How to Analyze Focus Group Data: A Step-by-Step Guide with a Worked Example

Why a focus group is not six interviews

Jenny Kitzinger made the case in 1994, and it still holds. The thing that separates a focus group from a one-to-one interview is the interaction between participants. People question each other, tell stories that trigger other stories, tease, disagree and change their minds. That interaction is why you chose a focus group in the first place. Her complaint was that published studies then reported focus group data as a pile of individual quotes, as if the conversation had never happened.

Wendy Duggleby (2005) asked the practical follow-up: if interaction matters, how do you analyse and report it? Her answer was that it depends on the study. Sometimes the interaction is the finding. More often it is context you need in order to read the content correctly. David Morgan (2010) pushed the point further. Interaction is what produces focus group data, he argued, but that does not mean every report has to be about interaction.

You can take a position anywhere on that spectrum. What you cannot do is ignore it. If you do, agreement gets counted as corroboration: when P2 describes a problem and P4 says “same”, you have one account and one endorsement, and counting them as two makes the idea look more widespread than it is. Disagreement also tends to vanish. Pushback shows where an idea stops being true, which makes it some of the most useful data in the transcript, and coding each speaker in isolation smooths it away.

A worked example

The example used through this guide is invented for illustration: a focus group of five final-year students about why they did or did not use the university counselling service. Here are about forty seconds of the discussion.

Moderator: What stopped people from booking an appointment?

P2: Honestly, the waiting list. I tried in October and got a slot in January.

P4: Same. Eleven weeks for me.

P1: I didn't even try, because everyone said the wait was months.

P5: I'd push back a bit. I got seen in a week, but I said it was urgent.

P3: That's the thing though. You have to know to say that. Nobody tells you.

P4: Yeah.

Coded the interview way, this excerpt produces a tidy result: four of five students mentioned long waits. That sentence is accurate and misleading at once. Look at what each person actually contributed:

Read with the stance column, the excerpt says something different. Two students waited a long time. A third never tried because of what they had heard, which is a finding about reputation, not about waiting lists. A fourth got in quickly by using a word the others didn't know mattered. And the most interesting idea in the exchange, that access depends on knowing an unwritten rule, came from P3 reacting to that disagreement. No individual interview would have produced it.

The codes are the ones you would make in an interview anyway. The stance column is the part that makes it focus group analysis.

The analysis, step by step

1. Prepare the transcripts

Every turn needs a speaker label, and each person must keep the same label for the whole session. “P3” on page one and “Sarah” on page six will be counted as two people by you and by any software you use. Label the moderator clearly (“Moderator:” or “Facilitator:”). Where it matters, note overlapping talk, laughter and long pauses in square brackets. If an assistant moderator took notes on body language or who was nodding, add those notes to the file. Krueger points out that intensity is often carried by tone of voice and is hard to see in a transcript alone.

Write a debrief straight after each session: the main points, any surprises, who dominated and who stayed quiet. Ten minutes of notes the same evening are worth more than an hour of reconstruction a month later.

2. Decide your unit of analysis

Before you code, decide whether you are analysing individuals, groups or the interaction between people. This decision changes what you count and how you write it up, so it gets its own section below.

3. Read each session as a whole

Read the whole transcript once without coding. In an interview you follow one person's account. In a focus group you are following the shape of a conversation: where it settled quickly, where it split, which questions fell flat. Write a half-page summary per session. Those summaries make the later comparison across groups much easier.

4. Code with speaker and stance

Code the content the same way you would code an interview. Our guide to coding interview transcripts covers code types, code length and codebooks, and all of it applies here. Then add two things to every coded passage: the speaker, and their stance towards what came before. Four stances cover most turns:

• Raises: introduces the idea.
• Endorses: agrees with someone, with or without their own example.
• Adds to: extends or qualifies someone else's point.
• Challenges: disagrees, or offers a counter-example.

A spreadsheet with columns for speaker, code, stance and “responding to” is enough. In NVivo or MAXQDA you can store speakers as cases and stance as a separate set of codes.

5. Weigh the evidence

Not every mention carries the same weight. P2's dated, first-hand account is stronger evidence than P1's “everyone said”. Krueger's criteria, set out below, give you a checklist for this.

6. Compare across groups

The group, not the person, is the natural unit of comparison. Draw a table with your emerging themes as rows and your groups as columns, and note in each cell whether the idea came up, how much agreement it got, and whether anyone challenged it. If you ran groups for different populations (first-years and final-years, say, or nurses and doctors), this is where differences between them show up.

7. Develop themes and report them

Group related codes into themes using whichever method you chose (see which method to use). A theme is an idea about the data, not a topic heading, and our guide to codes versus themes covers that move. Then write it up with the group context attached, as shown in the reporting section.

Choosing your unit of analysis

This is the decision examiners ask about most, and the one most methods sections skip. There are three options.

For most dissertations, group-level analysis with attention to interaction is the defensible middle. You report themes per group, you keep track of who endorsed and who challenged, and you quote short exchanges where the interaction changes the meaning. State the choice in one sentence in your methods section, for example: “The group was treated as the unit of analysis; individual contributions were coded with the speaker and their response to preceding turns, so that agreement and disagreement could be distinguished from independent accounts.”

If you want a formal tool for the interaction side, Onwuegbuzie, Dickinson, Leech and Zoran (2009) proposed what they called micro-interlocutor analysis. For each question you record which participants responded, in what order, what their responses were like and what nonverbal signals went with them. The result shows at a glance whether a theme was shared across the group or carried by the same two talkative people every time.

Weighing what was said: Krueger's criteria

Richard Krueger's guidance on focus group analysis has been the practical standard since the 1990s. Later editions written with Mary Anne Casey group the criteria under fewer headings, but the ideas are the same. His original list works well as a check on each emerging theme. Here it is, applied to the counselling example:

Krueger also warns about the opposite of frequency: notice what people did not say, or what got little attention when you expected a lot. And his advice on big ideas is practical. Step away from the data for a day before writing the summary, because the big ideas tend to appear once you stop looking at individual quotes.

Which analysis method to use

Focus groups are a way of collecting data, not a method of analysis. You still need to choose one. The common choices:

• Thematic analysis. The usual choice for student projects. Braun and Clarke's approach works on focus group transcripts without modification, as long as you add the speaker-and-stance step above.
• Framework analysis. Popular in health and policy research. You chart summaries into a matrix of cases by themes, and with focus groups the cases are usually the groups. Fatemeh Rabiee's (2004) short paper walks through framework analysis on focus group data and is a good methods-section citation.
• Qualitative content analysis. Suits descriptive questions and structured category systems. See our guide to qualitative content analysis.
• Conversation or discourse analysis. For studies where the interaction itself is the object, such as how a group builds a shared view of a risk. These need detailed transcription conventions and are rarely the right fit for an applied question.

If you are still choosing, our comparison of thematic analysis frameworks and the broader guide to qualitative data analysis methods set the options side by side.

How many focus groups do you need?

The best evidence comes from Guest, Namey and McKenna, who analysed 40 focus groups from a study of health-seeking behaviour among African American men in North Carolina. More than 80% of all the themes in the full dataset turned up within the first two to three groups, and 90% within three to six. The most common themes were all there by the third group.

Two caveats keep that from being a universal rule. It is one study, on one topic, with a fairly similar population. And it counts themes, which says nothing about how well you understand them. A reasonable reading is three to six groups for each distinct type of participant. If you are comparing two populations, plan for at least three groups of each, and expect to need more if the population is varied or the topic is sensitive.

How to report focus group findings

The reporting rules follow from everything above.

• Label quotes by participant and group, for example (P3, Group 2), so readers can see when one person is quoted repeatedly.
• Quote exchanges when the exchange is the point. A three-line sequence where one person challenges another tells the reader more than either line alone.
• Count groups, not heads. “Raised in all three groups” is meaningful. “Mentioned by 11 of 17 participants” usually is not, because the 11 heard each other.
• Report disagreement. Say where a view was contested and by whom. It strengthens the finding.

Here is how the counselling excerpt might read in a findings chapter:

Finally, add a limitation statement. Something like: “Because participants responded to and were influenced by one another, the number of speakers supporting a theme is not a count of independent respondents; contributions that only endorsed another speaker are reported as endorsements.” Our guide to thematic analysis for a dissertation has more wording for the methods and findings chapters.

Common mistakes

• Inconsistent speaker labels. One person under two labels quietly doubles their weight in every count.
• Coding the moderator. The moderator's questions are context, not data. Code what participants said in reply.
• “Most participants agreed.” Often means one person spoke and others nodded. Say who spoke and who agreed.
• Letting the loudest voice set the themes. If one participant supplies half the quotes for a theme, check whether anyone else held the view at all.
• Ignoring silence. A question that got no answer, or a topic that people steered away from, can be a finding.
• Pooling all groups into one dataset. You lose the comparison between groups, which is often the most useful part of the analysis.

Doing the first pass with software

The stance column is the step people skip, because doing it by hand for six sessions of ninety minutes each is slow. It is also the step most general-purpose AI tools get wrong. Paste a focus group into a chatbot and it will usually treat each speaker as a separate respondent and tell you that five of six people agreed.

We built focus group handling into thematicanalysis.ai for this reason. When you paste or upload a transcript with labelled speakers, it recognises the moderator, splits the discussion by participant, and marks turns that reply to someone else. Codes record whether a speaker raised an idea, endorsed someone, added to a point or pushed back. Each theme then carries a line such as “Group evidence: raised by P2; endorsed by P4; challenged by P5”, and the limitations always include a sentence about group influence, so the counts can't be read as independent agreement.

Treat it as a first pass to check against the transcript, not as the analysis. Very short agreements (“Yeah.”) are easy for any coder to miss, human or machine, so look at who is listed as endorsing each theme and correct it where needed. Every code links back to the quote it came from, and you can rename, merge or remove themes. The judgement calls in this guide, such as what counts as a challenge or which unit of analysis to report, stay yours. If you are unsure whether AI belongs in your analysis at all, our guide to AI qualitative coding sets out the debate, and reporting AI-assisted analysis has wording for your methods section.

Paste one session below to see how it handles your transcript. Keep the speaker labels in.

Frequently asked questions