Road safety
Which streets have more collisions than they should, and how sure we can be.
Counting
A busy road has more collisions because it has more traffic. That alone does not mean anything is wrong with it. The question that matters is whether a street has more collisions than a street like it would. Ask that, and you stop defending a list that only shows where the traffic is.
Expected
One model learns from every street in Great Britain, including the ones with no collisions. Traffic, speed limit and road type go in. So do junctions, lorries and buses. Each street is predicted by a version of the model that never saw it. That is how the side street with a real problem shows up, where a plain count would bury it.
Chance
Collisions are rare, so counts jump about. The best estimate blends the prediction with what happened. It leans on the record where plenty of collisions are expected, and on the prediction where few are. That is the Empirical Bayes method, and it allows for regression to the mean. So the scheme budget is not spent on a street that was only unlucky.
Tested
Each run ranks streets on the first three years only, then looks at the next two. In those two years, its picks had nearly twice as many collisions above expected as a plain count's picks. That is a list with a clear reason for every street on it.
Ways to rank
Rank against a street like it on injury collisions or on killed or seriously injured. Or rank what was recorded by count, KSI, cost to society or points. Or rank on risk alone: where the model expects the most collisions a year.
The data
Each injury collision the police record (DfT STATS19) is placed on the nearest street within 30 metres, over the latest five full years. KSI is as each force records it. The rankings update when new collision data or traffic estimates arrive.
Count the collisions on each street, sort the list, and you mostly find busy roads. A busy road has more collisions because it has more traffic. That alone does not mean anything is wrong with it. The useful question is different. Does this street have more collisions than a street like it would have?
Every street in Great Britain goes into one model: over a million of them, including the many with no collisions at all. It learns how collisions per kilometre go with traffic and the speed limit. Road type, carriageway and junctions go in too. So do lorries and buses. Each street is predicted by a version of the model that never saw it. That way a bad street cannot teach the model that it is normal.
Collisions are rare, so counts jump about. A quiet street with three collisions in five years may simply have been unlucky. So the best estimate for each street blends the prediction with what actually happened. The record gets more weight where more collisions are expected on each kilometre of street. The prediction gets more where few are. This is the Empirical Bayes method in the Highway Safety Manual. It allows for regression to the mean: a street picked for one bad year usually looks better the next, whether anything is done or not. Streets are ranked on how far that best estimate is above a street like it.
Each run checks itself. It ranks streets on the first three years only, then looks at what happened in the next two. In those two years, the streets it picked had nearly twice as many collisions above expected as the streets a plain count picked.
Against a street like it, you can rank on all injury collisions or on killed or seriously injured (KSI). There is also a ranking per kilometre for all injury collisions, on streets of 200 metres or more.
On what was recorded, you can rank four ways:
Or rank on risk alone: where the model expects the most collisions a year, whatever has been recorded.
Each street's place is plain: “12th of 6,234”, with 1st the most. Streets on the same figure share a place, written “joint 40th of 6,234”: one more than the number of streets with more.
Most streets have no injury collisions in five years. A street with none just says “Collisions recorded: None”, with no place for them, and shows only its expected risk, which every street has (for example “1,416th of 6,234”).
We use the injury collisions the police record (DfT STATS19), for the latest five full years. Each is placed on the nearest street within 30 metres. Police forces record serious injuries in different ways. The KSI figures use the severity the police recorded, without DfT's adjustment for those differences. Streets come from Ordnance Survey. Traffic comes from our estimates, built on DfT counts and any surveys you upload. The rankings update automatically when new collision data or traffic estimates arrive.
A ranking says where to look, not what is wrong. It does not replace a site visit or an engineer's judgement. A street near the top may already have had something done about it.