Toronto is the flagship. Seven other jurisdictions on three continents run the same direction.
Eight jurisdictions on three continents
Toronto is the flagship; the others are the Lab's own analyses of each jurisdiction's public data. Every jurisdiction with crash or speed data runs the same direction: enforcement present, fewer crashes or lower speeds; enforcement withdrawn, the reverse. Magnitudes vary from near zero in Toronto's calendar-matched designs to −22 percent in New York.
| Jurisdiction | Headline numbers | Direction |
|---|---|---|
| Toronto (era, 2020–25) | −4 to −7% crashes (defensible range); injuries OR 0.920, p = 2.3 × 10⁻⁵ | Same |
| Toronto (camera-free, 2026) | Crashes null at former sites; injuries −28.6 pp vs control, calibrated p < 0.0001, not fading | Consistent |
| New York City | −14 to −22% near 375 fixed cameras (p < 10⁻¹⁵); 66% of locations improved; injury probability −3.7%; 2.24M records | Same |
| Waterloo Region | Toronto injuries −28 to −55 index points vs no-camera control (p < 0.006); parallel trends holds for totals, fails for injuries | Same |
| Vaughan | Speeding −56% in 42 days; +59% rebound within weeks of cancellation; speed evidence only | Same (speed only) |
| Copenhagen | −8.4% crashes in camera municipalities (p = 0.052; rural −12.2%, p = 0.011); effect growing after 6 fixed years | Same |
| Alberta / Edmonton | Speeding +3.0 pp citywide after ~70% of sites removed (+10.8 on 80 km/h roads); five prior studies −11 to −25% | Same |
| London | KSI ratio near cameras 1.09 → 1.02 after installation; 304 cameras, 652,697 records | Same |
| Tel Aviv | 120 hidden section cameras tendered Dec 2025; design tests technology, road type, secrecy | No data yet |
Vaughan, and why the distinction matters
Vaughan's evidence is about speed, not crashes or injuries. Its value is the cleanest measured speed rebound after a programme cancellation that we have.
Vaughan's cameras ran 42 days before council cancelled the programme. Inside that window daily speeding fell 56 percent and average speeds fell 19.4 percent, and within weeks of cancellation daily speeding rebounded 59 percent. Only 15 collisions occurred near those camera sites during enforcement, far too few to test crashes or injuries in either direction, so Vaughan shows neither a crash decline nor its absence.
Vaughan gives us the behavioural half of the Toronto story, measured directly. Toronto's departure rebound needs drivers to speed up again once enforcement stops; Vaughan measures exactly that, on speeds rather than crashes, in a window short enough that nothing else changed. The speed-to-crash link itself rests on the other cities.
| Measure | Value |
|---|---|
| Days of enforcement before cancellation | 42 |
| Daily speeding during enforcement | −56% |
| Average speeds during enforcement | −19.4% |
| Daily speeding after cancellation | +59% within weeks |
| Collisions near camera sites during enforcement | 15 (too few to test) |
New York City
Our largest companion analysis, and the one that established the method we brought to Toronto. We matched 2,240,376 crash records from July 2012 to February 2026 against 375 geocoded camera locations, derived from 1,253,508 Department of Transportation Notices of Liability, using the same 250-metre buffer. The staggered rollout between June 2015 and May 2016 supplies the before-and-after.
| Measure | Value | p |
|---|---|---|
| Crash frequency near cameras after installation | −13.9% | 0.048 |
| Per-camera paired analysis (Wilcoxon) | −22.2% | <10⁻¹⁵ |
| Camera locations showing a decrease | 66% of 375 | n/a |
| Injury probability per crash | −3.7% (RR 0.966) | <10⁻¹¹ |
| Fatality probability per crash | −13.3% (RR 0.867) | 0.064 (n.s.) |
| Crashes within 250 m of a camera | 149,235 of 1,992,722 | n/a |
The fatality result points the right way and does not reach significance, because fatal crashes are rare even in two million records. The crash-level difference-in-differences interaction is not significant either (p > 0.3), so the camera-specific change in severity is not separable from the citywide trend in that framework. Toronto's rotation let us go further.
London
London is the place where speed cameras grew up. Toronto's cameras are five years old; London tells us what the next twenty might look like.
We analysed 652,697 collision records from 2000 to 2024, drawn from the UK's STATS19 series, against 304 camera locations. The first Gatso cameras appeared in the early 1990s, roughly 700 were operating by 2005, and cameras now detect 97 percent of all speeding offences in the Metropolitan Police area. Twenty-five years of continuous operation lets us ask not only whether cameras reduce crashes when installed, but whether they keep working.
| Finding | Value |
|---|---|
| KSI ratio near cameras, before and after installation | 1.09 → 1.02 |
| Regression to the mean, share of the headline reduction (Mountain & Maher 2005) | ≈ 60% |
| Camera effect once RTM is removed | 15–22% |
| Drivers who brake for the camera and speed up again (Corbett & Simon 1999) | 33% |
| Average-speed (SPECS) cameras vs fixed (Owen et al. 2016) | 36% vs ~22% casualty reduction |
| Dedicated roads policing officers, England and Wales, since 2015 | −21% (≈4,000 → 3,100) |
London and New York produce opposite raw patterns, and both are right. In New York, collisions near cameras are less severe than the city average (risk ratio 0.979). In London they are more severe (1.027). The contradiction dissolves once you know the placement rules: New York puts cameras near schools, which are relatively safe locations, while UK policy requires cameras at sites with four or more killed or seriously injured in three years. Control for site selection and both cities converge on the same 15 to 22 percent reduction. That convergence across countries, methods and decades is among the stronger reasons to think the association is real, though separating camera effects from regression to the mean remains hard everywhere.
Preliminary: we hold roughly 38 percent of London's 800-plus cameras and have no per-camera installation dates, so the London figures sharpen the questions we ask of Toronto rather than settling them.
Denmark
The first empirical crash-count evaluation of Denmark's fixed ATK camera programme, which had been evaluated on speeds but not on collisions. We ran a municipality-level difference-in-differences on Statistics Denmark's StatBank series across 11 treated municipalities and 88 controls, 2014 to 2024, excluding 2018 as the transition year so the comparison runs 2014–2017 against 2019–2024.
| Measure | Value | p |
|---|---|---|
| Difference-in-differences estimate | −4.35 accidents per municipality per year (−8.4%) | 0.052 |
| Rural municipalities | −12.2% | 0.011 |
| Effect after six years of fixed operation | still growing | n/a |
Denmark matters for a specific reason: its cameras are fixed and long-lived, and the effect appears to grow rather than decay over six years. Toronto's rotation cannot test that, and the two designs answer different halves of the durability question. Standard errors are HC1-robust; the headline sits just outside conventional significance and the rural result inside it.
Alberta and Edmonton
Alberta gives us the other cancellation, alongside Vaughan and Toronto. Edmonton's Intersection Safety Device cameras supply the crash analysis, and the province's April 2025 photo-radar restrictions supply a second natural experiment on what happens when enforcement is pulled back.
| Measure | Value |
|---|---|
| Crash counts at ISD intersections, 2021 to 2023–24 | −14% (Empirical Bayes mean −14.2%, median −17.5%) |
| Intersections improving | ≈ 79% |
| Driver-feedback-sign halo within 500 m | ≈ 5.5 pp lower speeding |
| Speeding after the April 2025 restrictions | +3.0 pp citywide |
| Same, on 80 km/h roads | +10.8 pp |
| Monitored sites worsening after the restrictions | ≈ 72% |
We used an Empirical Bayes before-and-after estimator here precisely because ISD cameras sit at high-crash intersections, where regression to the mean is strongest. Even so, the crash association cannot be cleanly separated from secular trends and post-COVID normalisation, and we say so. A recording change complicates the injury series: Edmonton inflated minor-injury counts roughly 2.4-fold in 2022–23 while serious and fatal counts stayed stable, so the injury analyses lean on fatal-plus-major or property-damage-only and model the break explicitly.
Tel Aviv
Israel tendered 120 hidden section-control cameras in December 2025. Section control measures average speed over a distance rather than at a point, which removes the brake-and-accelerate behaviour London documented, and the cameras are hidden, which tests whether visible warning is doing the deterrent work. The design varies technology, road type and secrecy together. No collision data exist yet; this is the study we expect to learn the most from next.
Why Toronto is the flagship
The rotation makes Toronto unusual. Permanent-camera programmes, including New York's and London's, can measure what happens when a camera arrives. Only a rotating programme can measure what happens when one leaves, at scale, across hundreds of sites, with each site serving as its own control. Toronto's 627 enforcement periods supply 383 testable departures. Bill 56 then switched off all 520 at once, which converts the design into a prospective test that we sealed before the data existed.