Every survey instrument produces two data sets. The first is the substantive variable — age, education, caste, monthly income, fertility history, land ownership, opinion on the state of the ration shop — that the analysis will summarise, cross-tabulate, and model. The second is the response pattern — who answered which questions, who declined, who said "don't know," who claimed not to know something they clearly did know, at which point in the interview the respondent's willingness dropped. The first data set is what the report presents. The second is what most reports discard.
The response pattern is not noise. It is structured information about what the respondent found comfortable to answer to the enumerator who asked and to the household members present when it was asked. Studying it as data changes what a survey can tell you.
The mechanics are straightforward. Every questionnaire records item-level response codes: response given, response refused, don't know, not applicable, skip pattern completed. Most analyses collapse these into "response" or "missing" and impute the missing. A more informative treatment keeps the non-response codes as their own variable and asks the following questions in the analysis. Which items have elevated refusal rates? Which subgroups refuse those items disproportionately? Where in the questionnaire does the refusal rate rise, and does the rise correspond to a topic transition?
The patterns we have seen enough times to write down are these.
Refusal is item-specific in ways that align with topic sensitivity. Income, land ownership, household violence, sexual and reproductive health, caste of neighbours, and opinions on named political figures produce refusal rates well above the survey average. Age and education produce refusal rates near zero. The relationship between refusal rate and topic sensitivity is monotone. A survey in which the refusal rate on income is under three per cent is either a survey with an unusually cooperative sample or a survey whose income data is being confabulated by respondents who did not want to say "refused" out loud. Usually the second.
Refusal has a subgroup structure. Women in patriarchal households refuse political and financial-decision items more often when male household members are present, and refuse them less often when the interview is conducted privately. Caste-sensitive items produce different refusal patterns from upper-caste and lower-caste respondents even when the item is neutrally phrased. When the question is about whether the respondent's family has faced discrimination, the refusal pattern by caste is itself a finding, often more reliable than the substantive responses given.
Refusal accumulates within an interview. If a respondent refuses an item at question fifteen, the probability of refusal on subsequent sensitive items in the same interview rises. Interviewers who are skilled at rapport recovery reduce this drift; interviewers who move ahead through the script accelerate it. The correlation of refusal across items within an interview is a check on interviewer effects that most analyses do not do and could easily do.
The don't-know code is not the refusal code. "Don't know" is used by respondents who are unwilling to commit publicly to a position, or who are unsure whether the interview is confidential enough to permit a specific answer. Its rate rises on items that are politically or socially costly to be seen holding a position on. Treating "don't know" as identical to "refused" collapses two different behaviours; treating "don't know" as random measurement error understates its patterning.
What is the practical implication? Two things.
First, report response and refusal rates by item, subgroup, and interview position. The analytic weight of this is small; the interpretive weight is large. A published table of refusal rates by item is often more informative than the corresponding table of substantive responses, particularly on sensitive topics. Reviewers who take non-response seriously read that table first.
Second, model the response indicator alongside the substantive variable. If the outcome is presence of household violence in the last year, the analysis benefits from a joint model of the substantive response and the refusal indicator: a respondent who refused the item is not analytically equivalent to a respondent who reported no violence, and the two categories should not be pooled in the summary statistic. This is not novel; multiple-imputation with an indicator for missingness is the standard reference. It is under-used in the applied literature we read.
A note on ethics. Non-response is the respondent's exercise of a right. Analysing it should not become a device for extracting the information the respondent chose not to give. The analytic interest is in the pattern of refusal across items and subgroups, not in inferring what any individual would have said. The distinction matters. A study that treats non-response as a puzzle to be solved with imputation ignores what the respondent was communicating by refusing. A study that treats non-response as data listens to the refusal as a signal about the survey's relationship to the population.
The best survey enumerator we ever worked with told us, at the end of a long day in a difficult district, that the most useful thing she had learned was to write down which items respondents paused on before answering. The pause was almost as informative as the refusal, she said, and she was right. Timing data from digital data collection platforms — SurveyCTO, ODK, KoboToolbox — makes this available at almost no marginal cost. Most analyses discard it. Which is a shame, because the pattern of pauses at sensitive items is data, and it is often the finding.
Useful references: Little and Rubin's Statistical Analysis with Missing Data remains the standard reference on non-response modelling; Groves et al.'s Survey Methodology covers the field-level dynamics; and van Buuren's Flexible Imputation of Missing Data is the practical treatment.