THE ACTUARY LENS · SOCIAL-FACTOR RESEARCH · 10 OCTOBER 2026
What does social context add to road-risk prediction?
Social context carries a temporal signal. Geographic transfer remains unresolved. Recorded burglary adds no consistent predictive value in the matched force study.
Adaptive merging found no stable extra predictive value
Training-only geographic validation selected complete pooling in all three annual tests and all nine regional tests. Income deprivation, employment deprivation and education therefore add zero social coefficients to the selected point models. This is a finding about this modelling procedure, not proof that underlying risks are equal.
Income deprivation: 1–10; Employment deprivation: 1–10; No qualifications: 1–10.
Selected-model temporal gain beyond baseline: 0.0%. Fixed quintiles and continuous-score alternatives are also reported. 199 of 199 full nested regional resamples completed; the predictive acceptance gates are not all met.
Deciles remain useful for illustration; their predictive use is not validated.
Combined deciles improve the pooled temporal score, but do not meet every frozen validation criterion and do not establish non-inferiority to smooth continuous scores. Keep continuous-score research comparisons and use deciles to explain held-out predictions. No rating factor is deployed.
Combined deciles give +7.7% temporal deviance gain, with a refitted 95% interval of -1.8% to +31.1%. Their unseen-region gain is -39.1%. The frozen persistence, geographic-transfer and training-support requirements are not all met.
LSOA profiles remain descriptive: no credible matched all-road mileage denominator is available. This follow-up has its own version and finding register; the original findings below remain frozen.
The signal is useful for research, but not yet validated for pricing
Police-reported personal-injury collisions at road location per 100,000 matched motor vehicle miles. The models estimate collision counts using matched mileage as the exposure offset. These findings describe area context around roads.
Read the percentages as model performance. Percentage reduction in held-out Poisson deviance versus each stated comparator. Positive means better model fit; negative means worse. It is not a percentage change in collision frequency, claims or premiums.
133 English local authorities
Deprivation and education together
Temporal signal; regional transfer unresolved
Original later-year gain 2023–2025
+11.0%95% interval: +5.99% to +15.02%
Original unseen-region gain 2025
-4.35%95% interval: -15.58% to +10.59%
Added: Income and employment deprivation plus the 2011 no-qualifications share. Compared with: Road-context baseline and collision history.
Fixed-parameter diagnostic and calibration
The later diagnostic retains the 2023 model parameters and gives +12.2% combined temporal gain. It uses already-inspected outcomes and does not confirm the original result independently.
Original aggregate observed / expected collisions: 1.030 for later years; 1.141 for unseen regions. A value of 1 indicates aggregate calibration.
Original decision: Not consistently supported. The diagnostic does not change this status.
Later-year improvement; little original transfer gain
Original later-year gain 2023–2025
+6.8%95% interval: +3.00% to +10.03%
Original unseen-region gain 2025
-0.01%95% interval: -7.25% to +8.99%
Added: Income and employment deprivation. Compared with: Road-context baseline and collision history.
Fixed-parameter diagnostic and calibration
The later diagnostic retains the 2023 model parameters and gives +7.7% combined temporal gain. It uses already-inspected outcomes and does not confirm the original result independently.
Original aggregate observed / expected collisions: 1.034 for later years; 1.146 for unseen regions. A value of 1 indicates aggregate calibration.
Original decision: Not consistently supported. The diagnostic does not change this status.
Added: 2011 share of residents aged 16+ with no qualifications. Compared with: Road-context baseline, deprivation and collision history.
Fixed-parameter diagnostic and calibration
The later diagnostic retains the 2023 model parameters and gives +4.9% combined temporal gain. It uses already-inspected outcomes and does not confirm the original result independently.
Original aggregate observed / expected collisions: 1.030 for later years; 1.141 for unseen regions. A value of 1 indicates aggregate calibration.
Original decision: Not consistently supported. The diagnostic does not change this status.
Added: Two-year-lagged burglary intensity relative to fixed 2011 population. Compared with: Force-level road-context baseline, deprivation and collision history.
Fixed-parameter diagnostic and calibration
The later diagnostic retains the 2023 model parameters and gives -4.1% combined temporal gain. It uses already-inspected outcomes and does not confirm the original result independently.
Original aggregate observed / expected collisions: 0.974 for later years; 0.683 for unseen regions. A value of 1 indicates aggregate calibration.
Original decision: Not consistently supported. The diagnostic does not change this status.
Four decision-relevant comparisons beyond collision history are summarised. They are not the full analysis: all 11 primary comparisons and the 29-force recording sensitivity remain available in v2. Authority and force studies have different rows and are not a league table. 95% intervals use nine region blocks and fixed predictions. They omit refitting and input measurement uncertainty, may be optimistic under regional dependence, and are unadjusted for multiple exploratory comparisons.
WHY THE DISTINCTION MATTERS
Performance over time and performance in new places differ
+12.2%
Combined deprivation and education gain with 2023 parameters held fixed. Later years still use updated lagged inputs and realised mileage.
24 of 33
London authorities outside at least one training predictor range in the original combined-model geographic test. This covers 63.9% of London’s matched mileage.
Alternative assumptions change the geographic result. They were examined after seeing the original outcomes; a better diagnostic score does not replace the original failed transfer test. London remains included. Being outside a training range is a warning about extrapolation, not proof of why a model fails.
How much does geographic performance change?
Combined deprivation and education versus collision history · same 2025 regional assessment rows
Specification
Unseen-region deviance gain
Evidence status
Original v2
-4.35%
Original exploratory assessment
Poisson count family
+6.51%
Post-result diagnostic
Fixed 2015 deprivation
+3.90%
Post-result diagnostic
Without regional intercepts
+7.11%
Post-result diagnostic
The fixed-parameter test is temporal only. These separate sensitivities are not independent replications.
Retain the combined block as a research candidate. Resolve road-context confounding and test a frozen specification on fresh outcomes before operational use.
Deprivation
Test the deprivation block alongside road mix and history. The combined evidence does not identify separate causal effects of income and employment.
Education
Keep qualifications as an area-context descriptor for investigation. It is neither a current annual education measure nor a driver-level effect.
Recorded burglary
Do not promote burglary as a stable predictor from this study. This finding cannot settle the predictive value of all crime or smaller local crime measures.
Publish the research interpretation and negative findings with exact comparison links. No risk score, postcode ranking, individual-driver relativity or automatic pricing factor is created. Area education or deprivation cannot be attributed to individual drivers. The study does not establish causation or a premium adjustment.
What would change the decision?
Comparable road context. Add historically matched joint road-class and urban/rural mileage controls, or design a study on a network where matched exposure is observable.
Fresh validation. Freeze the next specification and acceptance criteria before inspecting new outcomes or independent data; assess temporal performance, geographic transfer and calibration together.
Insurer applicability. Test relevant features against geocoded insurer mileage/exposure and developed claims, with portfolio calibration and review of intended use.
TRACE THE INTERPRETATION
The original results remain the evidence
No comparison met every original exploratory criterion. The full v1 hypotheses remain untested under their intended road-context specification. This publication records an interpretation of v2 and v3; it does not record human actuarial approval or promote any hypothesis.
Original models were refitted before each annual assessment. The later diagnostic holds 2023 parameters fixed while updating lagged inputs and using realised mileage. Its design followed inspection of the original results.
Methods and unresolved data gaps
The reduced baseline omits historical joint road-class and urban/rural controls; they could explain some social signal.
2011 Census context is held fixed. Historical input series were retrieved in 2026 and exact forecast-origin vintages have not been reconstructed.
Area residents are not necessarily the drivers on those roads. The study excludes damage-only events and does not measure insured claims.
The fixed complete-history cohort is a restricted sample and uses knowledge of later data availability. It does not establish a whole-England result.
All 2023–2025 outcomes have been inspected. Reusing them cannot provide fresh confirmation.
Baseline controls include fixed 2011 population density and age structure, lagged vehicle mix and traffic intensity, region and trend. Deprivation uses income and employment domains. Crime means recorded burglary in fully matched force areas, with a separately reported recording-break sensitivity.
Publication register and source fingerprints
Publication social-findings-publication-2026-10-10-v1 Base data release lens-2026-10-09-cbec43fe76c4
At this frozen publication: Phase 7 was complete. Phase 8 remains pending. Recurring source monitoring and model refresh are not enabled by this publication. Current operations are shown on the separate maintenance page.