Axial Coding in Grounded Theory: Relating Categories After Open Coding

The commonest failure in axial coding is producing a tidier code list instead of a set of relationships. If the output of your axial coding phase is a code tree — parent categories with child codes nested underneath — you have done code management, not axial coding. Axial coding asks a different question. Not what belongs with what, but under what conditions does this happen, what do people do about it, and what follows from what they do. The output of real axial coding is not a list. It is a set of statements you could argue with.

That distinction is the whole page. Everything below — which version of the paradigm to cite, how to run the procedure on one category, why Glaser called it forcing, when to skip it entirely — follows from it. For the surrounding sequence that axial coding sits inside, see CASRAI’s guide to grounded theory as a method; for the definitional criteria, the grounded theory dictionary entry.

Axial coding relates categories; it does not tidy them

The term comes from the image of an axis. Anselm Strauss described axial coding as analysis done “around one category at a time” — that category is the axis, and the work consists of establishing what surrounds it. In Qualitative Analysis for Social Scientists (1987) he put it as “intense analysis done around one category at a time in terms of the paradigm items” (p. 32). One category. Not the whole scheme at once.

Two things are worth noticing about the origin, because both are routinely lost in the textbook version.

• The paradigm was introduced as scaffolding for beginners. Strauss (1987) observed that novices had trouble generating genuine categories — “the common tendency is simply to take a bit of the data (a phrase or sentence or paragraph)” and stop there — and offered the coding paradigm as a device “especially helpful to beginning analysts” (p. 27). It was a remedy for a specific novice error, not a mandatory apparatus.
• It is a repair operation. Open coding deliberately breaks the data apart, line by line, to force the analyst off the participants’ own framing. That leaves you with fragments. Axial coding is what puts them back together along analytic lines rather than along the lines of the original transcript.

So the sequence is not small codes → medium codes → big codes. It is fracture → reassemble → integrate. If your axial phase only moved you up a level of abstraction, the reassembly step has not happened.

The paradigm: three versions, and why the difference matters for your methods section

“The paradigm model” is not one fixed thing. It appears in at least three forms across the Strauss and Corbin literature, and reviewers who learned one version will read your methods section against that version. Name the one you used.

Strauss (1987) — four items

Conditions; interaction among the actors; strategies and tactics; consequences (p. 27). Compact, and closest to the pragmatist theory of action underneath it.

Strauss and Corbin (1990) — six elements

The version most commonly taught. As Kelle (2005) enumerates it, categories from open coding are examined for whether they relate to: (1) the phenomenon at which action and interaction are directed; (2) causal conditions leading to that phenomenon; (3) attributes of the context; (4) intervening conditions that influence it; (5) action and interactional strategies actors use to handle it; and (6) the consequences of those actions. Strauss and Corbin described this model as a way “to think systematically about data and to relate them in very complex ways” (1990, p. 99).

Later editions — a reduced three-part form

Later editions of Basics of Qualitative Research present the paradigm in a stripped-down form: conditions, actions–interactions, and consequences or outcomes. This is how most current published studies actually apply it. Fekonja and colleagues’ 2024 grounded theory study of emergency triage nurses, for instance, describes using “the paradigm analysis tool and its three characteristics (conditions, actions-interactions and consequences)” to develop categories and identify the central phenomenon.

Practical implication: if you write “axial coding was conducted following Strauss and Corbin” and cite the fourth edition, but your results table has columns for context and intervening conditions, you have mixed editions without saying so. Cite the edition, state the elements you used, and say plainly if you dropped any. That single sentence prevents most methods-section queries on this topic.

A working procedure for one category

Axial coding is done category by category, not scheme by scheme. A workable loop for a single category:

1. Choose the axis. Start with the category carrying the most coded data and the most variation across participants — not the one you find most interesting. Variation is what makes relationships visible.
2. Specify properties and dimensions before relationships. A category like withholding concerns from a supervisor has properties (who is withheld from, what is withheld, for how long) and each property has a dimensional range (occasionally ↔ routinely; minor detail ↔ safety-relevant). Skipping this step is why so many axial models are unfalsifiable: without dimensions you cannot say the relationship is stronger under some conditions than others, only that it exists.
3. Work the paradigm as questions, not as boxes. When does this occur and when does it not? What is going on around it that is not causing it but is shaping it? What do people actually do? What happens next, for them and for others?
4. Go back to the data for each answer. Every clause in your emerging statement should have coded excerpts behind it. A clause with no excerpts is a hypothesis you have imported, and it should be marked as such in a memo rather than quietly promoted.
5. Hunt the negative case. Find the participant for whom the relationship does not hold. Either the statement gains a condition (“except where the supervisor had previously acted on a concern”) or it was never a relationship in the first place.
6. Write it as a memo, in sentences. The memo, not the diagram, is the deliverable. Diagrams hide missing logic; sentences expose it.
7. Let it drive sampling. An axial statement with a thin condition tells you exactly who to recruit or which document to read next. Axial coding that does not change what you collect next is being run as a filing exercise.

Note that steps 1–7 are not a phase with a start and end date. They interleave with open coding and with selective coding, and the same category will pass through them more than once as new data arrives.

The test that separates axial coding from code-sorting

One check settles it. Write your result as a sentence and see whether it can be contradicted.

• Code-sorting output: “Communication breakdown” with sub-codes “withholding information”, “unclear escalation route”, “fear of blame”. This is a nested list. Nothing in it can be false.
• Axial output: “Where staff have previously raised a concern and seen no visible response (condition), they stop volunteering information (action), which removes the supervisor’s early warning of problems (consequence) — except where an escalation route bypasses the immediate supervisor (intervening condition).” This can be wrong, and your data can show it is wrong.

Apply the same test to a diagram: if every arrow can be replaced by the words “is a type of” without loss, the diagram is a taxonomy. Three further tells that a phase labelled axial coding was actually code-sorting: the number of categories fell but nothing gained a condition; no participant contradicted anything; and nothing about the analysis changed what was collected next.

The forcing critique, and the part of it that is actually actionable

Axial coding is the specific target of the best-known dispute in qualitative methodology. Barney Glaser’s Emergence vs. Forcing: Basics of Grounded Theory Analysis (1992) argued that by using concepts such as axial coding and the coding paradigm, researchers would force categories onto the data rather than allowing them to emerge. Kendall (1999), writing in Western Journal of Nursing Research, examined both approaches side by side using data on families raising children with ADHD, and concluded that researchers need not treat one as superior — the choice can follow the goal of the study.

Kelle’s (2005) analysis is the most useful thing a practitioner can read on this, because it separates the part of the critique that is overblown from the part that is real:

• The forcing risk is limited, and for a specific reason. The paradigm terms — conditions, actions, consequences — “carry only limited empirical content”, so “the risk is not very high that data are forced by its application”. A framework so contentless that almost any social phenomenon can be described with it cannot easily bend the data; that is precisely what a sensitising concept is for.
• The real cost is a hidden theoretical commitment. The paradigm is “linked to a certain micro-sociological perspective” — a general model of action rooted in pragmatist and interactionist social theory. Kelle notes that researchers wanting a macro-sociological or systems perspective may find the paradigm steers them away from it. The paradigm does not distort your data so much as quietly decide that your explanation will be about actors, intentions and strategies.

The actionable mitigations Kelle offers are concrete: deliberately apply different and even competing theoretical perspectives to the same data, and pay explicit attention to whether the chosen framework is excluding phenomena in the data. If your study is about resource allocation, regulation, or institutional structure, run that check before committing — an actor-centred paradigm can make a structural explanation invisible rather than refuted.

When not to do axial coding

Axial coding is a Straussian procedure, and two of the three major grounded theory traditions do without it.

• Glaserian / classic GT substitutes theoretical coding using “coding families” — an extended list of theoretical terms (Glaser 1978) that the analyst draws on ad hoc rather than applying systematically. Kelle frames the whole controversy as exactly this choice: a single well-defined paradigm applied systematically, versus a large fund of coding families applied opportunistically.
• Constructivist GT treats axial coding as optional and often declines it. King and colleagues (2025), reporting a constructivist study of paramedic-led care, state that they deliberately rejected axial coding, following Charmaz in preferring theory construction that happens organically “rather than offering a procedural step-by-step method”; their sequence was line-by-line initial coding on the first 11 interviews, then focused coding from interview 12 onward, with continuous memoing.

Declining axial coding is a legitimate, citable choice — but it has to be a choice, stated and justified, not a gap. And if your study is not building theory at all, the honest answer is that you do not need axial coding: descriptive work is better served by a named descriptive method such as framework analysis, whose matrix is a summarising device rather than a claim about relationships.

Doing it in CAQDAS without mistaking a hierarchy for a relationship

Software is where the code-tree confusion becomes structural, because every major package makes nesting easy and relating harder. Two rules:

Nesting a code under a parent is not a relationship. It records membership. To express “A is a cause of B” you need an explicit typed link between codes, which is a different feature in every package.

Use typed links and say which type you used. ATLAS.ti, for example, ships six default relations for linking codes in networks: is associated with and contradicts (symmetric, drawn with arrowheads at both ends), and is part of, is cause of, is a and is property of (asymmetric, drawn with a single arrow to the target). Those defaults can be substituted, modified or supplemented with user-defined relations — and defining relations that match your paradigm elements is usually worth the five minutes. The exact feature names and defaults differ by package and version, so check your own tool’s current documentation rather than assuming parity.

Whatever the tool, the memo is still the artefact that carries the argument. A network view is a summary of reasoning done elsewhere.

What published studies report — realistic numbers

There is no correct number of categories, and any source that gives you one is guessing. What is available is calibration from published studies that reported their counts:

• Yu and colleagues (2024, Heliyon), studying knowledge anxiety among researchers, extracted 66 initial concepts in open coding, synthesised these into 24 categories in axial coding, and arrived at 6 core categories. Notably, they organised the final model with a stimulus–organism–response structure rather than the classic paradigm — a reminder that the relating framework can be substantive rather than generic.
• Fekonja and colleagues (2024, BMC Nursing) reported a far more compact scheme: two main categories with four and three subcategories respectively, a single core category, and data saturation at 19 participants. Selective coding integrated the categories via a conditional matrix.

The spread between those two is the point. Category counts are driven by the grain of the research question and the heterogeneity of the sample, not by a methodological rule, and reviewers should be reading your relationships rather than counting your boxes. For worked examples of the underlying coding mechanics on real transcript excerpts, see CASRAI’s guide to coding qualitative interview data.

Reporting axial coding so a reviewer can follow it

A methods paragraph that survives review typically contains all six of these:

1. Which grounded theory tradition you followed, and the specific edition cited.
2. Which paradigm elements you used — and which you dropped, with a reason.
3. That coding was iterative and interleaved with collection, if it was; if axial coding genuinely ran as a discrete pass, say so, because that is a design choice with consequences.
4. At least one relationship stated in full, with its conditions, so the reader can see the form your results take.
5. How negative cases were sought and what happened to them.
6. How saturation (or theoretical sufficiency, if you prefer that framing) was judged — against categories, or against relationships, which is a stricter and more defensible bar.

Frequently asked questions

Is axial coding the same as developing themes in thematic analysis?

No, and conflating them is the source of most weak axial coding. Theme development groups codes by shared meaning; the output is a set of patterns. Axial coding specifies conditional and consequential relationships between categories; the output is a set of contestable statements. A theme answers “what is in the data”. An axial statement answers “what accounts for what, and when”. If you are doing thematic analysis, you do not need axial coding, and borrowing the label does not upgrade the analysis.

Do I have to use the paradigm model?

No. Strauss (1987) offered it as a device “especially helpful to beginning analysts”, not a requirement, and Kelle (2005) points out that its terms carry little empirical content precisely so that they can serve as a heuristic. Some studies substitute a substantive framework — Yu et al. (2024) used a stimulus–organism–response structure. What you cannot do is skip the relating work itself and still call the result grounded theory.

What is the difference between axial coding and selective coding?

Axial coding works around one category at a time, establishing its conditions, actions and consequences. Selective coding works on the scheme as a whole: it identifies or refines the core category and integrates the others around it into a single explanatory account. In Fekonja et al. (2024), for example, selective coding is the step where all categories are integrated into one central phenomenon via a conditional matrix. In practice the two overlap heavily; the boundary is analytic, not chronological.

How many open codes should become how many categories?

There is no defensible ratio. Published studies vary widely — 66 concepts to 24 categories in one study cited above, two main categories in another. The number that matters is how many of your relationships have conditions attached and survived a search for negative cases.

Can I use axial coding outside grounded theory?

People do, and the term has drifted into general qualitative use as a label for second-cycle coding. It is defensible if you say what you mean by it and do not imply you produced a grounded theory. It is not defensible as a synonym for “grouping codes into categories” — that is category development, and it already has a name.

Does axial coding happen once, after all the data is collected?

It should not. Grounded theory interleaves collection and analysis, and an axial statement with an under-specified condition is one of the strongest signals for what to sample next. A study that collected everything first and coded afterwards can still be good qualitative research, but it has given up theoretical sampling, and the methods section should say so rather than describe a loop that did not occur.

References

• Kelle, U. (2005). “Emergence” vs. “Forcing” of Empirical Data? A Crucial Problem of “Grounded Theory” Reconsidered. Forum Qualitative Sozialforschung / Forum: Qualitative Social Research, 6(2), Art. 27. https://www.qualitative-research.net/index.php/fqs/article/view/467 — source for the 1987 and 1990 paradigm formulations, the quoted passages from Strauss (1987, pp. 27–32) and Strauss & Corbin (1990, p. 99), Glaser’s (1992) forcing charge, coding families, and the limited-empirical-content and micro-sociological-commitment arguments.
• Kendall, J. (1999). Axial coding and the grounded theory controversy. Western Journal of Nursing Research, 21(6), 743–757. https://pubmed.ncbi.nlm.nih.gov/11512211/
• Fekonja, Z., Kmetec, S., Fekonja, U., Mlinar Reljić, N., Pajnkihar, M., & Strnad, M. (2024). Emergency triage nurses’ perceptions of caring behaviors and the safety of the patient during triage encounters: a grounded theory study. BMC Nursing, 23, 453. https://pmc.ncbi.nlm.nih.gov/articles/PMC11221186/
• Yu, Z., Ma, B., Song, L., Guo, C., & He, Q. (2024). Identification of knowledge anxiety factors among researchers based on grounded theory. Heliyon, 10(4), e25752. https://pmc.ncbi.nlm.nih.gov/articles/PMC10900946/
• King, R., Downer, T., Lord, B., Flanagan, B., & Oprescu, F. (2025). A practical example of how to apply constructivist grounded theory methodology: Exploring patient experiences during paramedic led healthcare. Research in Nursing & Health, 48(4), 508–521. https://pmc.ncbi.nlm.nih.gov/articles/PMC12217405/
• Welch, P., Plummer, D., Young, L., Quirk, F., Larkins, S., Evans, R., & Sen Gupta, T. (2017). Grounded theory — a lens to understanding clinical reasoning. MedEdPublish, 6, 2. https://pmc.ncbi.nlm.nih.gov/articles/PMC10885213/
• ATLAS.ti. About Relations, ATLAS.ti 26 Mac user manual. https://manuals.atlasti.com/Mac/en/manual/Networks/NetworksAboutRelations.html