Data quality
We tell you who we removed, and why
Quality claims are cheap because nobody checks them. A rejection log is checkable: it names every respondent taken out of your file, the rule that removed them and the moment it fired. You can reconcile it against your own counts, which is the only way to know whether quality meant anything on a given project.
Removing bad data is the easy half. The hard half is proving to a buyer that the removals were real, that they happened before delivery rather than after a complaint, and that the rate is reported the same way on the projects where it looks good and the projects where it does not.
Where a respondent can be removed
- Before the invitation Duplicate accounts, device clusters, and members whose profile was edited immediately before a high-paying study never receive the invitation at all. Nothing about them reaches your file, and nothing about them appears in your costs.
- At the screener The screener re-asks the attribute the study depends on. A stored profile that disagrees with the live answer ends the session and records the contradiction against the member.
- During the questionnaire Completion time under a threshold derived from the study’s own median, straight-lining down a matrix, and failed attention checks. These are judged against the study, not against a fixed rule, because a ten-minute survey and a thirty-minute one have different floors.
- At settlement Open-ended answers that repeat across sessions, and sessions whose callback signature does not verify. Removals here are the ones a buyer would otherwise never see.
- After delivery, on request If your own processing flags a respondent we passed, tell us. We reverse the payment, credit the complete and record the reason, and the case feeds the member’s quality score.
What the log contains
| Field | Why it is there |
|---|---|
| Respondent id | Lets you match the removal to a row you can see in your own export |
| Stage | Which of the five points above the removal happened at |
| Rule | The named check that fired, not a generic label like “quality” |
| Observed value | The measurement that triggered it, so you can judge the threshold yourself |
| Timestamp | Proves the removal preceded delivery rather than following a complaint |
| Source | Whether the respondent came from our own recruitment or partner supply |
What we do not claim
No screening catches everything, and any provider who says otherwise is describing a marketing position rather than a pipeline. Determined fraud adapts, and the checks that catch it today are public knowledge tomorrow.
What can be promised is narrower and more useful: the checks are named, the thresholds are visible, the removals are logged, and the rate is published whether or not it flatters us.