Thematic Coding Explained (With Examples)

A thematic code is a short label saying what a chunk of your data is about. Thematic coding is the pass where you attach those labels across the whole dataset, so that patterns become findable at all.

It is one phase of thematic analysis rather than another name for it, and that distinction sounds pedantic until you see what it costs. Most of the trouble I see in qualitative chapters starts with a coding pass that was really a summarizing pass, or with a list of codes presented as findings. Both are cheap to fix if you catch them early and expensive if you don’t.

Here I’ll set out what actually counts as a code, code a real extract, walk the four levels from data point to theme, and deal with the question that decides how you write the whole thing up: whether themes emerge.

Notice how little the code adds to the quote, and how much the theme does. That gap is where the interpretation lives, and it’s the part you have to defend.

---

Thematic coding vs thematic analysis

The two get used interchangeably, including by people who should know better, and the confusion costs students marks. Here’s the difference.

	Thematic coding	Thematic analysis
What it is	Labeling segments of data	The whole analytic process
Scope	One phase	Six phases, coding among them
What you produce	Coded extracts, and a record of what each code means	Named themes and a written account of them
When it happens	Early, and again on later passes	Across the whole project

Coding is phase two of six in Braun and Clarke’s framework, between familiarizing yourself with the data and searching for themes; National University’s guide to thematic data analysis sets the six out in order. So thematic coding is a thing you do; thematic analysis is the method you’re doing it inside.

You can also code thematically without running a full thematic analysis, which is common in mixed methods work where the qualitative strand is a supporting one.

If it’s the whole method you need rather than this one step, start with thematic analysis. If you’re weighing thematic coding against the other techniques instead, our qualitative data coding guide covers the main ones.

What counts as a code

A code is a short label for what a piece of data is about. That is the whole definition, and almost every coding problem I see comes from quietly replacing it with something else.

Four things a code is not:

• Not a summary of the quote. If your code is nearly as long as the extract, you have paraphrased rather than coded. “Participant explains that she finds out her shifts late and cannot plan around them” is a sentence. “Late shift notification” is a code.
• Not a research question. Coding your data by the questions you asked is a legitimate technique, but it’s structural coding and it does a different job. A thematic code names content, not location.
• Not a theme. Themes come later and are far fewer. A code applied to nine extracts is a code; the claim you eventually build across several codes is the theme.
• Not a demographic. Age, role and site are attributes of the case, and they belong in a column of their own rather than in your code list, where a label applied to a third of your data says nothing about content. That holds even when your question is about one of them: if you’re comparing how part-time and full-time staff experience something, employment status is the axis you compare codes across, not a code. What someone says about their age or role is a different matter and does get coded, so “I’m too old for these shifts” earns something like Age as a barrier.

“Isn’t a code just a summary of the quote?”

No, and the test is portability. A good code can be applied to a second extract that uses completely different words. A summary can only ever describe the one passage it came from, which is why a summarizing pass produces as many codes as extracts and no patterns at all.

Codes also don’t have to be short or to use the participant’s exact words. Using their words is in vivo coding, a specific technique with its own tradeoffs, not a rule you’re breaking by writing your own label.

---

Coding a real extract

Here is a stretch of an interview from a study of hospitality workers and shift scheduling.

Interview extract	Code
“My shifts land on a Thursday for the following week.”	Late shift notification
“You can’t book anything. My sister’s birthday, I just said I’d let her know.”	Declining social plans
“If you say no to a shift once, you notice you get offered fewer.”	Fear of being passed over
“The managers are fine, honestly. It’s the system that’s the problem.”	Separating people from process
“I’ve got a second job, so two rotas have to line up.”	Juggling two employers

Notice that none of those codes is a summary, and none is a theme. Each names what its extract is about, in words that could apply to somebody else’s extract about the same thing. That portability is what makes the next step possible, and there are worked examples of other techniques applied to the same kind of data if you want to compare.

Notice too the fourth one. “Separating people from process” is doing more interpretive work than the others, and that’s worth flagging in your notes rather than pretending it isn’t. A code that required you to read between the lines is a code you will have to justify.

---

How to code, step by step

1. Read a transcript through without coding anything. You’re getting the shape of it. Coding on a first read produces codes for the first ten pages and fatigue for the rest.
2. Code a second time, one extract at a time. Keep the extract beside the code, always: a code with no evidence attached is unusable when you write up and indefensible if anyone asks.
3. Keep an alternative-code column. When two labels both fit, record the one you didn’t use. It costs a second, and when the same choice recurs on transcript eleven you’ll want to know what you decided on transcript two.
4. Write your code definitions as you go, not afterwards. Each one needs the code, a line saying when it applies, and an example extract. Definitions written at the end are reconstructions, and they show. Whether that record has to become a formal codebook depends on which approach you’re following, which the last section covers.
5. Recode the earlier transcripts once the list stabilizes. Your codes at transcript twelve are better than your codes at transcript one, and the earlier cases deserve the better version.

There’s no single right number of codes to end up with, but there is a wrong feeling: if the list has grown past what you can hold in your head and every new transcript adds five more, read our guide on what to do with too many codes.

---

From codes to categories to themes

This is the part clients most often skip, and skipping it is why so many results chapters read as a list of topics.

The four levels, in order:

• Data point. The extract itself, in the participant’s words.
• Code. What that extract is about. Descriptive, portable, close to the data.
• Category. A group of codes that belong together. Still descriptive, just tidier.
• Theme. A claim you are making about the data, built from categories.

The jump that matters is the last one, because it changes what kind of statement you’re making. A category names a topic; a theme says something about that topic that a reader could disagree with.

That last jump is the one people find hardest to defend, and it raises the question the rest of your methodology chapter hangs on. If you built that sentence rather than found it, does anything in a thematic analysis “emerge” at all?

---

Do themes actually “emerge”?

No. And this matters more than it sounds, because the answer decides what belongs in your methodology chapter.

Braun and Clarke, whose 2006 paper is the standard citation for the method, are blunt about it. In their 2022 paper on good practice they write that in reflexive thematic analysis, themes “are generated, created or constructed (for example), they are not identified, found or discovered, and they definitely don’t just ‘emerge’ from data”. The objection is not really about style. The passive phrasing hides the researcher, which makes an active choice sound like something that happened to you.

That’s not just a wording preference. Thematic analysis is a family of approaches, not one method, and they disagree about what coding is:

Approach	Coding is	So it uses
Coding reliability	Measurement, to be done consistently	A fixed codebook, several coders, an agreement statistic
Codebook	Structured but interpretive	A structured codebook, developed early and applied consistently
Reflexive	An interpretive act by a situated researcher	No agreement statistic, and a reflexivity trail instead

One wrinkle worth knowing if you go to the source: Braun and Clarke’s own typology has a fourth member, which they call thematic coding. In their usage it names a whole approach rather than the step this post describes. The step is the common meaning, and it is the one you will meet everywhere outside their typology.

“My advisor asked for an inter-rater reliability score.”

Then find out which approach your study is claiming, because an agreement statistic and a reflexive analysis do not belong together. Braun and Clarke call consensus coding and measuring intercoder agreement “conceptually incoherent additions” to a reflexive analysis: if meaning is constructed by the researcher, a second researcher agreeing with you measures similar training, not truth.

The practical version: cite the approach you actually followed. A chapter that cites Braun and Clarke for the six phases and then reports a Cohen’s kappa has described two incompatible studies, and that is the kind of thing an examiner notices and asks about. If your advisor (supervisor, in the UK) wants the reliability figure, use a coding-reliability or codebook approach and cite accordingly. Either is defensible. The mixture is not.

Still have questions?

Do I need a codebook for thematic analysis?

It depends which version you’re doing, which is why the approach question above is worth settling. Coding-reliability and codebook approaches require one by definition. Reflexive thematic analysis doesn’t, though keeping a record of your code definitions is still sensible, because you’ll need to explain your decisions whether or not the method obliges you to.

Should I report an inter-rater reliability score?

Only if your approach calls for one. A kappa or percentage-agreement figure belongs in a coding-reliability study and sits awkwardly in a reflexive one, where the researcher’s interpretation is the instrument rather than a source of error. If a second coder is involved for developmental reasons, describe what the discussion changed instead of scoring the agreement.

How is thematic coding different from content analysis?

Thematic coding looks for patterns of meaning and is usually interpretive. Qualitative content analysis is more systematic and often counts: how often a category appears, across how many cases. The two overlap in practice, and the honest way to choose is to ask whether frequency is part of your claim.

How do I know when to stop coding?

When new transcripts stop producing new codes, and the ones you’re adding are variations on labels you already have. That point is usually earlier than people expect. Stopping is a judgment call rather than a threshold, so record when it happened and what the last few transcripts added, because that record is what makes the decision defensible later.

Can’t find your answer here? Ask a Grad Coach directly – the initial chat is free.

Speak with a friendly coach →