{
  "id": "20mph-phase4-v4",
  "specifiedAt": "2026-11-04",
  "status": "Retrospective extension; choices recorded before phase4 subgroup results were calculated. Published phase3 and external results were already known.",
  "question": "How do observed casualty patterns and exploratory controlled associations differ across broad groups?",
  "causalGate": "Not met in phase3; subgroup models cannot repair identification.",
  "data": "Pinned DfT 2015\u20132025 casualty records joined one-to-many to the phase3 collision records; all-road location-based Wales plus nine English regions. No provisional2026 splice.",
  "primaryPre": "2015\u20132019 and January2022\u2013August2023:80 months",
  "post": "October2023\u2013December2025:27 months; September2023 excluded",
  "outcome": "All-severity casualty counts. No exposure rate or adjusted serious-injury claim.",
  "groups": {
    "road_user": [
      "Pedestrian",
      "Cyclist",
      "Motorcyclist",
      "Car occupant including taxi",
      "Other known road user",
      "Unknown"
    ],
    "age": [
      "0\u201315",
      "16\u201324",
      "25\u201364",
      "65+",
      "Unknown"
    ],
    "setting": [
      "Urban",
      "Rural",
      "Unknown"
    ],
    "road_class": [
      "Motorway / A road",
      "B / C / unclassified",
      "Unknown"
    ],
    "road_type": [
      "Single carriageway",
      "Dual carriageway",
      "Other known layout",
      "Unknown"
    ],
    "income": [
      "Below10% income deprived",
      "10\u2013below20%",
      "20% or more",
      "Unmatched / excluded"
    ],
    "speed": [
      "Recorded20/30mph",
      "Recorded40mph or above",
      "Other / unknown"
    ]
  },
  "groupingNotes": "Mutually exclusive within each family, overlapping across families. Use official current codebook; unexpected non-missing codes fail closed. Income uses fixed pre-policy2019 comparable England/Wales income deprivation percentages at collision LSOA; no cross-country overall-index ranking and no assumption about the casualty's household income. Income geography uses an independent ONS2011-to2021 exact-fit crosswalk: retain unchanged or uniquely split2011parents for all years; exclude merged/complex parents and corresponding children throughout. This avoids creating post-only missingness from the2024 STATS19 geography change. Display unmatched/excluded share in each period.",
  "model": "Reuse phase3 collision-trained regional weights unchanged for every subgroup. Normalize each series to its own2015 mean. Log Wales-to-comparison index on month indicators, calendar trend and post indicator. Paired pooled-England sensitivity, same specification. Calendar-lag Newey-West12 covariance, finite-sample n/(n-k).",
  "supportGate": {
    "minPreCount": 100,
    "minPostCount": 50,
    "minWelshCountEveryIncludedMonth": 5,
    "maxConditionalMdeReductionPct": 30,
    "mde": "100*(1-exp(-(1.96+0.841621)*SE_post)); require <=30% under both weighted and pooled-England comparisons; conditional nominal80% power at5% threshold, not causal design power.",
    "rule": "All checks required. Also positive2015 means and positive comparison every included month. Unknown/unmatched groups descriptive only. Suppress model estimates for failed groups; retain counts and reasons. No narrow intersections or street rankings. Broad pooling chosen before results; no post-hoc pooling to obtain significance."
  },
  "multiplicity": "Bonferroni95% simultaneous intervals across all support-eligible non-unknown subgroup main estimates. Show paired pooled-England sensitivity with same critical value. No claim of different effects from nonoverlap or separate significance tests. Family heterogeneity Wald tests use calendar-HAC cross-outcome covariance, with Holm correction across eligible families; exploratory only.",
  "diagnostics": "Publish subgroup pretrend estimate,95% interval, zero/small-month counts, conditional MDE, comparator sensitivity. Unknowns reconcile to total. Sparse counts shown without an unstable fitted coefficient.",
  "diversion": "All-network outcomes plus complementary recorded-speed categories are a broad screen, not adjacent-road displacement identification. Adjacent-road causal test requires dated treated geometry, buffer definition, compatible traffic exposure and zero-event roads. If absent, record unidentifiable rather than infer diversion from category counts.",
  "reversal": "Reproduce published Wrexham speed observations and disclose actual observation windows, missing effective dates and lack of matched casualty follow-up. No reversal casualty effect without dated sites and sufficient follow-up.",
  "secondSetting": "Attempt recovery of Edinburgh historical implementation dates and zone/network boundaries. Published primary research is an external benchmark, not an independent replication by this lab.",
  "severity": "Fatalities remain separate descriptive annual counts; serious/slight subdivision withheld from subgroup comparisons until adjustment and force transitions validated.",
  "publication": "Archive phase3; publish supported subgroup associations and an explicit feasibility result for every phase4 question. No where-it-works causal ranking. Phase5 economics is outside this release.",
  "amendments": [
    {
      "date": "2026-11-04",
      "beforeSubgroupModelResults": true,
      "reason": "Source audit found2024\u201325 LSOA vintage changes; use stable U/S parent universe for income instead of naive2011-code join. Data guide duplicated code23; official STATS20 resolves23 electricmotorcycle,33 poweredpersonaltransporter. Code99 unknown."
    },
    {
      "date": "2026-11-04",
      "beforeSubgroupModelResults": false,
      "reason": "Review showed family heterogeneity tests also depend on comparator; publish weighted and pooled tests side-by-side with separate Holm families. Clarify21-estimate Bonferroni families are separate by comparator, not joint42-estimate coverage."
    }
  ]
}
