Beyond Text

2.1. The Microarray: A Potential Tool from an Unlikely Source

What might potential tools and techniques related to the microarray look like? Although ethnographers use a unique set of methods in their studies of social life, sharing and analyzing large volumes of context-dependent data is not a challenge unique to the social sciences. In molecular biology, a technique known as microarray analysis has proved powerful because it uses an interpretable heatmap visualization to help analyze and depict complex multilevel biological systems and processes to varied audiences. The term microarray refers to both a process for analyzing biological samples—typically patterns of gene expression in tissues—and a graphical product displaying results (Belacel, Wang, and Cuperlovic-Culf 2006; Eisen et al. 1998; Schena et al. 1995). The introduction of microarrays and their exploratory use has led to important advances. For example, microarrays helped scientists identify genetic patterns (overexpression vs. underexpression) in breast cancer tumors by analyzing and displaying expression profiles for a large number of tumor samples simultaneously (see Figure 1), differences that can help explain the course of illness, distribution within populations, and responsiveness to different types of therapy (Prat and Perou 2011).

OPEN IN VIEWER

Figure 1.

Microarray based on gene expression profiling data from 337 breast samples (in columns; 320 tumors, 17 normal tissues) and approximately 1,900 genes (rows).

Using microarrays, biologists can display, aggregate, analyze, and share complex, multilevel data using exploratory statistical procedures, such as principal component analysis (PCA), that allow systemic inductive identification of group boundaries and pattern recognition (Stears, Martinskey, and Schena 2003). Figure 1 provides an existing example using gene expression data from breast tissue specimens. ⁷ At the same time, the microarray retains microlevel information about individual specimens so that analysts do not lose context. Analysts can “zoom in” to examine characteristics of an individual case in the array as readily as they can “zoom out” to see how that case fits within the array’s overall pattern. Incorporating individual-level data within the microarray means that molecular biologists can use arrays not only to share findings but also to share the data and process by which those findings have been generated from voluminous underlying data.

Ethnographic field notes are quite different from the gene expression profiles found in microarrays. The former are typically more interpretative; the latter are expressed via quantitative reduction. Ethnography necessarily involves self-reflection; there is no directly comparable activity for biologists. But analyzing either kind of data requires scholars to shift their gaze between distinct analytical levels and to represent their interpretations to a wider audience. Biologists use the microarray to examine genes, markers, individuals, and populations. Ethnographers examine microlevel interactions, emergent themes, theoretical constructs, and social contexts, and in this way they engage their sociological imagination to explore connections among and between behavior and narrative, group and organization, institution and society (Mills 1959).

3.1. Developing an Ethnoarray

Using preliminary data from the PtDelib study, we developed a small-scale model of an ethnoarray. All of the data we use in this mock-up are drawn from field notes and interviews we had previously entered and coded via an iterative interpretive analysis using the ATLAS.ti software platform. An example of the coded database is presented in Figure 2, which shows a single paragraph from the transcript of a PtDelib patient interview. As this figure illustrates, passages of text typically include many codes. The software allows analysts to flexibly search the database to retrieve passages of interest using Boolean search procedures and even basic inductive tools such as co-occurrence tables. However, given that the ethnoarray involves a new approach, current software packages are not designed at this time to directly facilitate the production of ethnoarrays. In translating the data into an array, our first—and most analytically consequential—decision for this mock-up was to focus the ethnoarray on analyzing patients’ trajectories into clinical trials. That is, we chose to organize this array to facilitate understanding differences and similarities between individuals. ⁸ After selecting the unit of analysis for the array, we chose five substantive domains that prior literature and our early iterative interpretive analyses suggested were relevant, discussed with team members how to properly represent those domains, ultimately selected three to four measures for each domain, and arranged the domains and measures as the array’s 16 rows. ⁹

OPEN IN VIEWER

Figure 2.

A single paragraph from the transcript of a Cancer Patient Deliberation Study patient interview, coded as an ATLAS.ti data set.

For columns, we selected a sample of patients for whom we had sufficient data (i.e., for whom we had at least two interviews and field observations before they either died or left the study). We then debated how to represent time variation in their experiences. In Figure 3, we show an array in which all information has been aggregated into a single column (to create a 10-column array); Figures 4 to 6 show arrays in which patients’ experiences and statuses at different times are shown in distinct columns (baseline [T1] and first follow-up [T2]), which creates a larger and perhaps less intuitively interpretable array that includes greater richness about patients’ experiences and trajectories. The key to Figures 3 to 6 is provided below. Each cell reflects all interview and participant observation data associated with that individual, domain, and measure. The domains, measures, and rows are ordered with a temporal logic following our particular research questions, but future arrays need not follow this model. The rows of a grounded theory approach, for example, might include general emergent themes generated entirely inductively. Columns could represent organizations, events, interaction sequences, or other units depending on the analyst’s goals. Appendix A shows a brief example of how ethnoarrays can use other units of analysis (such as neighborhood contexts) in a comparative participant observation study, but for the sake of clarity, we focus on individual trajectories from PtDelib data in the main text.

OPEN IN VIEWER

Figure 3.

An ethnoarray based on data from 10 participants in the Cancer Patient Deliberation Study. The “Key to Figures 3–6” on p. 288 provides additional information regarding the domains, measures, and color assignment.

OPEN IN VIEWER

Figure 4.

An ethnoarray based on data from 10 participants in the Cancer Patient Deliberation Study, with baseline (T1) and follow-up (T2) data. The “Key to Figures 3–6” on p. 288 provides additional information regarding the domains, measures, and color assignment.

OPEN IN VIEWER

Figure 5.

A sorted ethnoarray based on data from 10 participants in the Cancer Patient Deliberation Study. The “Key to Figures 3–6” on p. 288 provides additional information regarding the domains, measures, and color assignment.

OPEN IN VIEWER

Figure 6.

A sorted ethnoarray with selected quotations from linked data. The “Key to Figures 3–6” on p. 288 provides additional information regarding the domains, measures, and color assignment.

OPEN IN VIEWER

Key to Figures 3–6.

Overview of domains, measures, and color assignment.

The sample array we present is organized by discrete units (characteristics and experiences of individuals at a given point in time evidenced by field notes and interviews). This corresponds with the analytical goals of the PtDelib project but raises important points about array construction. First, which type of data can be included in arrays? In any study, analysts must answer this question on the basis of the particular research question. For illustrative purposes here, we included only traditional ethnographic data derived from field observations and interviews, but the larger study also includes data from medical records, focus groups, and surveys, which could also be incorporated. A related question is which portions of data become part of the array. Again, analysts will make this decision on the basis of the nature of their study and question. Their decision will likely reflect the specific epistemic and intellectual tradition within which they situated their work.

For the sample array in this article, we used a broadly interpretive approach. We examined coded data in the ATLAS.ti database and narrative summaries of each patient’s experience, and we held discussions among the team of researchers and fieldworkers who had firsthand knowledge of the patients, providers, and clinics represented in the array. Within this broad contextual framework, we interpreted specific interview passages and field notes according to whether and how they were related to array domains. We used all such passages in constructing the arrays in this article. The temporal structure of the array reflects the longitudinal design of the study, in which interviews and observations were conducted in a coordinated sequential fashion. Others who use ethnoarrays need not follow this model. Analysts might use only a single form of data (e.g., interviews or field notes). They might choose to deal with the question of inclusion differently as well. They might use formal linguistic tags to aid in categorization rather than relying on interpretive coding. “Uncoded” data (e.g., data that are not formally categorized according to substantive domains but are still associated with the columnar unit of analysis, such as individuals or neighborhoods) could be linked under a broad category labeled “other” (again, see Appendix A). Finally, researchers might decide that including all data for a person, site, and so on, is not feasible, ethical, or relevant. In this case, they would still be able to share summaries and still provide data beyond those in the traditional ethnographic report, but the array would not be inclusive. In short, like the quite varied notion of “coding,” the array is a flexible tool whose use depends largely on researcher decisions and justifications (i.e., what to examine, how to measure it, and how to represent it) that parallel those found in other forms of social science (Cicourel 1964).

With a unit of analysis selected, inclusion criteria identified, and rows and columns defined and ordered, a final design decision concerns how to assign colors to the resultant cells in an analytically useful way. The goal of color coding in this array is not simply aesthetics; it is also to enable a visual summarization to facilitate examining data patterns. To this end, the model ethnoarray features a three-color matrix that indicates the degree to which a given characteristic is present (blue = less; gray = typical or unremarkable among study participants; red = more). These colors were based on the team’s review, discussion, and interpretation of each patient’s case and associated transcripts and field notes. That is to say, similar to the procedures used for deciding on data inclusion, color assignment in this particular example was based on the interpretations of fieldworkers who were deeply familiar with the site and individuals rather than formalistic procedures or automated approaches for text mining. Existing coding in our CAQDAS data set, which already tagged text according to themes of interest in the project, were used to help identify information about each cohort member and provide a level of confidence that we were not overlooking relevant data as we developed our interpretations of patients’ experiences and understandings within the measures of each domain.

Embedding a traditional interpretation of rich ethnographic data within a structured tabular framework of domains and measures is, of course, only one approach to achieving a balance between more interpretive and formalistic approaches to analysis. ¹⁰ Other approaches may involve an explicit scoring procedure for determining cell color, for example, density or co-occurrence of codes from a CAQDAS database, linguistic algorithms, or word frequency counts. Color assignment schemes could be developed to indicate the absence of data for a particular theme of interest, for example, to capture different degrees of theoretical saturation or other types of unevenness in data collection that must be addressed in analysis. The tabular format may not be suitable for some projects, including those that do not have a clearly identified unit of analysis or take a more humanistic approach. Still, even within the broad spectrum of approaches falling under the umbrella of sociological participant-observation, many studies maintain a clear unit of analysis and examine variation within and across groups or contexts (cf. Abramson 2015; Cicourel 1968; Dohan 2003; Lareau 2011; Lutfey and Freese 2005; Sallaz 2009; Sánchez-Jankowski 2008).

Inset 1 outlines several different approaches to color assignment and illustrates how the ethnoarray can be used with a variety of analytical styles, from formal rules geared toward the quantification of observed behavior to flexible integration of interpretative insights. The ability to integrate and even combine these divergent approaches exemplifies the ethnoarray’s ability to accommodate different analytical approaches, goals, and styles from diverse intellectual and epistemic traditions.

OPEN IN VIEWER

3.2. Reading the Ethnoarray: The Experiences of Wayne Burley

Once constructed, the ethnoarray can be used in numerous ways to understand and represent large volumes of data. For analysts interested in examining and visually representing the experiences of specific patients to understand an outcome (e.g., whether they enter a clinical trial), data can be read along a single column or across the adjacent columns of a single patient’s baseline and follow-up data. This can be useful to both contextualize interpretive insights and to provide information about the validity of inferences drawn using statistical methods that can group like trajectories such as sequence analysis (SA).

For example, the first columns of Figure 4 illustrate the experiences of Wayne Burley ¹² (the array shows his study identifier number, 4020) derived from analysis of two in-depth interviews with him, an interview with his live-in girlfriend Heather Okeefe, and direct observation of two appointments around the time of the interviews with his oncologist, Antonio Akin, who was also interviewed (as well as observed in numerous other interactions with patients and colleagues). Wayne and Heather had been living on opposite coasts prior to his cancer diagnosis, and over the course of several months, he lived for short periods in three cities as he sought diagnosis and treatment for his cancer (a rare form of the disease; note the red cell under Health and Illness>Zebra diagnosis). He finally settled in northern California to make it easier for Heather to care for him, and the couple moved from a small one-bedroom apartment (where we conducted his baseline interview) to a larger two-bedroom apartment by the time of his follow-up interview (at which point we also interviewed Heather). Given this history, we characterized his Insurance and Finance>Housing as relatively unstable with respect to the others in our cohort at the baseline interview (blue cell) and more typical by follow-up (gray cell). Wayne had a long career with a government employer, retired early, and had begun a second career teaching public school when his cancer was discovered. Although no longer working because of his illness, Wayne’s employment history provided a stable pension and generous health benefits; we thus classified his Insurance and Finance>Finances and Insurance and Finance>Health insurance as higher than typical (red cells). Our interpretation of Wayne’s situation in the domains of Social Support, Health and Illness, Communication, and Decisions, drawn in comparison with dozens of other study participants, can be read in a similar fashion by examining the color of each relevant cell.

Wayne and Heather’s first visit with Dr. Akin was among the most contentious we have seen in the PtDelib study. Wayne relocated to northern California in part because Dr. Akin is acknowledged as an expert in his unusual cancer, but Wayne also received treatment from other oncologists. Before meeting Wayne and Heather for the first time, Dr. Akin reviewed Wayne’s medical record in the clinic hub room (a physicians’ work room out of earshot of patients) and commented to one of his colleagues that the other oncologists had been overly aggressive “cowboys” in their treatment approach. Although frank commentary is the norm in the hub room, we observed Dr. Akin repeat his “cowboy” comment to Wayne and Heather in the exam room. Their interactions became tense as a result, something Wayne and Heather commented on after the encounter. They both discussed this uncomfortable visit during their interviews but acknowledged that their relationship with Dr. Akin improved with time. They characterized him as a “straight talker” whose frank assessments of Wayne’s progress and prospects were valuable, and they brushed aside his more insensitive remarks. In the ethnoarray, this trajectory in their relationship is reflected in the Communication domain, which we coded blue at baseline (indicating atypically poor communication) and gray (typical communication) at follow-up.

The ethnoarray also reflects other changes between baseline (T1) and follow-up (T2) observations. Initially Wayne’s daily activities continued uninterrupted, and he believed that his life span would be unaffected by his illness (Health and Illness>Daily activities and Health and Illness>Live long time). At follow-up, he was experiencing substantial fatigue and was unable to do many of the things he had enjoyed just a few months previously (we characterized this as a blue cell for Health and Illness>Daily activities); like some of the patients in our study, at this point he acknowledged that his cancer would not be cured (Health and Illness>Live long time is now gray). Finally, examining the Decisions domain, we note that Wayne remained aggressive in his approach to his illness but that he seemed to be less interested in finding other doctors to manage his treatment; at his follow-up interview, he said he planned to stick with Dr. Akin. Although Wayne initially had said he was not interested in participating in a clinical trial during his baseline interview, a few months later he had begun to actively research trials to join (Decisions>Clinical trial is gray at T1 and red at T2). The resulting representation in the array summarizes key aspects of Wayne’s trajectory and provides a useful visualization that helps contextualize his experience relative to other subjects. This also facilitates further pattern analysis and possibilities for data sharing that we now examine. ¹³

3.3 Relational Mapping

The colored cells of the array can be used not only for reading narratives but also for mapping and understanding relationships among actors, institutions, and concepts, a fundamental goal for many ethnographic approaches. Take, for example, Wayne Burley’s relationship with his physician. The array visualization summarizes that Wayne’s trust in his physician changed over time, and on the basis of preliminary study data, it appears other patients have experienced similar shifts. Moreover, the array allows analysts to readily see that these relationships of trust occur not only between patients and their physicians but also in patients’ experiences as members of the health care team that includes physicians, nurses, and other health care providers. Analytic memos provide one way of documenting and interpreting an individual patient’s experiences of these relationships. The array provides a way to supplement those analyses by considering broader contextual elements that might also influence these experiences. ¹⁴

Our preliminary analyses suggest that patients’ experiences of trust and team membership reflect their estimations of physicians’ competence and the congruence between clinicians’ treatment preferences and their own, but these factors do not operate in a simple or mechanistic way. The array allows analysts to examine patients’ experiences of trust and team membership within a much broader context, for example, whether their cancer is affecting their daily activities, how the progression of illness over time shapes these relationships, and how patients’ own beliefs about whether their illnesses are life limiting color the trust and connection patients feel with their clinicians. For Wayne Burley, the progression of his illness, its impact on his daily activities, and the exhaustion of available treatments appeared to reshape his engagement with his oncologist and care team. Our preliminary data suggest that Burley is not alone in this experience of illness progression—an element of physiology that can force patients of diverse values and expectations to rethink coping, engagement with family caregivers, and their relationship to the illness of cancer itself.

3.4. Arranging and Sorting to Examine Patterns

In addition to summarizing and representing ethnographic data and facilitating the interpretation of relationships, arrays can help bridge quantitative and qualitative analytical techniques by allowing researchers to combine statistical techniques for pattern recognition with interpretation of the underlying field notes and transcripts. It is important to recognize that although observations are tagged, they are not “reduced” to numbers or codes. That is to say, code patterns are meant to be orienting rather than reductive. ¹⁵ Depending on how they are sorted, ethnoarrays can also help facilitate either explanatory or confirmatory analysis. Examining cells within a column still facilitates interpretation of patient experiences or narratives, and sorting the ethnoarray can bolster interpretative insights through exploratory logics, for example, helping reveal or stimulate interpretations in the data that the analyst might have otherwise missed while reviewing or searching field notes and transcripts. That is, arrays can also be useful in examining whether a typology or pattern implied by a researcher or theory maps onto his or her data. Like the microarray on which it is based, the sorted ethnoarray would ideally allow analysts to identify patterns of similarity and difference in data and explore how these patterns resolve and translate into socially meaningful behaviors and theoretically meaningful categories and constructs.

Dating back to the popularization of exploratory data analysis (Tukey 1977), numerous quantitative techniques have been used to identify patterns in data that lack the strong sampling assumptions, claims to directionality, or assumptions about generalization that are typically associated with techniques like ordinary least squares regression. For instance, PCA, SA, latent structure analysis (LSA), multiple correspondence analysis (MCA), qualitative comparative analysis, and various applications of linguistic algorithms for mining large quantities of text data each provide useful ways of investigating patterns of colored cells that might be fruitfully integrated with the array approach. An in-depth discussion of the procedures involved in these techniques can be found elsewhere. However, it is worth noting that PCA, SA, LSA, and MCA are the most directly comparable with the simpler model of clustering we use in Figure 5. PCA is commonly used in biological microarrays to define groups on the basis of nominal, ordinal, or interval gene expression data. PCA requires only a shared and directional scale. LSA and MCA allow categorical data to be clustered without assuming directional scale. These techniques are more common in the social sciences. SA groups like sequences and trajectories of longitudinal data. Because any of these approaches can be applied to ethnoarray data, researchers must decide which approach to clustering (if any) is most useful for their projects, as well as whether the use of interval approaches provides more worthwhile insights than the categorical approaches. Depending on how domains are measured and organized within the ethnoarray, statistical patterns revealed via these techniques could address a range of research questions such as ascertainment of temporal sequence, explication of causal mechanisms, and discovery of new grounded-theoretical constructs.

In our mock-up, we used a simple scale and sorting procedure based on interpretive color assignment to show how inductive techniques for pattern recognition might be useful even in interpretive analyses. In Figure 4, patients are arranged arbitrarily. In Figure 5, the ethnoarray is based on patients’ structural characteristics (the bottom two domains, Social Support and Insurance and Finance) to examine whether and how those characteristics might shape pathways and tendencies related to seeking aggressive care. Each cell in these domains was assigned an ordinal value on the basis of color (red = 1; grey = .5; blue = 0). For each patient, an index value was calculated as the sum of the 12 cells in the two domains at times 1 and 2; the index had a potential range of 0 to 12; the actual range in the 10-patient array in Figure 4 was 1 to 11.5. The ethnoarray was then sorted according to these scores. Patients with higher index scores had their columns of data moved to the right side of the array; those with lower scores had their columns placed toward the left. Thus, reading Figure 5 from left to right roughly corresponds to examining experiences of patients with fewer to greater social structural resources. ¹⁶

In the case of the PtDelib project, prior research had suggested that more advantaged patients were more likely to be seen as “good study patients” whom clinician-investigators targeted for clinical trials recruitment (Joseph and Dohan 2009), and the ethnoarray provides an opportunity to examine this expectation in our preliminary PtDelib data. To examine the plausibility of this notion, we turned to the clustered array. The patients on the left side of Figure 5 tend to have less security in terms of Insurance and Finance and weaker Social Support. The distribution of the red cells in the Decision domain (at the top of the ethnoarray) may suggest that these patients are more aggressive in their pursuit of treatment and participation in clinical research. Given that we have arranged the domains in causal-temporal order, analysts and readers can then scan the array to try to identify patterns of color in the “intervening” domains—Health and Illness and Communication—that might suggest plausible social mechanisms. Analysts can then examine the underlying data (in this case interview transcripts and field notes) to see if these associations are likely real or spurious and to explain how the linking mechanisms operate in specific social contexts, a classic goal of ethnography. ¹⁷

3.5 Representing Data

Ethnoarrays have the advantage of being able to summarize large amounts of data in a compact yet flexible form, a key feature of many forms of sociological analysis that has been underdeveloped in many ethnographic approaches. Figures 3 to 6 each summarize data from hundreds of pages of interview transcripts and participant-observation field notes from multiple sites. Just the data from Wayne Burley include dozens of pages of text—too much evidence to include in a journal article or even a monograph. As in the microarray, each color-coded cell reflects a rich storehouse of meaningful information. The color assignment both summarizes the data as an interpretable visual representation and enables new analytics, such as using clusters to identify new patterns in data or verify whether the typologies ethnographers create map onto the underlying data they represent. These visual summaries are meant to supplement, but do not replace, the narratives that form the standard for ethnographic representation.

Arrays also provide new possibilities for sharing information. Ideally, the data underlying cells could be bundled and shared along with the array, and interested readers could explore the underlying data for any array cell. We refer to this type of array as a “data-linked” ethnoarray, in contrast to the arrays shown in Figures 3 to 5, which we characterize as “flat” (i.e., a noninteractive summary representation). We do not yet have the technology to produce and publish a data-linked array, though tabular and XML output functions of current CAQDAS platforms could facilitate this production, as we discuss and illustrate below. Inset 2 presents segments of underlying data from interviews with three participants that informed our coding of the High Social Capital measure in the ethnoarray. Figure 6 illustrates how these excerpts are situated within the ethnoarray. In a fully data-linked array, each cell of the ethnoarray would be associated with one or more fragments of ethnographic data, perhaps a quotation from an interview, a document, or an extract from a field note. Here, we use quotations from patient interviews to illustrate the kind of data that would underlie each cell in a data-linked array.

OPEN IN VIEWER

Data-linked arrays are dynamic and interactive and thus a poor fit for paper. However, modern computer interfaces are well suited to publishing and sharing such arrays, and we are working on developing the tools and techniques that will allow the construction and electronic publication of data-linked ethnoarrays. Using such an application, readers could explore particular cells or groups of cells within the ethnoarray by reading through the underlying data. Such an interface would also allow readers to sort or reorder the ethnoarray’s rows (domains) and columns (cases) using various procedures to highlight or discover patterns. Computer applications and tablet “apps” would ideally allow the reader to navigate a data-linked ethnoarray as one currently navigates online maps: clicking or tapping on cells to reveal the underlying data, zooming in and out of the ethnoarray to focus on particular patterns of data, dragging domains and cases to explore alternative patterns in the data. However, before implementing a data-linked array, important questions of how qualitative data might be adequately deidentified for sharing must be addressed, an issue we discuss in the next section. These discussions are consistent with calls for more transparent “open-source” social sciences (e.g., Freese 2007) and the shift toward digital models of publication that can facilitate new connections between scholarship and underlying data.

4.1. Research Infrastructure to Support Ethnographic Arrays

A robust ethnoarray research infrastructure would include (1) computer, Web, or tablet applications to facilitate the creation, distribution, and examination of arrays; (2) policies and procedures for anonymizing ethnographic data; and (3) servers to store and share data. Software to support ethnoarrays would differ from—though ideally integrate with and extend the capabilities of—presently available analysis programs. Current software helps experts manipulate data using technical procedures. CAQDAS platforms help analysts code, sort, and explore data as well as tag or memo excerpts that will ultimately be presented to readers. Statistical analysis software allows researchers to fit models and produce tables or graphs. These packages all focus on manipulating data and producing output, and in most instances the output, not the underlying data, is all that is distributed to audiences. Array software would include similar tools to manipulate and analyze data (e.g., clustering and search functions) as well as provide a new form of output in “flat arrays.” However, this software would also provide a mechanism for ethnographers to distribute findings. In short, ethnoarray software would help ethnographers not only produce output but also be part of their output and allow them to engage in communal research activities currently common to quantitative research such as archiving data and replicating analysis. ¹⁸

As a data management tool, ethnoarray software would help analysts link data to array cells and to arrange and rearrange the cells to explore alternate definitions of domains or configurations of cases. Links between data and cells occur when assigning cell color (we describe multiple strategies for color assignment in Inset 1). Ideally, applications would remain agnostic about the process of color assignment to allow analysts flexibility in how they link cells to data. This would also allow analysts the freedom to arrange the data set on their own terms, albeit in a way that aids in making their work more transparent. Some analysts might rely on interpretation alone, while others might develop an automated formal process for coding, sorting, and linking data to cells. No matter how the data-cell link is created, however, array applications should help analysts rearrange data to examine patterns or explore new relationships.

The ethnoarray also allows the representation of ethnographic data in two key ways. First, analysts can publish arrays online or in printed articles or monographs. Used this way ethnoarrays, like any other visual representation of data such as graphs or charts, allow researchers a way to summarize a large volume of information. For some researchers, such a summary might represent a key finding of an ethnographic study. Other ethnographers might use the arrays color-coded tabular representation to supplement interpretative analyses of field notes, interview transcripts, and other data that are presented using more traditional narrative approaches. The second way is to share an entire array, including cells and linked data, with readers. Readers then have the ability to examine the array, to iteratively explore and arrange the cells, and to examine the links between cells and data.

Such dissemination strategies differ substantially from the dominant ethnographic practice of publishing monographs and research articles with solely narrative evidence, but it is not unprecedented. The Human Relations Area Files (HRAF), a nonprofit international consortium, aimed to provide a resource of ethnographic data focused on comparative societal analysis, and full text from early ethnographies exists online. However, to provide comparable data across societies, HRAF used rigid coding and analytical constructs, and the archive has been interpreted as a historical document illustrating the challenges—and perhaps folly—of a cumulative approach to knowledge production in cultural analyses (Clifford and Marcus 1986; Marcus 1998). The ethnoarray model shares HRAF’s interest in making data available to a wide community of scholars, but scholars who use the ethnoarray need not format or categorize their data according to rigid preexisting conceptual schemata. They need not even agree about epistemic assumptions underlying ethnographic scholarship. They need only to specify what they do. In this sense only, data-linked ethnoarrays are more akin to publicly available quantitative data sets, such as the Integrated Public Use Microdata Series (IPUMS) or the General Social Survey, than the HRAF.

Researchers could produce array data sets as a part of their scholarly activities, but they need not adhere to a unified epistemic logic in doing so. They could then provide data sets to the sociological community with full documentation of how they were produced but without placing rigid boundaries on how the data are intellectually deployed. Similarly, data-linked ethnoarrays would not follow the rigid proscriptions of HRAF standardization but would instead exist as a series of independent data sets. Access could be provided via Web portals such as those seen at the Interuniversity Consortium for Political and Social Research (ICPSR). Researchers would ultimately have to decide if and how to reuse data sets and whether they were comparable with other data sets. ¹⁹

This raises the question of how to handle data governance. Sharing IPUMS- or ICPSR-housed data relies on policies for depositing, storing, and distributing data that ensure the safety and rights of research participants. Sharing arrayed ethnographic data would require producing new policies or extending current policies. Although data warehouses that host qualitative data are beginning to emerge, such policies are in their nascent stages, and the lack of a shared format for summary representation and sharing of ethnographic data remains a major limitation. Providing a comprehensive solution is beyond the scope of this article, but the development of ethnoarray approaches could potentially advance work in this arena. Policies for protecting microlevel quantitative social science data or protected health information may offer some further guidance for ethnographic policies.

Ethnographers’ own habits and practices regarding treatment of informants and other data, which have generally been passed along as craft rather than codified in policy, would need to be made more explicit. Sharing ethnographic data via arrays would also require a physical computing infrastructure, which could be provided via Internet-connected servers. Finally, even if the secure research infrastructure developed to accommodate arrays were never used to share data-linked ethnoarrays, it might prove to be a valuable resource for ethnographers to store, analyze, and reanalyze their own data, especially as the ethnoarray provides an analytical approach for linking data from multiple studies and points in time. In other words, although the array approach does not provide a universal solution to the challenges of sharing ethnographic data, it does provide a tool that can advance discussions about if and how this aim might eventually be reached. ²⁰

4.2. Using Arrays to Support Ethnographic Claims-Making

Evaluating ethnographic claims can be circuitous. Often, ethnographers collect and analyze data by themselves, and they can share only a fraction of their data with readers. Readers rely on self-reports of how field notes, interviews, and other data were collected; how these data were analyzed; and how insights were obtained and conclusions drawn. Ethnographers have long recognized that their authority derives in part from these reports of fieldwork and readers’ trust in those reports (Rabinow 1997; Whyte 1993). Marked by interpretation and iteration, ethnographic data often gain legitimacy when the insights they produce appear plausible and comprehensible—when, in essence, the data take on the appearance of speaking for themselves. Thus, the quality of their presentation—including richly evoked empirical context and well-developed theoretical framing—helps establish the legitimacy of the data that produce those results. ²¹

For many, a description of research procedures is a necessary first step, but an inadequate proxy, for a more direct examination of the links between data and claims. The limitations of this proxy become apparent when ethnographers debate whether the data support the claims made and even whether the data were collected. Such debates can flounder on irreconcilable divergences about the contextual or historical specificity of evidence and argument (Boelen 1992; Duneier 2002; Orlandella 1992; Sánchez-Jankowski 2002; Wacquant 2002). Sometimes a lack of standardization is associated with a lack of rigor. The combination of the requirement of trust without access to data to reconstruct analyses and the often charged nature of ethnography’s research topics can lead to scholarly exchanges that generate as much heat as light. Given that the ethnographer is the data collection instrument, it is not entirely surprising that controversies over the validity of ethnographic claims can devolve into attacks on analysts’ legitimacy or even morality (cf. Duneier 2002, 2006; Katz 2010; Wacquant 2002). Explicitly revealing how analysts link data and claims and encouraging readers to assess how the former sustains the latter could provide a more productive dialogue. We believe the ethnoarray represents a potentially useful tool to support ethnographic claims-making by facilitating such examinations. In Appendix B, we provide an example of how an array might be applied to examine the claims made in a well-known comparative ethnography.

Ethnoarrays can facilitate a more detailed examination of claims-making and, ideally, generate explicit discussion of how ethnographic data are invoked for causal or narrative purposes. This does not put the research community on an inexorable road toward an ethnographic equivalent of a p < .05 threshold for theoretical or substantive significance or even the conceptual standardization of HRAF, nor do ethnoarrays necessarily privilege causally or hypothesis-oriented research. Rather, we hope new tools can provide a way to examine how and why interpretations overlap or differ. Such discussions may provoke new ways of exploring fertile ethnographic questions.

4.3. The Scope of Ethnography

The ethnoarray may provide new capacities for analysis, but these capacities may come at the price of new burdens on those who choose to use the approach. A historical characteristic of ethnographic practice has been minimal barriers to entry; the lone ethnographer requires little more than time, a notebook, and access to enter the field and potentially contribute to the literature. In contrast, developing and contributing to ethnoarrays introduces new burdens for data collection and analysis. Consider the potential new burdens of an ethnoarray approach for ethnographers who conduct participant observation. When lone ethnographers collect notes in the field, they rely on their own judgment to decide what to observe and document. ²² Although many approaches encourage specifying a unit of analysis and identifying conceptual domains, this is not universally the case. Field notes may include everything from contextualized individual behaviors, to reflexive musings, to researcher descriptions of physical space (Emerson et al. 1995). In the midst of fieldwork, researchers decide what types of data to record and how to record it, but they rarely have the time, energy, or foresight to completely document how these decisions were made. Key background information in the form of schemata and headnotes may still remain unarticulated (Sánchez-Jankowski 2002). Even arrays thoughtfully constructed to include research questions, domains, and units of analysis may be incomplete when it comes to crucial details of how and why particular data were collected or recorded. Teams of ethnographic researchers may strive for more consistency in their procedures for conducting and documenting field sites, but the team’s shared understandings of the site and the project may not be formally documented. In short, ethnographers currently write field notes for themselves or for small audiences of fellow fieldworkers. They consider broader audiences when designing a study, when deciding what data to collect, or when writing up results. But the ethnographers themselves are the usual audience for most study data that remains largely private.

In contrast, data bound for arrays must continually consider a broader audience. The broader audience may be unknown, but generally it does not know the field site. Data included in an ethnoarray must be clear to a naïve audience lacking the presumed Verstehen of the ethnographer. They must have a defined unit of analysis. Formal field notes may thus require greater attention to detail and context, be longer, and take more time to write. They may contain greater redundancies than field notes that are destined for more traditional ethnographic uses, and ethnographers may feel self-conscious about array-bound notes. In short, ethnographers producing arrays may collect data differently than ethnographers producing traditional monographs or articles.

Using arrays also requires analytical transformations after the field data have been collected. Developing and distributing a flat array means using computer software, while using a data-linked array requires a series of steps to anonymize and secure data. These steps have the potential to make ethnography more expensive and less nimble, and it seems certain that anonymizing data for use in data-linked arrays will lead to the development of new research tasks, infrastructure, and personnel that could change aspects of ethnographic production for those using the array method. ²³