Quality begins with the instrument
Many field errors originate in questionnaire design rather than enumerator behaviour. Ambiguous wording, overlapping response categories, inconsistent recall periods and unnecessary complexity make reliable interviewing difficult even for experienced teams.
Before fieldwork, the instrument should be reviewed for logic, respondent burden, translation, variable naming, validation rules and alignment with the analysis plan. Every question should have a clear reason for being collected.
Training should create consistent interpretation
Large teams can introduce interviewer variation if training focuses only on reading the questionnaire. Enumerators need shared definitions, examples and decision rules for difficult cases. Supervisors should be trained on the same interpretations and on how to resolve questions consistently.
Standardised exercises, role plays and certification or readiness checks can identify who needs additional coaching before deployment. This is particularly important when teams are recruited across several regions.
Pilot data should be analysed, not only discussed
A pilot is most useful when the team reviews the actual pilot dataset. Interview duration, missing values, response distributions, skip patterns and interviewer notes can reveal problems that are not obvious during a training-room debrief.
The pilot should end with a documented list of changes to the form, manual, translations and field procedures. Once full data collection starts, uncontrolled changes should be avoided.
Use layered supervision
Quality control is stronger when responsibility is distributed across enumerators, field supervisors, coordinators and a central data-quality function. Supervisors can observe interviews and manage sampling, while central reviewers examine patterns across teams and locations.
Layered supervision also reduces dependence on one person’s judgement. Clear escalation rules help the team decide when a record should be corrected, re-contacted, flagged or excluded.
Monitor interviewer-level patterns
Large datasets make it possible to compare interviewers on indicators such as completion, duration, missingness, use of ‘other’ categories, repeated values and unusual distributions. These indicators can identify where coaching or verification is needed.
They should be interpreted carefully. An interviewer assigned to a remote or unusual population may legitimately have different data. Pattern monitoring is a diagnostic tool, not an automatic fraud detector.
Document cleaning decisions
Final data cleaning should preserve an audit trail. Recodes, derived indicators, outlier decisions, exclusions and corrections should be reproducible through syntax, scripts or a clear cleaning log rather than undocumented manual editing.
Clients should receive enough documentation to understand how raw field records became the analytical dataset. This is essential when findings will be reviewed by donors, auditors, researchers or future evaluation teams.
Separate correction from fabrication risk
When an error is identified, teams need a rule for what can be corrected and on what evidence. Some mistakes can be resolved through a documented respondent callback, supervisor verification or reference to a source record. Other values should remain missing or flagged if no defensible correction is possible.
The goal of cleaning is not to make the dataset look perfect. It is to make decisions transparent and preserve the distinction between observed information, verified correction and analyst-created variables.
Quality metrics should be reviewed by geography as well as interviewer
Large surveys can have location-specific problems that are hidden in overall averages. One region may have higher non-response, a translation issue or a logistics problem that affects interview duration. Reviewing quality indicators by cluster, language and geography can reveal patterns that interviewer-level monitoring alone misses.
This is particularly important for multi-region studies, where operating conditions and field-team composition can vary substantially across locations.