SUPPORTED BROAD SUBGROUP MODELS
21
Pass the count and detectability checks
Intervals allow for testing several groups. Sparse or unknown categories retain their counts without an unstable fitted estimate.
Retrieving and verifying the requested analytical release. Result data loads before optional map boundaries and source checks.
Your filters and release stay in the address. Navigation remains available.
← STATS 19 research overview / Current publication and overview / Phase 3 comparisons / Phase 2 observations / Phase 1 protocol
20MPH EVIDENCE LAB · PHASE 4 · 4 NOVEMBER 2026
Explore who was injured, the roads involved and neighbourhood income. Broad groups can support useful comparisons, but the unresolved causal limitations from phase 3 still apply.
Phase 4 complete · subgroup evidence and feasibility · causal gate not met
WHAT THE EVIDENCE CAN SUPPORT
Certain: 1,620,726 casualty records reconcile to their collision totals, and every subgroup family sums to the all-road total. Likely: some differences reflect real changes in recorded injury burden. How much the 20mph policy caused remains unresolved.
SUPPORTED BROAD SUBGROUP MODELS
Intervals allow for testing several groups. Sparse or unknown categories retain their counts without an unstable fitted estimate.
ALL CASUALTIES · MODEL SENSITIVITY
The same overall outcome changes with the comparison. Neither number establishes a causal effect; subgroup results inherit this limitation.
EXPLORATORY ASSOCIATIONS · ALL SEVERITIES
All-road casualty counts; 80 pre-policy months in 2015–2019 and January 2022–August 2023, compared with 27 full post months from October 2023–December 2025. Negative estimates mean a lower fitted Welsh trajectory relative to that comparison.
| Group | Welsh pre / post count | Weighted comparison | Pooled England sensitivity |
|---|---|---|---|
| Pedestrian | 4,784 / 1,059 80 /27 months | -19.2% -29.9% to -6.9% | -19.4% -29.0% to -8.5% |
| Cyclist | 2,760 / 610 80 /27 months | -9.6% -31.6% to +19.3% | -9.3% -30.1% to +17.8% |
| Motorcyclist | 3,984 / 1,110 80 /27 months | -9.3% -27.1% to +12.9% | +7.2% -15.3% to +35.6% |
| Car occupant including taxi | 25,742 / 5,597 80 /27 months | -3.5% -11.1% to +4.8% | +4.6% -6.8% to +17.5% |
| Other known road user | 2,408 / 681 80 /27 months | -6.5% -24.9% to +16.6% | -0.1% -20.6% to +25.6% |
| Unknown | 0 / 0 80 /27 months | Estimate withheld Unknown / unmatched category: descriptive only; Fewer than100 pre-period casualties; Fewer than50 post-period casualties; Fewer than5 Welsh casualties in at least one included month; Nonpositive baseline or monthly series | Descriptive counts only |
Exploratory tests of differences among supported groups (Holm-adjusted within each comparison): Weighted p=less than 0.001 · Pooled England p=less than 0.001. These test model associations, not causal policy effects. A large value does not establish that groups respond equally.
Wales: solid green. Weighted English comparison: dashed ochre. Each series has its own 2015 baseline; the chart is not a fitted counterfactual.
| Quarter | Wales index | Weighted comparison index |
|---|---|---|
| 2015 Q1 | 109.43 | 100.44 |
| 2015 Q2 | 82.55 | 91.53 |
| 2015 Q3 | 87.26 | 88.26 |
| 2015 Q4 | 120.75 | 119.77 |
| 2016 Q1 | 87.26 | 107.53 |
| 2016 Q2 | 89.62 | 83.50 |
| 2016 Q3 | 87.74 | 83.71 |
| 2016 Q4 | 109.91 | 109.99 |
| 2017 Q1 | 95.75 | 102.75 |
| 2017 Q2 | 73.58 | 81.52 |
| 2017 Q3 | 81.13 | 81.50 |
| 2017 Q4 | 98.11 | 99.03 |
| 2018 Q1 | 90.57 | 83.73 |
| 2018 Q2 | 68.87 | 82.61 |
| 2018 Q3 | 74.06 | 82.17 |
| 2018 Q4 | 102.83 | 113.52 |
| 2019 Q1 | 99.06 | 97.04 |
| 2019 Q2 | 73.58 | 71.51 |
| 2019 Q3 | 75.00 | 75.31 |
| 2019 Q4 | 107.55 | 117.08 |
| 2020 Q1 | 72.17 | 85.57 |
| 2020 Q2 | 22.17 | 26.33 |
| 2020 Q3 | 60.85 | 63.08 |
| 2020 Q4 | 60.85 | 77.37 |
| 2021 Q1 | 33.96 | 48.34 |
| 2021 Q2 | 59.43 | 76.15 |
| 2021 Q3 | 69.81 | 68.92 |
| 2021 Q4 | 81.13 | 89.27 |
| 2022 Q1 | 68.87 | 91.85 |
| 2022 Q2 | 56.60 | 70.27 |
| 2022 Q3 | 54.72 | 71.55 |
| 2022 Q4 | 88.68 | 97.57 |
| 2023 Q1 | 67.92 | 86.45 |
| 2023 Q2 | 60.85 | 73.67 |
| 2023 Q3 | 70.75 | 71.71 |
| 2023 Q4 | 64.15 | 97.02 |
| 2024 Q1 | 58.96 | 84.47 |
| 2024 Q2 | 46.23 | 76.80 |
| 2024 Q3 | 50.00 | 68.57 |
| 2024 Q4 | 66.51 | 97.19 |
| 2025 Q1 | 53.30 | 88.04 |
| 2025 Q2 | 50.94 | 76.38 |
| 2025 Q3 | 41.98 | 69.50 |
| 2025 Q4 | 67.45 | 104.96 |
Smallest included Welsh month: 25; months below 5 casualties: 0. Approximate nominal detectable reduction under the weighted model: 12.2%. This conditional precision diagnostic does not establish power to identify the policy effect.
Pre-policy relative trend: -2.3% per year (nominal 95% interval -3.5% to -1.1%). A trend difference warns against assuming parallel trajectories.
| Year | Wales | England |
|---|---|---|
| 2015 | 848 | 21,525 |
| 2016 | 794 | 21,093 |
| 2017 | 739 | 21,706 |
| 2018 | 713 | 20,472 |
| 2019 | 753 | 19,774 |
| 2020 | 458 | 13,493 |
| 2021 | 518 | 15,366 |
| 2022 | 570 | 17,846 |
| 2023 | 559 | 17,760 |
| 2024 | 470 | 17,809 |
| 2025 | 453 | 18,657 |
SUPPORT BEFORE CLAIMS
Each model needs at least 100 pre-period and 50 post-period Welsh casualties, at least 5 in every included month, and a conditional detectable reduction no greater than 30% under both comparisons. Broad categories pool small groups in advance; no narrow age-by-road-user or street rankings are fitted.
The intervals use 12 calendar lags and a separate Bonferroni correction for each comparison across all 21 supported subgroup estimates. Family difference tests use joint covariance and a separate seven-test Holm correction for each comparison. Neither adjustment provides joint coverage across both comparisons. These adjustments address sampling and repeated testing; they cannot remove confounding, uncertain policy histories or comparator contamination.
| Family / country | Pre share | Post share | Change |
|---|---|---|---|
| Road user · Wales | 0.00% | 0.00% | 0.00 points |
| Road user · England | 0.01% | 0.01% | +0.01 points |
| Age · Wales | 0.66% | 1.16% | +0.50 points |
| Age · England | 1.93% | 2.20% | +0.27 points |
| Urban / rural setting · Wales | 0.03% | 0.09% | +0.06 points |
| Urban / rural setting · England | 0.01% | 0.00% | -0.01 points |
| Road class · Wales | 0.00% | 0.09% | +0.09 points |
| Road class · England | 0.00% | 0.00% | 0.00 points |
| Road layout · Wales | 0.26% | 0.11% | -0.15 points |
| Road layout · England | 1.60% | 2.10% | +0.50 points |
| Neighbourhood income deprivation · Wales | 2.57% | 2.48% | -0.09 points |
| Neighbourhood income deprivation · England | 0.90% | 0.79% | -0.11 points |
| Recorded speed category · Wales | 0.00% | 0.00% | 0.00 points |
| Recorded speed category · England | 0.01% | 0.00% | -0.01 points |
These annual Welsh fatality counts are descriptive, with no fitted subgroup or policy-effect claim. Serious/slight comparisons are withheld while severity adjustments and force recording transitions remain unresolved.
| Year | Fatalities |
|---|---|
| 2015 | 105 |
| 2016 | 103 |
| 2017 | 103 |
| 2018 | 103 |
| 2019 | 98 |
| 2020 | 71 |
| 2021 | 86 |
| 2022 | 95 |
| 2023 | 101 |
| 2024 | 84 |
| 2025 | 91 |
ALL-NETWORK SCREEN · ADJACENT ROADS UNRESOLVED
The recorded-speed view splits all casualties into 20/30mph,40mph-or-above and other/unknown categories, which reconcile to the full network. It is a broad screen only: it does not locate roads neighbouring a changed limit or distinguish diverted traffic from secular changes.
The current Welsh map is not a verified complete historical network. The bounded source review did not establish exact effective dates and adjacent-road coverage sufficient for a displacement estimate. All-network and complementary speed-category comparisons are available; an adjacent-road causal test remains unsupported.
Required next evidence: a complete dated treated-road inventory including zero-event sections, a fixed adjacent-road buffer, and compatible traffic counts. An unmatched map feature stays unknown; it is never assigned treatment from its collision speed label.
WREXHAM · EXPLORATORY SPEED FOLLOW-UP
July2025 · 20MPH LIMIT
85th percentile: 30.4mph. Vehicle observations: 526,530. These are repeated vehicle passages, not independent policy experiments.
September2025 · two weeks · 30MPH LIMIT
85th percentile: 31.8mph. Vehicle observations: 533,767. These are repeated vehicle passages, not independent policy experiments.
At five sites, the published mean increased2.4mph between the July20mph and September30mph observations. September covers only two weeks. The source check found no later published site series, exact site-level reversal dates or matched injury follow-up sufficient for an injury-effect estimate. The short before/after comparison does not isolate the legal-limit effect from other changes. The five remaining Wrexham20mph sites averaged24.1 then24.0mph, but they were not assigned or matched controls. The published combined-volume footnote differs from its two source rows by180 observations; the two rows are retained here.
Read the official monitoring report. This is a different five-site sample from the 38 retained20mph monitoring sites in phase2.
SECOND SETTING · REPLICATION ATTEMPT
The authors’ repository contains seven historical zone polygons with phased dates and7,478 road features. That enables a concrete second-setting check. We reproduced aggregate calculations from the primary paper and author workbook; original collision-to-road assignment, matched controls, uncertainty intervals and causal attribution were not independently replicated. The historical research map is not a verified legal-order inventory, and its reuse licence remains unverified.
| Zones | Implementation |
|---|---|
| 1 · two polygons | 31 July2016 |
| 2 and3 | 28 February2017 |
| 4 and5 | 16 August2017 |
| 6 | 5 March2018 |
The published Table2 rates reproduce differences in percentage changes of−22.31 percentage points for collisions and−19.71 points for casualties. Author zone counts and rounded follow-up durations give−22.13 points for collisions when combined with the paper’s control rates. These are percentage-point contrasts, not relative risk reductions. A separate Table1 casualty contrast gives−9.15 points versus the reported−10 using printed inputs; that discrepancy remains explicit. Table2 omits zone2 while the workbook total includes it; the underlying control pairing is not reconciled.
Primary study · Authors' analysis and historical GIS · Our reconstruction and limitations
REPRODUCIBLE RELEASE · V4
Every retained casualty joins exactly once to a collision; no orphans, duplicate casualty keys or declared-count mismatches. All 1320 monthly checks against phase3 match, and all 10,560 exhaustive subgroup totals reconcile. Location geography governs the analysis.
The specification is retrospective. Phase3 collision-trained English regional weights are fixed for every subgroup, without fitting to subgroup post-policy outcomes. Each series is normalized to its own 2015 mean; the log Wales/comparison ratio is regressed on calendar month, linear calendar time and a post-period indicator. Pooled England uses aggregate counts with the same model.
The 2019 income release measures 2015–16 conditions and uses 2011 LSOAs. An independent ONS crosswalk retains unchanged areas and uniquely split children within their original parent, excluding merged/complex areas in every period. Directly matching old codes would have created artificial missingness after the 2024 STATS19 geography change. Income results describe that consistent eligible geography, not every Welsh road.
The current data guide repeats code 23 with conflicting labels. The official STATS20 specification and conversion table identify 23 as electric motorcycle and 33 as personal powered transporter; the source inconsistency is logged. Unknown type 99 remains unknown. Car occupants include taxis; e-scooters are in other known road users.
No adjusted serious-injury, road-user exposure, insurance premium or economic benefit claim is made. Recorded serious/slight changes can reflect police reporting changes. Downloadable results include every supported and withheld group, with the same labels and precision inputs used here.