You have 800 answers to “Why did you cancel?” and a meeting on Friday. Qualitative coding turns that pile of text into counts you can compare by segment, while keeping the quotes that explain them. It is the core step in thematic analysis, the wider method of finding patterns in qualitative data. This guide covers inductive and deductive coding, a codebook template, a worked example and intercoder agreement.
Key takeaways
- Qualitative coding labels segments of text with short codes so you can count, compare and quote them.
- Inductive coding builds codes from the data; deductive coding applies a codebook written in advance. Most projects use a hybrid.
- A good codebook gives each code a name, a definition, include and exclude rules and an example quote.
- One response can carry several codes, so code percentages can add up to more than 100%.
- Check intercoder agreement on a sample before coding everything.
What is the difference between inductive and deductive coding?
The difference is where the codes come from. In inductive coding, you read the data with no list in hand and create codes as ideas appear. In deductive coding, you start from a predefined codebook, often based on earlier research. A hybrid approach starts with a short codebook and adds codes when the data shows something it misses.
| Inductive coding | Deductive coding | Hybrid | |
|---|---|---|---|
| Starting point | The data | A predefined codebook | A short codebook, open to additions |
| Best when | The topic is new | You track a question over time | Most survey and interview projects |
| Strength | Finds unexpected issues | Fast and comparable across waves | Comparable, still open to new issues |
| Risk | Slow; codes drift between coders | Misses anything outside the codebook | Codebook grows without discipline |
Use inductive coding for exploratory interviews or a new question, and deductive coding for a tracker that asks the same question every quarter. In our experience, hybrid qualitative coding is the right default: fixed codes keep waves comparable, and a watched Other bucket catches what is new.
How do you code qualitative data step by step?
- Write down the question the coding must answer. For example, “Why do customers cancel, and does it differ by plan?”
- Read a sample first. Read 50 to 100 responses to learn the language people use.
- Draft the codebook. Start from known reasons, from what the sample showed, or both.
- Do first-cycle coding. Assign every code that fits each response.
- Check agreement. Have a second person code the same subset and compare.
- Do second-cycle coding. Merge, split and rename codes, then group them into categories and themes.
- Count and cut. Count responses per code, compare segments and pull quotes per theme.
First-cycle coding labels the raw text, staying close to what people said. Second-cycle coding works on the codes themselves, grouping “late box” and “missing items” under a category such as delivery reliability. For the full survey workflow, see how to analyze open-ended survey responses.
How do you build a codebook for qualitative coding?
A codebook is the rulebook for qualitative coding: it lets two people code the same text the same way. Each code needs a short name, a one-line definition, include and exclude rules for edge cases, and an example quote. Here is a five-code template for a meal-kit cancellation survey.
| Code | Definition | Include | Exclude | Example quote |
|---|---|---|---|---|
| Price and value | Cost too high for what you get | Discount ending, portion size | Quality complaints alone | ”Cheaper to just buy the groceries myself.” |
| Delivery | Problems getting a complete box on time | Late or damaged boxes, missing items | Delivery area (code Other) | “Missing ingredients twice in a month.” |
| Food quality | The food itself disappoints | Freshness, taste, bruising | Portion size (code Price and value) | “The produce was often bruised.” |
| Variety | Menu feels repetitive or limited | Same recipes, few diet options | Recipe difficulty | ”Same five recipes every week.” |
| Time and effort | Cooking takes more time or skill than expected | Prep time, complex steps | Delivery wait | ”Cooking took 45 minutes, not the 20 on the card.” |
The exclude column does most of the work. Without it, “portions too small” could land in Price and value or Food quality, depending on who is coding.
What does qualitative coding look like in a worked example?
These responses are illustrative, not from a real study. Ten customers answered “What is the main reason you cancelled?” We coded them with the codebook above, then inductively with no list.
| # | Response | Deductive codes | Inductive codes |
|---|---|---|---|
| 1 | ”Too expensive for what you get, especially once the discount ended.” | Price and value | Discount ended |
| 2 | ”Boxes kept arriving a day late, so the fish wasn’t fresh.” | Delivery; Food quality | Late delivery; Not fresh |
| 3 | ”Same five recipes every week. Got bored.” | Variety | Repetitive menu |
| 4 | ”Cooking took 45 minutes, not the 20 on the card.” | Time and effort | Longer than promised |
| 5 | ”So much plastic and ice packs. I felt guilty every week.” | Other | Packaging waste |
| 6 | ”Cheaper to just buy the groceries myself.” | Price and value | Cheaper to shop myself |
| 7 | ”Missing ingredients twice in a month.” | Delivery | Missing items |
| 8 | ”Portions too small for two adults, and pricey.” | Price and value | Small portions |
| 9 | ”We moved and they don’t deliver to our new area.” | Other | Moved out of area |
| 10 | ”Recipes were too complicated for weeknights and the produce was often bruised.” | Time and effort; Food quality | Too complex; Not fresh |
Responses 2 and 10 each carry two codes. That is multi-coding: one answer can hold two reasons, and you keep both.
Now count the deductive codes. Price and value appears in 3 of 10 responses (30%); Delivery, Food quality and Time and effort in 2 each (20%); Variety in 1 (10%); Other in 2 (20%).
That is 12 code assignments across 10 responses, so the percentages add up to 120%. Report them as “share of respondents mentioning,” never as a pie chart.
The inductive pass found two issues the codebook missed: packaging waste and moving out of the delivery area. The hybrid move is to add “Packaging waste” as a code and watch whether it grows.
How do you measure intercoder agreement?
Before coding all 800 responses, have a second coder code the same 10. The simplest measure is percent agreement: responses where both coders assigned the same codes, divided by the total.
Say the coders matched on 8 of the 10. Coder B coded response 7 as Food quality and added Food quality to response 8. Percent agreement is 8 ÷ 10 = 80%. Both disagreements sit on rules the codebook already states, so the fix is to discuss them and add those quotes as examples.
Percent agreement ignores agreement that happens by chance, which is common with only a few codes. Cohen’s kappa corrects for it: kappa = (observed agreement − chance agreement) ÷ (1 − chance agreement). A kappa of 1 means perfect agreement, and 0 means no better than chance. Agree on the measure and the bar you will accept before coding starts.
What are the common mistakes in qualitative coding?
- Coding before reading. The codebook fits your assumptions, not your data.
- Vague code names. “Service” or “Experience” catch everything and explain nothing.
- Forcing one code per response. Answers with two reasons undercount the second one.
- Letting Other grow unread. A large Other bucket often hides the most important new issue.
- Changing the codebook mid-wave. Recode earlier responses so counts stay comparable.
How do you put coded data to work?
For customer feedback, teams often need a continuous view of satisfaction drivers, journey friction and emerging issues. When the same codes run across several studies, connecting findings across studies and waves helps you reuse what earlier coding found. For large volumes, see how AI thematic analysis speeds up the first cycle alongside a human-checked codebook.
As a next step, take one open-ended question from your latest survey, read 50 answers, draft a five-code codebook, and ask a colleague to code 20 of them with you.