Loading approved evidence
Retrieving and verifying the requested analytical release. Result data loads before optional map boundaries and source checks.
Your filters and release stay in the address. Navigation remains available.
Retrieving and verifying the requested analytical release. Result data loads before optional map boundaries and source checks.
Your filters and release stay in the address. Navigation remains available.
DATA LAB · EXPLORATORY RESEARCH · V2
A restricted public-data study of injury collisions per 100,000 motor vehicle miles. Later years and unseen regions test whether added information lasts.
New decile study: Adjacent bands learned and tested · inspect the grouping and predictive results →
What do these findings mean for insurance? Read the executive interpretation and finding register →
Persistence and robustness checks are now available. Compare fixed models, alternative assumptions and regional calibration →
Models fitted · Exploratory evidence · Human actuarial review not recorded
A narrower test, with its assumptions visible. Protocol v2 uses a reduced road/traffic baseline and fixed 2011 Census context. It is separate from the original registered H1–H6 tests, which remain untested under their full specification. These results do not establish causation or insurer pricing lift.
Deprivation and qualifications together reduce later-year deviance by 11.0% beyond a baseline with past collision history. That comparison worsens by 4.3% in unseen regions. In the separate force-area study, adding burglary beyond deprivation and history worsens temporal deviance by 5.3%.
No comparison meets every predeclared exploratory criterion. The results support further investigation of deprivation and education, with regional transfer and missing road context still unresolved.
INCREMENTAL PREDICTIVE VALUE
Positive gain means lower held-out Poisson deviance. A negative result means adding the factor worsened predictions. Every comparison uses the same rows for its two models.
Baseline + deprivation compared with Context baseline.
An exploratory candidate must improve all three temporal holdouts by at least 1%, meet calibration criteria, have a positive combined uncertainty interval and transfer to at least five regions. It is not a promoted v1 finding.
Why this comparison falls short: 3 annual holdout(s) fail the gain or calibration criteria. Combined performance worsens in unseen regions. Combined observed/expected lies outside 0.90–1.10.
| Year | Gain | Observed / expected | Observed rate / 100k | Predicted rate / 100k |
|---|---|---|---|---|
| 2023 | +20.1% | 1.145 | 0.03664 | 0.03201 |
| 2024 | +30.4% | 1.140 | 0.03465 | 0.03039 |
| 2025 | +27.1% | 1.124 | 0.03408 | 0.03032 |
| Comparison | Gain | 95% interval | Supportive years | Assessment |
|---|---|---|---|---|
| Add income and employment deprivation | +25.7% | +19.7% to +30.1% | 0/3 | Not consistently supported |
| Add 2011 qualifications | +17.4% | +11.7% to +22.0% | 1/3 | Not consistently supported |
| Use Level 4+ qualifications instead | +19.8% | +12.2% to +27.8% | 2/3 | Not consistently supported |
| Add deprivation beyond collision history | +6.8% | +3.0% to +10.0% | 3/3 | Not consistently supported |
| Add qualifications beyond deprivation and history | +4.4% | +1.9% to +7.3% | 3/3 | Not consistently supported |
| Add deprivation and qualifications beyond history | +11.0% | +6.0% to +15.0% | 3/3 | Not consistently supported |
Each region is removed from all pre-2025 training, then assessed in 2025. Combined gain: -1.1%; observed/expected: 1.233. An unfamiliar region receives the average fitted regional effect.
| Excluded training region | Test areas | Gain | Observed / expected |
|---|---|---|---|
| East Midlands | 8 | +17.8% | 1.167 |
| East of England | 11 | +43.8% | 1.173 |
| London | 33 | -0.6% | 3.247 |
| North East | 10 | -135.7% | 0.534 |
| North West | 22 | -23.3% | 0.748 |
| South East | 18 | +9.3% | 1.361 |
| South West | 7 | -37.9% | 0.967 |
| West Midlands | 12 | +14.2% | 0.888 |
| Yorkshire and The Humber | 12 | +44.0% | 1.137 |
These tests use only 2018–2019 training. They are deliberately reported separately and do not satisfy the three-year primary training requirement.
| Year | Gain | Observed / expected |
|---|---|---|
| 2020 | +7.7% | 1.018 |
| 2021 | +15.5% | 1.082 |
2,000 paired bootstrap resamples of English-region blocks using frozen out-of-sample predictions. Nine authority-study blocks; intervals are conditional on fitted models, do not include source measurement error or refitting uncertainty, and are not multiplicity-adjusted.
Education, age structure and density are fixed 2011 descriptors. Qualification shares use residents aged 16+; age shares use all residents. They are not estimates of each later year’s population. Changed codes without an exact match are excluded.
Population density, age shares, lagged vehicle mix, traffic per road kilometre, region and trend enter the baseline. Detailed joint road class and urban/rural controls are missing; they could explain some apparent social signal.
Traffic, burglary and collision history use two-year lags. The archived series were retrieved in 2026, so these are retrospective tests using revised history. Exact January-origin data vintages have not been reconstructed. Actual target-year miles supply the scoring exposure.
Crime is tested in 28 fully matched whole police-force areas, with a 29-force sensitivity including Greater Manchester’s flagged recording period. Burglary counts are scaled by fixed 2011 population, with complete quarters and explicit offence categories. This measures a baseline-population intensity, not current annual crime rates or all crime. Deprivation aggregation uses resident-weighted contextual proxies.
Area context describes people living around roads; it does not identify the drivers using them. The outcome excludes damage-only incidents and insurance claims. No individual-driver education, deprivation or crime relativity follows.
| Year | Included collisions | Full-panel collisions | Coverage |
|---|---|---|---|
| 2016 | 108,377 | 123,350 | 87.9% |
| 2017 | 105,054 | 118,299 | 88.8% |
| 2018 | 99,549 | 111,978 | 88.9% |
| 2019 | 95,867 | 107,510 | 89.2% |
| 2020 | 75,509 | 84,484 | 89.4% |
| 2021 | 84,236 | 93,873 | 89.7% |
| 2022 | 88,696 | 98,546 | 90.0% |
| 2023 | 86,404 | 96,764 | 89.3% |
| 2024 | 82,941 | 93,847 | 88.4% |
| 2025 | 83,147 | 94,418 | 88.1% |
Known-mile totals exclude authorities whose mileage is unavailable and do not establish whole-England mileage coverage. The original 95% nationwide publication gate is not met.
| Authority | Code | Reason |
|---|---|---|
| Stoke-on-Trent | E06000021 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Bath and North East Somerset | E06000022 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Bristol, City of | E06000023 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| North Somerset | E06000024 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| South Gloucestershire | E06000025 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Isles of Scilly | E06000053 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Northumberland | E06000057 | No exact-code 2011 Census match |
| Bournemouth, Christchurch and Poole | E06000058 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Dorset | E06000059 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Buckinghamshire | E06000060 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| North Northamptonshire | E06000061 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| West Northamptonshire | E06000062 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Cumberland | E06000063 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Westmorland and Furness | E06000064 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| North Yorkshire | E06000065 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Somerset | E06000066 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Gateshead | E08000037 | No exact-code 2011 Census match |
| Barnsley | E08000038 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Sheffield | E08000039 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
| Staffordshire | E10000028 | Incomplete 2016–2025 eligible count/mileage/deprivation history |
Police-force membership, complete-force exclusions and source warnings are provided in the crime input audit below.
294 model fits · all numerical convergence checks passed. Protocol and exact fold membership were frozen before fitting; fitted transformations and predictions are retained. Publication of this research page does not record human actuarial approval.
Qualifications · KS501EW · Age structure · KS102EW · Population density · QS102EW
Official qualifications definitions · Home Office crime tables
Study social-exploratory-2026-10-10-v2 · base evidence lens-2026-10-09-cbec43fe76c4 · fitted 2026-10-10.