{
  "version": "ew-glm-review-v1",
  "frozenDate": "2026-10-13",
  "sourceCommit": "0ee1c697a43a4c599ed9adab35752790a6ffa0b9",
  "userInstruction": "Include all features for which a GLM can be fitted; user decides drops and simplifications.",
  "primaryFeatures": [
    "no_qualifications",
    "level4_qualifications",
    "social_rented",
    "private_rented",
    "owner_occupied",
    "overcrowded_bedrooms",
    "unemployment",
    "terraced",
    "flats",
    "no_car",
    "cars_per_household_lower_bound",
    "population_density",
    "age15_24",
    "age65_plus",
    "age0_14",
    "motorway_share",
    "a_road_share",
    "minor_road_share",
    "junction_density",
    "roundabout_density",
    "road_sinuosity",
    "air_frost_normal",
    "stop_density",
    "heavy_rain",
    "multi_car",
    "three_plus_car",
    "car_commute",
    "long_commute",
    "work_from_home",
    "walk_commute",
    "cycle_commute",
    "connectivity_public_transport",
    "connectivity_walking",
    "connectivity_cycling",
    "connectivity_driving",
    "air_no2",
    "air_pm25",
    "bankruptcy_la",
    "insolvency_la",
    "heat_summer_days_gt25"
  ],
  "diagnosticFeatures": [
    "claimant_rate",
    "claimant_change",
    "road_gradient",
    "steep_road_share",
    "vehicle_crime",
    "burglary",
    "criminal_damage_arson",
    "non_vehicle_crime",
    "speed_le20",
    "speed_le30"
  ],
  "primary": "Unpenalised Poisson log-link frequency GLM, all40 frozen nationally banded features simultaneously, offset log(vehicleMiles/100000). All eligible zero counts retained. No screening on prior p-values or gains.",
  "core": [
    "intercept",
    "year factor",
    "Wales x year",
    "sum-zero region",
    "urban/rural",
    "four allocation-quality shares",
    "estimated-major-mileage share"
  ],
  "form": "Primary features categorical, replacing their raw copies in the old baseline. Reference D5 or structural zero where available; nearest observed band otherwise. No automatic merging, dropping or shrinkage.",
  "benchmark": "Frozen Phase4 rich_poisson refit on identical observations. Different functional form, so non-nested held-out comparison; not a likelihood-ratio test.",
  "development": [
    2022,
    2023
  ],
  "validation": "2022 fit ->2023 test;10 leave-one-region-out2022\u201323 folds. Refit representations and core within each training fold. Unsupported held-out categorical levels withheld from scoring for both comparator and candidate, counted explicitly.",
  "laterYears": "2024/2025 previously inspected diagnostics only. Latest trained year effect carried forward. No untouched validation.",
  "relativities": "Pooled2022\u201323 conditional coefficient contrasts and cluster-sandwich95% intervals. Report exact reference, raw rates, summed exposure/collisions, fitted rates, weak support, rank and estimability. No causal interpretation of counterfactual changes in compositional features.",
  "dropOne": "Refit without each of40 features in2022; score2023 on the full-model supported cohort. Gain=1-D(full)/D(without). Conditional on every other feature; may be negative. Also report development deviance change; not out-of-sample evidence.",
  "diagnostics": "Each of10 unbanded raw numeric features added separately to same40-feature model, refitting both on its complete-case cohort. Restricted cubic spline on source values, training quantiles5/27.5/50/72.5/95. Four crime rates and mean terrain gradient log1p; other diagnostics identity. Also show linear per-unit sensitivity. Raw grid values are not national deciles. Speed and terrain are unvalidated context proxies; current speed is not historic speed.",
  "diagnosticFailure": "If sparse subset prevents a finite/identifiable joint fit, retain explicit failure; do not invent neutral effects or silently replace with reduced controls.",
  "inference": "162 permanent source-area/authority unions inherited from Phase5 where available. Finite-sample clustered sandwich. No full-model omnibus Wald because coefficient count may exceed cluster covariance rank. Per-feature conditional Wald p and BH q over40 primary or10 exploratory diagnostics separately, unestimable p=1. These are review aids, never drop rules.",
  "adjacent": "Pairwise adjacent-band contrasts use full covariance; intervals including1 do not establish equivalence or approve a merge. Structural zero remains labelled.",
  "overlap": "Unweighted unique-development-LSOA Spearman raw-value correlations and adjusted GVIF^(1/(2df)) where model full-rank. These are overlap diagnostics, not drop criteria.",
  "uncertainty": "Paired fixed-prediction block intervals,1000 draws seed20261012; no refit/selection/exposure uncertainty. Exact null-space contrast checks; zero-event levels and unidentifiable contrasts withheld.",
  "sensitivities": "Refit primary model under benchmark and road-length offsets, same cohort; matched-counted-route outcome/physical-mileage fit with allocation quality set zero. These are sensitivity models, not new selection.",
  "decisionWorkflow": "All feature decisions initially Undecided. User can mark Keep, Drop or Simplify and describe grouping. Browser-local/exported proposals; changes require explicit refit, updating all other coefficients and validation. No multiplying away dropped factors or averaging relativities.",
  "scope": "User-led inclusive GLM review, not completion of the earlier automated F8 penalised/nested-selection protocol. No commercial index/insurer-claims validation. National values and bands and previous results immutable."
}