Data quality is designed in, not cleaned in. Five control points determine whether a household dataset can be trusted.
Control point 1: the instrument
Ambiguous wording produces unreliable answers no matter how well fieldwork is managed. Test every question for a single interpretation, a defined recall period and mutually exclusive response options.
Control point 2: training and certification
Do not assume attendance equals competence. Certify enumerators through practice interviews scored against a checklist, and retain a slightly larger training cohort than required so that weaker performers can be released before deployment.
Control point 3: field supervision
Supervisors should observe live interviews, not only review submissions. Observation catches protocol drift, leading questions, skipped consent, paraphrasing, that data alone cannot reveal.
Control points 4 and 5: verification and analysis-stage checks
Back-checks on a defined share of interviews, GPS plausibility review and duration outlier analysis form the verification layer. At analysis, run internal consistency checks and compare distributions against known external benchmarks where they exist.
Document every cleaning decision in a cleaning log. A dataset whose transformations cannot be reproduced is not a defensible evidence base.