Introduction and Overview

You arrive at Lesson 6 having done real analytic work. In Lesson 5 you developed a codebook from the loneliness transcripts and applied it systematically. You have a coded dataset. You have some emerging themes. You have, for the first time in this course, a stack of analytic material that is yours — categories you defined, applied, and revised. And now, predictably, the question hits: what next? Bernard, Wutich, and Ryan call this the moment of the analytic fork. You have to choose an analytic strategy, and that choice will determine what your eventual findings look like.

Bernard, Wutich, and Ryan (2017, Ch. 7) open their chapter with a deceptively simple figure — a map of qualitative analytic strategies. The figure has “text” on the left and “findings” on the right, and between them are five paths. Each path is a different family of methods. Most students, in my experience, do not realize they are choosing one. They drift down whichever path their advisor happens to know, and they never see the others. The point of this section is to show you the whole map so that, when you choose a path in your capstone, you are choosing it because it fits your research question and not because it is the only one you know.

This section also makes a second move that is harder to teach but more important: it explains why the choice of analytic strategy is downstream of the research question, not upstream. Bernard, Wutich, and Ryan are explicit that the analytic strategy is dictated by what you want to find out (2017, p. 162). If you want to know what people say, content analysis is the right path. If you want to know what people mean, thematic analysis or grounded theory. If you want to know how people think, schema analysis. If you want to know how people tell, narrative analysis. If you want to know how people do things with words, discourse or conversation analysis. The error students most often make is choosing a method first and then deciding what their question was — usually because that method is what they have a software licence for, or what their committee chair publishes in. We will resist that error explicitly.

1.1 Why a Map Is Needed at All

Qualitative methods are not a single thing. They are a federation of distinct intellectual traditions, each with its own founding figures, philosophical commitments, vocabulary, and conventions about what good work looks like. Grounded theory emerged from medical sociology (Glaser and Strauss, 1967); content analysis from communication research (Berelson, 1952; Krippendorff, 2018); ethnographic discourse analysis from linguistic anthropology (Hymes, 1974); conversation analysis from a sociology that took ordinary talk to be the foundation of social order (Sacks, 1992); narrative analysis from a movement across psychology, sociology, and folklore studies (Riessman, 2008); schema and cognitive-anthropological methods from a small community working on how culture is mentally represented (D'Andrade, 1995; Quinn, 2005). These traditions developed in parallel and they only recently began to talk to each other.

The result is that the methods literature is a thicket. A student new to qualitative analysis reads three books and gets three incompatible vocabularies. The Bernard, Wutich, and Ryan stance — the one we adopt — is that the federation can be mapped: at the highest level, qualitative analysis is about going from text to some kind of structured representation, and the kind of representation you build determines the family of methods you are in. The map is not a hierarchy. It is not a sequence. It is a fork.

1.2 The Five Paths

Bernard, Wutich, and Ryan organize the field into five analytic paths. Each is named by what the path turns text into. The names are simple. The traditions behind them are deep.

It is worth saying explicitly: many real studies use more than one path. A grounded-theory study (text-to-themes) might include a content-analytic count of how often each theme appears across cases (text-to-counts). A narrative analysis (text-to-narratives) might include attention to discourse markers in the telling (text-to-talk). The paths are not mutually exclusive; the point of distinguishing them is so that you know which is your dominant path and which is supporting work.

1.3 Path 1 — Text to Counts

The text-to-counts path treats the text as data that can be reduced to frequencies and analyzed quantitatively. The classical method is content analysis, which Krippendorff (2018) defines as “a research technique for making replicable and valid inferences from texts (or other meaningful matter) to the contexts of their use” (p. 24). The procedure is: develop a coding scheme; apply it; count; analyze the counts. Module 8 of this course is dedicated to content analysis.

The strengths of the text-to-counts path are exactly what you would predict from a quantitative methods background. It is transparent, replicable, and produces findings that are directly comparable across groups. The weaknesses are equally predictable. Counts cannot distinguish a thematic mention that is central to a participant's account from one that is incidental. Counts of words or codes do not capture the meaning the words carry. A participant who says “I am lonely” once at the heart of their interview is not equivalent to one who uses the word fifteen times as a discourse filler.

In a loneliness study, a text-to-counts approach might code every transcript for the presence/absence of (a) a structural attribution (“society is set up to make new mothers lonely,” in Sarah's words at P03), (b) a phenomenological description of an embodied state (Sarah: “heavy. Different. I don't recognise it”), and (c) a coping strategy. The analyst could then count the prevalence of each across the 20 transcripts and compare across age and gender. The findings would be of the form: “Structural attributions appeared in 14/20 transcripts and were significantly more common among women caregivers (8/8) than among men in late-life relationship rupture (2/4).”

1.4 Path 2 — Text to Themes

The text-to-themes path treats the text as a field in which patterns of meaning are to be identified and elaborated. The dominant methods here are thematic analysis (Braun & Clarke, 2006, 2019, 2022) and grounded theory (Glaser & Strauss, 1967; Charmaz, 2014; Corbin & Strauss, 2015). They differ in important ways — Module 7 will spend serious time on grounded theory specifically — but they share an analytic commitment: the unit of analysis is the theme, and the work is the identification, elaboration, and patterning of themes across the corpus.

Text-to-themes is by far the dominant family of methods in contemporary qualitative health research, for both good and bad reasons. Good: it is intuitive, it produces findings non-specialists can read, and it bridges easily into intervention design. Bad: “thematic analysis” has become a default residual category in journal articles, often invoked when no specific method was actually used, and the rigour of the work varies enormously.

For a loneliness study, a text-to-themes approach might develop themes like (a) loneliness inside companionship — the experience of feeling lonely while in the constant company of another person (Sarah at P03, Diana at P07, Jacob at P17); (b) identity loneliness — loneliness produced by a change in how one understands oneself (Kenji at P14 after coming out at 60, Rose at P19 in her identity work); (c) structural loneliness — the participant's interpretation of their loneliness as caused by social arrangements rather than personal failure (Sarah at P03, Diana at P07, Amira at P15). Lesson 5 was your introduction to this work; Lessons 7 and 9 deepen it.

1.5 Path 3 — Text to Schemas

The text-to-schemas path treats the text as evidence about the underlying mental structures people use to organize their understanding of a domain. A schema, in cognitive anthropology, is an abstract knowledge structure that lets a person recognize a situation, fill in defaults, and act appropriately. Schemas are not directly observable. They are inferred from speech, behaviour, and consistency of response across stimuli.

Methods in this family include cultural-domain analysis (Borgatti, 1994), free-listing followed by pile-sorts, cultural-consensus analysis (Romney, Weller, & Batchelder, 1986), and mental-model interviews (Kempton, Boster, & Hartley, 1996). These are less common in mainstream public-health qualitative work but are powerful when the question is about cultural cognition — for example, what laypeople understand climate change to be, what categories of medical treatment they recognize, what kinds of relationships they distinguish. Module 9 introduces schema analysis; Module 12 returns to it with computational tools.

For loneliness, a text-to-schemas analysis might ask: what kinds of loneliness do participants distinguish? Sarah at P03 distinguishes (a) the loneliness of being awake while everyone sleeps, (b) the loneliness of being in a marriage and unable to say the true thing, and (c) the loneliness of mourning a previous self. These are not three instances of one schema; they are three distinct schemas that Sarah is articulating. Kenji at P14 distinguishes the loneliness of romantic absence (which he says he does not have, because of his partner David) from the loneliness of his sons' refusal (which he does). The schema work is identifying these distinctions, formalizing them, and asking whether they are shared.

1.6 Path 4 — Text to Narratives

The text-to-narratives path treats each transcript as a story that the participant is telling, and asks what kind of story it is. The foundational text is Riessman's Narrative Methods for the Human Sciences (2008). Riessman distinguishes thematic narrative analysis (what is the story about?), structural narrative analysis (how is the story built — orientation, complicating action, resolution?), dialogic/performance analysis (what is the story doing for the teller?), and visual narrative analysis (how do images figure in the telling?).

Narrative analysis is the right path when the unit of meaning is the story-as-told, rather than themes that can be abstracted from it. A loneliness study that asked “how do people construct a narrative of becoming lonely?” would be a narrative project. Look at Kenji at P14 again: his entire interview is structured as a redemption-and-cost narrative — an old self that was inauthentic, a turning point (coming out at 60), a present that contains specific lonelinesses (the sons), and an emerging future the protagonist is working toward (“to be a man my sons might want to know someday”). That narrative arc is itself the analytic unit. Module 9 covers narrative analysis.

1.7 Path 5 — Text to Talk

The text-to-talk path treats the transcript not as content to be interpreted but as a record of action being performed through language. The classical method is conversation analysis (CA), which emerged from the work of Harvey Sacks at UCLA in the 1960s (Sacks, 1992; Sidnell & Stivers, 2013). CA asks: at this turn in the conversation, what is the speaker doing? Repairing? Agreeing? Resisting? Hedging? Asking for affiliation? CA pays close attention to features of talk that other methods discard — pauses, overlaps, false starts, repair sequences — because these are the materials through which conversational action is built.

Related methods include discourse analysis (Potter & Wetherell, 1987; Edwards & Potter, 1992) and critical discourse analysis (Fairclough, 2010; Wodak & Meyer, 2016). These are less micro than CA and attend more to ideological and political work being done through language. Module 10 covers all three.

In a loneliness study, a text-to-talk analysis might look at Sarah's repeated rhetorical management of her own claim that her loneliness is structural: she says it, and then immediately performs a kind of evaluative work — “Which is — I don't know if that helps me feel better or worse. Worse maybe. Because I can't fix structural.” That sequence is not a content claim; it is a piece of self-management in talk, and a discourse-analytic lens would treat the sequence itself as the unit of analysis.

1.8 The Choice Is Downstream of the Question

You will, in your capstone, eventually choose a dominant analytic path. Bernard, Wutich, and Ryan are emphatic that the choice should be made after you have specified your research question, and not before. The mapping is approximately:

For most students in this course, the dominant path will be text-to-themes, with secondary content-analytic counts. That is the most common path in applied health qualitative research. But it should be your choice because it fits your question, not because it is the default. The capstone milestone in this lesson is a first conceptual model, and which kind of model you build will be heavily shaped by which analytic path you are committing to.

1.9 From Codes to Interpretation — the Transition Most Students Skip

Before we leave Section 1, it is worth naming a specific failure that Bernard, Wutich, and Ryan return to repeatedly: the failure to make the transition from codes to interpretation. Many qualitative studies stop at a list of codes. They report what was coded, how often, by whom, with what inter-rater reliability. They are, in Bernard, Wutich, and Ryan's vocabulary, descriptive but not analytic. The promised explanation — why the patterns are there — is missing.

The transition from codes to interpretation is the moment where the analyst stops being a librarian and starts being a theorist. Codes are categories. Interpretation is the claim about what those categories tell us. A coded dataset with the codes structural attribution, embodied exhaustion, witness absence, and maternal identity grief is a stack of labelled material. The interpretation is the claim: postpartum maternal loneliness, in this dataset, is best understood as the intersection of an embodied state of depletion, a witness deficit (no adult present during the hardest hours), and a grief for a pre-maternal self that the structural arrangement of the nuclear family is no longer adequate to support. That is interpretation. It is what your methods section will eventually justify, and what your discussion will defend.

Sections 2 and 3 of this lesson give you the tools that make this transition possible: memos (Section 2) and conceptual models (Section 3). They are the mechanisms by which codes turn into interpretation. Without them, the codes are just a filing system.

Introduction and Overview

Section 1 ended on the diagnosis. The transition most students skip is the move from codes (a labelled corpus) to interpretation (a claim about what the labels tell us). Section 2 provides the mechanism. There are two of them, and they work together. The first is memo writing: the systematic generation of analytic prose alongside the coded data. The second is triangulation: the deliberate checking of an emerging interpretation against multiple sources, analysts, theories, or methods. Memos are the engine of interpretation. Triangulation is the brake that keeps interpretation from running over what the data actually support.

Bernard, Wutich, and Ryan treat memos as a methodological invention of grounded theory (Glaser & Strauss, 1967; Strauss, 1987; Charmaz, 2014) that has migrated into every other qualitative tradition. Charmaz, in particular, treats memo-writing as the “pivotal” intermediate step between data collection and theory development — the place where the analyst's thinking happens visibly, on paper, and becomes inspectable rather than hidden inside the analyst's head (Charmaz, 2014, ch. 7). The methodological function of memos is exactly this: to externalize analytic thinking so that it can be revised, critiqued, and eventually published.

2.1 The Memo as the Cognitive Workshop

A memo is a piece of analytic writing addressed to yourself. It is dated, named, and stored in a way that lets you retrieve it later (most coding software has a memo function; if you are working in plain text, a folder of .md files named by date and code works). The memo records your in-process thinking about a code, a comparison, a theoretical possibility, a methodological problem, or an emerging interpretation.

What makes memo-writing analytic, rather than diaristic, is that the memo is anchored in specific data. A memo about “structural attributions in loneliness narratives” cites which transcripts use this framing, what the participants actually say, and what the analyst is making of it. A memo without specific data citations is not yet a memo; it is a daydream. Charmaz (2014) is explicit on this point: “memos catch our thoughts, capture the comparisons and connections we make, and crystallize questions and directions for us to pursue” — but they do so by staying close to the empirical material.

Bernard, Wutich, and Ryan (2017, pp. 167–172) distinguish three kinds of memo, each serving a different function in the analytic workflow. Treating them as distinct kinds is useful because the writing voice is different in each, and many students confuse one with another.

2.2 Code Memos

A code memo is the running definition of a code as it develops. When you create a new code — let's say witness absence for the family of references to having no adult presence in moments of acute loneliness — you should immediately write a code memo. The code memo records: (a) the working definition of the code, (b) what counts as an instance and what does not, (c) the first instance you tagged with it, (d) any boundary cases that pushed you to refine the definition, (e) revisions to the definition over time and the date of each revision.

A code memo serves the audit-trail function Bernard, Wutich, and Ryan treat as essential to transparency. Without it, your codebook entry says “witness absence: when participants describe being alone in a hard moment.” With it, you can reconstruct that the code originally meant “literal solitary moments” but was expanded after Sarah at P03 said “and Adam's asleep, which he has to be because he has work,” and the code now covers functional witness absence (someone is present but not available as an adult witness). That kind of refinement is what a code memo records.

Notice what the code memo does: it makes the code's history inspectable. A reviewer of the eventual paper can see how this category formed. That is the working content of transparency in qualitative analysis.

2.3 Theoretical Memos

A theoretical memo is a different beast. It is not about a single code; it is about a relationship, a possibility, an explanatory claim. Theoretical memos are where interpretation happens. They are the place where you write sentences of the form “the data suggest that...” or “what unites these three cases is...” or “this pattern cannot be explained without invoking...”

Strauss (1987) describes theoretical memos as the form of writing in which “the analyst thinks on paper.” The voice is exploratory but the content is empirical: every theoretical claim cites the cases that support it and (importantly) the cases that do not. A theoretical memo that names only the supporting cases is a hypothesis-confirmation exercise, not analysis. The discipline is to look for the disconfirming case and either reconcile it or revise the claim.

This is what an analytic memo looks like. It names cases, cites them, acknowledges what does not fit, and ends with a next step. It is the kind of writing that, over the course of a project, accumulates into the body of an eventual paper's findings section.

2.4 Operational Memos

An operational memo is about the analytic process itself, rather than the substantive content. It records methodological decisions, problems, and revisions. “I am going to stop double-coding every transcript and shift to double-coding every fifth, because we have reached saturation on the codebook.” “I am going to reconcile the disagreement between coder A and coder B by doing a third pass with both present.” “I noticed that I have been giving more analytic weight to the longer transcripts because there is more material; I need to check this and possibly correct.”

Operational memos are the most under-used of the three kinds. They feel like meta-commentary, not analysis. But they are what allows the methods section of the eventual paper to be precise rather than retrospective fiction. A study that reports “we revised the codebook three times during analysis” needs operational memos to explain when and why and what was re-coded. Without them, the claim is hand-waving.

2.5 When a Memo Becomes a Finding

One of the most underappreciated features of disciplined qualitative analysis is that, by the end of a project, you have effectively drafted your findings section in the form of theoretical memos. The work of writing the paper is partly the work of assembling memos into argument. A theoretical memo that has been revised, cross-checked against new transcripts, and survived the discovery of disconfirming cases is no longer a tentative musing — it is a finding. The paper renders it as one.

Charmaz (2014) uses the metaphor of memos as “the building blocks” of grounded-theory papers. Strauss (1987) is even more direct: he describes a polished theoretical memo as “ready to be inserted with little revision into a publication.” The discipline is to write memos that are good enough that, at the end, the paper writes itself out of the memo corpus. Most papers that feel laboured to write feel that way because the memo discipline was skipped and the findings have to be invented from scratch at the end.

2.6 Triangulation — Denzin's Typology

Memos are the engine of interpretation. Triangulation is the brake. The term comes from surveying: to locate a point precisely, take bearings from three positions. Norman Denzin (1978, expanded in 1989 and 2009) imported the metaphor into qualitative research and offered the most-cited typology in the methods literature. Denzin distinguishes four kinds of triangulation, each addressing a different threat to interpretive validity.

Data triangulation

Data triangulation is checking the emerging interpretation against multiple sources of data on the same phenomenon. In a loneliness study, this might mean: interviews and participant-generated diaries; interviews from participants and from family members of those participants; interviews and direct observation in adult day-care or moms-group settings. The logic is that an interpretation supported across multiple kinds of source is more robust than one supported by a single source.

For your capstone, full data triangulation may not be possible — you are working with interview transcripts alone. But a kind of internal data triangulation is available: an interpretation that holds across transcripts from very different life-stages (Maya at 22, Margaret at 73, Kenji at 60) is more robust than one that holds only in a single subgroup. The capstone milestone in Section 4 asks you to think explicitly about which transcripts don't fit the model you build, which is a form of data triangulation.

Investigator triangulation

Investigator triangulation is having multiple analysts code or interpret the same data. This is the form most familiar to public-health audiences, because it overlaps with inter-rater reliability statistics. But Denzin meant something broader: not just inter-coder agreement on a codebook, but multiple analytic eyes on the same material, ideally bringing different perspectives.

In real qualitative projects, investigator triangulation looks like: two analysts code the same five transcripts independently and meet to discuss; the disagreements are themselves productive analytic material, often surfacing assumptions in the codebook that neither analyst could see alone. A study with intercoder reliability of 0.85 is not necessarily a study with good investigator triangulation — the 15% of cases of disagreement may be exactly where the analytic action is.

Theory triangulation

Theory triangulation is interpreting the data through multiple theoretical frames and seeing whether the conclusions converge. A loneliness study could interpret its dataset through (a) social-isolation theory (Cacioppo & Hawkley, 2010); (b) attachment theory (Bowlby, 1980; Ainsworth, 1989); (c) existentialist phenomenology (Heidegger, 1962; Yalom, 1980); (d) feminist theories of relational care (Tronto, 1993; Held, 2006). Each frame highlights different aspects of the same material. If your central interpretive claim survives translation across theoretical frames, it has earned theoretical robustness.

This is less common in applied health qualitative research than in interpretive sociology, partly because most applied studies commit to one frame upfront. But the discipline of asking “what would this look like read through a different theory?” is one of the best tests of whether your interpretation is bound to your starting assumptions.

Methodological triangulation

Methodological triangulation is checking the qualitative findings against findings from a different methodology — usually a quantitative one. In a loneliness study, this might mean: your interview-based finding that postpartum maternal loneliness has a strong structural dimension is methodologically triangulated against the survey literature showing that countries with longer paid parental leave have lower postpartum-depression prevalence (Heshmati et al., 2023; Smith et al., 2021); your interview-based finding about adult day-care as caregiver oxygen is triangulated against quantitative caregiver-burden literature.

This is the form of triangulation that mixed-methods designs explicitly engineer. For your capstone, full methodological triangulation is out of scope, but a strong discussion section will gesture at it: my qualitative finding is consistent with / inconsistent with / extends the quantitative literature on...

2.7 Critiques of Triangulation — Richardson's Crystallization

Triangulation, as Denzin formulated it, sits inside a postpositivist assumption: there is a phenomenon out there, and multiple readings of it should converge on something real. Laurel Richardson (2000; Richardson & St. Pierre, 2005) argued that this metaphor is wrong for much qualitative work. Triangles are flat and they have only three sides. Real social phenomena are multifaceted, refracted differently from every angle, and the goal of an interpretive analysis is not to nail down one point but to show the phenomenon's multiple shapes.

Richardson proposed crystallization as an alternative metaphor. A crystal — a prism — refracts light in different directions depending on the angle. Different methods, different theories, different analysts, different writing forms (article, poem, ethnographic vignette, performance) reveal different facets. The goal is not convergence on one finding; it is a richer, multidimensional sense of the phenomenon. The valid analysis is the one that shows the phenomenon's complexity rather than reducing it.

The methodological stakes of the disagreement are real. Triangulation, especially methodological and investigator triangulation, fits public-health audiences trained in convergent validity. Crystallization fits interpretive and critical traditions where the point of qualitative work is to expand what is sayable about a phenomenon, not to test whether different methods agree. Most contemporary qualitative health research lives somewhere in between — it uses triangulation language because reviewers expect it, but the substantive analytic practice has crystallization features.

2.8 Worked Example: Triangulating a Loneliness Claim

Let's take the interpretive claim being developed in the theoretical memo above — loneliness inside companionship is a distinct kind — and walk it through triangulation.

Data triangulation: Does the claim hold across very different participants? Sarah (postpartum, 34), Diana (caregiver, 52), Jacob (mid-life partnership, 49) all describe it. Different life-stages, different demographic profiles, different proximal causes. Sarah's husband is asleep; Diana's mother has dementia; Jacob's partner is emotionally absent. The shared structural feature — companioned but witnessless — survives translation across these differences. That is meaningful internal data triangulation.

Investigator triangulation: Have other analysts looking at the same transcripts converged on a similar reading? If your co-coder, blind to your theoretical memo, also coded these passages with a kindred concept (perhaps under a different name — “present-but-absent companionship” or “witnessless intimacy”), that is investigator triangulation in operation. If your co-coder read the same passages differently — say, as straightforward complaints about relational quality rather than as a distinct phenomenological kind — that disagreement is itself analytic material to be reconciled.

Theory triangulation: Does the claim translate across theoretical frames? Attachment theory would read “loneliness inside companionship” as activation of the attachment system in the presence of a non-responsive attachment figure — consistent. Existentialist phenomenology would read it as the inescapable solitude of the self made visible by the felt absence of recognition — consistent. Feminist relational ethics would read it as the gendered burden of caring without being cared for — consistent and adding the dimension that the three cases (postpartum, eldercare, partnership) all involve women's care work. Three theoretical lenses converge on the same kind of loneliness but illuminate different aspects. The claim survives theoretical translation.

Methodological triangulation: Is the phenomenon visible in quantitative loneliness research? Hawkley and Cacioppo (2010) and the postpartum-isolation literature (Razurel et al., 2017) both report that marital status alone is a poor predictor of loneliness in some populations — consistent with the qualitative finding that the presence of a partner does not preclude loneliness. The methodological convergence is partial but real.

This is what triangulation looks like in operation. Notice that it does not produce a yes/no verdict on the claim. It produces a richer specification of where the claim holds, how it is read under different lenses, and what kinds of evidence support it. That is the analytic work the memo was leading toward.

Introduction and Overview

Bernard, Wutich, and Ryan (2017, Ch. 8) open the chapter on conceptual models with a claim that is easy to underestimate: the purpose of qualitative analysis is, in the end, to produce a model of the phenomenon. By “model” they do not mean a regression equation. They mean a structured visual or verbal representation of how the parts of the phenomenon hang together — what the kinds are, how they relate, in what order they unfold, under what conditions one leads to another. A taxonomy, a typology, a flowchart, a concept map, a decision tree are all kinds of conceptual model. Each renders an interpretation visible. Each is a finding.

The chapter is also explicit about why visualization matters. The cognitive limits on holding twenty interconnected qualitative codes in your head at once are real. Most students try, fail, and end up writing flat findings sections that report codes one after another without showing how the codes relate. A conceptual model is the analytic move that turns a list of codes into a structure. Miles, Huberman, and Saldaña (2020) make this point throughout their classic Qualitative Data Analysis: matrix and network displays are not decoration; they are the medium through which qualitative interpretation gets done. Module 7 of this course will return to matrix displays specifically; Section 3 here introduces the building blocks.

3.1 Taxonomies — Hierarchical Category Systems

A taxonomy is a hierarchical organization of categories: a top-level category divided into subordinate categories, which may themselves be further divided. The biological taxonomy of life (Domain → Kingdom → Phylum → ... → Species) is the classical example. In qualitative analysis, taxonomies usually have only two or three levels — the goal is interpretive structure, not exhaustive Linnaean classification.

Taxonomies answer the research question: what kinds of X are there? A loneliness study that asks “what kinds of loneliness do these participants experience?” is asking for a taxonomy. The analytic work is to build a category system in which every coded instance can be placed somewhere, and in which the categories are mutually exclusive and collectively exhaustive (the two classical desiderata, though most real qualitative taxonomies relax them).

Worked taxonomy: kinds of loneliness in the dataset

Drawing on the coded transcripts, a defensible first-pass taxonomy of loneliness might look like this:

Notice three things about this taxonomy. First, it has structure: four top-level kinds, with nine subordinate kinds. Second, it is grounded: every category has named exemplars. Third, it is provisional: this is a first pass, and a careful re-reading of, say, Priya (P09) might force a fifth top-level category or push Marcus (P08) from existential-developmental into a situational subcategory. The taxonomy is a hypothesis about the data's structure, not a finished claim.

Standards for a defensible taxonomy

What makes a taxonomy defensible rather than arbitrary? Bernard, Wutich, and Ryan (2017, p. 187) and Spradley (1979) before them identify several standards:

1. Grounded in data. Every category corresponds to coded instances; no category is included because it “ought” to be there theoretically without evidence.
2. Inclusion criteria are explicit. A reader can determine, from your description, whether a given instance belongs in category X or category Y.
3. The categories are at the same level of abstraction within a tier. “Existential” and “loss loneliness” should not be siblings; they are at different levels.
4. The taxonomy makes empirical claims that are testable against the data. “Every transcript can be assigned to one of these categories.” If not, the taxonomy needs revision (a residual category, a new top-level kind, or a recognition that the taxonomy is non-exhaustive and why).
5. The taxonomy is revisable. Disconfirming evidence prompts revision, not denial.

3.2 Typologies — Cross-Classification by Dimensions

A typology is a different beast. Where a taxonomy is hierarchical (sub-kinds of a kind), a typology is dimensional (cases located in a space defined by two or more variables). The classical sociological typology is Max Weber's (1922) ideal-types: cases located at intersections of dimensions like rational/traditional, instrumental/value-oriented. In qualitative analysis, typologies are powerful when you have two dimensions that crosscut each other in a way no single hierarchical sort can capture.

Typologies answer the research question: how do cases vary along multiple dimensions, and what kinds emerge from their intersection?

Worked typology: trigger kind × response kind

Consider two dimensions that emerged in the loneliness data: the kind of trigger that brings loneliness on (existential, relational, structural, situational), and the kind of response the participant has developed (reframing, behavioural change, withdrawal, no consistent strategy). Crossing these dimensions produces a 4×4 typology with 16 cells. Not every cell will be populated — that itself is a finding — but the populated cells map an interpretively interesting structure.

Notice what the typology lets you see that the taxonomy does not. The taxonomy told you what kinds of loneliness exist. The typology tells you about the relationship between trigger and response. Reading the table by row reveals that existential triggers produce predominantly reframing responses. Reading by column reveals that “no consistent strategy” clusters with relational and structural triggers — a finding that has intervention implications (these are the participants who would benefit most from external support). The cell-by-cell reading is what makes the typology analytically valuable.

An important note on typologies: the dimensions you choose are themselves an interpretation. Choosing “existential / relational / structural / situational” as the trigger dimension reflects a particular reading of the data; another analyst might use “internal / external” or “acute / chronic.” The typology renders this interpretive choice visible. Your methods section needs to justify the dimensions you used, the same way it justifies your codebook.

3.3 Concept Maps — Relational Models

A concept map (Novak & Cañas, 2008) is a visual representation of concepts and the relationships among them. Concepts are nodes; relationships are labelled arrows between nodes. Where taxonomies are hierarchical and typologies are dimensional, concept maps are network-like: they can show many-to-many relationships, feedback loops, and indirect connections.

Concept maps are the right model when the research question is about how things relate rather than what kinds there are. A concept map of loneliness might show structural conditions (poverty, isolation, immigration status) as nodes feeding into triggers (events, losses, transitions) which feed into experienced states (witness absence, embodied exhaustion, identity disruption) which feed into responses (reaching out, withdrawal, reframing) which feed back into structural conditions in a recursive loop.

A concept map need not have hundreds of nodes — in fact most useful maps have between five and fifteen. The discipline is choosing which nodes to elevate. Every node in a concept map should be a code or category for which you have a defensible empirical grounding. Arrows should be labelled with the kind of relationship (causes, enables, blocks, transforms, is-a, part-of). Unlabelled arrows are the most common failure of concept maps; they communicate nothing.

3.4 Process Models — Temporal/Sequential Maps

A process model (sometimes called a flowchart or pathway model) renders a phenomenon as a temporal sequence: it starts here, moves through these stages, branches at this decision point, ends in one of these outcomes. Process models are common in implementation science, clinical-pathway documentation, and theory development in grounded theory (Strauss & Corbin's “conditional matrix” is a process model). Case-study research traditions (Yin, 2018; Eisenhardt, 1989) similarly emphasize the construction of process and within-case models from empirical material.

For loneliness, a process model might trace how a single episode unfolds: trigger → onset (embodied recognition) → interpretation → response selection → response enactment → outcome (relieved / unchanged / worsened). Each stage is a node; transitions are arrows. The model invites the analyst to ask, at each stage, what shapes the transition. Why do some participants move from onset to a reaching-out response, while others move from onset to withdrawal?

Worked process model: the loneliness-coping cycle

A defensible process model for the loneliness-coping cycle, grounded in the dataset, has six stages:

1. Triggering event or condition. A specific event (Maya's breakup, Sarah's 4 a.m. nursing, Diana's birthday party, Kenji's son returning a gift) or a chronic condition (Margaret's long-term care, Frank's demographic isolation).
2. Embodied onset. The participant notices loneliness in the body before they name it. Sarah's exhaustion fusion (“the loneliness and the exhaustion are kind of fused”); Diana's back pain plus drinking; Maya's chest tightness. Most participants describe this stage even when they do not call it “onset.”
3. Naming and interpretation. The participant labels the experience as loneliness (or refuses to — Aarav, P06, initially resists the term) and attributes it (to personal failure, to specific others, to structural conditions, to fate, to a life-stage). This stage matters: the interpretation determines the response.
4. Response selection. The participant chooses (often implicitly) a response category: reach out, withdraw, distract, reframe, seek help, suppress. The chosen response is shaped by past experiences with each option.
5. Response enactment. The participant does the thing. Sarah goes to the moms group; Diana posts in the Facebook caregivers group; Maya cries in the bath; Kenji takes a tea-ceremony lesson.
6. Outcome and learning. The response produces some outcome (relieved, partial relief, no change, made worse), which then feeds back into stage 3 (interpretation of what loneliness is for this participant) and stage 4 (which response they are likely to select next time).

The feedback loop in stage 6 is what makes this a process model rather than a flowchart. Loneliness, in this rendering, is not a linear sequence with a stopping point; it is a cycle that the participant lives in over years, with each cycle modifying the next. That feedback structure is itself a finding — it is what makes the analytic claim “loneliness has a process structure” rather than just “loneliness is a state.”

3.5 Decision Trees — Preview of Module 11

A decision tree is a special kind of process model in which the branches represent choices and the leaves represent outcomes. Decision trees are particularly powerful for modelling how a person navigates a complex situation: the structure is “if X then Y else Z,” recursively. Bernard, Wutich, and Ryan dedicate substantial space in Chapter 11 (which Module 11 of this course covers) to decision-tree modelling. For now, recognize that decision trees are available as a kind of conceptual model and that they are particularly useful when your research question is about how participants make a specific kind of decision (e.g., whether to seek help for loneliness, whether to disclose loneliness to a friend).

3.6 The Role of Visualization — Why Diagrams Matter

Why bother to draw a conceptual model when you could write a paragraph? Three reasons.

First, visualization forces specificity. When you write “loneliness is shaped by structural conditions, interpersonal relations, and individual response,” the sentence is glib. When you draw a diagram with nodes labelled “structural conditions,” “interpersonal relations,” and “individual response” you have to draw arrows, and the arrows force you to specify what the relationships are. Does structural condition cause individual response or only mediate it? Is there a feedback loop? You can fudge in prose. You cannot fudge in a diagram.

Second, visualization aids comparison. Once you have a diagram of your interpretation, a co-analyst can see exactly what you are claiming. They can disagree with a specific arrow or node, rather than gesturing vaguely at your prose. This is investigator triangulation in operation. Miles, Huberman, and Saldaña (2020) make this case throughout: diagrams are the medium of qualitative discussion.

Third, visualization aids communication. A reader of your eventual paper will hold the diagram in mind as they read the surrounding prose. The diagram becomes a navigation aid for your interpretation. This is why every grounded-theory paper of consequence ends with a diagram, and why papers without diagrams often feel diffuse.

3.7 Tools — What to Use for What

The choice of tool for building conceptual models matters less than the discipline of building them. But informed choice of tool helps. Here is the working set.

For your capstone, the recommended workflow is: hand-draw first (fast, exploratory); transfer to draw.io or DiagrammeR once the model has stabilized (publication-quality); revise in the same tool as the model evolves through co-coder discussion. Some students find that drawing-by-hand on a tablet (e.g., GoodNotes, Notability, Concepts) gives them the speed of paper with the shareability of digital files.

3.8 R Tools — ggplot Heatmap and DiagrammeR Process Diagram

Two short R demonstrations follow. The first shows how to visualize a code×participant matrix as a heatmap — useful for seeing at a glance which codes apply to which participants, and where the gaps are. The second shows how to render a process diagram in DiagrammeR.

library(tidyverse)

# Long-format coded data: one row per (participant, code) coded instance
coded <- tribble(
  ~participant, ~code,
  "P01_Maya",     "situational",
  "P01_Maya",     "acute-event",
  "P03_Sarah",    "witness-absence",
  "P03_Sarah",    "structural-attribution",
  "P03_Sarah",    "companioned-loneliness",
  "P03_Sarah",    "maternal-identity-grief",
  "P07_Diana",    "witness-absence",
  "P07_Diana",    "companioned-loneliness",
  "P07_Diana",    "structural-attribution",
  "P10_Daniel",   "existential",
  "P13_Margaret", "structural-isolation",
  "P14_Kenji",    "identity-disruption",
  "P14_Kenji",    "family-rupture",
  "P15_Amira",    "migration-cultural",
  "P15_Amira",    "structural-attribution",
  "P17_Jacob",    "witness-absence",
  "P17_Jacob",    "companioned-loneliness"
)

# Build the wide presence/absence matrix
mat <- coded |>
  mutate(present = 1) |>
  pivot_wider(names_from = code, values_from = present, values_fill = 0) |>
  pivot_longer(-participant, names_to = "code", values_to = "present")

# Plot
ggplot(mat, aes(x = code, y = participant, fill = factor(present))) +
  geom_tile(color = "white", linewidth = 0.5) +
  scale_fill_manual(values = c("0" = "#F3F4F6", "1" = "#0B7B6B"),
                    labels = c("absent", "present"),
                    name = "Code") +
  theme_minimal(base_size = 11) +
  theme(axis.text.x = element_text(angle = 45, hjust = 1)) +
  labs(title = "Codes by participant — loneliness capstone",
       x = NULL, y = NULL)

library(DiagrammeR)

grViz("
digraph loneliness_cycle {
  graph [layout = dot, rankdir = LR, fontname = 'Open Sans']
  node  [shape = box, style = 'rounded,filled', fillcolor = '#E6F3F0',
         color = '#0B7B6B', fontname = 'Open Sans']
  edge  [color = '#065C50', fontname = 'Open Sans', fontsize = 10]

  trigger    [label = 'Trigger\\n(event or condition)']
  onset      [label = 'Embodied onset\\n(noticed in body)']
  interp     [label = 'Naming & interpretation\\n(attribution)']
  select     [label = 'Response selection']
  enact      [label = 'Response enactment']
  outcome    [label = 'Outcome & learning']

  trigger -> onset
  onset   -> interp
  interp  -> select
  select  -> enact
  enact   -> outcome
  outcome -> interp [label = 'feedback', style = dashed, color = '#CC0033']
  outcome -> select [label = 'feedback', style = dashed, color = '#CC0033']
}
")

3.9 What Makes a Conceptual Model Defensible

Bernard, Wutich, and Ryan (2017, p. 196) close Chapter 8 with three criteria for a defensible conceptual model. Reproduce them with care; they will be your benchmark in your capstone.

1. Grounded in data. Every node, every category, every arrow has a defensible empirical basis. You can name the coded instances that justify it. A node with no empirical anchor is a hypothesis, not a finding, and should be marked as such if included.
2. Transparent in construction. The reader can follow how you got from the raw transcripts to this model. The path goes through codes, code memos, theoretical memos, draft diagrams, revision history. The audit trail makes the model inspectable.
3. Revisable. The model is provisional. It accommodates disconfirming evidence by revision, not by denial. A model that has not been revised after its first articulation is suspect — either the data are unusually uniform (rare) or the analyst has stopped looking.

A model that meets these three standards is a finding. A model that does not is decoration. The discipline of model-building is in the verification, not the drawing.

Introduction and Overview

The previous three sections gave you the conceptual furniture: the analytic-strategy map, memos as the engine of interpretation, triangulation and crystallization as checks on that interpretation, and the four kinds of conceptual model with their criteria of defensibility. Section 4 is where you put the furniture into a room. The Week 6 capstone milestone is to build a first conceptual model of the loneliness phenomenon — not your final model, not the model that will eventually be in the discussion of your paper, but a defensible first articulation, grounded in the dataset, that the rest of the term will revise.

Bernard, Wutich, and Ryan are explicit on the value of producing a first model early. The exercise forces you to commit to an interpretation in a visible form. It exposes the gaps in your coding (cases that don't fit are immediately visible). It gives the rest of the term something to revise, rather than a blank page. The most common failure of capstone projects is not producing a model at all — students stop at coded data, write a flat findings section, and never make the move to interpretation. Producing a first model at Week 6 prevents that failure.

4.1 Worked Model #1 — A Taxonomy of Kinds of Loneliness

The taxonomy worked out in Section 3.1 used four top-level kinds (existential, relational, structural, situational) with nine sub-kinds. Let's stage what this would look like as a finished first-pass deliverable.

What this taxonomy claims, in one sentence

The interpretive claim the taxonomy makes is: loneliness in this dataset has at least four distinct kinds (existential, relational, structural, situational), each with its own internal structure of sub-kinds, and the kinds differ in what kind of intervention would even make sense. That last clause matters for a public-health audience: a taxonomy that distinguishes kinds with different intervention implications is doing applied work, not just classificatory work.

4.2 Worked Model #2 — A Typology Cross-Classifying Triggers and Responses

Section 3.2 sketched a 4×4 typology crossing trigger kind with response kind. As a worked capstone deliverable, this would proceed as follows.

The dimensional choice. The typology requires you to commit to two dimensions. The two we chose are trigger and response. Other defensible choices: kind of loneliness (existential/situational) × relational presence at the moment (alone/companioned); onset (acute/chronic) × attribution (personal/structural); life-stage (young/mid/late) × support availability (high/medium/low). The choice of dimensions is itself an interpretive move and must be defended in the memo.

The cell-by-cell reading. Once cases are located in cells, the typology yields findings that no taxonomy could:

• Existential triggers cluster with reframing responses. Daniel (P10) and Kenji (P14) both face triggers that no behavioural intervention can address (the human condition; a fundamental life change). Their adaptations are interpretive — the tea ceremony, the support group of similarly-situated men. The empty cells in the existential row (no behavioural-change-only response, no withdrawal-only response) are themselves a finding: existential loneliness does not have a behavioural fix.
• “No consistent strategy” clusters with structural triggers. Margaret (P13) in long-term care has no consistent strategy because the structural condition (institutional placement) does not yield to her individual response. This is the typology surfacing a finding about where interventions are needed at the structural rather than individual level.
• Situational triggers populate every response column. Acute-event loneliness produces the most response variability — people try things, some work, some don't, the variability itself reflects the open-endedness of acute coping.

These cell-level findings would each get a paragraph in the eventual paper. None of them would be visible from a flat list of codes. The typology is doing analytic work.

4.3 Worked Model #3 — A Process Model of the Loneliness-Coping Cycle

Section 3.4 sketched the six-stage cyclical process model: trigger → onset → interpretation → response selection → enactment → outcome (with feedback to interpretation and response selection). As a worked capstone deliverable, this becomes:

The process model claims something the taxonomy and typology don't: loneliness is dynamic, cycles repeat, and each cycle modifies the next. That claim is testable against the data — the transcripts should show the cyclic structure if the model is right. Sarah's account of cycles of crying, joining the moms group, finding Tessa, and using the moms group differently afterward is exactly the kind of evidence the process model predicts.

4.4 Choosing Which Model to Build

For your capstone Week 6 milestone, you will build one of these models. The choice depends on which kind of model best fits the research question you have been developing.

Some students find it productive to build two models — for example, a taxonomy of kinds and a process model of how each kind unfolds. That is allowed and often illuminating. The Week 6 minimum is one defensible model with its 500-word justifying memo.

4.5 The Discipline of Disconfirming Cases

The single most important methodological move in this lesson is the discipline of naming which transcripts do not fit your model. This is what separates a defensible interpretation from a confirmation exercise.

Bernard, Wutich, and Ryan are explicit (2017, p. 192) that negative-case analysis — the deliberate search for cases that disconfirm the emerging interpretation — is one of the strongest tests available in qualitative analysis. It does in qualitative work what falsification does in quantitative work. A model that has not been tested against disconfirming cases is no stronger than a hypothesis with no power calculation.

For your capstone deliverable, the memo must include a paragraph honestly naming the transcripts that don't fit. There are three productive responses to such cases:

1. Revise the model. If a transcript doesn't fit, perhaps the model is wrong. Revising to accommodate it strengthens the model.
2. Bound the model's scope. Perhaps the model fits a subset of the data and not others. Articulating the scope explicitly is honest and useful.
3. Identify the case as boundary-marking. The non-fitting case shows where the model's reach ends — itself an analytic claim.

What is not a productive response is ignoring the case or quietly dropping it. That move would not survive review and is the kind of error that disciplined memo-writing, applied across the term, is supposed to prevent.

4.6 The Week 6 Capstone Milestone

Bringing It All Together

Lesson 6 sits at the analytic hinge of HSCI 841. The first five lessons gave you definitions, design, sampling, data, and codebooks. Lesson 6 introduced the analytic-strategy map, the memo discipline, triangulation and crystallization as checks on interpretation, and the four kinds of conceptual model with their criteria of defensibility. The Week 6 milestone asks you to commit to a first model and to engage honestly with the cases that don't fit it. The next six lessons — on grounded theory (7), content analysis (8), schema and narrative analysis (9), discourse analysis (10), analytic induction and QCA (11), and computational text analysis (12) — will give you specific techniques for elaborating and revising that first model.

The single move this lesson is asking you to internalize: codes are categories; interpretation is the claim about why those categories are there; and a defensible interpretation is one that has been written as memos, checked through triangulation, rendered as a conceptual model, and tested against disconfirming cases. That move is what separates qualitative work that public-health audiences take seriously from qualitative work that they politely set aside.

The final reflection below asks you to step out of method-mode and articulate the analytic stance you carry forward into the second half of the term.

This glossary collects the key concepts, people, and methodological stances introduced in Lesson 6. Use it as a reference while you work through the material, or as a review before the final assessment. Type in the search box to filter entries.