Guides / discourse analysis
Discourse analysis: a practical guide
Discourse analysis treats language as something people do rather than a window onto what they think — and that single shift is what most dissertations claiming the method never actually make. This guide explains the approaches, shows what analytic attention looks like on a real extract, and covers what an examiner expects to see.
Written and reviewed by Elaine Halliburton, Professor of Applied StatisticsDiscourse analysis is a qualitative approach that examines how language is used to accomplish things — to justify, position, persuade, deflect or construct a version of events. Rather than treating what people say as a report of their inner views, it asks what a particular way of saying something is doing in that context, and what it makes possible or forecloses.
Definition
What discourse analysis is
Discourse analysis asks what language is doing, not what it reveals. When a participant says something, the question is not primarily “what does this tell me about their beliefs?” but “what is being accomplished by putting it this way, here, to this listener?”
This is a genuine shift in what counts as data. In most qualitative work, an interview transcript is treated as a route to something behind it — experiences, attitudes, perceptions. In discourse analysis the talk is the object of study. How something is phrased, what is treated as obvious, what is left unsaid and what work a hesitation does are the findings, not the noise around them.
The intellectual roots explain why the method looks so different from its neighbours. It draws on ordinary language philosophy, where Austin observed that many utterances do not describe the world but act on it — saying “I promise” is not a report of a promise, it makes one. Extending that insight, discourse analysts argue the same holds far more widely than the obvious cases: describing a situation as inevitable, or a group as reasonable, does work in the conversation regardless of whether it is accurate. That is why the analysis attends to construction rather than accuracy, and why asking whether a participant was telling the truth is beside the point.
Because the analysis is of language use, you cannot tidy the transcript. Repetitions, false starts, hedges and pauses are not transcription errors to be cleaned up — they frequently carry the analysis. A smoothed transcript has had the data removed from it.
The distinction
How it differs from thematic analysis
| Thematic analysis | Discourse analysis | |
|---|---|---|
| Treats talk as | Evidence of experience or view | Social action in its own right |
| Asks | What are they saying? | What are they doing by saying it this way? |
| Output | Themes across participants | Discursive patterns, repertoires, positions |
| Extracts used to | Illustrate a theme | Demonstrate the analytic claim itself |
| Transcript detail | Clean, readable | Detailed — pauses, overlaps, emphasis retained |
| Context | Often backgrounded | Central — who is speaking to whom, and why |
The clearest test is what your extracts are for. In thematic analysis an extract illustrates a theme you have already stated, and a different extract making the same point would serve equally well. In discourse analysis the extract is the analysis: you are showing the reader specific features of this piece of language and arguing about what they accomplish, so it cannot be swapped for another.
Coding transcripts for themes, reporting the themes, and calling it discourse analysis because the data are talk. If your findings could be summarised as “participants felt X” or “three themes were identified”, you have done a thematic analysis. Describing it accurately is a stronger position than claiming a method you did not use.
Send your research question and a sample of your data. A named researcher advises which method fits and what it will require.
Get a fixed quoteVariants
The main approaches
| Approach | Focus | Associated with |
|---|---|---|
| Discursive psychology | How psychological states are constructed and deployed in talk | Potter, Edwards |
| Foucauldian | How discourses constitute what can be said and known | Foucault, Parker, Willig |
| Critical discourse analysis | How language reproduces power and ideology | Fairclough, van Dijk, Wodak |
| Conversation analysis | The sequential organisation of interaction | Sacks, Schegloff |
| Narrative | How accounts are structured as stories | Riessman |
These differ substantially in what they treat as data and what counts as a finding. Conversation analysis works on naturally occurring interaction with extremely fine transcription and would regard a research interview as a peculiar object of study. Foucauldian analysis often works on policy documents and institutional texts and pays little attention to turn-by-turn detail. Choosing between them is a decision about your question, not a matter of preference.
Software is of limited help. NVivo and MAXQDA will store transcripts and let you tag stretches of text, which is useful for assembling every instance of a pattern once you know what you are looking for. What they cannot do is notice that a disclaimer is doing work, because that requires reading. Expect the analysis to be done by hand on paper or on screen, with software used afterwards to organise what you found rather than to find it.
CDA is explicitly political: it assumes language reproduces power relations and sets out to expose that. This is a legitimate position, but it must be declared rather than smuggled in. An analysis that arrives at conclusions about power without having stated the framework that made them findable will be challenged on exactly that point.
Analytic attention
What to look for in a transcript
There is no coding frame in the thematic sense. What there is instead is a set of features that reliably repay attention.
| Feature | What to notice |
|---|---|
| Extreme case formulations | “always”, “never”, “everyone” — strengthening a claim against challenge |
| Disclaimers | “I'm not being difficult, but…” — pre-empting an unwelcome reading |
| Footing shifts | Moving between I, we, you and one — relocating responsibility |
| Nominalisation | “the closure of the unit” rather than “they closed it” — agency disappears |
| Modality | “must”, “might”, “obviously” — how much certainty is claimed |
| Consensus building | “everyone knows” — making the claim hard to dispute |
| Interest management | Handling the risk of sounding self-serving |
| Contrast structures | Setting up a version by opposing it to an alternative |
| Hesitation and repair | Where the talk becomes difficult, and what is being managed |
A related concept worth knowing is the interpretative repertoire: a recognisable, culturally available way of talking about something that speakers draw on and that comes with its own vocabulary and implications. In accounts of health, for example, a biomedical repertoire and a lifestyle repertoire are both readily available, and which one a speaker reaches for shapes where responsibility lands. Identifying the repertoires in circulation, and noticing when speakers switch between them mid-account, is often the most productive route into an interview corpus.
What is not said often matters as much as what is. If every participant discusses resourcing and none mentions who decides it, that shared silence is a finding — but only if you can evidence the pattern across the corpus rather than asserting it about one extract.
Worked example
Worked example: analysing an extract
A study examines how senior staff account for provision decisions. One extract, and what analytic attention finds in it.
What is being accomplished
The speaker faces a problem: acknowledging a limitation on what students receive while not appearing indifferent to them. The extract manages that problem in four moves.
Taken together, these construct constraint as external and self-evident, and position the speaker as a reasonable person responding to circumstance rather than as an actor making choices. The analytic claim is not that the speaker is being dishonest — that would be a claim about their beliefs, which is not what this method examines. It is that this way of talking makes the decision difficult to contest.
Notice too what the interviewer contributes. The tag question — “haven't you” — invites agreement, and whatever the interviewer did next will have shaped what followed. In interview-based work the researcher is a participant in the interaction, not a neutral collector of it, so extracts should normally include the surrounding turns rather than presenting the participant's words in isolation. Presenting a monologue where there was a conversation removes the context the analysis depends on.
One extract demonstrates a possibility. The finding is that this construction recurs: that agency is routinely removed when constraint is discussed, across speakers and contexts. Show two or three further instances, and note any case where a speaker does the opposite — deviant cases sharpen the claim rather than weakening it.
Practicalities
Transcription and data selection
| Level | Retains | Use for |
|---|---|---|
| Orthographic | Words only | Thematic work; too coarse for most DA |
| Jefferson-lite | Pauses, emphasis, audible breath, overlaps | Discursive psychology, most interview-based DA |
| Full Jefferson | Timed pauses, intonation, volume, pace | Conversation analysis |
Transcription is slow: a Jefferson-lite transcript takes roughly six to eight hours per hour of audio, and full Jefferson considerably longer. Budget for it explicitly. It is also analytic work rather than clerical work — decisions about what to represent are decisions about what counts as data, which is why outsourcing transcription is a poor fit for this method.
Ethical considerations differ slightly from other qualitative work. Because extracts are reproduced at length and analysed in fine detail, anonymisation is harder: a distinctive phrase or a specific institutional reference can identify a speaker within a small professional community even when the name is removed. Plan for this at consent, be explicit that verbatim extracts will be published, and consider whether any detail in a quotation needs altering — noting where you have done so, since altering the words is not a neutral act in a method that analyses them.
How much data
Discourse analysis works on much smaller corpora than thematic analysis, because the analysis is far more intensive. Six to ten interviews is a common range for a doctoral study, and published papers frequently work from a handful of extracts examined in great depth. Volume is not the currency here; analytic depth is.
Interviews are interactions with a researcher, and much of what you analyse will be the participant managing that encounter. If your question concerns how something is talked about in practice — meetings, consultations, policy documents, online forums — naturally occurring data removes that layer and is generally preferable where ethics allow.
Quality
Rigour without reliability coefficients
Inter-rater reliability makes no sense here: the analysis is interpretive, and two analysts producing identical readings would not demonstrate anything. Rigour is established differently.
Reflexivity carries more weight here than in most methods, because your presence is part of what produced the data. A statement that names your position, your relationship to participants and what that plausibly made speakable or unspeakable is not a formality — it bears directly on how the extracts should be read.
Every claim should be traceable to something observable in the text — a word choice, a structure, a placement. “The speaker feels defensive” is a claim about a mental state you cannot access. “The disclaimer structure pre-empts a challenge that has not yet been made” is a claim about the text, and a reader can check it.
Send your transcripts and analysis so far. A named researcher reviews the analytic claims against the extracts and advises on evidencing the pattern across your corpus.
See qualitative codingPitfalls
Six mistakes examiners look for
1. Thematic analysis with a different label
If the findings are “participants felt X”, the analysis has not made the shift to language as action.
2. Not naming the approach
Discursive psychology, Foucauldian and CDA make different assumptions and answer different questions. Name yours and cite it.
3. Cleaning the transcript
Hedges, repairs and pauses are frequently where the analysis lives. Removing them removes the data.
4. Claiming things about speakers' minds
The method examines talk, not cognition. Anchor claims in features of the text.
5. Building a finding on one quotation
A single extract shows a possibility. Evidence the pattern, and address cases that do not fit.
6. Reporting inter-rater reliability
It misrepresents what the method claims. Use the rigour criteria that fit interpretive work.
Reporting
Writing it up
The findings section will look markedly different from a thematic one. Expect fewer, longer extracts with detailed commentary attached to specific lines, rather than short illustrative quotations grouped under theme headings. If your findings chapter reads like a list of themes with quotes beneath them, the analysis has probably not become discursive.
Expect the write-up to take longer than a thematic one of equivalent length, because the commentary must be built around specific lines rather than summarised. Allow for it in your timetable.
“Across the corpus, constraint was routinely constructed through nominalisation, with funding, capacity and demand appearing as states of the world rather than as outcomes of decisions. This construction positions speakers as responding to circumstance rather than exercising discretion, and makes the decisions difficult to contest without appearing unrealistic. Two speakers departed from this pattern; both held budgetary responsibility.”
Answers
Frequently asked questions
What is discourse analysis in simple terms?
An approach that studies what people are doing with language rather than what their words reveal about their thoughts. It asks what a particular way of phrasing something accomplishes in context — justifying, deflecting, building consensus, removing agency — and treats the talk itself as the object of analysis.
What is the difference between discourse analysis and thematic analysis?
Thematic analysis treats talk as evidence of experiences or views and produces themes. Discourse analysis treats talk as social action and produces accounts of what particular ways of speaking accomplish. In thematic work an extract illustrates a theme; in discourse analysis the extract is the analysis and cannot be swapped for another.
Which type of discourse analysis should I use?
It depends on your question. Discursive psychology suits how psychological states are constructed in interaction; Foucauldian analysis suits how discourses shape what can be said, often using documents; critical discourse analysis suits questions about power and ideology; conversation analysis suits the fine sequential organisation of naturally occurring talk.
How much data do I need for discourse analysis?
Far less than for thematic analysis, because the analysis is much more intensive. Six to ten interviews is common for a doctoral study, and published papers often work from a handful of extracts in great depth. Analytic depth rather than corpus size is what carries the work.
Do I need detailed transcription?
For most interview-based discourse analysis, a Jefferson-lite convention retaining pauses, emphasis and overlaps is sufficient. Conversation analysis requires full Jefferson notation. Orthographic transcription of words alone is generally too coarse, because it removes features the analysis depends on.
How do I establish rigour without inter-rater reliability?
By presenting enough extract for readers to evaluate your reading, demonstrating patterns across the corpus rather than relying on a single quotation, attending to deviant cases, grounding every claim in observable features of the text, and being reflexive about your own role in generating the data.
Can I use discourse analysis on documents or social media?
Yes, and naturally occurring data is often preferable to interviews, since interview talk includes the participant managing the research encounter itself. Policy documents, meeting transcripts, forum threads and media coverage are all standard sources. Check the ethical position on public online data with your committee.
Is critical discourse analysis political?
Explicitly so. It assumes language reproduces power relations and sets out to make that visible. This is a legitimate position provided it is declared as the framework rather than presented as a neutral finding that emerged from the data.
Send the data. Get a fixed quote.
Attach your dataset or just describe the project. A named statistician replies with a price and a deadline, usually within one working day.