Coding and the codebook: making interpretation inspectable

Unit 1 argued that a transcript is already a theory of the data, since the choices about what to record decide which claims the record can carry. This second of ten units begins where the transcript ends, with the analyst marking the text and deciding that this passage is an instance of that. Each mark is a judgement, not a filing move.

The claim here is that a code is a decision, and that the codebook makes the analyst's decisions inspectable, teachable and contestable. A coding procedure that cannot be reconstructed from its codebook has not been specified at all. A reader cannot see why a segment took its code, or what would not count under it, and has been handed a result without its warrant. The unit that follows turns that standard into a number.

Coding is a decision the analyst can be asked to defend

The word code makes the activity sound smaller than it is. Elliott (2018) argues that coding is ubiquitous in qualitative work yet under-considered in methods training and in the literature. He offers a conception of it as decision-making. Three things are choices: the density of coding, the frequency with which a code is applied, and the size of the pieces of data coded. Each follows the design and the research questions and belongs to a study rather than a rulebook, so the choices are not made once and for all. A code that cannot be defended in the terms of the study is a preference wearing a label.

Saldaña (2021) refuses to make the activity mechanical. His manual catalogues first-cycle and second-cycle coding methods. He argues that coding is cyclical and analytic rather than labelling. The analytic memo carries the analysis from code to category to theory, and there the analyst records why a pattern is being claimed. Weston and colleagues (2001) reach a compatible conclusion from a real project, tracing successive versions of a coding scheme across several years and concluding that coding is not what happens before analysis but an important part of it.

The codebook is the instrument of inspectability

That artefact is the codebook. MacQueen and colleagues (1998) describe it as the means by which a team makes its analytic categories explicit. Each code gets a definition, a boundary and example segments. Coders then apply it consistently, and someone who was not in the room can examine the interpretive basis of the analysis. Their team settled on five parts for every definition. A brief definition prompts the analyst's memory, and a full definition follows. Then come a rule for when to use the code, a rule for when not to use it, and an example. The when-not-to-use rule is where a boundary becomes checkable. It names the case that looks like the code and belongs elsewhere.

DeCuir-Gunby, Marshall and McCulloch (2011) turn the principle into a procedure, setting out steps for building a codebook for interview data, comparing theory-driven with data-driven codes, and discussing how to train other people to code from the book. Schreier (2012) gives the parallel account for qualitative content analysis, where the coding frame is the core instrument, built from the data or from theory, with the material segmented into units of coding and the frame piloted and evaluated before the main analysis. On both accounts a codebook is a set of instructions complete enough for another person to follow. The test of completeness is whether that person's coding would match the analyst's. Since 2024 that other person can be a model. Tai and colleagues (2024) gave a large language model a codebook and the interview text, ran the same text through it 160 times, and treated each pass as a new coder working deductively from the book. The test is now cheap to run, and the question moves to what a model's agreement shows.

What a codebook has to specify, and what it cannot

The hardest part of the codebook is not the label but the boundary. Ryan and Bernard (2003) survey the techniques by which themes are identified, from word repetition and indigenous typologies to cutting and sorting. They compare them on six dimensions, including required labour, required expertise and stage of analysis. Their finding states the unit's problem: theme identification is fundamental yet, they write, mysterious. A codebook is where that mystery has to be resolved, since a code whose boundary is unstated cannot be applied twice.

Basit (2003) shows that the tool does not settle the boundary. She compares manual and electronic coding across two interview studies: the method shapes how the data are organised and analysed, while making sense of the text remains the analyst's. Software can hold the codebook, but not decide whether a segment belongs under the code, and the completeness test assumes a coder who is not working in a second language.

Inductive and deductive coding are not rival camps

Fereday and Muir-Cochrane (2006) describe a hybrid that combines inductive, data-driven coding with deductive, theory-driven coding. They show it on a doctoral study of performance feedback in nursing self-assessment, coding ten focus groups and sixteen policies against a template drawn from social phenomenology. The hybrid produced three core themes, the first being familiarity: whether a feedback message was credited depended on the source's contact with the nurse and its grasp of the context. The codebook is where the combination becomes visible. A reader can see which codes were brought to the data and which were found in it.

Bradley, Curry and Devers (2007) give the applied version for health services research. They set out five code types: conceptual, relationship, perspective, participant characteristics and setting. Conceptual codes build taxonomies, and relationship and perspective codes build themes, carrying the analysis from raw data to taxonomy, themes and theory. The type of code an analyst writes decides what the analysis can produce.

Iteration is the analysis, not a failure of the book

Weston and colleagues (2001) show the scheme evolving through successive versions as the team's shared understanding changed, and draw a further lesson: a team builds codes, and coding builds a team. Collaborative analysis therefore requires a rigour an independent researcher may not need, and the codebook is both a product and a means of coordination. Its revisions are evidence of the analysis, not proof of its failure.

Locke, Feldman and Golden-Biddle (2022) ask what iteration consists of when it works. They answer that it is not a fixed sequence. Studying how projects report their coding, they identify three coding moments: making codes, organising to code, and putting patterns together. The order is organised by the next needed analytic input rather than by a template. A codebook that shows only the final codes therefore hides the iteration in which the analytic work happened.

Contested question: should an analysis be coded at all?

The codebook tradition treats coding as the way an interpretation becomes inspectable. The sharpest disagreement is not about technique but about whether coding should be central at all, and both sides argue from position rather than from comparison.

St. Pierre and Jackson (2014) hold that coding should not be central. Their evidence is an editorial that opens a special issue, and it reports no data set and no comparison. They argue that reducing analysis to coding forecloses the theoretical work analysis should do, and that the coding habit has become an unexamined orthodoxy. On this reading the codebook is the mechanism of a loss. Deterding and Waters (2021) press a narrower version from inside the practice, on the evidence of methodological argument and worked examples on large interview samples rather than an experiment. The procedures students are taught, they hold, were shaped by the tools of the 1960s, fit large samples and team coding poorly, and hamper transparency. Their remedy is to index large portions of the text and code analytically only where it matters.

Against them stands the codebook tradition, and its evidence is also prescription rather than trial. MacQueen and colleagues (1998) describe one team's practice. Schreier (2012) sets out a method in a textbook, and Saldaña (2021) in a manual. They make the categories explicit and require the frame to be piloted, so the basis of the analysis can be examined. St. Pierre and Jackson are right that coding can be thoughtless. They are wrong that the remedy is to stop coding, because the alternative to an inspectable procedure is not freedom but unaccountability.

The strongest counter-argument to this unit's own claim is that specification is not the same as validity. A codebook can be complete, followable and stable, and still encode a category formed before the data arrived. The unit's test, that another coder would match the analyst, certifies reproducibility rather than warrant. On that reading the codebook makes the analysis inspectable and leaves the interpretation as unexamined as before.

What would settle it is harder to state than in unit 1, because the two sides may not want the same thing. A study on one data set comparing a codebook-reconstructable analysis with a coding-free one would test what an analysis can claim. No further evidence would test what analysis is for.

What to do while the dispute is open

Code, and make the trail answer the objection rather than ignore it. Record in the codebook which codes were brought to the data and which were found in it. Version the book, so a changed boundary stays visible. Index one subsample and code a second, then report what each procedure supported and what it lost.

Where the argument is heading

Deterding and Waters (2021) and Locke, Feldman and Golden-Biddle (2022) both move away from exhaustive, line-by-line template coding. They move towards flexible, index-based procedures whose iteration is reported rather than concealed. The codebook is also migrating into tools. Basit (2003) found that the choice between manual and electronic coding shapes how data are organised. Schreier (2012) gives a chapter to software. Saldaña (2021) shows how many coding decisions are of a kind a machine could help to apply. St. Pierre and Jackson (2014) have asked whether coding should be central, and once that question is open, a codebook that records only the final codes has to say what it was for. As with the transcript in unit 1, the demand is not for a standard codebook, but for an account of how the categories were made and revised.

The sources this unit rests on

• Basit, T. N. (2003). Basit compares manual and electronic coding across two interview studies and finds that the method shapes how the data are organised while making sense of the text remains the analyst's task. Software can hold a codebook but not apply one.
• Bradley, E. H., Curry, L. A. and Devers, K. J. (2007). Bradley, Curry and Devers set out five code types for health services research, conceptual codes build taxonomies, relationship and perspective codes build themes, so the type of code decides what the analysis can produce.
• DeCuir-Gunby, J. T., Marshall, P. L. and McCulloch, A. W. (2011). DeCuir-Gunby, Marshall and McCulloch give steps for building a codebook for interview data, comparing theory-driven with data-driven codes and showing how to train coders from the book, which presupposes a book someone else can follow.
• Deterding, N. M. and Waters, M. C. (2021). Deterding and Waters argue that the grounded-theory coding procedures students meet fit poorly with large samples, team coding and mixed-method analysis. They propose indexing text and coding analytically only where it matters.
• Elliott, V. (2018). Elliott argues that coding is under-taught and under-theorised and sets out coding as decision-making, in which density, frequency and the size of the pieces coded are choices made within a study. Her article supplies the unit's claim that a code is a decision.
• Fereday, J. and Muir-Cochrane, E. (2006). Fereday and Muir-Cochrane describe a hybrid combining inductive, data-driven coding with deductive, theory-driven coding, and use a worked study to show the two can be integrated without losing rigour.
• Locke, K., Feldman, M. and Golden-Biddle, K. (2022). Locke, Feldman and Golden-Biddle identify three coding moments, making codes, organising to code and putting patterns together, and find the sequence is organised by the next needed analytic input rather than a template, so codebooks showing only the final codes hide the iteration.
• MacQueen, K. M., McLellan, E., Kay, K. and Milstein, B. (1998). MacQueen and colleagues describe the codebook as the means by which a team makes its analytic categories explicit, giving each code a definition, a boundary and example segments so that the interpretive basis of the analysis can be examined. It is the unit's central artefact.
• Ryan, G. W. and Bernard, H. R. (2003). Ryan and Bernard survey techniques for identifying themes and compare them on six dimensions including required labour, required expertise and stage of analysis, arguing that theme identification is fundamental yet rarely described.
• Saldaña, J. (2021). Saldaña's manual catalogues first-cycle and second-cycle coding methods and argues that coding is a cyclical and analytic rather than mechanical activity, with analytic memos carrying the move from codes to theory. It supplies the unit's account of coding as judgement.
• Schreier, M. (2012). Schreier treats the coding frame as the core instrument of qualitative content analysis, covering how to build it, how to segment the material, and how to pilot and evaluate it before the main analysis.
• St. Pierre, E. A. and Jackson, A. Y. (2014). St. Pierre and Jackson argue against the centrality of coding, holding that reducing analysis to coding forecloses the theoretical work analysis should do and that the coding habit has become an unexamined orthodoxy. The note rests on the title, the framing and the reported argument.
• Tai, R. H., Bentley, L. R., Xia, X., Sitt, J. M., Fankhauser, S. C., Chicas-Mosier, A. M. and Monteith, B. G. (2024). Tai and colleagues gave a large language model a codebook and interview text, ran the same text 160 times, treated each pass as a new coder, and compared the result with human coding. A codebook is now an instruction set a machine can execute.
• Weston, C., Gandell, T., Beauchamp, J., McAlpine, L., Wiseman, C. and Beauchamp, C. (2001). Weston and colleagues trace the development of a coding system across an interview study, showing successive versions of the scheme and arguing that coding is not what happens before analysis but part of it.