Skip to main content
Back to articles
The Voxaa blog

Quality control for survey data: our approach

QualityJune 20269 min read

Survey data is not reliable just because it is complete. We score every interview on a bundle of signals, and we set aside whatever falls below the threshold, before the analysis, not during it.

The signals we track

Latency per question, spread of answers, consistency between linked questions, the answer to the trap question, and total duration against the fieldwork median. None of these is enough on its own: a fast respondent is not necessarily careless, but a fast, monotonous and inconsistent one is.

From signal to score, from score to decision

Each signal feeds a score per interview. Above the threshold, the interview enters the analysis; below it, the interview is set aside and replaced within the same stratum. The threshold is fixed before fieldwork and documented in the report: moving it afterwards amounts to choosing your results.

What we do not do

We do not correct an answer we find improbable, and we do not impute missing values on opinion questions. A rejected interview is removed, never repaired: repairing means substituting the analyst's assumption for the respondent's statement.

Key takeaways

  • No signal is enough on its own; the bundle decides.
  • The threshold is set before fieldwork and published.
  • A doubtful interview is set aside, never corrected.

Read next

Get our next articles

One email a month: methods, figures and field notes, no advertising.