The digital leadership emerging construct: a multi-method approach

Systematic literature reviews are “an increasingly used review methodology to synthesize the existing body of literature in a field” (Kraus et al. 2020, p.1023). Typically, this review focuses on a particular topic, while being transparent and reducing bias (Williams et al. 2021). This approach is used to flexibly analyze data to systematically and rigorously integrate, interpret, and synthesize qualitative findings from multiple studies (Finfgeld-Connett 2014).

Our review follows Corley and Gioia (2004) regarding the coding process who structure data according to hierarchical categories such as first-order concepts, second-order themes, and aggregate dimensions. The categories were developed inductively through constant revision of their labels as necessary. The first-order concepts highlight key aspects about the discerned coding process but not the patterns or relationships in the data (Clark et al. 2010). Second-order themes help to assemble a more structured view of the relationships among the first-order concepts by viewing the data at a higher level of theoretical abstraction (Gioia et al. 2010). In the third stage of analysis, major themes are assembled into aggregate dimensions indicate the interrelationships among the major concepts, themes, and dimensions (Gioia et al. 2013). This systematic method leads to a credible interpretation of the data and provides a qualitative rigor to the analysis (Curado et al. 2019; Gioia et al. 2010; Nag and Gioia 2012).

Data were collected from the Web of Science and Scopus (Cobo et al. 2011; Schäfer 2022). First, we performed a search on the terms “e-leadership” or “virtual leadership” or “digital leadership” or “leading online communities” in the Scopus database by focusing on three main subject areas (i.e., social sciences; business, management, and accounting; and psychology) and by using article title, abstract, and keywords for the search. We decided to limit our search to articles published between 2000 and 2020, because Avolio et al. (2000) coined the term e-leadership to connect leadership and technology in 2000. We restricted our search only to journals published in English and eliminated documents such as letters, editorials, conference papers, and book chapters, since these works are a body of certified knowledge providing reliable results (Bhatt et al. 2020; Haleem et al. 2020). The Scopus database had 173 articles and, subsequently, we performed the same search in the Web of Science database to find missing articles. This search turned up 86 articles, most of them overlapped with those from Scopus.

The data retrieved form the bibliographic databases normally contain errors, such as misspelled elements. Consequently, an analysis of the retrieved data was necessary (Cobo et al. 2011). Our exclusion criteria follows Soriano et al. (2018), as articles unrelated to the scope of our research (e.g., articles not related to organizational studies), were excluded. In this way, we excluded thirty-nine articles from non-organizational domain (e.g., educational studies, healthcare topics), three articles that were non-English manuscripts, and sixty-eight in which the search term was missing (e.g., results in the databases provided articles with the word “leadership” instead of “e-leadership”). This process generated 79 articles to be more deeply examined. Appendix A shows the articles belonging to the research dataset and whether they appear in Scopus and/or the Web of Science databases. The decrease in the number of articles from the initial dataset is not unusual, as in Galvagno and Giaccone (2019), Palmaccio et al. (2021), and Keathley-Herring et al. (2016).

The aim of data coding is to identify relevant codes or themes in the data without worrying about broader categories (Williams and Shepherd 2017). This study was started by reading all the articles from our selected database with an open mind. We started the first-order codes categorization process by classifying the elements and regrouping them due to their common characteristics (Bardin 2011; Gioia et al. 2012). Sentences were named and assigned to nodes that emerged during the coding process after using an inductive approach with no coding framework in mind (Finfgeld-Connett 2014). A single classification was used for each code that meant the categories were mutually exclusive, as recommended by Weber (1990), to restrict the categories to those codes that unmistakably reflect them and to maximize validity.

As the research progressed, we started seeking similarities and differences among the first-order codes; a process that eventually reduced the number of categories to a more manageable number, namely second-order themes. The first-order concepts and second-order themes were reviewed and renamed, when appropriate. All the articles from the dataset were coded. Finally, we were able to name the aggregate dimensions, building a data structure. The DL capabilities were identified through the second-order themes aggregated into different dimensions. As the search process was based only on leadership keywords concerning digital aspects (i.e., digital leadership, e-leadership, virtual leadership, leading online communities), all the identified capabilities came from articles related to the digital realm.

A qualitative content analysis should be evaluated based on reliability and validity (Duriau et al. 2007). According to Krippendorff (2004), a research procedure is reliable when it replies to the same phenomena in the same way regardless of the circumstances of its implementation. First, stability was achieved when the same content was coded the same way by the same coder (Williams and Shepherd 2017). This process was interactive, with one author constantly going back and forth from the text to the emerging concepts and themes while coding and recoding the data as many times as needed to adequately capture the desired domain. Second, intercoder reliability was achieved when different coders coded the same data the same way (Campbell et al. 2013). Further, the other two authors randomly recoded 10% of the dataset (Potter and Ware 1987; Ader 1995; Potter and Levine-Donnerstein 1999). Agreement among coders was reached through discussion over the coding discrepancies, and intercoder reliability ranged from 92 to 98%. The validity of the findings was evaluated based on data saturation and fit (Finfgeld-Connett 2014). When we finished coding the forty-seventh article, we reached theoretical saturation (Glaser and Strauss 1967). Additional data did not increase understanding or meaning and thus, no new codes arose. Consequently, clear and unambiguous items were developed based on the second-order themes. The use of simple language and short sentences was a concern relating to the creation of items that aimed to be neither double barreled nor leading (Fisher et al. 2001).

The aggregate dimensions and their definitions are detailed in Table 1. Four different dimensions were identified, namely strategic focus, delivery related, personal attribute and interpersonal oriented. Table 2 shows the second-order themes identified for each dimension with their definitions. There are 32 constructs in total, being 9 related to strategic focus, 8 for delivery related, 8 for interpersonal oriented, and 7 for personal attribute.

A Delphi study is a scientific process designed to elicit and organize the opinions of experts to reach consensus (Miller 2001), and to validate the information from the literature (Ahmad and Wong 2019). Therefore, this method requires a panel of experts on the subject under study (Schmidt et al. 2001). One benefit of the panel is that the experts asynchronously participate in group communication and they can be geographically dispersed as they do not meet each other nor interact directly (Linstone and Turoff 2011). The Delphi study can combine academic expertise with practitioners’ perspectives that enables each participant to express their views independently while providing information generated by an entire group (Foth et al. 2016). It can obtain a reliable consensus of the panel of experts through systematic interactives approaches, or “rounds”, using a questionnaires with controlled feedback (Clayton 1997; Linstone and Turoff 2011). Some authors call it “e-Delphi” when referring to the use of an online platform to administer the questionnaire (Massaroli et al. 2017; Bec et al. 2019). The number of experts (n = 24) recruited in this study was within the scope of the most frequent size in the literature (n = 15–30) (de Jesus et al. 2019). Each expert was required to have leadership experience and to possess specialized knowledge and interest in the studied topic. We drew experts who worked in different countries, such as Portugal, China, France, Germany, Switzerland, and Brazil to obtain a greater diversity in the outcome.

Standard Delphi studies usually have three rounds of data collection (Brady 2015), but some authors argue that the classic technique has four rounds (Hasson et al. 2000). Erffmeyer et al. (1986) argue that before limiting Delphi to less than four rounds, the researcher should confirm if they have reached stability in the previous rounds with an acceptable level of accuracy. Usually, Delphi studies starts with open or semi-open questions to collect data from experts and the next rounds are made up of items generated by experts in the first round combined with items identified from the literature that are aimed at reaching stability and consensus in the next rounds (Dewar et al. 2017).

Our study consists of four rounds and starts with online interviews using semi-open questions. The next rounds were conducted through an online questionnaire system (www.​qualtrics.​com) and were based on the responses of the interviewees combined with the data gathered in the literature (Ahmad and Wong 2019). Regarding the available time that experts dedicated to the Delphi study, the experts in Round-02 were separated into two groups receiving different questionnaires due to the large number of items to be evaluated. For Round-02 onwards, the experts had a week to respond, and two reminders were sent per round: the first, 1 day before the deadline; and the second, on the deadline day itself. The expected time for completion for each questionnaire varied from 5 to 15 min. There was an option to give comments in the questionnaire for all rounds. Different measures were used to assess the items across the questionnaire rounds, including picking-type (Round-02) and Likert-type (Round-03 and Round-04). These measures were designed to measure the items from different perspectives and for different purposes to identify the most important ones (Okoli and Pawlowski 2004; Bec et al. 2019). It is essential to differentiate between the concept “consensus” and “stability” in Delphi studies. Stability refers to the consistency of experts’ answers between successive rounds of a Delphi study and it is considered a necessary criterion to evaluate consensus (Dajani et al. 1979; Von der Gracht 2012). Although there are different approaches to defining and reporting consensus (von der Gracht 2012), systematic reviews have shown that consensus is mainly defined based on the percentage of experts who rate items at the upper extremes of the Likert scales used (e.g., items scored as 4 and 5 on a 5-point Likert scale) (Diamond et al. 2014; Foth et al. 2016). There is no agreement about the percentage of answers by experts that is considered consensus. Some authors state that items with 50% or more have reached consensus (Trevelyan and Robinson 2015), while others argue that 60% or higher is preferable (Foth et al. 2016). This differentiation is important since stopping a Delphi study based only on a specific number of rounds may lead to invalid and meaningless results (Barrios et al. 2021). Measures of group stability also varies in the literature, varying from a coefficient of variation (CV) of 15% change in two Delphi rounds (Von der Gracht 2012) up to 20% (Massaroli et al. 2017; Shi et al. 2020).

In the first Delphi round, we asked open-ended questions of 24 experts during in-depth on-line interviews that aimed to obtain practical insights about which leadership capabilities they considered to be the most relevant to thrive in a digital environment. The interviews were composed of open-ended questions (de Jesus et al. 2019), which are acknowledge to increase the richness of the information collected (Powell 2003). All the experts consented to having the interviews recorded as per our protocol (Massaroli et al. 2017). The interview responses were analyzed to compare with the capabilities identified in the systematic literature review, adding the practical insights of the experts concerning DL capabilities.

The qualitative analysis of the first round improved the spectrum of capabilities and provided more themes to be processed in the following rounds, as performed in Ahmad and Wong (2019) and Dewar et al. (2017).

Round-02 involved narrowing down the list of items based on the new capabilities identified by the experts in Round-01, added to the items found in the systematic literature review (Okoli and Pawlowski 2004). Each expert received a unique code that identified them in the questionnaire throughout the rounds. We sought to restrict the number of items in Round-02 to facilitate future consensus (Hsu and Sandford 2007; Schmidt et al. 2001). A list of items was sent to the experts with instructions to indicate those which were most important for the leadership capabilities in the digital environment. The list was randomly ordered to reduce any bias (Paré et al. 2013; Schmidt 1997). We requested that the experts select at least 40% of the items for each dimension they considered the most important for a leader in digital era, as the literature has stated that the experts should independently select at least 10% of the items as the most important (Schmidt 1997). Regarding consensus, we eliminated all items that were not selected by at least 60% of the respondents, following Foth et al. (2016). In addition, if a capability had less than three remaining items, it was eliminated as some authors recommend against retaining constructs with less than three items (Worthington and Whittaker 2006). This process reduced the lists to a manageable size.

In Round-03 and Round-04, the questionnaires were used for the purpose of attaining consensus on items by the panel of experts. Items were measured by using a Likert scale that ranged from 1 (not important) to 5 (very important). The Round-03 questionnaire was developed based on the items retained in Round-02. The items were again randomly presented within each dimension to avoid bias. The consensus on the inclusion of items for the next round was based on their means. The items that scored 4 or 5 on the Likert scale were kept separate as they had already achieved consensus, as suggested by Barrios et al. (2021). Therefore, only the items that received a mean rate of less than 4 were considered for Round-04 to be re-evaluated by the panel of experts. Furthermore, at least three remaining items for each capability should remain to retain the construct, as in the previous round.

For Round-04, to ascertain the level of collective opinion, we attributed the mean scores of the expert’s rate for each item that had not obtained consensus to each expert, to remind them of their previous responses (Hasson et al. 2000). Questionnaires were prepared individually for each expert and they were asked to re-consider their rates maintaining or modifying their previous rate (Trevelyan and Robinson 2015). Finally, the items were analyzed to identify consensus. Applying the same criteria used in Round-03, the items with scores of less than 4 had not reached consensus and were eliminated. The experts´ group stability from Round-03 to Round-04 was 15%, which is in line with Diamond et al. (2014) for the Delphi process to reach stability.

While we had 24 experts who participated in the first round, only 16 completed all four rounds. This reduction was is in line with other Delphi studies in which the abstention rates varied from 20 to 50% in each round (Massaroli et al. 2017). Concerning demographic information, the characteristics of the experts in the first round had an average leadership experience (± standard deviation) of 17.0 years ± 5.84 years and all of them had at least a bachelor degree, while 46% had a master´s degree, and 8% had a doctorate. Of the experts, 50% were male and 50% were female. The average age was 45.96 years ± 6.64 years. Moreover, 21% of the experts were between 30 and 40 years, 54% were between 41 and 50 years and 25% were more than 50 years old.

In Round-01, no new dimension was identified through the interview process. The experts identified 11 capabilities: 6 for personal attribute, 3 for interpersonal oriented, 2 for delivery related, and none for the strategic focus dimension. New items were elaborated and added to the initial pool creating the baseline for Round-02. The number of constructs identified, either in the literature review process or by the experts, together with the consensus information for each round are shown in Appendix B regarding all four dimensions.

In Round-02, there were a total of 292 items concerning all four dimensions to be evaluated among the experts. Among these items, 68 belonged to the strategic focus dimension, 68 to delivery related, 80 to personal attribute, and 76 to interpersonal oriented. Each expert judged two dimensions that they received in the questionnaire and could choose a certain number of items. There was between 144 and 148 items for them to select as the most important concerning two dimensions. Some second-order themes were eliminated in this round, as the experts did not choose the correspondent items and, so, the theme did not reach the minimum of three items to be maintained for the next rounds. For the interpersonal dimension, the second-order themes that were eliminated were “able to resolve conflicts” and “willingness to listen”; for strategic focus, the theme was “sustainability acumen”; for personal attribute, they were “humility”, “self-awareness”, and “self-regulation”; and, for delivery related, they were “creative problem solving”, “provide resources to the team”, and “able to manage work-life balance”.

In Round-03, the quantity of items rated by the experts were 40 strategic focus, 34 delivery related, 52 personal attribute and 32 interpersonal oriented. For each dimension, the number of items that reached consensus in this round were 36 strategic focus, 31 delivery related, 41 personal attribute and 27 interpersonal oriented. In this round, there were some second-order themes for which that all the items reached consensus; consequently, these were designated as final, and the other items were excluded from Round-04.

In Round-04, the number of items that have reached consensus were 2 for strategic focus (out of 4), 2 for delivery related (out of 3), 2 for personal attribute (out of 11) and 2 for interpersonal oriented (out of 5). The second-order themes that were eliminated in the final round were “inspire and engage people” (interpersonal oriented dimension), “curiosity” and “enthusiastic and passionate” (personal attribute dimension) and none for strategic focus and delivery related. In all four dimensions, the change in distribution of the responses was less than 15%. This procedure is a crucial part of the Delphi process in determining consensus and stability in the answers (Shi et al. 2020) and, therefore, the stopping criteria rule was reached. After this fourth round, the remaining capabilities and their definitions can be seen in Table 3.

A conceptual model of the final digital leadership capabilities is presented in Fig. 1. Concerning the four dimensions and the relationship within each theme, an interpersonal oriented leader interacts successfully with others (Giromini et al. 2016). Therefore, this dimension relates to social interaction between leaders and followers, such as establishing relationships, communicating, being trustworthy, coaching, inspiring, empowering, and engaging people (Buchan et al. 2008; Schiuma et al. 2022; Walvoord et al. 2008). The personal attribute dimension informs how leaders manage themselves and their emotional reactions to situations (Scott et al. 2008). This dimension and its constructs concern the leaders’ inner self: curiosity, initiative, autonomy, and self-awareness (Owens et al. 2013; Shamim et al. 2016). The strategic focus dimension establishes how leaders form and deploy strategic decisions toward the organization’s future goals (Orndoff 2002). Therefore, this dimension refers to leadership capabilities that affect a long-term organizational objective and the response to changes (Vecchiato 2015). Some leadership constructs identified within this dimension are vision and direction, innovation, sustainability, knowledge sharing, and leading change (Kane et al. 2019; Lawrence 2015; Swanson et al. 2020). Lastly, the delivery related dimension refers to the leadership capability of managing the team toward a desired outcome (Cordery et al. 2009), and the leadership capabilities for this dimension include results orientation, creative problem solving, and establishing and managing team performance (Agarwal et al. 2017; Taggar 2001).

Following Cronin and George (2023), it is indicated in the conceptual model three different categories (literature, experts, literature/experts). It is informed if a construct was, proportionally (i.e., difference in percentage of the number of times the construct appeared in the literature compared to the number of times the experts mentioned it), more heavily based on articles from the research dataset (literature) or deriving more from the experts’ assessment, or if the construct was balanced reported between the literature information and experts opinion (i.e., difference less than 10% between the literature and the experts).

While our review is timely and includes both an academic literature analysis and a practitioner perspective, it is not without its limitations. The background of each panel member may directly influence their decision-making about the topics; this bias is generally not within the control of a Delphi study. Additionally, panel members are often limited in the amount of time they could dedicate to the Delphi study, therefore reducing their ability to consider all dimensions under study. Although the results from a panel of experts can produce practical and thoughtful insights, they may not be an exhaustive nor an all-inclusive set of ideas. Regarding the systematic literature review, despite analyzing 21-years of peer-reviewed articles, the choice to limit the study to those that explicitly mentioned the search terms and may have missed some important articles concerning the theme.

The avenues for future research are plentiful. Upcoming literature reviews and Delphi studies should consider expanding the keywords for different searches on the digital phenomena. Focusing on digital technologies (e.g., artificial intelligence, machine learning) could provide some new insights for the field, and new leadership characteristics can emerge as the topic becomes more mature in organizations and the literature. Furthermore, it will be necessary to empirically validate the proposed capabilities through new and different studies. Nonetheless, this study can be regarded as an insight into the main leadership capabilities for digital times. Other future studies could strengthen the studies on the competencies that were highlighted most strongly by the experts compared to the literature. Perhaps these topics have nuances to better explore due to the digital transformation.

DL operationalization could be established by elaborating a scale to measure the DL capabilities expected of the digital leader. Table 4 suggests items that could be tested using an exploratory factor analysis followed by a confirmatory factor analysis to verify which items will remain in the final DL scale. Moreover, a causal relationship between different DL constructs could be explored. As leadership is context-related, what are the most significant changes in DL behaviors pre- and post-COVID-19? How important is each dimension concerning the industry of DL studied? Is there any difference?

Leadership is context-related (Osborn et al. 2002). Therefore, the DL needs to perform in a time of disruptive technological change, in which strategic importance of ethics in the context of DL is important (Vial 2019). With the increased use of generative AI systems (e.g., ChatGPT) to provide content or outputs (e.g., text, images, audio, etc.) based on the data they are trained on (Eke 2023), DL is needed to operate and guide the organization towards an acceptable and ethical behavior using these systems (Dwivedi et al. 2023). AI has been used to accomplish different HR activities (e.g., recruitment, selection, training) in organizations (Banks et al. 2022). AI relies on data, training models, and human input, being possible to inject bias into the system (Dwivedi et al. 2023). Therefore, another study could concentrate on the DL role to avoid bias amplification and to ensure diverse representation (e.g., see Bolukbasi et al. 2016 as an example of gender bias in machine learning datasets). As more machine-based human interactions are being used as recruiters (e.g., AI chatbots), the findings may have implications for identifying the impact on HR recruiting and engagement practices (Nawaz and Gomes 2019). Therefore, calling leaders to concentrate their attention on it. Another future study can combine case studies and in-depth interviews to understand of employees’ AI usage contribute to employee performance and engagement (Wijayati et al. 2022) and how DL influences employees’ behavior. In addition, the qualitative examination could provide different elements of the DL’s capabilities. Finally, leaders must increasingly interact between employees and AI, adding another component to the leadership process (Peifer et al. 2022).

While Industry 4.0 is about automation (i.e., AI, machine learning, big data), Industry 5.0 focuses on the synergy between humans and autonomous machines (Nahavandi 2019). The introduction of robots in the workplace may affect social interactions and can be more challenging with robots in management or superior positions (Demir et al. 2019). Therefore, another potential study could examine the leader–follower interaction when one part is a robot. How different will the employee’s social behavior be in interaction with human leaders and robot leaders, if any? What are the challenges for managers to build a team based on collaboration, having humans and robots working together?

Furthermore, another future study could analyze the interaction between leaders and followers in the metaverse ecosystem using avatars. The metaverse with different design elements is recognized as a valuable means to deliver emotional content and can elicit distinct emotions in people (Dozio et al. 2022). So, how does a leader build trustworthiness with their followers in a different scenario? As leadership is context-related, are there any differences between building trustworthiness in the metaverse and in person?