A question without a substantive answer can mean very different things. The question may not have applied to this respondent, the respondent may not have known the answer or declined to give it, or the answer may be missing because of an interruption or an error. If the questionnaire does not record these cases distinctly, they all look the same in the data: an empty cell. Analysis depends on telling them apart: a refusal on an income question says something about the respondent, while “not asked” says something about the questionnaire's routing.
Four kinds of nonresponse
| Kind | What happened | How to record it |
|---|---|---|
| Not asked | Skip logic did not route the respondent to the question | Empty by design, or a dedicated “not applicable” code, explained by the skip condition |
| Don't know | The respondent was asked but could not answer | An explicit answer code |
| Refused | The respondent was asked and declined | An explicit answer code |
| Missing | An answer should exist but does not | Empty, and investigated |
Make nonresponse an answer, not a gap
- Offer “Don't know” and “Refused” as answer options wherever they are realistic, so that interviewers record what happened instead of leaving a gap or guessing.
- Make these options exclusive in multiple-choice questions: “Don't know” cannot be selected together with a substantive answer.
- Decide deliberately how these options are presented, because presentation affects how often they are chosen, and the effect differs by mode. In interviewer-administered surveys, respondents often volunteer “Don't know” even when it is not read out; many self-administered web surveys do not show it and let respondents skip instead. Pilot the approach and keep it consistent across the survey.
- Use the same codes for the same meaning throughout the questionnaire, and document them.
Special codes in numeric questions
Some surveys record “Don't know” in a numeric question with a code such as -98 or 9999. These values look like real numbers and can quietly distort averages if an analyst misses them. Where possible, record nonresponse as a separate answer next to the numeric question. If codes are used, choose values that are impossible as real answers and label them clearly in the data documentation.
Document the routing
Analysts need the skip conditions to interpret empty cells, so publish them with the data and check them against it. An empty answer to a question that should have been asked points to a recording or routing problem; an answer to a question that should have been skipped points to a routing error.
An example
Suppose 1,000 households are asked whether anyone in the household received money or goods from someone living elsewhere in the past 12 months, and only those who say yes are asked for the amount. For 700 households the answer is no, so the amount was not asked: these cells are empty by design. Of the 300 households asked, 240 give an amount, 30 answer “Don't know” and 20 “Refused”, each recorded with its own code, and 10 have no answer because the interview was interrupted. Only those 10 are truly missing; the 700 empty cells are explained by the routing, and the 50 coded answers are informative nonresponse.
Limitations
Explicit nonresponse options do not reduce nonresponse; they make it visible. A high “Don't know” rate on one question can signal a wording or translation problem that is worth investigating in a pilot.
Further reading: Pew Research Center's methods notes on how survey mode affects “Don't know” answers.