{
  "id": "social-factors-exploratory",
  "version": "v2",
  "amendedOn": "2026-10-10",
  "status": "Exploratory amendment; frozen before model fitting",
  "parent": "/research/social-factors/v1/protocol.json",
  "rationale": "The full v1 baseline lacks historical joint road strata and several as-of inputs. A reduced, explicitly retrospective model can test whether social factors add predictive information within a restricted, exactly matched sample. V1 and its H1\u2013H6 statuses remain unchanged. This amendment does not redefine a v1 pass.",
  "priorInspection": "Historical outcomes have already been displayed on the website and inspected for counts, eligibility and coverage. No social-factor model or new feature/outcome association was fitted before freezing this amendment and its design. This is website registration, not independent preregistration.",
  "outcome": "Police-reported personal-injury collision counts at road location; log(matched motor miles / 100000) offset. Predicted intensity is collisions per 100000 motor vehicle miles. Not all accidents, claims, driver fault or resident risk.",
  "cohort": "Fixed authorities with all ten 2016\u20132025 count/mileage rows eligible, both income/employment domain scores, and exact-code 2011 Census education/age/density matches. No missing outcomes, exposures or crime are imputed. Changed unmatched codes are excluded. Exact geographic codes are assumed to represent sufficiently stable whole-area geography; minor same-code boundary adjustments are an unquantified limitation.",
  "modelYears": [
    2018,
    2019,
    2020,
    2021,
    2022,
    2023,
    2024,
    2025
  ],
  "temporalHoldouts": [
    2023,
    2024,
    2025
  ],
  "latestReservedModelHoldout": 2025,
  "training": "Expanding windows strictly before each holdout, excluding 2020\u20132021, with at least three non-pandemic training years. Two-year lagged traffic/history begins in 2018, so the first primary holdout is 2023 (training 2018, 2019, 2022). Pandemic stress fits train on 2018\u20132019 only and cannot satisfy the primary minimum.",
  "spatial": "Leave each English region out entirely from pre-2025 training, then predict that region in 2025. Report every represented region. No random row split. Unknown-region intercept is the training-row-weighted mean fitted regional effect.",
  "family": "Negative-binomial NB2 maximum likelihood for all compared models, with variance mu + alpha*mu^2 and log link. Estimate alpha separately within each training fit; alpha bounds 1e-6 to 10. Fixed numerical ridge 1e-6 on non-intercept coefficients for stability. Linear feature terms, no interaction or outcome-selected spline/regularisation search. Use the same count family and exact rows for nested comparisons.",
  "baseline": [
    "log 2011 population density per square kilometre",
    "2011 population share aged under25 and share65plus",
    "two-year lagged car/taxi share of motor mileage",
    "log two-year lagged motor miles per road kilometre, an area traffic-intensity proxy",
    "linear calendar-year trend and categorical English region",
    "log realised target-year matched motor miles / 100000 offset"
  ],
  "socialFeatures": {
    "deprivation": "Income and employment domain average scores from the frozen panel, using the published 2015/2019 vintage available before the target-year origin. No overall IMD or traffic-derived domains.",
    "education": "2011 Census share of residents16plus with no qualifications. Level4plus replaces it in a separately reported sensitivity; never include both together. Fixed baseline descriptors, not reconstructed annual education.",
    "crime": "Separate whole-police-force extension only. Two-year lagged total recorded burglary across its explicitly coded subcategories, transformed log(1 + count/2011population*1000). This is intensity relative to a fixed2011population, not an annual resident crime rate. Include only forces fully constructed from compatible whole authorities; sum collision counts and matching miles. Reject incomplete crime quarters. Primary fixed-cohort crime study excludes any force with a marked recording-system-break window (Greater Manchester), across all study years. Separately repeat with that force included, retaining the source warnings. Historical numerical revisions remain usable only under the retrospective-vintage assumption. Exclude collision/traffic offences by choosing burglary specifically. Do not generalise a burglary result to all crime.",
    "forceAggregation": "Population and age counts sum. Density uses summed population over summed area. Education shares use summed16plus denominators. Income/employment use2011resident-weighted authority scores as contextual proxies, not official force-level deprivation statistics or reconstructed deprivation numerators. Lag vehicle share and road intensity use summed matching miles/road lengths."
  },
  "historyBenchmark": "Add log credibility-adjusted two-year lag collision rate. Prior rate is the training aggregate collision rate; prior exposure equals median training lagged annual exposure. All prior parameters are re-estimated only inside each training fold. Pandemic lag values remain observed rather than replaced.",
  "comparisons": {
    "authority": [
      "B+deprivation vs B",
      "B+deprivation+education vs B+deprivation",
      "Level4plus education alternative",
      "Repeat deprivation and education with history",
      "Combined deprivation+education against history baseline"
    ],
    "crime": [
      "B+deprivation+burglary vs B+deprivation",
      "Repeat burglary with history",
      "Add education after deprivation+burglary",
      "Repeat education after deprivation+burglary with history",
      "All three social blocks against history baseline"
    ]
  },
  "transformations": "Training-only centring and scaling of continuous terms; store parameters with every model. No holdout recalibration. Realised test mileage is an exposure for conditional-frequency scoring, not an available-at-origin mileage forecast.",
  "timingAssumption": "Two-year lags avoid obvious publication delay, but archived inputs are current retrieved/revised historical series. Exact historical bytes at every January forecast origin are not certified. Results are retrospective out-of-time validation and not a simulated live-vintage backtest. Census2011 tables were available byJanuary2013; these retrieved bytes are not a2013snapshot.",
  "metrics": "Held-out Poisson deviance reduction versus comparator; observed/expected calibration; exposure-weighted absolute rate error. Report temporal annual and combined results, leave-region-out results, and pandemic stress failures. Keep all outcomes including zero counts.",
  "uncertainty": "2000 paired resamples of English-region blocks over fixed out-of-sample predictions, seed20261010. Keep all years of an authority together. Conditional intervals exclude model-refitting and input measurement uncertainty. Nine regions provide limited independent blocks; nearby-region dependence can make intervals optimistic. One-region table rows do not provide independent uncertainty; only their point estimates are interpreted. Intervals are exploratory, unadjusted for multiple comparisons; no confirmatory p-values or FDR pass is claimed.",
  "candidateLabel": "Exploratory candidate only if combined temporal deviance gain >=1%, regional-bootstrap95%lower bound>0, all3annualholdouts gain>=1% and observed/expected0.90\u20131.10, positive leave-region-out gain in at least5regions and overall, and combined temporal observed/expected0.90\u20131.10. Otherwise label Not consistently supported. This exploratory label never promotes v1 hypotheses or claims proof.",
  "publication": "Publish the full exploratory research package and null/negative results separately from approved analytical releases. Do not record a human actuarial review unless actually performed. Wider-site cards link to the same results without causal, individual-driver, insurer-pricing or whole-England claims. The original95%countrywidecoverage gate remains intact. Detailed joint road-mix controls, currentannualpopulation and prospective vintage validation remain outstanding.",
  "freeze": "Before fitting, archive exact row IDs, fold membership, source/protocol/code hashes. Runtime model transforms are estimated on training rows only and archived afterwards. Any outcome-informed change after fitting requires a new version and disclosure; preserve failed specifications.",
  "crimeSensitivity": "Repeat every crime comparison with the full29completeforce sample, including Greater Manchester. This does not replace the primary28force result or create an additional confirmatory test."
}