Layered quality control: why one fraud check is never enough

Fraud prevention in survey research is often sold as a single capability: a detection system, a score, a percentage. That framing is the problem. Every check has a specific blind spot, and a determined respondent only has to defeat the one you rely on.

The checks and what each one misses

• Digital fingerprinting catches the same device twice, and misses a respondent with several devices
• Speeder detection catches the careless, and misses anyone who has learned to wait
• Straight-lining detection catches the lazy pattern, and misses a randomised one
• Geolocation catches the obvious mismatch, and misses a nearby VPN
• Open-end plausibility catches the empty answer, and increasingly misses a generated one

Why layering works

The layers are useful because their blind spots do not overlap. Defeating one is straightforward. Defeating five simultaneously, consistently, across a whole interview, is enough work that the economics stop making sense for the respondent.

That is the actual mechanism: not perfect detection, but raising the cost of misrepresentation above what the incentive is worth.

The layer people skip

Human review is the layer that gets dropped first when a timeline compresses, and it is the one that catches what the automated checks were not designed for. A generated open end that scores as plausible still reads oddly to an analyst who has read four hundred responses to the same question that week.

We treat a minimum manual review share as a floor rather than a target, and report it with the data.

Report the removals

Whatever the stack, the number that matters to a client is how many completes were removed and why. A supplier reporting a two percent removal rate and a supplier reporting fourteen percent are describing different studies, and the difference is worth understanding before the data is analysed rather than after.