How to analyse qualitative interview data
By Hey Juno · Published · Updated
Qualitative data analysis is the work of turning interview transcripts into findings you can defend. In practice you read closely, label what people said, group those labels into themes, test each theme against the evidence that does not fit, and write the result as claims that point back to quotes. The hard part is not finding themes. It is stopping a plausible theme from becoming a conclusion before it has been challenged.
This guide explains thematic analysis in plain terms, then covers coding, dissent, patterns versus single stories, counting people, and writing findings. It also describes how Juno, the AI researcher from Hey Juno, handles analysis.
Thematic analysis in plain terms
Thematic analysis is the most widely used way to analyse interview transcripts. Braun and Clarke's 2006 paper in Qualitative Research in Psychology set out a clear version of it. They refined it in a 2019 paper on what they call reflexive thematic analysis. The steps, simplified:
-
Read everything first. Get familiar with the whole set before you decide what matters.
-
Code. Label the parts of the transcripts that seem relevant.
-
Build themes. Group related codes into candidate patterns.
-
Review the themes. Check them against the coded material and against the full set of transcripts.
-
Define and name them. Say what each theme is about, and what it is not.
-
Write up. Tell the story of the data, with evidence.
The steps are a loop, not a line. You will go back and re-code once a theme looks wrong. Braun and Clarke also stress that themes do not simply "emerge" from the data. Someone interprets them, so a good analysis states its choices and stays open to revision.
Coding qualitative data
A code is a short label for something a participant said or did. "Called back within the hour" might be coded as responsiveness. "Never heard from the other site" might be coded as follow-up gap. Codes stay close to the words. Themes are broader and carry an idea, such as "speed of response is how people judge whether a provider cares".
Two habits make coding more reliable:
-
Code everything relevant, not only what fits. If you tag only the passages that support your hunch, the codes will confirm it.
-
Keep the code tied to its source. You should be able to go from any code back to the exact passage, with enough surrounding conversation to read it fairly.
Illustrative example: in a study of how people choose a dentist, "asked a friend" is a code. A theme might be "people borrow trust from someone they already know".
Looking for disconfirming cases and dissent
Once a theme forms, everything after it starts to look like support. This is confirmation, and it is the usual failure of qualitative synthesis. Fluent writing hides it, because a weak theme and a strong one read equally well.
The countermeasure is to ask what would have to be true for the theme to be wrong, and then look for it on purpose.
-
Search for exceptions. Who did something different, and why?
-
Treat dissent as information. A segment that behaves differently is not an inconvenience. It often marks where the theme stops being true.
-
Let dissent change the claim. The test is not whether disagreement is mentioned. It is whether the finding got narrower or more precise because of it.
Agreement deserves suspicion too. If everyone seems to say the same thing, check whether the question invited it, or whether grouping answers together hid real differences. Our guide to leading questions covers how the questions themselves can plant agreement.
One story or a pattern
A vivid story can dominate a report. The person told it well, it is quotable, and it feels like it explains something. But one story is one story.
To separate a pattern from an anecdote, look at the conditions underneath. What happened, in what circumstances, and would those circumstances recur for other people? A repeatable pattern has a mechanism you can describe. An anecdote has a plot.
Present single stories as single stories. An edge case can still show what the typical case hides.
How many people sit behind a theme
Qualitative analysis is not statistics, and a count of people does not give you a percentage you can project to a population. But the number still matters, because it tells readers how much weight a theme can carry.
Some practical guidance:
-
State how many participants support each theme. "Most" and "many" are vague. Name the number of interviews, and say how many were included in total.
-
Count people, not mentions. One person repeating a point five times is still one person. A topic everyone mentions may be shallow, and one raised by a few may bear directly on the decision.
Guest, Bunce and Johnson's 2006 study in Field Methods analysed 60 interviews with women in Africa and looked at when new themes stopped appearing. It does not settle how many interviews your question needs, which depends on how varied your participants are.
Writing findings as claims with evidence
A finding is stronger when it is written as a claim that can be checked, not as a topic heading. "Pricing" is a topic. "Customers who cancel in the first month describe the price as unclear, not too high" is a claim.
A useful claim has four parts:
-
The statement, in plain words.
-
The evidence: the participants and quotes behind it.
-
The reasoning: why that evidence supports the statement.
-
The limits: who it does not cover, what is unknown, and how confident you are and why.
Match your wording to the evidence. "Customers want" and "clearly" promise strength. If the evidence is a handful of interviews, write "several participants described" instead. Provisional findings that state their limits beat confident ones that fall apart under a question. For a checklist on testing someone else's findings, see how to check AI research findings.
Two habits guard against confirmation. Write down what you expect to find before you read, so you can see where you were surprised. And ask someone who did not build the analysis to try to break it.
How Juno analyses interviews
Juno interprets rather than counts. It does not tally how often words come up. It organises what participants said around what matters for the study's objectives, and points out absences when they mean something.
-
It reads every interview included in the analysis. You can see which interviews sit behind each theme, so a vivid quote cannot stand in for a pattern that is not there.
-
It reads and codes every interview, so a theme can be traced to where it came from.
-
It looks for what breaks a theme. Juno seeks disagreement, alternative explanations and negative cases, and can narrow or revise a finding when dissent changes how far it can be defended.
-
It states each finding as a claim with its reasoning and evidence, linked to the participant, the quote and the transcript moment, and to the audio where available. Report quotes are checked against the original transcripts.
-
It names its gaps. When evidence is thin, Juno says so, limits how far the finding can travel and says what evidence would help.
Afterwards you debrief with Juno as you would with a person. You can ask what it found, who disagreed and what did not hold up, and challenge a claim until it re-examines the evidence. More detail is on the reveal page and in the methodology standards.
Where this doesn't apply
These methods suit interview and open-ended text data. If you need to estimate how common something is across a population, you need a properly sampled survey, and qualitative themes should not be turned into percentages. Thematic analysis is also one approach among several, and others such as grounded theory suit some projects better.
Frequently asked questions
What is the difference between coding and theming?
Coding labels specific passages with short descriptions. Theming groups related codes into broader patterns that carry an idea. Coding stays close to the words, and themes involve interpretation.
How many interviews do I need before I can trust a theme?
There is no universal number. It depends on how varied your participants are and what the decision needs. Report how many people support each theme, and be cautious where the count is small.
Can AI do qualitative data analysis well?
It can read and code large volumes quickly. Whether the result is sound depends on whether it looks for disconfirming evidence, shows its reasoning and traces claims to quotes. Judge it on those points, not on how fluent the summary is.
Should I report percentages from interviews?
Usually not. Qualitative samples are not designed to estimate how common something is across a population. Report the number of participants and describe the pattern, with its limits.