Advanced · Data and analysis · module 138 of 155
Analysing surveys and open answers
Let AI sort open answers into themes, and check that classification with a sample and an honest count.
After this lesson
You prepare survey data in a privacy-friendly way, let AI suggest themes and check the classification and the conclusions.
In view · Step by step
From open answers to fair figures
Names and recognisable details removed, the original stored safely.
Themes with examples, checked against twenty answers.
Numbers per theme, and stating who they refer to.
Part 1 of 3 · Prepare
Your preference for less movement is on. Choose each part yourself.
Prepare the data first
Open answers often contain more than you think: names, contact details or details from which someone can be identified. Remove them before you put answers into an AI tool, and use a tool your organisation approves (lessons 5 and 117). Watch out for combinations: age, place of residence and job title can together point to one person. Keep a copy of the original answers in a safe place, so you can verify the classification later.
Suggesting and checking themes
Ask AI to sort the answers into a limited number of themes, with a short description and two verbatim examples for each theme. Such a list of themes is called a codebook. Check the codebook with a sample: read twenty answers yourself and see whether you would choose the same themes. An answer can touch on several themes, and irony or a negation is sometimes misread. Adjust the codebook and run the classification again.
Count honestly and draw cautious conclusions
Count how many answers fall under each theme; do not let AI estimate percentages without a count. State how many people completed the survey and how many gave an open answer. The people who respond are not always representative: satisfied or, on the contrary, dissatisfied people respond more often. So write “of the 84 people who gave an open answer, 31 mentioned the car park”, and not “most visitors find the car park a problem” (lesson 129).
New practice example
From explanation to a usable result.
AI can sort open answers into themes. You take care of privacy, check the grouping and count honestly.
Fictional practice situationFictional: a school surveyed parents about the new school report. The open answers include sentences such as: “As Lotte’s mum from 3B, I find the comments too short.”
View a possible solution
This is how it can be done
For the analysis, I replace names and classes with [child] and [class]: “As a parent, I find the comments too short.” The original is kept safely by the school management. Then I have AI sort the answers into themes and check twenty answers myself.
Why this works
The analysis needs the content, not who wrote it. Without names and classes, the risk is smaller, and the themes remain usable.
Make it slightly more difficult
Which combination of data in a survey from your own organisation could make someone identifiable?
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
The analysis needs the content, not who wrote it. Without names and classes, the risk is smaller, and the themes remain usable.
A step further
Which combination of data in a survey from your own organisation could make someone identifiable?
Try & discover · no obligation
Choose which data to omit
View the fictitious data. Omit sensitive practice fields and retain only what the task requires.
Fictional survey row with practice values. Your task is to sort the open answers into themes; who gave the answer is not needed for that. Note: a combination of data can also make someone recognisable.
What would you include as input?
Name: Fictional respondent Bram Voorbeeld Email address: bram@example.invalid Age, street and occupation: 52 years old, Kerkstraat, the only pharmacist in the village Open answer: The car park was too small and the music started too late.
Still visible in this exercise: name, email address, age, street and occupation. Are those fields necessary for the general task?
Omitting or masking is not proof of anonymity. This fixed exercise does not check real personal data, files, or policies.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
Below are fictional, anonymised open answers to the question “What could be better about our neighbourhood party?”. Propose a codebook of at most six themes, with a description and two verbatim examples for each theme. Then assign each answer to one or more themes and count per theme. Do not estimate percentages without a count. [Paste the answers here]
Apply it
An AI tool summarised 300 open answers as: “Most respondents are very satisfied.” Nothing was counted and nobody reread the answers.
Which approach best fits this situation?
- Have the answers categorised using a codebook, check a sample yourself and count per theme before you draw a conclusion.
- Copy the summary into the report.
- Ask the tool to phrase the summary more positively.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
Of 400 participants, 50 gave an open answer, and 20 of those complained about the car park. Which conclusion is justified?
- Most participants found the car park a problem.
- 20 of the 50 people who gave an open answer mentioned the car park; that is a signal to look into further.
- 40% of all participants complained about the car park.
Extra practice
Take ten fictional or anonymised open answers. Have AI create a codebook, classify five answers yourself and compare with AI’s classification.
Codebook: car park (too few spaces), music (too loud or too quiet), food (choice and waiting time), children (activities), communication (announced late or unclearly). Sample: AI put the answer “the music wasn’t too loud, finally” under music as a complaint; I classified it as a compliment. Count: 84 open answers, of which 31 mentioned the car park.
