Thematic Analysis in Qualitative Research: A Step-by-Step Guide to Coding and Themes

Thematic analysis is a family of approaches for identifying and interpreting patterns of meaning across qualitative data. Those patterns need to matter in relation to the research question.

Suppose you interview postgraduate students about receiving feedback from supervisors. During coding, you might identify unclear comments, fear of asking questions, repeated revisions, waiting for supervisor replies and learning through trial and error. These are useful codes, but they are not necessarily themes.

A theme should say something more substantial about the pattern connecting the data. For example:

Students have to decode feedback before they can act on it.

That theme brings several observations together around a central analytical idea: feedback is not automatically usable simply because it has been given. Codes capture more specific ideas within segments of data; themes explain a broader pattern organised around a central concept.

A common problem is that researchers start coding immediately and only later try to work out what methodological approach they used. It is better to make the main analytical choices earlier, because those choices shape what you pay attention to and how you interpret the data.

Are you working inductively or deductively?

An inductive analysis is primarily driven by patterns developed through engagement with the dataset. A deductive analysis is more explicitly guided by an existing question, theory or framework. These are not absolute opposites. A study can be largely deductive while still allowing unexpected patterns to appear.

If you analyse health-worker interviews using the Consolidated Framework for Implementation Research, much of your coding may be deductive because the framework gives you concepts of interest. If your study is exploratory and you deliberately avoid imposing a pre-existing coding frame, the analysis may be more inductive.

The useful question is not, “Which one sounds more qualitative?” It is: what was driving my analytical attention?

Are you coding semantic or latent meaning?

Semantic coding stays relatively close to what participants explicitly say. Latent analysis moves further into underlying assumptions, meanings or social processes.

Consider the statement:

“I don’t tell my partner when I go for the injection because he doesn’t agree with family planning.”

A semantic code might be concealing contraceptive use from partner. A more latent interpretation might be contraceptive decision-making constrained by relationship power.

Both may be legitimate. They answer slightly different analytical questions.

What version of thematic analysis are you using?

If you are using reflexive thematic analysis, say so and use procedures that make sense within that approach. Do not automatically add intercoder agreement, consensus coding or a reliability statistic simply because you think qualitative research requires it.

In reflexive thematic analysis, researcher interpretation is part of the analytical process. Treating coding differences as errors that must be eliminated can conflict with that position. Other approaches to thematic analysis may legitimately use structured codebooks, multiple coders and agreement procedures. The key is methodological coherence.

Do not begin analysis by immediately attaching labels to every sentence. Read the transcripts, then read them again. Listen to recordings where useful. Notice contradictions, repeated ideas, silences, tensions and unexpected explanations. Record questions that occur to you.

At this stage, you might write an observation such as:

Several participants describe receiving feedback but still not knowing what action to take.

Or:

Participants rarely describe openly refusing partner preferences. They instead describe delay, secrecy and negotiation.

These are not final themes. They are analytical observations that help you begin seeing the dataset as a whole.

The goal of familiarisation is to understand individual excerpts in context rather than treating each quotation as an isolated unit.

A code identifies something analytically interesting in a segment of data.

Consider this fictional interview excerpt:

“My supervisor wrote ‘clarify this section’. I knew something was wrong, but I didn’t know whether she meant the argument, the writing or the methods. I changed the whole section and then she told me that wasn’t what she meant.”

Possible codes include:

• ambiguous supervisory feedback
• uncertainty about expected revision
• substantial work created by unclear feedback
• interpreting comments through trial and error

Good codes help you retrieve analytically related material later. Weak coding often produces labels that are too broad, such as supervision, feedback or problems. Those labels identify topics but say very little about what is happening.

Compare partner issues with anticipating partner opposition. The second begins to capture a process.

Do not build a major interpretation from one memorable quotation and then search only for material that confirms it. Work systematically across the dataset.

At the same time, coding does not require every line to receive a code. What matters is that relevant material is examined consistently in relation to the research question.

A segment can receive multiple codes. For example:

“I stopped the pills because I was bleeding. The nurse told me it was normal, but I couldn’t keep going to work like that.”

Possible codes include:

• side effects affecting daily life
• provider reassurance not resolving concern
• discontinuation despite counselling

The same excerpt contributes to several dimensions of the participant’s experience.

This is where thematic analysis becomes analysis.

Lay out your codes and ask which ones appear to describe the same underlying process. Imagine you have:

• ambiguous supervisory comments
• fear of asking for clarification
• guessing what the supervisor means
• repeated revisions
• learning expectations over time
• asking peers to interpret feedback

A weak theme would be Supervisor feedback. That is only the topic.

A stronger candidate might be:

Feedback creates a second task: learning how to interpret the supervisor.

Now the codes are connected by an analytical proposition.

This distinction is critical. Themes should not merely act as filing cabinets.

At this stage, ask of each candidate theme:

• Does it have a clear central idea?
• Does the data within the theme belong together?
• Is it meaningfully distinct from the other themes?
• Does it answer the research question?
• Is there enough evidence across the dataset to support it?

Suppose you initially create:

• Theme 1: Poor communication
• Theme 2: Supervisor expectations
• Theme 3: Revision difficulties

On inspection, the three may be different versions of the same underlying issue. You might reorganise them into:

Students first have to decode what “good work” looks like.

Possible subthemes might include interpreting brief comments, learning expectations through revision and using peers to translate feedback.

This produces a more coherent analytical story.

Go back to the extracts, then go back to the full dataset. Ask whether your theme exaggerates consistency.

Perhaps six participants described avoiding clarification because they feared appearing incompetent, while others actively challenged feedback and negotiated revisions. Do not erase that difference. It may become an important variation within the theme, a subtheme, a boundary condition or evidence that the original theme needs refining.

Strong qualitative analysis is not produced by making everybody sound the same.

You should be able to finish this sentence:

This theme shows that...

If you cannot, your theme may still be a topic.

Consider:

Theme title: Managing contraception in the shadow of partner opposition

Definition:
This theme shows how women adapted contraceptive decisions in anticipation of partners’ reactions, using secrecy, delay and selective disclosure to preserve access while limiting relationship conflict.

That definition tells the reader what the theme actually argues.

A theme title does not need to be theatrical, but it should be informative.

Compare:

Barriers

with:

Access depended on being able to return without being noticed

Or:

Side effects

with:

Side effects became an access problem when switching support was difficult to obtain

The second version makes the analytical contribution visible.

A weak qualitative findings section often follows this structure:

Participants reported side effects. Participant 3 said... Participant 7 said... Participant 11 said...

The quotations are doing almost all the work.

Instead, start with the analytical claim:

Side effects did not automatically lead to discontinuation. Discontinuation became more likely when women could not obtain timely reassurance, switching or removal support. Several participants initially attempted to continue their method despite bleeding or pain, but persistence became difficult when symptoms interfered with work or daily responsibilities.

The quotation can then illustrate or complicate the claim.

The researcher remains responsible for the analysis.

Research question: How do women navigate difficulties while using contraception?

Transcript:

“I wanted to change because the bleeding had continued, but when I went back they told me to wait. I waited another month. Eventually I just stopped going.”

Initial codes:

• attempting to switch
• request for help deferred
• continued side effects
• withdrawal from service after unmet request

Code cluster: Difficulty obtaining method change

Weak candidate theme: Poor side-effect management

Stronger analytical theme:

Continuation depended on whether services could support change, not only whether a method was available.

Why is the final theme stronger? Because it connects the individual experience to a broader mechanism. Availability at initiation is not sufficient. Meaningful access also depends on whether a client can return, switch or discontinue with support when needs change.

That is an analytical finding.