Case note · NSW speed-zone changes · connected-vehicle data

Did lowering the limit lower the speeds?

Six NSW speed-limit changes evaluated against control roads with weighted difference-in-differences — speed spread tracked as well as the mean, and the bias in averaged data shown rather than hidden.

Treatment vs controlWeighted DiDRolling CoVPath-level vs aggregateInstrumented-car validation

Pooled, six sites · cars

−2.5km/h average operating-speed change, net of control roads
(published Dec 2025)

Study design

Every site gets a control, a window and a weight.

Six treated sites

Rural and peri-urban NSW roads where the posted limit changed (e.g. 100 → 80 km/h), taken from TfNSW's permanent-speed-limit change notices.

Matched controls

Comparable roads nearby with no limit change over the same period, so seasonal, weather and fuel-price effects are differenced out.

Before / after windows

Equal-length windows either side of the change date with a buffer around the changeover; vehicle class separated (cars vs all vehicles).

What is measured

1Average and median speedper direction, per window

285th and 15th percentilesthe fast tail and the slow tail

3Rolling 7-day coefficient of variationspeed spread — are drivers agreeing with each other?

Where the data come from

Connected-vehicle speed observations on each road, aggregated to road sections and hours, plus path-level trajectories for the aggregation-bias check. Speeds are validated against two instrumented SafeFuture research cars driven on the same roads.

No individual vehicle or driver is identified; site-level extracts are not published.

Published results · net of control

Two sites we have put on the public record.

Culburra Road, Shoalhaven

100 → 80 km/h, October 2025. Controls: Coonemia Road and Gerroa Road. Result published 11 December 2025.

−11.0km/h

westbound average speed, cars, net of controls

−8.9km/h

eastbound average speed, cars, net of controls

The Escort Way, Borenore

Controls: Peabody Road and Bowen Park Road. Result published March 2026.

−1.1 / −3.2km/h

cars, by direction, net of controls — a small change

−7 to −10km/h

all vehicles — heavy vehicles responded more than cars

Read these together. Same method, two very different responses. One site's drop is not a rule; which is exactly why each change is evaluated on its own, against its own controls, before anyone claims "lower limits work".

The method

Difference-in-differences: subtract what would have happened anyway.

Method diagram, not site data. The net effect is the observed after-speed minus what the control roads say it would have been without the change.

1Raw before/afterwhat most reports stop at — confounded by season, weather, traffic

2Control-adjusted (DiD)subtract the change on matched roads with no new limit

3Weighted + spreadweight by observations; track CoV so a calmer road is distinguished from a slower one

Every estimate is reported with its control comparison. When the control moves as much as the treated road, the honest answer is "no evidence of change" — and that is what gets printed.

Aggregation bias · Eltham Road flagship

Averaged segment speeds can mislead. We measure by how much.

Schematic. A section average mixes vehicles on different paths and journeys; the same section can look faster or slower than any vehicle actually travelling through it.

ΔBias = aggregate − trajectorycomputed for average speed, 85th percentile, and for the before/after change itself

✓Validated against instrumented carstwo SafeFuture research vehicles driven on the same roads

→Why it mattersa before/after result built on aggregates can overstate or understate the real change; we report both and the gap

Open the connected-vehicle speed audit → to see probe speeds compared with the instrumented cars.

Limits · what we will not claim

Operating speed changed. Crashes are a different question.

Speed is the outcome here — not crash reduction. Crash effects need years of after data and their own evaluation.

Only sites already on the public record are shown by name; the remaining sites are in the pooled figure.

Connected-vehicle data are a sample of the fleet; validation against instrumented cars bounds, but does not remove, that limit.

Reproducible

Same pipeline, same controls, re-runnable from the raw observations by a second analyst.

Transferable

Built for NSW; runs wherever connected-vehicle speeds and a change register exist.

Bring one speed-zone change. We evaluate it the same way — controls, windows, spread, bias — and tell you what the data can and cannot support.