{
  "id": "social-deciles-2026-10-10-v1",
  "registeredAt": "2026-10-10T16:36:29.784805+00:00",
  "status": "Exploratory after prior outcome inspection; not fresh confirmation",
  "outcome": "Police-reported injury collision events at road location per 100,000 all-motor vehicle miles, including non-motor-only collision events; not claims or resident-driver risk",
  "primaryHypothesis": "DE improves held-out count prediction beyond B; categorical deciles are not assumed linear or monotone",
  "secondaryComparisons": [
    [
      "deprivation",
      "D",
      "B"
    ],
    [
      "education",
      "E",
      "B"
    ],
    [
      "education-added-last",
      "DE",
      "D"
    ],
    [
      "deprivation-added-last",
      "DE",
      "E"
    ],
    [
      "deciles-vs-linear",
      "DE",
      "C-linear"
    ],
    [
      "deciles-vs-smooth",
      "DE",
      "C-smooth"
    ]
  ],
  "geography": "Exact-code 133-authority inherited complete-history cohort; separate England LSOA21 profiles",
  "cohort": "Frozen exploratory-v2 authority rows, 2018\u20132025. Same observations and offsets in all six models. No new outcome-dependent exclusions.",
  "models": {
    "B": "Context + two-year-lagged credibility-adjusted collision history",
    "C-linear": "B + continuous income, employment and 2011 no-qualifications share",
    "C-smooth": "B + natural cubic spline of each of the same three scores, 3 df each",
    "D": "B + categorical income and employment deciles",
    "E": "B + categorical education deciles",
    "DE": "B + all three categorical decile features"
  },
  "context": [
    "logDensity2011",
    "ageUnder25Share2011",
    "age65PlusShare2011",
    "lagCarTaxiShare",
    "lagLogMilesPerRoadKm",
    "year",
    "logHistoryRate",
    "region intercepts"
  ],
  "likelihood": "NB2; log(motorMiles/100000) offset, jointly estimated dispersion",
  "estimation": "Existing frozen v2 NB optimiser, fixed ridge 1e-6 on non-intercept coefficients; no hyperparameter selection. Failure if optimizer unsuccessful, nonfinite, or gradient max >0.05. All failures archived; no substitute fit.",
  "deciles": "1 highest deprivation/no-qualifications; 10 lowest. Income and employment remain separate. Official national LSOA deprivation domain deciles are not authority quantiles. Overall IMD is descriptive only.",
  "trainingBands": "Nine empirical inverse-CDF quantiles of negated scores from unique training area \u00d7 feature-vintage records; separately by edition. Apply unchanged to assessment. Ties stay in lower numbered band; missing remains missing; no reranking after filtering.",
  "displayBands": "Authority display cuts use the fixed eligible cohort separately by feature vintage; no outcome inputs. LSOA education cuts use all available England LSOA21 profiles. Display bands are separate from model-fold bands.",
  "spline": "Four knots at 0, 1/3, 2/3 and 1 quantiles of unique training area \u00d7 vintage values, linear quantile interpolation; restricted cubic basis with linear tails, 3 df excluding intercept. Training-only scaling.",
  "support": "For each feature \u00d7 vintage \u00d7 training fold \u00d7 band, require >=10 distinct authorities and >=100 summed training collisions. Flag every failure, including empty bands; fit without merging and mark unsupported. No fallback quintiles in this version.",
  "referenceDecile": 5,
  "interactions": "None in primary model",
  "assessment": "Expanding annual 2023, 2024, 2025 assessment; train on earlier 2018+ years excluding 2020/2021. Outcome years were inspected in earlier studies; realised test mileage is used.",
  "descriptive": "Each year 2018\u20132025 plus pooled non-pandemic years 2018/2019/2022\u20132025. Sum counts / sum matched miles. Exact conditional Poisson intervals; not clustered or adjusted effects. Annual charts retain pandemic years with labels.",
  "phase4Metrics": "Held-out Poisson deviance (proper common count score, despite NB fitting), observed/expected and mileage-weighted absolute rate error. Point estimates only; refitted block uncertainty and geographic transfer reserved for Phase 5.",
  "futurePhase5Gates": {
    "candidate": "DE vs B gain >=1% pooled, refitted paired spatial-block 95% lower >0; >=1% in each assessment year; O/E 0.90\u20131.10 pooled and each year; positive total leave-region-out gain and >=5/9 positive regions; overall absolute rate error no worse than B; all support gates met",
    "nonInferiority": "DE vs prespecified C-smooth: upper 95% relative excess deviance <0.5%, with no deterioration in overall calibration distance |O/E-1| or absolute rate error",
    "superiority": "DE vs C-smooth gain >=1% and refitted paired-block 95% lower >0; linear comparison secondary",
    "validation": "Nine leave-region-out refits and paired spatial-block refit uncertainty remain Phase 5; no pass/fail promotion at Phase 4. Record access to future outcomes before allocating a fresh holdout."
  },
  "limits": [
    "Historical joint road-class/urban-rural exposure controls unavailable; baseline remains reduced.",
    "Complete-history cohort selected retrospectively; narrower than England. Resident characteristics are assigned to road locations, not drivers.",
    "2011 education is old; 2021 qualifications and 2025 deprivation never backdated into historical models.",
    "Current LSOA profiles mix explicit 2021 education and 2025 deprivation vintages. Nomis TS067 v1 not reconciled to later ONS editions; profiles provisional.",
    "No matched all-road LSOA mileage; LSOA collision-rate modelling remains blocked. No zero mileage, population proxy or invented exposure.",
    "Decile coefficients are conditional associations with correlated area features, not causal or individual insurance effects."
  ]
}