How 852 campaigns became 652 sites
Attrition from gross harvest to the locked canonical cohort
How 852 campaigns became 652 sites
852 campaigns went in. 652 survived every stage, and every figure in this series reports against those same 652.
The funnel
| Stage | Step | Removed | Left |
|---|---|---|---|
| 1 | Gross harvest | — | 852 |
| 2 | English jurisdiction only | 35 | 817 |
| 3 | IMD health-domain record present | 25 | 792 |
| 4 | Geocoded to a neighbourhood (LSOA) | 49 | 743 |
| 5 | LSOA matched to IMD File 7 boundaries | 42 | 701 |
| 6 | Repeat submissions collapsed (strict rule) | 49 | 652 |
That is 76.5% of the harvest.
Why it was rebuilt
The v6 rebuild fixed an index misjoin that had attached another site’s health and deprivation record to 583 of the earlier 641 sites. Stages 3–5 of the old cohort had been decided on the wrong sites, so every v5 figure is superseded.
What it doesn’t settle
116 sites were dropped at stages 3–5. They are not a random loss. The stage-5 sites (42) are in neighbourhoods whose boundaries were redrawn after 2011: 52% are in the South East against 28% of the sites that survived every stage, and 57% carry our “infill” label against 35% (both p < 0.001). That stage is needed only for the health measures; without it the cohort would be about 639 sites under the broader de-duplication rule. A broader de-duplication rule gives 608 sites and is kept as a sensitivity cohort.
Revised 12 November 2026: Stages are now numbered, and the title describes the funnel.
Corrected again 12 November 2026 after a check of the data pipeline: added how the sites dropped at the neighbourhood-match stage differ from the rest.
Evidence type: Direct count. Source: Manuscript ch. 5. No claim in this series is causal.