From Petrolhead Porkies by Cowboy Cops to Elephantine Eighteens via Bayesian Bananas

Independent 15jun04 News

Death rate 'down at speed camera sites' The Government hit back at … critics today by publishing figures ... Produced by University College London … showed deaths and serious injuries had fallen by an average of 40 per cent…. equivalent of 100 fewer deaths a year - a figure hailed by Transport Secretary Alistair Darling, and safety and environment groups. But …Safe Speed …said the figures were "deeply flawed" …Association of British Drivers said …far lower than that of some European countries. …42 police force areas …are now in … money raised goes into camera activity and the Treasury receives any surplus...  870 fewer KSIs per year…33 per cent fall in injury accidents - 4,030 fewer per year …a 35 per cent reduction in … pedestrians killed or seriously injured. Average speeds … fell by 7 per cent 2.4mph …at urban sites … around 8 per cent  …vehicles speeding at new camera sites dropped by 71 per cent …79 per cent of people surveyed supported … benefit … through casualties saved … £221 million a year Mr Darling …recognised …controversy … people felt resentful …figures prove that cameras save lives. …"Up to 10 people are killed on our roads each day. We owe it to them and their families to do everything we can to improve road safety even further. "Most …sites …achieved good results. …where results were not as good …see what more could be done ..." …published the location of every site …show why …installed and …effect  …ensure …are needed ..."  Conservatives said …suspicion …to raise money… Liberal Democrats said …would silence …advocated the removal …Association of Chief Police Officers said …cameras did …save lives AA Motoring Trust said …cameras were about safety and not revenue. Paul Smith… Safe Speed Safety Campaign, said …were flawed.

Independent 16jun04 News

Drivers dismiss speed camera safety claims A furious dispute … Association of British Drivers (ABD) said … "meaningless" …no substantial reduction in accidents …government adviser warned sceptics …"blood on their hands" … University College London, published by the Government, showed people killed or seriously injured …fell by 40 per cent over three years to mid-2003 …benefit to society of £221m a year. ...reduction in …collisions involving injury varied from 11 per cent …to 72 per cent … Mark McArthur-Christie ABD, … no …decrease in accidents since …installed…."cluster" …at one site in one year… at another site the next year.  David Begg, chairman Commission Integrated Transport, a government advisory body, …:"Today's report proves that speed kills, that cameras reduce speeds at accident spots and they prevent injuries and save lives." Critics of speed cameras would "end up with blood on their hands". Department for Transport said casualties had risen at 743 out of the 5,000 sites… spokesman said …not been in place long enough to judge….

Independent 21jun Letters

Speed cameras should be controlled by independent traffic police:  Sir: …official report concluded speed cameras save lives …virtually all in the right place should come as no great surprise. The Government likes cameras. I am surprised that … effort and fanfare … only saving 100 lives a year. …I view it as a poor return on the investment. …not assessed how many accidents speed cameras cause….having to concentrate more on speed and less on driving is dangerous and causes fatigue, a major cause of accidents. …greater emphasis on cameras and reduction of traffic police …more drink drivers, more dangerous vehicles and other serious offences going undetected. How many lives has this cost? Without considering these issues the report is meaningless. …a powerful lobby determined to increase the number of cameras no matter what. A balanced and sensible debate seems unlikely. …an independent national traffic police force …control of the cameras … unaccountable "safety partnerships" would aid public confidence. This would allow ….money from cameras to fund additional officers and education…. some confidence …used sensibly …not just as cash cows. …halt the real damage which is being done between the police and the public.




REPORT The sites that were accepted for inclusion in the cost recovery programme conformed to the handbook requirements that are specified in Appendix A of the present research report. The requirement for a record of PICs during recent years is a central criterion for selection. However, there was also a requirement to identify speed as a contributory factor to these collisions, a clear indication of motorists speeding, suitability of the site for treatment by enforcement, and unsuitability of further engineering remedial measures. Because of this requirement to identify the sites as suitable for this particular safety measure, and in particular that the collision record was not the sole criterion for selection, the established statistical phenomenon of regression to the mean (also known as bias by selection) will not apply in full measure. Furthermore, the results of the statistical analysis of casualties and collisions at speed cameras are consistent with the observed reductions in speeding, showing that the enforcement measures are working as intended.| We cannot, in all cases, make comparisons with the results from the previous report as there have been refinements in the modelling techniques used and an increase in data. Where there have been substantial changes to the results found last time, these are highlighted and, where possible, explained. | modelling gobbledegook p93-5

MELDREW In round terms say 3kpa deaths, 30kpa injuries, 300kpa incidents, 3000kpa speeders (projected), 30000kpa drivers corresponding to equivalent intervals of 10y/speeder, 100y/incident, 1000y/injury, 10000y/death. Current fines £60Mpa from 1M speeders of which 60%+ is spent on supply-support. As simple suggestion suppose 3600 deaths site-risk partitioned as 1800 high, 1200 medium, 600 low and suppose high and medium are cameraed but not low ~ perhaps plausibly associate each category with 1000 sites so high equates to 2- deaths, medium to 1+ and low to 1- (yearly). These segmented flat distributions could then be seen as segmented mean values of a linear profile over all events ranging down from 2100 to 300, with high band 2100-1500, medium band 1500-900, low band 900-300 and divide by 1000 for incidence per site.

Doing this we could then recover the 100 camera attributed reduction plausibly by shifting the high-medium scale from 2100-900 to 2050-850 such that the supposed uniform 50 reduction divided by 1000 for incidence times 2000 sites returns the observed 100 reduction. However a subsidiary observation was increased incidence at 1/6th or so of cameraed sites. Some maybe all of this increase reflects annual variability but if we want to assign it all within the simple representation then one of two simple limit attributions would be to put it all in medium range and suppose for example that the high mean had shifted down by 200 from 1800 to 1600 and the medium mean had shifted up by 100 from 1200 to 1300. Segmented linear distributions would then run from 1900-1300 for high set and 1600-1000 for the medium set such that a single linear fit running down from 1900 to 1000 would have mean value 1450 against 1500 before and so this 50 reduction divided by 1000 times 2000 sites again recovers 100 fewer deaths. This linear profile (1900-1000) would then cross the original one (2100-900) at 1300 on the death rate corresponding to 2/3 sites having a reduced rate and 1/3 having an increased rate.

You can see how this might be easily tuned to return the 1/6th or so recorded but it’s just not worthwhile given uncertain variability. However commonsense suggests that if medium risk sites returned a higher than expected rate then this behaviour quite probably wasn’t special but spilled over into the low risk range originally suggested as 900-300 (mean 600). Indeed, not just commonsense but also nigh on steady state overall records ~ meaning that if the net 100 reduction obtained over all cameraed sites was attributed entirely to variability (not causative consequence) then an increase of 100 deaths would have occurred at uncameraed ones taking their mean up to 700 which here would be simplistically assigned a linear profile 1000-400. If the case for installing more cameras could only be made on grounds of death rate exceeding a threshold then 1pa would threshold opening this door to the top end of the uncameraed ones.

By the same token even with all of the 1/6th “mysteriously increased” sites dumped into the medium bin some at least likely deserve decamering being borderline to the same threshold ~ “mysterious” that is to Ali Darling in his Toady spot though not to grumpy old gits like me traditionally trained in turbulence! In fairness, Ali did indicate that some cameras would be pulled although also stated that total numbers would be increased in view of their successful demonstration ~ and that’s cheating because you really can’t claim causation on one side of the fence but merely mystery on the other. 

As another indicator of the fallaciously claimed causation consider that deaths broadly carry companion ratios of injury, perhaps 6:1 for major each with 6:1 for minor, in all then 36 injuries for each death, these figures as overall numbers for all events from fatal accidents to minor bumps. This ratio was reproduced in the summary claims (deaths down by 100, injuries by 4000) when it really shouldn’t have been because the camera assignment was biased toward the most dangerous sites as dictated by approval regulations. It’s not clear which way injury numbers should have gone because some death reduction must have been converted into injury amplification whilst some presumably had no associated conversion.

But observed ratio conservation accompanied by claimed camera causation seems an implausible pairing without any assay that details the trade-off between push and pull (none was delivered). Indeed, the two going together arguably points to prima facie indication that the data were dominated by variability alone!  And that’s the rub here as indeed with most mantras of the “speed kills” variety ~ statistical massaging can always provide what the politicos want whether or not it connects with commonsense. Apologetically adapting Einstein and Rutherford adages into the twinned tag “God didn’t dice Man merely to play pinhead fairies with elephantine eights on Bayesian Bananas” is about the neatest way of summing up this dicing of data into an elephantine eighteen whose complexity means it can’t be constructively deconstructed even for plausibility never mind sensitivity or reproducibility.

Nonetheless I had a crack at accounting for the 1/6th sites supposedly where cameras had caused increased incidence of casualties ~ yes, caused because you can’t claim causative connection on your side of the fence without conceding it on the other. This will appear very soon as postscript suggesting what was claimed to be a conundrum may well be instead an embarrassing explanation of statistical trickery played to provide promotional purpose for what in reality was always a political policy, one that has always been motivated and maintained by no more than a mantra pandering to familial grief of unfortunate victims not so much of speed killers but of senseless stupidities ~ some as extreme excesses but mostly by far almost certainly inane inattention or insensitivity to impact consequences comparable .with those conveyed in carrying of a lethal weaponry.


            My meldrewism portrays Peter & Jane’s rendering of the same story as one that allows Humpty-Dumpty to be both cracked open and patched up again. Inspired as the simplified picture posed above it provides a parametric portrait that is probably about as much as Einstein and Rutherford would have tolerated given all the variability and uncertainty in capturing and collating such data. It is of course self-evident that realities of different driving behaviours broadly reflect sensitivities on a spectrum from disdain to discomfort, this scaling in turn broadly aligned with actuarial actualities of age dependence for accident incidence. The most meaningful generic metric of course is frequency spectrum of g-force but that’s rather fancy not to say fanciful for general guidelines digestible by other than sensibly schooled scientists. A crude cartoon can be conveyed in terms of g-bands ranging downwards from excitingly accessible to grudgingly tolerable, dominantly discriminated within a triad of decadal age aggregates respectively from young follies (20-40) adrenalising on 4/3g twitching tricks through middling fifties (40-60) fearful of 2/3g abrupt encounters to old farties (60-80) discomforted by 1/3g neck benders.

          These things are the decisive discriminators that not only set speeds seen as standard away from camera cavorting but also provided practical policies for policing in earlier eras before focus follies eliminated this established equilibrium. As seen below the numbers are pretty much on the button as indicators for speeds seen as endemic excesses, upwards from 20% for the farties through 30% for fifties to 40% for the follies ~ respectively 36, 39, 42 in  30- zones and pro rata for 60- and 70- zones, all in line with established experience for drivers and policers as realities that have stupidly been sidelined by partnerships poseurs pontificating from their privileged hi-profiled platforms.

          Anyway I did this simple stuff within a week of that ridiculous report and it would have appeared at the time had I not been distracted by disrupted email courtesy of antispam silliness adding to already seriously depleted accessibility courtesy of a virus mailer impersonating my URL (woes to be related in a future meldrewism). Amusingly ironic then that during the delay Ali D has judged multi-million alienation of respectable law abiders would be too much to Blair in the upcoming general election and acted to reduce the penalties that were threatening the livelihoods of not just many movers but many makers who wouldn’t sit idly by, indeed were already banging their angry drums well within hearing range of Westminster’s Woolly Wilderness.

          By extraordinary coincidence Ali’s new brainwave is bulls-eyed on my own portrayal (bang-on at 30% = 39!)~ well, well actuarial actualities finally found their place in policies that previously had been pirated by partnership pontificateurs whose leading lights number that obnoxiously opinionated ACPO pop-idle with the funny Scandinavian-sounding name ~ the man whose main claim to fame has been having not just the most cameras  per mile with the most stringent settings but also the most ruthless readiness to prosecute the hapless Henries who happened to ruffles his sensitivities on his patch, a biggish regional outback and no doubt a happy hunting ground during the summer season. You don’t have to take my word ~ replay his Toady spot 26Jun04 (BBC R4 website ~ google it!) and judge for yourself whether you’re listening to an obsessive carer or merely a maker moving up his poky political pole!

          This is where I sign off for now having finally finished it 3 months later than planned ~ prettification is a gargantuan gorger of time and time used to be so precious a commodity that prettification was seldom seen outside sales and advertising arenas. Nowadays of course with so much of slender substance, prettification has displaced pontification as primary point of promotion, mostly of course pointless pap with such shortened shelf-life as to provide a presumed primary purpose to the repetitive renderings of anything and everything however flawed or fatuous.



      PA-C’s pro camera interest predisposed prejudice (at best covert with micro-c) in their pejorative portrait, a betrayal which triggered my googling for evidence of conjectured conflict to retrieve a trio of top tenners as prima facie pointers, suspicions endorsed also by absolute absence of anything via their own site searched for “speed cameras” when I expected to see site screamers for successful selection in a panful of pro projects procured from present partial politicos. DoT seriously contaminated credibility in their Sec State trumpeting of the report as a product of UCL impartial excellence, especially as no attention (nor mention even) was drawn to PA participation ~ extraordinary considering the glossy navel-glazer itself was topped and tailed with PA adverts for their incomparable arena acumen. Integrity wasn’t merely evaporated by this bunch of bungling bandits as ambushed and extinguished in alliance with ambitious adventurers and avaricious academics. Poor old GB ~ yet another splash down the Suwannee and not even a paddle in sight to possibly prevent impaling on the debris of our dilapidated Dome and all the other doomed dogmas that Downing Street has spewed since they stole the showboat from a comparably crackpot crew.


  National safety camera programme             Three-year evaluation report June 2004


93 G.5 Analysis In order to estimate the part of the variations in the observed occurrence of PIC and KSI at camera sites that can be associated with introduction of safety cameras, we undertook a statistical modelling exercise. The model that was developed is log-linear in form, to estimate the mean frequency (number per unit time) of a Poisson process. The modelling was undertaken using the GenStat statistical analysis package (GenStat Committee, 2002). 94 Because the observations are reported in units of various durations, the durations were accommodated by using the GenStat offset facility. We supposed that the data have a Poisson distribution with mean frequency that is modelled as follows. The same model form was developed for each of KSIs and for PICs, but with different parameter values fitted for each. The description below is that for KSIs. Where Ntp is the recorded number of KSI casualties for observation dated t at site p, Otp is the logarithm of the duration of the observation period t at site p, Pp is a parameter to allow for the differing number of KSI casualties between sites p due to their sizes, populations and other fixed attributes. (Note that in this model, the whole of non partnership areas are treated as single sites). T is a parameter that represent the general change in frequency of KSI casualties over time t, which is measured from the start of the study period. Qq,u is a parameter to represent the seasonal variation in KSI casualties during each year with a value that varies between quarters q at sites in location of kind u (u = 1 for urban, u = 0 for rural), q(t) is the quarter year into which the observation falls: in cases where the observation spans several quarters, the quarterly effects were averaged, Af,u is a 2 ´ 2 parameter to represent the effect associated with a camera of kind f (f = 1 for fixed, f = 0 for mobile) in location of kind u (u = 1 for urban, u = 0 for rural), a(p,t) is the proportion of the period of observation t at site p for which the camera was established, B is a parameter to represent the effect associated with operation under cost recovery, b(p,t) is an indicator of whether site p operated under cost recovery during observation t (when b(p, t) = 1) or not (when b(p, t) = 0), C is a parameter to represent the effect associated with increased conspicuity, c(p,t) is an indicator of whether site p was recorded as or required to be conspicuous during observation t (when c(p, t) = 1) or not (when c(p, t) = 0), f(p) is an indicator of whether the camera as site p is fixed (when f(p) = 1) or not (when f(p) = 0), u(p) is an indicator of whether the location of site p is urban with speed limit < 40 mph (when u(p) = 1) or not with speed limit > 40 mph (when u(p) = 0) ept is an error term that is assumed to have Poisson distribution. ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )tp p u p f p u t q p tp tp p f t p c C t p b B t p a A Q Tt P O N ε + + + + + + + = , , , exp , , Use of a separate parameter Pp for each of the sites p means that comparisons are made for each site individually according to its collision record. Use of the temporal parameter T allows for long-term trend in the mean frequency of KSI casualties, and use of the parameters Qq allow for seasonal variation in the mean collision frequency through the year. The effect of the interventions at a camera site is represented through the parameters A, B and C . In the case of fixed cameras, all three components of the effect are taken to apply, whilst at mobile sites, the additional conspicuity requirements are believed not to have has any substantial influence, so that the conspicuity component represented by parameter C was not modelled at them and is not applied in estimating their effect: this is included in the structure of the model by multiplying the conspicuity parameter C by c(t, p) to represent the conspicuity requirement and by f(p) to represent its application only at fixed sites. Similarly, where a camera was established before the start of the baseline period, the camera component represented by parameter A was not modelled at them; this was controlled by the presence indicator a(p, t) , which takes the value 1 throughout at such sites. Thus the proportional effect on the mean frequency of occurrence of KSI casualties of establishing a conspicuous fixed camera operating under cost recovery at a site p is estimated as Similarly, the proportional effect on the mean frequency of occurrence of KSI casualties of establishing a mobile camera operating under cost recovery at a site p is estimated as Because of the nature of the data from which this model was estimated, estimates of the parameters A, B and C are correlated. For this reason, the standard error s of the sum of parameters (given generically as r and s ) was calculated using the formula where denotes the standard error of estimation of parameter r, and denotes the correlation between estimates of parameters r and sp . The GenStat software provided values for the parameter estimates, their standard errors of estimation, and the correlation between estimates: these values were used in the analysis of results presented here. ( ) ( )C B A p u + + , 1 exp( ) ( )B A p u + , 0 exp 2 2 2 s s r rs r σ σ σ ρ σ σ + + = r σ σ rs σ ρ


PA Consulting Group is a leading management, systems and technology consulting firm. Operating worldwide in more than 35 countries, PA draws on the knowledge and experience of 3,000 people, whose skills span the initial generation of ideas, insights and solutions all the way through to detailed implementation. PA focuses on creating benefits for clients rather than merely proposing them. PA’s results-focused approach is founded on a unique commitment to excellence, value and independence. PA has a considerable track record in supporting the police and central government. This often involves multi-agency working at a national and at a local level – taking ideas right through from policy to implementation. We work on a number of high profile assignments in the transport sector that are making real contributions to reducing congestion, tackling criminality and improving road safety. Acknowledgements This report has been produced by PA Consulting Group and University College London. … The Centre for Transport Studies (CTS) at UCL (University College London) is one of the leading inter-disciplinary transport research and teaching centres in Europe. It consists of six permanent academic staff with backgrounds in Mathematics, Geography, Economics, and Engineering, and over 20 research staff and doctoral students. The Centre is based within the Department of Civil and Environmental Engineering, which has achieved the distinction of receiving the top rating (5) in the Government’s most recent research assessment exercise. The CTS undertakes research across a broad range of transport-related topics, including transport safety, transport planning, traffic management and control, infrastructure design and accessibility. Although this report was commissioned by the Department, the findings and conclusions are those of the authors and do not necessarily represent the views of the Department for Transport. …PA Consulting Group is a leading management, systems and technology consulting firm, operating worldwide in more than 35 countries…. Principal national offices Argentina, Australia, Czech Republic, Denmark, Finland, France, Germany, Indonesia, Ireland, Japan, Malaysia, Netherlands, New Zealand, Norway, People’s Republic of China (offices in Beijing and Hong Kong), Russian Federation, Sweden, United Kingdom, United States