Start with a transcript you can actually code
The analysis is only as good as the transcript under it. Before you code a word, you want a clean, speaker-labeled, timestamped transcript – ideally line-numbered, so you can point a co-coder to "line 214." Producing that artifact is a job of its own; the qualitative research transcription workflow covers making and formatting it. This guide picks up once the transcript exists.
Decide your verbatim level before you analyze, because it shapes what you can claim. Strict verbatim – every "um," pause, and false start – matters for discourse or conversation analysis, where how something is said is the data. For most thematic work, clean verbatim (filler removed, words intact) is enough. Apply one style consistently across every interview in the set.
Keep the audio close. A transcript is a reduction of the recording, and you'll want to re-hear ambiguous lines while coding. Timestamps make that a two-second jump instead of a scrub through an hour of tape. Flag anything genuinely unclear as [inaudible] with its timestamp, rather than guessing at a word you might later quote.
Coding turns interview transcripts into analyzable data
Coding means labeling segments of the transcript so you can retrieve and compare them. Johnny Saldaña frames it in cycles: First Cycle coding assigns codes to portions from a single word to a full page, and Second Cycle coding recodes and regroups those into categories, themes, and concepts (Saldaña, 2013). Coding is cyclical – the first pass is rarely the last.
Grounded theory supplies the classic coding vocabulary. Strauss and Corbin's approach moves from open coding to axial coding, where categories are reconnected around a "coding paradigm" (Kelle, 2005). Open coding fractures the data into concepts; axial coding builds the relationships between them.
Not everyone accepted that. In Emergence vs. Forcing (1992), Glaser argued that axial coding and coding paradigms "force" categories onto the data instead of letting them emerge (Kelle, 2005). The practical lesson: a coding structure is a tool, not a verdict. Hold it loosely enough to notice what it hides.
However you code, tie each code to evidence you can attribute. Pull the exact line, with its speaker and timestamp, as you tag it – a timestamped quote is what turns a code into something you can defend in the write-up. Codes without traceable extracts are just opinions wearing labels.
Braun and Clarke's six phases of thematic analysis
Thematic analysis is the most common route for interview data, and Braun and Clarke's six-phase guide is its reference procedure. The phases are: familiarizing yourself with the data, generating initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the report (Braun & Clarke, 2006). They're recursive, not strictly linear.
A theme is more than a topic that keeps coming up. Braun and Clarke define it precisely: a theme "captures something important about the data in relation to the research question, and represents some level of patterned response or meaning within the data set" (Braun & Clarke, 2006). Frequency alone doesn't make a theme.
One reframing matters. Braun and Clarke reject the idea that themes "emerge" from data; in reflexive thematic analysis, themes are "analytic outputs, created from codes and through the researcher's active engagement with their data" (Braun & Clarke, 2019). You build themes; you don't find them lying in wait.
That's also why they don't advocate inter-rater reliability scores or a fixed coding frame for reflexive TA (Braun & Clarke, 2019). Coding "inescapably bears the mark of the researcher," so there's no single accurate coding to agree on. If your design needs multiple independent coders, that's a coding-reliability approach – a different tradition. Be clear which you're using.
What CAQDAS software does, and what it can't
Software organizes coding; it doesn't do the thinking. A peer-reviewed reflection on NVivo puts it plainly: the main function of CAQDAS "is not necessarily to analyse data, but rather to aid the analysis process, which the researcher must always remain in control of" (Zamawe, 2015). The interpretation stays yours.
What these tools do well is store, tag, and retrieve. To get a transcript in, export a clean DOCX first. ATLAS.ti imports .doc, .docx, .rtf, and .txt, and separately imports .srt and .vtt caption files as transcripts; NVivo imports Word and text sources and PDFs, converting formats like PowerPoint to PDF first. An audio-to-DOCX export lands cleanly in either.
Speaker structure is worth preserving on import. Both packages can auto-code by speaker when your transcript's turns are formatted consistently – the payoff for tidy speaker labels upstream. The dissertation transcription guide walks through the speaker-turn import mechanics if you're coding in NVivo or ATLAS.ti.
Manual coding is still legitimate. For a handful of interviews, colored highlighters, a spreadsheet, or a word processor's comments work fine. Software earns its place when the data set is large, several people code together, or you need to retrieve every extract under a code across dozens of transcripts.
How many interviews before the themes hold up?
Fewer than most people expect, if your sample and question are tight. A well-known Field Methods study found saturation within the first twelve interviews, with the basic elements of metathemes present by six (Guest, Bunce & Johnson, 2006). Scope, not a magic number, drives it.
A 2022 systematic review put a range on it. Hennink and Kaiser found studies reached saturation within 9–17 interviews, or 4–8 focus groups, for homogeneous populations with narrowly defined aims (Hennink & Kaiser, 2022). Broader questions and more varied participants push the number up.
Saturation is about having enough data; rigor is about using all of it. The temptation is to quote the three vivid lines that fit your argument. Silverman's rule cuts against that: do not look for telling examples, but "analyse your data thoroughly and fairly" (Silverman, 2011). Test each theme against the whole set.
Look for the disconfirming case on purpose. A theme that survives the interview that seemed to contradict it is stronger than one built only from supportive quotes. Report the counter-examples and how you accounted for them – that's what separates analysis from a highlight reel.
