Seventy-nine pre-specified tests. Twelve passed. Every result is recorded, including the fifty failures, and the full register is available under NDA.
Everything below is built from published NHS data. Gates were written down before the data was acquired, and the significance threshold was set at p < 0.001 in advance, stricter than convention, because a battery this size would produce about four passes by chance alone at the usual threshold.
What we found
The cost of an agency cut
When NHS trusts cut agency spend across FY 2020-21 to FY 2022-23, each £100 less agency spend was associated with about £2.70 of pay-bill-equivalent sickness absence across the six months that followed, on the whole panel. On a £100 million annual pay bill, a two point reduction corresponds to about £54,000 over those six months. That is a valuation of additional absence at pay-bill rates, not cash leaving the organisation, not a net saving, and not a demonstrated cause. Measured over twelve months instead the figure is about £73,000, but at p=0.0041 it does not meet our threshold of p < 0.001 and is reported only as a sensitivity.
The underlying coefficient: cutting agency staffing share by 1 percentage point was associated with permanent-staff sickness rising by approximately 0.054 percentage points within six months (β = -0.054, p = 0.0003, N = 519 trust-financial-year pairs across 178 trusts). Reported in the past tense deliberately: the coefficient nearly vanishes when the model is trained only on the FY 2020-21 and FY 2021-22 transitions, which we read as a COVID-era distortion. The limitation is documented, not hidden.
"That period ended in FY 2022-23. The world has changed." Correct, and it is why the claim is written in the past tense and labelled as a historical observation rather than a forecast. We also state openly that sector turnover fell sharply after 2022, mean trust turnover moving from 12.89% in FY2022 to 9.68% in FY2025, which means absolute figures from that period should not be quoted forward. What transfers is the direction and the mechanism, not the exact number. If you want a current-period position for your own organisation, that is what a Structural Review produces.
Full findings page, with the per-sector table and confidence labels →
The gap between intending to leave and leaving
Across English NHS trusts, around 22% of staff report an intention to leave their organisation. Around 10% actually go the following year. The gap sat at a median of 11.5 percentage points across 203 trusts, and a trust's position relative to its peers largely held between 2021 and 2023: rank correlation rho = 0.62 across 203 trusts (p < 0.0001). Across single years it is lower, 0.49 from 2021 to 2022 and 0.37 from 2022 to 2023. This is persistence of rank, not of the gap's absolute size.
A wide gap is not a warning about the future. It is a property of the organisation, not of any individual. The two rates come from different measurements on different bases, and this analysis does not link a survey answer to a leaving date. The gap is a trust-level measurement and nothing below that level was observed.
Distress that stopped triggering escalation
Eighteen trusts, 6.4% of the 282 trusts scored, show persistent multi-year distress in NHS Staff Survey and workforce measures without corresponding regulatory escalation, flagged in two or more of three consecutive survey years.
Scored against all English NHS trusts. Ambulance services are over-represented because sickness absence and intention to leave run higher across that sector; on a within-sector comparison they are not unusual.
The Normalised Fragility watch list contains 18 trusts, 6.4% of the 282 English NHS trusts with enough years of data to be scored, including 7 of 10 English ambulance services. These are scored against all English NHS trusts. Ambulance services are over-represented on this measure because sickness absence and intention to leave run higher across that sector as a whole; on a within-sector comparison they are not unusual. Their agency use is in fact well below the national median.
Escapes from agency dependence mostly hold, but not as often as never being there
Trusts that get out of the heaviest agency quartile are still out two years later 77.8% of the time, against 91.1% for trusts that were never in it. Most hold. The 13.3 point gap is the cost of having been there, averaged across three historical windows. If a reduction target is going into your plan, that relapse rate belongs in the plan too.
Positions persist
Flagged trusts stay flagged (rho 0.72, n 429). The normalisation gap is itself stable (0.80, n 358), as is the departure lag gap (0.57, n 283). This does not self-correct. Waiting is a decision with a cost.
What failed
A vendor with a black-box model would have hidden these. Our protocol required publishing them.
| Outcome | Count | Meaning |
|---|---|---|
| Passed | 12 | Survived a pre-specified gate at p < 0.001 |
| Failed | 50 | Did not survive. Published anyway. |
| Underpowered | 2 | Too little data to score either way |
| Not scored | 15 | Could not be run on public data, six of them blocked outright |
The named failures
- Capability Collapse by Cluster. Our fourth pre-specified measure. The doctrine direction was confirmed (d = -0.22) but the magnitude fell below the pre-specified threshold. It is not shipped as a predictor. It is documented as a watch list only.
- Geographic accessibility. Road drive-time catchments were computed for 327 NHS organisations, 212 of them trusts, and the population reachable within 30 minutes spans a 32-fold range. Tested against agency dependency on the 178 trusts with both measures: rho = -0.034, p = 0.65. We specified in advance the expectation that harder-to-reach trusts would depend more on agency staff. It does not hold across the English panel.
- Local labour market competition. Adult social care vacancy rate against agency dependency: rho = +0.004, p = 0.96. Nothing.
- Housing cost. Passed its primary gate, then failed its robustness test and was retracted the same day. Controlling for region made it non-significant. What survives is only that the Departure Lag gap is widest in London and the South East and narrowest in the North East and Yorkshire, and that is an association, not a cause.
- A single blended fragility score. Tested and failed. This is why we report three separate axes rather than one index.
What public data cannot answer
Fifteen of the seventy-nine could not be scored on public data, and six of those were blocked outright. They fail for the same reason: NHS public data does not publish tenure, age by trust, pay band in accounts, per-profession survey results, trust-level vacancies, or off-framework agency rates.
That is the clearest available statement of where public benchmarking stops and your own data begins. We will tell you which of your questions we cannot answer from outside.
The six mechanisms we test against
Six named mechanisms that explain why healthcare workforce problems persist despite intervention. Three are validated against public NHS data; three remain constructs. Every workforce challenge an organisation faces sits somewhere on this map. Trusts differ in which mechanism is running, and that difference is what we measure.
This is a map, not a sequence. We tested whether these mechanisms follow one another in time. Four of the five consecutive links were testable and all four failed. We do not claim a trust moves through them in order, and no single mechanism is the dominant one for more than 26% of organisations, however we count it: 21.6% of trust-years on the panel the measure was developed from, 24.6% across 167 trust-years from 57 organisations held out before the analysis began and examined only once afterwards, and 25.6% counted one row per organisation, a later check on the unit rather than a gate fixed in advance. That is the one finding in our register that has been checked outside the data it came from.
Agency Debt Cycle
Temporary cover becomes load-bearing, and the service stops being able to run without temporary labour. Trusts that escape heavy agency dependency fall back into it more often than trusts that were never there.
Pressure Displacement
Where agency share fell without the work being redesigned, permanent-staff sickness absence ran higher across the following six months. The two move together; the analysis does not establish that one caused the other. β = -0.054, p = 0.0003, N = 519.
Burnout Escalation Loop
Sustained pressure raises absence, absence raises pressure on those remaining. Tested for upgrade to validated status in July 2026 and did not pass.
Departure Lag
The gap between the share of staff who report an intention to leave and the share who actually go. Median 11.5 percentage points across 203 NHS trusts, and a trust's position relative to its peers largely held between 2021 and 2023, rank correlation rho = 0.62. It is a trust-level measurement and nothing below that level was observed.
Structurally Generated Turnover
Turnover produced by how work is organised rather than by individual choice. Remains a construct; three separate framings around it have been withdrawn after testing.
Normalised Fragility
Warning signs stop triggering action. The distress becomes the baseline, and the trust stops measuring itself against anything else. 18 trusts, 6.4% of the 282 trusts scored.
How we compare
NHS benchmarking tells you where you rank. Consultancies tell you what to start doing. We are the only people who will tell you, from tested evidence, what to stop doing, and show you the tests that failed on the way there.
Full comparison against national benchmarking, dashboards and consultancies →