Income, employment and education deciles together reduce held-out deviance by 7.7% beyond context and collision history. Their deviance is 3.6% higher than linear scores and 2.0% higher than smooth scores. These are paired point estimates on the same 399 authority-year assessments.
Education deciles added after deprivation increase deviance by 0.42%. Easier presentation has not established an advantage over continuous features.
Exploratory evidence; no validated pricing finding. The 2023–2025 outcomes were already inspected in earlier studies. Geographic transfer, uncertainty from refitted models and the registered non-inferiority test remain Phase 5. These are area associations, not causal effects or the characteristics of individual drivers.
PHASE 3 · AUTHORITY DESCRIPTIVE PATTERNS
How do observed rates vary across deciles?
Decile 1 means the highest deprivation or no-qualifications share; decile 10 the lowest. These are custom authority bands within the 133-area study cohort. They are separate from official LSOA deprivation deciles and training-fold model bands.
Cohort coverage in 2025: 88.1% of target collisions and 84.3% of known target motor mileage. Coverage describes the cohort, not completeness of police reporting. Missing authorities are not assigned to a decile.
Income deprivation · 2019 edition
Observed collisions per 100,000 motor miles. Whiskers: conditional 95% Poisson intervals. Equal authority weights define bands; mileage weights determine rates.Counts, mileage, coverage and uncertainty for income deprivation
Income deprivation · 2025 · same matched count and exposure rows
Decile
Areas
Area-years
Regions
Collisions
Billion miles
Rate / 100k
95% interval
1
14
14
4
7,274
15.618
0.0466
0.0455–0.0477
2
13
13
5
6,014
11.419
0.0527
0.0513–0.0540
3
14
14
5
6,459
14.774
0.0437
0.0427–0.0448
4
13
13
6
6,280
10.373
0.0605
0.0591–0.0621
5
13
13
7
6,384
15.221
0.0419
0.0409–0.0430
6
13
13
6
9,260
20.252
0.0457
0.0448–0.0467
7
16
16
6
14,251
43.720
0.0326
0.0321–0.0331
8
11
11
4
7,553
27.884
0.0271
0.0265–0.0277
9
14
14
8
11,883
53.070
0.0224
0.0220–0.0228
10
12
12
5
7,789
31.640
0.0246
0.0241–0.0252
Employment deprivation · 2019 edition
Observed collisions per 100,000 motor miles. Whiskers: conditional 95% Poisson intervals. Equal authority weights define bands; mileage weights determine rates.Counts, mileage, coverage and uncertainty for employment deprivation
Employment deprivation · 2025 · same matched count and exposure rows
Decile
Areas
Area-years
Regions
Collisions
Billion miles
Rate / 100k
95% interval
1
14
14
3
4,029
11.403
0.0353
0.0342–0.0364
2
13
13
5
6,335
17.528
0.0361
0.0353–0.0370
3
13
13
7
5,645
15.941
0.0354
0.0345–0.0363
4
14
14
8
8,086
21.862
0.0370
0.0362–0.0378
5
13
13
7
10,762
23.489
0.0458
0.0450–0.0467
6
13
13
5
10,225
20.841
0.0491
0.0481–0.0500
7
15
15
7
13,502
45.359
0.0298
0.0293–0.0303
8
13
13
5
6,691
24.431
0.0274
0.0267–0.0281
9
12
12
5
9,831
30.858
0.0319
0.0312–0.0325
10
13
13
5
8,041
32.256
0.0249
0.0244–0.0255
No qualifications · 2011 Census
Observed collisions per 100,000 motor miles. Whiskers: conditional 95% Poisson intervals. Equal authority weights define bands; mileage weights determine rates.Counts, mileage, coverage and uncertainty for no qualifications
No qualifications · 2025 · same matched count and exposure rows
Decile
Areas
Area-years
Regions
Collisions
Billion miles
Rate / 100k
95% interval
1
14
14
4
5,112
14.855
0.0344
0.0335–0.0354
2
13
13
6
7,035
16.470
0.0427
0.0417–0.0437
3
13
13
6
9,014
30.469
0.0296
0.0290–0.0302
4
14
14
7
7,311
25.632
0.0285
0.0279–0.0292
5
13
13
7
8,903
28.857
0.0309
0.0302–0.0315
6
13
13
6
10,959
42.782
0.0256
0.0251–0.0261
7
14
14
5
8,761
22.649
0.0387
0.0379–0.0395
8
13
13
7
9,590
34.204
0.0280
0.0275–0.0286
9
13
13
2
9,498
20.980
0.0453
0.0444–0.0462
10
13
13
2
6,964
7.070
0.0985
0.0962–0.1008
Rate = 100,000 × total collisions / total matched motor miles. It is never the mean of area rates. The intervals condition on fixed exposure and omit spatial clustering, overdispersion and source measurement error. Counts include all police-reported injury collisions, including events without a motor vehicle; the denominator is all-motor mileage on those roads. Raw patterns do not adjust for road mix or other confounding.
Inspect the annual pattern for any decileShaded years are 2020–2021. Deprivation editions change from 2015 to 2019 for outcome years 2020 onward; membership may change at that boundary. Education is fixed at 2011.
Annual observed rate for the selected feature and decile
Year
Areas
Collisions
Billion miles
Rate / 100k
95% interval
2018
14
10,748
14.336
0.0750
0.0736–0.0764
2019
14
10,520
14.509
0.0725
0.0711–0.0739
2020
14
6,185
12.393
0.0499
0.0487–0.0512
2021
14
7,261
13.528
0.0537
0.0524–0.0549
2022
14
7,722
14.726
0.0524
0.0513–0.0536
2023
14
7,243
15.009
0.0483
0.0472–0.0494
2024
14
7,189
15.376
0.0468
0.0457–0.0478
2025
14
7,274
15.618
0.0466
0.0455–0.0477
PHASE 4 · SAME ROWS, SAME EXPOSURE, SIX SPECIFICATIONS
What predictive value survives assessment?
Negative-binomial count models use an offset of log(motor miles / 100,000). Each assessment year is predicted using earlier non-pandemic outcome years. A common Poisson deviance scores every model; positive gain means lower deviance. Assessment mileage is realised mileage, so this is conditional rate prediction.
Combined 2023–2025 temporal assessments · point estimates
Model
Feature specification
Deviance
Gain vs B
Observed / expected
Absolute rate error
B
Context + collision history
8,810.4
0.00%
1.037
0.00497
C-linear
Continuous scores · linear
7,844.3
+10.97%
1.030
0.00468
C-smooth
Continuous scores · smooth
7,970.8
+9.53%
1.028
0.00480
D
Income + employment deciles
8,094.8
+8.12%
1.025
0.00486
E
Education deciles
8,622.0
+2.14%
1.031
0.00495
DE
All three decile features
8,129.1
+7.73%
1.027
0.00487
Observed / expected = 1 indicates aggregate calibration. Absolute rate error is mileage-weighted, in collisions per 100,000 miles. Better values on one metric do not establish superiority. All 18 fits converged; rank, conditioning, dispersion and coefficients are downloadable.
B + all three categorical decile features. Pooled observed / expected: 1.027.
DE · earlier-year training, later-year assessment
Year
Assessment rows
Deviance
Gain vs B
Observed / expected
Observed rate
Predicted rate
2023
133
2,396.5
+0.33%
1.005
0.03664
0.03648
2024
133
2,714.3
+10.02%
1.016
0.03465
0.03411
2025
133
3,018.4
+10.95%
1.062
0.03408
0.03208
Primary and secondary comparisons, including each block added last
Registered comparisons · pooled temporal point estimates, no hypothesis promoted
Comparison
Model / comparator
Deviance gain
All deciles beyond baseline
DE / B
+7.73%
Deprivation beyond baseline
D / B
+8.12%
Education beyond baseline
E / B
+2.14%
Education added after deprivation
DE / D
-0.42%
Deprivation added after education
DE / E
+5.72%
All deciles versus linear scores
DE / C-linear
-3.63%
All deciles versus smooth scores
DE / C-smooth
-1.99%
Phase 5 decision pending. Combined deciles currently have 1.99% excess deviance relative to the prespecified smooth comparator. The registered non-inferiority margin is 0.5% at the upper 95% bound. The combined-decile gain in 2023 is 0.33%, below the registered 1% annual threshold. A favourable decile claim is not supported by these point estimates; the planned geographic tests and refitted uncertainty still need to run.
PHASE 2 · SEPARATE LSOA21 PROFILE PRODUCT
Explore neighbourhood characteristics
All 33,755 English LSOA21 areas have an exact-code deprivation/education join. Income and employment use official national 2025 domain deciles. Education uses custom national deciles of the 2021 no-qualifications share, based on the archived Nomis TS067 v1 edition. This provisional mixed-vintage snapshot is kept separate from the historical authority models.
National ranks stay fixed when an area is selected. LSOA bands describe resident characteristics; they do not rank collision frequency or individual driving risk.
Check the number of LSOAs in each national band
National profile support · ties can make education bands uneven
New decile study · separate from the original eight-phase social-factor programme
Phase
Delivery
Status
1
Versioned hypotheses, model specifications, cohort, folds, source ledger and decision thresholds
Frozen before these fits
2
Authority and LSOA deciles, source hashes, exact-code joins, vintage controls and support audits
Complete; LSOA frequency exposure unavailable
3
Three descriptive decile charts, matched totals, uncertainty and annual patterns
Complete at authority level
4
Six count-model specifications, 18 fits, coefficients and 399 paired assessment rows
Complete; exploratory point estimates
5
Geographic transfer, refitted uncertainty, persistence and non-inferiority decisions
Pending
6
LSOA frequency modelling on matched all-road mileage
Blocked by exposure data
Banding, leakage prevention and model assumptions
Each model fold estimates cutpoints from unique training area × feature-vintage values, with equal area weights. Repeated outcome years do not overweight the reference distribution. Assessment values use those frozen cutpoints.
Income and employment scores are binned separately within each known edition. Ties stay together; empty or missing bands remain visible. All 150 training band checks passed at least 10 distinct authorities and 100 training collisions.
Decile 5 is the categorical reference. No ordinal slope, forced monotonic trend or 10 × 10 interaction is fitted. Overall IMD is excluded from the model.
Smooth models use natural cubic splines with three degrees of freedom per social feature and training-only knots and scaling. The fixed numerical ridge is 10⁻⁶; there is no holdout-based tuning.
Baseline context includes 2011 density and age structure, year, region, two-year-lagged traffic mix, mileage per road kilometre and credibility-adjusted collision history.
Data limitations and the next evidence gates
The cohort uses complete-history availability known retrospectively. Historic joint road-class/urban-rural controls are missing. Pandemic outcome years are excluded from training, but pandemic lagged inputs remain. Education is from 2011 and resident characteristics may poorly describe people travelling through an area.
Phase 5 requires at least 1% pooled and annual gain over baseline, a positive refitted 95% lower bound, calibration between 0.90 and 1.10, improvement in at least five of nine unseen regions and no worse absolute rate error. Non-inferiority to smooth scores uses a separate 0.5% margin. The full frozen protocol contains the exact rules. New, genuinely uninspected outcomes still require a recorded reserve.
LSOA frequency cannot be recovered by treating population as miles or by redistributing authority mileage without validation. No such assumption has been made. No human actuarial review or pricing-factor approval is recorded.
Inspect and reproduce the study
Protocol frozen 2026-10-10; fits completed 2026-10-10. Source archives and prior source hashes are included in the reproduction bundle. Earlier social-factor findings remain unchanged.