Integrity
The known failure mode of survey research everywhere is fabrication at the enumerator level, interviews filled in at a tea stall rather than a doorstep. Every agency knows it happens and none can show a clean audit of it, because the evidence was never captured. Here it is captured by default and 3,914 interviews left this wave before anyone asked.
The second half of this screen is the more important one. A field enumerator visits once; if nobody answers, that person vanishes and their absence is invisible. This system tries up to eight times, which lifts the response rate but, far more usefully, lets us watch the estimate move as we reach harder and harder respondents. If it drifts, the drift is a direct measurement of nonresponse bias and it can be extrapolated to the people never reached. That drift is the shape of the 2024 error, and the technique is only affordable when a callback costs nothing.
| Rule | Held | Share | Catches |
|---|---|---|---|
| Duplicate respondent | 1,284 | Same sealed key completing twice in the wave window | |
| Duration below floor | 962 | Completed under 96 seconds, faster than the instrument can be read | |
| Straight-lined responses | 701 | Identical answer position across nine or more scale items | |
| Geography mismatch | 488 | Declared constituency inconsistent with routing record | |
| Third party on line | 301 | Second voice answering for the respondent | |
| Instrument deviation | 178 | Question order altered mid-call by supervisor override |
Incumbent vote share by how many attempts it took to reach the respondent. Easy contacts favour the incumbent by 4.6 points over hard ones. Extrapolated to the never-reached, the raw estimate overstates by roughly 2.4.