Thematic analysis: a complete guide, with a worked example

Thematic analysis is a method for identifying, analysing and reporting patterns of meaning — themes — across a qualitative dataset. You read the data closely, label segments of it with codes, group those codes, and develop the groups into themes that answer your research question.

It was formalised for the social sciences by Virginia Braun and Victoria Clarke in 2006, in a paper that has since become one of the most cited in qualitative research. Their central argument was that thematic analysis was already being used everywhere but rarely described properly, so it looked like an absence of method rather than a method in its own right.

Three properties distinguish it from the methods it is most often confused with.

Thematic analysis is flexible, which is its strength and its trap. Because it can be applied to almost any qualitative dataset, it is often chosen by default rather than because it fits the question.

The honest test: if you cannot say what thematic analysis gives you that another method would not, you have probably chosen it because it is familiar. That is a defensible reason to change method at the design stage and an uncomfortable one to defend at a viva.

The six phases are widely quoted and widely misunderstood. They are not a linear pipeline: phases four and five routinely send you back to phase two, and Braun and Clarke are explicit that the process is recursive.

Phase 1 — Familiarisation

Read the entire dataset before coding anything. If you transcribed the interviews yourself, you have already begun; if you did not, read each transcript at least once without a highlighter in your hand. Take notes on early impressions, but resist turning them into codes — the point of this phase is to know the shape of the whole dataset before you start fragmenting it.

Practical marker: you should be able to describe what each interview was broadly about from memory before you code.

Phase 2 — Generating initial codes

Work systematically through the dataset, labelling segments that are relevant to your research question. Codes are short and descriptive: felt dismissed, waiting without information, relied on other parents. Code generously at this stage — it is far easier to merge codes later than to notice a pattern you never labelled.

Two rules save the most time. First, code the extract with enough surrounding text that it still makes sense when read in isolation, because in phase four you will read hundreds of extracts stripped of context. Second, allow the same extract to carry more than one code; real speech rarely does one thing at a time.

Phase 3 — Constructing candidate themes

Group codes that seem to speak to the same underlying idea. A theme is not a bucket of similar codes — it is a claim about what those codes mean taken together. If you can only describe a theme by listing its codes, it is a category, not yet a theme.

Phase 4 — Reviewing themes

This is the phase most often skipped, and the one that separates a defensible analysis from a plausible one. Review at two levels: first against the coded extracts, then against the entire dataset.

Phase 5 — Defining and naming

Write a short definition of each theme — two or three sentences saying what it captures and, importantly, what it excludes. If you cannot write the boundary, the theme is not yet defined. Names should be informative rather than clever; a reader should know what a theme is about from its name alone.

Phase 6 — Writing up

The write-up is analysis, not reporting. Extracts illustrate a claim you are making; they do not make it for you. A results section that is mostly quotations with connecting sentences is a common and heavily penalised pattern.

Consider a study interviewing parents about accessing a support service. Below are six extracts, the codes applied to them, and the two candidate themes they were grouped into. The dataset in the real study was 22 interviews; this is a fragment.

Why these codes and not others

Take the extract “I found out from another parent”. It could have been coded word of mouth, which is descriptive and accurate. It was coded informal information instead, because the analytic interest was in how information reached parents at all, not in the specific channel. That decision is defensible either way — what matters is that it was made deliberately and recorded.

This is what a codebook entry looks like for one of these codes:

From codes to a theme

The three codes no obvious route in, unexplained process and informal information were grouped as Navigating in the dark. The name makes a claim: that these parents were not merely uninformed but were required to find their own way through a system that assumed knowledge they did not have.

That claim is what makes it a theme. Had it been named “information” it would have been a category — a place to put things rather than something the analysis says.

What phase 4 changed

On review against the full dataset, a third candidate theme — “waiting” — was dissolved. Its codes distributed between the two themes above: waiting without information belonged to Navigating in the dark, and waiting while being treated as a nuisance belonged to Having to fight to matter. The waiting itself turned out not to be what participants were talking about.

A separate decision from which variant of thematic analysis you use is where your codes come from. Most methods chapters claim one and describe the other.

In practice almost every real analysis is hybrid. You begin with sensitising concepts from your literature, code openly, and find that some prior categories survive while others do not. That is a perfectly defensible approach — but it needs saying, because a chapter claiming purely inductive coding while using a framework lifted from a published model is an easy criticism to make.

Where a study is deductive, name the framework, cite it, and say what you did with codes that did not fit it. Data that falls outside an a priori framework is often the most interesting finding in the dataset, and discarding it silently is the single biggest weakness of deductive coding done badly.

Braun and Clarke have since been explicit that “thematic analysis” names a family of methods with genuinely different assumptions. Saying which one you used is now expected, and mixing them is the most common methodological criticism in review.

Six of these seven are write-up problems rather than analysis problems, which is worth noticing: most people do more analytic work than their methods chapter gives them credit for.

What belongs in the methods chapter

What belongs in the results

One subsection per theme is conventional. Open each with the claim the theme makes, develop it across two or three paragraphs, and use extracts to evidence specific points rather than to carry the argument. Attribute extracts consistently (P07, or a pseudonym) and keep them short — a half-page quotation is almost always doing less work than three lines would.

A thematic map is optional but often earns its place, particularly where themes relate to each other rather than sitting in parallel. If you include one, it must match the themes as written; a map showing four themes above a results section describing three is a preventable inconsistency.

Advice to “make the claim, then evidence it” is easier to give than to follow. Below is the opening of the Navigating in the dark theme from the worked example, written the way it would appear in a results chapter, with the moves annotated.

The written version

“Participants consistently described the service as something they had to find rather than something offered to them. This was not simply a matter of missing information: the process assumed a level of prior knowledge that most parents did not have, and the gap was filled informally or not at all. As P07 put it, ‘I found out from another parent at the school gate, nobody had told me it existed.’ Where information did arrive through official channels it was often unusable — P12 described a form that ‘might as well have been in another language’. The result was a period, sometimes months long, in which parents were nominally eligible for support but practically unable to reach it. Two participants who had navigated the system successfully both attributed this to prior professional experience of similar services rather than to anything the service itself had done.”

What that paragraph is doing

How much extract to include

The two quotations above total nineteen words. That is deliberate. Long extracts feel like evidence but usually contain one usable clause surrounded by context the reader has to do the work of discarding. If an extract needs more than about thirty words, consider whether you are quoting because it proves the point or because you found it striking.

A practical ratio: across a results chapter, extracts should occupy perhaps a quarter of the words. Much more than that and the chapter is a collection; much less and the claims are unevidenced.

Braun and Clarke published a fifteen-point checklist alongside the original paper. The version below is adapted for what UK examiners and reviewers actually query, grouped by phase.

Thematic analysis does not require software. It requires organisation, and software provides it once a dataset passes roughly fifteen interviews.

Whatever you use, the requirement is the same: at the end you should be able to retrieve every extract behind every theme in seconds. If you cannot, you will not be able to answer the question an examiner is most likely to ask.