Open-Ended Response Coding

How do I build a code frame for open-ended survey responses?

Read a random or stratified sample of 100 to 200 responses before writing a single code, so the frame is not shaped by whatever appeared first in the file. Then write codes at the smallest distinct level people are actually talking about, give each one a definition, an inclusion rule, an exclusion rule and a boundary example, and net upward for reporting. Pilot the frame on fresh responses before applying it to everything.

Should I report open ends as a percentage of everyone or a percentage of those who answered?

Show both, and state which one the commentary uses. On a question with a 45 percent answer rate the two figures differ by more than a factor of two, and a reader who assumes the wrong one will misread every finding on the page. This is the commonest reporting error in open-end work, and it happens in both directions.

Can one response have more than one code?

Usually yes, and it should. Most people say more than one thing. Multi-coding means the column will sum above 100 percent, which is fine as long as you state the convention and never convert it into a "share of mentions" table, because that quietly changes the denominator from people to codes.

What do I do with blank and "n/a" responses?

Classify them rather than deleting them. A blank, an explicit refusal, "nothing to add" and a junk string are four different findings. Deleting them silently changes the denominator without a log. "Nothing to add" after a satisfaction rating is often a substantive statement about the absence of problems and belongs in the coded data.

How big can the "other" code be before there is a problem?

Under 10 percent is normal residue. Between 10 and 15 percent the frame is missing something specific and you should find it. Above 15 percent the frame is wrong and should be rebuilt from a fresh sample. A large other bucket is not a category, it is the frame telling you it was built on the wrong material or at the wrong level of granularity.

How do I keep a code frame consistent across tracker waves?

Freeze it. Add new codes beneath the existing structure and log every addition, but do not restructure because you would have built it differently. A restructured frame converts frame change into apparent market change, and the trend is usually the client's entire reason for asking. If a rebuild is genuinely necessary, declare it as a break and run both frames in parallel for at least one wave so the discontinuity is measured.

Does a low mention rate mean people are not concerned about something?

No. An unprompted open end measures what was top of mind under that specific prompt. Nine percent mentioning waiting times does not mean 91 percent are content with waiting times: it means 91 percent raised something else. Absence is not disagreement. If you need the incidence of a belief, that requires a prompted closed question.

Is coding open ends the same as thematic analysis?

No, and the difference is the material. Thematic analysis works on rich, discursive text where meaning sits in how people explain themselves, and it builds themes with organising concepts, tested across the dataset. Open-end coding works on short text where the deliverable is a distribution: what was said, by how many, out of how many. Running thematic analysis on one-line answers produces a topic list dressed as an analysis.

Can AI code open-ended responses accurately?

It codes consistently, which is not the same as correctly. Two AI passes over the same data measure stability of application, not validity, and reporting an agreement coefficient over them dresses one up as the other. The controls that matter are sequencing and audit: a human approves the frame before it is applied at scale, and a sample of the coded output is re-coded blind afterwards, with disagreement reported per code rather than overall.

How do I compute net codes correctly?

On unique respondents. A respondent who gave two price-related codes counts once in the price net. Building a net by summing its member codes overstates it, sometimes badly, and the error grows with how freely you multi-coded. Nets are reporting objects, never coding objects, and nobody should ever code directly to one.