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.
ORIGINAL PHASE 4 · ENGLAND & WALES · RESEARCH BASELINE
Geography, road context and mileage explain part of collision frequency. This phase establishes what future features must improve, and shows where the baseline still falls short.
IMPLEMENTED · CONDITIONAL RESEARCH USE
The selected model improves development Poisson deviance by 1.6% in the next-year test and 2.0% in held-out regions versus the core comparator. National feature values and deciles remain fixed. No commercial risk index is released in this phase.
Wales is overpredicted in both 2024 and 2025. London and the North East also fall outside the calibration tolerance in 2025. Residual spatial dependence remains. The mileage is provisional, and the 2025 outcomes have already been inspected. These results are development evidence, not independent validation or insurer-claims evidence.
The response is distinct police-recorded injury collisions involving a motor vehicle. Each model uses log(annual vehicle miles / 100,000) as an offset. Zero-collision rows stay in the analysis; invalid or withheld rows stay in the national register with no prediction.
Core controls cover year and region, Wales-by-year context where estimable, urban/rural status, population density, age composition, motorway/A-road shares and outcome-free mileage-quality fields.
The richer model adds nonlinear density terms, urban-density interaction, junction and roundabout density, and road sinuosity. Terrain and vehicle traffic mix remain unsupported controls because their feature candidates are deferred.
Road measures are retrospective April 2026 context. The frozen road-share features include restricted-access roads, while the mileage network is access-screened. Density and road context also enter the mileage allocation, so sensitivity tests remain essential.
Selection uses 2022–23 only: fit 2022 and predict 2023, then leave each English region or Wales out across both development years. The selected 2022–23 fit is frozen for 2024 and 2025, carrying the latest fitted year effect forward. Welsh time effects absorb changes where estimable; they do not identify the causal effect of the 20 mph policy.
| Candidate | 2023 deviance reduction | Geographic deviance reduction | NB2 dispersion α | Decision |
|---|---|---|---|---|
| Core Poisson | 0.00% | 0.00% | 0.000 | Retained comparison |
| Core negative binomial | -2.73% | -0.78% | 0.393 | Retained comparison |
| Richer road Poisson | 1.57% | 2.00% | 0.000 | Selected comparator |
| Richer road negative binomial | -1.12% | 1.60% | 0.381 | Retained comparison |
Poisson dispersion is 2.36, so a Poisson variance assumption is inadequate. Selection concerns predictive means. Coefficient intervals use authority-cluster covariance, with the Cumbria successor authorities joined. Negative-binomial comparisons and all fitted parameters are downloadable.
O/E below 1 means the model overpredicts. The shaded range is the declared 0.90–1.10 calibration tolerance. Intervals resample whole authority blocks around fixed predictions; they exclude model refitting and mileage uncertainty.
| Group | Rows | Observed | Expected | O/E | Conditional 95% interval | Within tolerance |
|---|---|---|---|---|---|---|
| East Midlands | 2,847 | 7,956 | 7,817 | 1.018 | 0.829–1.230 | Yes |
| East of England | 3,758 | 9,709 | 9,981 | 0.973 | 0.883–1.101 | Yes |
| London | 4,992 | 19,860 | 22,314 | 0.890 | 0.824–0.958 | No |
| North East | 1,682 | 2,764 | 3,253 | 0.850 | 0.745–0.937 | No |
| North West | 4,567 | 9,449 | 10,123 | 0.933 | 0.821–1.024 | Yes |
| South East | 5,237 | 14,543 | 16,068 | 0.905 | 0.778–1.010 | Yes |
| South West | 3,406 | 8,809 | 9,701 | 0.908 | 0.837–0.980 | Yes |
| Wales | 1,916 | 2,951 | 3,311 | 0.891 | 0.665–1.218 | No |
| West Midlands | 3,574 | 8,774 | 8,982 | 0.977 | 0.809–1.164 | Yes |
| Yorkshire and The Humber | 2,864 | 7,976 | 8,547 | 0.933 | 0.862–1.039 | Yes |
Compared with fitted-flow exposure on exactly the same rows: prediction-rate rank correlation 1.000; the central 90% of fitted-rate ratios spans 1.00–1.00. Coefficients are refitted for each allocation. This range is a sensitivity result, not an exposure confidence interval.
A separate history benchmark adjusts the baseline using earlier observed versus expected collisions, shrunk with a fixed pseudo-count of 10. It uses only earlier eligible outcomes. Missing history receives no adjustment.
| Evaluation | Geographic/road baseline O/E | Frozen-history O/E | History deviance reduction |
|---|---|---|---|
| 2023 · earlier 2022 history | 0.950 | 0.976 | 20.9% |
| 2024 · earlier 2022–23 history | 0.937 | 0.964 | 26.0% |
| 2025 · frozen 2022–23 history | 0.927 | 0.954 | 25.5% |
The additional rolling-history 2025 benchmark also uses 2024 outcomes and is reported separately. Local-history forecasts are distinct from geographic transfer into an unseen region. Future feature indices must show useful incremental value against both comparators.
The counted-major-route sensitivity uses 7,386 accepted, non-overlapping section-years. It matches the collision numerator to the same route and polygon fragments as the annualised counted-AADF mileage. It retains 12,202 zero-collision cells and 26,192 matched motor injury collisions before the modelling eligibility gates.
A separate restricted-scope model is fitted to 22,888 eligible observations. This avoids minor-road allocation, while retaining uncertainty from annualised counts, retrospective geometry and road matching. It does not validate the national all-road denominator.
| Evaluation | Eligible rows | Matched collisions | O/E |
|---|---|---|---|
| 2023 development | 5,098 | 6,032 | 0.966 |
| 2024 diagnostic | 6,430 | 6,863 | 1.033 |
| 2025 exploratory | 6,060 | 6,133 | 0.941 |
The geographic holdouts retain material residual dependence: development mean Moran’s I is 0.344. Authority clustering and block intervals acknowledge part of that dependence; they do not remove cross-boundary structure.
Phase 5 now tests individual features against this fixed comparator. Open the feature evidence register and research shortlist. Promotion still requires better calibration, exposure robustness, uncertainty that refits model selection, and genuinely untouched outcome evidence. All 18 deferred candidates remain deferred.
Reported performance intervals condition on fitted predictions. They do not satisfy the later promotion bootstrap requirement. No causal, residence-risk or all-claims interpretation is supported.
The reproduction kit contains pinned inputs and scripts to refit the models, rerun validation and rebuild numerical outputs. National predictions account for all 142,688 rows. In-sample development predictions and later-year predictions are explicitly labelled.
| Download | Purpose | Size |
|---|---|---|
| National baseline predictions | All 142,688 rows; predictions withheld where ineligible. | 3.91 MB |
| All validation predictions | Every candidate/scenario/fold; observed counts, expected counts and Poisson deviance. | 75.53 MB |
| Observed-to-expected diagnostics | Countries, regions, urban/rural, exposure support and years. | 0.25 MB |
| Model comparison | Temporal, geographic and history benchmark performance. | 0.00 MB |
| Mileage sensitivity | Identical rows; three independently refitted allocation scenarios. | 0.00 MB |
| Poisson and NB2 specifications | Development fit, dispersion, coefficients, cluster covariance intervals and transformations. | 0.04 MB |
| Selected baseline models | Scorable parameters and design for each mileage scenario. | 0.03 MB |
| Baseline coefficient comparison | Scenario-specific fitted effects; conditional UTLA-cluster intervals. | 0.01 MB |
| Eligibility register | All national rows retained with quality flags. | 0.62 MB |
| Counted major-route panel | Matched road scope: annual AADF×physical route length and matched injury collisions. | 0.36 MB |
| Counted route diagnostics | Separate refit; no modelled minor allocation. | 0.02 MB |
| Conditional block intervals | 1,000 UTLA block resamples of fixed predictions; excludes refitting/selection/exposure uncertainty. | 0.03 MB |
| Residual spatial dependence | Moran diagnostics by year and across development. | 0.00 MB |
| Reproduction kit | Pinned source inputs and code to refit all models and rebuild checksummed numerical outputs. | 65.51 MB |
Model specificationMethods and reproductionPinned input filesOutput checksumsVerification record
Browser interaction QA was unavailable in this build environment. Numerical reproduction, rendered-page checks, production build and download verification are recorded separately.