Wednesday, September 29, 2010

Anti Speed Camera Arguments #5: Questionable Statistical Methods

Yet another in my series of posts debunking the arguments of 'Safe Speed', the anti-speed camera website.

Safe Speed claim here that the case for speed cameras is founded on data that is not statistically sound. In particular, they point to the fact that the 'regression to mean' effect has nother been taken into account.

Very briefly, 'regression to mean' is the effect where if a variable such as the number of accidents at a point on the highway is extreme on first measurement, it will tend towards the average on a second measurement. There is a clear danger that, given speed cameras are sited at accident blackspots, any reduction in accidents in subsequent years may have happened anyway due to normal variations.

Safe Speed note that in the DfT's Three-Year Evaluation Report on speed cameras, selection effects such regression-to-mean effects had not been accounted for. Which is true, and the DfT admit to this (see here) although they also point out here that "An analysis was carried out on a subset of camera sites to estimate the size
of any regression-to-mean effects. Whilst regression-to-mean does appear to account for some of the reduction in collisions at cameras, the safety effects of cameras remain substantial.".

What Safe Speed are not in a hurry to point out is that the possible presence of distortion due to regression-to-mean effects does not show that speed cameras don't work. It merely means that there is some doubt that the  improvement in crash rates can be wholly or partly attributed to the presence of speed cameras. In other words, it's "case not proven" rather than "guilty". Safe Speed are also not keen to point out that cameras are not sited simply because there is a history of crashes at a site. They are sited where the authorities consider that speed has contributed to the cause and/or severity of the crash, and that slowing down traffic would have some benefit. If you do not pick your data points solely on their value, and instead establish an underlying cause for the data points having a high value, regression-to-mean becomes less significant as an explanation for the improvements.

There is one other thing that Safe Speed definitely don't want you to know, because it seriously holes their case below the waterline. There was a further study, over 4 years, which you can buy here if you have a spare £15, or here for free. According to the blurb [my emphasis]:

"The principal methodology adopted for this research was a "before" and "after" study comparing collision numbers at safety camera sites with those at suitably matched control sites. The study design enabled the effects of cameras to be isolated from regression-to-mean and trend effects."

Oh dear. That could mean that speed cameras do actually work. And indeed, the report states: “the presence of safety cameras has reduced the number of collisions involving fatal or serious injury by a statistically significant 21%  (and)…the total number of collisions involving fatal, serious or slight injury by a statistically significant 12%”. (source)

There comes a point where, if you are a scientist rather than an ideologue, you have to let go of a theory when the evidence against it becomes overwhelming. You could be forgiven for thinking that Safe Speed are not an organization that are motivated by road safety and interested in "what works", but instead have a libertarian agenda, and believe that people should be allowed to drive at higher speeds regardless of the safety consequences. It's a shame Britain's roads can't be allowed to be a playground for petrolheads, but it's an truth inconvenient to Safe Speed that a lot of lives are at risk from speeding.

No comments:

Post a Comment