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.
F7 · STATS19 COLLISION VALIDATION · ENGLAND & WALES
Public transport access-site density improves collision-frequency prediction beyond the frozen baseline. Heavy-rain climatology remains uncertain. All eleven planned candidates stay visible.
Retrospective STATS19 development evidence, using distinct motor-involved police-recorded injury collisions per 100,000 motor miles. National bands stay fixed across all 35,672 LSOAs. These results do not establish insurer-claims value, a commercial index or a causal effect.
| Feature | 2023 gain | Geographic gain | Mean gain | Regions positive | FDR q | Decision |
|---|---|---|---|---|---|---|
| Public transport access-site density near roads | 3.56% | 3.43% | 3.49% | 9/10 | <0.0001 | retained for joint testing |
| Road-weighted heavy-rain days | 0.24% | -0.69% | -0.23% | 5/10 | 0.0745 | uncertain |
Access-site density passes all declared shortlist gates, including positive alternative-mileage and matched counted-route checks. Heavy-rain days do not pass the FDR, geographic, interval, mean-gain or exposure-sensitivity requirements. A small p-value alone cannot put a feature on the shortlist.
The common-cohort comparison uses 32,789 LSOAs and 65,578 observations in 2022–2023. Baseline and single-feature models are refitted on exactly the same rows. National bands are unchanged.
| Feature | 2023 gain | Geographic gain | Mean gain | 2023 O/E | Geographic O/E |
|---|---|---|---|---|---|
| Public transport access-site density near roads | 3.57% | 3.45% | 3.51% | 0.951 | 1.135 |
| Road-weighted heavy-rain days | 0.24% | -0.73% | -0.25% | 0.950 | 1.129 |
Geographic O/E remains above 1: the access-site model predicts fewer collisions than observed in pooled geographic holdouts. Better relative deviance does not resolve the baseline’s calibration limitations.
Ranked by the equal-weight mean of temporal and pooled geographic primary gains. Original Phase 5 results and family-specific q-values are preserved. Cohorts differ across features, so small rank differences are descriptive. The common-cohort comparison above covers the two F7 additions.
| Rank | Feature | Release / FDR family | Development rows | 2023 gain | Geographic gain | Mean gain | Decision |
|---|---|---|---|---|---|---|---|
| 1 | Car/van driver commuting share | Phase5 · 28 tests | 66,216 | 5.44% | 13.64% | 9.54% | retained for joint testing |
| 2 | Public transport connectivity | Phase5 · 28 tests | 66,216 | 5.64% | 12.99% | 9.32% | retained for joint testing |
| 3 | Walking connectivity | Phase5 · 28 tests | 66,216 | 6.96% | 8.66% | 7.81% | retained for joint testing |
| 4 | Households without a car or van | Phase5 · 28 tests | 66,216 | 7.05% | 7.49% | 7.27% | retained for joint testing |
| 5 | Overcrowded households | Phase5 · 28 tests | 66,216 | 3.61% | 10.28% | 6.94% | retained for joint testing |
| 6 | Cars/vans per household · lower bound | Phase5 · 28 tests | 66,216 | 6.66% | 7.21% | 6.94% | retained for joint testing |
| 7 | Households with 2+ cars or vans | Phase5 · 28 tests | 66,216 | 6.28% | 7.26% | 6.77% | retained for joint testing |
| 8 | Flats and converted/commercial accommodation | Phase5 · 28 tests | 66,216 | 3.71% | 8.73% | 6.22% | retained for joint testing |
| 9 | Cycling connectivity | Phase5 · 28 tests | 66,216 | 4.45% | 7.87% | 6.16% | retained for joint testing |
| 10 | Households with 3+ cars or vans | Phase5 · 28 tests | 66,216 | 4.29% | 4.56% | 4.43% | retained for joint testing |
| 11 | Owner occupied households | Phase5 · 28 tests | 66,216 | 4.10% | 4.61% | 4.35% | retained for joint testing |
| 12 | Public transport access-site density near roads | F7 · 11 tests | 66,170 | 3.56% | 3.43% | 3.49% | retained for joint testing |
| 13 | Driving connectivity | Phase5 · 28 tests | 66,216 | 2.63% | 4.02% | 3.33% | retained for joint testing |
| 14 | Unemployment among economically active adults | Phase5 · 28 tests | 66,216 | 3.03% | 2.66% | 2.84% | retained for joint testing |
| 15 | Private rented households | Phase5 · 28 tests | 66,216 | 3.10% | 2.10% | 2.60% | retained for joint testing |
| 16 | Road-weighted air-frost days | Phase5 · 28 tests | 65,622 | 1.54% | 1.59% | 1.57% | uncertain |
| 17 | Social rented households | Phase5 · 28 tests | 66,216 | 1.07% | 1.94% | 1.50% | retained for joint testing |
| 18 | Road-weighted summer days >25°C | Phase5 · 28 tests | 65,622 | 1.06% | 1.85% | 1.46% | uncertain |
| 19 | Commutes of 20 km or more | Phase5 · 28 tests | 66,216 | 0.70% | 1.39% | 1.04% | uncertain |
| 20 | Walking commuting share | Phase5 · 28 tests | 66,216 | 2.47% | -0.70% | 0.88% | uncertain |
| 21 | Road-weighted NO2 | Phase5 · 28 tests | 66,216 | 0.99% | 0.44% | 0.72% | uncertain |
| 22 | Cycling commuting share | Phase5 · 28 tests | 66,216 | 0.32% | 0.73% | 0.52% | uncertain |
| 23 | No qualifications | Phase5 · 28 tests | 66,216 | 2.95% | -2.41% | 0.27% | uncertain |
| 24 | Road-weighted heavy-rain days | F7 · 11 tests | 65,622 | 0.24% | -0.69% | -0.23% | uncertain |
| 25 | Terraced housing | Phase5 · 28 tests | 66,216 | 0.04% | -0.53% | -0.24% | uncertain |
| 26 | Bankruptcy rate · authority context | Phase5 · 28 tests | 66,216 | 0.37% | -1.38% | -0.51% | uncertain |
| 27 | Level 4+ qualifications | Phase5 · 28 tests | 66,216 | 1.07% | -2.50% | -0.71% | uncertain |
| 28 | Personal insolvency rate · authority context | Phase5 · 28 tests | 66,216 | 0.51% | -3.18% | -1.33% | uncertain |
| 29 | Working mainly at or from home | Phase5 · 28 tests | 66,216 | 2.10% | -4.81% | -1.35% | uncertain |
| 30 | Road-weighted PM2.5 | Phase5 · 28 tests | 66,216 | 0.54% | -3.43% | -1.45% | uncertain |
Development fits use 2022–2023: fit 2022 and evaluate 2023, plus ten leave-one-region-out folds across nine English regions and Wales. The richer Phase 4 Poisson baseline is refitted within every fold, on the candidate’s eligible observations, with log(miles / 100,000) as the offset. Eligible zero-collision rows remain.
Benjamini–Hochberg adjustment includes all eleven primary F7 tests. Nine untested candidates receive conservative p = 1 placeholders; these are not measured null results. Original test families remain unchanged. The combined 39-candidate FDR sensitivity is deferred until all eleven additions are available.
Coefficient intervals use clustered covariance. Gain intervals resample 1,000 fixed-prediction geographic blocks and omit model refitting, selection and exposure uncertainty. They do not meet the formal promotion gate.
The 2026 access-site snapshot is retrospective context, not service frequency or passenger flow. Heavy-rain days are a 1991–2020 climate normal, not observed annual weather. Mileage remains provisional. Previously inspected 2024 and 2025 results are diagnostics only and cannot select features or forms.
Evidence cards include adjusted band relativities, raw collision rates, intervals, subgroup support, exposure sensitivities and decision gates. Untested features receive no estimated collision-risk direction.
| Feature | Decision | National banded LSOAs | 2023 gain | Spatial gain | Positive regions | Primary q |
|---|---|---|---|---|---|---|
| Traffic and activity · F7 eleven-candidate family | Not testedDeferred source | 0 | — | — | — | Not tested |
| Traffic and activity · F7 eleven-candidate family | retained for joint testing | 35,646 | 3.56% | 3.43% | 9/10 | <0.0001 |
| Socio-demographic · F7 eleven-candidate family | Not testedBands withheld | 0 | — | — | — | Not tested |
| Socio-demographic · F7 eleven-candidate family | Not testedBands withheld | 0 | — | — | — | Not tested |
| Road layout · F7 eleven-candidate family | Not testedBands withheld | 0 | — | — | — | Not tested |
| Road layout · F7 eleven-candidate family | Not testedBands withheld | 0 | — | — | — | Not tested |
| Weather · F7 eleven-candidate family | Not testedDeferred source | 0 | — | — | — | Not tested |
| Weather · F7 eleven-candidate family | Not testedDeferred source | 0 | — | — | — | Not tested |
| Weather · F7 eleven-candidate family | uncertain | 35,365 | 0.24% | -0.69% | 5/10 | 0.0745 |
| Weather · F7 eleven-candidate family | Not testedDeferred source | 0 | — | — | — | Not tested |
| Weather · F7 eleven-candidate family | Not testedDeferred source | 0 | — | — | — | Not tested |
Count of eligible unique physical sites in the LSOA within100m of eligible roads / eligible road km in the LSOA.
Source period: NaPTAN 13 October 2026; eligible OS roads April 2026 · Unit: sites per eligible road km
All declared research shortlist gates passed; F8 joint incremental value remains untested.
Retrospective physical access-site density, not service frequency, ridership or pedestrian flow. F6 source and national band gates passed; 26 LSOAs remain unavailable. F7 tests a 2026 snapshot against earlier collisions, not an as-of-2022 forecast.
135,811 eligible annual observations; 66,170 in development. 92 otherwise eligible rows lack this feature. Every national LSOA remains in the frozen feature lookup.
| Representation | 2023 next-year gain | Held-out region gain | 2024 diagnostic gain | 2025 exploratory gain |
|---|---|---|---|---|
| Frozen categorical bands | 3.56% | 3.43% | 3.50% | 3.49% |
| Continuous linear | 3.73% | 3.82% | 3.70% | 3.74% |
| Restricted cubic spline | 3.74% | 3.80% | 3.67% | 3.74% |
Best average development representation: linear. Later-year results do not change the decision. All source editions are retrospective; these are not forecasts using only information available in 2022.
| Band | National LSOAs | 2023 collisions / miles | 2023 frequency / 100k | Adjusted RR (95% interval) | Effect sample support |
|---|---|---|---|---|---|
| D1 | 3,565 | 8,756 / 45,184,248,522 | 0.0194 | 0.756 (0.709–0.806) | 3,339 LSOAs; 17,722 collisions; 158 blocks |
| D2 | 3,565 | 10,106 / 52,712,497,559 | 0.0192 | 0.810 (0.769–0.854) | 3,323 LSOAs; 20,371 collisions; 161 blocks |
| D3 | 3,564 | 9,351 / 38,481,297,459 | 0.0243 | 0.909 (0.869–0.949) | 3,280 LSOAs; 18,832 collisions; 162 blocks |
| D4 | 3,565 | 8,468 / 29,860,694,344 | 0.0284 | 0.955 (0.917–0.995) | 3,311 LSOAs; 17,243 collisions; 161 blocks |
| D5 | 3,564 | 8,000 / 23,663,679,730 | 0.0338 | 1.000 (1.000–1.000) | 3,302 LSOAs; 16,381 collisions; 161 blocks |
| D6 | 3,565 | 8,404 / 21,068,374,695 | 0.0399 | 1.060 (1.022–1.100) | 3,309 LSOAs; 17,070 collisions; 162 blocks |
| D7 | 3,565 | 8,526 / 18,718,490,564 | 0.0455 | 1.124 (1.079–1.171) | 3,293 LSOAs; 16,995 collisions; 161 blocks |
| D8 | 3,564 | 8,767 / 16,801,764,636 | 0.0522 | 1.163 (1.111–1.217) | 3,304 LSOAs; 17,985 collisions; 160 blocks |
| D9 | 3,565 | 9,592 / 15,539,925,786 | 0.0617 | 1.268 (1.214–1.323) | 3,291 LSOAs; 19,306 collisions; 159 blocks |
| D10 | 3,564 | 12,648 / 13,781,546,391 | 0.0918 | 1.492 (1.425–1.562) | 3,333 LSOAs; 25,525 collisions; 158 blocks |
| 2023 next-year check | Deviance gain | Conditional 95% interval | O/E | Rows |
|---|---|---|---|---|
| Fitted-flow mileage | 3.56% | 2.89% to 4.18% | 0.951 | 33,085 |
| Traffic-group mileage | 4.43% | 3.67% to 5.13% | 0.954 | 33,085 |
| Road-length mileage | 4.52% | 3.76% to 5.26% | 0.956 | 33,085 |
| Counted major routes | 2.33% | 1.33% to 3.38% | 0.966 | 5,091 |
Intervals resample 1,000 fixed-prediction spatial blocks. They omit refitting, selection and exposure uncertainty, so they do not satisfy the E8 promotion gate. Subgroup effects are exploratory; only the primary categorical test enters its declared FDR family.
All fitted model parameters · Paired validation block totals · Original feature source
| Download | Contents | Size |
|---|---|---|
| Expanded feature ranking | 30 novel candidates ranked by mean primary gain; ten baseline form checks separate, with original families preserved. | 0.02 MB |
| All eleven F7 decisions | Coverage, primary p/q, gains, testing status and research disposition. | 0.00 MB |
| All F7 predictive comparisons | Primary and secondary forms, geographic and temporal gains, conditional intervals and calibration. | 0.02 MB |
| Same-observation comparison | The two new features refit on exactly the same development observations. | 0.02 MB |
| Eligibility and missingness | Feature support and exclusions by year, region and urban/rural context. | 0.01 MB |
| Adjusted band relativities | Development, annual, regional and urban/rural effects, cluster intervals and support. | 0.02 MB |
| Raw band frequencies | Summed collisions and miles, national membership and frequency per 100,000 miles. | 0.01 MB |
| Held-out calibration by subgroup | Year, geography, urban/rural, observed-mileage subset and frozen band. | 0.03 MB |
| Common-cohort membership | Exact LSOA-year keys for the same-observation comparison. | 0.15 MB |
| Complete F7 reproduction kit | Pinned inputs and executable code to refit all 134 models and verify exact output hashes. | 50.02 MB |
Frozen F7 protocolMethods & reproductionPinned inputsOutput checksumsVerification record
F8 joint testing and F9 report refresh are not part of this release. Statistical tests, exact numerical reproduction, rendered-content checks and the production build are recorded. Browser interaction QA was unavailable in this build environment.