{
  "version": "20mph-phase3-v3",
  "writtenAt": "2026-11-04",
  "status": "Retrospective exploratory specification, written before calculating the phase3 comparison estimates; published Welsh outcomes already known. Not prospective preregistration.",
  "amendment": "Use the approved all-road area fallback. Dated legal road inventory and policy-clean English donors remain unavailable, so estimates are controlled associations and the causal release gate remains closed.",
  "geography": "Wales versus nine English regions; stable national/regional boundaries using current highway authority lookup, checked against RAS0402 annual controls. Police force country only an audit, never the primary geography.",
  "pre": [
    "2015-01/2019-12",
    "2022-01/2023-08"
  ],
  "post": "2023-10/2025-12",
  "transitionExcluded": "2023-09",
  "outcomes": [
    "all-road injury collision counts",
    "all-road total casualties"
  ],
  "weights": "Nonnegative English regional weights summing to one. Fit collisions indexed to each region's mean2015, against Wales index, using80 pre months only. Minimise mean squared index error plus0.01*sum((weight-1/9)^2). Same weights used for casualties. Separately fit2015\u20132019 and assess2022\u2013Aug2023 as a pre-policy holdout, without tuning against that holdout.",
  "model": "OLS of log(Wales index / weighted comparison index) on intercept, eleven calendar-month indicators, calendar-time linear trend and post indicator. Main coefficient is a constant multiplicative level association. Time does not compress pandemic gaps. No mileage denominator or severity subgroup claim.",
  "uncertainty": "Headline95% calendar-lag Newey-West interval with12-month maximum lag and finite residual-degrees adjustment. Secondary1999 deterministic-seed stationary-bootstrap draws, expected block length12 months, sampling within contiguous included calendar runs, fixed comparison weights; sensitivities999 draws. Residuals centred within each run.95% percentile intervals conditional on model and weights; short runs remain limited. Neither method captures unmeasured confounding or donor-policy uncertainty.",
  "diagnostics": "Indexed quarterly pre/post plot, pre-only annual relative trend with HAC uncertainty, weighted pre-fit RMSE, holdout RMSE, annual geographic reconciliation, temporal placebo dates2017-01,2017-10,2018-10 with12-month follow-up and weights re-fitted only before each fake date.",
  "sensitivities": [
    "no relative linear trend",
    "separate post slope",
    "include2020\u20132021",
    "recent pre2022\u2013Aug2023 only",
    "end follow-up2024-12",
    "omit London",
    "omit London,South East,South West",
    "pooled England count index",
    "leave each English region out",
    "recorded20/30 category proxy, not a fixed treatment network"
  ],
  "absolute": "Sum of fitted post counts with the policy-period terms minus the same fitted model with those terms zero; a model contrast, not an observed or causal count of prevented events.",
  "power": "Conditional80% power approximation from block-bootstrap SD of level coefficient: 1-exp(-(1.96+0.842)*SD). Describes model sensitivity, not causal design power.",
  "decision": "Publish all diagnostics and sensitivity estimates together even if signs disagree. No causal wording unless verified counterfactual and road or area identification gates are met. Policies and incomplete inventory currently prevent that gate.",
  "amendmentsAfterDataAudit": [
    {
      "date": "2026-11-04",
      "reason": "62 of220 annual comparisons differ, all at English regional level, maximum6 collisions/7 casualties; Wales/England and phase1 match exactly. Current highway region versus table geography differences retained, not silently corrected.",
      "addedSensitivity": "Scale each English region-year2016\u20132025 to official RAS0402 total, retaining monthly profile; re-fit weights from scaled pre data.2015 unchanged. This sensitivity was added after first estimates, for a data-audit reason."
    },
    {
      "date": "2026-11-04",
      "reason": "Independent review found fixed-length circular blocks degenerate for the isolated12-month actual-policy follow-up. Before publication, changed to random geometric block lengths and made calendar-HAC the headline interval. Point estimates unchanged.",
      "status": "Review-driven amendment after initial exploratory estimates, not prospective selection."
    }
  ]
}
