No dataset is better than the interviews behind it. Automated checks can show where answers look unusual, but on their own they cannot show whether an interview took place with the right respondent, or whether the questions were asked as written. Fieldwork quality control adds human verification, in person or remotely, through three complementary methods, carried out respectfully and in line with the study's consent and privacy protocol.
Accompanied interviews
A supervisor sits in on an interview and observes how questions are read, how answers are probed and recorded, and how consent is handled. Tell the respondent who the observer is and ask whether they agree to the observer being present; on sensitive topics, or where an extra person could affect privacy or safety, observe differently or not at all. Accompanied interviews are most valuable early in fieldwork and for new interviewers, while coaching can still change habits. A short observation form keeps feedback consistent across supervisors.
Unannounced spot checks
A supervisor visits an area where an interviewer is working, without warning, to confirm that the interviewer is where the schedule says and that interviews are taking place as planned. Spot checks deter shortcuts, such as interviewing by phone or in groups when interviews should be face to face and private.
Back-checks
A separate team contacts a sample of respondents again, in person or remotely by phone where the protocol and the respondent's consent allow, and asks a few questions a second time. Comparing the two sets of answers helps assess whether the interview took place and whether key answers were recorded reliably. The comparison is evidence for review, not conclusive proof either way.
- Select respondents at random, and add targeted cases: new interviewers, interviewers whose data looks unusual and interviews flagged by automated checks.
- Choose three kinds of questions: facts that should not change, such as household size; questions that test interviewer behaviour, such as whether a question was asked; and questions that test the understanding of complex items.
- Back-check soon after the interview, so that real change between the two contacts is unlikely.
- Keep back-checkers independent of the interviewers they check. Tell interviewers that back-checks are part of the protocol, but not which interviews will be checked.
Acting on the results
Differences are signals for investigation, not evidence of misconduct. A different household size may reflect a real change, a different respondent, a misunderstanding of the membership rule or a recording error. Review differences by interviewer and by question, hear the interviewer's explanation, and only then decide what to do: retraining, revisiting households or, in serious and confirmed cases, replacing interviews. Record every decision and its reason.
An example
In the first week, back-checks find that one interviewer's households are consistently smaller than the back-check team reports. An accompanied interview, with the respondent's agreement, shows that the interviewer leaves out members who are away during the day. After a refresher on the membership rule, the difference disappears in the next week's back-checks, and the affected households are revisited.
Limitations
- Back-checks repeat only a few questions; they cannot verify every answer.
- Respondents may answer a second contact differently for legitimate reasons, especially on sensitive topics or in a different mode.
- Back-checks need consent and data protection of their own: respondents should know they may be contacted again, and back-check data needs the same care as the main survey.
Further reading: J-PAL's research resource “Data quality checks” and the World Bank DIME Wiki page “Back Checks”.