Building novel capabilities to enable business intelligence agility: results from a quantitative study

Abstract

The class of business intelligence (BI) systems is widely used to guide decisions in all kinds of organizations and across hierarchical levels and functions. Organizations have launched many initiatives to accomplish adequate and timely decision support as an important factor to achieve and sustain competitive advantage. Given today’s turbulent environments it is increasingly challenging to bridge the gap between establishing a long-term strategy and quickly adopting to the dynamics in market competition. BI must address this field of tension as it was originally used to retrospectively reflect an organization’s performance and build upon stability and efficiency. This study aims to understand and achieve an agile BI from a dynamic capability perspective. Therefore, we investigate how dynamic BI capabilities, i.e., adoption of assets, market understanding, and intimacy as well as business operations, impact the agility of BI. Starting from a literature review of dynamic capabilities in information systems and BI, we propose hypotheses to connect dynamic BI capabilities with the agility to provide information. The derived hypotheses were tested using partial least squares structural equation modeling on data collected in a questionnaire-based survey. The results show that the lens of dynamic capabilities provides useful means to foster BI agility. The study identifies that technological advancements like in-memory technology seem to be a technology enabler for BI agility. However, an adequate adoption and integration of BI assets as well as the focus on market orientation and business operations are crucial to reach BI agility.

Appendices

Appendix A: Survey measures

Table 10 contains the items of our constructs and literature sources. All items are presented on a 7-point Likert scale. First, the rating about the importance in general is asked for the item. Second, the participants are asked to rate their organization.

Table 10 Survey measures

We are well aware of the discussion about the pros and cons of “don’t know responses” (Beatty et al. 1998; Dillman et al. 2009). As the respondents’ best subjective estimation adds value to our analysis, and to avoid missing values, we used mandatory questions in the survey.

Appendix B: Survey statistics

Table 11 illustrates the work experience of the participants. Fifty-three percent have more than 10 years of work experience. Further information about the participants, such as industries or consultant rate, is shown in Table 12.

Table 11 Work experience of participants

Table 12 Industry and consultant distribution of participants

Appendix C: Common method bias

Common method bias (CMB) is a potential problem in survey research. It exists if a significant amount of spurious variance is attributed to the measurement or data collection method rather than to the constructs that the measures represent. Ex-post statistical analysis, e.g., correlation analysis, helps to identify common method variance in the collected data (Chen and Siau 2012; Urbach and Ahlemann 2010; Podsakoff et al. 2003).

According to Pavlou et al. (2007), Chen and Siau (2012) and Bagozzi et al. (1991) correlation analysis is a useful means to determine common method bias. A high correlation among the main constructs of a model of 0.9 and above (Pavlou et al. 2007; Chen and Siau 2012) indicates the evidence of common method bias. The highest correlation among the constructs in this study is 0.73 for the general view (see Table 13) and 0.83 for the organization-specific implementations (see Table 14). This suggests that no common method bias exists. The sample size is n = 110.

Table 13 Correlation among main constructs (general view)

Table 14 Correlation among main constructs (organization-specific implementations)

Appendix D: Model validation (main study): general view

We conducted a series of tests to analyze the validation of our outer measurement model (n = 110).

4.1 Indicator reliability

Indicator reliability describes the extent of consistency in what an item (or a variable or a set of variables) measures and what it intends to measure. Indicator reliability can be assessed by the standardized outer loading of an item. Indicator loadings should be significant at the 0.05 level and should exceed 0.7 or 0.5 if the squared indicators are used (Urbach and Ahlemann 2010; Hair et al. 2014; Henseler et al. 2009). For exploratory research with newly developed items lower thresholds have been proposed. Hulland (1999) suggests dropping an item if the loading is below 0.4 or 0.5. But, eliminating reflective indicators has been done with care and only if the reliability of the item is low and the elimination substantially increases the composite reliability (Henseler et al. 2009). Table 15 contains the item loadings as well as the t values and p levels.

Table 15 Item loadings and cross-loadings

4.2 Internal consistency reliability

We computed two criteria to assess internal consistency reliability, Cronbach’s alpha (CA) and composite reliability (CR). A high value for CA assumes that the correlation of a set of items within a construct is a good estimate for the correlation of all items within this construct (item inter-correlation) and that the items have the same meaning and range (Urbach and Ahlemann 2010; Henseler et al. 2009). CA is said to have some limitations. Therefore, we also considered composite reliability (CR) as suggested by Hair et al. (2014). Both, CA and CR should exceed the threshold of 0.7 and must not be lower than 0.6 (Urbach and Ahlemann 2010). While Nunnally and Bernstein (2010) postulate a CA of 0.95 as a desirable standard in specific cases, Straub et al. (2004) consider values above 0.95 to be suspicious. As our constructs are newly developed, we follow the threshold of 0.70 for “early stage” research of CA and CR (Nunnally and Bernstein 2010; Henseler et al. 2009). The results of the tests for internal consistency reliability are shown in Table 4.

4.3 Convergent validity

A commonly accepted criterion of convergent validity is the average variance extracted (AVE) proposed by Fornell and Larcker (1981). It measures the average amount of variance which a construct captures from its indicators in relation to the amount of measurement error. In other words, convergent validity tests whether an item measures the construct that it is supposed to measure. An AVE value of 0.5 or higher is deemed acceptable as it indicates that a construct explains more than half of the variance of its indicators (Urbach and Ahlemann 2010; Hair et al. 2014). Table 4 contains the AVE values.

4.4 Discriminant validity

Discriminant validity describes the extent to which a construct is distinct from other constructs. If discriminant validity is established, it implies that a construct is unique and items of a construct do not unintentionally measure constructs they are not supposed to measure. Discriminant validity is established to be assessed in two ways (Hair et al. 2014; Urbach and Ahlemann 2010). One way is to examine cross-loadings of indicators, and the other approach is to use the Fornell–Larcker criterion (Fornell and Larcker 1981). If cross-loadings are used, each loading of an indicator should be higher for its designated construct than for other constructs and each of the construct’s highest item is among its own items compared to the corresponding constructs’ items (Urbach and Ahlemann 2010). Gefen and Straub (2005) suggest 0.1 as a threshold for the difference between own loading and loadings on other constructs.

The Fornell–Larcker criterion postulates a construct to share more variance with its indicators than with other constructs. It is fulfilled if the square root of a construct’s AVE is higher than the correlation with any other construct.

Recently, Henseler et al. (2015) introduced a new criterion for assessing discriminant validity, the heterotrait-monotrait (HTMT) ratio of correlations criterion. Values that meet the threshold of 0.9 are deemed acceptable (Henseler et al. 2015; Gold et al. 2001). The analysis for assessing discriminant validity with HTMT was done using SmartPLS 3 (Ringle et al. 2015).

The cross-loadings are shown in Table 15, whereas the discriminant validity check according to Fornell–Larcker is depicted in Table 4. The HTMT ratio values are shown in Table 16.

Table 16 Discriminant validity (HTMT ratio)

Tables 17, 18 show the path coefficients, indirect effects and the corresponding significance.

Table 17 Path coefficients and significance

Table 18 Indirect effects and significance

Appendix E: Model validation (main study): organization-specific view

Tables 19, 20, 21, 22 show the validity assessments for the second part of our study, i.e., the specific implementations of BI at the organizations (n = 110).

Table 19 Item loadings and cross-loadings

Table 20 Discriminant validity (HTMT ratio)

Table 21 Path coefficients and significance

Table 22 Indirect effects and significance

Appendix F: Model validation (main study): impact of in-memory experience

The group “non-experienced IM organizations” contains participants that work at organizations using no IM technology for BI at all. For this group (n = 31) we analyzed their expectation, i.e., general view (Fig. 7).

Fig. 7

Hypotheses test for non-experienced IM organizations (general view), n = 31. ns not significant; ***significant at p < 0.01; **significant at p < 0.05; *significant at p < 0.1 (one-tailed tests)

The group “experienced IM organizations” contains participants that work at organizations that have been using IM technology for BI for at least three years. We took the organization-specific view (Fig. 8) for this group (n = 39).

Fig. 8

Hypotheses test for experienced IM organizations (organization-specific view), n = 39. ns not significant; ***significant at p < 0.01; **significant at p < 0.05; *significant at p < 0.1 (one-tailed tests)

Appendix G: Model validation (pre-study): results of scale development validation

Tables below contain the results of the outer measurement model validation of the preliminary study (n = 16). Tables 23 and 24 show the validation of general view, whereas Tables 25 and 26 consider the statements regarding organization-specific implementations. We assessed internal reliability using Cronbach’s alpha and composite reliability (see Tables 24, 26). All values are above the proposed threshold of 0.7 for both indicators (Urbach and Ahlemann 2010). We also checked for convergent reliability with average variance extracted (AVE). Tables 24 and 26 show that the criterion is met as all values exceed 0.5 (Urbach and Ahlemann 2010). We used the Fornell–Larcker criterion to check discriminant validity. The criterion is missed in one of six cases (Tables 24, 26). Tables 23 and 25 show the item and cross-loadings.

Table 23 Items and cross-loadings (general view)

Table 24 Internal consistency and convergent reliability (general view)

Table 25 Items and cross-loadings (organization-specific view)

Table 26 Internal consistency and convergent reliability (organization-specific view)

Appendix H: Control variable: size of organization

We assessed the robustness of the research model by using the size of organization, i.e., the number of employees, as control variable (n = 110). This analysis is shown in Table 27 and Fig. 9 (general view) and Table 28 and Fig. 10 (organization-specific view).

Table 27 Path coefficients and significance incl. control variable (general view)

Table 28 Path coefficients and significance incl. control variable (organization-specific view)

Fig. 9

Hypotheses test incl. control variable (general view). ns not significant; ***significant at p < 0.01; **significant at p < 0.05; *significant at p < 0.1 (one-tailed tests)

Fig. 10

Hypotheses test incl. control variable (organization-specific view). ns not significant; ***significant at p < 0.01; **significant at p < 0.05; *significant at p < 0.1 (one-tailed tests)