Compare deprivation and education deciles with continuous scores, using the same injury collisions, motor mileage and assessment rows.
CURRENT CONCLUSION
The selected merged bands return to the baseline: no incremental gain was retained.
Evidence strength
Exploratory adjacent-band analysis. Candidate and replacement gates fail; the selected fusion pools each social feature completely.
Pricing implication
The tested merging procedure does not support adding these social bands to this baseline. It does not establish that every social representation is uninformative.
Evidence needed next
A new hypothesis needs a prospectively frozen specification and uninspected assessment evidence, with independent exposure checks.
What this study measures
133 exact-code English authorities, 2018–2025. Authority-year injury collisions and realised all-motor miles; temporal assessments 2023–2025 and leave-region-out checks. Outcomes were previously inspected.
Read the bands correctly. Authority social bands use their frozen source vintages, including Census 2011 qualifications. Read the feature-specific boundaries and direction; do not transfer D1–D10 labels to the separate 2021 national LSOA study.
Follow the reasoning · why each test comes next
1 / Question
Can simpler bands retain useful information?
Adjacent groups may be combined only if the predictive evidence supports the simpler representation. A lack of significance is not proof of equivalence.
Compare representations on matched authority-year observations and mileage. The national joint GLM uses a different population and is a separate study.
Cutpoints, tuning and fusion are relearned within the training procedure. Compare the selected grouping against continuous, spline, ten-band and fixed-pair alternatives.
Compare temporal and geographic assessments and bootstrap refits. The selected solution returns to the baseline; earlier outcome inspection limits confirmation.
Adaptive adjacent grouping · research version v3 · no incremental gain
Validation selected one band for each social feature.
In each annual assessment, training-only geographic validation selected complete pooling of income, employment and education deciles. The resulting model has the same predictions as the baseline. This tests predictive usefulness; it does not establish equal underlying risk across deciles.
2025 social coefficients0 of 27 retainedOriginal deciles preserved in the source dataTemporal gain beyond baseline0.00%Refit 95% range: 0.00% to 0.00%Geographic gain beyond baseline0.00%All nine regions assessed separately
No predictive factor promoted. Outcomes from 2023–2025 had already been inspected. This exploratory experiment does not justify deploying a deprivation or education rating factor. Groupings apply to the authority study; no LSOA frequency model is available.
133 ENGLISH AUTHORITIES · MODEL TRAINED THROUGH 2024
The learned grouping
Income deprivation
1–10
1 band · decile 1 is highest deprivation / no qualifications
Employment deprivation
1–10
1 band · decile 1 is highest deprivation / no qualifications
No qualifications
1–10
1 band · decile 1 is highest deprivation / no qualifications
Original source deciles remain 1–10. Income and employment use their historical deprivation editions; education uses 2011 Census no-qualifications share. These are custom authority quantiles, separate from official LSOA deciles.
Solid teal: learned bands. Dashed brown: ten separate decile effects. Vertical lines: exploratory 95% refit ranges. These are conditional model coefficients relative to decile 5, holding other predictors fixed. They are not observed rates or causal effects. Both point models use training data through 2024.Inspect effects, intervals and boundary stability
2025 assessment model · each feature remains separate
Original decile
Learned band
Adaptive relative rate
Refit 95% range
Ten-decile relative rate
Next boundary retained
1
1–10
1.000
1.000–1.000
1.149
1.0%
2
1–10
1.000
1.000–1.000
1.109
0.5%
3
1–10
1.000
1.000–1.000
1.025
0.5%
4
1–10
1.000
1.000–1.000
0.977
0.0%
5
1–10
1.000
1.000–1.000
1.000
0.5%
6
1–10
1.000
1.000–1.000
1.092
0.5%
7
1–10
1.000
1.000–1.000
1.137
0.0%
8
1–10
1.000
1.000–1.000
1.049
1.0%
9
1–10
1.000
1.000–1.000
1.291
0.0%
10
1–10
1.000
1.000–1.000
1.331
—
Boundary retention repeats cutpoints, grouping selection and fitting. The training population and score cutpoints therefore vary across refits. It is a stability measure, not a significance test. 199 of 199 paired regional resamples completed.
Training support and all selected groupings
2025 model · merged training bands need ten original authorities and 100 collisions
Feature
Vintage
Deciles
Authorities
Training collisions
Billion miles
Support
Income deprivation
2015
1–10
133
195,416
479.50
Pass
Income deprivation
2019
1–10
133
258,041
705.91
Pass
Employment deprivation
2015
1–10
133
195,416
479.50
Pass
Employment deprivation
2019
1–10
133
258,041
705.91
Pass
No qualifications
2011
1–10
133
453,457
1,185.42
Pass
No single full-data grouping is reused in assessment folds
Assessment fold
Income bands
Employment bands
Education bands
Social coefficients
temporal-2023
1–10
1–10
1–10
0
temporal-2024
1–10
1–10
1–10
0
temporal-2025
1–10
1–10
1–10
0
spatial-East Midlands
1–10
1–10
1–10
0
spatial-East of England
1–10
1–10
1–10
0
spatial-London
1–10
1–10
1–10
0
spatial-North East
1–10
1–10
1–10
0
spatial-North West
1–10
1–10
1–10
0
spatial-South East
1–10
1–10
1–10
0
spatial-South West
1–10
1–10
1–10
0
spatial-West Midlands
1–10
1–10
1–10
0
spatial-Yorkshire and The Humber
1–10
1–10
1–10
0
Does merging improve prediction?
Six models use the same 399 authority-year assessments in 2023–2025. Lower deviance is better. A positive gain means the model improves on the baseline. All fits use a mileage offset; rates and absolute error are per 100,000 all-motor vehicle-miles.
Pooled future-year assessments · models trained separately before each year
Model
Deviance
Gain vs baseline
Observed / expected
Predicted rate
Absolute rate error
Baseline
8,810.4
0.00%
1.037
0.03387
0.00497
Continuous · linear
7,844.3
+10.97%
1.030
0.03410
0.00468
Continuous · smooth
7,970.8
+9.53%
1.028
0.03416
0.00480
Ten deciles
8,129.1
+7.73%
1.027
0.03420
0.00487
Fixed quintiles
8,142.2
+7.58%
1.026
0.03421
0.00486
Adaptive adjacent bands
8,810.4
0.00%
1.037
0.03387
0.00497
The 0–0% percentile range against baseline reflects a concentration of bootstrap selections at exactly the same baseline model. It is not certainty that social effects are zero, and these post-selection ranges are exploratory.
The fixed-quintile comparison merges 1–2, 3–4, 5–6, 7–8 and 9–10. Adaptive merging is selected for geographic transfer inside training data; it may sacrifice same-authority future-year performance. Every comparator is reported, with no switch to whichever assessment result looks best.
Adaptive bands versus each comparator · positive gain favours adaptive bands
Comparator
Temporal gain
Refit 95% range
Geographic gain
Replacement supported?
Baseline
0.00%
0.00% to 0.00%
0.00%
No
Continuous · linear
-12.32%
-26.28% to +2.77%
+4.17%
No
Continuous · smooth
-10.53%
-39.13% to +13.53%
+11.12%
No
Ten deciles
-8.38%
-48.93% to +3.51%
+28.08%
No
Fixed quintiles
-8.21%
-20.52% to +6.24%
+7.24%
No
Every predictive acceptance gate
Adaptive bands versus baseline · all gates must pass
Requirement
Result
At least 1% pooled gain
Fail
Complete paired interval above zero
Fail
At least 1% gain in every year
Fail
Pooled O/E between 0.90 and 1.10
Pass
Every annual O/E between 0.90 and 1.10
Pass
Positive pooled geographic gain
Fail
Positive gain in at least five regions
Fail
Absolute rate error no worse
Pass
Every selected training band supported
Pass
Converged, full-rank selected point models
Pass
Replacement additionally requires the upper 95% excess-deviance bound below 0.5%, with no deterioration in calibration or absolute error. A merged pair is not proof of equivalence.
Persistence and unseen regions
Annual gain vs baseline · each model uses the same assessment rows
Year
Continuous · linear
Continuous · smooth
Ten deciles
Fixed quintiles
Adaptive adjacent bands
2023
+5.70%
-7.95%
+0.33%
+1.74%
0.00%
2024
+11.57%
+11.38%
+10.02%
+10.88%
0.00%
2025
+14.17%
+20.29%
+10.95%
+8.80%
0.00%
2025 geographic assessment · each target region excluded from all training years
Held-out region
Areas
Adaptive gain vs B
Ten-decile gain vs B
Quintile gain vs B
Adaptive O/E
East Midlands
8
0.00%
-39.80%
-8.35%
1.199
East of England
11
0.00%
+12.18%
+14.00%
1.314
London
33
0.00%
-83.07%
-20.15%
1.244
North East
10
0.00%
-87.27%
-75.23%
0.832
North West
22
0.00%
-63.65%
-34.26%
0.977
South East
18
0.00%
+3.45%
+7.87%
1.174
South West
7
0.00%
+8.12%
+15.11%
1.181
West Midlands
12
0.00%
+11.44%
+22.81%
1.102
Yorkshire and The Humber
12
0.00%
-46.15%
-3.23%
1.052
Geographic baseline O/E is 1.152; adaptive O/E is 1.152. Pooling does not resolve omitted transport context or all calibration problems.
How often did bootstrap refits retain more bands?
2025 assessment model · whole-region resampling with complete nested refitting
Number of bands
Income deprivation
Employment deprivation
No qualifications
1
196 / 199
197 / 199
196 / 199
2
0 / 199
0 / 199
0 / 199
3
2 / 199
2 / 199
0 / 199
4
0 / 199
0 / 199
0 / 199
5
1 / 199
0 / 199
1 / 199
6
0 / 199
0 / 199
2 / 199
7
0 / 199
0 / 199
0 / 199
8
0 / 199
0 / 199
0 / 199
9
0 / 199
0 / 199
0 / 199
10
0 / 199
0 / 199
0 / 199
There are only nine independent regional blocks. These 199 exploratory resamples provide limited tail precision. All failures are retained in the downloadable record. 0 selected bootstrap designs and 108 comparator designs were rank deficient; they remain in the disclosed analysis using the fixed numerical ridge.
How adjacent bands were selected
Start with ten ordered deciles for each feature and estimate adjusted effects in the negative-binomial count model.
Penalise differences between neighbouring coefficients. Increasing the penalty produces wider bands, without imposing a monotonic risk trend.
Within each outer training set, hold out three groups of regions in turn. Rebuild cutpoints and every transform using only the remaining regions, then score the held-out regions in the latest training year.
Select the strongest penalty within 0.5% of the best eligible inner deviance. This tolerance favours simplicity and is not a statistical equivalence test. All selected bands must meet support requirements.
Assess on the untouched outer fold, then repeat the full process in 199 regional bootstrap samples.
Inner tuning measures geographic transfer. The outer annual tests measure future-year prediction. Seven penalty ratios from zero to the fully merged endpoint were frozen before fitting; this experiment does not rule out every possible grouping or modelling approach.
Why did validation select complete pooling?
2025 model: inner geographic validation on training data only
Penalty ratio
Social coefficients
Inner deviance
Change vs best eligible
Selected?
1
0
3,975.2
0.00%
Yes
0.3
9
4,452.3
+12.00%
No
0.1
21
4,805.6
+20.89%
No
0.03
25
5,237.0
+31.74%
No
0.01
26
5,371.0
+35.11%
No
0.003
27
5,418.7
+36.31%
No
0
27
5,439.5
+36.84%
No
Lower inner deviance is better. Ratio 1 is complete pooling; ratio 0 is ten separate deciles. The intermediate choices were available to the selection procedure. These scores are from training-only region splits, separate from the outer assessment results above.
A recorded numerical correction makes the fully merged endpoint use the exact baseline fit. This removes tiny optimizer-roundoff boundaries. The first attempt, correction and unchanged selection rules are available below.
LSOA scope: Authority coefficients and groupings do not transfer to neighbourhoods. Credible matched all-road LSOA mileage is still required. Inspect LSOA profiles and exposure gaps.
Study limitations and future confirmation
Authority groupings cannot be applied to LSOAs without separate validation. LSOA matched all-road mileage remains unavailable.
Correlated area characteristics, omitted historical joint road-class/rurality exposure and retrospective cohort selection remain limitations.
No significance-based adjacent testing. A merged pair is not proof of equal underlying risk.
No fresh confirmation: retain existing prospective 2027–2029 reserve and access audit.
Bootstrap coefficient intervals are exploratory refit summaries, not post-selection confidence guarantees.
Inspect and reproduce
Computed 2026-10-10. The same finding register supplies this page and the Actuary Lens summary. Earlier v1 and v2 studies remain available.
Read the Actuary Lens summary. Automated browser and mobile visual review were unavailable; calculation, rendering, query validation and publication checks are recorded.