Framework Analysis: The Matrix Approach to Qualitative Data

Framework Analysis is a matrix-based method for analysing qualitative data: rows are cases (participants, sites, documents), columns are themes, and each cell holds a summarised, page-and-line-referenced note back to the original transcript. It was developed by Jane Ritchie and Liz Spencer at the Qualitative Research Unit of the National Centre for Social Research (NatCen) in the United Kingdom in the late 1980s, and formalised in their 1994 chapter ‘Qualitative Data Analysis for Applied Policy Research.’ Unlike grounded theory, which builds categories inductively from the data with no predetermined structure, Framework Analysis starts from the research questions themselves and lets an a priori thematic structure organise the first pass of analysis — while still leaving room for themes that emerge from the data to be added in. That combination is why it took hold in applied and policy research: a funder or commissioning body has already specified what the study needs to answer, a multi-disciplinary team needs a shared, auditable structure to work from, and a fixed reporting deadline rules out a fully open-ended inductive process.

The five stages

Ritchie and Spencer’s original method runs in five stages. Each stage produces a concrete output that the next stage works from, and — unlike some qualitative methods — the intermediate outputs (the framework, the index, the matrix itself) are artefacts a reader can actually inspect, which is part of why the method audits well in team and commissioned research.

1. Familiarisation

The analyst (or, on a team project, several analysts) reads and re-reads the full data set — transcripts, field notes, documents — without yet coding anything, noting recurring ideas, language, and initial impressions. On a team project this stage usually includes each analyst independently familiarising themselves with a subset of transcripts, then comparing notes, so the framework that follows is grounded in the whole data set rather than one person’s reading of it.

2. Identifying a thematic framework

The team drafts an initial set of themes and sub-themes — an index — combining a priori issues drawn directly from the research questions and topic guide with concepts that emerged from familiarisation. This draft framework is deliberately provisional: it is a working structure, refined against the data in the next stage rather than fixed in advance and imposed on it.

3. Indexing

Every transcript is worked through systematically and each relevant passage is labelled with the framework codes that apply to it — analogous to coding in other qualitative methods, but applied against the shared index from stage 2 rather than built up code-by-code from a blank slate. Indexing is where the framework gets tested against real data: passages that don’t fit any existing code are exactly what should trigger revising the framework itself, not forcing an awkward fit.

4. Charting

Indexed data is lifted out of each transcript and rearranged into the matrix itself — one chart per theme (or, for smaller studies, one combined matrix), with a row per case and a cell per case-theme intersection. Each cell holds a summary in the participant’s own words and phrasing wherever possible, not the analyst’s paraphrase, together with a reference back to the exact page and line in the source transcript. This is the step that makes Framework Analysis distinctive: the whole data set becomes visible at once, case-by-case and theme-by-theme, in a form that can be read across a row (everything one participant said) or down a column (everything every participant said about one theme).

5. Mapping and interpretation

With the matrix built, the analyst studies it as a whole — comparing across cases, looking for patterns, ranges, and clusters, and building typologies or explanations that account for what’s in the matrix. This is the analytic payoff of the first four stages: because the matrix already exists as a structured object, the search for patterns is a search across a shared, page-referenced grid, not a re-immersion in dozens of separate transcripts.

A worked example: the matrix, and a double-indexing consistency check

The illustration below is a composite for teaching purposes, not a real study — a hypothetical 6-participant evaluation of a support programme, indexed against five themes drawn from the (also hypothetical) research questions. A real charted matrix would hold full summarised quotations per cell; this shows the shape only.

Reading down a column shows how many of the six cases raised each theme (awareness: 4/6; access: 2/6; trust: 3/6; affordability: 2/6; competing priorities: 3/6) — the kind of pattern that stage 5 (mapping) works from. Reading across a row shows the full picture for one participant, which is what makes typology-building across rows possible.

Framework Analysis’s team-based indexing (stage 3) is usually paired with a consistency check: two analysts independently index a shared sample of excerpts against the same framework, then compare. The numbers below are not illustrative guesses — they come from a small script that hand-assigns two independent category codes to 30 excerpts against a 5-category framework (Awareness / Access / Trust / Affordability / Competing priorities) and computes agreement programmatically, so the result is exactly reproducible rather than a plausible-sounding number:

• 24 of 30 excerpts (80.0%) received the same code from both coders.
• Cohen’s kappa, which corrects raw agreement for the rate you’d expect by chance given each coder’s mix of category assignments, came out at 0.748 — on the commonly cited Landis and Koch scale, “substantial” agreement, one band short of “almost perfect.”
• All 6 disagreements were between semantically adjacent codes (e.g. Access vs Trust, Affordability vs Competing priorities) rather than random noise — the kind of pattern that, in a real study, is the actual signal to revisit the framework’s category boundaries in stage 2, not just a number to report and move past.

That last point matters more than the topline number: a mixed, non-perfect result like this is the normal, useful outcome of an indexing consistency check. It doesn’t mean the analysis failed — it means the framework had a genuine ambiguity the team can now resolve with a documented decision rule before charting proceeds.

Framework Analysis vs. grounded theory

The two methods sit at different points on the same inductive-to-structured spectrum. Grounded theory deliberately withholds any predetermined structure — categories are built up from the data through open, axial, and selective coding, and the analysis continues, sample by sample, until theoretical saturation. Framework Analysis starts with a structure drawn from the research questions and refines it against the data, but the structure exists before indexing begins. Neither is “more rigorous” than the other; they answer different kinds of research questions. A study with genuinely open questions about how a phenomenon works, with time to sample iteratively, fits grounded theory. A study commissioned to evaluate a specific policy or programme against pre-defined objectives, on a fixed timeline, with a multi-disciplinary team that needs a shared working structure from day one, fits Framework Analysis.

Framework Analysis vs. thematic analysis

Framework Analysis and Braun and Clarke’s thematic analysis both produce theme-based findings, and both can combine inductive and a priori elements. The practical difference is the matrix itself: thematic analysis’s output is a set of named, defined themes supported by illustrative extracts, without a required intermediate step that lays every case against every theme in one visible grid. Framework Analysis’s charting stage forces that grid into existence as a distinct, inspectable artefact — which is exactly what makes it suit team-based and commissioned research (a client or co-investigator can review the matrix itself, not just the final write-up) at some cost in flexibility, since restructuring an already-charted matrix is more disruptive than renaming a theme.

Where it’s used, and where it’s been extended

Framework Analysis’s original home was large-scale UK social policy research at NatCen. In 2013, Nicola Gale, Gemma Heath, Elaine Cameron, Sabina Rashid, and Sabi Redwood published ‘Using the Framework Method for the analysis of qualitative data in multi-disciplinary health research’ in BMC Medical Research Methodology, extending and slightly re-sequencing the method (they separate transcription and coding as their own steps, for seven stages in total) specifically for health services research teams that mix clinicians, patients, and lay representatives with professional qualitative researchers. Their central argument is that the method’s explicit, auditable structure is what lets a team without shared qualitative training work from the same matrix — provided, as they’re careful to note, that an experienced qualitative methodologist still leads the process. That caveat is worth taking seriously: a visible structure makes indexing decisions easy to delegate, but it doesn’t remove the judgment calls in building the framework and interpreting the matrix in the first place.

When to reach for it

• The research questions are defined before fieldwork starts — a commissioned evaluation, a funded policy study with fixed objectives, an audit against known criteria.
• A multi-disciplinary or mixed-experience team needs a shared structure — clinicians, policy staff, or lay co-researchers working alongside qualitative specialists.
• The findings need to be auditable to someone outside the analysis team — a commissioner, a co-investigator, a committee — who can review the matrix itself rather than take the write-up on trust.
• The timeline is fixed, which favours a structured, stage-based process over an open-ended iterative one.

It’s a weaker fit when the research questions are genuinely open (a predetermined framework forecloses exactly the emergent structure grounded theory is built to find), or for a single-analyst project small enough that a full matrix adds process overhead without an auditability payoff a lone researcher doesn’t need.

Software

The matrix structure maps directly onto spreadsheet software for small studies, and several dedicated qualitative data analysis (QDA) packages, including NVivo, offer a purpose-built framework-matrix view that keeps each cell linked back to its source passage in the coded transcript — the digital equivalent of the page-and-line reference in stage 4. See content analysis and narrative analysis for how coding-driven and structure-first qualitative approaches compare more broadly, and the research methods hub for the full set of qualitative and quantitative method guides on this site.

FAQ

Is Framework Analysis the same as the Framework Method?

They’re closely related, not identical. “Framework Analysis” usually refers to Ritchie and Spencer’s original five-stage version from applied policy research. “The Framework Method” usually refers to Gale et al.’s 2013 seven-stage adaptation for multi-disciplinary health research teams, which separates transcription and coding into their own explicit steps. Most people use the two names loosely interchangeably; if you’re citing a specific source, cite the version you actually followed.

Do you need qualitative software to do Framework Analysis?

No. The method predates dedicated QDA software by decades and works with a word processor or spreadsheet for the matrix itself, though software that keeps each cell linked to its source transcript passage removes a real amount of manual cross-referencing on a large data set.

How is Framework Analysis different from just building a codebook?

A codebook is a list of codes and their definitions — it’s an input to Framework Analysis’s stage 2 and 3, not the whole method. What makes Framework Analysis distinct is the charting stage: rearranging the indexed data into a case-by-theme matrix that can be read across cases as well as within them, which a codebook alone doesn’t produce.

Can new themes be added after indexing has started?

Yes, and this is expected rather than a sign the framework was wrong. The method treats the thematic framework as provisional through indexing — a passage that doesn’t fit any existing code is a signal to revise the framework, document the change, and, where practical, re-check already-indexed transcripts against the revision.