In which type of qualitative research do the researchers intend to generate a theory that based on data systematically gathered and analyzed?

5 Current Emphases and Future Directions

The grounded theory method has significantly influenced the development of qualitative research in the social sciences and professions, particularly nursing and education. Strauss and his colleagues trained several generations of graduate students in sociology and nursing, whose students and colleagues have subsequently adopted the grounded theory method. In sociology, this method has, perhaps, had most influence in medical sociology and social studies of science (see Baszanger 1998, Clarke 1998, Star 1989). However, recent debates between Glaser and Strauss and Corbin, as well as larger epistemological debates about scientific inquiry, have intensified interest in examining and using the method. Although questions have been raised concerning objectivist premises within grounded theory, a constructivist revision of the method provides answers to them and points to new directions. In short, the grounded theory method can be expected to gain renewed support from those social scientists and professionals who value conceptual analysis of rigorous empirical qualitative research.

11.4.1.1 Grounded theory and emergent coding

If you are working on a new topic that has very limited literature to build on, you may not be able to find established theories that allow you to develop the coding categories in advance. In this case, the emergent coding approach, based on the notion of grounded theory, is appropriate. Grounded theory was first proposed by Glaser and Strauss (Glaser and Strauss, 1967), who described a qualitative research method that seeks to develop theory that is “grounded in data systematically gathered and analyzed” (Myers, 2013). Grounded theory is an inductive research method that is fundamentally different from the traditional experimental research methods described in Chapters 2 and 3. As demonstrated in Figure 11.1, when conducting experimental research, we normally start from a preformed theory, typically in the form of one or more hypotheses, we then conduct experiments to collect data and use the data to prove the theory. In contrast, grounded theory starts from a set of empirical observations or data and we aim to develop a well-grounded theory from the data. During the process of theory development, multiple rounds of data collection and analysis may be conducted to allow the underlying theory to fully emerge from the data (Myers, 1997; Corbin and Strauss, 2014). Therefore, some researchers refer to the theory generated using this method as the “reverse-engineered” hypothesis.

Grounded theory can be applied to a variety of research methods discussed in this book such as ethnography (Chapter 9), case studies (Chapter 7), and interviews (Chapter 8). The major difference between qualitative research strategies that are mainly descriptive or exploratory and grounded theory is its emphasis on theory development from continuous interplay between data collection and data analysis.

Because grounded theory does not start from a preformed concept or hypothesis, but from a set of data, it is important for researchers to start the research process without any preconceived theoretical ideas so that the concepts and theory truly emerge from the data. The key to conducting successful grounded theory research is to be creative and have an open mind (Myers, 2013). Since grounded theory was first proposed in 1967, opinions on how to conduct research using grounded theory have diverged (Glaser, 1992; Strauss, 1987; Corbin and Strauss, 2014). The founders disagree on whether grounded theory can be formalized into a set of clear guidelines and procedures. Glaser believes that procedures are far too restrictive and may contradict the very basis of this method: creativity and an open mind. Even with the public disagreement, the procedures and guidelines proposed by Strauss and Corbin have been widely used in the field of social science, probably partly due to the fact that the procedure makes grounded theory more tangible and easier to implement. We briefly introduce the procedures of grounded theory according to Corbin and Strauss (Corbin and Strauss, 2014).

The grounded theory method generally consists of four stages:

open coding;

development of concepts;

grouping concepts into categories;

formation of a theory.

In the open coding stage, we analyze the text and identify any interesting phenomena in the data. Normally each unique phenomenon is given a distinctive name or code. Given a piece of text to analyze, you would read through, trying to identify the patterns, opinions, behaviors, or other issues that sound interesting. Since you are not constrained by preestablished theories, frameworks, or concepts, you are open to all possibilities that reside in the data.

During this process, you need to find terms to describe the interesting instances that emerge from the data. Sometimes the participants may provide terms that describes the instances or key elements so vividly or accurately that you can borrow the term directly. Coding categories generated in this manner are called in vivo code. In vivo coding can help ensure that the concepts stay as close as possible to the participants’ own words. These types of codes are largely adopted when using the grounded theory method. In one survey that the authors conducted on computer usage by children with Down syndrome, we borrowed many terms (e.g., curriculum integration) directly from parents’ response and used them as low-level themes (Feng et al., 2010).

When the original text does not contain a key term to describe the instance of interest, the researcher will need to find an appropriate term to describe the instance. Those terms are called “researcher-denoted concepts.” For example, if you read the following descriptions in the data, you may use the term “frustration” to describe the underlying theme of both responses:

My son just sits there and sobs when the computer does not do what he wants.

He becomes irritated and keeps pushing the Enter button when the web page loads slowly.

In the second stage, collections of codes that describe similar contents are grouped together to form higher level “concepts,” which can then be grouped to form “categories” (the third stage). Definitions of the concepts and categories are often constructed during this phase of the analysis. The identification and definition of relationships between these concepts—a process often referred to as “axial coding” (Preece et al., 2015; Corbin and Strauss, 2014) is often a key step in this process. As analysis continues, we are constantly searching for and refining the conceptual construct that may explain the relationships between the concepts and categories (Glaser, 1978).

Although this description implies a linear process with well-defined phases, analyses might not be quite so clear-cut. The identification of new codes through open coding, the grouping of these codes into categories, and the definition of relationships between these codes is a complex process involving the evolving construction of an understanding of the data. Iterative review of the data is often a key part of the process, as identification of new codes and categories might lead you to rereview documents from the perspective of codes identified in later documents. This rereview might also suggest multiple categorizations or types of relations between codes.

In the last stage, theory formulation, we aim at creating inferential and predictive statements about the phenomena recorded in the data. More specifically, we develop explicit causal connections or correlations between the concepts and categories identified in the previous stages. This process might be followed by selective coding, in which previously coded data might be revisited from the context of the emerging theory. Of course, further iteration and identification of open codes or axial codes is also possible.

A study of the issues involved in building information technology support for palliative care provides an instructive example of the use of emergent coding and grounded theory. Noting that palliative care differs significantly from other forms of medical care in a focus on the individual needs of each patient, Kuziemsky and colleagues conducted a grounded theory analysis of multiple data sources, including 50 hours of interviews with seven professionals (nurses, physicians, and counselors), patient charts, and research literature. Figure 11.2 provides an example of open and axial codes used in this analysis. This coding process was used to form a more detailed map of relationships between factors important to palliative care (Kuziemsky et al., 2007).

While conducting research using grounded theory, it is important to fully understand the advantages and limitations of this research method. Grounded theory obviously has a number of advantages. First, it provides a systematic approach to analyzing qualitative, mostly text-based data, which is impossible using the traditional experimental approach. Second, compared to the other qualitative research methods, grounded theory allows researchers to generate theory out of qualitative data that can be backed up by ample evidence as demonstrated in the thorough coding. This is one of the major attractions of the grounded theory and even novice users found the procedure intuitive to follow. Third, grounded theory encourages researchers to study the data early on and formulate and refine the theory through constant interplay between data collection and analysis (Myers, 2013).

On the other hand, the advantages of grounded theory can become disadvantages at times. It is not uncommon for novices to find themselves overwhelmed during the coding stage. The emphasis on detailed and thorough coding can cause researchers to be buried in details and feel lost in the data, making it difficult to identify the higher-level concepts and themes that are critical for theory formulation. In addition, theories developed using this method may be hard to evaluate. Unlike the traditional experimental approach in which the hypothesis is clearly supported or rejected by quantitative data collected through well-controlled, replicable experiments, grounded theory starts from textual information and undergoes multiple rounds of data collection and coding before the theory fully emerges from the data. The evaluation of the outcome depends on measures that are less direct, such as the chain of evidence between the finding and the data, the number of instances in the data that support the specific concept, and the familiarity of the researcher with the related topic. Lastly, the findings of the grounded theory approach may be influenced by the researchers' preconceived opinions and, therefore, may be subject to biases. In order to avoid these issues from happening, researchers should always keep in mind the key of this approach: being creative and open minded; listening to the data. When there is a gap between the concept and the data, additional data need to be collected to fill in the gap and tighten the linkage between the concept and the data. Due to these limitations, some researchers prefer to use grounded theory just as a coding technique, not as a theory generation method. For a detailed exploration of these and many other issues relating to the use of grounded theory in qualitative analysis, see “The SAGE Handbook of Grounded Theory” (Bryant and Charmaz, 2007) and Corbin and Strauss’ classic text Basics of Qualitative Research: Techniques and Procedures for Developing Grounded Theory (Corbin and Strauss, 2014).

13.4.4 Data analysis strategy

Our data analysis was guided by the Grounded Theory method of Charmaz (2008). It is a qualitative approach applied broadly in social sciences to construct general propositions (called a “theory” in this approach) from verbal data. It is exploratory and well suited for situations where the researcher does not have preconceived ideas, and instead is driven by the desire to capture all facets of the collected data and to allow the theory to emerge from the data. In essence, this was a process of making analytic sense of the interview data by means of coding and constant/iterative comparison of data collected in the case study (Figure 13.1).

Constant comparison means that the data from an interview is constantly compared to the data already collected from previously held interviews. We first read the interview transcripts and attached a coding word to a portion of the text—a phrase or a paragraph. The “codes” were selected to reflect the meaning of the respective portion of the interview text to a specific research question. This could be a concept (e.g., “willingness to pay”) or an activity (e.g., “operationalization,” “quantification”). We clustered all pieces of text that relate to the same code in order to analyze it in a consistent and systematic way. The results of the data analysis are presented in the next section, after which a discussion is added.

2 Other Methods Based on Mead

While Blumer's adaptation of Mead's theories is the methodological mainstay of SI, there are other methodologies based on SI, and these will be mentioned next. Analytic induction was first discussed by Znaniecki (1928). It was later used, with minor variations by Lindesmith (1937, 1968) (he was a graduate student of Blumer), Cressey (1950) (a student of Lindesmith), Becker (1963) (see Hammersley 1989), and others. Analytic induction, according to Znaniecki, recognizes the fact that objects in the world are open to an infinite number of description and, thus, our account of them must be selective; this selectivity will be based on the interest at hand, which for sociologists is primarily social and cultural systems; commonly used sociological methods relying on pre-identification (deductive) or superficial description (inductive) will not work, only analytic induction will accomplish the task. The researcher will select a small number of cases (10–12, usually) and study them in depth, continually defining and redefining the event and formulating and reformulating theoretical propositions until they will fit all cases. Negative cases must also be examined (this was Lindesmith's idea).

Another student of Blumer, Strauss, together with Glaser, developed another SI method, grounded theory (Glaser and Strauss 1967). Closely related both to Blumer's methodology and to analytic induction, grounded theory placed more emphasis on the generation and development of theory. Relying on the inductive method, grounded theory is akin to Blumer's inspection, only much more elaborate. Rather than relying on a priori population, in analytic theorizing one continues to study new cases until the point of saturation, generating theoretical categories.

Not all SI methods followed the constructionist approaches outlined above. A notable exception came from the Iowa School of Sociology. Kuhn (1964a) adopted a much more deterministic approach to Mead's discussion of the self and the nature of the ‘me,’ the various roles and images we have of ourselves. The methodology he adopted to discover the nature of the self was called the Twenty Statements Test (TST), a series of open-ended questions about the self. Kuhn felt that rather than use the oblique method of observing people one ought to ask them directly about the nature of their inner feelings and they would honestly disclose them to the researcher. The results of TST would be used, by Kuhn, to outline generic laws that would apply to human beings in different situations. Other positivistic oriented symbolic interactionists are Sheldon Stryker, described as a ‘structural role theorist,’ who influenced numerous students at the University of Indiana and Carl Couch, who was a stalwart of the discipline, with his ‘Behavioral Sociology’ at the University of Iowa (cf. Reynolds 1993).

In the 1960s and 1970s a plethora of theoretical approaches, largely based on the naturalistic method, appeared. Some were based on basic Meadian tenets, such as dramaturgy (Goffman 1959), and labeling (Becker 1963). Others based their constructionist approach not only on the ideas of Mead but on those of the phenomenologists (Husserl, Schutz, Heidegger, Dilthey) and the existentialists (Merleau-Ponty, Sartre), and ordinary language philosophers (Wittgenstein). They are phenomenological sociology, existential sociology, ethnomethodology, and the sociology of emotions (see Douglas et al. (1980) for a survey of these sociologies and a list of references to them; also, see Adler et al. (1987)).

3.3 Manual Coding

Once we have collected textual data for analysis and interpretation, it needs to be structured and classified. This classification is often referred to as coding, where we identify patterns in texts, having an explanatory or a exploratory purpose [1] and serving as a basis for further analysis, interpretation, and validation. Coding can be done in two ways: manually or automated. In this section, we introduce coding as a manual process. A detailed example for applying the manual coding process is provided in Section 3.5.1.

Although manual coding is often associated with interview research, the data we code is not limited to transcripts as we can structure any kind of textual data given in documents, Wikis, or source code (see also Section 3.2.2). This kind of structuring is used in social science research and is also gaining attention in software engineering research. An approach commonly used in these research areas is grounded theory. We briefly describe grounded theory in Sidebar 1, but its theoretical background is not necessary for many practical text analysis contexts.

Sidebar 1

Grounded theory in a nutshell

Manual coding, as discussed in this chapter, has its origins in grounded theory. Because grounded theory is the most cited approach for qualitative data analysis [1], which comes at the same time with a plethora of different interpretations, we briefly clarify its meaning in a nutshell. Grounded theory describes a qualitative research approach to inductively build a “theory”—that is, it aims to generate testable knowledge from data rather than to test existing knowledge [1]. To this end, we thus make use of various empirical methods to generate data, and we structure and classify the information to infer a theory. A theory, in its essence, “provides explanations and understanding in terms of basic concepts and underlying mechanisms” [2, 3]. In empirical software engineering, we mostly rely on the notion of a social theory [4], and refer to a set of falsifiable and testable statements/hypotheses. As most qualitative research methods, grounded theory has its origins in the social sciences, and it was first introduced in 1967 by Glaser and Strauss [5]. A detailed introduction to the background of grounded theory and the delineation with similar concepts arising along the evolution of grounded theory is given by Birks and Miller [1]. For the remainder of the chapter where we introduce a manual coding process, we rely on the terms and concepts as introduced in the context of grounded theory.

3.3.1 Coding Process

Manual coding is a creative process that depends on the experiences, views, and interpretations of those who analyze the data to build a hierarchy of codes. During this coding process, we conceptualize textual data via pattern building. We abstract from textual data—for example, interview transcripts or commit comments stated in natural language—and we build a model that abstracts from the assertions in the form of concepts and relations. During this coding process, we interpret the data manually. Hence, this is a creative process which assigns a meaning to statements and events. One could also say that we try to create a big picture out of single dots.

There are various articles and textbooks proposing coding processes and the particularities of related data retrieval methods such as why and how to build trust between interviewers and interviewees (see, e.g., Birks and Mills [1]). The least common denominator of the approaches lies in the three basic steps of the coding process itself followed by a validation step:

1.

Open coding aims at analyzing the data by adding codes (representing key characteristics) to small coherent units in the textual data, and categorizing the concepts developed in a hierarchy of categories as an abstraction of a set of codes—all repeatedly performed until a “state of saturation” is reached.

2.

Axial coding aims at defining relationships between the concepts—for example, “causal conditions” or “consequences.”

3.

Selective coding aims at inferring a central core category.

4.

Validation aims at confirming the model developed with the authors of the original textual data.

Open coding brings the initial structure into unstructured text by abstracting from potentially large amounts of textual data and assigning codes to single text units. The result of open coding can range from sets of codes to hierarchies of codes. An example is to code the answers given by quality engineers in interviews at one company to build a taxonomy of defects they encounter in requirements specifications. During open coding, we then classify single text units as codes. This can result, for example, in a taxonomy of defect types, such as natural language defects which can be further refined, for example, to sentences in passive voice. During axial coding, we can then assign dependencies between the codes in the taxonomies. For example, the quality engineers could have experienced that sentences in the passive voice have frequently led to misunderstandings and later on to change requests. The axial coding process then would lead to a cause-effect chain that shows potential implications of initially defined defects. The final selective coding then brings the results of open and axial coding together to build one holistic model of requirements defects and their potential impacts.

We subsequently form a process that we applied in our studies and which worked well for us. Figure 3.3 depicts the basic coding process and further steps usually (or ideally) performed in conjunction with the coding process.

Figure 3.3. Coding process.

The idea of (manual) coding—as it is postulated in grounded theory—is to build a model based on textual data—that is, “grounded” on textual data. As the primary goal is to gather information from text, we need to follow a flexible process during the actual text retrieval and the coding process as well. For example, in the case of conducting interviews, we perform an initial coding of the first transcripts. If we find interesting phenomena for which we would like to have a better understanding of the causes, we might want to change the questions for subsequent interviews; an example is that an interviewee states that a low quality of requirements specifications is also connected with low motivation in a team, leading to new questions on what the root causes for low motivation are. We thereby follow a concurrent data generation and collection along with an emerging model which is also steered according to research or business objectives.

Figure 3.4 shows the coding steps for our running example. During the open coding step (lower part of the figure), we continuously decompose data until we find small units to which we can assign codes (“concept assignment”). This open coding step alone shows that the overall process cannot be performed sequentially. During the open coding step, we found it useful

Figure 3.4. Manual coding of our running example.

•

to initially browse the textual data (or samples) before coding them to get an initial idea of the content, meaning, and finally, of potential codes we could apply,

•

to continuously compare the codes during coding with each other and especially with potentially incoming new textual data, and

•

to note down the rationale for each code to keep the coding process reproducible (of special importance if one is relying on independent recoding by another analyst).

Having a set of codes, we allocate them to a category as a means of abstraction. In our running example, we allocate the single codes to the categories “entities” and “dependencies.” During axial coding, we then assign directed associations between the codes. Finally, the last step in the coding process is supposed to be the identification of the core category, which often can also be predefined by the overall objective; in our case, it is “Operation.”

The overall coding process is performed until we reach a theoretical saturation—that is, the point where no new codes (or categories) are identified and the results are convincing to all participating analysts [1].

3.3.2.1 Data Analysis with the Grounded Theory.

The grounded analysis was used to provide insight into the software organizations, their software processes and testing activities. By interviewing people in different positions from the software organization, the analysis could gain additional information on testing-related concepts, such as different testing phases, test strategies, testing tools and case selection methods. Later, this information was compared between organizations, allowing hypotheses on the test process components from several viewpoints and from the test process itself as a whole.

The Grounded Theory method contains three data analysis steps: open coding, axial coding, and selective coding. The objective for open coding is to extract the categories from the data, whereas axial coding identifies the connections between the categories. In the third phase, selective coding, the core category is identified and described [56]. In practice, these steps overlap and merge because the theory development process proceeds iteratively. Additionally, Strauss and Corbin state that sometimes the core category is one of the existing categories, and at other times no single category is broad enough to cover the central phenomenon.

The objective of open coding is to classify the data into categories and identify leads in the data, as shown in Table V. The interview data were classified into categories based on the main issue, with any observation or phenomenon related to it being the codified part. In general, the process of grouping concepts that seem to pertain to the same phenomena is called categorizing, and it is done to reduce the number of units to work with. In this study, this was done using ATLAS.ti software [77] that specializes on the analysis of qualitative data. The open coding process started with “seed categories” [75] that were formed from the research subquestion the publication was studying and prior observations from the earlier publications. Overall, the analysis process followed the approach introduced by Seaman [68], which notes that the initial set of codes (seed categories) come from the goals of the study, the research questions, and predefined variables of interest. In the open coding, we added new categories and merged existing categories into others, if they seemed unfeasible or if we found a better generalization.

Table V. Example of Codification Process

After collecting the individual observations into categories and codes, the categorized codes were linked together based on the relationships observed in the interviews. For example, the codes “Software process: Acquiring 3rd party modules,” “Testing strategy: Testing 3rd party modules,” and “Problem: Knowledge management with 3rd party modules” were clearly related and therefore could be connected together in the axial coding. The objective of axial coding is to further develop categories, their properties and dimensions, and find causal, or any other kinds of connections between the categories and codes. For some categories, the axial coding can also include actual dimension for the phenomenon, for example, “Personification–Codification” for “Knowledge management strategy”, or “Amount of Designed Test Cases vs. Applied” with dimension of 0%–100%, where every property could be defined as a point along the continuum defined by the two polar opposites or numeric values. Obviously, for some categories, which were used to summarize different observations such as enhancement proposals, opinions on certain topics or process problems, defining dimensions was unfeasible.

Our approach to analysis of the categories included Within-Case Analysis and Cross-Case Analysis, as specified by Eisenhardt [76]. Basically, this is a tactic of selecting dimensions and properties with within-group similarities coupled with intergroup differences based on the comparisons between different research subjects. In this strategy, one phenomenon that clearly divided the organizations into different groups was isolated, and looked into for more details explaining differences and similarities within these groups. As for one central result, the appropriateness of OU as a comparison unit was confirmed based on our size difference-related observations on the data; the within-group- and intergroup comparisons did yield results in which the company size or company policies did not have strong influence, whereas the local, within-unit policies did. In addition, the internal activities observed in OUs were similar regardless of the originating company size, meaning that in this study the OU comparison was indeed a feasible approach.

Each chain of evidence was established and confirmed in this interpretation method by discovering sufficient citations or finding conceptually similar OU activities from the case transcriptions. Finally, in the last phase of the analysis, in selective coding, the objective was to identify the core category—a central phenomenon—and systematically relate it to other categories and generate the hypothesis and the theory. Overall, in theory building, the process followed the case study research described by Eisenhardt and its implementation examples [69,78].

The general rule in Grounded Theory is to sample until theoretical saturation is reached. This means, until (1) no new or relevant data seem to emerge regarding a category, (2) the category development is dense, insofar as all of the paradigm elements are accounted for, along with variation and process, and (3) the relationships between categories are well established and validated [56]. In this study, saturation was reached during the third round, where no new categories were created, merged, or removed from the coding. Similarly, the attribute values were also stable, i.e., the already discovered phenomena began to repeat themselves in the collected data.

6.3.1.4 File-level proxy measures of effort

Toward our goal of providing an empirical foundation to support decisions on modularization activities, we first propose and justify the adoption of three-dimensional proxy measures for effort: actions—the total number of patches and commits made to a file for the purpose of addressing issues; churn—the number of lines of code changed in a file for the purpose of addressing issues; and discussions—the number of textual comments made to address issues and related to a file. These measures can be obtained from both open-source and closed-source projects. We hypothesize that these proxy measures have the potential to bridge the gap between the variation of file metrics and file-level maintenance effort.

In our search for proper measures of effort, we considered various factors, including granularity and credentials. To fit the need of the envisioned DSS system, proper measures of effort should be at the file level because source files are the target of refactoring and manifest modularity decisions. The traditional measure of effort, in the unit of person-hours, person-days, or person-months, is associated with tasks or subsystems, and is too coarse-grained for file-level decisions and hard to attribute to files. We chose actions and churn based on the prior work on software cost estimation. The use of discussion as a measure of effort is supported by application of the grounded theory method to open-source software.

Churn. Lines of code represents probably the most widely and longest used measure of effort in software engineering. Although this measure of effort is not perfect, it has been the output of traditional software effort estimation models, such as the CoCoMo family of models (Boehm, 1981). We thus use the total lines of code added, deleted, or changed, commonly referred to as churn, as one type of effort measure. Recently, churn has been intensively investigated to predict error-proneness in source files (Giger et al., 2011).

Actions. Although code churn can be easily identified and counted, the change volume of a file from one version to the next does not always accurately represent the effort spent on the modification. It is possible that some lines of code require far more effort to create or understand than others. In open-source software (OSS) projects, in order to address an issue, contributors may propose many patches, each of which may change some code in some source files. Some of the patches may be accepted and others may be rejected. Simply counting the code churns between versions would not capture the full complexity of multiple patches, which also represent effort. Accordingly, we chose to use the number of patches and commits as another complementary measure of effort: the more patches/commits that are needed to address an issue, the greater the effort. We refer to this effort measure as actions.

Discussions. As we have discussed, not all lines of code are equal. There are times when programmers struggle over dozens of lines of code, or hundreds lines of code can be created relatively easily. This difference is partially caused by the inherent complexity of the issue being addressed. In an open-source project, such issues are often discussed among the developers. Accordingly, we consider that the amount of discussion associated with a file can be another useful measure of effort, which we call discussions.

Our reasoning is based on a grounded theory (GT) investigation. Grounded theory was first proposed by (Glaser and Strauss, 1967) as a methodology in the social sciences to discover theory by forming hypothesis and concepts from collected data. In OSS projects, development activities are usually logged in textual form via issue-tracking systems, commit messages, and mailing lists. These textual corpuses formed the dataset needed for GT methodology. We first investigated the discussions in Apache Lucene and then extended our dataset to include multiple Apache projects. The raw data for the analysis include commit titles, JIRA archives, IRC discussions, and mailing lists.

Using several GT analysis techniques, we first obtained groups of similar concepts needed to generate a theory. After that, we augmented the data and refined the derived concepts using additional projects. Using this method, a core category (theory)—iterative informed consent (IIC)—emerged. IIC identifies a behavior pattern of contributors: They iteratively discuss and provide solutions (tests and patches) to achieve consent. After the core category was identified, we observed many forms of supporting evidence in many OSS communities. For example, specific patches would show up as part of proposed implementation solutions, for the purpose of clarifying or strengthening the chances of consent. Although is not always the case since a solution can be directly submitted by developers with commit permission, it is the norm in OSS communities. We thus believe that the results of our GT analysis support the choice of discussions as another effort proxy.

We also considered other possible proxy measure of maintenance, including the time spent to resolve an issue and the number of bugs within a file. We decided not to use time as an effort measure first because the time elapsed during the life cycle of issues often reflect not only the actual effort caused by file complexity, but in most cases, the priority of a bug. For example, a ticket may remain open for a long time because of its low priority rather than inherent difficulty.

Unlike other work in the area of defect prediction (e.g., Hassan, 2009) we chose not to use defect rate as an effort proxy first because not all defects are addressed and bug reports may have duplications. Moreover, if a defect is fixed, then fix has to be manifested as changes in source file, patches, commits, and discussions, which will be captured by our three effort measures. On the other hand, if a defect is not fixed (no churn), or if no attempt is made to fix it (no actions and/or no discussions), then there is no evidence of any effort spent on fixing these bugs. Accordingly, only considering the number of bugs is not sufficient to capture the effort caused by source file complexity.

Stance on philosophy of science

The success of grounded theory (GT) in the social sciences has led to the development of various variants with different underlying ontological and epistemological assumptions and coding procedures (Goldkuhl and Cronholm, 2019; Wiesche et al., 2017). Researchers sometimes distinguish between Glaserian GT, which is perceived as rather objectivistic and thus closer to a positivistic worldview, whereas Straussian GT is perceived as rather constructivist and thus closer to an interpretive worldview (Goldkuhl and Cronholm, 2019). Yet, “the assertion that GTM [grounded theory method] is positivist, interpretive, critical realist, or constructivist is neither supported by the grounded theory literature, nor based on research practice. GTM is in many ways neutral and should be seen as a container into which any content can be poured” (Urquhart and Fernández, 2013, p. 229). Thereby, GT may be able to “bridge traditional positivistic methods with interpretative methods” (Charmaz, 2010, p. 30) and it is the responsibility of the researcher to clarify their underlying epistemological assumptions (Urquhart and Fernández, 2013). In the context of this literature review, we adopted an interpretivist stance since the approach to reviewing literature as proposed by Wolfswinkel et al. (2013) is organized around two key concepts: comparative analysis (constantly comparing the underlying data with emerging concepts, categories, and their interrelations) and theoretical sampling (insights from already read articles inform the analysis of the remaining articles). Both concepts are inconsistent with purely positivist assumptions since they challenge the clear separation of data collection and data analysis, leading to an iterative understanding of the topic, rather than an a priori defined understanding in form of hypotheses (Suddaby, 2006).

3.2 Synthesis results

In this review, we put emphasis on mHealth data protection and argue that it should be understood as a complex phenomenon with substantial impacts in the context of healthcare eco-systems. To present a realist view of mHealth data protection failures and successes we have systematically analyzed and coded the content of the included articles, guided by Template Analysis and Grounded Theory methods.

In the coding process, we sought insights into the mechanisms of data protection failures and successes and how these mechanisms led into outcomes in the context of mHealth. We followed Gioia, Corley and Hamilton [34]’s approach for qualitative rigor and cycled between emergent concepts and themes from our synthesis and the relevant cybersecurity literature. The results of our grounded theory coding are presented in Table 2.

Table 2. Grounded theory-informed coding.

To demonstrate a theoretical model, we integrated the result of template analysis, contextual factors (see Appendix Table A4), and our grounded theory-informed coding. With this integrative approach, we were able to identify the patterns of failures and successes (see Appendix Table A5). A diagrammatic overview of the key results of the theoretical integration is provided in Fig. 2. Our approach conceptually formulates a grounded model of failures and successes in mHealth data protection. The model is emerged from a realist synthesis and is structured by CMO configurations. This theoretical model represents a subset of the phenomenon (mHealth data protection) in the real digital world [50]. In the following sections, we contextually explain and discuss the distinctive mechanisms that lead to the failures and successes, as well as the subsequent impacts.