1. Modelling of Accident Data from Fixed Safety Camera Sites in London
Mike Maher
University College London

2. Background to the talk
Around 500 fixed safety cameras installed since 1990

3. Analysis of London data
Data available for camera sites operated by TfL
annual data on FSCs and PICs at cameras 1990 – 2011
.. plus totals for all non-camera sites
TfL press release, Sept 2014:
“Cameras have proved
successful in reducing road casualties in recent years. 58% reduction in KSIs at camera sites”

4. Analysis of London data
Data available for camera sites operated by TfL
annual data on FSCs and PICs at cameras 1990 – 2011
.. plus totals for all non-camera sites
TfL press release, Sept 2014:
“Cameras have proved
successful in reducing road casualties in recent years. 58% reduction in KSIs at camera sites”
But – similar reduction at non-camera sites too
camera
sites
non-camera
sites

5. Method of analysis Clear that trend must be allowed for
comparison with annual totals at non-camera sites
to gauge what would have happened
But regression to the mean (RTM) also
sites selected on basis of previous accident record
typically number of accidents in three year period
the site selection period (SSP)
generally selected sites show a reduction in after period, even if treatment ineffective
DfT national safety camera study (2005)

6. 4-year evaluation report: FSCs
before
after
FSCs/site/year:
(-54%)
Overall reduction = =
trend RTM camera
54% = 10% % %
50% allowing for trend
19% allowing for trend + RTM
relative to what would have been:

7. Four time periods method
Report by Allsop (2013) for RAC Foundation
For each camera, years divided into periods
before
site selection period (SSP): typically 3 years
transition (year camera installed): specified
post-installation (camera period)
So - if camera installed in mid-2000
SSP assumed to be 97-99
before period is then 90-96
camera period is 01-10

8. Accs/yr (adjusted for trend)
SSP
RTM
ASBiC
Camera effect
Pre-SSP
Post-installation
Installation of camera
time

9. Timing of the SSP? No record in data of actual SSP for any site
Not safe to assume the three preceding full years
Different SSP for each site?
if SSP is earlier, some RTM in the before period
hence inflates camera benefit
important to get timing of assumed SSP right
Or play safe and find a range to cover all SSPs
Model to estimate factors Fd : d = t – s(i)

10. Way to identify range of SSP
d = t – installation year
d = 0 is installation year
d > 0 is camera period
d < d* is before period
d = d*, … -1 is range of SSP
Clear signs of inflated level for more than 3 years
Best choice of d* = -6. SSP*: d = (-6, -5 … -1)

11. Form of data used
For each camera (eg KSIs at camera 240: installed in 2000)
Before
SSP*
After
No. years 4 6 11
Site total 5 8
Regional total
27934
36197
37784
4/37784
Estimate of camera effect =
= 0.591
5/27934

12. Form of model Binomial model to split accidents between periods
Probability factored in camera period by 
Otherwise given by regional totals (trend)
Then maximise log likelihood wrt 

13. Results Max Likelihood Estimate β = 1.014 (se = 0.030)
95% CI: (0.958, 1.074)
So, no significant camera
effect.

14. Conclusions and further work?
No apparent camera effect
modelling confirms initial graphical evidence
important to allow for trend and RTM
and to avoid making spurious claims
not just TfL, but Scotland, Wales …
But other modelling possibilities to investigate
are the actual SSPs known for each site?
6-year SSP* used to encompass possible SSP
so some data being lost
was there a change in selection procedure at some point?
monthly data available?

15. Thank you!
Any questions?