Field Data Quality Checks: Back-checks, Audio Audits, GPS
How to check field survey data quality: back-checks, audio audits and GPS
Good field data quality comes from three layers: rules inside the form that block impossible answers, daily checks on what comes in, and independent verification of a sample through back-checks, audio audits and GPS. Most NGO surveys have the first layer and skip the other two. This guide shows how to add them without a research-firm budget. If you are still choosing a tool, start with our 2026 guide to offline data collection apps for NGOs, which covers which tools support these checks out of the box.
Layer 1: rules inside the form
Catch errors before they are saved:
- Range constraints: age 0–110, household size 1–30, no future dates.
- Required GPS at the start of the interview, with a minimum accuracy (for example, reject readings worse than 30 metres; tune this to your terrain).
- Consistency checks: a 14-year-old cannot be listed as a widow with three children without a confirmation prompt.
- Required consent before any personal question, recorded in the submission.
Layer 2: daily high-frequency checks
Every morning, someone in the office reviews yesterday’s data by enumerator. The useful flags:
| Check | What to flag | Why it matters |
|---|---|---|
| Interview duration | Under half the median for the form | Questions skipped or invented |
| "Don't know" / "Refused" rate | Enumerator rate double the team average | Shortcutting long modules |
| GPS distance | More than 2 km from the assigned village centroid | Interview not done on site |
| Interviews per day | Far above the team's normal pace | Rushing or fabrication |
| Identical answer patterns | Same answers across a long grid | Straight-lining |
| Submission time | Interviews at 11 pm or overlapping start times | Forms filled after the fact |
Most XLSForm tools record start and end times and GPS automatically. SurveyCTO adds built-in monitoring; with Kobo or ODK you can run the same checks in a spreadsheet or a dashboard on the exported data.
Layer 3: independent verification
Back-checks. A separate team revisits or phones a sample of respondents and repeats a short set of questions. The World Bank DIME wiki recommends back-checking 10–20% of observations and completing 20% of back-checks within the first two weeks, so problems surface while you can still retrain. Each back-check should take 10–15 minutes. DIME groups back-check questions into three types:
- Type 1, facts that should match exactly (household head’s name, dwelling type). Repeated mismatches can lead to action against an enumerator.
- Type 2, questions with tricky skip logic. Mismatches point to a training need, not misconduct.
- Type 3, key outcome questions, to test whether answers are stable. Mismatches point to a questionnaire problem.
Keep back-checkers independent of the survey team, and do not tell enumerators which questions will be checked.
Audio audits. The app records part of an interview silently, with the respondent’s consent. SurveyCTO’s guidance on audio audits suggests setting a probability that a recording happens rather than recording everything, recording only key sections such as the consent script, and mentioning audio audits in your ethics submission and consent statement. Recording every interview fills devices and makes uploads expensive on weak networks. Budget time for a reviewer who speaks the interview language; unheard recordings are wasted storage. Under India’s DPDP Act a voice recording is personal data, so state it in the consent notice and set a deletion date.
GPS and photos. A GPS point proves the device was somewhere, not that the right household was interviewed. Pair it with a photo of the house exterior (never of people without consent) for a small sample, or with a check against the household’s location from the listing exercise.
Worked example: a quality plan for a 30-enumerator household survey
Say your nonprofit runs a livelihoods baseline: 30 enumerators, 3,000 households over four weeks, interviews in Marathi and Hindi.
| Element | Plan |
|---|---|
| Back-check sample | 10% = 300 households, stratified so every enumerator has at least 8 back-checked |
| Early back-checks | 60 of the 300 (20%) done in week 1–2 |
| Back-check team | 3 people not involved in the main survey, mix of phone and revisit |
| Back-check form | 12 questions: 5 Type 1, 4 Type 2, 3 Type 3; rotated weekly |
| Audio audits | 15% probability, consent and income module only, 3–4 minutes each |
| Audio review | 1 Marathi/Hindi-speaking reviewer, 2 hours a day |
| Daily checks | Duration, don't-know rate, GPS distance, interviews per day, by enumerator |
| Escalation | 2 flags in a week = supervisor accompanies the enumerator; Type 1 mismatch above 10% = formal review |
The total cost is roughly three extra staff for a month plus a reviewer’s time, usually a small share of a baseline budget. The return is a dataset you can defend to a funder or evaluator.
Common mistakes
- Back-checking only the “good” villages because they are easy to reach. Stratify by enumerator and area.
- Waiting until the end to compare back-checks. By then, the bad data is the dataset.
- Treating every mismatch as cheating. Type 2 and Type 3 mismatches usually mean the form or the training needs work.
- Collecting audio with no plan to listen.
Quality checks also depend on clean sync. Duplicate submissions inflate counts and confuse back-check samples; our guide to offline sync conflicts and how to prevent them covers that side. And because recordings, GPS and photos are all personal data, check your set-up against the DPDP compliance checklist for NGOs.
Not sure your field set-up will survive the next survey round? Take the free nonprofit technology assessment. Start the free technology assessment →