SOCIAL-FACTOR DECILE STUDY — DRAFT FOR THE NEXT ANALYTICAL VERSION
Recorded: 10 October 2026. Requested by the site owner during Phase 8.
Status: design candidate; no decile features have been fitted or validated.
This draft does not amend the frozen v1/v2/v3 studies or their conclusions.

QUESTION
Do deprivation and educational-attainment bands provide stable, interpretable
prediction of police-reported injury collisions per 100,000 motor vehicle miles,
and how much predictive information is lost relative to continuous features?

FEATURE DEFINITIONS
1. Keep separate income-deprivation and employment-deprivation domain deciles.
   Use the official English LSOA deciles from an edition available at the
   forecast origin: 1 is the most deprived tenth; 10 is the least deprived.
   Keep source scores, ranks, edition and geographic codes alongside each band.
   Overall IMD may be shown as descriptive context, but is not the primary
   deprivation predictor alongside education/crime: it already includes those
   domains. Do not invent a deprivation index by subtracting domains from IMD.
2. Define a separate educational-attainment band using the Census proportion
   of usual residents aged 16+ with no qualifications. Band 1 has the highest
   no-qualifications share; band 10 the lowest. It is a custom band, not an
   official IMD education-domain decile. Level 4+ is a separate sensitivity.
3. For the descriptive map, use a declared, complete England LSOA reference
   population and a source vintage available at the illustrated date. Bands
   group areas, not equal numbers of residents or equal traffic exposure.
   For learned model cutpoints, use unique training LSOAs only, freeze the nine
   thresholds, and apply them unchanged to validation areas. Keep ties together,
   disclose uneven/empty groups, and retain missing as missing. Never use
   collision outcomes to choose thresholds or relabel the bands as risk.

GEOGRAPHY AND EXPOSURE GATE
The current fitted cohort is 133 authorities, not LSOAs. LSOA scores alone do
not provide LSOA collision frequency. A neighbourhood study requires matching
road-location collision counts and all-road motor vehicle miles in the same
LSOA boundaries and years, including zero-collision areas with measured miles.
Do not allocate authority mileage to LSOAs by resident population or road
length alone. If segment traffic estimates are used, document their coverage,
uncertainty and reconciliation to authority totals before fitting.
Crosswalks must preserve numerator/denominator counts; never average decile
codes or treat a changed boundary as an exact match.
An earlier authority-level comparison can bin the existing authority features,
but must be labelled authority deciles, not LSOA deciles or neighbourhood risk.

PRESPECIFIED COMPARISON TO COMPLETE BEFORE FITTING
Use identical eligible rows, controls, count family and mileage offset for
baseline, continuous social features, and separate categorical decile models.
Treat deciles as categories initially: a 1–10 linear term assumes equal effects
between adjacent ranks. Do not assume a monotonic effect or add a 10-by-10
interaction after inspecting the results. Archive all cutpoints and contrasts.
Freeze the exact model design, any shrinkage, minimum support rules and cohort
before estimation. Benchmark against both road-context and lagged-rate models.
Assess temporal and geographic transfer, deviance, observed/expected ratios,
exposure-weighted errors and area-cluster uncertainty. Report decile models
against the continuous comparator as well as the baseline, including failures.
Reanalysis of 2023–2025 remains exploratory: these outcomes are already seen.
A new confirmatory claim needs fresh assessment data reserved before inspection.

DATA LAB AND WEBSITE
Data Lab: add score/decile controls, source-vintage and boundary labels, a
descriptive LSOA band map, and a decile chart of observed and held-out predicted
collisions per 100,000 miles. Compute each plotted observed rate as 100000 times
summed collisions divided by summed matching miles, never the mean area rate.
Show uncertainty, area counts, collision counts, miles, exclusions and support.
Adjusted model contrasts must be distinguished from raw observed differences.
Downloads must include source hashes, band definitions, thresholds, counts,
denominators, fold assignments and the new protocol/results version.
Wider website: link a plain-language explanation to the same finding register.
Promote a predictive claim only after its predefined validation gates pass;
otherwise show exploratory or inconclusive results. Area-level relationships
do not identify individual driver risk, fault, causation or insurance prices.

PRIMARY DEFINITIONS
https://www.gov.uk/government/statistics/english-indices-of-deprivation-2025/english-indices-of-deprivation-2025-statistical-release
https://www.gov.uk/government/statistics/english-indices-of-deprivation-2025/english-indices-of-deprivation-2025-frequently-asked-questions
https://www.ons.gov.uk/datasets/TS067/editions/2021/versions/3
These definitions do not authorize backdating 2025 deprivation or Census 2021
releases into earlier forecast origins. Exact source vintages still need audit.