Qualitative analysis for interaction research

A study comes back with hours of recorded talk, a notebook of field notes and a folder of video. None of it is a finding yet. Analysis turns a record into a claim another person can examine, and it takes most of the effort in a qualitative project.

The contested question can be stated first. Can the quality of a qualitative analysis be judged by how far two coders agree? One tradition says yes and asks for a figure with every paper. Another says no: interpretation is the point of the work, and agreement between two people trained in the same frame measures the training rather than the finding (Braun and Clarke, 2021; O'Connor and Joffe, 2020). Know what an agreement figure can show, what it cannot, and choose deliberately.

A recording is not yet data

Transcription is the first analytic act, not a clerical one. Any transcript is a selection: it fixes what was said, in what order, and it decides which features of the setting to keep. Overlap, pause length, emphasis, a change of pitch, a hand reaching for the mouse. Each is marked or dropped, and the choice depends on the question the analyst intends to ask.

The cost of a careless choice appears later. Braun and Clarke (2006) treat transcription as a way of starting to become familiar with the material, which is also the reason to do it deliberately. Heath, Hindmarsh and Luff (2010) give transcription notations their own appendix in a book about video, and their point holds in another medium: the recording is a record and not the setting, so the setting must be recovered by other means.

A constructed case, not a study result. A team studying a telephone helpline transcribes its calls as a clean document with one speaker per line. The analysis proceeds well until someone claims that advisers cut callers short. The claim may well be right and the transcript cannot support it: there is no timing and no overlap in it. The recording could. Transcribe for the question you will ask, and return to the recording for claims the transcript cannot carry.

Coding is a decision you can be asked to defend

Coding means attaching a short label to a segment because something in it is analytically interesting. That is all it means at the start. It becomes more when segments are compared with one another, which the grounded theory tradition calls constant comparison. A code becomes a category when it stops naming content and starts making a claim about how things relate.

The instrument that carries this work is the codebook, and O'Connor and Joffe (2020) describe it precisely: a list of codes, often grouped into higher level categories, each with a definition and example segments. They also name the problem that catches beginners: when coders decide for themselves where a segment starts and ends, the boundaries differ, and some take larger chunks while others code short ones. That difference alone moves an agreement figure, before any disagreement about meaning has occurred.

A constructed case: two coders agree on seventy per cent of units, which sounds reassuring until the disagreements are read. Almost all sit on one code, and all concern the same boundary question. The useful output is not the percentage but the sentence the coders then write to draw that boundary sharply. O'Connor and Joffe report the same: the content of intercoder disagreements is often more valuable than the final score.

Thematic analysis, and which version you are doing

Braun and Clarke (2006) gave the field the version most students meet first. Six phases: familiarising yourself with the data, generating initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the report. They also name the failures that make an analysis worthless. The first is a failure to analyse at all: extracts strung together with commentary that paraphrases them. The second is using the data collection questions, such as the interview schedule, as the reported themes, when no analytic work across the data set has been done. The language of themes emerging from the data suggests that themes sit in the data waiting to be found, and Braun and Clarke argue the opposite, that themes reside in the thinking of the analyst who connected the material.

The 2021 paper is the revision. Braun and Clarke (2021) point out that thematic analysis now exists in several iterations differing paradigmatically, philosophically and procedurally, and that this plurality is often unrecognised by editors, reviewers and authors who promote coding reliability measures as universal requirements of quality. They list ten common problems in published work citing their guidance, and close with twenty critical questions for reviewers and editors. The consequence is a reporting duty: say which version you used, what it assumes about knowledge, and why it suits the question.

A constructed case shows the failure the 2021 paper targets. A dissertation reports four themes: views on trust, experience of waiting, problems with reminders, suggestions for improvement. Those are the headings of the interview schedule, and the report's fluency hides that nothing was analysed. The repair is a return to the codes and a search for a patterning the schedule could not have predicted.

Grounded theory and the promise in its title

Glaser and Strauss (1967) made the strongest claim for the systematic generation of theory from data. Their programme is theory discovered from data that has been systematically obtained and analysed, through the deliberate choice and study of several comparison groups. They argue that a grounded theory fits empirical situations and supplies explanations and predictions, while devoting part of the book to the credibility of that claim. Constant comparison, the search for negative cases and the move from codes to categories all come from here, and they remain good practice whether or not the label is used.

The title is where the dispute lies. A theory that is discovered suggests a finder rather than an author, and the metaphor has been questioned ever since. Braun and Clarke's (2006) argument that themes are made rather than found is one version of the objection. What a grounded theory study can claim is therefore narrower than the title suggests: it can describe the conceptual structure it built by comparison and show the data that forced each step, and it cannot show that the structure was already there.

A constructed case: a study of a booking system offers five named themes and calls them a theory of booking anxiety. There is no comparison across groups, no negative case, and no account of how any category changed. The label does no work.

Content analysis and what counting can and cannot buy

Hsieh and Shannon (2005) separate three approaches, and the separation is the useful part. In conventional content analysis the categories are derived directly from the text. In the directed approach the analysis starts from a theory or earlier findings that guide the initial codes. In the summative approach the analyst counts keywords or content and then interprets the context behind the count. The three differ in their coding schemes, in where their codes come from, and in the threats to their trustworthiness.

Counting buys reproducibility and speed. It does not buy meaning. Suppose a team counts the word sorry across the transcripts of a support chat and finds it clustered in one week. The count cannot tell you whether an apology repaired a failed turn or was a formula used every few lines. The summative approach is honest about that two step structure: the task is to make the step from count to interpretation visible enough to be disagreed with.

Interaction as an achievement, not a message

Conversation analysis starts from a different premise. Talk is not a channel carrying meanings that the analyst decodes. It is an order that participants produce together, turn by turn, in the time available. Sacks, Schegloff and Jefferson (1974) built a model of turn-taking and examined it against grossly observable facts about conversation. The properties they settle on are worth learning by name: the model is locally managed, party-administered, interactionally controlled, and sensitive to recipient design.

Suchman (1987) carried the same scepticism into design. Her target was the planning model that treats a course of action as the execution of a plan, and her argument is that plans are better understood as resources for action than as its determinants. Her case was a recorded interaction between novice users and an intelligent machine whose design lacked resources that human communication normally supplies. The question she leaves a designer is precise: what does this system assume about the person on the other side of it?

A constructed case: a voice assistant asks which day the caller wants. The caller answers "the same as last time". The system registers a failure and repeats the question, with an emphasis a hearer would read as impatience. Nobody has misunderstood anything. The caller has answered as people who share a history do, and the design has nowhere to put it.

Video: turns, gaze, gesture and the thing on the table

Heath, Hindmarsh and Luff (2010) wrote the practical guide to video in qualitative research. Its chapters follow the work, from access, ethics and project planning through collection to matters of context and the presentation of findings. Their examples come from control centres, operating theatres, medical consultations, auction houses and museums, so the same apparatus applies from work to leisure. The appendix of transcription notations is the reminder that video analysis inherits the transcription problem too.

Jordan and Henderson (1995) set out interaction analysis as the study of what people do with one another and with the objects around them. Their account is usually summarised through seven foci: the structure of events, the temporal organisation of activity, turn taking, participation structures, trouble and repair, the spatial organisation of activity, and artefacts and documents. The turn is not only a speech turn, because a reach for the mouse, a glance and a pause all take part in the coordination. And the artefact is not scenery, because what sits on the screen or the desk is part of how participants organise what they do.

A constructed case: two people share a screen and one stops speaking while reaching for the mouse. From the audio alone that is a gap in the talk. From the video it is a turn, and the other person waits. An analysis reading only the transcript records the silence and loses the sequence.

Rigour without a number, and the analyst in the account

What replaces the reliability figure? O'Connor and Joffe (2020) work with a different vocabulary: credibility, transferability, dependability and trustworthiness. They set out the practices that carry that vocabulary: transparent reporting of the analytic procedure, thick description with plentiful raw data, triangulation, attention to deviant cases, and asking participants to validate an interpretation. They also record the objection, attributed to Guba and Lincoln, that reliability is a criterion imported from a paradigm assuming a single reality. They add a counter risk: research showing that nonsensical mathematical information can inflate judgements of a report's quality. Their conclusion is measured: agreement assessment has real uses, and should never become the point of the analysis.

Reflexivity is the other half, and Finlay (2002) is the best short guide to it. She describes the process as negotiating a swamp and maps five variants of reflexive practice: introspection, intersubjective reflection, mutual collaboration, social critique, and discursive deconstruction. The variants are not interchangeable. What they share is that reflexivity is work done on the analysis rather than a confession at the front of a report: which doors opened, which questions were askable, which reading of an extract you resisted, and what would have to be true for a different reading to win.

Writing is where this becomes visible. The participant's own words are evidence, so they belong in the finding, long enough to be judged and with the setting around them. The analyst's claim belongs beside them and must be separable from them, because a claim that paraphrases a quotation has not been made. Braun and Clarke (2006) warn against the opposite failure too, in which one or two instances are reified into a pattern, and that warning holds whatever method was used. All of it meets one question that the next stretch of the programme takes into design and measurement: what would have to be observed for this claim to be wrong, and what would the numbers then have to show?