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.
SOCIAL FACTORS · DECILE VALIDATION V2 · PHASES 5–7
Compare deprivation and education deciles with continuous scores, using the same injury collisions, motor mileage and assessment rows.
Phase 5 complete · Phase 6 exposure gate blocked · Phase 7 authority findings published
Combined deciles improve the pooled temporal score, but do not meet every frozen validation criterion and do not establish non-inferiority to smooth continuous scores. Keep continuous-score research comparisons and use deciles to explain held-out predictions. No rating factor is deployed.
Exploratory; no hypothesis promoted. The outcomes were inspected before this study. The interval spans no gain, 2023 fails persistence, geographic transfer fails, and seven employment-band support cells fall short. No causal effect, individual-driver claim or insurance pricing approval follows.
AUTHORITY RESEARCH · VERSION V2
Geography: 133 exact-code English authorities. Assessment: 2023–2025, using earlier non-pandemic training years. Features: 2019 deprivation and 2011 no-qualifications share. Bands: fixed custom authority display deciles; 1 highest, 10 lowest. These are separate from model-fold and national LSOA bands.
Eligible-cohort coverage for the selected period: 88.6% of target collisions and 84.3% of known target mileage. Counts include all police-reported injury collisions at road location; the denominator is all-motor vehicle-miles. The continuous model remains a research comparator, not a deployed risk score.
| Decile | Areas / area-years | Regions | Collisions | Billion miles | Observed rate [95%] | Predicted rate [95%] | Valid refits |
|---|---|---|---|---|---|---|---|
| 1 | 14 / 42 | 4 | 21,706 | 46.002 | 0.0472 [0.0466–0.0478] | 0.0483 [0.0325–0.0644] | 995 / 999 |
| 2 | 13 / 39 | 5 | 18,278 | 33.789 | 0.0541 [0.0533–0.0549] | 0.0530 [0.0275–0.1018] | 999 / 999 |
| 3 | 14 / 42 | 5 | 20,403 | 43.561 | 0.0468 [0.0462–0.0475] | 0.0464 [0.0235–0.1098] | 997 / 999 |
| 4 | 13 / 39 | 6 | 19,136 | 30.644 | 0.0624 [0.0616–0.0633] | 0.0673 [0.0333–0.1165] | 999 / 999 |
| 5 | 13 / 39 | 7 | 19,882 | 45.197 | 0.0440 [0.0434–0.0446] | 0.0459 [0.0290–0.1015] | 999 / 999 |
| 6 | 13 / 39 | 6 | 27,972 | 59.271 | 0.0472 [0.0466–0.0477] | 0.0467 [0.0262–0.0850] | 999 / 999 |
| 7 | 16 / 48 | 6 | 42,721 | 128.334 | 0.0333 [0.0330–0.0336] | 0.0330 [0.0248–0.0445] | 999 / 999 |
| 8 | 11 / 33 | 4 | 22,644 | 82.209 | 0.0275 [0.0272–0.0279] | 0.0280 [0.0171–0.0648] | 994 / 999 |
| 9 | 14 / 42 | 8 | 36,466 | 156.410 | 0.0233 [0.0231–0.0236] | 0.0199 [0.0159–0.0284] | 999 / 999 |
| 10 | 12 / 36 | 5 | 23,284 | 93.746 | 0.0248 [0.0245–0.0252] | 0.0213 [0.0189–0.0625] | 998 / 999 |
Rates use 100,000 × summed counts / summed matched miles. Predicted counts are summed over exactly the same rows. Refitted prediction intervals use 999 paired regional samples, keeping every area’s years together. Some resamples have empty bands; valid-refit counts are shown. Only nine independent regional blocks are available, and exposure measurement error is not included.
| Comparison | Temporal gain | Refitted 95% interval | Geographic gain | Positive regions | Verdict |
|---|---|---|---|---|---|
| All deciles beyond baseline | +7.73% | -1.78% to +31.10% | -39.05% | 4 / 9 | insufficient evidence |
| Deprivation beyond baseline | +8.12% | -8.55% to +22.49% | -21.15% | 4 / 9 | insufficient evidence |
| Education beyond baseline | +2.14% | -4.38% to +15.90% | -27.40% | 1 / 9 | inconsistent |
| Education added after deprivation | -0.42% | -6.68% to +19.81% | -14.78% | 1 / 9 | no incremental gain |
| Deprivation added after education | +5.72% | -3.55% to +25.39% | -9.14% | 5 / 9 | insufficient evidence |
| Deciles versus linear scores | -3.63% | -11.39% to +27.22% | -33.26% | 0 / 9 | no incremental gain |
| Deciles versus smooth scores | -1.99% | -15.89% to +30.26% | -23.59% | 1 / 9 | no incremental gain |
“Insufficient evidence” includes inadequate training support; “inconsistent” means the available tests do not meet every criterion. “No incremental gain” records a nonpositive pooled point estimate, not proof of no possible association. Secondary comparisons are exploratory and not adjusted for multiple testing.
| Requirement | Result |
|---|---|
| At least 1% pooled gain | Pass |
| Paired 95% lower bound above zero | Fail |
| At least 1% gain in every year | Fail |
| Pooled O/E 0.90–1.10 | Pass |
| Annual O/E 0.90–1.10 | Pass |
| Positive pooled geographic gain | Fail |
| Positive gain in at least five regions | Fail |
| Absolute rate error no worse | Pass |
| Every band meets training support | Fail |
| Converged, full-rank point fits | Pass |
No replacement claim is supported. Excess deviance is 1.99%; its refitted 95% interval is -30.26% to +15.89%. The registered upper-bound margin is 0.5%. The primary candidate gate must also pass, and calibration and absolute rate error must not deteriorate.
| Model | Deviance | Gain vs B | Observed / expected | Absolute rate error |
|---|---|---|---|---|
| B | 8,810.4 | 0.00% | 1.037 | 0.00497 |
| C-linear | 7,844.3 | +10.97% | 1.030 | 0.00468 |
| C-smooth | 7,970.8 | +9.53% | 1.028 | 0.00480 |
| D | 8,094.8 | +8.12% | 1.025 | 0.00486 |
| E | 8,622.0 | +2.14% | 1.031 | 0.00495 |
| DE | 8,129.1 | +7.73% | 1.027 | 0.00487 |
The selected model is C-smooth · Continuous scores · smooth. Annual training expands through the preceding year. Each regional test excludes that region entirely from training and assesses 2025; unseen-region effects use the training-only average.
| Year | Deviance | Gain vs B | Observed / expected | Absolute rate error |
|---|---|---|---|---|
| 2023 | 2,595.4 | -7.95% | 1.010 | 0.00489 |
| 2024 | 2,673.5 | +11.38% | 1.015 | 0.00470 |
| 2025 | 2,701.9 | +20.29% | 1.059 | 0.00482 |
| Held-out region | Areas | Deviance | Gain vs B | Observed / expected | Absolute rate error |
|---|---|---|---|---|---|
| East Midlands | 8 | 268.1 | +18.95% | 1.154 | 0.00442 |
| East of England | 11 | 515.7 | +40.40% | 1.246 | 0.00514 |
| London | 33 | 2,887.4 | -53.49% | 1.288 | 0.02439 |
| North East | 10 | 343.7 | -93.20% | 0.746 | 0.00751 |
| North West | 22 | 476.3 | -17.57% | 0.887 | 0.00552 |
| South East | 18 | 585.5 | +12.62% | 1.155 | 0.00504 |
| South West | 7 | 220.5 | +17.65% | 1.110 | 0.00384 |
| West Midlands | 12 | 287.6 | +13.52% | 1.010 | 0.00397 |
| Yorkshire and The Humber | 12 | 63.0 | +29.42% | 1.028 | 0.00251 |
| Sensitivity | Combined gain | Annual gains |
|---|---|---|
| poisson | +6.56% | 2023: +5.86%; 2024: +7.88%; 2025: +5.98% |
| fixed-imd2015 | +8.76% | 2023: +5.28%; 2024: +11.74%; 2025: +8.57% |
| education-level4-complement | +11.93% | 2023: +1.04%; 2024: +12.89%; 2025: +18.80% |
| include-pandemic | +7.30% | 2023: +1.65%; 2024: +8.00%; 2025: +11.23% |
| fixed-2023 | +6.69% | 2023: +0.33%; 2024: +6.25%; 2025: +11.15% |
| without-region:East Midlands | +6.75% | 2023: +3.08%; 2024: +8.73%; 2025: +7.69% |
| without-region:East of England | +5.14% | 2023: -2.97%; 2024: +7.23%; 2025: +9.21% |
| without-region:London | +12.37% | 2023: +14.18%; 2024: +14.16%; 2025: +8.49% |
| without-region:North East | +12.54% | 2023: +2.05%; 2024: +12.36%; 2025: +20.30% |
| without-region:North West | +12.48% | 2023: +7.83%; 2024: +11.86%; 2025: +15.95% |
| without-region:South East | +14.13% | 2023: -8.95%; 2024: +17.62%; 2025: +26.66% |
| without-region:South West | +11.82% | 2023: -5.82%; 2024: +15.83%; 2025: +20.08% |
| without-region:West Midlands | +8.74% | 2023: -1.71%; 2024: +10.78%; 2025: +14.81% |
| without-region:Yorkshire and The Humber | +9.71% | 2023: +2.35%; 2024: +8.94%; 2025: +15.54% |
| pandemic-stress | Unavailable | 2020: Unavailable; 2021: Unavailable |
Fixed-2023 freezes parameters and preprocessing; lagged inputs and realised mileage still update. Level 4+ uses its complement as a separate education definition. Pandemic D/DE stress fits are unavailable because the assessment vintage has no training cutpoints; the failed attempt and correction are archived. Boundary reallocation and historical rural/urban interactions remain blocked by matched-data gaps.
All 54 geographic point fits converged. The bootstrap refitted 17,982 models with 0 convergence failures; 870 bootstrap designs were rank deficient, and 995 resamples contained undersupported bands. These samples are retained with the frozen numerical ridge and disclosed; they do not repair the support gate.
| Fold | Feature | Vintage | Band | Authorities | Collisions |
|---|---|---|---|---|---|
| spatial-London | employment | 2015 | 2 | 9 | 10,709 |
| spatial-London | employment | 2015 | 5 | 9 | 7,037 |
| spatial-London | employment | 2015 | 9 | 9 | 18,783 |
| spatial-London | employment | 2019 | 9 | 9 | 27,271 |
| spatial-North West | employment | 2019 | 6 | 9 | 24,813 |
| spatial-South East | employment | 2015 | 6 | 9 | 24,987 |
| spatial-South East | employment | 2019 | 4 | 9 | 10,500 |
2027–2029 reserved prospectively; access ledger and programme-wide prior-inspection audit required before opening each year.
33,755 English LSOA21 profiles use official 2025 income/employment domain deciles and custom national 2021 education bands. The map coordinate download could not be completed, so a neighbourhood map is unavailable in this release. Exact-code profiles and all data downloads remain available.
The segment-splitting and collision-assignment pipeline is implemented and tested. It preserves missing flow, counts each event once, checks duplicate and overlapping roads, uses the correct calendar-year length, and requires independent traffic validation by road class and rurality. It has zero eligible LSOA frequency rows because the necessary real exposure inputs are absent.
| Input | Availability | What remains |
|---|---|---|
| Social profiles | available | 33,755 exact-code LSOA21 profiles; 2021 education and 2025 deprivation. Descriptive only. |
| Collision records | partial | Archived STATS19 records contain positions and published LSOA codes; unique event-to-LSOA21 allocation and positional audit remain to be run with compatible boundaries. |
| LSOA21 allocation boundaries | missing | Full compatible polygons not loaded into this exposure pipeline; population centroids are display-only. |
| Historical all-road segments | missing | The 2026 road-network snapshot used elsewhere is not a year-matched 2018–2025 network. |
| Major-road flow/link matching | partial | DfT publishes major-road link estimates; no audited link-to-LSOA mileage allocation is present. |
| Minor-road AADT | missing | DfT count points cover a sample. Unsampled-road estimates and independent holdout validation are absent. |
| Road-class × rurality totals | missing | No matched authority/year/stratum reconciliation set is present. |
| Exposure uncertainty | missing | No segment-level uncertainty bounds or independent traffic validation set is present. |
Population-centroid coordinates and descriptive polygons cannot supply vehicle-miles. A current road snapshot, population allocation or a sample of count points would not justify an all-road neighbourhood frequency layer. A major-road-only study would need a separate matched collision and mileage panel.
Protocol frozen 2026-10-10; findings computed 2026-10-10. All views use finding register social-deciles-validation-2026-10-10-v2. The original phases 1–4 release remains pinned as v1. No human actuarial review is recorded.
Read the Actuary Lens finding · Original sources, dictionary, features and hashes
Publication checks cover calculations, pinned versions, rendering, selections, downloads and source consistency. Automated browser interaction and mobile visual review were unavailable in this environment; the layout uses responsive controls and accessible chart tables.