Inductive vs deductive coding: what the labels decide and what they don't

Scope: coding interview and focus group data across the main qualitative traditions. All example data are invented.

Short answer. The difference is where your codes or categories come from. In inductive coding they are developed from the data during analysis. In deductive coding they are derived beforehand from a theory, a framework, earlier research or a client brief, and are then used to read the data. Both are interpretive: deciding that a passage fits a predefined category is a judgement, just as naming a new pattern is. Most applied projects combine the two. A hybrid works only if you record which codes came from where.

The label answers one question only. It does not tell you how open the codes are to revision, which analytic tradition you are working in, or whether intercoder agreement makes sense. Those are separate decisions, covered below.

The difference at a glance

These are orientations on a continuum, not two methods. Braun and Clarke (2006) describe thematic analysis as more inductive (data-driven) or more theoretical (analyst-driven). Hsieh and Shannon (2005) separate conventional content analysis, where categories come from the text, from directed content analysis, which starts from theory. Elo and Kyngäs (2008) set out an inductive and a deductive route through content analysis. Some researchers add a third orientation, abduction, in which surprising findings are worked through against existing theory to produce new explanations (Timmermans & Tavory, 2012).

Four decisions the label does not make

1. Origin. Where do the categories come from: the data, prior theory, both (hybrid), or a back-and-forth between data and theory (abduction)?
2. Openness. Once you have categories, are they locked (applied unchanged, with uncoded text reported as a residual), extendable (new categories may be added from uncoded text), or iteratively revised throughout?
3. Analytic family. Is the coding reflexive (the analyst's interpretation is the resource, with no fixed codebook), codebook-based (a structured but revisable frame shared by a team), or coding-reliability oriented (a fixed frame applied consistently by several coders)? Braun et al. (2019) distinguish these three families of thematic analysis, and each has its own quality criteria.
4. The operation at hand. At any moment, are you generating an interpretation (naming what is going on), indexing (marking what a passage is about), or assigning a passage to a defined category?

You can combine these freely. A reflexive thematic analysis can be quite deductive, read through a theoretical lens, and still have no codebook. A coding frame can be developed inductively and then fixed and applied by several coders. Directed content analysis starts deductively but extends the frame from text the theory does not cover. Asking all four questions is more useful than asking "inductive or deductive?" alone.

How the traditions use inductive and deductive codes

Note what the table does not say. Reflexive thematic analysis is not grounded-theory open coding, and directed content analysis is not the same as applying a closed list.

One passage, two readings

Invented example. Rovina Regional Hospital (fictional) pilots a new rota on three wards: twelve-hour shifts and a shared team of nurses who move between wards. Two focus groups follow: FG-A with eight ward nurses (N1–N8) and FG-B with six ward managers (W1–W6).

[A-13] N3: On the noticeboard. One day it was just up there. Nobody asked us,
           and then they were surprised that we swap shifts among ourselves.
[A-14] N5: It suits me, actually. I have a small child, and twelve-hour shifts
           mean more days off. If only we knew more than a week ahead.
[A-16] N3: […] Someone from the shared team comes in, doesn't know the
           patients, and I have to show her everything anyway. So I'm doing
           two jobs.
[A-17] N6: Mhm. That's exactly it.

[B-32] W4: It's gone down on the spreadsheet. When they swap a shift between
           themselves, nobody records it as overtime.

Data-led reading

First-pass codes stay close to what people do, a habit taken from grounded theory's initial coding (Charmaz, 2024): learning from the noticeboard, not being asked, swapping shifts among ourselves (A-13); gaining days off, needing to know ahead (A-14); inducting a colleague, doing two jobs (A-16); agreeing (A-17).

Memo: "Swapping shifts" comes straight after "Nobody asked us". Is it resistance or a practical fix? What else do nurses do to make the rota work?

Then FG-B: W4 describes the same swaps from the manager's side (B-32). Comparing the two suggests a new code, reworking the rota where management cannot see it, and the analyst goes back to A-13. A-16 shares a property with it: it is work the rota does not count. A candidate concept takes shape: invisible work that keeps the rota running. Nothing in the brief asked for it. It came from comparison, and it changed how earlier passages were read.

Theory-led reading with a codebook

Now suppose the brief asks which job demands and resources participants link to the rota. The codebook follows the job demands–resources model (Bakker & Demerouti, 2007):

What happens next depends on openness. In a locked application, informal swapping stays in the residual and is reported as something the theory did not capture. In directed content analysis, the uncoded passages are examined and may become a new subcategory, marked as derived from the data. The swaps might be explored as job crafting, employees' own changes to their work (Tims & Bakker, 2010), if further context shows how nurses actively reshape their work; that concept was absent from this initial codebook. In either case the residual is a finding to analyse, not a bin.

The two readings answer different questions. The deductive reading makes demands and resources comparable across groups and against a theory. The inductive reading found a concept that links the groups but that the codebook did not name.

The hybrid approach, done properly

Fereday and Muir-Cochrane (2006) show a staged hybrid: theory-driven codes are set out first, data-driven codes are added alongside, and themes are built from both. Hybrid designs are common in applied work, and they carry a specific risk: something found in the data gets presented later as if it had been expected, or an assumption gets presented as a finding. To guard against that:

• give every code an origin: a priori, from the data, or revised (an a priori code whose definition the data changed);
• date changes to definitions and say what was recoded;
• in the report, keep "the theory or brief expected this, and the data support it / do not support it" apart from "this came from the data".

Where intercoder agreement fits

Agreement depends on the analytic family, not on whether codes are inductive or deductive. It belongs to coding-reliability approaches and is often useful in team codebook work (MacQueen et al., 1998), and its use is debated (O'Connor & Joffe, 2020). In reflexive thematic analysis, agreement scores do not fit the method's logic (Braun & Clarke, 2021). An inductively developed frame can be checked for agreement once it is fixed. A theoretically oriented reflexive analysis would not use agreement at all.

Common mistakes

• Calling a topic summary "inductive". Grouping passages under the interview questions describes the data. The result is not a set of themes (Braun & Clarke, 2021).
• Saying codes "emerged". The analyst makes the codes, so say how.
• Labelling first and thinking later. In inductive work the thinking happens while you code. Codes applied without memos or comparison give you a topic index, not an analysis.
• Stretching a category so text fits, or drifting so that the end of the dataset is coded differently from the start.
• Counting agreement as new evidence. In focus groups, "Mhm, exactly" shows alignment with an earlier turn but is not a separate account; see how to analyse focus group data.

Which orientation should you start from?

When this distinction is not the main one

If your question is about how people talk, for example how a group negotiates blame or how a story is built, the choice of code origin matters less than working with sequence and language. Conversation, discourse and narrative analysis have their own procedures. SurveysAI Qualitative Analysis Workflow does not offer them.

Inductive and deductive analyses in the app

SurveysAI Qualitative Analysis Workflow offers one analysis for each orientation.

• AI-assisted inductive thematic analysis: the model proposes and iteratively refines working codes across the dataset and drafts candidate themes. Choose I will decide at key points (the researcher interprets) to set the analytic orientation yourself and decide on the themes and their definitions. The report states who made the analytic decisions and does not claim the full procedure of reflexive thematic analysis.
• Deductive qualitative content analysis: you supply the codebook. Only my codebook applies your categories unchanged and reports uncovered text as a residual. My codebook plus new categories from the data lets the analysis add categories marked as derived from the data.

In both, the app produces draft analytic outputs and program code checks quotations word for word against the transcripts. You remain responsible for reviewing, revising and owning the interpretation. The app does not produce intercoder agreement.

• Create an account and run an inductive analysis. If the analysis is not preselected, choose AI-assisted inductive thematic analysis in the start dialog.
• Create an account and run a deductive analysis. If the analysis is not preselected, choose Deductive qualitative content analysis in the start dialog.

References

• Bakker, A. B., & Demerouti, E. (2007). The Job Demands-Resources model: State of the art. Journal of Managerial Psychology, 22(3), 309–328. https://doi.org/10.1108/02683940710733115
• Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa
• Braun, V., & Clarke, V. (2021). One size fits all? What counts as quality practice in (reflexive) thematic analysis? Qualitative Research in Psychology, 18(3), 328–352. https://doi.org/10.1080/14780887.2020.1769238
• Braun, V., Clarke, V., Hayfield, N., & Terry, G. (2019). Thematic analysis. In P. Liamputtong (Ed.), Handbook of Research Methods in Health Social Sciences (pp. 843–860). Springer. https://doi.org/10.1007/978-981-10-5251-4_103
• Brooks, J., McCluskey, S., Turley, E., & King, N. (2015). The utility of template analysis in qualitative psychology research. Qualitative Research in Psychology, 12(2), 202–222. https://doi.org/10.1080/14780887.2014.955224
• Charmaz, K. (2024). Constructing grounded theory (3rd ed.). SAGE.
• Elo, S., & Kyngäs, H. (2008). The qualitative content analysis process. Journal of Advanced Nursing, 62(1), 107–115. https://doi.org/10.1111/j.1365-2648.2007.04569.x
• Fereday, J., & Muir-Cochrane, E. (2006). Demonstrating rigor using thematic analysis: A hybrid approach of inductive and deductive coding and theme development. International Journal of Qualitative Methods, 5(1), 80–92. https://doi.org/10.1177/160940690600500107
• Gale, N. K., Heath, G., Cameron, E., Rashid, S., & Redwood, S. (2013). Using the Framework Method for the analysis of qualitative data in multi-disciplinary health research. BMC Medical Research Methodology, 13, 117. https://doi.org/10.1186/1471-2288-13-117
• Hsieh, H.-F., & Shannon, S. E. (2005). Three approaches to qualitative content analysis. Qualitative Health Research, 15(9), 1277–1288. https://doi.org/10.1177/1049732305276687
• MacQueen, K. M., McLellan, E., Kay, K., & Milstein, B. (1998). Codebook development for team-based qualitative analysis. Cultural Anthropology Methods, 10(2), 31–36. https://doi.org/10.1177/1525822X980100020301
• O'Connor, C., & Joffe, H. (2020). Intercoder reliability in qualitative research: Debates and practical guidelines. International Journal of Qualitative Methods, 19. https://doi.org/10.1177/1609406919899220
• Tims, M., & Bakker, A. B. (2010). Job crafting: Towards a new model of individual job redesign. SA Journal of Industrial Psychology, 36(2), Art. 841. https://doi.org/10.4102/sajip.v36i2.841
• Timmermans, S., & Tavory, I. (2012). Theory construction in qualitative research: From grounded theory to abductive analysis. Sociological Theory, 30(3), 167–186. https://doi.org/10.1177/0735275112457914