Conventional Qualitative Content Analysis: categories from the data

Hsieh and Shannon's Three Approaches to Qualitative Content Analysis is one of the most cited methods papers in health research, and its central contribution is a distinction people routinely collapse.

Conventional content analysis derives coding categories directly from the data. It is used when existing theory or literature on the phenomenon is limited, and it avoids preconceived categories on purpose.

Directed content analysis starts from existing theory or prior research, which supplies initial codes that the analysis then extends or challenges. It is used to validate or extend a theoretical framework.

Summative content analysis counts occurrences of words or content, then interprets the underlying context of their use. It is the closest of the three to quantitative content analysis and the furthest from thematic work.

Saying which one you are doing is the first requirement of reporting. A paper that cites Hsieh and Shannon without specifying the approach has not told the reader what was done.

Alongside Hsieh and Shannon, the other standard reference is Elo and Kyngäs's The qualitative content analysis process, which sets out the inductive and deductive routes in more procedural detail and is the paper most nursing studies follow.

Elo and colleagues later addressed the field's trustworthiness problem in Qualitative Content Analysis: A Focus on Trustworthiness, which is the more useful of the two for anyone preparing a study for review — it identifies the specific reporting failures that recur.

The method's roots are in mid-twentieth-century communication research, which is why its vocabulary (units of analysis, coding units, categories) reads more systematically than the interpretive traditions.

A study of how patients experienced a newly introduced remote-monitoring service used conventional content analysis, appropriately — nothing had been published on this service and there was no framework to test.

The unit of analysis was the meaning unit: a passage expressing one idea, which could run from a clause to a paragraph. That choice was stated, because a sentence-level unit would have fragmented accounts that only made sense across several sentences.

Open coding produced 187 codes, many in participants' own words: "the machine tells them before I do", "phoning about a number", "no one to ask about the number". Grouping produced subcategories — anticipated contact, unexplained readings, absent interpretation — and then three categories: being monitored rather than cared for, data without meaning, and the disappearance of the appointment.

The analysis stopped there, and stopping was the right call. The categories describe the experience clearly and are directly usable by the service. What the study did not do was claim an interpretive account of what remote monitoring means for the patient relationship — that would have required a different method and a different design, and asserting it on this analysis would have overreached.

The reported scheme gave each category a definition, its subcategories, and two example meaning units, which is what allows a reader to judge whether the categories hold.

• Not naming the approach. Citing Hsieh and Shannon without saying conventional, directed or summative leaves the method unspecified.
• Unstated unit of analysis. It determines what the codes can be, and it is omitted more often than not.
• Categories that are topics from the interview guide. If the categories reproduce the questions, the data has been sorted rather than analysed.
• Silent drift into directed analysis. Starting from literature-derived codes while claiming a conventional approach misdescribes the study.
• Frequency as finding. Counting is summative content analysis; presenting counts as the result of a conventional analysis conflates the two.
• Claiming interpretive depth. Conventional content analysis is descriptive by design, and a discussion section that reads as latent interpretation needs the analysis to have supported it.
• No trustworthiness account. Elo and colleagues specify what to report; most studies still do not.

The expected reporting is more explicit than in interpretive traditions: state the approach, the unit of analysis, whether latent content was included, how categories were formed, and how the scheme was checked against the data.

A category table with definitions and exemplar units is the standard evidence, and its absence is a reasonable reason for a reviewer to ask for revision. Where multiple coders were used, say what their role was — consistency checking is coherent within this method in a way it is not within reflexive TA.

The method is descriptive and does not generate theory. Studies wanting theory should use grounded theory; studies wanting interpretive depth should use thematic or phenomenological approaches, and choosing content analysis for its procedural comfort while wanting those outputs is the field's recurring mistake.

It is also weak on context. Segmenting into meaning units and grouping them across participants loses the shape of the individual account, and there is no case axis to recover it from.

And its apparent objectivity is partly illusory. Deriving categories "from the data" still requires judgement at every grouping, and the procedural vocabulary can obscure how much interpretation went into a category scheme.

Directed content analysis is the deductive sibling, starting from theory-derived codes.

Summative content analysis counts and then interprets the context of use.

Framework analysis offers a comparable level of structure with a case-by-theme matrix that preserves individual accounts.

Reflexive thematic analysis is the right choice where interpretive depth rather than description is the goal.

The organisation phase is hierarchical grouping of a large code set, revised repeatedly — 187 codes into subcategories into categories, with the whole scheme checked back against the material each time it changes. That is mechanical work that a tool absorbs entirely.

The reporting requirement helps too: a category table with definitions and exemplar units is a by-product of coding in a platform and a separate chore otherwise. Evidano supports conventional qualitative content analysis as a named methodology and derives categories from the material rather than applying a stored framework. Choosing the unit of analysis, and deciding whether the analysis stops at description, remain the researcher's calls — and the second one is where most studies go wrong.