{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-11-16T00:58:22.017Z"},"content":[{"type":"documentation","id":"32520506-b6f9-4596-b2d9-f9f7d085d32b","slug":"deductive-coding-qualitative-research","title":"Deductive Coding: How to Code Interviews With a Start List of Codes","url":"https://www.koji.so/docs/deductive-coding-qualitative-research","summary":"Deductive coding is a qualitative analysis approach in which codes are defined before reading the data, from a theory, framework, earlier research or the study's questions, and then applied to transcripts. Saldaña calls it provisional coding and stresses that the start list should be revised. The guide covers when to use it, how to build and define a start list with MacQueen's six-part codebook entry, a six-step process with a pilot and revision, checking agreement with Cohen's kappa, keeping the method honest, a worked churn example, and how Koji's grounded first-pass codes and evidence links help check a start list.","content":"**Short answer:** Deductive coding is a way of coding qualitative data in which you write the codes before you read the transcripts, from a theory, a framework, earlier research or the questions your study has to answer, and then apply them to the data. Johnny Saldaña calls the same idea provisional coding. It is the fastest way to answer a question you can already frame (\"which of our five onboarding steps do people complain about?\"), and the easiest way to compare across studies. Its risk is that you only find what you went looking for, so most practitioners keep room for new codes and end up with a hybrid of deductive and inductive coding.\n\n**Try it this week:** take the last study you ran, write five codes from the question it was meant to answer, give each one a definition and a \"when not to use\" line, and code three transcripts with them. Anything you can't code goes on a separate list. That list tells you more about the limits of your framework than any amount of planning.\n\nThis guide covers when deductive coding is the right choice, how to build a start list, the six steps of coding with it, how to check that two people apply it the same way, and how to keep it from turning into confirmation bias. It sits in Koji's analysis series next to [the qualitative research codebook](/docs/qualitative-research-codebook) and [framework analysis](/docs/framework-analysis-qualitative), which uses a deductive frame by design.\n\n## What deductive coding is (and what it isn't)\n\nEvery coding approach answers one question: where do the codes come from?\n\n- **Inductive coding** builds codes from the data. You read a transcript, label what you see, and the codebook grows as you go. Grounded theory and most first-cycle open coding work this way.\n- **Deductive coding** starts with codes you defined in advance. In their widely cited 2005 paper, Hsieh and Shannon describe three approaches to qualitative content analysis. Their _directed_ approach is the deductive one: it starts from a theory or earlier research, which supplies the initial coding categories.\n- **Hybrid coding** does both. Fereday and Muir-Cochrane (2006) published the best-known worked example. They combined Boyatzis's data-driven inductive coding with Crabtree and Miller's template of a priori codes, so the theory shaped the analysis while themes could still come straight from the data.\n\nSaldaña's _Coding Manual for Qualitative Researchers_ describes provisional coding this way: it \"begins with a 'start list' of researcher-generated codes based on what preparatory investigation suggests might appear in the data before they are collected and analyzed.\" The same entry adds the part people forget: provisional codes \"can be revised, modified, deleted, or expanded to include new codes.\"\n\nThat second sentence is the difference between deductive coding done well and deductive coding done badly. A start list is a hypothesis about what you'll find, and it's meant to change.\n\nDeductive coding isn't the same as a closed survey question. A survey forces every answer into your categories. Deductive coding lets people say whatever they want and then asks how much of it your categories explain. The part they don't explain is often the finding.\n\n## When to code deductively\n\nUse a start list when at least one of these is true:\n\n1. **You're testing or extending a framework.** Jobs to Be Done forces, the four trustworthiness criteria, a service blueprint's stages, a technology acceptance model. The framework already names the concepts.\n2. **You need to compare across studies.** If last quarter's churn study coded \"price\", \"missing feature\" and \"switched to competitor\", this quarter's study should use the same codes, or the trend line means nothing.\n3. **The study has a narrow, decision-shaped question.** \"Which of the four setup steps causes the most drop-off?\" has four codes before anyone speaks.\n4. **Several people will code.** A shared, defined codebook is what lets a team split transcripts without ten private coding schemes.\n5. **You have little time.** Applying 12 defined codes is quicker than building a codebook from nothing, which is why rapid qualitative analysis in health services leans on templates built from the interview guide.\n\nCode inductively instead when you don't yet know the shape of the problem: early discovery, a new market, a \"why does nobody use this?\" question with no good theory behind it. If you're unsure, use the decision path below.\n\n![Flowchart for choosing deductive, inductive or hybrid coding: three questions about an existing framework, comparison with past studies and room for surprises.](/visuals/deductive-inductive-or-hybrid-coding-flowchart/deductive-inductive-or-hybrid-coding-flowchart.png \"Most product studies end up hybrid: a start list for the question you must answer, open codes for the rest.\")\n\n## How to build a start list\n\nA start list has three sources, and the best ones use all three.\n\n**The framework.** Turn each concept into a code. If you're using the JTBD forces of progress, that's push, pull, anxiety and habit. Keep the framework's own words so anyone who knows it can read your codebook.\n\n**The research questions.** Each question in your brief should map to at least one code. If a research question has no code, you won't be able to answer it at the end. If a code maps to no question, ask why it's there.\n\n**What you already know.** Last study's themes, support ticket tags, sales call notes. These give you codes in your customers' language, which saves translation later.\n\nThen write each code down properly. MacQueen and colleagues' 1998 guidance on team codebooks is still the standard: every entry has six parts.\n\n![Framework of the six parts of a codebook entry from MacQueen and colleagues: code, brief definition, full definition, when to use, when not to use and example.](/visuals/macqueen-codebook-entry-six-parts/macqueen-codebook-entry-six-parts.png \"Write the when-not-to-use line before you start. It settles most coder arguments in advance. · Source: MacQueen, McLellan, Kay and Milstein, Cultural Anthropology Methods, 1998\")\n\nThe \"when not to use\" field matters more in deductive coding than anywhere else. Pre-defined codes look clear on paper and blur on contact with real speech. \"Price\" and \"value for money\" will collide in your first transcript unless you've decided in advance which is which.\n\nKeep the start list short. Ten to twenty codes is a workable range for a product study. With fifty, coders stop checking definitions and start pattern matching on keywords.\n\n## Deductive coding in six steps\n\n![Process of deductive coding in six steps: build a start list, define codes, pilot, revise, code the full set and report what fell outside.](/visuals/deductive-coding-process-six-steps/deductive-coding-process-six-steps.png \"The pilot and the revision are the steps people skip, and they're the ones that make the codes mean something. · Source: after Fereday and Muir-Cochrane (2006) and Hruschka and colleagues (2004)\")\n\n1. **Build the start list** from the framework, the research questions and prior findings.\n2. **Define every code** with the six codebook parts. Add one real example from earlier data if you have it.\n3. **Pilot on two or three transcripts.** Code them, and keep a separate \"uncoded\" list of anything that matters but fits no code.\n4. **Revise.** Merge codes nobody can tell apart, split codes that hold two ideas, and add codes for whatever keeps landing on the uncoded list. Write down every change and why. This log becomes part of your [audit trail](/docs/qualitative-research-validity).\n5. **Code the full set.** New codes are still allowed, but each one needs a definition and a note on when it appeared, so you can go back and check earlier transcripts for it.\n6. **Report what fell outside.** Say how much of the data your framework explained and what it missed. A reader trusts deductive findings more when you show where the frame didn't fit.\n\nStep 4 is where a deductive study either becomes credible or becomes a rubber stamp. If you finish the pilot with zero changes to the codebook, either your framework is remarkably good or you weren't looking.\n\n## Checking that two people apply the codes the same way\n\nDeductive coding usually comes with a reliability check, because a pre-defined codebook makes the claim that anyone trained on it would code the same text the same way. That claim can be tested.\n\nThe standard measure is Cohen's kappa, which corrects simple percent agreement for chance. Landis and Koch (1977) proposed the bands still used to read it: 0.41 to 0.60 is moderate agreement, 0.61 to 0.80 substantial, and 0.81 to 1.00 almost perfect. Their bands are a convention, not a law, and plenty of methodologists consider them generous.\n\nExpect a low number first. Hruschka and colleagues (2004) worked through coding on three HIV behavioural studies with the US Centers for Disease Control and Prevention and reported what most teams discover the hard way: \"a team of coders will initially produce very different codings, but … it is possible, through a process of codebook revision and recoding, to establish strong levels of intercoder reliability.\" After those rounds, most of their codes reached a kappa of 0.8 or higher.\n\nA practical routine for a small team:\n\n- Both coders code the same two transcripts independently.\n- Compare code by code. For each disagreement, decide whether the definition was unclear or one coder misread the text.\n- Fix the definitions, not just the codes, then code two new transcripts and compare again.\n- Once agreement holds, split the remaining transcripts and spot-check one in five.\n\nOne caution. Reliability scores belong to deductive and codebook approaches. If you're doing [reflexive thematic analysis](/docs/reflexive-thematic-analysis), Braun and Clarke advise against kappa, because that method treats coding as interpretation, not measurement. Pick the quality check that matches your approach. For a full treatment, see [inter-rater reliability in qualitative research](/docs/inter-rater-reliability-qualitative-research).\n\n## Coding as a team: calibrate before you split the transcripts\n\nMost deductive coding in product teams isn't done by one researcher. It's a PM, a designer and someone from support, each taking a third of the transcripts on a Friday afternoon. That's fine, as long as they spend the first hour together.\n\nSit down with one transcript everyone has read. Code it out loud, one passage at a time, and argue about the borderline cases. Every argument that ends in \"fine, that's _onboarding friction_, not _missing feature_\" should become a sentence in the codebook. An hour of this prevents a week of reconciling three people's private definitions.\n\nThen agree how new codes get added. The simplest rule: anyone can propose one, but it goes into the shared codebook with a definition before anyone uses it, and the person who added it checks the transcripts already coded.\n\n## How to keep deductive coding honest\n\nThe main criticism of deductive coding is that it finds what it looks for. Four habits counter that.\n\n**Track the uncoded share.** If a large part of what people said fits none of your codes, your framework is missing something. Say so in the report.\n\n**Look for disconfirming cases.** For each major code, search for participants whose experience contradicts it. [Negative case analysis](/docs/negative-case-analysis) is the formal version of this.\n\n**Count with care.** Deductive codes make counting easy, and counting is useful (\"price came up in 9 of 14 interviews\"). But a code that appears often isn't automatically important, and twelve interviews is not a sample you can turn into percentages for a market.\n\n**Know when you've seen enough.** Codebooks stabilise quickly in a focused study. In Guest, Bunce and Johnson's 2006 analysis of 60 interviews, 92% of the codes they eventually developed had appeared within the first twelve. That's a reason to stop adding interviews only if your population is fairly uniform and your question is narrow, as theirs was. See [data saturation](/docs/data-saturation-qualitative-research) for the caveats.\n\n## A worked example: coding churn interviews against a start list\n\nA B2B SaaS team wants to know why accounts cancelled in the last quarter. They have a previous churn study, so they start from its codes:\n\n| Code               | Brief definition                              | When not to use                                                  |\n| ------------------ | --------------------------------------------- | ---------------------------------------------------------------- |\n| Price              | Cost was too high for the budget they had     | They thought it cost too much _for what it did_: use Value gap   |\n| Value gap          | Product didn't deliver enough for its cost    | The product lacked a specific capability: use Missing capability |\n| Missing capability | A named feature or integration they needed    | A competitor offered it: also add Competitor switch              |\n| Competitor switch  | They moved to a named alternative             | They built it in-house: use Internal build                       |\n| Champion left      | The person who bought it left or changed role | The team shrank but the champion stayed                          |\n\nThey pilot on three transcripts. Two things happen. \"Value gap\" and \"Price\" keep colliding, so they sharpen the definitions. And two of the three participants talk about a reorganisation that moved the product to a team who never asked for it. That fits no code, so it becomes a new one, _Ownership moved_, with its own definition and a note that it was added after transcript three.\n\nBy the end, _Ownership moved_ is the third most common reason. The old start list would have hidden it inside \"Champion left\". The deductive codes kept the trend comparable with last quarter, and the open code found the news.\n\n## How Koji helps\n\nKoji does the parts of this that take the most time, and leaves the framework decisions to you.\n\n**It structures the data before you code it.** Each Koji study has a [research brief](/docs/how-to-write-research-brief) and a fixed set of questions, so every transcript is organised by the same questions. Questions that map to codes you already know can be [structured questions](/docs/structured-questions-guide): a single choice or multiple choice question whose options are your categories, with an open follow-up asking why. The categories get counted for you, and the reasons come back as words.\n\n**It gives you an inductive pass to check your start list against.** For each open-ended answer, Koji's analysis writes short codes for each distinct point a participant made, either a descriptive label or one that keeps the participant's own framing (an in vivo code), and attaches the verbatim quote behind each one. Across interviews, it then clusters those codes into a single theme list per question, the second-cycle step. Theme charts appear once at least ten interviews have answered a question. Put that list next to your start list. Themes with no matching code are your candidates for new codes. Codes with no matching theme may not be in the data at all.\n\n**It keeps every code tied to the evidence.** Each finding in a Koji report links to the responses behind it, and each quote is highlighted in its transcript, so you can check whether a passage really belongs under a code before you count it.\n\n![Koji report summary with a side panel of supporting participant quotes behind one finding](/email/shots/shot-evidence-panel.jpg \"Each claim in the report opens the quotes behind it, which is what you check a code against.\")\n\n**It exports for your own coding.** The responses table exports to CSV with one row per participant and one column per question, plus Koji's themes for open questions. Apply your own codebook in a spreadsheet or your QDA tool, then compare.\n\nKoji doesn't apply a codebook you upload, and it isn't a substitute for the pilot-and-revise step: deciding that \"Value gap\" and \"Price\" are different codes is your call. What changes is the time it takes. Interviews run in parallel, by voice or text, in the participant's language, and transcripts come back already organised by question with a first coding pass attached. Compared with booking and transcribing calls one at a time, most of the waiting goes away. You can [start a study free](/docs/creating-your-first-study) and have transcripts to code the same week.\n\n## Frequently asked questions\n\n**What is the difference between deductive and inductive coding?**\nDeductive coding starts with codes defined before analysis, from a theory, framework, earlier research or the study's questions, and applies them to the data. Inductive coding builds codes from the data as you read it. Most applied studies combine both: a start list for what you need to answer and open codes for whatever the list misses.\n\n**Is deductive coding the same as provisional coding?**\nClose enough for most purposes. Saldaña uses provisional coding for coding that begins with a researcher-generated start list based on preparatory work, and he notes that the codes can be revised, deleted or expanded. Hypothesis coding, a related method in his manual, applies a predetermined list to test specific hypotheses.\n\n**How many codes should a deductive codebook have?**\nThere's no fixed rule. Ten to twenty well-defined codes works for most product and UX studies. With many more, coders stop checking definitions. Group codes under a few parent categories if the framework is large.\n\n**Can I add new codes during deductive coding?**\nYes, and you should when the data demands it. Give each new code a full definition, record when it was added and why, and recheck transcripts coded before it existed. That turns a purely deductive study into a hybrid one, which is what Fereday and Muir-Cochrane recommend.\n\n**Do I need inter-rater reliability for deductive coding?**\nIf more than one person codes, checking agreement is good practice, because a defined codebook claims that trained coders would apply it consistently. Cohen's kappa is the usual measure. If you're doing reflexive thematic analysis rather than codebook coding, don't use it: that method treats coding as interpretation.\n\n**Is directed content analysis deductive coding?**\nYes. Hsieh and Shannon (2005) describe directed content analysis as the approach that starts from a theory or earlier research to define the initial codes, which is deductive coding applied within content analysis.\n\n## Related Resources\n\n- [How to Build a Qualitative Research Codebook](/docs/qualitative-research-codebook)\n- [Framework Analysis for Qualitative Data](/docs/framework-analysis-qualitative)\n- [Inductive vs Deductive Research](/docs/inductive-vs-deductive-research)\n- [Inter-Rater Reliability in Qualitative Research](/docs/inter-rater-reliability-qualitative-research)\n- [First and Second Cycle Coding](/docs/first-and-second-cycle-coding)\n- [Structured Questions Guide](/docs/structured-questions-guide)\n\n## Sources\n\n- Saldaña, J. _The Coding Manual for Qualitative Researchers_. SAGE (provisional coding and hypothesis coding).\n- Hsieh, H.-F., & Shannon, S. E. (2005). Three approaches to qualitative content analysis. _Qualitative Health Research_, 15(9), 1277–1288.\n- 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.\n- MacQueen, K. M., McLellan, E., Kay, K., & Milstein, B. (1998). Codebook development for team-based qualitative analysis. _Cultural Anthropology Methods_, 10(2), 31–36.\n- Hruschka, D. J., Schwartz, D., St. John, D. C., Picone-Decaro, E., Jenkins, R. A., & Carey, J. W. (2004). Reliability in coding open-ended data: Lessons learned from HIV behavioral research. _Field Methods_, 16(3), 307–331.\n- Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. _Biometrics_, 33(1), 159–174.\n- Guest, G., Bunce, A., & Johnson, L. (2006). How many interviews are enough? An experiment with data saturation and variability. _Field Methods_, 18(1), 59–82.\n","category":"Research Methods","lastModified":"2026-11-15T05:57:44.387806+00:00","metaTitle":"Deductive Coding in Qualitative Research: A Practical Guide","metaDescription":"Learn deductive coding: build a start list of codes, define them, pilot and revise, check coder agreement and combine it with inductive coding. With a worked example.","keywords":["deductive coding","deductive coding qualitative research","provisional coding","a priori codes","hybrid inductive deductive coding","directed content analysis","deductive thematic analysis"],"aiSummary":"Deductive coding is a qualitative analysis approach in which codes are defined before reading the data, from a theory, framework, earlier research or the study's questions, and then applied to transcripts. Saldaña calls it provisional coding and stresses that the start list should be revised. The guide covers when to use it, how to build and define a start list with MacQueen's six-part codebook entry, a six-step process with a pilot and revision, checking agreement with Cohen's kappa, keeping the method honest, a worked churn example, and how Koji's grounded first-pass codes and evidence links help check a start list.","aiPrerequisites":["Basic familiarity with qualitative interviews","Understanding of what coding qualitative data means"],"aiLearningOutcomes":["Decide when deductive, inductive or hybrid coding fits a study","Build a start list from a framework, research questions and prior findings","Write codebook entries with clear when-not-to-use rules","Pilot, revise and apply a deductive codebook in six steps","Check coder agreement and report what the codes missed"],"aiDifficulty":"intermediate","aiEstimatedTime":"14 min read"}],"pagination":{"total":1,"returned":1,"offset":0}}