Qualitative Content Analysis: What It Is, When to Use It, and How to Do It

What qualitative content analysis is

Content analysis started life as a counting method. Early twentieth century researchers tallied words in newspapers and propaganda, and a quantitative version of it is still used in media studies. The qualitative version keeps the discipline of the original (explicit categories, written rules, a clear trail) but cares about what the words mean in context rather than how often they appear.

The definition most dissertations cite comes from Hsieh and Shannon (2005): a research method for the subjective interpretation of the content of text data through the systematic classification process of coding and identifying themes or patterns. Notice the two words pulling against each other there. Subjective, because you are interpreting. Systematic, because you do it by rules that someone else could check. Mayring (2000) leans harder on the systematic side and describes step-by-step models and rules for assigning text to categories, applied without rushing to quantify.

One distinction you will be asked about in a viva is manifest versus latent content. Manifest content is what the text plainly says. Latent content is what it implies: the hesitation before an answer, the joke that covers a complaint, the topic a participant keeps steering away from. Elo and Kyngäs (2008) ask you to decide before you start whether you will analyse only the manifest content or the latent content too, and to say so in your write-up. Latent analysis gets you further but invites over-interpretation, so it needs more evidence per claim.

The output is a category system. Each category has a name, a definition, the codes inside it, and example quotes. Depending on the aim, the categories may sit in a hierarchy (main category, generic categories, sub-categories) that you can draw as a tree. Our thematic map generator draws that kind of hierarchy for you as a figure.

The three approaches: conventional, directed and summative

Hsieh and Shannon's 2005 paper in Qualitative Health Research is the reason most people now talk about three types. The difference between them is where your codes come from. To make that concrete, imagine the same study run three ways: twelve interviews with international students about their first term at a UK university.

Conventional content analysis suits a topic where theory is thin. You read everything several times, highlight the exact words that seem to carry key ideas, label them, and let the labels settle into categories. The advantage, in Hsieh and Shannon's words, is getting information from participants without imposing preconceived categories. Any theory you know about goes in the discussion chapter, where you compare your categories against it.

Directed content analysis is for extending or testing something that already exists. You turn the theory into a coding scheme with an operational definition for each category, then code the data against it. Hsieh and Shannon describe two ways to start: highlight everything that looks relevant first and code it afterwards (safer, you miss less), or code straight into the predefined categories and park whatever doesn't fit for later. Elo and Kyngäs call the coding scheme a categorisation matrix, which can be structured (only the theory's categories) or unconstrained (the theory's categories plus new ones built inductively).

Summative content analysis begins with counting and then moves to interpretation. If you stopped at the counts it would be a quantitative study. It becomes qualitative when you go back into the text and ask why a word is used, by whom and in what circumstances. Hsieh and Shannon's own illustration is how often clinicians, patients and families say “death” and “dying” versus euphemisms like “passing.”

In student projects the conventional approach is by far the most common. If you aren't sure which one you are doing, you are probably doing conventional content analysis. The same data-first versus theory-first choice runs through every qualitative method, and our guide to inductive vs deductive coding explains how to decide.

When to use it (and when not to)

Content analysis is a good fit when your research question starts with “what”: what do nurses report about night shifts, what concerns do parents raise in open survey comments, what do these policy documents say about inclusion. It handles large amounts of text well and gives a clear, describable answer. It is common in nursing, health, education and communication research, partly because examiners in those fields value its step-by-step transparency.

Reach for it when:

• You want to describe a phenomenon, not build a theory of it.
• You have a lot of fairly short text, such as open-ended survey answers, where deep interpretation of each response isn't possible anyway.
• You want to report how widespread each category is across participants or documents.
• You have an existing framework to test or extend (the directed approach).
• You are reviewing literature and want to categorise what studies have found (see synthesising findings across studies).

Pick something else when your question is about process or explanation (“how does trust develop between…”), which is grounded theory territory, or about lived experience and meaning, which points to IPA or phenomenology. Hsieh and Shannon are direct about this: the most a conventional content analysis delivers is concept development or a model, not a theory. If you are still choosing, our overview of qualitative data analysis methods compares the main options.

The comparison you will most likely have to defend is against thematic analysis. Vaismoradi, Turunen and Bondas (2013) note that the two share a lot and are often confused. The practical difference is this. Content analysis stays closer to the surface of the text, uses tightly defined categories, and is comfortable reporting frequencies. Braun and Clarke's reflexive thematic analysis treats themes as the researcher's interpretation of shared meaning and does not use counts as evidence that a theme matters. If you want the longer version, our guide to reflexive TA versus other qualitative methods goes further.

The process, step by step

Elo and Kyngäs (2008) organise both the inductive and the deductive routes into three phases: preparation, organising and reporting. That structure is the one most methods sections borrow, and it is the one used here. The steps below are for the conventional (inductive) route. Differences for the directed route are noted where they matter.

Phase 1: Preparation

1. Choose your unit of analysis. This is the chunk of text you treat as one thing: a whole interview, one answer, a paragraph, a sentence. Elo et al. (2014) put the trade-off neatly. Too broad and a unit carries several meanings at once. Too narrow and the text fragments into pieces that no longer mean anything. For interviews, most students treat each transcript as the unit of analysis and code at the level of meaning units inside it.

2. Decide on manifest, latent, or both. Write the decision down. It changes what counts as evidence later.

3. Get a sense of the whole. Read every transcript through, once without a pen. Elo and Kyngäs suggest keeping a few questions in mind as you read: who is telling this, where and when it happened, what is happening, and why. Only then start marking.

Phase 2: Organising

4. Open coding. Go through the text and write notes and headings in the margin for everything relevant to your question. Graneheim and Lundman (2004) give a useful sequence here. Lift out a meaning unit (the words that belong together), shorten it into a condensed meaning unit that keeps the core, and give it a code. Hsieh and Shannon suggest open coding three or four transcripts, settling on preliminary codes, then coding the rest (and recoding the first ones) with new codes added whenever something doesn't fit. A code labels one idea in the text; a category groups many codes. The code vs theme guide covers the same distinction in thematic analysis, and how to code an interview transcript covers code types and codebooks.

5. Group codes into categories. Put codes that belong together into sub-categories. A category is a group of content that shares something. The point is to decide which codes belong together and which don't, so each category should be internally consistent and clearly different from its neighbours.

6. Abstract. Keep grouping upward: sub-categories into generic categories, generic categories into main categories. Name each level using content-characteristic words. Stop when further merging would lose something that matters to your question. Elo and Kyngäs admit that part of this step rests on the researcher's insight and is hard to describe, which is exactly why you keep notes as you go.

7. Define every category. Write a definition and pick one or two anchor quotes for each. Mayring calls these anchor examples. They are what a second coder would use to check whether a new passage belongs.

Directed route: steps 4 to 7 run the other way round. You build the categorisation matrix from the theory first, write the definitions and coding rules, pre-test the matrix on a transcript or two, then code. Data that won't fit gets abstracted inductively into new categories, and those new categories are often the most interesting part of the findings.

Phase 3: Reporting

8. Report the categories and how you got them. Present the category system (a table or tree diagram works well), describe what each category contains, and support it with quotes. Then describe the analysis itself in enough detail that a reader could judge it. Elo et al. (2014) stress that readers judge the trustworthiness of a content analysis by how clearly the path from data to categories is shown, so this part carries real weight.

A worked example: from quote to category

Back to the twelve international students, analysed with a conventional approach and manifest content only. The research question is: What challenges do international students describe in their first term? Here is how four excerpts move through the steps.

After coding all twelve transcripts the student has 41 codes. Grouping and abstraction then goes like this:

The definition written for the second main category might read: “Learning the unwritten rules: accounts of expectations that staff or home students treated as obvious but that were never stated, covering contact with staff, assessment language and study habits. Excludes difficulties with English itself, which are coded under Language.” That last sentence is a coding rule. It tells a second coder where the boundary is, and it is the kind of detail that makes content analysis checkable.

Notice what the analysis doesn't do. It doesn't explain why money and hidden rules combine to push some students to the edge of university life, or build a model of how belonging develops. Those would be natural questions for the discussion chapter, or for a follow-up study using a different method.

Should you count?

Reporting that a category appeared in 9 of 12 interviews is normal in content analysis and often expected. It helps a reader see whether a category describes most of the sample or a vocal few. Hsieh and Shannon suggest rank-ordering code frequencies in directed studies rather than running statistical tests, because qualitative samples are rarely designed for them.

Two cautions. First, count participants, not mentions. One person who talks about rent eleven times is still one person. Second, a rare category can matter more than a common one. If one student describes being racially abused on a night bus, that belongs in your findings even though the count is one.

Limitations you should name in your thesis

Examiners like to see that you know where your method is weak. The literature is fairly specific about content analysis:

• It describes more than it explains. Content analysis has no built-in technique for connecting concepts to each other (Elo & Kyngäs, 2008). You get categories, and perhaps a hierarchy, but not a causal or process account.
• Conventional analysis can miss the context. Hsieh and Shannon warn that without a full understanding of the context you may fail to spot key categories, producing findings that don't represent the data. They also note it is easily confused with grounded theory or phenomenology, so say clearly which one you are doing.
• Directed analysis leans toward confirmation. Starting from a theory is an informed bias. You are more likely to find support for it than evidence against it, and interview probes built from the theory can cue participants to agree. An audit trail and a second reviewer of your definitions help.
• Summative analysis can stay shallow. Word counts show usage, not meaning, unless you do the interpretive work that follows.
• Abstraction often stops too early. Elo et al. (2014) observe that a long list of overlapping categories usually means the grouping is unfinished. The results then read like a list of what participants said instead of an analysis of it.
• Latent content invites over-reading. Pauses, laughter and silences are data, but the further you move from the words, the more evidence you need for each claim.
• It looks easy and isn't. Elo and Kyngäs point out that its reputation as a simple method leads researchers to underestimate it. The mechanics are simple. Getting clean, non-overlapping categories out of messy text takes several rounds.

Most of these are handled the same way: define categories tightly, keep an audit trail of your decisions, have someone check a sample of your coding against your definitions, and show your working in the write-up. Elo et al. (2014) include a checklist of questions for each phase that is worth running through before submission.

Getting the first pass done faster

Look back at the worked example. The judgement calls (what counts as a meaning unit, where one category ends and another begins, whether “money worries” is really about money) are yours and should stay yours. The slow part is everything around them: reading twelve transcripts line by line, writing out 41 codes, copying quotes into a spreadsheet, and counting which participant said what. That can take weeks, and it is usually the part where students run out of time.

thematicanalysis.ai runs a qualitative content analysis framework built on Hsieh & Shannon and Elo & Kyngäs. You paste or upload your interview transcripts, survey answers or published papers and it does the conventional route: open codes grounded in the text, grouped into categories, with the verbatim quote behind every code and a note of how many sources support each category. You then read it the way a careful supervisor would, merge, split and rename, and export it as a Word report. On a paid pass you can also export to NVivo or MAXQDA and carry on coding there.

Try it on one transcript and see whether the categories it finds match the ones you would have found. The first three sources are free.

Writing it up in your methods section

Name the approach, cite the process you followed, state your unit of analysis and whether you analysed latent content, and say how you checked the analysis. Something like this, adapted to your study:

Data were analysed using conventional qualitative content analysis (Hsieh & Shannon, 2005), following the preparation, organising and reporting phases described by Elo and Kyngäs (2008). Each transcript was treated as a unit of analysis, and analysis focused on manifest content. Transcripts were read in full several times before open coding. Meaning units were condensed and coded (Graneheim & Lundman, 2004), and codes were grouped into sub-categories and abstracted into main categories. Each category was given a written definition, coding rules and anchor quotes. A second researcher coded a 20% sample against the category definitions, and disagreements were resolved through discussion.

If you used software or AI for any part of the coding, say which part and how you checked it. Our guide on reporting AI-assisted analysis has wording you can adapt. For how the methods section fits with the findings chapter, see qualitative analysis for a dissertation, and if your analysis includes latent content, a short positionality statement helps examiners see where your interpretation comes from.

Frequently asked questions