The effect of business strategy on the relation between big data analytics capability and firm performance

(1)

The effect of business strategy on the

relation between big data analytics

capability and firm performance

Bart Visser – 11409185 22-06-2018

MSc. in Business Administration – Strategy Track University of Amsterdam

(2)

1

Statement of originality

This document is written by Bart Visser who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

2

Abstract

Big data analytics (BDA) is believed to become a driving force for innovation and competition across many industries. This study aims to discover how it can be effectively utilized and if some firms, depending on their strategy, can benefit from it more than others can. Scholars have proposed that, in order to benefit from BDA, businesses should go beyond the BDA technology itself and build BDA capability by incorporating relevant human an intangible resources. Two research questions are addressed in this study; 1. Does BDA capability lead to increased firm performance, and 2. does business strategy of a firm influence this relation. For addressing these questions, data is collected using the survey research method. 63 usable responses were collected from of big data analytics managers. BDA capability has a positive and significant effect on both Operational Performance (β= 0,54) and Market Performance (β= 0,74), explaining 34,5% and 55,9% of the variance respectively. For Prospectors, the effect of BDA capability on relative Market Performance is lower than for Low-Cost Defenders and Reactors (Δβ = -0,43). Other strategic types did not significantly affect the relationship with respect to Low-Cost Defenders and Reactors. However, these findings may not be conclusive as the study is limited in terms of its validity and reliability.

(5)

4

1 Introduction

“Information is the oil of the 21st_{century, and analytics is the combustion engine”} - Peter Sondergaard, Gartner Research

This statement speaks volumes about the importance of data and analytics. Data will become increasingly more valuable, as it becomes widely available as a result of trends like massive popularity of mobile devices in everyday life and the Internet of Things. The latter is a worldwide trend in which technologies like Radio Frequency Identification and sensor network technologies are used to acquire data from everyday objects in the environment by connecting them to the internet (Gubbi, Buyya, Marusic, & Palaniswami, 2013). By 2025, 75 billion devices are estimated to be connected to the internet worldwide, which will be a significant increase over the 23 billion devices connected in 2018 (“Internet of Things - number of connected devices worldwide 2015-2025,” 2015). The voluminous, fast flowing and diverse data generated through the Internet of Things is generally called ‘big data’ (McAfee & Brynjolfsson, 2012). Big data is a very popular topic among scholars and practitioners, as amounts of data of such magnitudes can offer businesses valuable insights, help to create significant value and give rise to a competitive advantage over competitors (Manyika et al., 2015; Morabito, 2015).

However, following the analogy of Peter Sondergaard, big data on its own will not be of much use without analytics. In order for businesses to gain access to the benefits of this ever increasing pool of data, the data has to be mined and analysed (Y. Chen et al., 2014). This is generally referred to as big data analytics (BDA). Popular belief is that big data is relevant for virtually all businesses (H. Chen, Chiang, & Storey, 2012) and that BDA will be a large driver of innovation and competition across many industries (Manyika et al., 2011). A survey from 2014 found 89% of responding companies to believe that businesses that choose not to adopt a BDA strategy will lose market share and momentum (Columbus, 2014). Another indication of

(6)

5 the rising popularity of BDA is the growing market for BDA related software and services, which is expected to exceed $210 billion by 2020 (IDC, 2017).

Scholars and practitioners seem to agree on the huge potential of BDA. However, trends of such magnitudes often bring along risks of a hype resulting in uninformed investments by organizations that do not comprehend how BDA should be integrated and whether or not it can actually provide value for their business (McAfee & Brynjolfsson, 2012; Ross, Beath, & Quaadgras, 2013). In other words, BDA is certainly highly interesting for many, but might not be for everyone.

Discovering how and for what kind of businesses BDA can be used as a competitive force makes for an interesting field of research. In studying how it does so, scholars have made an effort to go beyond the basis of data and analytics technologies by developing the concept of BDA capability. Following the reasoning of Grant (1991) within the resource based view (RBV), capabilities are the product of coordinating and combining resources. BDA capability comprises tangible, intangible and human resources needed for effectively utilizing big data and is defined as “a firm’s ability to assemble, integrate, and deploy its big data-specific resources” (Gupta & George, 2016, p. 1049). Akter et al. (2016) came up with a similar concept for BDA capability. Both studies found a positive relation between BDA capability and relative firm performance.

Thus far, only a limited amount of empirical evidence, based solely on US samples, has been put forward for establishing this relation. This study will attempt to broaden this empirical basis in a European context by addressing the following research question:

RQ1. What is the relation between big data analytics capability and firm performance? Building the concept of BDA capability and establishing its relation with firm performance is a first step in discovering how BDA can add value. This study will make a first step in discovering what kind of firms can most effectively use BDA capability as a competitive

(7)

6 force. For doing so, while investigating the relation between BDA capability and firm performance, a distinction will be made between firms on the basis of their general business strategy. Business strategy is described using a combination of the Miles and Snow (1978) and the Porter typology (1980) (Walker & Ruekert, 1987).

Due to the massive interest in BDA, it is likely that firms across many strategic types will invest in BDA. In other words, BDA capability likely is adopted by different firms pursuing different business strategies. However, contingency theory predicts that internal fit of an organizations’ structure and processes is paramount for its effectiveness (Burns & Stalker, 1961; Lawrence & Lorsch, 1967; Miller, 1992; Van de Ven, Ganco, & Hinings, 2013). Due to the structural differences among organizations pursuing different business strategies (Miles et al., 1978; Porter, 1980; Walker & Ruekert, 1987), it is expected that BDA capability might not fit equally well within the internal structure and processes of these organizations. Thus, there is likely to be heterogeneity in firm performance obtained from BDA capability among firms based on business strategy. To investigate this line of reasoning, this study will attempt to address the following research question:

RQ2. What is the effect of business strategy on the relation between big data analytics capability and firm performance?

This study will attempt to answer these research questions using a quantitative research approach using the survey research method. The target population consists of firms that have already invested in BDA. The survey is sent to BDA managers within these firms that are selected via LinkedIn. This study will add to the scientific literature by broadening the empirical basis of the relation between BDA capability and firm performance. Also, it will make a first step in assessing what kind of firms have the right fit for BDA. For practitioners that are considering to invest in BDA for their firms this can be of great value, as it can shed more light onto the value that it can bring to their business in specific.

(8)

7 This study is structured in the following way: in chapter 2 the main body of relevant literature to BDA capability and business strategy is elaborated upon and hypotheses are drafted to address the research questions introduced in this chapter. Chapter 3 contains the methodology of the empirical study. In chapter 4 the results are displayed, addressing validity and reliability of the measurement model before evaluating the structural model, in which the hypotheses are tested. In chapter 5 the results are discussed, limitations of the research are pointed out and future research is proposed. Finally, chapter 6 concludes with a summary of all important findings of the study.

2 Literature review

In this chapter, the main body of relevant literature for this study is reviewed. First, the resource based view is introduced followed by a theoretical description of BDA capability. Then, contingency theory is described followed by the literature on business strategy. Finally, hypotheses are drafted for addressing the research questions.

2.1 Resource based view

The RBV has in part originated through the work of Wernerfelt (1984), who argued that a business should be analysed by looking at its resources rather than solely looking at its products. Briefly after this, Barney (1986b) suggested internal resources to be greater determinants of strategic advantage than a firms’ external environment. He went on to propose a framework of strategic factor markets to explain competitive differences among firms. His idea was that resources needed for implementing a certain strategy can be purchased on these markets. In perfect strategic factor markets, it is impossible for firms to obtain a competitive advantage over their competition as the costs of the resources needed for such a strategic position will be equal to the economic value that these resources will generate. A real strategic advantage can only be obtained through imperfections in factor markets. These imperfections arise when

(9)

8 different firms have diverging expectations about the future value of the strategic resources in the market (Barney, 1986b).

Dierickx and Cool (1989) added to this framework by arguing that the strategic factor markets are incomplete, implying that not all strategic resources can be bought on strategic factor markets. For example, the reputation of a company or its organizational culture cannot simply be bought on a market, but must be internally accumulated over time. Dierickx and Cool (1989) talk about asset stocks and assets flows of non-tradable assets. “Strategic asset stocks are accumulated by choosing appropriate time paths of flows over a period of time. (…) It takes a consistent pattern of resource flows to accumulate a desired change in strategic asset stocks.”(Dierickx & Cool, 1989, p. 1506).

From the viewpoint of the RBV, the task of a firms’ management is to adjust and renew its stock of unique resources over time in order to establish and maintain a specific competitive position (Conner, 1991). Resources provide a sustained competitive advantage if they are valuable, rare, inimitable as well as non-substitutable (VRIN) (Barney, 1991). Barney (1991) describes resources as assets as well as capabilities (among other things). Grant (1991) however, distinguishes between resources and capabilities. He defines capabilities as “the capacity of a team of resources to perform some task or activity” (p. 119) . He argues that resources on their own typically do not amount to competitive advantage, but coordination and cooperation of a combination of resources can. In other words, resources help build capabilities, and capabilities can create a competitive advantage. He also distinguishes between tangible, intangible and human resources.

2.2 Big data analytics capability

Gupta and George (2016) have constructed the concept BDA capability. They define it as: “a firm’s ability to assemble, integrate, and deploy its big data-specific resources” (p. 1049). They argue that, in assessing whether a firm is able to utilize their BDA in an effective way, it is not

(10)

9 merely relevant to measure its ability to acquire and analyse big data. It is at least as important to assess the firm’s ability of effectively integrate and exploit the insights obtained through BDA. Following the perspective of Grant (1991), Gupta and George (2016) build BDA capability from seven different resources which are segregated into three categories; tangible, intangible and human resources. In the following sections, these resources will be defined and elaborated upon per category.

2.2.1 Tangible resources

Tangible resources are assets such as raw materials, equipment or financial capital that, conforming to the RBV, can be bought in a market. According to Barney (1991) these resources tend to be readily available and therefore do not meet the VRIN requirements. This indicates that tangible resources on their own cannot amount to a competitive advantage. However, these resources are a necessary condition for building BDA capability. According to Gupta and George (2016), tangible resources needed to build BDA capability are; data, technology and basic resources. Even though data does not seem tangible at first, it certainly is as it is stored in physical locations on physical devices. In the following sections, these tangible resources will be expanded upon.

2.2.1.1 Data

The resource ‘data’ within the construct BDA capability is no regular data, but rather ‘big data’. Big data is not fundamentally different from regular data. However, big data is so voluminous and complex that traditional data processing tools are unable to extract meaningful insights from it. Therefore, it is considered a different class of data for which more sophisticated types of analysis are required (Gupta & George, 2016).

Although there exists no consensus about the exact definition of big data, it is often described by its volume, velocity and variety (Davenport, 2014; McAfee & Brynjolfsson, 2012). Volume stands for the sheer amount of data that is present, which is continually increasing through global digitalization. An example is the development in the Internet of

(11)

10 Things, in which technologies such as sensor network technologies are able to generate data points from objects and devices by connecting them to the internet (Gubbi et al., 2013). Velocity is the speed at which data is created. In contrast with regular data, big data is often generated and analysed real-time or nearly real-time (McAfee & Brynjolfsson, 2012).

Variety stands for the diversity of data types such as GPS locations, raw text from social media platforms like Facebook or Twitter or data from Radio Frequency Identification devices that are becoming increasingly more prevalent. The wide variety of data that is available nowadays creates one of the biggest challenges of big data as it can lack structure, which can make it difficult for organizations to organize the data in a meaningful way (DeVan, 2016). Zhao, Fan and Hu (2014) make a general distinction between internal and external data. Internal data refers to data that originates within the firm which often created as a result from e.g. its internal operations, transactions or sales. Internal data is often limited in its variety and does not provide firms with new insights that can be used for accurate business predictions. External data originates from the general population, often from internet based sources such as social media. This type of (big) data can come in many different forms, such as plain text from posts on social media, internet browsing behaviour or social circles. Therefore it has a high variety and can provide insights in e.g. market trends (Zhao et al., 2014).

2.2.1.2 Technology

As described in the prior section, big data comes in many different forms, is very large in volume and is often fast-moving. In order for businesses to gain access to the benefits of this ever increasing pool of data, the data has to be mined and analysed (Y. Chen et al., 2014). Traditional relational database management systems (RDMS) are not able to handle these kinds of data. To extract valuable insights from this data, more sophisticated analytics technologies are needed such as Hadoop, NoSQL and NewSQL (Storey & Song, 2017).

(12)

11

2.2.1.3 Basic resources

Apart from the availability of big data and sophisticated data analytics technologies, data analytics projects require sufficient financial investments and time in order to be successful. Gupta and George (2016) have chosen to refer to these resources as basic resources.

2.2.2 Intangible resources

Intangible resources can be divided into two classes, namely ‘assets’ and ‘skills’. Examples of assets are patents or copyrights which can be owned and bought. Skills can be present on the level of the individual employee as specific ‘know-how’ or can be entrenched within the organization as a whole in the form of organizational culture. Skills cannot be bought in a market and need to be built within the organization and thus accumulated over time (Hall, 1992). They often meet Barney’s (1991) VRIN criteria and thus amount to a competitive advantage (Teece, 2014). Gupta and George (2016) argue that the presence of the intangible resources ‘data-driven culture’ and ‘organizational learning’ are important factors in explaining differences in the effective utilization of big data analytics among businesses. These intangible resources will be expanded upon in the following sections.

2.2.1.1 Data-driven culture

Although there is no clear agreement on the definition of organizational culture, Barney (1986a) defines it as “a complex set of values, beliefs, assumptions, and symbols that define the way in which a firm conducts its business” (p. 657). Alike Gupta and George (2016), many scholars suggest that big data analytics on its own has little value for businesses. It is argued that for effectively utilizing BDA, having a data-driven culture throughout the organization is paramount (Davenport, Harris, De Long, & Jacobson, 2001; McAfee & Brynjolfsson, 2012; Mikalef, Framnes, Danielsen, Krogstie, & Olsen, 2017; Ross et al., 2013). They state that a data-driven culture within a firm requires decisions throughout the organization to be made on the basis of analytical evidence instead of intuition. McAfee and Brynjolfsson (2012) describe data-driven culture in the following way: “The first question a data-driven organization asks

(13)

12

itself is not “What do we think?” but “What do we know?” This requires a move away from acting solely on hunches and instinct.” (p. 9)

Data-driven culture can be present at the beginning of an organization, but often evolves over time instead (Kiron & Shockley, 2011). In order for a data-driven culture to develop within an organization, support from top management in formulating decisions based on data is paramount (McAfee & Brynjolfsson, 2012; Mikalef et al., 2017). The importance of having a data-driven culture within an organization is underlined by the findings of Brynjolfsson, Hitt, & Kim (2011) who have found a direct and positive relationship between data-driven decision making and firm performance. This is in line with the findings of McAfee and Brynjolfsson (2012). They found data-driven companies to perform better than non-data-driven companies.

2.2.1.2 Organizational learning

Organizational learning has been described as the “accumulation, sharing, and application of knowledge” (Huber, 1991, p. 90). It is regarded as a dynamic capability (Bhatt & Grover, 2005), which is the ability to reconfigure resources for adapting to changing market conditions (Eisenhardt & Martin, 2000). Eisenhardt and Martin (2000) have extended the RBV with the concept of dynamic capabilities, as the RBV does not explain why businesses are heterogeneous in adapting to rapid environmental changes. Organizational learning provides absorptive capacity which is defined as the ability of a business for acquisition, assimilation transformation and exploitation of knowledge (Cohen & Levinthal, 1990; Zahra & George, 2002). Therefore, businesses with a high intensity of organizational learning should be better able to assemble, integrate and deploy big data related knowledge to help build BDA capability.

2.2.3 Human resources

Human resources include the knowledge, experience, insights and networks of employees working at an organization (Barney, 1991). Based on prior IT literature (Bharadwaj, 2000; Chae, E. Koh, & Prybutok, 2014; Mata, Fuerst, & Barney, 1995), Gupta and George (2016)

(14)

13 argued technical and managerial skills to be crucial human resources for building BDA capability. These will be expanded upon in the following sections.

2.2.3.1 Technical skills

Technical skills refers to the proficiency of utilizing sophisticated data analytics technologies such as Hadoop, NoSQL and NewSQL (Storey & Song, 2017). Employees possessing these analytical skills are needed for extracting insights from big data. However, due to the novelty of the sophisticated data analytics technologies, there seems to be a shortage of these workers (Gallagher, 2015). For this reason, technical skills can provide businesses with a competitive advantage as these employees are valuable and rare. However, this competitive advantage cannot be sustained as education for advanced data analytics is publicly available and therefore not inimitable.

2.2.3.2 Managerial skills

Not only do organizations require technical skills for effectively utilizing IT applications, but having the right managerial skills is also crucial (Mata et al., 1995). Managers must be able to recognize value in insights obtained through data analytics and know how to apply them throughout the organization (Kiron & Shockley, 2011). Managerial skills are firm specific tacit skills that are accumulated over time by individual managers within an organization through experience and trial-and-error (Mata et al., 1995).

2.3 Contingency theory

Between the 1960s and 1980s there has been a lot of interest towards contingency theory, resulting in a large quantity of both empirical and theoretical body of literature on the theory (Van de Ven et al., 2013). Contingency theory describes organizational design and at its core lies the premise that there is no single best way to structure organizations (Lawrence & Lorsch, 1967; Mintzberg, 1989; Van de Ven et al., 2013). Within contingency theory several different theoretical movements have developed over time, called perspectives (Van de Ven et al., 2013). Following the configuration perspective, organizations require internal ‘fit’ as well as a ‘fit’

(15)

14 with the external environment in order to be effective (Burns & Stalker, 1961; Donaldson, 2001; Lawrence & Lorsch, 1967; Van de Ven et al., 2013). An organization has internal fit if there is complementarity among its processes and the internal structure (Miller, 1992). Van de Ven, Ganco and Hinings (2013) describe internal fit as “internal coherence or interdependence among components of a particular design configuration” (p. 414). An organization can obtain external fit by aligning their structure and processes to the external environment (Miller, 1992). For example, Lawrence and Lorsch (1967) found that organizations operating in dynamic environments were most effective if they had a decentralized and informal organizational structure. However, some scholars have criticized the simplistic view of combining internal and external fit for maximizing performance (Child, 2015; Khandwalla, 1973; Miller, 1992). They argue that increasing external fit sometimes comes at the cost of internal fit and vice versa. Thus, increasing internal or external fit is not always an improvement and sometimes internal and external fit must be carefully balanced.

2.4 Business strategy

Business strategy is about how firms compete within industries (Walker & Ruekert, 1987). Two dominant frameworks have emerged within the literature of business strategy. Namely, the Porter typology (1980) and Miles and Snow typology (1978) (Olson, Slater, Hult, & Olson, 2018). Below, both typologies will be explained individually. Then, the hybrid topology of Walker and Ruekert (1987) is introduced, which combines both typologies into one framework.

2.4.1 The Miles and Snow typology

The Miles and Snow typology describes different ways in which businesses approach their product-market domains through adopting specific configurations of organizational design (Slater & Olson, 2000). In these approaches, Miles and Snow (1978) argue that organizational behaviour and structure are not only determined by the external environment, but also influenced through choices of top management. The consideration needed for making these choices are reduced to three general ‘problems’; the ‘entrepreneurial problem’, the ‘engineering

(16)

15 problem’ and the ‘administrative problem’. The entrepreneurial problem is about which market or market segment to enter or develop and how to develop products for these markets. When a solution has been found, the engineering problem arises. This problem associates how a system can be created that operationalizes the solution of the entrepreneurial problem. An appropriate technology has to be selected and new information and control channels have to be formed. Finally, the administrative problem involves organizing the activities that solve the entrepreneurial and engineering problems in such a way that they can be successfully executed with as little uncertainty as possible (Miles et al., 1978).

The strategic typology of Miles and Snow (1978) is based on ways that firms can solve these problems. They distinguish between three strategic types; Prospectors, Defenders and Analysers. Each type has its own market strategy and a corresponding composition of technologies, structure and organization of activities. There is also a fourth strategic type, called the Reactor which is more a residual category. The Reactor is considered a strategic failure with inconsistencies in its strategy (Miles et al., 1978).

2.4.2 The Porter typology

The Porter typology (1980) revolves around external competition and businesses have to decide on the type of competitive advantage and their market scope. Competitive advantage can be obtained through lowering costs or by differentiation to obtain a product that is of higher value to the customers. According to Porter (1980), businesses need to choose a specific market scope by either focussing on a specific market segment, or offering products across different market segments. Broadly speaking, business strategy falls into three generic types; cost leadership, differentiation and focus.

2.4.3 Hybrid typology

The idea of combining both typologies has initially been proposed by Walker and Ruekert (1987). They argue that both the Porter (1980) and the Miles and Snow (1978) typologies have their strengths and weaknesses. The main issue with the Miles and Snow typology is that the

(17)

16 Defender type is broadly defined and heterogeneous with respect to several strategic aspects such as market position in mature markets (Walker & Ruekert, 1987). The generic strategies from Porter (1980) are combined with the Miles and Snow (1978) typology by discriminating between Low-Cost Defenders and Differentiated Defenders. Slater, Olson and Hult promoted this idea by validating it through multiple studies (e.g. Olson, Slater, & Hult, 2005; Slater & Olson, 2000, 2001, 2002; Slater, Olson, & Hult, 2006). Incorporating these different types of Defenders results in the following strategic types: Prospectors, Analysers, low-cost Defenders, differentiated Defenders and Reactors.

Prospectors are primarily characterized by their focus on innovation and market development. They excel in discovering new market opportunities and developing new products. Prospectors need to invest in R&D and in means of scanning the environment for the right market and product opportunities. Managers of these organizations focus heavily on changes in the environment. Also, Prospectors tend to prioritize the successful entry or creation of new markets over increasing in profitability or return on investment (Miles et al., 1978).

Low-cost Defenders focus on profitability by improving efficiency and being cost leaders within the industry. These organizations must excel in efficient production, distribution and process engineering in order to undercut the prices of the competition while maintaining profitability. Economies of scale is often a significant driver for increased efficiency. To accommodate this high efficiency, these organizations limit the number of market segments they operate in and often operate in mature and stable markets (Walker & Ruekert, 1987).

Differentiated Defenders, like the Low-Cost Defenders, also operate in a limited number of market segments which are mature and stable,. They aim to maximize value to their customers. They do not achieve this by lowering prices, but rather by improving product quality (Slater & Olson, 2000). It is very important for these organizations to fit their products to

(18)

17 consumer demand. Therefore, they need to pay close attention to this demand and keep close ties with consumers (Slater & Olson, 2000; Walker & Ruekert, 1987).

Analysers fall right between the strategies of Prospectors and Defenders. “A true Analyser is an organization that attempts to minimize risk while maximizing the opportunity for profit” (Miles et al., 1978, p. 553). In order to do so, the Analyser must find a unique balance between maintaining efficiency and profitability within mature and stable markets and being able to follow Prospectors in new product or market opportunities (Miles et al., 1978)

Reactors are ineffective to uphold a definite strategy. The Reactor strategy is a result of an incorrect enactment of one of the other strategies. Miles and Snow (1978) propose three causes for the existence of this strategic type within organizations. First, top management can be unclear in communicating their proposed strategy downstream. Second, management may not structure the organization in a way to fit a selected strategy. Third, management may fall short in recognizing necessary adjustments in business strategy based on environmental changes.

2.5 Hypotheses development

Based on the body of literature that is described, hypotheses are now drafted for addressing the research questions. The first research question regards the relation between BDA capability and firm performance. For theorizing whether BDA capability leads to increased performance, the RBV can give useful insights as it describes how resources and capabilities can lead to competitive advantage (Barney, 1986b, 1991; Wernerfelt, 1984). Resources build capabilities and can lead to a competitive advantage if they meet the VRIN criteria (Barney, 1991; Grant, 1991). BDA capability is built from seven tangible, intangible and human resources (Gupta & George, 2016).

Tangible resources comprise data, technology and basic resources. Data that has a large volume, velocity and variety is regarded as big data and forms the cornerstone for BDA

(19)

18 (Davenport, 2014; McAfee & Brynjolfsson, 2012). It can provide valuable insights and assist in substantial value creation across many industries (Manyika et al., 2015; Morabito, 2015; Zhao et al., 2014). For businesses to access the insights from big data, the data has to be mined and analysed (Y. Chen et al., 2014), requiring sophisticated data analytics technologies (Storey & Song, 2017). Finally, BDA projects require sufficient funding and time in order to develop and succeed. Gupta and George (2016) refer to time and capital as basic resources. Tangible resources often do not amount to a competitive advantage as they can be freely bought on markets and thus are not VRIN (Barney, 1991). Data is becoming widely available as a result of global digitalization and the Internet of Things (Gubbi et al., 2013) and BDA technologies can be bought on an ever growing market (IDC, 2017). Thus, even though tangible BDA resources are required for building BDA capability, they will amount to a competitive advantage by themselves.

Intangible resources comprise data-driven culture and organizational learning. Having a data-driven culture within an organization requires decisions throughout the organization to be made on the basis of analytical evidence instead of intuition. It is paramount for organizations in order to effectively utilize the insights provided by BDA (Davenport et al., 2001; McAfee & Brynjolfsson, 2012; Mikalef et al., 2017; Ross et al., 2013). Organizational learning provides absorptive capacity (Cohen & Levinthal, 1990; Zahra & George, 2002) and thus enables firms to assemble, integrate and deploy big data related knowledge to help build BDA capability (Gupta & George, 2016). Most intangible resources cannot be bought on markets and need to be built within an organization instead (Dierickx & Cool, 1989; Hall, 1992). They often meet Barney’s (1991) VRIN criteria and thus amount to a competitive advantage (Teece, 2014). This seems to be true for data-driven culture because it evolves over time within an organization (Kiron & Shockley, 2011) as well as for organizational learning as there is heterogeneity across firms in terms of organizational learning (Bhatt & Grover, 2005).

(20)

19

Human resources comprise managerial and technical skills. For effective utilization of BDA, managers must possess the right skills for recognizing value in insights obtained through BDA and must know how to apply these throughout organization (Kiron & Shockley, 2011). Technical skills refer to the proficiency of utilizing sophisticated data analytics technologies such as Hadoop, NoSQL and NewSQL for analysing big data (Storey & Song, 2017). Although human resources can often be bought on the labour market, managerial skills are highly firm specific and need to be accumulated over time through experience and trial-and-error (Mata et al., 1995). Thus, managerial skills meet Barney’s (1991) VRIN criteria. Technical skills are valuable and rare as there seems to be a shortage of these workers due to the novelty of the sophisticated data analytics technologies (Gallagher, 2015). However, they are not inimitable as education for advanced data analytics is publicly available.

All seven resources are essential for building BDA capability and several of them meet Barney’s (1991) VRIN criteria. Thus, BDA capability is a VRIN capability and can provide firms with a competitive advantage which should result in an increased relative firm performance (Barney, 1991; Grant, 1991). Prior research has already hinted towards a positive relation between BDA capability and firm performance (Akter et al., 2016; Gupta & George, 2016). These findings do not seem controversial within the scientific literature as a positive relation between several IT capabilities and firm performance has a wide empirical basis (Y. Chen et al., 2014; Dale Stoel & Muhanna, 2009; Melville, Kraemer, & Gurbaxani, 2004). Thus, the following hypothesis is drafted:

H1. There is a positive relation between big data analytics capability and both operational and market performance.

For addressing research question 2, the effect of business strategy on the relation between BDA capability and firm performance needs to be examined. The RBV seems to be unable to explain differences caused by general strategic types, as general strategy is by no

(21)

20 means rare and thus cannot amount to a competitive advantage (Barney, 1991). From the perspective of contingency theory however, the influence of business strategy can be explained on the basis of internal and external ‘fit’ (Burns & Stalker, 1961; Donaldson, 2001; Lawrence & Lorsch, 1967; Van de Ven et al., 2013). The increasing importance and value of big data can be regarded as a trend in the external environment of an organization. Building BDA capability within an organization will increase its ability to utilize this big data and therefore increase external fit. However, because external fit can sometimes come at the cost of internal fit (Child, 2015; Khandwalla, 1973; Miller, 1992), BDA capability might not necessarily strengthen the internal fit of all organizations adopting it.

Due to the structural differences among organizations pursuing different business strategies (Miles et al., 1978; Porter, 1980; Walker & Ruekert, 1987), it is expected that BDA capability might not fit equally well within the internal structure and processes among these organizations. Thus, there is likely to be heterogeneity in firm performance obtained from BDA capability among businesses based on their strategy.

To devise for which strategic types BDA capability has a better internal fit, it is useful to reflect on the definition of BDA capability by Gupta and George (2016): “a firm’s ability to assemble, integrate, and deploy its big data-specific resources” (p. 1049). For BDA capability to fit the internal structure and processes of an organization, assembly, integration and deployment of big data-specific resources should be complementary to the structure and processes (Miller, 1992). These are aligned with the business strategy for firms that fit the criteria of Walker and Ruekert (1987). Thus, for these organizations BDA capability has internal fit within the organization if it also aligns with business strategy.

Prospectors tend to operate in a more dynamic environment and are continuously trying to find and exploit new product and market opportunities (Miles et al., 1978). BDA capability contributes towards this strategic goal as big data enables businesses to create new products or

(22)

21 enhance existing ones. Also, it assists firms in exploiting new market opportunities by enabling them to invent new business models by providing insight into environmental conditions, trends and events (Manyika et al., 2011). A positive effect on firm performance is expected on the basis of this fit between the business strategy of Prospectors and BDA capability. However, as Prospectors tend to prioritize innovation over profitability (Miles et al., 1978), only an increase in market performance is expected as a result of BDA capability without a significant increase of operational performance.

Differentiated Defenders aim towards maximizing product value for their customers and therefore their willingness to pay. This enables them to increase profits by increasing product prices (Walker & Ruekert, 1987). Inherent to maximizing product value to customers is to know what it is that they value. BDA capability contributes to this strategic goal as big data can provide for better segmentation of consumer demands and utilize this to more accurately tailor products to their desire (Manyika et al., 2011). Thus, BDA capability assists Differentiated Defenders in introducing superior products or altering existing products to improve both market and operational performance. A positive effect on firm performance is expected on the basis of this fit between the business strategy of Differentiated Defenders and BDA capability

Low-Cost Defenders strive towards costs minimization of their products in a limited amount of markets. Minimizing costs is mainly done through increasing economies of scale and operational efficiency (Walker & Ruekert, 1987). Although data analytics has proven useful in optimizing operational efficiency by analysing internal data, the emphasis of BDA lies in the analysis of external data and therefore BDA capabilities might not contribute much with the needs of Low-Cost Defenders. (Zhao et al., 2014). Therefore, due to limited internal fit with the strategic goal of Low-Cost Defenders, no positive effect on firm performance is expected due to BDA capability.

(23)

22

Analysers balance traits of both Defenders as Prospectors as pursue high risk product and market development while maintaining core products and consumers (Miles et al., 1978). Since there seems to be low internal fit between the strategy of Low-Cost Defenders and BDA capability and higher internal fit with the strategy of Prospectors and Differentiated Defenders, it is possible for Analysers to obtain internal fit as long as BDA capability is integrated correctly.

Reactors are not likely to obtain internal fit with any particular capability as these businesses do not have a coherent strategic goal to begin with (Miles et al., 1978). Thus, no increase in performance is expected as a result of BDA capability. Based on the differences of internal fit between BDA capability and strategic goals among strategic types, the following hypotheses are drafted:

H2. The relation between big data analytics capability and market performance is moderated by business strategy such that: for Prospectors, Analysers and Differentiated Defenders the relation between big data analytics capability and market performance is more positive than for Low-Cost Defenders and Reactors

H3. The relation between big data analytics capability and operational performance is moderated by business strategy such that: for Analysers and Differentiated Defenders the relation between big data analytics capability and operational performance is more positive than for Low-Cost Defenders and Reactors.

The conceptual model of all concepts and their reciprocal hypothesized relations is depicted in figure 1.

(24)

23 Figure 1

Conceptual model

Note: P - Prospector, A= Analyser, DD= Differentiated Defender, LD= Low-Cost Defender, R= Reactor.

3 Methodology

In this chapter, the research method will be expanded upon. First, the decision for choosing survey as a means of collecting data is substantiated and the design of the survey is described. Then, descriptions of the measures used in the research model are given. This is followed by an explanation of the control variables and the software used for data analysis. Finally, relevant information about the sample and data collection process is provided.

3.1 Survey design

Due to a limited amount of time and funds available for carrying out this research, the survey research method has been selected due to its cost and time effectiveness. A possible alternative to this would be the interview research method. However, this would greatly reduce the sample size and thus the external validity of the study, limiting the possibility for significant findings. The online survey consists of 45 different items. In order to prevent for missing data to occur, the survey items cannot be skipped by the respondents (except the control variables). The first question of the survey filters out employees that work at firms that did not invest in big data analytics at least six months ago. Also, to reduce the effect of common method bias, the order of the survey items have been randomised (Chang, Witteloostuijn, & Eden, 2010).

(25)

24

3.2 Measures

The dependent variable BDA capability is a three dimensional formative construct developed by Gupta and George (2016). It encompasses big data specific tangible, intangible and human resources. The survey items of the construct are listed in appendix 1. They are measured using a seven point Likert scale. Their validity and other characteristics from Gupta and George (2016) are depicted in table 1. The first order constructs Data, Technology, Basic Resources, all second order constructs (Tangible, Intangible and Human resources) and the third (BDA capability) order constructs are all formative. Formative constructs are formed by a linear combination of causal indicators (Bollen & Bauldry, 2011). The concept of the causal indicator is introduced by Bollen and Lennox (1991) and describes indicators which do not reflect their corresponding construct, but define it. The three dimensional formative construct BDA capability is formed using the repeated indicator approach, meaning that all indicators (measured by individual scale items) weigh on both their corresponding first order, second order constructs as well as the third order construct (Gupta & George, 2016). A visual representation of the three dimensional BDA capability construct is shown in Appendix 4. Dimensionality of the indicators and constructs (appendix 1) are assumed the basis of the work of Gupta and George (2016).

The dependent variable firm performance is a reflective construct that has been validated by Ravichandran and Lertwongsatien (2005). Based on the previous literature on firm performance this study describes it as a combination of a firm’s Operational and Market Performance (Rai, Patnayakuni, & Seth, 2006). Operational Performance includes financial results such as sales, profit and return on investment. Market Performance describes a firm’s success in entering new markets and introducing new products or services (Gupta & George, 2016; Rai et al., 2006; Ravichandran & Lertwongsatien, 2005; Wang, Liang, Zhong, Xue, & Xiao, 2012). The survey items of the construct are listed in appendix 2. Firm performance is

(26)

25 Table 1

Characteristics of first order BDA capability constructs (Gupta & George, 2016).

Construct α

Formative/

reflective Origin of scale items

Data - Formative (Davenport, 2014)

Technology - Formative (Davenport, 2014; Gordon, 2007)

Basic Resources - Formative (Davenport, 2014; Wixom & Watson, 2001) Data-Driven Culture 0,86 Reflective (McAfee & Brynjolfsson, 2012; Ross et al., 2013) Organizational Learning 0,92 Reflective (Bhatt & Grover, 2005)

Managerial Skills 0,87 Reflective (Davenport, 2014; Mata et al., 1995)

Technical Skills 0,89 Reflective (Carmeli & Tishler, 2004; Mata et al., 1995; Wixom

& Watson, 2001)

regarded as relative performance in comparison with competitors, as this fits within the framework of the RBV that revolves around competitive advantage. Each type of performance is measured with 4 scale items using a 7 point Likert scale. Ravichandran and Lertwongsatien (2005) established the validity of the measure with a Cronbach’s Alpha of 0.76.

The moderating variable business strategy is a categorical variable, measuring the strategy type using the ‘Self typing paragraph’ and has been validated by Conant, Mokwa and Varadarajan (1990) with a Cronbach’s α of 0.75. It consists of one question in which the participant has to choose which one out of five paragraphs best describes the firm’s strategy (Appendix 3). Each answer describes one of the five hybrid strategic typologies by Walker & Ruekert (1987).

3.3 Control variables

Three control variables are included to control for; firm size, industry and big data analytics experience. These control variables should be accounted for as they could have a significant impact on the relationship between BDA capability and firm performance. Firm size could influence this relationship as it impacts overall organizational practices and firm performance (Baum & Wally, 2003). Industry could affect the relation between BDA capability and firm performance as advantages obtained through BDA capability could be more or less valuable depending on the industry. For example, insights obtained through BDA capability could be more valuable in highly dynamic and uncertain industries than in industries that are more stable.

(27)

26 Finally, the amount of experience an organization has with BDA could influence its ability to align it with organizational processes, especially if this is done by trial and error.

3.4 Structural equation modeling

In this study structural equation modeling (SEM) is used for analysis of the data. SEM is combination of different statistical methods and can be used to determine the relationship between independent variable(s) and dependent variable(s) (Ullman & Bentler, 2012). In this study, the partial least squares structural equation modeling (PLS-SEM) method is used. The objective in PLS-SEM is to maximize the variance of dependent latent constructs explained by the causal model (Hair, Sarstedt, Hopkins, & Kuppelwieser, 2014). Many researches use covariance based structural equation modeling (CB-SEM) instead of SEM. Although PLS-SEM has been subject of criticism, it is gaining popularity in field of business research as it is superior to CB-SEM in dealing with small samples and constructs with few indicators (Hair et al., 2014).

3.5 Sample and data collection

The sample used in this study is a subset of the target population of European firms that have already invested in BDA. To obtain data about these firms, the online survey tool has been sent to managers that are employed at these firms and manage their BDA. These people are likely to have sufficient knowledge to answer the survey questions in a meaningful way. A non-probability sampling method has been used for selecting the BDA managers. 1050 LinkedIn members with the terms ‘big data analytics’ and ‘manager’ in their profile description were contacted by email. After the initial request, two reminder emails were sent, each one being one week apart. Email addresses were obtained through LinkedIn using RocketReach software. Since this study is focused on differences among firms, each firm could only be represented once in the dataset and therefore a maximum of one employee per firm is added to the sample. Only LinkedIn members living in one of the following countries were considered: Denmark,

(28)

27 the Netherlands, Belgium, France, Finland, Spain, Norway, Switzerland, Ireland, Sweden and Germany. A total of 170 responses were collected, 41 of which did not pass the filter question. Another 26 did pass the filter but did not fill in any answers beyond that point. 36 respondents answered some questions but stopped before reaching the end of the survey. In total, 67 complete responses were collected.

4 Results

In this chapter, the data gathered with the survey is evaluated. First, the data is analysed for unengaged responses, normality and other characteristics. Then, the measurement model is evaluated on its validity, reliability and common method bias. Finally, the structural model is evaluated by checking the validity of the hierarchical model and then testing the hypotheses.

4.1 Data analysis

Because respondents could not skip the survey items, no values are missing in the dataset. However, some responses seem fairly unengaged. These are identified using two methods. First, a cut off time is established. Because deciding on a cut off time lacks scientific rigor, a very conservative minimum time of 2 seconds per item is used (Huang, Curran, Keeney, Poposki, & DeShon, 2012). It is seems very difficult to argue that respondents answered the survey items in a meaningful way if they took less than two seconds for doing so. With a total of 70 survey items, the minimum time of filling in the survey should be 140 seconds. 2 responses were excluded following this approach.

Consequently, a minimum threshold is established for the variance among all data per respondent. Because this method also lacks scientific rigor, a very conservative lower variance threshold of 0,2 is set. Responses falling below this threshold are almost certainly unengaged as virtually all answers were identical. 2 responses were excluded following this approach, leaving a total of 63 responses for further analysis. The characteristics of this sample is depicted in table 2.

(29)

28 Table 2 Sample characteristics N=63 Industries Computer/Software 24% Manufacturing 10%

Finance, Insurance, Real estate 14%

Retail, Wholesale 3%

Services 13%

Healthcare 6%

Others, (e.g. Transportation, HR, Media) 30%

Total BDA experience

Less than 3 years 29%

3-6 years 40%

More than 6 years 32%

Number of employees in the organization

Fewer than 100 3% 101-250 6% 251-500 2% 501-1000 43% 1001-2500 6% 2501-5000 32% More than 50000 8% Strategic type Prospector 40% Low-Cost Defender 14% Differentiated Defender 11% Analyser 24% Reactor 11%

The skewness and excess kurtosis of the data per variable is depicted in Appendix 5. Excess kurtosis and skewness is deemed acceptable if it has a value between -2 and 2 (Trochim & Donnely, 2001). The variables DD1, DD2, OL1 and OL4 have excess kurtosis above 2 and therefore may lack variance. In the next section, it will be assessed if this causes problems during factor analysis.

4.2 Evaluation of the measurement model

The measurement model will now be evaluated regarding its construct validity and reliability. The validity gives the extent to which things that are intended to be measured, are actually measured. The reliability is about repeatability of measurements. In other words, it gives the

(30)

29 likelihood that repetition of the measurement yields the same results. Below, reflective construct validity and reliability as well as formative construct validity is discussed.

4.2.1 Construct validity of reflective indicators

For establishing construct validity, both convergent and divergent validity should be established. Convergent validity is established by checking whether the reflective indicators significantly load on their theoretical constructs. They should have loadings of at least 0,7 (Hair, Hult, Ringle, & Ciambotti, 2013). Firstly, after bootstrapping the model, indicator loadings are highly significant (p<0,01). Secondly, for assessing the factor loadings a confirmatory factor analysis is performed (Appendix 6). Except for TS1 and TS2, each factor loads on their respective constructs with loadings above 0,70. TS1 and/or TS2 should only be removed as long as doing this increases the composite reliability of the Technical Skills construct (Hair et al., 2013). As only the removal of TS1 increases reliability, it is removed from the model.

Discriminant validity of the reflective indicators is established using both Fornell and Larcker’s (1981) criterion and the heterotrait-monotrait ratio (HTMT), brought forward by Henseler et al. (2015). The square root of the AVE of each reflective construct is greater than the correlations with any other construct (table 4) and thus Fornell and Larcker’s (1981) criterion is satisfied. The HTMT ratio uses correlations between reflective constructs to establish discriminant validity; all values should be below 0,85. Indeed, HTMT ratios turn out to be below 0,85 (table 3). To further establish discriminant validity of the reflective indicators, the average variance extracted (AVE) is determined. The AVE of each reflective construct exceeds 0,5 (table 4), confirming discriminant validity for these items (Hair et al., 2014). As both convergent and discriminant validity have been established for the reflective constructs, construct validity is confirmed.

4.2.2 Reliabilities of reflective constructs

Reliability of the reflective constructs is assessed using composite reliability and Cronbach’s (1951) α. Cronbach (1951) describes α as “an estimate of the correlation between two random

(31)

30

samples of items from a universe of items like those in the test” (p. 297). However, scholars argue that Cronbach’s α has several limitations (Shook, Ketchen, Hult, & Kacmar, 2004). Shook et al. (2004) argue that composite reliability is a superior choice as it uses factor loadings as well as measurement error for each item. The composite reliability and α of all reflective constructs above 0,80 (table 4), signifying great reliability (Hair et al., 2013; Nunnally, 1978; Shook et al., 2004). However, a composite reliability higher than 0,95 is not desired as it indicates significant overlap between survey items which may diminish the measures’ content validity (Hair et al., 2013). The construct Managerial Skills has a composite reliability of 0,96 and therefore needs inspection of potential overlap of its underlying survey items. MS1 and MS4 seem to be very similar (Appendix 1). Removing MS4 reduces the composite reliability more than if MS1 were removed and brings it down below 0,95. Therefore, MS4 is removed from the model.

4.2.3 Formative construct validity

Means for determining reflective construct validity cannot be used for formative constructs (Lowry & Gaskin, 2014; Petter, Straub, & Rai, 2007; Straub, Boudreau, & Gefen, 2004). Traditionally, scholars have not used any tools other than theoretical reasoning to argue for validity of formative constructs (Diamantopoulos & Siguaw, 2006; Lowry & Gaskin, 2014).

However, new approaches that test construct validity are surfacing (Lowry & Gaskin, 2014; Petter et al., 2007). A basic approach for assessing the construct validity of formative indicators is by simply assessing if the indicator weights are significant (Lowry & Gaskin, 2014; Ringle, Sarstedt, & Straub, 2012). After bootstrapping, the following indicators on their respective theorized first-order construct turn out to be non-significant (p > 0,05): D3, T1, T2, T3 and T5. Before deciding on which indicators to drop, the potential causes for this non-significance should be assessed (Cenfetelli & Bassellier, 2009). One cause for insignificant weights can be a large number of formative indicators for one construct. If this is the case, formative indicators with non-significant weights can be kept, as long as their contribution to

(32)

31 Table 3

HTMT values of reflective constructs

Table 4

Reliability and validity measures of reflective constructs and inter-correlations of all first-order constructs1

1_{Square root of the AVEs on the diagonal.}

the formative construct can be argued for (Cenfetelli & Bassellier, 2009). The Technology construct has many indicators and each item captures a distinctive data-related technology (Gupta & George, 2016). Thus, its non-significant indicators are not removed from the model. However, the indicator T3 is an exception as it is also non-significant in both the pilot study and the main study by Gupta and George (2016). When indicators are repeatedly non-significant across multiple studies this is evidence against their inclusion (Cenfetelli & Bassellier, 2009). Thus, T3 is removed from the model. The indicator T5 also has a slightly negative weight besides being non-significant. While this is surprising, it does not grant sufficient reason to remove it as long as it is not multicollinear (Cenfetelli & Bassellier, 2009). Multicollinearity is tested for by checking if the variance inflation factor (VIF) does not exceed 3,33 (Diamantopoulos & Siguaw, 2006). T5 has a VIF of 2,22 and is therefore not removed from the model. Finally, the Data construct has only three indicators and the reasoning used for

Data-Driven Culture Managerial Skills Market Performance Operational Performance Organizational Learning Technical Skills Data-Driven Culture Managerial Skills 0,56 Market Performance 0,47 0,54 Operational Performance 0,51 0,41 0,78 Organizational Learning 0,63 0,56 0,57 0,35 Technical Skills 0,60 0,56 0,35 0,19 0,44 Items CR α AVE 1 2 3 4 5 6 7 8 9 Data NA NA NA NA Tech NA NA NA 0,62 NA BR NA NA NA 0,49 0,47 NA DDC 0,88 0,83 0,60 0,40 0,36 0,33 0,77 OL 0,93 0,91 0,74 0,55 0,48 0,40 0,55 0,86 TS 0,91 0,87 0,67 0,30 0,36 0,30 0,51 0,41 0,82 MS 0,94 0,93 0,78 0,39 0,21 0,45 0,52 0,52 0,52 0,77 MP 0,89 0,83 0,67 0,30 0,15 0,50 0,40 0,51 0,31 0,49 0,82 OP 0,92 0,88 0,73 0,20 -0,01 0,19 0,44 0,32 0,17 0,39 0,65 0,85

(33)

32 retaining non-significant indicators of large formative constructs cannot be applied. Therefore, due to non-significance, D3 is removed from the model.

The validity of the formative first order constructs are corroborated using Edwards’ (2001) adequacy coefficient (R2a) and by checking for multicollinearity suggested by Petter et

al. (2007). R2a represents the variance in the first order construct that is explained by its

indicators and is calculated through summing the squared correlations between the indicators and their formative constructs and then dividing by the number of indicators (Edwards, 2001; MacKenzie, Podsakoff, & Podsakoff, 2011). All R2a values exceed 0,50 (table 5) further

establishing validity of the formative indicators (MacKenzie et al., 2011). Furthermore, multicollinearity is tested for by checking if the VIF of each indicator is below the threshold of 3,33 (Diamantopoulos & Siguaw, 2006). Each indicator meets this criteria (table 5) which further establishes validity of the formative constructs (Diamantopoulos & Siguaw, 2006).

For establishing the validity of the higher order formative constructs, the same method is applied as for the first order constructs. Multicollinearity is ruled out as the VIFs are all below 3.33. However, the majority of the indicators weigh non-significantly on the higher order constructs. Also, most of the Edwards’ (2001) adequacy coefficients do not exceed the threshold of 0,50 (table 5). These problems are likely to be caused by the small sample size and addressing them would result in removing the majority of the model. However, Gupta and George (2016) used the same measures and did establish the validity of the reflective constructs using a significantly larger sample. In combination with the theoretical underpinning of each formative construct from chapter 2 and a lack of consensus among the scientific community on establishing formative construct validity (Diamantopoulos & Siguaw, 2006; Lowry & Gaskin, 2014), validity is assumed for the higher order formative constructs regardless.

4.2.4 Common method bias

After establishing construct validity and reliability, the level of common method bias is checked. As each measurement for the BDA capability and performance constructs is taken at

(34)

33 Table 5

Formative construct validation.

Construct Measures Weights Significance VIF R2

a Data D1 0,66 p < 0,01 1,14 0,68 D2 0,56 p < 0,01 1,14 Technology T1 0,50 ns 3,18 0,50 T2 0,12 ns 1,38 T4 0,73 p < 0,01 1,80 T5 -0,08 ns 2,25 Basic Resources BR1 0,65 p < 0,05 1,47 0,78 BR2 0,48 p < 0,05 1,47 Tangibles Data 0,61 p < 0,01 1,78 0,31 Technology -0,24 p < 0,01 1,74 Basic Resources 0,69 p < 0,01 1,39

Intangibles Data-Driven Culture 0,52 p < 0,01 1,44 0,50

Organizational Learning 0,60 p < 0,01 1,44

Human Technical Skills 0,19 p < 0,01 1,38 0,22

Managerial Skills 0,87 p < 0,01 1,38

BDA capability Tangibles 0,21 p < 0,01 1,60 0,34

Intangibles 0,58 p < 0,01 1,99

Human 0,30 p < 0,01 1,90

the same time, by the same respondent and using the same scales, common method bias possibly distorts the measurements and therefore should be tested for. Two methods are used in order to test common method bias following the approach of Lowry and Gaskin (2014). First, Harman’s single factor test is applied to an exploratory and unrotated factor analysis. This test assesses if the majority of the variance in the model can be explained through one single factor. The factor analysis produced 8 different factors. The largest factor only explained 32,54% of the variance of the model, suggesting that common method bias is not a concern in this dataset. The second method for assessing common method bias is through a Pearson’s correlation matrix. The correlations among formative indicators are below 0,90 which also signifies that common method bias is not likely to be an issue in this dataset (Pavlou, Liang, & Xue, 2007).

4.3 Evaluation of the structural model

Now that validity and reliability of individual constructs has been established, validity of the hierarchical model will be assessed. Then, the hypotheses will be tested.

(35)

34

4.3.1 Hierarchical model evaluation

The hierarchical model is assessed by evaluating the weights and significance levels of the first-order constructs on the second-first-order constructs and those of the second-first-order constructs on the third order construct (table 5). After bootstrapping, the weights of the first-order constructs Technology and Technical Skills on their respective second-order constructs turn out to be non-significant. However, the construct Technical Skills is not removed from the model because, 1. validity and reliability of the construct have been established, 2. validity of the hierarchical model has been established twice by Gupta and George (2016) and 3. removal of the item does not increase significance of the weight of its corresponding second-order construct on BDA capability.

However, the weight of Technology on the second-order Tangible Resources is not only non-significant, but also negative (β= -0,24). Negative formative indicators that are intended to be positive (Gupta & George, 2016) can lead to severe misinterpretation of the results (Cenfetelli & Bassellier, 2009). Two possible causes for this problem can be multicollinearity and suppression effects (Cenfetelli & Bassellier, 2009). Earlier in this study, multicollinearity has already been ruled out by establishing that all VIF values are below 3.33 (table 5). Cenfetelli and Bassellier (2009) state that “suppression effects occur when one of the predictor variables explain significant variance in other predictor variables not otherwise associated with the criterion” (p. 696). Indicators require a significant negative weight in order to qualify as suppressors. As the weight of the indicator Technology is negative but not significant, it is not a suppressor. As the indicator neither has multicollinearity nor suppressor effects, it should not be excluded from the model (Cenfetelli & Bassellier, 2009).

4.3.2 Hypotheses testing

Before testing the hypotheses, the categorical control variables are coded as dummy variables and added to the model. However, due to the small sample size and the large amount of categories within the firm size and industry variables, SmartPLS fails to run the model. To

The effect of business strategy on the relation between big data analytics capability and firm performance