One in three

Paul Smith asserts, based on his analysis of road fatality data, that speed cameras are now responsible for a third of road deaths. He justifies this by reference to a "loss of trend" in fatality reductions starting after 1993, when, according to him, speed cameras became widespread. Does this argument have merit?

This is my view with some comments which occurred to me as I investigated the matter. For a more authoritative discussion see this document from PACTS.

The statistics for road casualties are available from the DfT website for free download. I thought it would be interesting to look at the casualty trend since cameras became common. I downloaded the data and had a look to see if anything leapt out at me regarding the change in trend Smith observes in the early 1990s. I was expecting to see a change related to the growth of mobile phone usage or something like that - what I actually found was a surprise to me, so let me walk you through the process and see if you agree with my tentative conclusions.

It is very important to understand that the reporting mechanisms changed in 1993, and figures either side are not properly comparable, although efforts have been made by the DfT to make them so. The precise text is: From 1993 the data has been estimated using the expansion factors and the new methodology for measuring road lengths, they are not directly comparable with the figures for 1992 and earlier.

According to some of Smith's fans, the fact that I have written this document showing where Smith's claims are wrong means that I am a murderer because I support measures which I "know" cost lives. That claim is so bizarre as to be hard to comprehend: I show that Smith's reasoning is faulty while somehow "knowing" that it is not, and that means I am actively colluding in causing deaths which I have shown do not have the cause they attribute. Very odd. No, I do not believe that speed cameras cost lives, I don't think there has ever been any credible evidence to support that idea and certainly no evidence that passed peer review. I don't know whether they save any meaningful number of lives, but I don't care very much because the speed limit has been in force since long before I passed my driving test so the question is one of enforcement of laws which I knew about and to which I must be presumed to have consented by virtue of deciding to apply for a license - and the same will apply to most British drivers.

The Data
First, I plotted the current trend in fatal and KSI (Killed or Seriously Injured) crashes. I have plotted many of the charts with a false zero for clarity. This is a statistician's ploy which is open to abuse, so remember to check the axes every time! The first obvious thing about Smith's analysis is that he uses some pretty arbitrary date ranges ("some funny dates" I believe is his phrase) and fails to control out a large blip in the fatality figures in the early 90s due to a recession (indeed uses it to create the projected trend if cameras had not been introduced), so I'm going to get as close to his date ranges as I can with the readily available data.

The years 1994-2002 are chosen for convenience: 1994 was the start date of the detailed series on the DfT website and 2002 was when I first did the analysis.

Notice that the fatality trend is much flatter than the KSI trend. I will return to this later. KSI is the measure usually used by road safety professionals because it is larger, so less prone to fluctuations due to exceptional events (with around 3,500 fatalities annually, a single coach crash would be visible as a significant blip). Looking at the above chart, note that both fatality and KSI figures are down year-on-year, every year, since 1994. Notice also that the more accurate KSI figure shows a much sharper drop.

Now look at 1b. Here I've plotted the rates, per billion passenger km, as well as the absolute numbers. Notice that the decline looks quite marked when assessed by rate. This is not a reason for complacency, but it does rather undermine the idea that things are getting much worse. Smith uses both sets of figures.

The devil is in the detail
One obvious way to prove Smith's hypothesis would be to see if the trends correlate in any way with the prevalence of cameras. There isn't much data about that, it's not easy to analyse out which crashes happened on roads with cameras and which on roads without, but there are two sorts of roads where cameras have become much more common - urban roads and rural A roads - and two sorts of road where cameras are still rare - motorways and country lanes. Outside of the M25 Variable Speed Limits the British motorway network has very few fixed cameras.

Notice that two sorts of roads trend above the others - motorways and rural minor roads. These are types of roads with few cameras. The two road types where fatalities have trended best are urban roads and rural A roads - both road types where the number of fixed camera sites has increased most. It would be unwise to infer a causal link at this stage, though. These are numerical values, not rates per unit exposure.

I did say that the fatality figures were vulnerable to freak events, and motorways are the safest roads per passenger mile (and so they should be, with no right turns, very few junctions, restricted access, no pedestrians and so on). So let's have a look at the larger, less accurate but more statistically amenable series for KSI:



From fig. 3 it may be seen that the UK has experienced a significant improvement in KSI numbers, except for motorways (and rural minor roads, not shown on this chart, that is shown elsewhere though). We also know that motorways are a class of roads where speed cameras are much less to be found, outside the M25 variable speed limits. In the first 18 months of operation of the M25 VSLs, the accident rate reduced by 28%. So the one stretch of motorway where there are significant numbers of cameras, 26,000 fines in this period, has experienced a very significant improvement in crash rate. At this point any reasonable person could be excused for thinking that cameras are quite obviously not to blame, and may possibly be highly beneficial. But experience shows that Smith doesn't give up that easily when he has a preconception to support!



Fig. 4 shows the growth of motorway fatality rates against all road fatality rates, and the growth of motorway traffic, against a 1994 base. This is not inconsistent with what we know about lengthening commuting distances, and also brings to mind Joksch's rule of thumb relating the probability of fatality in a crash to the mean speed of the road. Motorways are faster roads, more mileage is being done on motorways, so it is not a surprise to see that this has translated into increases in the proportion of fatalities, although more evidence is required to prove a case.

OK, this is getting a bit nebulous. So, here's a single, simple chart which shows risk over time going back to 1950.



Once again it is apparent from this chart that fatality and KSI trends, which have hitherto tracked each other quite closely, seem to be diverging. That can't be accounted for by cameras, unless a credible mechanism could be suggested whereby cameras could increase crash severity without affecting the probability of the crash happening, but it could be due to some population effect in a group where crash outcomes are more likely to include fatality.

By now a couple of things have become pretty clear, in my view:


 * Whatever we are looking for, it is something which increases the rate of fatality given a crash, but does not increase the probability of crashing, or at least not to anything like the same extent. Arguably we could have just superimposed a hundred additional crashes, all of which are fatal, on top of the existing data, but I really don't buy that.
 * Whatever we are looking for, it is something which has affects motorways and minor rural roads more than urban roads and rural major roads.

But preconceptions are the enemy of understanding, so let's plough on and see if we can mine any more gems from the data.

Killed vs. KSI
So what about the distinction between killed and killed or seriously injured? On the face of it, it seems that the chances of being killed in a crash are rising in relation to the chances of being involved in a serious crash. There is a third series of data, recording minor injuries, which is much less used. Almost all fatalities are recorded, the majority of serious injuries are recorded, but there is a historical and substantial under-reporting of minor injuries, running close to 100% for the least serious injuries. It is also difficult to quantify the point at which recording might start - a bruise sustained in a very low speed crash might be reported if an insurance claim is at stake, or it might not.

So why should one choose KSI over fatality figures? The answer is simply that the number is much larger, which makes statistical analysis more dependable. It also allows comparison of figures which are in such low numbers that individuals would show on the charts - the number of child cyclist fatalities is in the tens only, which makes it difficult to plot and unreliable to draw inferences. So KSI is the most widely used measure, offering the best balance between statistically significant sample size and accuracy of recording.

Why would you choose fatality figures, and use absolute values rather than trends? I don't know. It would be one way of working back from a conclusion to reach a set of supporting data, I suppose, but it's undoubtedly not the way any reputable transport statistician would approach the subject.



In case you weren't pretty much charted out by now, 5 is a chart showing all fatalities and KSI rates (per billion passenger km) broken down by road type. Some caution is required, the definitions of road types are not identical between the data sources, but the picture is clear enough. KSI figures are going down strongly, fatality figures less strongly, and the difference is found in two of the series: motorways and rural minor roads. Your starter for ten and no conferring, where are most of the cameras? University of Life, Cleverclogs: Yes, rural major roads and urban roads. So if speed cameras are genuinely degrading safety they are doing it indirectly, by driving people onto other roads. It is not terribly believable that people will choose to use a motorway rather than a rural road simply because of cameras, when the motorway is wider, straighter, has fewer conflicts, a higher speed limit - but when you have a preconception to support, any mechanism however incredible must be given full weight.

Any other hypotheses?
Measuring the effect of speed cameras or any other road safety measure requires great care. When compulsory seat belts were introduced the driver casualty rate dropped, but this was later found to be due to the simultaneous introduction of evidential breath testing. At the same time levels of cyclist, pedestrian and rear passenger casualties rose steeply. Drivers felt safer, and apparently consumed that safety improvement as a performance benefit.



Some of the user types have very small samples (bus and goods vehicle are all under 2% of the total), I'll come back to that in a moment, but note the bold lines: this are road user classes which have seen an apparent rise in KSI trends. The groups are motorcyclists and their passengers, and pedal cyclists. These figures span a slightly longer period, from 1992 to 2002 - there was a change in the counting method in 1993, but DfT have recalculated these figures (although this is probably not entirely accurate as the old data collection methods were used). See 7a and 7b.



On the face of it, pedestrians are safer than in 1992, cyclists and drivers are at about the same risk, and motorcyclists are at much higher risk. So the entire change in the trend, in as much as it exists, might be due to an effect in the motorcyclist population. Transport Statistics Great Britain (TSGB) records that motorcycle use has risen from 0.71% to 0.73% of all transport by passenger kilometre, so this is not explained by increasing motorcycle use (there was a sharp dip and then a recovery in the 1990s, see below). Note that this was before the London congestion charge caused a significant shift to two wheels in the Capital. So I'll use Smith's chosen data, fatality trends 1978-1993 and 1990 - 2002 (why these particular years? I don't know) - that's fig 8.



What I notice here is that the decline in motorcycle usage has flattened off, and the trend in the fatality figure has also flattened off. That's not unexpected given what we know about the dangers of motorcycling, some of which are self-imposed and some are not. In fact the fatality trend and the motorcycling trend seem to track quite closely. So have we found the cause? I think we need more information. Let's look at the earlier data, as fig. 9.



That blows that one out of the water! OK the road environment has changed rather a lot since the 60s, and motorcyclists do have a unique profile on the roads: they are vulnerable when crashed into, and they are disproportionately likely to be involved in single vehicle crashes. Motorcyclists are the most likely to die of all road users, by any measure. But there's not much to see here, no apparent correlation. So we could undoubtedly be misled if we allow our earlier conclusion to colour our interpretation of these results. Certainly if we wanted to work back from the conclusion that motorcycling is responsible then it would be straightforward to choose dates and series which support that conclusion. That would still not make it an accurate conclusion.

What else might it be?
While I was plotting that chart I had a thought: motorcyclists are not the only vulnerable road users. I've already shown that the problem is not (as Smith claims) a worsening in road safety, but a worsening chance of a fatality in the event of a crash. There is a wealth of anecdotal evidence from cyclists that driver behaviour towards them is getting worse. So I plotted cycling (and motorcycling) against the risk of fatality in a crash. The result is shown as fig. 10.



See how, from 1990 on, the rise in fatalities changes with the rise in miles travelled by vulnerable road users? But Maybe my eyes deceive me Do the lines really trend together or is there just some vague similarity of shape? Here's a refinement, the risk of fatality (fatality / KSI) against the proportion of all mileage done by cycle - I've plotted it as the year-on-year percentage change in each case, fig. 11.



There appears to me to be a significant correlation, probably enough to explain the disparity between trends in the two most accurate measures of risk, fatality per billion passenger km and KSI per billion passenger km. This is, of course, entirely inconsistent with the motorway hypothesis, as cyclists are banned from motorways, but that is the problem with fatality-only data - the numbers are often too small for reliable inferences to be drawn. Looking at the rural minor roads, which have experienced almost as large a change in fatality trend, as noted above, there were 636 fatalities on such roads compared with 175 on motorways. Cyclists undoubtedly do use rural minor roads - indeed we prefer them, and much of the National Cycle Network runs on such roads.

What about the drivers?
Smith uses proof by assertion to discount any possible cause of his observed loss in trend other than cameras and camera policy. I don't have figures yet for factors such as the growth of mobile phone use or 4x4 vehicles, and he doesn't either, but still he is entirely confident that these are not to blame.

A couple of press reports underline the problems with this position.

The Sunday Times of 10 July 2005 carried in its Driving section the following: "Drivers of 4x4s are more likely to have convictions for dangerous driving, speeding and driving without due care than other motorists. They are also more likely to be at fault in an accident, according to a survey by Admiral, the insurance company. Admiral looked at the driving records of 38,000 4x4 owners and compared them with those of other motorists on its database of more than 1m policy holders. [...] 4x4 drivers were found by the survey to be 27% more likely to be the ones at fault in an accident and 15% more likely to be convicted of driving without due care and attention. Convictions for speeding and dangerous driving were 11% more common among 4x4 owners". So there is a strong correlation between bad and dangerous driving, and ownership of vehicles with disproportionate potential to cause harm. It is not clear whether this is a reflection of bad drivers choosing vehicles in which they are less likely to suffer the consequences of their driving; whether the poor handling and feeling of protection conspire to make 4x4 drivers less careful; or some other factor or factors. Whatever: there are plausible reasons for supposing that the growth of large 4x4s might lead to more, and more serious, casualties.

What is more likely, though, is the mobile phone connection. There has never been much doubt that mobiles impair driver control, and a report from Australia has highlighted this graphically. Researchers looked at the phone bill records of 456 drivers needing hospital treatment after road crashes in Perth, Australia. For each driver, they assessed phone use immediately before a crash and on trips at the same time of day 24 hours, three days, and seven days before the crash for comparison. Mobile phone use in the 10 minutes before a crash was associated with a four-fold increased likelihood of crashing, irrespective of whether the driver was using a hand-held or hands-free phone. Similar results were found for the interval up to five minutes before a crash.

From the mid 1960s onwards the casualty trend shows an exponential decline. The rate of decline in fatalities does reduce around 1990, but this predates by over a decade the real growth in cameras. It coincides, however, with a period when mobile phone use was increasing sharply. The major growth in speed cameras followed the introduction of camera partnerships in 2000, and it was a couple of years before the numbers ramped up. There were under 800 fixed cameras in 2000, by 2003 this had risen to over 4,300. Phone use, by contrast, ramped up sharply in the early to mid 1990s and by 2000 most drivers would have had a phone. Hard numbers are hard to come by, but a period of change the early 1990s followed by levelling off early in the new century is descriptive of both mobile phone ownership (especially among higher-mileage drivers) and of the fatality trend; it does not describe at all the growth of cameras. Finally, the fatality figures for 2004 show a sharp drop from the previous year. This change, while within the limits of statistical variation for such small numbers, does follow a change in the law banning the use of hand-held mobile phones. By contrast the change in the rules for camera placement and warnings of enforcement, in 2002, are accompanied by no such change in trends.

Much of this evidence is weak, but it does highlight the undoubted fact that Smith is ignoring at least two potential causes which have a proven link with crash risk, in favour of a supposed cause which, in addition to lacking a plausible mechanism, is not, in fact, correlated with the road types or road user types experiencing the observed effect.

Conclusions
Have I proved anything? Possibly. Taking this data at face value I would draw the following inferences:


 * the UK's roads are now safer by all measures than they were before cameras became widespread
 * the biggest improvements in trend have been on those roads most likely to have cameras, and the least good trends are on those roads least likely to have cameras
 * motorcyclists appear to be at increasing risk
 * the change in relative trends between KSI and fatalities may be the result of more people cycling

There is clearly no causal link between speed cameras and risk, because there is no link at all. Whatever the reason for the loss in trend in the fatality series, it is visible only in the fatality stats for classes of roads where the cameras are not. Smith's proposed mechanisms whereby cameras could cause increased risk all describe increased risk on the roads where the cameras are. He has not advanced any reason why the cameras should be responsible for more crashes ending in fatality on those roads where cameras are least likely to be, while not affecting the rate on those roads where cameras are. His proposed mechanisms would all increase the risk of all crashes, not just fatal ones. So it is easy to discount this hypothesis.

There is some apparent correlation between cycling and the proportion of road crashes which end in death. This is hardly a surprise. Even though cycling becomes safer the more we do it, being hit at 30mph by a SMIDSY is much easier to survive in a car than on a bike. The same goes for motorcycles. Nor is it a huge surprise to see a significant increase of risk on rural minor roads, as Transport Statistics Great Britain (TSGB) shows a strong growth in distance travelled on these roads, not least because improved trunk roads allow people who work in towns, to live in outlying villages. Not that this causes congestion or anything, you understand, that's all the Government's fault.