{
  "id": "social-factors-robustness",
  "version": "v3",
  "phase": "6: persistence and robustness diagnostics",
  "amendedOn": "2026-10-10",
  "parent": "/research/social-factors/exploratory-v2/protocol.json",
  "purpose": "Explain the temporal/spatial discrepancy and assess sensitivity of the existing exploratory findings. Preserve v1/v2 and unsuccessful results.",
  "priorInspection": "The complete v2 results, including 2023\u20132025 temporal holdouts and 2025 regional transfer, were inspected before this diagnostic amendment. London, North East and East Midlands transfer losses were visible. All v3 choices are post-result diagnostics, not a fresh preregistered confirmation. No unseen validation year remains in this data window.",
  "scope": "Reuse exactly133authorities and28primarycrimeforces, original eligible rows, same paired comparisons and v2 source definitions. The29force recording-break sensitivity remains available in v2; it is not the primary v3 study. No new outcomes, inputs or imputation. Keep realised target-year mileage as conditional exposure.",
  "freeze": "Record this protocol, v3 code, v2 model code, v2 frozen design, fits, predictions and results hashes before computing the diagnostics. Archive model transformations, folds, diagnostics and predictions. Source corrections require another version.",
  "comparisons": "All six authority comparisons and all five primary28force crime comparisons from v2. No selection of a winning alternative. Variants are individually changed assumptions, not their searched combinations.",
  "variants": [
    {
      "id": "fixed-2023",
      "name": "Fixed 2023 model",
      "definition": "Reuse coefficients, scaling, history credibility prior and regional coding from the existing2023fit. Predict2023,2024,2025 using each assessment year\u2019s two-year-lagged covariates. This tests parameter persistence with updated inputs, not a multi-year forecast made inJanuary2023. Compare with the original expanding fits on identical rows. No spatial result is invented for this temporal-only variant."
    },
    {
      "id": "poisson",
      "name": "Poisson count family",
      "definition": "Refit all compared models as Poisson log-link counts with the same offset, features, folds and numerical ridge1e-6. Use only training data. Report dispersion to make any lack of fit explicit. This checks sensitivity to count-family weighting; no claim that Poisson uncertainty is adequate."
    },
    {
      "id": "fixed-imd2015",
      "name": "Fixed 2015 deprivation",
      "definition": "Use each area\u2019s2015income/employment scores, recovered from its2018row, in every model year. For force units retain the original fixed2011population-weighted context proxy. All other inputs and folds unchanged. Tests dependence on the2019vintage switch, not a preferred specification selected from results."
    },
    {
      "id": "no-region",
      "name": "Without regional intercepts",
      "definition": "Remove regional indicator columns from every compared model and retain all other features, folds and negative-binomial family. Tests reliance on region offsets and their unseen-region fallback. Do not select it as a replacement winner using the already inspected holdouts."
    }
  ],
  "folds": "Retain v2 primary temporal2023\u20132025 and leave-one-region-out2025 folds. Fixed2023 uses the2023fit for all temporal assessments. Other variants refit on the same training rows as v2. Training excludes2020\u20132021; pandemic stress results remain in v2. No holdout recalibration.",
  "metrics": "Same paired held-out Poisson deviance, observed/expected, exposure-weighted absolute rate error and relative deviance gains as v2. Temporal annual and combined summaries; regional point results and combined spatial summary. Never interpret deviance reduction as percentage fewer collisions or insurance savings.",
  "regionalInfluence": "For each originalv2spatial comparison, report baseline and added-model deviance by region, absolute deviance saved/lost, exposure share and contribution to combined gain. Recalculate combined metric with each region omitted from scoring only, without refitting or removing it from the published primary result. These deletion diagnostics do not establish validity outside the omitted region.",
  "featureSupport": "For each v2spatial comparison, derive continuous feature ranges from that fold\u2019s training data. Flag assessment rows outside any expanded-model training minimum/maximum, excluding calendar year and region indicators. Compute thresholds and history transforms using training only. Record each feature and flagged count/mileage share; do not drop or clip assessment rows. Inside marginal ranges does not prove joint support.",
  "calibration": "For each originalv2comparison and fold, create five mileage-weighted risk bands from expanded-model training predicted rates. Assign assessment rows using those frozen thresholds. Compare observed and predicted rates and observed/expected for both models in the same bands. Pool bands across folds explicitly as within-fold ranks, with annual boundaries archived. Show row counts and region counts. No refitting or calibration correction on assessment data.",
  "uncertainty": "Reuse v2\u2019s2000paired region-block bootstrap over frozen predictions for combined temporal/spatial gain intervals, seed20261010. Only9region blocks; intervals conditional on fitted models, no refitting or input measurement error, unadjusted for multiple diagnostic comparisons. Region-only metrics are point estimates; calibration bands are descriptive and do not receive unsupported independent-row intervals.",
  "failedFits": "Record numerical convergence failures; never label missing/failed comparisons as zero gain. Abort incomplete output publication until the failure is explicitly represented or a numerical implementation correction is documented. Do not change hypotheses or data to obtain convergence.",
  "decision": "V3 never promotes v1/v2 hypotheses and does not turn an improved sensitivity into proof. Report sign changes, calibration and geographic failures as well as improvements. Promotion still needs the original controls, independent review and fresh validation;2025cannot be relabelled as an unseen holdout after inspecting these diagnostics.",
  "website": "Add a robustness view with study/comparison controls, frozen-versus-updated model results, assumption sensitivity matrix, regional contribution/range support/calibration tables, versioned downloads and an8phase progress register. Link from exploratory results, DataLab and RoadRiskLens. No new insurer pricing factor or causal claim.",
  "inheritedLimitations": [
    "The fixed complete-history cohort was selected using2016\u20132025availability, so it is retrospective, not an operational first-origin population.",
    "Excluding pandemic outcome years does not remove pandemic lagged predictors:2022uses2020history/traffic;2023uses2021.",
    "Income and employment scores are highly correlated. Joint block comparisons do not identify independent causal or pricing effects of either score.",
    "Poisson changes large-count weighting as well as its variance assumption; improved aggregate calibration does not independently validate a social factor."
  ]
}