All 40 nationally banded features fitted together, plus ten separate raw-value extensions. Find a feature, compare its predictive evidence and record what to keep, drop or simplify.
Outcome: motor-involved, police-recorded injury collisions at the collision location per 100,000 allocated vehicle miles. This retrospective evidence does not measure policyholder claim frequency or causal effects.
CURRENT CONCLUSION
The joint GLM improves predictive fit; calibration and allocated-mileage dependence remain unresolved.
Evidence strength
Exploratory joint model with 40 banded features. Ten raw-value extensions use separate restricted cohorts; 2024–2025 results are previously inspected diagnostics.
Pricing implication
Use this workspace to propose features for a refit. Keep, drop or simplify choices are research decisions, not approved rating changes.
Evidence needed next
Refit proposed specifications, challenge exposure and calibration, then assess prospectively frozen predictions and insurer-claims performance.
What this study measures
32,789 development LSOAs in England and Wales, 2022–2023. Distinct motor-involved injury collisions per 100,000 allocated vehicle miles. This numerator differs from the all-injury contextual rate; extension cohorts vary.
Read the bands correctly. National bands cover 35,672 LSOAs before model eligibility. Higher band = higher feature value, not necessarily higher risk. No qualifications uses Census 2021: D1 = lowest share; D10 = highest. Reference is normally D5, or Z0 for a structural zero.
Follow the reasoning · why each test comes next
1 / Question
Which features add to joint prediction?
An individual gradient can disappear once related features are fitted together. The question is additional predictive value beyond the named baseline and other terms.
Distinguish motor-involved collisions, allocated miles and eligible LSOAs. A restricted extension cannot be compared as though it used the full joint-model cohort.
Compare deviance, observed/expected and geographic and mileage checks. Then distinguish each feature’s contribution from whole-model performance and coefficient significance.
Expected collisions = miles / 100,000 × exp(core adjustments + feature effects). The model estimates 378 coefficients with a Poisson log link. It includes year, Wales × year, region, urban/rural context and mileage-quality controls alongside the 40 categorical features.
A relativity of 1.20 means 20% higher modelled frequency than the stated reference, holding the other terms fixed. D5 is the usual reference; features with a structural zero use Z0. For correlated household, road and area measures, some such comparisons have little real-world support.
Uncertainty and simplification
95% relativity intervals use 162 geographic/source-area blocks. Pearson dispersion is 1.86; clustered covariance allows for extra-Poisson dependence. Intervals exclude exposure, model-selection and source-measurement uncertainty. Conditional q-values and overlap diagnostics are review aids.
The feature contribution refits the model without that feature on 2022 and tests on 2023. It differs from the earlier single-feature ranking. Adjacent-band intervals help you inspect potential merges; an interval including 1 does not establish equivalence. Every chosen drop or merge requires a fresh joint fit.
Several sources describe later snapshots or climate normals. The 2024–2025 data were already inspected, so they are diagnostic rather than untouched validation. Speed coverage is highly selective and current; terrain gradients remain proxies. This review does not complete the earlier automated F8 selection protocol or establish a commercial index.
Exact development support and measurement record
40Features in the joint GLM10Separate restricted-data extensions32,789Development LSOAs, 2022–20230Automatic drops or merges
This models distinct motor-involved, police-recorded injury collisions at the collision location per 100,000 allocated vehicle miles. It is retrospective STATS19 evidence. It does not measure all crashes, policyholder claim frequency or causal effects.
The reference bands remain frozen across all 35,672 England and Wales LSOAs. Model eligibility reduces the development cohort to 32,789 LSOAs; it does not redefine the deciles. Twelve further candidates have no numeric observations and remain visible as unavailable.
32,789 eligible LSOAs; fixed national bands cover 35,672 LSOAs
Geography and period
England and Wales · collision location · Development 2022–2023; previously inspected diagnostics 2024–2025
Numerator
185,755 distinct motor-involved injury collisions in development
Denominator and unit
540,614,673,589.701 allocated vehicle miles in development · Collisions per 100,000 allocated vehicle miles; model effects are relative to the stated reference
Exposure provenance
Local modelled and allocated mileage
Adjustment
Joint Poisson GLM with a mileage offset and core controls
Coverage
Eligibility and exposure sensitivities shown per feature; restricted extensions use separate cohorts.
Interpretation limit: This motor-involved outcome differs from the STATS19 all-injury contextual rate. Retrospective public-road evidence does not establish policyholder claim frequency or causal effects.
Joint prediction, with calibration still to resolve
Compared with the existing rich Phase 4 baseline refitted on identical observations. Different functional forms make this a non-nested comparison. Positive gain means lower Poisson deviance.
Evaluation
Deviance improvement
Conditional 95% interval
Observed / expected
Fit 2022 → evaluate 2023
16.50%
14.90%–18.15%
0.951
Ten geographic holdouts, pooled
19.42%
14.97%–23.68%
1.085
2024 previously inspected diagnostic
16.11%
14.55%–17.86%
0.943
2025 previously inspected diagnostic
15.84%
14.17%–17.61%
0.933
The temporal model overpredicts total collisions; the pooled geographic model underpredicts them. Geographic results vary, with the North East showing negative deviance improvement. The matched counted-route check improves deviance by only 1.27%, with an interval spanning zero. Allocated-mileage gains therefore need further exposure validation.
Development contains 185,755 collisions over 540.61 billion miles, including 14,978 zero-collision observations. Aggregate observed frequency is 0.03436 per 100,000 miles. Its equality to the fitted development average follows from the intercept fit and is not validation.
Now inspect the individual features. Whole-model improvement and calibration provide the context. For each feature, distinguish its adjusted effect, additional predictive contribution and validation support before proposing a change.
Your feature review
Search the fixed register, inspect one feature and record a decision. All 40 banded features share the joint model; the ten raw-value extensions each use their own available cohort.
Exploratory analysis
No qualifications
Joint national-band model · decisions remain in this browser
Evidence ready for No qualifications. Joint national-band model.
The evidence link is shareable. It keeps the analytical release and selected filters. Browser review decisions and notes stay local; use Download decisions separately to share a proposed specification.
Held-out predictive contribution0.143%Conditional 95% interval: -0.007% to 0.298%Fit 2022 → evaluate 2023 · after the other 39 featuresAssociated model O/E: 0.951 · full 40-feature fit. See Validation for context.Coefficient test · conditional q-value0.0005Within-model coefficient evidence. This is a different test from held-out predictive contribution.Development support32,789 LSOAs65,578 observations · 2022–2023162 uncertainty blocks · adjusted GVIF 1.21
The interval includes zero: incremental predictive improvement is not resolved. A small coefficient q-value does not establish incremental predictive value.
Positive gain means lower 2023 Poisson deviance when the feature is included. The conditional interval holds predictions fixed and excludes exposure and model-selection uncertainty. 0 unsupported rows excluded from both scores.
Joint national-band model
No qualifications
100 × (Highest level of qualification: No qualifications) / (Highest level of qualification: Total: All usual residents aged 16 years and over)
Charts displayed; exact values remain available with each chart.
Selected comparison: Preferred mileage · 2022–2023 pooled
Change comparison 2 applied filtersPreferred mileage · 2022–2023 pooled
Fixed reference: D52022–2023 pooled · Preferred mileage · England and Wales32,789 eligible LSOAs · 185,755 motor-involved injury collisions · 540.61 billion miles
Band direction: D1 has the lowest feature values and D10 the highest; higher bands do not mean higher risk. Z0, where present, denotes zero values. This is the reverse of the London no-qualifications ranking, where D1 is highest.
Raw and adjusted patterns use the same selected band records and mileage basis. Raw frequencies retain each band’s covariate mix; conditional relativities hold the other model terms fixed. The two charts use different units. National band boundaries and reference D5 remain frozen.
Raw frequency
Unadjusted pooled frequency on this selection’s records. Raw-rate intervals are not supplied. Reference D5: 0.02959 per 100,000 miles.
Joint adjusted relativity
1.00 is the stated reference. Intervals are conditional on this model and the allocated mileage.
Exact values · also in Data (10 rows)
Same selected observations in both columns. Raw and fitted band averages can coincide by construction in a categorical fit; that equality is not validation. Conditional relativities are relative to the frozen reference.
Band
Frozen raw range
Reference
Adjusted relativity (95% interval)
LSOAs
Collisions
Million miles
Raw / fitted frequency per 100k
Support
D1
0.845–9.551
D5
0.865 (0.81–0.924)
3,212
20,706
46,459.99
0.04457 / 0.04457
No weak-support flag
D2
9.551–12.081
D5
0.904 (0.857–0.954)
3,294
20,560
75,316.58
0.0273 / 0.0273
No weak-support flag
D3
12.081–14
D5
0.91 (0.866–0.956)
3,251
18,785
74,178.25
0.02532 / 0.02532
No weak-support flag
D4
14.001–15.827
D5
0.966 (0.929–1.004)
3,300
19,841
71,979.5
0.02756 / 0.02756
No weak-support flag
D5
15.827–17.677
D5
1 (1–1)
3,285
18,782
63,473.43
0.02959 / 0.02959
No weak-support flag
D6
17.677–19.645
D5
1.03 (0.99–1.071)
3,284
18,028
54,111.69
0.03332 / 0.03332
No weak-support flag
D7
19.645–21.908
D5
1.039 (0.993–1.088)
3,295
17,361
45,503.81
0.03815 / 0.03815
No weak-support flag
D8
21.909–24.594
D5
1.016 (0.963–1.073)
3,302
17,039
43,468.53
0.0392 / 0.0392
No weak-support flag
D9
24.594–28.353
D5
1.084 (1.01–1.164)
3,304
16,443
33,786.08
0.04867 / 0.04867
No weak-support flag
D10
28.353–51.648
D5
1.131 (1.035–1.236)
3,262
18,210
32,336.81
0.05631 / 0.05631
No weak-support flag
Adjusted effects condition on all 40 banded features. Holding correlated or compositional features fixed may describe combinations rarely observed. Empty and unidentifiable bands remain unavailable.
Evidence details · adjacent bands and overlap
Adjacent-band contrasts below use preferred mileage, pooled 2022–2023, regardless of the displayed sensitivity. An interval containing 1 does not prove equivalence; any merge requires a refit and validation.
Exact values · also in Data (9 rows)
Preferred mileage; pooled 2022–2023 adjacent-band contrasts.
The table compares the full 40-feature model with the rich geographic/road baseline on matching observations. These gains and calibration values belong to the whole model, not the selected feature.
O/E = observed collisions divided by expected collisions. Below 1 means overprediction; above 1 means underprediction. The 2024–2025 observations were previously inspected and are diagnostics, not untouched validation.
Feature-specific geographic contribution and held-out contribution under alternative mileage are unavailable. Use the Effects tab for coefficient sensitivity, and the table for whole-model validation. The counted-route cohort differs from the allocated-mileage cohort.
Conditional intervals exclude exposure uncertainty, model selection and source measurement error. The public-road outcome does not establish claims pricing performance.
Choices are saved in this browser. Selecting a choice does not change the displayed model. Download the decision file to share a proposed refit or use another device.
0 choices recorded · browser-only · no automatic drops, merges or refits
Scientific checks, independent numerical comparison, rendered-content tests and the production build are recorded. Interactive browser QA was unavailable in this build environment.