Compare injury collisions or casualties against two English benchmarks. The estimates depend on comparison and trend assumptions; pre-policy differences and English policy changes prevent secure attribution.
Phase 3 complete · exploratory controlled analysis · causal gate not met
WHAT THIS RELEASE CAN SUPPORT
Both comparisons suggest a relative decline; attribution remains unresolved.
Certain: the national counts and calculations are auditable. Exploratory: the controlled estimates depend on an unverified comparison. They are not a measured number of collisions prevented by 20mph limits.
WEIGHTED ENGLISH REGIONS · SEASON + TREND
-9.7%
Modelled level association
95% calendar-HAC interval: -14.8% to -4.4%. The model allows Wales to have a different pre-existing linear trend.
Secondary stationary-bootstrap interval: -13.0% to -6.1%. Both intervals condition on the selected weights and model.
ALTERNATIVE · POOLED ENGLAND COUNTS
-3.1%
A materially different estimate
95% calendar-HAC interval: -10.1% to +4.5%. This comparison gives more influence to regions with larger event counts.
Comparison choice changes the magnitude and whether the interval excludes no change.
The primary window uses 80 pre-policy months (2015–2019 and January 2022–August 2023) and 27 post-policy months (October 2023–December 2025). September 2023 and the pandemic years 2020–2021 are excluded from the main fit. Counts cover all roads, not a verified inventory of roads changed to 20mph.
IDENTIFICATION CHECK · PRE-POLICY DIFFERENCES
The trajectories were already diverging.
The weighted comparison is chosen using pre-policy collision counts only. The same weights are used for casualties. It balances historical outcome trajectories, not traffic exposure, road mix, deprivation, weather, enforcement or local policy histories.
Solid green: Wales. Dashed amber: weighted English regional index. Grey: pandemic years, shown for context but excluded from the main model. Q 3 2023 is mixed; the monthly model excludes September alone. Both series are normalized separately to their 2015 mean.
Relative pre-policy trend
-1.2% per year in Wales relative to the comparison. Conditional 95% HAC interval: -1.9% to -0.5%. Parallel unadjusted trends are not supported.
A pre-policy holdout
Weights trained on 2015–2019 have an index RMSE of 6.3 points in training and 11.5 points in January 2022–August 2023. Mean Welsh index is -9.0 points below this held-out comparison.
Adding a linear trend is an assumption about what would have continued after 2023. It does not repair an unverified counterfactual or establish causation. The holdout precedes the policy, but the analysis was designed after published outcomes were known.
Comparison weights and pre-period balance
Nonnegative weights sum to 100%; pre-fit RMSE compares each regional 2015-normalized collision index with Wales. Final pooled fit RMSE: 7.3 index points.
English region
Weight
Individual pre-fit RMSE
Policy status
East Midlands
16.7%
9.5 points
Not verified untreated
East of England
0.0%
15.6 points
Not verified untreated
London
0.0%
24.9 points
Not verified untreated
North East
37.2%
8.0 points
Not verified untreated
North West
0.0%
16.6 points
Not verified untreated
South East
9.3%
11.6 points
Not verified untreated
South West
4.4%
14.6 points
Not verified untreated
West Midlands
24.6%
8.5 points
Not verified untreated
Yorkshire and The Humber
7.8%
11.3 points
Not verified untreated
Exact quarterly values and download
Display aggregates; the statistical model uses monthly observations.
Quarter
Welsh count
Wales index
Comparison index
2015 Q1
1,291
93.16
93.71
2015 Q2
1,343
96.92
98.20
2015 Q3
1,535
110.77
102.49
2015 Q4
1,374
99.15
105.61
2016 Q1
1,270
91.65
94.15
2016 Q2
1,164
84.00
91.02
2016 Q3
1,279
92.30
94.14
2016 Q4
1,209
87.25
93.39
2017 Q1
1,147
82.77
83.15
2017 Q2
1,107
79.88
81.69
2017 Q3
1,184
85.44
84.24
2017 Q4
1,116
80.53
85.72
2018 Q1
937
67.62
72.71
2018 Q2
1,027
74.11
79.05
2018 Q3
1,160
83.71
80.31
2018 Q4
1,090
78.66
83.62
2019 Q1
1,021
73.68
72.71
2019 Q2
1,082
78.08
73.48
2019 Q3
1,139
82.19
76.14
2019 Q4
1,075
77.58
84.54
2020 Q1
778
56.14
65.53
2020 Q2
441
31.82
37.66
2020 Q3
951
68.63
65.97
2020 Q4
704
50.80
64.04
2021 Q1
559
40.34
46.68
2021 Q2
889
64.15
70.22
2021 Q3
985
71.08
69.28
2021 Q4
868
62.64
69.76
2022 Q1
751
54.19
66.73
2022 Q2
815
58.81
63.42
2022 Q3
861
62.13
66.62
2022 Q4
888
64.08
70.63
2023 Q1
718
51.81
63.63
2023 Q2
875
63.14
65.99
2023 Q3
925
66.75
67.53
2023 Q4
737
53.18
68.84
2024 Q1
634
45.75
61.32
2024 Q2
783
56.50
67.53
2024 Q3
795
57.37
65.66
2024 Q4
701
50.59
69.25
2025 Q1
679
49.00
64.01
2025 Q2
788
56.86
66.90
2025 Q3
780
56.29
64.45
2025 Q4
757
54.63
69.00
ROBUSTNESS · SHOW THE ASSUMPTIONS TOGETHER
Which choices change the answer?
Negative values mean a lower fitted Welsh trajectory relative to the comparison. Intervals are 95% calendar-HAC except the separate post-slope model, whose average contrast uses the stationary bootstrap. These are conditional model intervals, not bounds on the causal effect.
Scale English region-years to official annual totals
-9.7%
-14.8% to -4.4%
80 / 27
Recorded 20/30mph is a category proxy, not a fixed treated network. Its result does not isolate the effect on roads converted from 30mph. Removing regions with documented rollouts does not certify the remaining regions as untreated. The short recent-pre window is especially sensitive to trend extrapolation.
Absolute model contrast and conditional detectability
The main model fits 6,647 collisions over the 27 post months, compared with 7,365 when its policy-period coefficient is set to zero: a model contrast of -718 (stationary-bootstrap 95% interval -986 to -441). Observed count: 6,654. This is not a count of events proved to have been prevented.
The bootstrap coefficient variability implies an approximate 5.2% reduction detectable at 80% power and a two-sided 5% threshold, conditional on this model. It does not establish causal design power for one national policy shock, and it excludes uncertainty in comparator selection.
FALSIFICATION · FAKE POLICY DATES
The model also detects breaks before the policy.
Each fake intervention has 12 months of follow-up wholly before the Welsh national change. Its comparison weights are fitted using only data before that fake date. The actual-policy row uses the first 12 full post months for a comparable horizon.
Temporal placebos are diagnostics, not independent randomized experiments or a calibrated permutation test.
Intervention month
Interpretation
Association
95% calendar-HAC interval
2017-01
Fake date
+6.8%
+3.0% to +10.6%
2017-10
Fake date
-0.9%
-5.7% to +4.1%
2018-10
Fake date
+7.4%
+3.9% to +11.1%
2023-10
Actual: first 12 full months
-10.4%
-15.5% to -5.1%
Positive placebo estimates are not evidence against a beneficial 20mph policy. They show that unrelated historical changes can produce apparent breaks in this model. One treated nation and a small set of dates cannot supply independent policy replications.
COUNTERFACTUAL ELIGIBILITY · GATE NOT MET
England had its own policy changes.
The primary-source check finds contemporaneous 20mph expansion. This is a bounded check, not a complete legal-order inventory. An unsuccessful search never counts as evidence that a comparison area stayed untreated.
Selected verified policy evidence. Dates of approval, implementation and reporting are kept distinct.
Causal release gate: not met. No exact dated treatment-road inventory or policy-clean area counterfactual has been verified. Differential reporting, traffic changes, concurrent measures and the historical trend breaks remain plausible explanations for part of the observed association. No premium, insurer-claims or cost-benefit conclusion follows from this release.
The downloadable ledger contains 7 policy records, source URLs, snapshots where available and date precision.
REPRODUCIBLE RELEASE · V 3
Inspect the data, model and remaining gaps.
The original DfT file contained 1,251,830 records for 2015–2025. The panel covers Wales plus nine English regions for 132 months, including pandemic sensitivity data. It retains records without coordinates and uses location geography rather than the reporting police force.
Source reconciliation
158 of 220 official annual comparisons match exactly. All 40 Wales and England national checks match. The 62 regional differences are at most 6 collisions or 7 casualties in a year; the annual-control sensitivity makes their impact visible.
All 240 phase 1 monthly checks match. 177 records use the source highway code where the current code is missing; 6 highway-country conflicts retain district-country classification.
Vintages stay separate
The model uses paired DfT data through 2025. Its Welsh 2025 totals are 3,004 collisions and 4,078 casualties; phase 2's later StatsWales series has 3,005 and 4,079. The unresolved recorded 20/30 category difference remains 18 collisions and 23 casualties.
Welsh provisional 2026 observations remain available in phase 2; they are not appended to a comparison ending in 2025.
Model, inference and reproducibility details
Each region and Wales are indexed to their own 2015 monthly mean. Nonnegative regional weights sum to one and minimize pre-policy collision-index error plus a fixed 0.01 penalty toward equal weights. No post-policy value selects weights. The model regresses the log of Wales divided by the weighted index on calendar-month indicators, calendar-year time and a post-period indicator. Gaps in the calendar are not compressed.
The headline interval uses Newey–West covariance with 12 calendar lags. The secondary stationary bootstrap uses random geometric blocks with expected length 12 months, centres residuals within uninterrupted calendar runs and keeps weights fixed. Main models use 1,999 draws; sensitivity models 999. These asymptotic/conditional intervals do not account for weight selection, unknown policies, unmeasured confounding or all uncertainty from short time series.
The specification is retrospective. Audit-driven amendments are logged, including the annual-table sensitivity and a review fix to avoid fixed-length bootstrap degeneracy in a 12-month run. Full original selected lines and source snapshots are retained with the versioned source; the public panel, hashes and Python script reproduce every displayed model estimate offline.
DfT and relevant government data: Open Government Licence where stated by the publisher. Attributed policy pages retain their publishers' rights. Analysis date 4 November 2026.