and automotive – making up nearly 50% of the 53 brands in total. This shows a consistent

view of the four brand ranking providers regarding industries in which brands are essential, as well as the general rise in (strong) brands within these industries. B2C brands constitute the most substantial part of the brands (57%), while mix brands with 26% and B2B brands with 17%, are somewhat less represented (see Figure 3). This may indicate a coherent view across rankings that B2B brands are not as relevant as opposed to B2C brands.

Figure 3: Distribution of brands per company type

The trend of especially B2C brands being dominant across rankings is also visible in the brand value variation and growth in brand value across the period considered. Figure 4 shows similar means across all types with B2C and mix companies demonstrating a higher variation across all rankings. B2B firms experience a rather small value range. This may imply that B2C and mix brands include a variety of potentially big players, such as Amazon and Apple, as well as less valuable brands (according to the rankings), such as eBay. Further, B2B brands experienced a less steep growth rate of brand value across time in all rankings compared to B2C firms (see Figure 5). Mix brands experience tremendous growth, which may again be ascribed to the rise of technology giants. Apparently,

22 BrandZ incorporated this rise most sensibly. Looking at the industry comparison (see Figure 16 &

Figure 17), the trend of dominant industries prevailing is also clear: Across all rankings, technology brands enjoy the highest brand value on average with large positive outliers. Most

industries’ brands demonstrate a similar average brand value, especially in the Brand Finance and Forbes ranking. At BrandZ and Interbrand, the variation is slightly more spread, with specifically

business services, financial services, FMCG, and restaurants showing a few more positive

outliers. Further, only the restaurant industry shows significant growth throughout the years. This could again be ascribed to large players and/or a focus of these rankings on specific industries.

Figure 4: Brand value variation per company type and ranking provider

23 On average, the sentiment including all company types is positive with an overall mean of 0.03366. However, as shown in Figure 6, only mix brands appear to be rather stable. Especially B2B brands show a considerable variation with many negative outliers. Having a closer look at the data reveals that SAP shows negative sentiment scores for all considered years, thus clearly polarizing the results. On an industry level, about half shows a rather stable average sentiment score without many outliers (see Figure 18, Appendix B). This could indicate that online WoM on average, is rather positive and constant in these cases. In the case of other industries, e.g., retail, financial

services, FMCG or business services, the average sentiment score is more spread, leading to the

assumption that these industries experience high fluctuations due to 1) particular brands in this group, or 2) general external events impacting the industry’s online WoM as a whole.

Figure 6: Sentiment variation per company type

The maximum

sentiment score achieved is 0.1516 in the retail industry. Especially interesting is the fact that only brands in the business

services, financial services, and energy industries contain observations which experience negative sentiment scores

on average, with business services clearly standing out and a minimum score of -0.2388.

Across time, it is visible that the average sentiment scores appear to converge and become more stable in recent years, as shown by the example of company types in Figure 7. While B2C and mix brands are generally perceived more positively across the years, B2B brands overall experienced tremendous growth, approaching the other company types’ scores. With regard to industries (see

Figure 19, Appendix B), especially in the early years the retail and business services industries

24 of negative WoM on brands within the same industry. Further, business services experienced massive growth, which may be explained by the initially negative perception of SAP.

Figure 7: Average sentiment score per company type across time

Lastly, the Google popularity score obtained through Google Trends shows a considerable variation when comparing company types, as opposed to the sentiment discussed previously. By definition, this score is seen relative compared to the peak of the brand in the period considered. As shown in Figure 8 and Figure 9, B2B brands show stable average popularity compared to B2C and mix brands, which experience more fluctuations. However, across time it is clearly visible that they experience the largest popularity decline, while mix brands suffer from this loss as well, and B2C brands stay rather constant. This could point out SAP’s negative sentiment influence on the initial years since negative publicity might attract users to search more heavily for a particular brand. A comparison on industry-level shows that brands part of the industries business services,

electronics, financial services, FMCG, media, restaurant, and technology experienced the largest

fluctuations in search volume (see Figure 20 & Figure 21). Brands in the energy and logistics sector demonstrate a rather stable popularity, which could be ascribed to the offering of commodities which are assumed to be rather less subject to online interest. This is supported by literature stating that experience as well as high-involvement products are more often subject to online WoM and search, respectively (Ha, 2002; Bei et al., 2004). Considering the search volume over time further demonstrates apparent variations across industries. Several industries show a sinus- or cosinus-curve path, meaning that after a period of popularity gain, they lost it, or after a period of popularity loss, they regained it. This includes the popularity development of the alcohol,

25

restaurants, media, and FMCG show a rather steady growth, while the search volume of diversified, business services, retail and financial services declines throughout the years.

Figure 8: Search volume variation across company types

Figure 9: Average search volume per industry across time

26 brands. For B2C and B2B brands, the relation appears to be rather stable. Further, correlation was tested: As shown in the correlation matrices (see Figure 24, Appendix B) the relations between the IVs with a range of -0.2 until 0.1 are rather weak, thus not necessarily implying a significant effect. However, in most cases the direction of the effect is as expected, i.e., negative in case of negative sentiment and positive in case of positive sentiment and search volume.

Figure 10: Scatter plot of the relation brand value – positive sentiment

4 RESULTS OF THE ANALYSIS

A summary of the data sets used for each model is provided in Table 2. The number of observations differs to some extent among the rankings, while B2C and mix brands are represented the most. Since not all brands are observed in each year and ranking, the dataset is considered as slightly

unbalanced. However, an initial check of the data using the punbalancedness() function in R

shows that the panel data of all rankings is > 0.7 in most cases, thus being somewhat balanced.

Table 2: Summary of data sets for panel models

Interbrand BrandZ

Main B2C B2B mix Main B2C B2B mix

# of observations 491 279 90 122 440 231 90 119

# of brands 53 30 9 14 53 30 9 14

27

Brand Finance Forbes

Main B2C B2B mix Main B2C B2B mix

# of observations 391 208 79 104 345 198 58 89

# of brands 53 30 9 14 53 30 9 14

Brand value mean, mio $ 29569 26581 24575 39340 26577 23907 26210 32756 Polarity mean 0.032 0.034 0.021 0.036 0.035 0.040 0.013 0.037 Google popularity mean 66.21 66.55 60.91 69.55 63.83 63.38 56.76 69.42 Balancedness gamma/nu 0.59/ 0.83 0.55/ 0.79 0.89/ 0.93 0.61/ 0.84 0.87/ 0.96 0.96/ 0.97 0.85/ 0.94 0.74/ 0.93

Before analyzing the results of the previously specified models, the model validation overall was conducted. First, a potential issue of heteroscedasticity was tested: Using the Breusch-Pagan test revealed insignificant results for each model (p-value > 0.05), thus allowing for the use of the

Hausman test. The latter was used to investigate whether a fixed effects or random effects model

28 in Appendix C, Table 18 - Table 25 for absolute brand value, as well as Table 34 - Table 41 for relative brand value as DV.

Furthermore, after including the lag of the DV as an IV in the model, the R²s are naturally much higher. It is to be noted, however, that the exclusion of the lag variable resulting in substantially lower R²s provides insights into the relevance of online WoM sentiment and search volume for the specific rankings and company types itself. Exclusion results in most models being more suitable as a random effects models then. As shown in Table 3 and Table 4, the R²s for both DVs would be substantially smaller, while the extent is even larger for relative brand value as DV. Across rankings, these figures are rather consistent. In turn, sentiment and search volume alone do not well explain any variance in relative brand value. It may thus be concluded that factors other than these IVs substantially affect relative brand value, which are accounted for by including the lag of the DV – as shown by the then substantially higher R²s. In the case of absolute brand value, the difference in R²s is present yet not as vast. Further, there are differences visible across rankings: Especially for BrandZ and Forbes, the R²s are comparably higher, thus indicating that sentiment and search volume alone do indeed explain some variance in absolute brand value. For Interbrand, these figures are slightly lower while it is substantially lower for Brand Finance. Hence, sentiment and search volume alone barely explain any change in the DV.

Table 3: R²s of panel regression models without lag - DV: Absolute brand value

Main effects models Interaction effects models

Interbrand BrandZ Brand Finance Forbes Interbrand BrandZ Brand Finance Forbes Main 0.1917 0.4007 0.0446 (FE) 0.3015 0.2032 0.4066 0.0735 (FE) 0.3273

B2C 0.2710 0.5074 0.0982 (FE) 0.2088 0.2659 0.5046 0.1429 (FE) 0.2170 B2B 0.2086 0.2786 0.6401 0.4683 0.2998 0.3573 0.6109 0.6972 mix 0.1841 0.3663 0.1147 (FE) 0.5373 0.2146 0.3701 0.1546 (FE) 0.5621 * FE = fixed effects model

Table 4: R²s of panel regression models without lag - DV: Relative brand value

Main effects models Interaction effects models

Interbrand BrandZ Brand Finance Forbes Interbrand BrandZ Brand Finance Forbes Main 0.0497 0.0128 0.0451 (FE) 0.0190 0.0601 0.0161 0.0729 0.0543

B2C 0.0539 0.0396 0.0823 0.0512 0.0612 0.0405 0.0770 (FE) 0.0617 (FE) B2B 0.2202 0.1608 0.4042 0.2965 0.3046 0.1919 0.4560 0.4267

29 4.2 Results of the Panel Regression on Absolute Brand Value

4.2.1 Main Effects

To estimate the main effects of sentiment and search volume on the log of brand value, a panel model for each ranking provider without interactions was developed. Additionally, models for B2C, B2B, and mix brands were created separately to investigate effects per company type. Further, estimates per IV were tested for significant differences across company types and rankings using a two-tailed Z-test since there are no assumptions regarding the direction of differences upfront. According to Paternoster, Brame, Mazerolle and Piquero (1998), the following formula can be used to test differences in estimates across regressions with different samples, where b is the parameter estimate and SEb the standard error:

Z =

√

22

Table 5: Panel regression absolute brand value – Main effects per ranking IV / Model Interbrand BrandZ Brand Finance Forbes

30 For the different rankings, the main effects on absolute brand value are rather consistent across models. Naturally, the lag of log(brand value) shows the highest significance and impact for all rankings. This is the case since the previous absolute brand value captures most of the brand-specific aspects, as well as all past events. Therefore, the R² is very high, ranging from 0.71 to 0.89 in all overall models (Appendix C Table 18, Table 19, Table 20, Table 21), meaning that all IVs jointly well explain any variance in the log of absolute brand value. Interestingly, Brand Finance shows the lowest R²s in all models across the rankings. This may provide an insight into this ranking’s brand valuation methodology to be more independent of the previous brand value. Considering the remaining main effects shows differences in terms of significance but mostly similarities in parameter signs. As expected, positive sentiment has a significantly positive yet rather marginal effect for both Interbrand and BrandZ, while the effects are slightly stronger in the latter case where a 1 unit (≙ %) increase in positive WoM changes the absolute brand value by 0.57%. For the remaining rankings, the effect is not significant overall. Splitting the models according to company type, however, shows some significant results: There is a positive main effect visible for B2C brands at Interbrand and B2B brands at Forbes. Looking at the heat map of main effects as shown in Figure 11 points out that in fact there are no significant differences among company types within each ranking except for Forbes, where B2C and B2B estimates differ. Thus, B2B brands benefit more than B2C brands. Further, the overall effect for BrandZ is significantly positive, while there are no significant differences among company types. The insignificance of estimates may be due to model power issues. Only Brand Finance’s brand values seem not to be affected by positive sentiment.

31 rankings, the effects do not differ substantially across company types. Across rankings, however, the estimates differ for B2C brands at Interbrand and Brand Finance where the latter does not show any significance at all, as well as for B2B brands at Interbrand compared to BrandZ and Forbes. Lastly, search volume is overall significant and positive for both Interbrand and Brand Finance, with a higher impact in the latter case: A 1% increase in search volume leads to a 0.126% increase in absolute brand value. For the remaining rankings, there seems to be no impact overall at all. Investigating differences across company types, however, reveals that search volume matters positively for B2B and mix brands at Interbrand, with a slightly higher growth of 0.244% in absolute brand value for mix companies. At Brand Finance, search volume matters for B2C brands only. Considering the heat map matrix in Figure 11 demonstrates that in fact the effect of search volume on the log of brand value differs across B2C and B2B as well as B2C and mix brands within the Interbrand ranking. Thus, the previously mentioned effect difference is not significant. It is further visible that estimates differ across rankings: Interbrand’s B2B estimate differs significantly from BrandZ.

Figure 11: Heat map matrix of significant differences among main effects on absolute brand value

Legend: green = insignificant, orange = significant, light orange = marginally significant, grey = irrelevant

32

4.2.2 Interaction Effects of Sentiment and Search Volume

In the next step, a panel model for each ranking provider with interactions was developed. With this, it could be investigated whether the effect of sentiment is moderated by search volume, thus potentially revising the previous interpretation of main effects. As shown in Table 6, there appear to be no interaction effects between sentiment and search volume overall. In all rankings, only the

lag of brand value is significant, except for search volume, which is significant in the Forbes

ranking only. Due to the absence of interaction effects, the main effects model is preferred for interpretation due to better statistical efficiency, shown in 4.2.1 Main Effects.

Table 6: Panel regression absolute brand value – Interaction effects of sentiment and search volume per ranking

IV / Model Interbrand BrandZ Brand Finance Forbes

33 Investigating these effects separately per company type, however, shows that only Interbrand and Forbes reveal interaction effects at all. For other rankings, the interaction of sentiment and search volume seems not to be interdependent, thus the main effects count. To be more concrete, only mix brands at Interbrand experience interaction effects of positive sentiment and search volume: With a negative interaction effect, the positive impact of positive sentiment is weakened to 0.1131 - 0.0243 = 0.0888 with a 1% increase in search volume. The heat map of significant differences in Figure 12 shows consistent results where only Interbrand shows significant differences among company types, i.e., mix and B2C brands within its ranking. Across rankings, there are only marginally significant differences for mix brands across Interbrand and BrandZ as well as Brand Finance, and B2C brands between Brand Finance and Forbes.

Concerning the interplays between negative sentiment and search volume, only Forbes shows significant results for both B2B and mix brands: The interaction effect of negative sentiment and search volume for B2B brands is negative, while it is positive for mix brands. Investigating these in combination with the main effects shows that for B2B brands the main effect of negative sentiment is positive which in turn is weakened through a negative interaction effect: If search volume of a B2B brand increases by 1%, the effect of negative sentiment on the log of brand value decreases to 0.0721 - 0.0164 = 0.0557, thus still being positive. In fact, B2B brands would require a minimum increase in search volume by x = | 0.0721/-0.0164 | = 4.4% in order to experience an adverse effect of negative sentiment on absolute brand value. In the case of mix brands, the interaction effect of negative sentiment and search volume is positive, however. This means that the negative effect of a negative sentiment is weakened if the mix brand’s search volume increases: If search volume increases by 1%, the effect of negative sentiment on the log of brand value is weakened to -0.1226 + 0.0262 = -0.0964, thus still being negative. With respect to the heat map of significant differences shown in Figure 12, it is clearly visible that in fact the interaction effects differ significantly for company types within the Forbes ranking only. Mix and B2C brands in the Interbrand ranking only differ marginally. Across rankings, there are differences visible for B2B brands only (between Forbes and Interbrand, and Forbes and BrandZ).

34 many significant IVs in the fixed effects model become insignificant. Those who stay significant, however, show a high similarity of estimates. Running the same models with change as DV shows clear similarities and thus internal consistency: IVs which are significant in both variants show the same estimate sign and similar strength. There are some exceptions in which IVs suddenly become positive, e.g., the interaction of positive sentiment and search volume at Forbes overall. However, models with change as DV show a substantially lower R² ranging from 0.05 to 0.40. Hence, it is assumed that some IVs become significant due to omitted variable bias, for instance.

Figure 12: Heat map matrix of significant differences among interaction effects on absolute brand value

Legend: green = insignificant, orange = significant, light orange = marginally significant, grey = irrelevant

With these findings, H₁ can be accepted for three out of four rankings, while the strength differs among rankings and company types: The effect of positive sentiment is positive overall for Interbrand and BrandZ, as well as Forbes where only B2B brands are significantly affected – yet there is no significant difference to other company types. H₂, on the other hand, can only conditionally be accepted: The effect of negative sentiment is negative yet only marginally significant at Interbrand overall and shows a negative effect for B2C brands at Interbrand. H₃ can hence be evaluated for Interbrand only: The overall as well as separate effects for B2C companies are similarly strong for both positive and negative sentiment. Thus, H₃ must be rejected. Further,

H₄ needs to be rejected for all cases in which interaction effects are significant due to opposite

35 shown that positive sentiment has no significantly different effect depending on the company type, as opposed to negative sentiment and search volume. However, this applies to Interbrand and Forbes only. At BrandZ and Brand Finance, there seem to be no distinctions.

4.3 Results of the Panel Regression on Relative Brand Value

4.3.1 Main Effects

Next to the absolute brand value, it was of interest to investigate the impact of sentiment and search volume on relative brand value, i.e., the proportion of a brand’s value compared to the total value of all brands within the same industry. This provides an insight into the competitive position. For the estimation of main effects, a restricted model was developed by excluding interaction effects. This was again necessary due to the low number of observations which may impact statistical power. Further, separate models were created per company type.

Table 7: Panel regression relative brand value – Main effects per ranking IV / Model Interbrand BrandZ Brand Finance Forbes

Intercept - - - - B2C - - - - B2B -4.6697 -11.6186 -7.0030 -16.3231** mix -4.6400 - - - Lag(log(rel_value)) 0.7967*** 0.6052*** 0.6399*** 0.7351*** B2C 0.7609*** 0.5938*** 0.6917*** 0.7733*** B2B 0.9912*** 0.8982*** 0.8558*** 0.9454*** mix 0.9028*** 0.4919*** 0.3232*** 0.1480 Polarity_pos_mean 0.0459 0.0702 -0.2807** 0.0270 B2C 0.0825* 0.3368. -0.5006* 0.1407* B2B 0.0367 0.0863 0.0573 0.0245 mix 0.0202 -0.3427 -0.3799* 0.0296 Polarity_neg_mean -0.0690 -0.1534 0.0566 0.0569 B2C -0.1020* -0.4147 0.4147. -0.1393. B2B 0.0059 0.0346 0.0079 -0.0243 mix 0.0825 -0.1851 -0.0633 0.4986** log(Google_popularity) 1.8539* 5.6808* -1.6246 2.9410** B2C 2.2289*** 7.0191* -5.9293. 2.0157. B2B 0.8763 2.5510 1.7442 4.1504** mix 1.0726 1.6765 0.0946 3.8865 Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

36 brand value captures most of the brand-specific aspects and past events, i.e., also those which determine the competitive position. For the remaining IVs, there is a rather inconsistent pattern visible: It appears that mainly search volume is relevant for the relative brand value, except for Brand Finance where it is not significant at all. However, the Brand Finance models demonstrate substantially lower R²s compared to the other rankings (Appendix C, Table 36). Thus, the significance levels may not be sufficiently accurate and should be analyzed with caution. On the other hand, BrandZ shows the comparably highest impact where a 1% increase in search volume results in a 0.057% increase in relative brand value. Investigating company types separately further demonstrates that search volume seems to not matter for mix brands at all, while it shows a significantly positive effect for B2C brands at Interbrand and BrandZ, as well as B2B brands at Forbes only. In all cases, the estimates are positive, thus showing that both B2C and B2B brands’ relative brand values benefit from increased search volume, while it is especially high for B2C brands at BrandZ. However, the preferred BrandZ fixed effects models show low R²s (Appendix C, Table 35), potentially leading to omitted variable bias. With regard to the heat map matrix of significant differences, shown in Figure 13, the estimates seem to not differ for company types within rankings except for Brand Finance where B2B differs significantly from B2C. Across rankings, mainly B2C estimates differ between Brand Finance and all other rankings.

37 to be beneficial for mix brands at Forbes. The heat map matrix of significant differences in Figure

13 shows estimates differ within rankings for Forbes, with significant differences among B2C and

mix as well as B2B and mix brands, i.e. the effect is significantly more positive for mix than any other brands. Across rankings, there are differences mainly for B2C brands.

Figure 13: Heat map matrix of significant differences among main effects on relative brand value

Legend: green = insignificant, orange = significant, light orange = marginally significant, grey = irrelevant

Comparing these outcomes with the un-preferred models, i.e., random effects models in nearly all cases show that in most models, all IVs become insignificant while the R²s increase substantially (Appendix C, Table 42, Table 43, Table 44, Table 45).

4.3.2 Interaction Effects of Sentiment and Search Volume

Further, interaction effects were investigated in order to assess whether the main effects may be interpreted as they are. Again, an overall model and models for each company type were created.

Table 8: Panel regression relative brand value – Interaction effects of sentiment and search volume per ranking

IV / Model Interbrand BrandZ Brand Finance Forbes

38 mix 0.0317 -3.9550 -5.2180** -4.1549** Polarity_neg_mean 0.0966 -0.2227 1.0032 -0.2096 B2C 0.0580 -1.2795 0.6868 -0.1902 B2B -0.4608 -0.4172 -0.3388 0.8193 mix -0.1287 0.2940 2.3123 1.3828 log(Google_popularity) 0.1246 2.9196 -4.6924 0.1272 B2C 0.3594 14.4362 -0.0151 2.9470 B2B -2.7934 -22.2002* -2.7694 11.1613 mix 0.5277 -13.5734 -16.6549. -14.0894. Polarity_pos_mean: log(Google_popularity) 0.0995 0.1173 0.2771 0.0935 B2C 0.1009 -0.4448 -0.2777 -0.0481 B2B 0.0807 0.9874* 0.1363 -0.1174 mix -0.0027 0.8212 1.1023** 0.9564** Polarity_neg_mean: log(Google_popularity) -0.0374 0.0201 -0.2231 0.0659 B2C -0.0363 0.1954 -0.0687 0.0120 B2B 0.1099 0.1306 0.0810 -0.1980 mix 0.0492 -0.1021 -0.5291 -0.1922 Significance codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

As shown in Table 8, there are no significant interaction effects of sentiment and search volume visible overall when it comes to investigating impacts on a brand’s relative brand value. Therefore, the main effects using the restricted and more efficient model are being used for interpretation, as shown in section 4.3.1 Main Effects.

39 other brands. For the other rankings, the effects seem not to differ substantially. Across rankings, there are differences mainly for B2B brands between BrandZ and Forbes as well as mix brands between Interbrand and Brand Finance.

When it comes to the interaction effect of negative sentiment and search volume, there is no significant effect visible at all. It can thus be inferred that search volume does not impact the effect of negative sentiment on relative brand value. In direct comparison with the un-preferred random effects models (Appendix C, Table 46, Table 47, Table 48, Table 49), these models show fewer significant estimates. However, those IVs which are significant in both cases show similar strengths, thus being internally consistent.

Figure 14: Heat map matrix of significant differences among interaction effects on relative brand value

Legend: green = insignificant, orange = significant, light orange = marginally significant, grey = irrelevant