Qualitative data is non-numeric information that captures qualities, meanings, and lived experience rather than quantities. It takes the form of text, audio, video, images, observation notes, and physical artefacts, and it answers “how” and “why” questions that counts alone cannot. This guide covers what qualitative data is, where it comes from, how it is coded and analysed, and what makes it rigorous and defensible.
What is qualitative data?
Qualitative data describes attributes, meanings, contexts, or experiences rather than measuring an amount. Where quantitative data can be summarised with a mean or a standard deviation, qualitative data is interpreted for its content and meaning: a transcript is read for what it says, not averaged.
It is closely related to but not identical with qualitative research, which is the broader methodological approach built around collecting and analysing this kind of data. This guide focuses specifically on the data itself: its forms, its collection, and its analysis. A companion page, qualitative vs quantitative data, contrasts the two paradigms directly if that is what you are looking for.
Examples of qualitative data by source
- Interview transcripts — verbatim records of structured, semi-structured, or unstructured conversations.
- Focus-group recordings and transcripts — group discussion capturing interaction between participants, not just individual responses.
- Open-ended survey responses — free-text answers to “why” or “describe” questions on an otherwise quantitative instrument.
- Field notes — a researcher’s written observations from time spent in a setting.
- Diaries and journals — participant-generated reflective accounts, often longitudinal.
- Documents and policies — organisational records, meeting minutes, or regulatory text analysed as data in their own right.
- Social media text and public posts — naturally occurring text data, increasingly used in discourse and content analysis.
- Photographs, video, and physical artefacts — visual and material data, common in ethnography and visual methods.
- Clinical and laboratory free-text notes — the narrative portions of a clinical record or lab notebook, as distinct from coded or numeric fields.
The measurement-scale connection: why some “numbers” are still qualitative
Measurement scales are usually split into four levels: nominal, ordinal, interval, and ratio. The first two — nominal (unordered categories, like blood type) and ordinal (ordered categories, like a pain scale of mild/moderate/severe) — are categorical, and categorical data is qualitative in nature even when it is recorded as a number.
This is the point most introductions to the topic skip past. When a survey codes sex as 1 = male, 2 = female, or a satisfaction item as 1 = strongly disagree through 5 = strongly agree, the digit is a label standing in for a category, not a quantity. Nothing about “2” being twice “1” is meaningful. Averaging those codes — reporting a “mean sex of 1.4” or a “mean agreement of 3.2” — treats an ordinal label as if it were measured on an interval scale with equal spacing between points, which it was not designed to be.
In practice this is contested rather than simply forbidden. Likert-type items (a single ordinal question) are widely treated by methodologists as qualitative/ordinal and analysed with medians, modes, and non-parametric tests. Likert scales — a composite score summed or averaged across several items — are more defensible to treat as approximately interval and analyse with means, because summing multiple ordinal items tends to produce a more continuous, better-behaved distribution. The debate over where exactly that line sits, and how much violation of strict interval assumptions is tolerable, has run in the methods literature for decades and has no single resolved answer. The safest practice is to be explicit about which one you have (a single item or a composite scale) and to justify the statistic you choose rather than defaulting to a mean because the codebook happens to use digits.
Collection methods
How qualitative data is collected shapes what it can and cannot show:
- Structured interviews use a fixed set of questions asked in the same order to every participant, maximising comparability but limiting the ability to follow up on unexpected answers.
- Semi-structured interviews use a guide of core questions but allow the interviewer to probe and reorder, trading some comparability for richer, more contextual responses. This is the most common format in applied qualitative research.
- Unstructured interviews are closer to guided conversation, with minimal predetermined structure — well suited to exploratory work but hardest to compare across participants and to code consistently.
- Focus groups generate data through group interaction, surfacing shared or contested views that individual interviews may not, but are vulnerable to dominant voices and social-desirability effects.
- Participant observation has the researcher embedded in the setting being studied, gaining depth and access at the cost of potential influence on what is being observed.
- Non-participant observation keeps the researcher outside the activity being observed, trading some access and rapport for less risk of the researcher’s presence changing behaviour.
- Document analysis uses existing text — policies, records, correspondence — that was not produced for the study, avoiding reactivity but constrained to whatever was recorded for other purposes.
Analysis approaches
Several distinct traditions exist for making sense of qualitative data, and they are not interchangeable — each carries different assumptions about what counts as a valid finding.
- Thematic analysis identifies and reports patterns (themes) across a dataset. Braun and Clarke’s widely used six-phase framework (Braun, V. and Clarke, V., “Using Thematic Analysis in Psychology,” Qualitative Research in Psychology, 2006) runs: familiarisation with the data, generating initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the report. See CASRAI’s dedicated guide to thematic analysis for the full walkthrough, including Braun and Clarke’s later distinction between coding-reliability, codebook, and reflexive variants of the method.
- Content analysis systematically categorises text, and can be run qualitatively (interpreting latent meaning) or quantitatively (counting the frequency of predefined categories or words) — making it something of a bridge between the two paradigms.
- Grounded theory builds theory inductively from the data itself rather than testing a pre-existing hypothesis, using open coding (breaking data into initial concepts), axial coding (relating those concepts to each other), selective coding (integrating around a core category), and constant comparison — continually comparing new data against codes and categories already developed, rather than coding everything and comparing only at the end.
- Framework analysis uses a structured matrix of cases against themes, developed for applied policy research where transparency and the ability to trace a finding back to its source data matter to a non-academic audience.
- Discourse analysis examines how language constructs meaning, identity, and power in text or talk, treating the way something is said as data in itself, not just a vehicle for content.
- Narrative analysis treats the structure of a story — sequence, plot, voice — as the unit of analysis, rather than breaking it into discrete themes.
- Interpretative Phenomenological Analysis (IPA) focuses on how individuals make sense of a specific significant life experience, working case by case in depth before looking for patterns across a (typically small) sample.
Coding mechanics
Coding is the process of labelling segments of data with a short descriptor that captures what is happening in that segment.
- Deductive coding applies codes derived in advance from theory, prior research, or the research question. Inductive coding lets codes emerge from the data itself during the reading process. Most real projects use both: a small deductive starter set plus inductively developed codes for what the starter set did not anticipate.
- The codebook is the documented list of codes, their definitions, and inclusion/exclusion criteria or examples — it is what makes coding traceable and, where required, replicable by someone other than the original coder.
- First-cycle coding is the initial pass assigning codes to data; second-cycle coding reorganises, merges, or groups first-cycle codes into broader categories or themes once the full picture of the dataset is visible.
- Intercoder reliability measures the extent to which independent coders assign the same codes to the same data, most commonly with Cohen’s kappa, which corrects for the agreement expected by chance alone. It is genuinely useful where the analytic goal is a stable, largely fixed coding scheme applied consistently across a large or team-coded dataset — for example, coding-reliability content analysis. It is philosophically inconsistent with approaches that treat coding as an interpretive act performed by a positioned researcher rather than a mechanical classification exercise: Braun and Clarke’s own reflexive thematic analysis explicitly rejects inter-rater reliability statistics as a quality marker, on the grounds that a single “correct” code a kappa score could validate against does not exist when meaning-making is understood to be the analyst’s interpretive contribution, not a fact to be extracted. Applying kappa to that kind of analysis does not make it more rigorous; it applies a criterion from a different paradigm.
Saturation
Saturation is the point in data collection at which new data stops producing new codes, themes, or insights — the traditional justification for stopping data collection and settling a qualitative sample size, in place of the a priori power calculations used in quantitative work (see CASRAI’s guide to sampling and sample bias).
The live methodological critique is that saturation is very often asserted in a methods section — “data collection continued until saturation was reached” — without being demonstrated: without reporting when new codes stopped appearing, how many interviews that took, or what specific evidence supported the claim. Because saturation is inherently retrospective (you only know you have reached it once further data confirms it), and because “no new themes emerged” is a judgment call rather than a measurable threshold, reviewers and methodologists increasingly ask for the reasoning to be shown, not just the conclusion stated.
Rigour: credibility, transferability, dependability, confirmability
Lincoln and Guba’s widely adopted framework (Lincoln, Y.S. and Guba, E.G., Naturalistic Inquiry, 1985) reframes the quantitative criteria of internal validity, external validity, reliability, and objectivity into equivalents suited to qualitative work:
- Credibility — confidence in the truth of the findings, built through techniques like prolonged engagement, triangulation across data sources or methods, and member checking (returning findings to participants to confirm they recognise their own experience in them).
- Transferability — the extent to which findings might apply in other contexts, supported not by statistical generalisation but by thick description detailed enough for a reader to judge fit for themselves.
- Dependability — the stability of the findings over time and conditions, supported by a clear, auditable account of how decisions were made during the study.
- Confirmability — the degree to which findings are shaped by participants and data rather than researcher bias, supported by an audit trail and explicit reflexivity.
Reflexivity — the researcher’s ongoing critical awareness of their own position, assumptions, and influence on the research — and a documented audit trail of analytic decisions are the practical mechanisms most of these criteria rest on. Neither is optional detail; both are usually what a reviewer is actually checking for when a methods section is judged as rigorous or not.
Data management: why qualitative data is harder to share
Managing qualitative data raises problems that a numeric dataset mostly does not:
- Transcription conventions need to be decided and documented up front — verbatim (including false starts, pauses, and filler words) versus cleaned/intelligent transcription (edited for readability) answer different research questions and are not interchangeable once collection is finished.
- Anonymisation and pseudonymisation of qualitative data is substantially harder than de-identifying a spreadsheet. A rich narrative account can reveal identity through an unusual combination of job title, location, timeline, and circumstance even after names are removed — the re-identification risk sits in the detail that makes the account valuable as evidence in the first place, which is exactly what makes qualitative de-identification a genuine trade-off against analytic richness, not a mechanical redaction task.
- Consent for reuse and archiving has to be obtained at the point of collection, not retrofitted afterward. Informed consent forms for qualitative studies should state explicitly whether anonymised transcripts may be archived, shared with other researchers, or reused for secondary analysis — participants who agreed to be interviewed for one study have not automatically agreed to have their words redistributed.
- Sharing plans belong in the study’s data management plan from the outset: what will be shared, in what redacted or synthesised form, under what access controls, and for how long raw recordings and transcripts are retained versus destroyed. Because full de-identification of rich narrative data is often impossible without destroying its analytic value, many qualitative DMPs specify restricted or mediated access — a data enclave, a request-and-review process — rather than open deposit, which is the default expectation for most quantitative datasets.
This is where qualitative data management genuinely diverges from the quantitative default that most institutional RDM guidance is written around: “make it FAIR and deposit it” is a much harder promise to keep when the data is a set of identifiable human accounts rather than a table of numbers. See CASRAI’s guide to data collection methods for the collection-stage decisions that determine what management problems you inherit later.
Software: organises the data, does not analyse it
Qualitative data analysis software — NVivo, ATLAS.ti, MAXQDA, and Dedoose among the most established — provides the infrastructure for coding, memoing, and querying a large qualitative dataset: attaching codes to text/audio/video segments, tracking a codebook, running queries like “show me everything coded X by participants who are also coded Y,” and supporting multiple coders working on the same project.
What none of these tools do is analyse the data for you. The software organises and retrieves; the researcher still has to decide what a pattern means, whether two codes should be merged, and what the resulting themes actually say about the research question. Automated or AI-assisted coding suggestions, where a tool offers them, are a starting point to be checked against the data, not a substitute for the interpretive work — a common and consequential misunderstanding among researchers new to the software.
Frequently asked questions
What is qualitative data?
Qualitative data is non-numeric information — text, audio, video, images, observations, or artefacts — that captures qualities, meanings, and experience rather than measurable quantities. It is analysed by interpretation rather than by calculation.
What are examples of qualitative data?
Interview transcripts, focus-group recordings, open-ended survey responses, field notes, diaries, policy documents, social media text, photographs and video, and the narrative free-text portions of clinical or lab records.
Is a number ever qualitative data?
Yes. Nominal and ordinal data are categorical and therefore qualitative in nature, even when recorded as digits (for example, 1 = male, 2 = female, or a 1-5 Likert code). The number is a label for a category, not a quantity, so operations like averaging it are only meaningful with justification — and are contested for single ordinal items, though more broadly accepted for summed multi-item scales.
What is the difference between qualitative data and qualitative research?
Qualitative data is the material itself — the transcripts, notes, and recordings. Qualitative research is the broader methodological approach: the design, collection strategy, and analytic tradition used to generate and make sense of that data. See CASRAI’s definition of qualitative research for more.
How much qualitative data is enough?
There is no fixed number; the traditional answer is data saturation — the point where further collection stops producing new codes or themes. Because saturation is a judgment call made in hindsight, reviewers increasingly expect the reasoning behind it to be documented, not just asserted.
