Generating impactful situated explanations through digital trace data

Using digital trace data to address real-world problems

Digital trace data present an exciting prospect for IS researchers to learn more accurately and precisely about socio-technical phenomena in an increasingly digitalized world. Not only can work be hardly meaningfully decoupled from digital technology anymore (Bailey et al., 2022; Orlikowski and Scott, 2008), but also most if not all aspects of our life are increasingly mediated by digital technologies (Alaimo and Kallinikos, 2017; Yoo, 2010). As the name suggests, digital trace data capture digitally mediated actions and events as they unfold over time, at an astonishing level of precision and granularity (Pentland et al., 2020). Every keystroke, every transaction, even every heartbeat and eye movement can potentially be logged (Power, 2022). This allows more precise and more voluminous data on actions and events than traditional modes of data collection, such as observations, interviews or archival data (Schensul et al., 1999).

Digital trace data provide plenty of opportunities for IS scholars, but serious methodological adjustments are required to consider the particularities of this new type of data. In consequence, recent works in IS have started to shed light on the implications of using digital trace data for research, especially for the development of theory (e.g., Berente et al., 2019; Lindberg, 2020; Miranda et al., 2022). For example, it has been suggested that digital trace data need to be framed within a theoretical lexicon in order to contribute to a specific discourse (Berente et al., 2019). Also, there have been works that focus on the interplay of human and computationally-enabled sense-making to detect patterns in digital traces (Kallio et al., 2022; Lindberg, 2020). Others have examined the quality and features of digital trace data sets, along with their opportunities and threats for theorizing (Howison et al., 2011; Pentland et al., 2020). Furthermore, there have been discussions about the general implications of digital trace data for theory building (Grover et al., 2020; Osterlund et al., 2020). What has received relatively little attention in this scholarly discourse is that digital trace data also provide ground for developing impactful solutions to real-world problems (Ram and Goes, 2021; Tremblay et al., 2021; vom Brocke et al., 2021). To this end, representing and analyzing digital trace data enables researchers and policymakers to gain an in-depth understanding about what is happening at a specific point in time, and what could be done to stir a given problem context into more desirable directions (Agerfalk, 2010; GovLab, 2021; Oliver et al., 2020).

We do not mean to suggest that impactful solutions to real-world problems based on digital trace data do not yet exist. Researchers and policymakers are already using all kinds of digital trace data to understand problem contexts as they evolve over time (Kogan et al., 2021; Mbunge et al., 2021; Oliver et al., 2020), sometimes in complement with other data sources, such as interviews or databases (Knippenberg and Meyer, 2020; Kogan et al., 2021). Researchers in China, for example, used digital trace data collected from Baidu, a national search engine, to understand how the COVID-19 pandemic was spreading as well as to evaluate the effectiveness of non-pharmaceutical interventions, such as travel restrictions and social distancing (Hu et al., 2021; Lai et al., 2020).

Situated explanations and impactful research

In recent years, there has been an increasing number of calls for researchers to address real-world problems and gain a deeper understanding of the contexts in which they occur (e.g., Pataro et al., 2021; Rosen, 2018). Such an understanding requires the creation of insights and solutions that consider the specific factors and circumstances relevant to a given problem and sharing them with relevant stakeholders, such as policymakers, at the appropriate time (Gaieck et al., 2020; GovLab, 2021; Lai et al., 2020). During the COVID-19 pandemic, for example, researchers, policymakers, and other stakeholders collaborated to address various emerging problems as they occurred. Some interventions were informed by available theory, while some new theory development occurred ex-post as a result of certain interventions (Ruggeri et al., 2022; Van Bavel et al., 2020). But irrespective of whether theory was the starting or end point, the primary focus has been to “bring about informed change” (Agerfalk, 2010: 251) in a specific context. The pandemic also showed that the quest for practical, impactful solutions can at times increase demand for timeliness to such an extent that traditional quality assurance processes in science were paused or altered (consider, e.g., the enormous uptake of pre-print and pre-review publication of scientific output (Hoy, 2020)). It appeared at times that the appetite for timely and impactful solutions trumped the demonstration of rigor (Oliver et al., 2020).

Impactful solutions thus create a need for situated explanations to be both specific and timely, two requirements that go beyond the traditional demands for theoretical explanations to be generalizable across time and space (Gregor, 2006; Hirschheim, 2019; Weber, 2012). The process of generalizing typically means that an explanation is abstracted from a given context so that it can be compared and integrated with a larger body of knowledge (Farhoomand, 1987; Hirschheim, 2019; Lee and Baskerville, 2003; Williams and Tsang, 2015). But this process may ignore important contextual details that, in turn, impede the application of an explanation to a specific problem context (Davison and Martinsons, 2016: 2), a limitation that the development of situated explanations can avoid.

The view that explanations should be specific and timely is not entirely new to the IS field. Several scholars have suggested in the past that generalizable explanations may not always be needed, or even useful, for solving real-world problems (e.g., Agerfalk et al., 2020; Hirschheim, 2019; Ram and Goes, 2021). In order to address real-world problems, these claimants suggest, information systems research should be primarily concerned with the development of an in-depth understanding of a given problem context by leveraging the increasing availability of digital trace data—regardless of whether they have theoretical value or significance (Hirschheim, 2019; Ram and Goes, 2021; Tremblay et al., 2021).

However, realizing that we can leverage the idiosyncratic features of a given problem-contexts to produce impactful solutions with situated explanations is only a first step. If our ambition is to leverage the potential of digital trace data for practical impact, we need to understand how situated explanations can be built. Previous works that called for using digital traces to develop “theories in flux” (Tremblay et al., 2021) or “local theories” (Leavitt et al., 2021) have been silent about the actual ways through which one can develop practical impact. This is what we set out to achieve with this article. Thus, in what follows, we present five principles to develop situated explanations based on digital trace data to generate timely and practical impact.

Embracing idiosyncratic and dynamic representations to develop situated explanations

Our starting point is the observation that we live in a world in flux (Mousavi Baygi et al., 2021; Tremblay et al., 2021), which means that we need to pay attention to the idiosyncrasies of problem contexts and how they evolve over time. The challenge is to represent both the idiosyncratic features and the temporal dynamics of real-world problems. A representation of a problem––be it in the form of a text, a figure or a picture, an animation or a video, a sound or an audio file, or any combination of these meaningful symbolic elements––can be useful for developing situated explanations as long as it depicts the problem in such a way that it is informative to a wide variety of different actors and stakeholders, and supports them in effectively reflecting on that problem in practice (Munger et al., 2021; Oliver et al., 2020; Tremblay et al., 2021) and as long as it describes the evolution of the relevant idiosyncratic features over time.

For example, during the COVID-19 pandemic, policymakers needed to define and evaluate effective measures to promote social distancing. To that end, researchers created representations of mobility patterns of populations and their changes in real-time and at scale (Whitelaw et al., 2020). To build these representations, they collected mobile phone data from sources, such as cell towers, dedicated smartphone applications, Bluetooth functionalities, and others ¹ (GovLab, 2021; Grantz et al., 2020; Oliver et al., 2020). Leveraging these data allowed policymakers and researchers to depict a highly complex, dynamic, and distributed phenomenon in a comprehensible way. It enabled fruitful discussions among a diverse group of stakeholders, including both scholars and policymakers (Knippenberg and Meyer, 2020; Oliver et al., 2020). All of this played a crucial role for designing effective policies in a collaborative and iterative process (Budd et al., 2020; GovLab, 2021; Oliver et al., 2020; Van der Drift, Wismans and Olde Kalter, 2022). Box 1 summarizes the approach taken by the European Commission’s Joint Research Centre in their efforts to restrict the transmission of COVID-19.

Box 1: Developing situated representations to understand real-world problems: The case of the European Commission’s Joint Research Centre during the Covid-pandemic.

At the beginning of the pandemic in March 2020, the European Commission faced the challenge of implementing social-distancing measures to prevent the spread of the virus, while at the same time, keeping the economic damage to a minimum (GovLab, 2021). Urgently and with little lead time, the European Commission negotiated with various European mobile network operators to obtain mobility data. Subsequently, the European Commission’s Joint Research Centre (JRC) carried out several initiatives to analyze real-time mobility in order to inform policymakers and other decision-makers (Santamaria et al., 2020).

The JRC received mobile positioning data which were recorded by cell towers servicing mobile phones as users make calls, write texts or retrieve data (GovLab, 2021). These data were delivered in the form of anonymized Origin-Destination Matrices, which depict the number of movements of mobile phone users from one area to another. To represent these data in meaningful ways, the JRC took two steps. First, they developed measures, such as Mobility Functional Areas (Iacus et al., 2020), to represent how populations move, regardless of fixed administrative boundaries (such as districts or cities) (Iacus et al., 2020). These areas were used to compute real-world mobility trends by means of various indicators, including those that occurred within this area, those that started in this area and led to another area, and those that started in different areas and ended in a given area. Second, members of the JRC produced a variety of informative representations in order to depict mobility trends—and thereby, the potential spread of the pandemic—from various angles. These representations showed, for instance, movements across regions and countries, and they could be used to evaluate the effects of specific interventions. Exemplary representations can be found in Iacus et al. (2020) and Santamaria et al. (2020).

But it is not enough that representations are informative about a given problem context. In a world that is in constant flux, most problems are going to be “moving targets” that change their shape over time. Since a problem context dynamically evolves, representations must capture idiosyncratic features at least at two points in time. Therefore, it will be more important to produce temporal sequences of representations quickly, rather than having a method that might produce a near-perfect representation but takes excessive computation time. Hence, the value of any representation should be evaluated by how timely and useful it is to depict a given problem context effectively (Agerfalk, 2010; Levy et al., 2020).

It is also important that we arrive at such representations in a reliable, that is, consistent and reproducible, way. Reliability is often high if algorithms and scientific workflows are used to produce representations in a deterministic and systematic way. This is because computational methods, at least deterministic ones, guarantee that different actors can independently arrive at the same representation of a problem, fostering the shared understanding necessary for effective collaborative problem solving processes (Okhuysen and Bechky, 2009). FAIR principles of data (Wilkinson et al., 2016) and research software (Lamprecht et al., 2020) support this ambition, in particular regarding reproducibility. For example, some process mining algorithms are known to produce representations (such as network graphs) with high reliability (Van Der Aalst, 2016). But not all computational methods guarantee reliable construction of outputs. For example, several pattern discovery algorithms are known not to be reproducible because they generate new visualizations of patterns in every execution (Indulska et al., 2012). Likewise, most traditional qualitative research methods produce representations (such as coding tables or concept maps) that can be subjective and unreliable (Silverman, 2013).

Taken together, what we envision is a complementary form of knowledge accumulation that centers around our ability to develop informative representations of idiosyncratic problem dynamics quicker (i.e., requiring less time to produce the representation) and more reliably (i.e., with higher chance of different scholars arriving at the same representation independently of each other). Together, this leads us to our first principle:

Principle 1:

In a world in flux, digital trace data should be used to develop informative representations of idiosyncratic and dynamic real-world problems in quick and reliable ways.

Processual patterns are the most useful type of representation for a world in flux

In order to systematically unpack the meaning of our first proposition, we need to specify the type of representations we are talking about. We thus address the following question: What type of representation will be most effective in helping actors understand the challenges that arise for them in a world where change is all there is (Mousavi Baygi et al., 2021)? We suggest that a processual pattern should be the primary form of representation around which situated explanations are built.

Generally speaking, a pattern describes recurrent aspects of a phenomenon or problem in a recognizable yet abstract form (see also Alexander, 1977; Miranda et al., 2022). That is, a pattern is a form of representation that focuses our attention on those aspects that are (at least somewhat) repetitive and, therefore, recognizable. Patterns have always played a key role in the social sciences (e.g., Gioia and Poole, 1984; Goffman, 1974), but patterns are not theories, and processual patterns are not process theories; instead they are empirical generalizations (Handfield and Melnyk, 1998). Still, patterns have become an important building block for theoretical contributions IS scholars seek to make when using digital trace data (Berente et al., 2019; Miranda et al., 2022). Patterns make latent structures in empirical data visible and thereby allow for more parsimonious and, therefore, potentially insightful descriptions of the world around us. Detecting a pattern thus provides an opportunity to intervene into a problem because one will be better able to see why and how a problem persists (Oliver et al., 2020; Tremblay et al., 2021).

While patterns are useful because they point to recognizable aspects of a problem, it would be misleading to focus only on those aspects of a problem which remain stable over time. Static patterns are not the best way of representing real-world problems because problem contexts can and will change their structure. This process typically has characteristic rhythms (Ancona and Chong, 1996) which, amongst other things, imply “windows of opportunity” during which interventions are more likely to be effective (Tyre and Orlikowski, 1994). Hence, to account for variation and change over time, we need to develop processual patterns. We define a processual pattern as a representation of one or multiple sequences of actions and events that captures their recurrent aspects in a recognizable form as the phenomenon of interest changes over time.

Processual patterns describe recurrent properties of a problem, but also reveal how a problem forms, changes, and dissolves over time. In analogy, while static patterns are pictures, processual patterns are movies. As a specific type of representation, processual patterns sensitize actors for the volatile, sometimes even unpredictable dynamics that the digital age confronts us with now and in the future (Mousavi Baygi et al., 2021). An understanding of these dynamics, we argue, will be key to design and evaluate solutions to these problems over time (Oliver et al., 2020; Tremblay et al., 2021).

The availability of digital trace data makes it feasible to generate such processual patterns. We gain more and more digitally logged evidence of activities and events that occur around us (Freelon, 2014; Lazer et al., 2020; Power, 2022), which allows us to infer patterns in peoples’ behavior (Leonardi and Treem, 2020) and how these patterns change over time (Pentland et al., 2020; Zhang et al., 2021). Together, they allow us to generate rich contextualized descriptions of actions situated in particular contexts such as time (now, later, …), place (here, there, …), subject (me, you, …), or technology (smartphone, app, computer, …) (Barnes and Law, 1976; Suchman, 1987). Moreover, processual patterns are insightful not because they are highly generalizable, but because they embrace the situated dynamics of real-world problems (Oliver et al., 2020; Pentland et al., 2020).

To illustrate: During the COVID-19 pandemic, policymakers across the globe recognized that digital traces stemming from mobile phone use were not only useful to monitor the dynamics of virus transmission, but also to evaluate the effectiveness of interventions. Various forms of mobile phone data offered them “specific, reliable, and timely data not only about infections but also about human behavior, especially mobility and physical co-presence” (Oliver et al. 2020: 1). Since different forms of mobile phone data had particular strengths and weaknesses (Grantz et al., 2020; Ojokoh et al., 2022), they were used for different areas of inquiry throughout the epidemiological cycle (Schlosser et al., 2020). For example, in the early stages of the pandemic, digital traces from mobile phone use, primarily collected through connections between mobile phones and cell towers, enabled situational analysis to understand if and to what extent the population adhered to lockdown orders (Perra, 2021) (see Box 2). At subsequent stages, other mobile phone data, such as data from smartphone apps and Bluetooth functionalities came increasingly into effect (GovLab, 2021; Van der Drift et al., 2022). In any case, these data allowed researchers and policymakers to examine human mobility and interaction patterns at various stages of the pandemic as they indicated, for example, the extent to which populations adhered to lockdown interventions over time (GovLab, 2021).

Regardless of the specific initiative and the type of mobile phone data that were used, key is that they allowed to track the temporal process through which the pandemic unfolded over time. Throughout the pandemic, the use of mobile phone data provided policymakers and other relevant stakeholders with the opportunity to produce quick and reliable representations of the problem context as they identified patterns about human behavior that changed at different points in time (Chang et al., 2021). This, in turn, allowed them to analyze and react to problem contexts as they unfolded over time. We illustrate this point in Box 2 by elaborating on the example by the European Commission’s Joint Research Center.

Box 2: Developing processual patterns to understand real-world problems: The case of the European Commission’s Joint Research Centre during the Covid-pandemic.

Members of the JRC developed a number of processual patterns that enabled policymakers to recognize changes in mobility patterns indicating how populations moved and potentially contributed to the spread of the virus (GovLab, 2021; Iacus et al., 2020). Santamaria et al. (2020) show mobility indicators for different regions in Spain over time, and how they changed after the lockdown had been imposed. Such representations, in turn, made it possible to see in which regions certain interventions were more or less effective. By comparing areas in Spain, for example, time-based representations showed that the first lockdown in March 2020 was leading to a significant decrease in mobility patterns, as represented through movements within, from and to specific areas (Santamaria et al., 2020). Over time, however, this effect remained to different extents and mobility patterns changed at different paces (Santamaria et al., 2020).

From a broader perspective, the JRC was also able to compare how mobility patterns changed across European countries (Santamaria et al., 2020), showing how different lockdown policies led to different effects over time. Also, it was possible to compute mobility indicators for areas and countries to evaluate the effectiveness of mobility restrictions (Santamaria et al., 2020). Insights obtained through such patterns, in turn, proved valuable for evaluating lockdown interventions, informing early warning applications and modeling-based analyses of the pandemic (GovLab, 2021; Iacus et al., 2020).

In short, we envision that the generation of processual patterns will aid the development of situated and impactful explanations. The second principle we propose is:

Principle 2:

Situated explanations should be based on representations of idiosyncratic and dynamic problem contexts by means of processual patterns generated from digital trace data.

Developing situated explanations at the intersection of problem, processual patterns, and data

In order to implement the two principles developed above, the development of situated explanations should focus on the relations between (1) real-world problems, (2) digital trace data, and (3) processual patterns (see Figure 1).OPEN IN VIEWER

Establishing data-problem fit

Digital trace data can only ever provide a partial representation of the real-world processes that left these traces (Power, 2022). In analogy, even a movie is just a series of snapshots; and every dataset however large and comprehensive is essentially only a case study at particular points in time and context (Indulska et al., 2012). Thus, every digital trace dataset allows us to see some aspects of a problem while hiding others (Freelon, 2014; Stier et al., 2020). In turn, to move from the recognition of a real-world problem to an understanding of the requirements for a digital trace dataset that will allow us to generate situated explanations requires the establishment of data-problem fit. Data-problem fit means selecting and analyzing digital trace from one or more digital trace datasets such that they allow scholars to (a) understand the problem’s core properties, (b) identify opportunities to mitigate it, and (c) monitor the effects of interventions over time (Latif et al., 2020; Oliver et al., 2020). It defines the task for the researcher to obtain data that fits the problem.

For example, the usefulness of mobile phone data to examine the transmission of the COVID-19 pandemic was quickly recognized by countries all over the world (GovLab, 2021; Klein et al., 2020; Lyons, 2020). This was because infection-relevant parameters could be directly obtained from various forms of these data, including mobility, interactions, and social proximity (Oliver et al., 2020). This approach established data-problem fit, which in turn, informed further actions by researchers and policymakers. Box 3 illustrates this point as it further elaborates on the example by the European Commission.

Box 3: Establishing data-problem fit: The case of the European Commission’s Joint Research Centre during the Covid-pandemic.

The European Commission requested data from mobile network operators for different reasons (GovLab, 2021; Iacus et al., 2020; Santamaria et al., 2020; Vespe et al., 2021). First, mobile phones are daily companions in peoples’ everyday life, and they are used for private as well as job-related purposes. Hence, the data obtained through mobile phone use was expected to depict mobility patterns within populations across various member states of the European Union, including Austria, Belgium, Denmark, Germany, Finland, France, and Greece, among others.

Second, these data could be used for various purposes, such as the evaluation of social distancing measures and the potential spread of the pandemic over time. An official document by the European Commission states that mobile phone data, “if pooled and used in anonymized, aggregated format in compliance with EU data protection and privacy rules, could contribute to improve the quality of modeling and forecasting for the pandemic at EU level” (European Commission, 2020: 5).

When relevant behaviors are not enacted with digital technologies, it can be challenging to determine which data sets can lead to meaningful representations about a specific real-world problem (Mureddu et al., 2020; Power, 2022). We still know next to nothing about how we can go about in evaluating the appropriateness of a digital trace dataset for such purposes (Alamo et al., 2020). One suggestion is to initiate the inquiry on the grounds of available data (e.g., from a work system in an organization), to subsequently frame the resulting patterns under a specific theoretical lexicon (Berente et al., 2019; Miranda et al., 2022; Pentland et al., 2021). This implies that the (theoretical) relevance of these data is evaluated after data have been accessed and evaluated.

Even if we have useful digital trace data sets at hand, it might not be immediately clear how data-problem fit could be achieved, and what kind of representations work well. Research on big data-based policy-making (Desouza and Jacob, 2017) and decision-making (Merendino et al., 2018) suggests that users need to explore and experiment with such digital trace data-based representations as they gradually understand problems and develop appropriate solutions. This happens, for example, when they explore a given set of digital traces with different analysis techniques (Desouza and Jacob, 2017; Elgendy and Elragal, 2016; Merendino et al., 2018). During the COVID-19 pandemic, for example, predictions were made through a variety of computational models and under different assumptions (Boland and Zolfagharifard, 2020). In this way, data-problem fit emerges over time from a co-evolution of problem understanding that informs data requirements and available data that contributes to problem understanding, similar to the coevolution of problem and solution observed in design studies (Dorst and Cross, 2001).

The question we need to address is: How can we predict with sufficient accuracy if we will be able to find processual patterns that are potentially insightful regarding a given real-world problem in a trace dataset with given characteristics? Accordingly, the establishment of data-problem fit in quick and reliable ways would need to be a central effort along which we should and could collectively accumulate knowledge. The principle we put forward is:

Principle 3:

Developing situated explanations is contingent on the establishment of data-problem fit.

Triangulating data sets

Because establishing data-problem fit means a digital trace data set must allow for (a) understanding the problem’s core properties, (b) identifying mitigation opportunities, and (c) monitoring intervention outcomes time (Latif et al., 2020; Oliver et al., 2020), single sources of data are not likely to be sufficient to develop a comprehensive representation of a real-world problem (Knippenberg and Meyer, 2020; Osterlund et al., 2020). We will need multiple datasets to identify the patterns that are most relevant to a problem at hand, such as patterns in how, when and where people move (e.g., through devices producing GPS data), with whom they communicate (e.g., through mobile phone use), how they work (e.g., through work systems), and what their motifs are (e.g., through social media posts), etc. (Whitelaw et al., 2020). Hence, if we want to obtain a more comprehensive, more informative representation of a real-world problem, we will need to engage in triangulation, that is, we need to relate multiple sources of evidence about a problem context (Jick, 1979). For situated explanations, triangulation is important because it helps establish a nuanced picture of the idiosyncrasies of any situated action. It also helps increase trust that any potential processual pattern is in fact recurrent and recognizable. Moreover, by triangulating digital trace data from different sources, we can better represent the dynamics of the situation in which a problem is embedded (Freelon, 2014; Pentland et al., 2020).

The use of mobile phone data alone, for example, points to changing mobility patterns within populations (Oliver et al., 2020) but it provides limited insights into the extent to which a virus is being passed on (Heiler et al., 2020; Ienca and Vayena, 2020; Ojokoh et al., 2022), which is why that data was triangulated with other data about the same content and timeframe. Box 4 discusses how the European Commission’s Joint Research Centre triangulated data.

Box 4: Establishing data-problem fit: The case of the European Commission’s Joint Research Centre during the Covid-pandemic.

Despite the various advantages of mobile phone data, they miss information surrounding the context in which they were collected; they show when and where mobility took place, but they do not explain why this is the case. Hence, the European Commission planned to include other digital trace data sources, such as data collected from social media, in order to obtain information about social interactions and early indications of disease spread (e.g., by means of search requests or postings) (European Commission, 2020). From a broader perspective, it proved crucial that initiatives had access to various data pools in order to complement and standardize different data sets for comprehensible analyses (GovLab, 2021; Vespe et al., 2021).

Triangulation is not easy to achieve. Given a problem and a first dataset to begin examining it, there are several issues to consider as we identify and integrate additional data sources. Since most of what we do is enabled through multiple digital technologies, it is challenging to define boundaries in regard to what should be considered relevant data (Hsiang et al., 2020; Latif et al., 2020). In our example, for instance, depending on the specific problem at hand, mobile phone data can be complemented through information from search engines (Ginsberg et al., 2009), social media platforms (Tsao et al., 2021), travel websites (Kogan et al., 2019), wearable devices (Radin et al., 2020), among others (Kogan et al., 2021), with the timeframe and focal objects (e.g., users) being the pillars around which to integrate the data. Kogan et al. (2021) exemplify how the integration of several disparate health and behavior-related digital trace data sources––such as data from Twitter, Google, Apple, clinical databases, and others––provides a multiproxy estimate of COVID-19 outbreak with much higher accuracy than mobile phone data alone.

Obtaining and combining multiple data sets for purposes of triangulation can be challenging. Relevant digital trace data sources may stretch across multiple digital platforms (Ienca and Vayena, 2020). Some data may be publicly accessible (e.g., open science initiatives or linked open data, see GovLab, 2021); others may be owned by companies and thus require a strong and trustful partnership between these companies and researchers (Power, 2022). Furthermore, data sets might be flawed (Alsudais, 2021) and data from different datasets are rarely standardized in terms of their semantics. They thus may need to undergo quality checks and considerable pre-processing before analyses can be performed (Jain et al., 2010). For example, events in digital trace data might be logged at different levels of granularity with unclear semantic linkages between datasets (Baier et al., 2018). Given all of this, it is key for the development of situated explanations to collect and combine multiple datasets in quick and reliable ways. Our fourth principle is:

Principle 4:

Situated explanations should include triangulated digital trace datasets that comprehensively capture real-world problems.

Moving from processual patterns to situated explanations together with relevant stakeholders

Until now, we have focused on highlighting several critical challenges in generating processual patterns as dynamic and informative representations more quickly and more reliably. However, situated explanations require us not only to represent but also to understand the idiosyncratic features of a given problem. This involves “investigating […] how we can make sense of what happened, and how such an understanding might help develop policy/strategy” (Hirschheim, 2019: 1340). To quickly and efficiently make sense of what happened and what is represented in data and patterns, we thus suggest including into the sensemaking process the actors who are involved in the problem that we are trying to understand. That is, we propose that only a substantial involvement of relevant stakeholders will allow us to leverage the processual patterns generated from digital trace data as a basis for arriving at situated explanations of what is going on.

Sensemaking allows establishing meaningful correlations in data (Desouza and Jacob, 2017). Various approaches to sensemaking exist (e.g., Desouza and Jacob, 2017; Elgendy and Elragal, 2016; Tay et al., 2018). Research on big data analytics points to different factors that facilitate sensemaking and the development of explanations on the grounds of various forms of representations. For example, representations are effective when those who are exposed to them are familiar with the problem context and pursue a certain goal. The interpretation of representations and the evaluation of patterns can be facilitated by means of graphical aids (e.g., using different colors for different data properties in visual representations) (Sinar, 2015). Furthermore, interdisciplinary teams––involving policy makers but also data scientists, among others––have been found to facilitate a joint understanding about a problem context based on representations of various kinds (Desouza and Jacob, 2017; Oliver et al., 2020).

Turning to the example of COVID-19 containment, several studies emphasize that digital trace data alone was not a sufficient means to design appropriate interventions (Benjamins et al., 2022). What proved to be important was that different stakeholders, such as data analysts and policymakers, formed partnerships where they jointly learnt about the problem context, defined relevant parameters, as well as mutual expectations (Coudin et al., 2021; Oliver et al., 2020; Vespe et al., 2021). Different forms of representations turned out to be useful for different purposes and at different points in time (GovLab, 2021). Box 5 illustrates this point in relation to the example by the European Commission’s Joint Research Centre.

Box 5: Moving from representations to explanations: The case of the European Commission’s Joint Research Centre during the Covid-pandemic.

In the analysis stage, it was crucial that the digital trace data obtained through mobile phone data were triangulated with other information to make sense of and explain the resulting representations. For example, considering information about various interventions, such as school closings or public information campaigns, was essential to interpret representations based on digital trace data and decide on further actions (GovLab, 2021; Santamaria et al., 2020).

Furthermore, it was recognized that various stakeholders were important in this process. The European Commission, for example, explicitly expected that actions and decisions are based on insights by policymakers as well as scientists (European Commission, 2020). At the same time, post-evaluations of the bespoken initiative by the EU as well as several others (Benjamins et al., 2022; GovLab, 2021) showed that various stakeholders were important for making sense of problem contexts as well as defining appropriate actions; these included public health bodies, government agencies and data scientists. These post-evaluations also found that various capabilities were needed to inform actions on the grounds of mobile phone data, such as the capability to translate insights gained from representations into accessible language, or the capability to form and lead collaborative networks that bring in various data sources and analyze them from different angles (GovLab, 2021).

We suggest that developing situated explanations should involve collaborations with relevant stakeholders as co-equal partners in the research process (Hovorka et al., 2014; Tremblay et al., 2021). Practitioners, such as policymakers, are less concerned about developing theoretical explanations for a given problem (Hirschheim, 2019) but they possess knowledge that lies “in the action” (Bartunek and Rynes, 2014). They are able to point to aspects of a problem that seem important while dealing with it, which, in turn, can help researchers gain a deeper understanding of that problem (McKelvey, 2006) and identify additional data sources. They can also help devising interventions from emerging explanations and they simultaneously allow evaluating the effectiveness of the explanations and treatments (Oliver et al., 2020; Schwartz, 2014). This translational process is common in the professions such as law or medicine but it has so far been neglected in information systems (Schwartz, 2014). But when researchers and other stakeholders work on a problem together, they can address, clarify and advance solution strategies on an on-going basis, as they are accounting for their mutual perspectives (Bartunek and Rynes, 2014). Immersing into practical contexts helps researchers to adequately understand ongoing dynamics and see what is not immediately obvious (Rosen, 1991). Taken together, this leads us to our final principle:

Principle 5:

Situated explanations should be developed through co-equal partnerships with relevant stakeholders.