45 Qualitative Inquiry: Interviews, Documents, and Themes
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
The research decision. Decide whose accounts you gather and why those cases, and then show that every theme you report traces back to what people actually said or wrote. A tool can search your transcripts, but it cannot choose your cases and it was not in the room, so no tool gets to decide what your participants meant.
45.1 Why this decision matters
The decision on the table: whose accounts you gather, and whether every theme you report can be walked back to the words that support it.
“You interviewed nine people and came back with three tidy themes. Fine. Now show me where each theme lives in the transcripts, show me the person who did not fit, and tell me why these nine and not nine others. If you cannot, what I am reading is your opinion with quotation marks around it.” — a qualitative-methods discussant, reading a conference paper on how small business owners set prices
Some of the best questions in research are not about how much. They are about how people make sense of a situation, what they actually do, and in what order. How does a restaurant owner decide when to raise prices? Why do some eligible students never register to vote? A number cannot answer those questions well, and a careful set of interviews can. The same interviews can also produce some of the most fluent nonsense in the social sciences: themes that are really the researcher’s hunches, quotes chosen to decorate them, and a sample nobody can explain. This chapter gives you the two moves that keep an interview study honest: a defended choice of cases, and an evidence trail from every claim back to the page.
45.2 The concept
Everything in this book so far has ended in a number with an interval around it. This route ends somewhere else, and it is worth saying plainly how the book’s framework stretches to reach it. The authors of RDSS state that the MIDA framework describes qualitative designs as well as quantitative ones, that qualitative data are data too, and that in some designs the answer strategy is the researcher reading documents and summarizing their meanings (Blair et al. 2023). Their observational descriptive library names classic case studies built on “thick description” as part of that family. This chapter keeps the four parts and adapts two of them. The Inquiry becomes a precisely worded question about meanings, practices, or processes. The Answer becomes a bounded account in which every claim traces to evidence. Diagnosis changes too: you will not compute power or bias for an interview study, and you should never invent such numbers. You diagnose it instead with questions a stranger could answer from your records. Did you choose cases that could show where each theme stops applying? Can a reader walk from every claim back to the words behind it? Did you write down how your own position shaped what you heard?
On the compass, most interview studies ask a descriptive question whose reach is the data at hand: the people and documents you actually studied. The warrant for the account is not a sampling design. It is an evidence trail a stranger can walk, cases chosen for the contrasts your question needs, and a written record of your own part in the analysis. That placement has a consequence you must respect.
Purposive case selection is choosing participants on purpose, because each one can show you something the question needs, rather than drawing them at random. Example: you interview owners of a food truck, a family diner, and a franchise café because cost shocks hit each kind of business differently. Purposive selection holds no sampling license, so the crossing to a population stays closed and no prevalence claim is available. You may not report that “most owners” do anything, because your nine owners were never a random slice of all owners. Small argues that interview studies should stop pretending to be small surveys and instead follow a case logic, choosing each new case to test or extend what the earlier ones suggested (Small 2009).
A semi-structured interview is a conversation guided by a written list of topics and open questions, with room to follow what the person raises. Example: “Walk me through the last time you changed a price on your menu” is a guide question; “What happened next?” is a probe, a neutral follow-up that asks for more without suggesting an answer (Weiss 1994). Neutral wording matters. “Did rising costs force you to raise prices?” hands the owner your theory; “Tell me about the last price change” lets the owner supply their own.
Thematic analysis is a method for building patterns of meaning across a set of transcripts or documents. Braun and Clarke describe it in phases: you read and reread until you know the material, attach short labels called codes to meaningful stretches of text, group codes into candidate themes, review those themes against the whole dataset, then define, name, and write them up (Braun and Clarke 2006). A theme is a pattern of shared meaning held together by one central idea, which Braun and Clarke call a central organizing concept (Braun and Clarke 2019). Example: “pricing” is a topic; “owners shield the dishes regulars watch and move increases onto items nobody tracks” is a theme, because one idea about how owners think unites its excerpts. A topic summary is the look-alike to avoid: it gathers everything people said about one subject, often one interview question, and reports it without an idea that ties it together.
Braun and Clarke later distinguished three families of thematic analysis (Braun and Clarke 2019, 2021a). Reflexive thematic analysis treats coding as an act of interpretation by a researcher whose position shapes what they see. Example: an analyst who once worked in a kitchen will notice supplier talk that an outsider misses, and says so. Coding reliability analysis fixes a coding scheme early and has several coders apply it independently, so that their agreement can be measured. Example: two coders label the same forty excerpts and report how often they chose the same code. Codebook thematic analysis sits between the two: it uses a structured scheme for practical reasons, such as a team that must share one set of codes, while still treating coding as interpretation. The choice decides your quality checks, so name it before you code. In coding reliability work, two coders agreeing is the check the method was built around. In a reflexive approach, an agreement statistic answers a question the method never asked, so do not import one to make reflexive work look rigorous. Reflexive work earns trust differently: through themes with a clear central idea, a trail a stranger can walk, contrary cases that mark each theme’s edge, a record of your own position, and description rich enough for a reader to judge fit. A second reader still helps, as a critical friend who offers another reading of the same excerpts, not as a vote you tally.
Reflexivity is the written practice of noticing how your own background, expectations, and presence shape what participants tell you and what you hear. Example: a short memo after each interview noting “I expected owners to blame suppliers; this owner blamed her landlord instead.”
An evidence trail is the chain that lets a reader walk from any claim in your report back to the words that support it: excerpt, then code, then theme, then claim. Example: excerpt E07 (“Now I raise the fries instead”) carries the code shift to sides, which supports the theme about shielded dishes. A contrary case is a participant or document whose account runs against a theme. Example: the one owner who changes prices on a screen every week, and so never waits for a menu reprint. You report contrary cases because they tell you where a theme stops applying.
Two words govern how far your account may travel. Generalization in the statistical sense is a claim about a population, licensed by sampling. Transferability is a reader’s judgment about whether your account fits another setting, which you make possible by describing your cases and their context in enough detail (Tracy 2010). Example: a reader in a coastal tourist town can judge whether your account of inland diners fits her cafés only if you said what those diners were like. Small adds that a well-chosen sequence of cases supports logical inference, an argument about how a process works and under what conditions, rather than a statistical claim about how often it happens (Small 2009).
Last, a word you will meet everywhere. Saturation is the idea that you can stop collecting when new interviews stop adding anything. It is contested. Braun and Clarke argue that in reflexive work, meaning is made through interpretation, so the point at which to stop cannot be fixed in advance by a count (Braun and Clarke 2021b). Never promise a reader that twelve interviews will be enough. Promise instead a stopping rule you can defend in terms of your question, such as “until each contrast my question needs has been heard, and each theme has met at least one account I sought out because it might complicate it.” Stopping is still your judgment, so say that it was.
Counting has one honest job here. Case coverage records which of your participants, and which kinds of participants, a theme draws on. Example: a theme heard only from sit-down diners has never met a counter-service owner, so you do not yet know whether it holds there. Coverage tells you where a theme has and has not been examined. It does not tell you how important a theme is, and Braun and Clarke warn that a theme’s weight need not rest on how often it appears (Braun and Clarke 2006). Nor does it tell you whether a theme is right.
45.3 Gathering accounts without harming anyone
Interviews put real people’s words on the record, so the ethics start before the first question. Settle your project’s permission status, as the ethics chapter describes it, before you recruit anyone: most interview studies need a determination from a human-subjects review office. Informed consent here means the person knows what you are studying, that the conversation is recorded, how the words will be stored and quoted, and that they may stop at any time. Record only with explicit permission. Anonymization is removing or changing details that could identify a person. Example: “a diner owner in a mid-sized town” rather than the restaurant’s name, street, or signature dish. In a small community a signature dish can identify an owner as surely as a name.
Transcripts are also a new exposure risk. Pasting a transcript into an AI tool can be a disclosure to a third party your participants never agreed to, as the ethics chapter explains. Check what your consent form and your review office permit before any transcript leaves your machine. When a tool is allowed, send the smallest de-identified fragment that does the job, and log what you sent.
45.4 A worked example
In the book’s first chapter you compared menu prices across years. This time you ask the owners. Your question, worded precisely: how do owners of independent restaurants decide when and how to change menu prices after their costs rise? That is a descriptive question about a practice. It does not ask how much prices rose, and it does not ask what caused the rises.
You choose eight owners on purpose. You include printed-menu and screen-menu restaurants, since the cost of reprinting might shape timing. You include owners who rent and one who owns the building, and a counter-service place beside a sit-down diner. That spread is your case logic: each owner can confirm or complicate what the others said. You write a guide with five open questions and neutral probes, interview each owner for about forty minutes, and transcribe.
You work alone and treat coding as interpretation, so your approach is reflexive thematic analysis. That choice sets your checks before you start: no agreement statistic, and instead a memo after each interview, a trail from every claim to its words, and a hunt for the owner who does not fit.
Then you code. Three candidate themes survive review. Owners reprice in one round because reprinting menus costs money. Owners shield the watched dishes and move increases onto sides. Owners wait out a passing shock before touching any price. One owner, O4, runs a screen menu and changes prices most weeks. That is a contrary case for the first theme, and it sharpens it: the one-round pattern belongs to printed menus, and your claim should say so. Your memo after O4 records that you had expected every owner to batch changes, and that this interview changed your mind, so a reader can see where the account moved and why.
Quality in interview research shows in things a reader can check, such as whether the account shows differences among participants and whether the researcher followed up on what people said (Small and Calarco 2022). The trail below makes those checks possible.
Every quote and transcript in the block below is constructed for the example. The block builds the evidence trail, confirms each quote appears word for word, shows which owners and which kinds of restaurant each theme draws on, and then checks a quote an AI tool offered.
import numpy as np, pandas as pd
SEED = 464
rng = np.random.default_rng(SEED)
# Constructed transcripts: eight owners, one line each, for the example only.
transcripts = {
"O1": "We reprint the menu once a year. Reprinting costs real money, "
"so everything goes up at once.",
"O2": "The breakfast burrito stays at eight dollars. People notice that one, "
"so the increase goes on the sides.",
"O3": "Eggs doubled last spring. I ate it for two months because I figured "
"it would pass.",
"O4": "Our prices live on a screen now. I change them most weeks when the "
"supplier invoice lands.",
"O5": "New menus cost me four hundred dollars, so I wait and do all the "
"prices in one go.",
"O6": "Nobody tracks what a side salad costs. They track the burger.",
"O7": "If beef stays high past the summer, then I move the price. Not before.",
"O8": "I raised the burger and lost two regulars in a week. Now I raise the "
"fries instead.",
}
# Two of the contrasts you chose the owners for: menu medium and service.
cases = pd.DataFrame({
"menu": ["printed", "printed", "printed", "screen",
"printed", "printed", "screen", "printed"],
"service": ["sit-down", "counter", "sit-down", "counter",
"sit-down", "sit-down", "sit-down", "counter"],
}, index=[f"O{i}" for i in range(1, 9)])
ONE, SHIELD, WAIT = ("reprice in one round", "shield the watched dishes",
"wait out a passing shock")
rows = [ # excerpt, owner, code, theme, quote
("E01", "O1", "reprint cost", ONE, "Reprinting costs real money"),
("E02", "O5", "reprint cost", ONE, "New menus cost me four hundred dollars"),
("E03", "O5", "batch the changes", ONE, "do all the prices in one go"),
("E04", "O4", "prices change weekly", ONE, "I change them most weeks"), # CONTRARY
("E05", "O2", "anchor dish", SHIELD, "The breakfast burrito stays at eight dollars"),
("E06", "O6", "anchor dish", SHIELD, "They track the burger"),
("E07", "O8", "shift to sides", SHIELD, "Now I raise the fries instead"),
("E08", "O2", "shift to sides", SHIELD, "the increase goes on the sides"),
("E09", "O3", "absorb a short spike", WAIT, "I figured it would pass"),
("E10", "O7", "judge if it lasts", WAIT, "If beef stays high past the summer"),
]
trail = pd.DataFrame(rows, columns=["excerpt", "owner", "code", "theme", "quote"])
trail["contrary"] = trail.excerpt.eq("E04")
# Every quote must appear WORD FOR WORD in its owner's transcript.
trail["verbatim"] = [q in transcripts[o] for o, q in zip(trail.owner, trail.quote)]
print(f"quotes found verbatim : {trail.verbatim.sum()} of {len(trail)}")
# Case coverage: WHICH of your owners, and which kinds, each theme draws on.
# It shows where a theme has and has not been heard. It is not a score.
def kinds(owners, col):
return "/".join(sorted(cases.loc[sorted(set(owners)), col].unique()))
coverage = pd.DataFrame({
theme: {
"owners": ",".join(sorted(set(g.owner[~g.contrary]))),
"against": ",".join(sorted(set(g.owner[g.contrary]))) or "-",
"menu": kinds(g.owner[~g.contrary], "menu"),
"service": kinds(g.owner[~g.contrary], "service"),
}
for theme, g in trail.groupby("theme")
}).T
print(coverage.to_string())
ai_quote = "I never pass cost increases on to my customers"
print(f"\nAI quote found in O3's transcript : {ai_quote in transcripts['O3']}")
audit = sorted(str(e) for e in rng.choice(trail.excerpt.to_numpy(), 3, replace=False))
print(f"excerpts for a critical reader : {audit}")
print("\ncoverage shows where each theme was heard among THESE eight owners;")
print("it is not a prevalence estimate and not a measure of a theme's importance")Read the output line by line. All 10 quotes appear verbatim, so every excerpt in the trail is real. The coverage table says where each theme has been heard, not how strong it is. The one-round theme comes from O1 and O5, both printed-menu sit-down diners, and O4’s screen menu marks its edge. The shielding theme comes from O2, O6, and O8, across counter-service and sit-down places, all with printed menus. The waiting theme comes from O3 and O7, across both kinds of menu, but only from sit-down diners. None of this says how many restaurant owners anywhere behave this way. What it does say is where to look next. No counter-service owner has spoken to the waiting theme, so your next interview goes to one, ideally an owner who raised prices the same week costs jumped. Two counter-service owners with printed menus, O2 and O8, never mentioned reprinting, so you follow up with them before claiming the one-round pattern for every printed menu. The quote the AI tool attributed to O3 does not appear in O3’s transcript, so it never enters your report. The three randomly drawn excerpts (E02, E04, and E05) go to a colleague who reads each one in its full transcript and offers their own reading. Drawing them at random keeps you from handing over only your best material. Where the colleague reads an excerpt differently, you note it in your memo and decide whether the theme needs to change.
45.5 What makes the account trustworthy
You cannot attach a confidence interval to a theme, and pretending otherwise would be a worse failure than reporting none. The uncertainty is still there, and you state it in the terms the route allows. Say which cases, and which kinds of cases, each theme draws on, where it has not yet been examined, and who does not fit. Say what kinds of owners you never reached: here, no owners who closed after the cost shock, which is a survivorship gap worth naming, and none of the owners who declined your invitation. Say what your own position may have led you to hear.
Tracy proposes a set of broad criteria for judging qualitative work, among them sincerity about the researcher’s role and credibility built from detailed, concrete description (Tracy 2010). Small and Calarco point to signs a reader can spot in the text itself, such as whether participants come across as different people rather than one voice, and whether the researcher followed up on surprises (Small and Calarco 2022). Their list also asks for self-awareness, the researcher’s grasp of how their own presence shaped what people said, which is reflexivity seen from the reader’s side. Neither list is a checklist to tick. Both are ways for a stranger to see whether your account earned their trust.
Two boundaries stay firm. First, prevalence: your coverage lines describe your cases, and a claim about how common a practice is needs a sampling design, such as a survey that asks a random sample of owners. Second, cause: owners can tell you the reasons they give for waiting, and those reasons are evidence about how they decide. They are not an estimate of what a cost shock does to prices. If your question turns into “does reprinting cost delay price increases?”, you are asking a causal question, and a count of owners who mention reprinting leaves it causal, currently unidentified. An argument about causes inside a single case needs its own design, such as the process tracing described in the observational-causal chapter, not a larger tally of themes.
45.6 An AI failure case
You upload the six de-identified transcripts whose consent forms allow it to an AI tool, and ask it to “identify the main themes, with supporting quotes.” It returns a clean table in seconds: Cost pressures, Customer reactions, Timing of changes, Future plans. Each theme has two neat quotes. It reads like a finished analysis.
Look closer and two failures appear. The four themes are four of your interview guide’s five questions, handed back as headings. They are topic summaries: each gathers what owners said on one subject, and no central idea ties any of them together. That is a plausible-but-wrong-method: the output looks like thematic analysis while skipping the interpretation that makes a theme. Then one quote under Customer reactions reads “I never pass cost increases on to my customers.” Owner O3 never said that. She said she absorbed a two-month spike because she expected it to pass. The tool compressed her words into a stronger, cleaner sentence and put quotation marks around it. That is confident fabrication, and in an interview study it puts words in a real person’s mouth. A tidy table is not an analysis.
You catch both by returning to the transcripts. Search every quote in its source, word for word, as the worked example does. Then ask of each theme, “what single idea holds these excerpts together, and where does it stop applying?” If the only thing uniting them is that they answer the same interview question, it is a topic. Rebuild the themes from your own codes, drop the invented quote, and log why.
45.7 It is your turn
A further route beyond the five pathways. This lesson extends Studio 5: Develop the pathway. Read it once you have declared your primary pathway and your question calls for this design. Studio 5’s milestone asks for the same decisions, answered for this route.
Your design is declared. If your question asks how people make sense of something or what they actually do, this is where you choose your cases and build the trail that lets a stranger check every theme.
The hands-on half of this section lives in the chapter’s companion notebook: open it in Colab with the badge at the top, and work the steps there.
Commit your own answer first, then delegate. Each prompt below is a checkable job, not a request for a verdict. Work them as a loop. The first answer is a draft: find the claim you cannot check, say so in your next message, and run it again. Some tools will run that whole loop unattended and hand you a finished set of themes. The finished look is exactly what makes the questions in this chapter worth asking before you accept it.
Three calls stay yours. You decide which cases you study and why each one earns its place, because only you know what your question needs to see. You decide what a participant meant, by reading their words in full and in context. And you decide which themes you are willing to defend, together with the contrary cases that bound them. A tool can suggest codes, draft a guide, and search for quotes. It cannot have been in the room, and it cannot sign your account.
Write your inquiry as one precise question about meanings, practices, or processes. Underneath it, write the answer form: “a bounded account of how [who] [do what], with every claim traced to excerpts.” Name the warrant in one line: an evidence trail, cases chosen for the contrasts your question needs, and a reflexive record, not a sampling design. Name your approach (reflexive, codebook, or coding reliability) and the quality checks that belong to it. Then write the prevalence claim and the causal claim you forbid yourself, in full.
Choose your cases and defend each one in a line. Name the contrasts your question needs (kinds of people, settings, or documents), then list the first cases you will approach and what each could show you. Add one case chosen to challenge your expectation.
When you are ready to delegate this step:
Act as a qualitative research advisor. My question is: "[paste your question]." Here are the cases I plan to interview and why: [paste list]. Do not add cases for me. Name the contrasts my list is missing, the kind of person whose absence would most bias my account, and one case that could contradict what I expect. Return a table.After running, verify: compare each suggested contrast with your own question and keep only those your inquiry needs. If every row praises your plan, ask “which case would most embarrass my expected answer?” Counters sycophantic agreement (praise that reviews your ego, not your evidence).
Draft your interview guide yourself: five to seven open questions, each with a neutral probe. Then have it checked for leading wording.
When you are ready to delegate this step:
Act as an interview methods reviewer. Here is my interview guide: [paste guide]. Flag every question that suggests an answer, assumes a cause, or uses words a participant might not share. Quote the exact words. Do not rewrite the guide for me.After running, verify: read each flagged question aloud as if you were the participant, and rewrite only the ones that push you toward an answer. Then check the questions it did not flag; a clean report on a guide you know has one leading question is an illusion of completeness (a tidy review that omits the one problem that matters).
Build your evidence trail as a table: excerpt ID, participant, code, theme, the verbatim quote, and a column marking contrary cases. Write each theme around one central idea in a full sentence, not a topic label. For each theme, add a coverage line naming which participants, and which kinds of participants, it draws on, where it has not yet been examined, and the contrary case that marks its edge. Coverage points your next interview; it does not rank your themes.
When you are ready to delegate this step:
Here are my themes and the excerpt IDs behind each: [paste table without names]. Act as a skeptical discussant. For each theme, say whether one central idea holds its excerpts together or whether it is only a topic summary, which excerpts sit uneasily under it, and which kinds of participants it has not yet been examined against. Do not propose new themes.After running, verify: check every coverage statement it makes against your own table, and if it treats a theme heard from fewer participants as weaker, set that aside, because coverage is not importance. If its answer quotes anyone, find that quote word for word in the transcript, as the worked example does in code. A quote you cannot find is a confident fabrication (a clean sentence nobody said), so delete it and trace the claim again.
Write the limits paragraph of your account: who you never reached, what your position may have led you to hear, how far a reader could transfer your account, the stopping rule you used, and why you judged it met. Before any transcript goes to an AI tool, write one line on what your consent form permits and what you removed.
Log the step in your AI Research Ledger, and verify at least one output with a named method from the Verification Guide. Peer reasoning is the natural check for this route: give a colleague three excerpts drawn at random, ask them to read each in its full transcript and offer their own reading, and record where it differs from yours and what you changed. Primary-source reading is the strong second check: every quote an AI tool touched goes back to its transcript before it enters your report. An AI reviewer may question your themes with you; the decision about what your participants meant stays yours.