Road Safety News

Speed cameras deliver ‘significant reduction in road traffic accidents’

Friday 24th March 2017

A new study has concluded that, on average, the number of road traffic collisions reduces by 30% at sites where speed cameras are deployed.

The study, led by Professor Dan Graham* from the Department of Civil Engineering at Imperial College London, is based on data for 771 camera sites in eight areas across England - Cheshire, Dorset, Greater Manchester, Lancashire, Leicester, Merseyside, Sussex and the West Midlands.  For control sites the researchers randomly sampled 4,787 points on the network across the same eight areas.

The researchers developed an ‘approximate Bayesian doubly-robust estimation method’ to quantify the causal effect of speed cameras on collisions.

The paper says that previous empirical work on speed camera effectiveness, which shows a ‘diverse range of estimated effects’, is based largely on ‘outcome regression models’ using the Empirical Bayes approach, or on simple before and after comparisons.

The paper’s conclusion reads as follows: “We have developed an approximate Bayesian doubly robust approach for estimation of average treatment effects to analyse the impact of speed cameras on road traffic accidents. This is the first time such an approach has been applied to study road safety outcomes.  

“The method we propose could be used more generally for estimation of crash modification factor distributions.

“Our case study results indicate the speed cameras do cause a significant reduction in road traffic accidents, by as much as 30% on average for treated sites. 

“This is an important result that could help inform public policy debates on appropriate measures to reduce RTAs.

“The adoption of evidence based approaches by public authorities, based on clear principles of causal inference, could vastly improve their ability to evaluate different courses of action and better understand the consequences of intervention.”

*Dan Graham is professor of statistical modelling in the Department of Civil & Environmental Engineering, Director of the Statistical Modelling and Economics Research group, and Research Director of the Railway and Transport Strategy Centre within the Centre for Transport Studies at Imperial College London. He holds doctoral degrees from the Department of Mathematics at Imperial and from the London School of Economics, and has published more than 80 papers in refereed journals and two books.

He is currently engaged in projects on Bayesian inference for road traffic accident prediction and modelling the effects of transport interventions for ex-post evaluation. He provides advice to Government departments around the world, including the UK Department for Transport and the Treasury, and was appointed specialist advisor to the UK Parliamentary Select Committee on Transport. 

 Photo: © Copyright Albert Bridge and licensed for reuse under this Creative Commons Licence.


Comment on this story
Report a reader comment

What's your view - comment on this story:

I confirm that I have read and accept the moderation policy and house rules relating to comments posted on this website.
Your comment:
Your name and location:
Your email:

Are we talking about 'fixed' or mobile sites? Fixed camera sites might just make people pay more attention to the road and hence check their surroundings. Whereas mobile units, if people are not aware of them, make no difference to their driving style. Better serve democracy by judging people on HOW they are driving, rather than their speed.
Terry Hudson. Kent

Agree (4) | Disagree (1)

Obviously I've long been aware of the Solomon curve and the subsequent doubts of its validity Idris, I just didn't mention it in my last comment. I wondered, as I still do, how it was possible for Mr Solomon to come to those conclusions - my findings are based on more evidence than would have been available to him at the time I suspect.

From your second para, how would you actually know that you '..have probably driven more miles significantly below average speeds..'? You can't possibly have known the average speed on every road you drove on. In any event, if you thought it involved 'greater risk', then why do it?

When you infamously got your speeding ticket many years ago, according to the press it was in a 1938 Alvis Speed 25 (ironic name don't you think?) and it was at 47mph in a 30 mph - I can only assume that must have been on one of those very rare occasions when you were not driving below the average speed - in this case, not far off 150% higher I would think.
Hugh Jones, Cheshire

Agree (0) | Disagree (4)

My brief final comments: Wrong again Hugh, I provided a link to the Solomon Curve article that I was aware of but you, despite your long experience, were not. Another example of why basing views on the maximum available evidence is preferable to basing them on inherently limited personal experience.

That said, my preference for older cars, including pre-WW2, means that I have probably driven more miles significantly below average speeds, in free-flowing traffic, than most readers here. I can assure you that doing to involves greater risk. Indeed, I was tempted the other day by a 1937 Wolseley 16 Tickford tourer that reminded me of the 1933 Austin 16 Tickford version, unused since WW2, that I bought in exchange for a (brand new) tennis raquet and, with 3 friends, drove to Monte Carlo and back in the summer of 62. I decided against the Wolsely for one primary reason, that its cruising speed of 50 to 55mph would have meant too much time well below the average speed of modern traffic, and with very little power in reserve. Faster is often safer.
Idris Francis Fight Back With Facts, Petersfield

Agree (4) | Disagree (2)

Idris: I won't protract the discussion too much, but as expected, all you've done is produce evidence that you have read something on the internet - hearsay evidence in other words - not evidence of something that you yourself could substantiate.

On the other hand, I've measured the speeds of thousand of vehicles at different times on different roads etc. and time and again it was clear to me that most people's speed centres around the average speed - and that gives the best balance of reasonable progress and a safety margin. Once speeds increase above that range, safety margins decrease. Needless to say speeds below and up to the average are safer still with more time to see, react and stop. No speeding motorist, when asked by me if they could have stopped in time to avoid a collision from their illegal and well above average speed said 'yes' - good enough evidence for me.
Hugh Jones, Cheshire

Agree (3) | Disagree (8)

David: Why did you claim that I am "relying on statistics that are not in themselves of good quality" when last sentence of my comment stated that "my analysis showing no camera benefit does not depend in any way on Stats19 causation analysis but instead records what actually happens at 3,336 camera sites"?

On the narrower point that it is literally impossible for cameras to reduce collisions by 30% I should first point out that as Professor Graham fails to differentiate between Fatal and Serious collisions and others his 30% claim appears to relate to ALL injury collisions of which, according to Stats19, only 5% involve (or might involve) speeds above limits, even then not necessarily as the most important of the 6 causal factors allowed.

If you had reviewed Partnership data as I have you would know that average speeds fall by only a few mph and if you had read Professor Allsop's analyses for the RAC Foundation (the last two including my corrections of his errors) you will know that he failed to find any identifiable relationship between changes in speeds and changes in accidents. To any engineer, though not to true camera believers, that can only mean that the first does not cause the second.

The 30% claim is some 6 times greater than the Stats19 analysis shows for all injury collisions or some 10 times greater if we allow for "possible" and minor contributions, or some 12 to 20 times greater if we allow for cameras far from eliminating speeding. Are you seriously suggesting that your former colleagues consistently under-report speeding by a factor of at least 6, thereby implying that speeding is a causal factor in 50% of collisions or more? Surely not?
Idris Francis Fight Back With Facts, Petersfield

Agree (7) | Disagree (5)

Hugh first - I would not deliberately quote you out of context and I do not believe I did: some time ago, you wrote in effect, that you preferred your own opinions to evidence. I accept now that you meant opinions based on your own experiences but the problem then is that no individual's experiences can begin to match, or be properly representative of, the sum total of available information.

I am surprised that as a former road safety professional you are not aware of the risk/speed relationship I mentioned and which Google identifies as the Solomon Curve of 1964. As one might expect, there are other opinions too but that does not mean that I was wrong. Incidentally, anyone who drives, as I do, older cars with relatively limited performance or which need running-in at 50mph knows full well that risk increases below average speeds .

I agree that evidence needs to be produce. That's why my analysis includes detailed explanations of the methods and all the data, providing a complete audit trail from raw official data to results in four Excel sheets that allow quick and easy generation of large number of graphs in two formats for different combinations of police area, distance from camera and camera type. All confirm that there is no sensibly identifiable camera benefit, let alone 30% reductions, except for those for some areas that show significant increases in collisions where cameras have been installed.

Hugh wrote that "yet a lot of road safety and collision prevention is informed by [causation analysis] seemingly without question". In reality, ever since speed cameras provided automated speed measurements, computers automated administration, and awareness courses ever-increasing cash flow and exorbitant profit margins, road safety has been very seriously compromised by wholly disproportionate concentration on speeding as the be-all and end-all of road safety. It is not and never was.

He also claimed that "The problem with campaigners is that very rarely are they called upon to substantiate or produce evidence of their assertions, claims and 'facts' in any formal setting like a Court or Public Enquiry etc." He might well be right about some campaigners but it is no coincidence that the most effective anti-camera campaigners I know are all engineers. No coincidence because engineers, more than most, know that if they err their bridges collapse, their ships sink, their aircraft crash, their electronic circuits burn out or (in my case, electric wheelchairs run out of control), their careers end and their businesses collapse. An ever-present and far more demanding task master, in my view, than the relatively distant prospect of facing an enquiry.
Idris Francis Fight Back With Facts, Petersfield

Agree (7) | Disagree (4)

Well said David. Simply ticking AGREE doesn't do a comment justice sometimes, but David's last comment on the shortcomings of STATS 19 is an important one and worth me endorsing.

I have long held doubts about the usefulness of STATs 19 - in terms of contributory factors at least - and yet a lot of road safety and collision prevention is informed by it, seemingly without question.

I have in the past been dismayed at the attitude to collisions and their causes by some officers, some writing them off as 'just accidents' without really thinking about or even being concerned about the actions that preceded them. I've read accident reports which were so useless as to be a waste of time (e.g. "V1 collided with V2" or, in circumstances that were simply impossible due to either inaccurate location or direction of travel etc.). I should hastily add, to redress the balance, that equally I've also read textbook accident reports which told you everything you needed to know from which one could make a reasonable asessment of 'how it happened' and reassuringly, I've worked with officers who do taken the business of collision reporting and collision prevention very seriously. My point is STATS 19 should not be regarded as gospel - sometimes our own common sense and daily observations are worth more.
Hugh Jones, Cheshire

Agree (7) | Disagree (7)

Idris, your argument is based on STATS19. My argument is based on STATS19 being rubbish, and it comes from a career spent in the Police. RTCs are regarded by most Response Officers as nothing other than a complete nuisance and a total irrelevance to their jobs. Reporting of collisions is usually the work of probationary officers, or the very few who think that such a task is worthwhile. In time those rare officers usually move to Roads Policing. The STATS19 figures are derived from reports of young officers with the detective ability of potatoes, doing a task for which they have received no specialist training whatsoever. As well meaning as they might be, they often turn out to be incapable of investigating the contents of their own pockets, even if allowed to use both hands.

When such officers are faced with a vehicle whose driver has pulled out onto a main road and collided with another vehicle, to them it is a clear case of failing to give way, or failing to see another vehicle. An officer with more experience might see the speed of the vehicle on the main road as the major cause, as the driver was unable to stop in time when the other car pulled out, or he came into view at such a speed that the driver in the side road had a clear road when he checked. Crashes such as this are highly unlikely to have excess speed recorded as a cause, but in reality it is the elephant in the room.

In short Idris, you are relying on statistics that are in themselves not of good quality.
David, Suffolk

Agree (9) | Disagree (6)

Hello Idris. First of all, genuinely glad to hear you're well - we've missed your unbiased views and comments on all the many diverse road safety related subjects appearing on this forum (well, the ones which mention speed cameras at any rate).

I'm flattered that you've quoted me (possibly out of context, but no matter) i.e. 'I prefer my own opinions to evidence' and I presume I must have meant opinions based on the evidence of my own eyes, or my own self-acquired knowledge - either way, probably not your evidence, however I'll give you the benefit of the doubt and ask you, as an example, to please produce evidence of your 'fact' that ".. it was established decades ago that the safest speed is a little higher than the average..." My own first-hand evidence says otherwise, but I'm always open to other genuine evidence (that doesn't include something you've read on the internet!) Otherwise, we'll do as you predicted and er... ignore it.

The problem with campaigners is that very rarely are they called upon to substantiate or produce evidence of their assertions, claims and 'facts' in any formal setting like a Court or Public Enquiry etc. In my former life, it was taken as read that at some later date, I may have to substantiate and justify my actions, including things I said and it is a good way of focussing the mind and helps separate the wheat from the chaff. I suspect the same applies to David as I believe he is, or was, a police officer and his comments on this subject I would suggest, are well-informed and should be taken note of.
Hugh Jones, Cheshire

Agree (5) | Disagree (6)

Sorry David, but the Stats20 manual confirms that Stats19 data for the involvement of speeding in collisions are based on the reporting officer's best assessment of whether it having been "very likely" or "possible", not on drivers’ admitting they were.

Further, I used FoI to ask every police force for any example of speeding later having been found to have been a factor despite not initially having been recorded as such. Not one was able to provide even one example.

In any case, given that cameras far from eliminate speeding, it would need to be the primary causal factor in almost every accident (which it clearly is not) for cameras to reduce collisions by anything like Professor Graham's absurd 30%.

Your comment therefore seems to be another example of what I long ago termed the "My mind is made up, please do not confuse me with the facts" syndrome, exemplified by Hugh Jones of this parish who wrote not long ago "I prefer my own opinions to evidence", or what professionals in such conditions describe as “cognitive dissonance” (see Wikipedia).

Here are some more facts for you and no doubt others to ignore: (a) it was established decades ago that the safest speed is a little higher than the average – i.e. that driving slowly and becoming a mobile chicane can be as dangerous as driving faster (b) the law of unintended consequences ensures that any measure implemented to reduce risk will also increase it in some way, for example when drivers frequently tailgate me or cross double white lines to overtake me on twisty local roads where speed limits have been unreasonably reduced, or (as a matter of record in at least ten inquests) people die when cameras distract them from their primary task of driving safely.

I should perhaps also emphasise that my analysis showing no camera benefit does not depend in any way on Stats19 causation analysis but instead records what actually happens at 3,336 camera sites.
Idris Francis

Agree (11) | Disagree (9)

Idris is right to mention - rightly with a note of cynicism - the vested interest angle of some, from the continuation of speed cameras. Apparently there is a multi-million pound industry which has grown up around them and which preys on the gullible motorist, by persuading them that they must have speed camera detectors and similar devices designed to avoid detection. One prominent motorists' organisation apparently even gets a commission from the sale of such items if purchased through their website. Scare tactics and emotive advertising and marketing about 'protecting your license' etc. are used but starngely, all the time never promoting the obvious alternative, which costs nothing and is not that difficult i.e. don't speed! Funny that.

On a more positive note, I'm glad David has rightly highlighted the fact that the all-important speed before impact of collisions is almost impossible to determine by investigators and yet the inevitable consequential under-reporting of its importance is glossed over - especially if one is anti-speed enforcement. Sometimes using our common sense and our eyes will tell us a lot more about collsion causation than official stats - You Tube helps as well!
Hugh Jones, Cheshire

Agree (4) | Disagree (7)

To use STATS 19 figures in questioning whether it all makes sense is wrong. In how many crashes is the speed of vehicles established? Only a small percentage of collisions are investigated to the level required to establish speeds. Drivers are hardly likely to admit to Police that they were speeding when the other driver pulled out in fron of them, so guess what - it doesn't get recorded on STATS 19. This leads people to erroneously conclude that excess speed is less of a factor than it actually is. This 'fact' is then used a big stick to beat the Professor about the head when he reaches his conclusions.

Long after I am dead and gone, I suspect that there will still be people arguing that driving faster does not result in more crashes, but I am with Professor Graham on this.
David, Suffolk

Agree (9) | Disagree (5)

Thanks Nick. I am extremely cross with myself that it took me so long to realise that, far from being as complex as statisticians would have us believe, camera analysis is essentially simple when tackled from first principles and a great deal more data. I am also cross with myself that, as you note, I have still not published my analysis (though in the last three years I have suffered significant health problems, thankfully now on the mend).

All of the analysis was completed quite some time ago, what has taken too much time since then is writing and re-writing text, Power Point presentations and video versions to be readily understandable to the layman for their intended distribution by all available means to all possible audiences, the only way to force those in authority finally to admit that cameras have been a complete waste of time, effort and not least money - except of course to vested interests.

I will of course ensure that you and RSGB see the final versions as soon as they are finished.
Idris Francis

Agree (16) | Disagree (7)

You've been talking about your own analysis for quite a while now. When you do finally publish, we will happily cover it in a news item here to give other readers the opportunity to comment on it, as you and others have done with regard to Professor Graham's work.
Nick Rawlings, editor, Road Safety News

Agree (26) | Disagree (4)

I am loath to break my self-imposed silence here but this article and the comments on my long-time special subject leaves me no choice:

I am in the process of finalising my own far larger scale analysis of the effects of speed cameras, based on avoiding at all costs the methods used until now by statisticians, few if any of whom ever ask themselves the question that (we) engineers have to ask if our buildings are to remain standing, our aircraft to remain airborne or our electronic devices to avoid meltdown. It is a simple yet vital question that I learned as long ago as 1953 and have practised ever since - "Do the results of my calculations make sense??

That Professor Graham's claims of 30% reductions make no sense at all whatever is made clear by (a) Stats19 causation analysis from 2005 that shows that no more than 8% of FSC and 4% of Slight Injuries/collisions involve speeds above speed limits in the first place (b) that even those small numbers include "possible" involvement as well as "very likely" (c) that much of that is in any case of secondary rather than primary significance (i.e. many of those collisions would still have happened due to the other factors (d) cameras far from eliminate speeding in any case and (e) close to 40 adverse effects of cameras, including tailgating, dangerous overtaking, distraction etc were identified 15 years ago.

That it would therefore be impossible for cameras to achieve anything like that 30% figure implies that Professor Graham has merely repeated the fundamental errors of analysis perpetrated by others who have such absolute confidence in their off-the-shelf theories that they don't even consider the possibility that the results that emerge could be not just wrong but wildly wrong. See also "groupthink" on Wikipedia for other explanations!

The article identifies two fundamental sources of error, the first of which is the assumption inherent in the Bayes method that collisions at camera sites would have followed quite closely the same trends as collisions at "control sites" had their cameras not been installed. In the real world the volatility of Stats19 data results in trends in different police areas being very different indeed, and therefore even more so for the much smaller numbers are camera sites.

The second, even more damning, is that in his response to comments he claims that "our approach does adjust for RTM", apparently unaware that when the Fourth national report of 2005 finally admitted that previous reports had overstated camera effects by ignoring the effects of RTM, the its stated in 3.3 that "while it would be desirable to include some allowance for regression to mean no reliable method has yet been established for doing so".

That comment at least remains as true now as it was then, for a very simple reason: The degree of selection bias and therefore RTM falls relates to how selection thresholds compare to normal levels and also on the volatility of the data, itself a function of accident rates and therefore traffic volume. While it is quite easy to estimate RTM effects on a statistical basis using random or much better-defined data it is, and will remain, literally impossible to estimate RTM from site data with anything like the necessary accuracy because selection thresholds are inconsistent and other essential data is rarely available.

My solution is therefore not to try to adjust for RTM but to isolate it entirely, as my two methods do. The first, as applied to TfL data and reported here, showing that no London camera have any identifiable effect at all, is to compare the long-term reductions of large numbers of accidents at sites and elsewhere, from before selection bias starts to after it ends. The results for the 22 police areas for which the relevant camera site data is available are unequivocal - at best, no effect.

The second method, using large volumes of monthly data instead of annual totals, looks only for the quite short-term reductions which would logically start at installation and reach close to maximum within a year or so later. Those results are at least as compelling if not more so, because they do not depend at all on comparing accident rates at sites with those elsewhere but only for any deviations from the trends at those sites. And they show the same - at best no identifiable benefit.

I am copying this comment to Professor Graham with my request for a meeting to discuss these important issues.
Idris Francis

Agree (19) | Disagree (18)

A (rather sheepish) hello to you Mr Graham! It is unusual and probably unexpected, to hear back from authors/creators of studies which occasionally appear on this website and to get immediate feedback from comments made (and possibly have the rugs pulled from under our collective feet!), but your response is genuinely appreciated.

As you have rightly guessed, I’m not a statistician and tend to go by empirical evidence, although I do bow to your superior expertise in this field and hope my comment didn’t come across as disrespectful.

I actually worked in speed management and with speed cameras for many years and have long been convinced of their effectiveness and necessity in preventing collisions anyway and it was satisfying therefore to read your conclusion - even if I barely understood a single word of the process which led up to that conclusion!

Anyway, thanks for taking my comment with such good humour.
(ps: loved the way you used the resampling scheme proposed by Muliere and Secchi (1996) which extends Rubin’s bootstrap in a general Bayesian nonparametric context, for your restricted moment model...Who saw that coming?)
Hugh Jones, Cheshire

Agree (10) | Disagree (3)

As author of this paper I am grateful for the comments here. In response:

To Hugh Jones: This comment made me laugh, and I do sympathise with your view. In our defence, I should stress that this work is intended primarily for an academic audience (it is an academic paper after all, not a "report") who will be familiar with the jargon, concepts and maths used in the paper. In fact, ultimately we hope it will be the statistical contribution of the paper, as much as the substantive contribution, that will be of interest to the academic community. Hopefully, by developing new methods that are appropriate to the problem, even if they might seem overly elaborate, we are able to improve analysis and get better answers.

To David Weston: no there is no RCT here. The objective of our method is to develop models that mimic an RCT. My understanding is that RCTs are not feasible in this area for cost / ethical reasons, but I could be wrong. Of course I agree, that RCTs offer the best way of deriving inference if its possible to conduct them.

To Dave Finney: I am not familiar with the FTP method so I can comment on it, but I can confirm that our approach does adjust for RTM, so that should not be a factor influencing our results.

To Duncan MacKillop: Actually the model was estimated in levels not percentages, and the results then converted to percentages to ease interpretation. So either way we do find a significant reduction in collisions. But I can't agree with your suggested conclusion because the data don't support it.

Thanks again for the comments. Much appreciated.
Dan Graham, London

Agree (21) | Disagree (3)

The instant the finding is quoted in percentage terms, you know that the study has not reached the conclusion that people think it's reached. A quoted 30% reduction most often means that there is a negligible effect from the intervention because these big percentages just show that the innate variation in the system will have a similar effect. For example if there were three crashes on a stretch one year and only two after the intervention then that would be no more than the sort of change that you would expect with the random variation in the system.

Although the veracity of the statistical methods used are beyond doubt, the form and presentation of the results leaves a great deal to be desired. A much better presentation would be to say that generally speed cameras have an effect that is occassionally better than the background variation, but overall any significantly beneficial effect is far from being proven.
Duncan MacKillop. No surprise - No accident

Agree (12) | Disagree (8)

Reports like this one are why I developed the FTP method back in 2009. There are 2 problems when trying to determine the effects of speed cameras. The 1st is that no scientific trials have been run and, because of that, the 2nd is the effect of site-selection (often referred to as RTM).

This report tries to estimate what effect speed cameras have by using models. Models are essentially just a set of equations and they can contain many variables. Each variable can require some degree of estimate or judgement and this can introduce errors.

This report, therefore, has to be considered with caution, especially as it achieves the opposite result of those reports that use the more accurate FTP method and it's estimates of collision reductions are substantially higher than even involve speeding.

There is one way that will clear up the speed camera confusion and it's cheaper, simpler and more accurate than anything done so far. Simply run scientific trials. Scientific trials are perhaps more necessary now than ever before as we enter potentially the largest speed camera roll out in history.

BTW, the above report did NOT use a "randomised controlled trial".
Dave Finney, Slough

Agree (14) | Disagree (10)

To be honest, the first thing that flew into my mind was whether or not a "randomised controlled trial" was used; I am still oblivious to whether or not that was used.
David Weston, Corby

Agree (9) | Disagree (4)

As soon as I saw the this headline relating to yet another speed camera appraisal, my first thought – which I’m sure was shared by many other readers – was naturally ‘All well and good, but have they used an approximate Bayesian doubly-robust estimation method?’

Also, have they developed ‘an approach within a novel approximate Bayesian framework to derive posterior predictive distributions for the ATE of speed cameras on road traffic accidents, bearing in mind that a causal DR approach combines propensity score (PS) and OR models to give an average treatment effect (ATE) estimator that is consistent and asymptotically normal under correct specification of either of the two component models?’ Reading the report’s summary, you can imagine my relief when I saw that they had.

I then went for a lie down, put on one of my favourite psychedelic albums from the 1960s and contemplated the universe for a bit. It’s all gone a bit too far.
Hugh Jones, Cheshire

Agree (18) | Disagree (4)