Joey Talbot 20/08/2021
The Saferactive project is funded by the Department for Transport. It aims to investigate spatial and temporal patterns in road safety for active travel.
High quality data is available on road safety through the Stats19 dataset. Published by the Department for Transport, this is a comprehensive dataset of road collision statistics, mainly as reported by police officers on the scene.
However, to estimate collision risk, we must also know the exposure, defined as the number of people walking and cycling on the road network. The exposure acts as the denominator in any estimate of road safety risk. The degree of uncertainty associated with estimates of exposure is high, and this uncertainty will be a key focus of the report.
The majority of the report focuses on cycling rather than walking. Firstly, this is because better data is available for cycle journeys, in particular journeys to work. Travel to work is recorded in the 2011 Census, and according to the NTS survey in 2011 it comprised around 38% of all cycle journeys but only 7% of journeys on foot. Therefore we decided it is acceptable to use travel to work as a proxy for cycling but not for walking. Networks of cycle counters from the DfT, TfL and others are also available, providing data on annual changes in cycle flows, while comparable data does not exist for walking.
A second reason for focusing on cycling is that since 2020 there have been large changes in how we use our streets, with unprecedented increases in cycling during the lockdown periods, accompanied in some cases by a rise in cycle casualties. It is important to take these changes in context and to recognise that an increase in casualties does not necessarily mean an increase in risk. This is why it is important to focus on exposure and the denominators of road safety.
We have focused mainly on the decade from 2010 to 2020, but there is some variability in the years covered by different datasets…
Where possible, we include data from England, Wales and Scotland. The sub-regional analyses are only available for England and Wales since they are based on the Propensity to Cycle Tool. NTS data are available for England only.
We have used a range of nationally available datasets to estimate and visualise road safety, including Stats19 collision data, DfT and TfL traffic counts, NTS survey returns and 2011 Census commute data.
TfL counter data (Central, Inner and Outer London count points only)
Mean count per Borough, over the period 2015 - 2020.
Change since 2015 in each Borough.
DfT counter data
Mean count per LAD, over the period 2010 - 2020.
{r} # knitr::include_graphics("") #At national and regional scales we have more reliable estimates of cycle volume than at sub-regional geographic scales, but there still remains differences between data sources.
The number of cyclists saw a small increase over the ten years prior to 2019, according to three different data sources.
In 2020, the mean number of cyclists passing DfT traffic count points was three times higher than in previous years.
Estimated change since 2011 in cycle flows at a national level, based on a GAM generated using the DfT and TfL counter data
Change in cycle KSI 2010 - 2019
{r} # knitr::include_graphics("") #In the period 2010-2019, risk of cycle KSI has reduced. This reduction has mainly occurred since 2014, and can be seen when casualty data are assessed using all three measures of cycle volume.
Sum of cycle KSI in each region (including Scotland and Wales)
Pedestrian KSI by region.
Total distance cycled per year in each English region (based on NTS survey)
Total distance cycled per year in each region (including Scotland and Wales), based on DfT regional traffic data.
Mean AADF cycle flow per DfT counter in each region (including Scotland and Wales). This is based on the raw flows per counter.
Mean change in AADF cycle flow per DfT counter in each region (including Scotland and Wales). This is based on the change in flow at each counter, calculated as the raw flow minus the minimum flow for the given counter.
Estimated mean change in cycle flows, according to GAM model based on the DfT and TfL cycle counters
Trends in KSI risk per km cycled, using NTS journey data
Trends in KSI risk per km cycled, using DfT regional traffic data
Trends in KSI risk per DfT mean cycle count
We investigate trends in KSI, exposure and risk for commuter cycling at the geographical scales of Upper Tier Local Authorities and Police Force areas. Data related to KSI and km cycled are available for the years 2010 - 2020, while data related to population estimates, such as km cycled per capita, are available for the years 2010 - 2019.
KSI and risk data are estimated based on the number of cyclists killed and seriously injured during weekday peak hours (07:00-10:00 and 16:00-19:00), and estimates of km cycled derived from Census 2011 travel to work by bicycle. These journeys are routed using the CycleStreets.net fast routes algorithm, as used in the Propensity to Cycle Tool.
For all other years, we adjust the 2011 commuter cycle flows using DfT counter data, based on mean AADF counts within the relevant geographical zone. As a baseline for the annual adjustments of cycle flows, we reduce volatility by using the mean AADF flows in a three year period centered on 2011, e.g:
K**m2020 = K**m2011 * (AAD**F2020/(Mea**n(AAD**F2010,AAD**F2011,AAD**F2012)))
where K**m2020 = estimated km cycled for travel to work in 2020; K**m2011 = km cycled for travel to work in 2011; AAD**F2020 = mean AADF of DfT counters within the geographical region in 2020; AAD**F2010 = mean AADF in 2010; AAD**F2011 = mean AADF in 2011; and AAD**F2012 = mean AADF in 2012.
Change in KSI at LA level is seen below.
This can be compared with changes in km cycled at LA level.
Combining the data from the previous two figures, we obtain estimates of change in risk.
Similarly, we investigate changes in peak hour KSI at Police Force level.
Km cycled has risen noticeably in many regions.
The result is a decrease in risk for cycle commuting across most Police Force areas.
The Upper Tier Local Authority level trends in peak hour cycling and risk can be found at https://github.com/saferactive/saferactive/releases/download/0.1.4/ksi.csv
The police force level trends in peak hour cycling and risk can be found at https://github.com/saferactive/saferactive/releases/download/0.1.4/ksi-pf.csv
Traffic levels on minor roads have increased dramatically in recent years…
Although the phrase ‘Low Traffic Neighbourhood’ is a recent invention, the concept is not new. Modal filters preventing vehicle movement along a street while allowing passage to pedestrians and cyclists have existed for many decades. They can take a wide range of forms, such as bollards, gaps, and continuations of pavement across the entrances of side streets.
It can also be argued that any street, or any network of streets, which does not provide a vehicle connection between two main roads, is in effect part of a low traffic neighbourhood. These streets do not provide worthwhile opportunities for ratrunning, so they are likely to be used mainly by residents and those wishing to access destinations on the street itself. This includes cul-de-sacs and wider networks of streets for which all of the access points are on the same main road.
Using this definition of low traffic neighbourhoods, CycleStreets.net have created a nationwide beta map of LTNs. All roads are assigned as main roads, low traffic neighbourhood roads, ratruns, or ratruns with traffic calming measures. Main roads are those classified as ‘A’, ‘B’ or ‘C’ roads. Data on traffic calming measures are obtained from OSM.
This map extract from Leeds gives an example of how roads are classified.
We have used this beta map with West Yorkshire as a case study region, to assess the impact of Low Traffic Neighbourhoods on road safety.
For this case study we have divided the West Yorkshire road network into a set of 1km grid cells. We see a clear correlation between total road length within each 1km cell and the number of pedestrian KSI over the period 2010-2019.
There is also a positive association between road length and cycle KSI but it is weaker, likely reflecting the fact that the cycling distribution is more uneven than the pedestrian distribution.
We can see a positive correlation between distance cycled for travel to work and cycle KSI.
Switching to maps, we can see how these patterns play out across West Yorkshire. The greatest concentration of cycle and walking KSI are found in the urban areas such as Leeds and Bradford. Cycle commute km are strongly concentrated within Leeds.
Now looking at the distribution of KSI risk, we see that for walking, KSI risk per km road remains highest in the urban areas, especially Bradford. For cycling, KSI risk per km road has a slightly more scattered distribution with high values in Leeds. However, when we use Bkm cycled as the denominator, the map is very different. Now some of the lowest rates are found in Leeds. This map shows very high KSI rates in some of the Pennine fringes of West Yorkshire, but these are probably artefacts of the fact that we are using commuter cycling as the denominator, while leisure and sport cycling dominate in these areas. A more reliable finding is that KSI risk appears higher in Bradford than in Leeds.
The next step is to investigate how these patterns relate to the distribution of road types. Unsurprisingly, the total road length within each 1km grid cell varies greatly across the region, being highest in towns and cities, but so too does the distribution of types of road within each grid cell.
Weighting by the road length within a grid cell, we look at how pedestrian and cyclist KSI vary with the proportion of roads of different types. We expect that KSI will increase as the proportion of ratruns increases, and decrease as the proportion of LTNs increases.
For pedestrians, a GLM shows a small increase in KSI per km road with the proportion of roads classified as ratruns.
For cyclists, GLMs show small decreases in KSI per km road and KSI per km cycled, with the proportion of roads classified as LTNs.
Rather than one simple answer to the question of how to measure road safety risk, it is clear that there are many different ways of approaching this topic. We have focused mainly on risk related to cycle commuting, which has some of the best data quality, but even here different data sources will provide different interpretations.
This uncertainty must be understood in the context that to assess risk, we need measures of both casualty rates and exposure. The casualty data is comprehensive and the core stats19 components (such as number of casualties) are highly reliable. Therefore it is uncertainty in the exposure, i.e. the number of km cycled on the roads, that is the most important factor.
Casualty severity adjustment…
When it comes to travel to work, we have comprehensive knowledge of spatial trends in cycling through the 2011 Census. Certain uncertainties remain even here, since we do not know the actual routes used to travel between origin and destination, but the CycleStreets.net fast algorithm provides a good approximation of these.
However, this only holds true for travel to work, which in 2011 accounted for 38% of cycle journeys. Similarly comprehensive data exists for travel to school, but not for other purposes such as leisure, personal business and shopping. In particular, leisure cycling accounted for 34% of journeys in 2011 and is likely to have considerably different spatial patterns to commuter cycling. In addition, we do not yet have evidence relating to multimodal journeys such as combined cycle-rail journeys.
For the sub-regional analyses in this report, one of the most prominent impacts of the uncertainty in spatial trends relates to this lack of data on other journey purposes. We have focused on travel to work by only counting collisions that occur during peak commute hours, but nevertheless, even during those hours not all journeys will be for commuting. Therefore, in the sub-regional analyses we will have overestimated KSI risk everywhere, but especially in places where commuting is a relatively smaller proportion of total cycle journeys. This could potentially explain the high risk rates seen in areas such as rural Wales, the fringes of London, the South Coast and the Pennines, which may have relatively high proportions of leisure/sport cycling.
The reliance on Census data which is now ten years old also means that spatial trends will not account for new build developments, although flows from these sites may be included in the traffic counter data for more recent years.
The spatial trends described above hold true for 2011, when the Census was conducted. When analysing change through time before and after this date, there is uncertainty related the measures used to assess change in cycling. At the regional level, it is possible to use DfT regional traffic data and NTS survey returns, but these are not available at higher geographic resolutions.
The sub-regional analyses rely on DfT traffic counters as a basis for estimating change through time. However, a key limitation is that the count locations are not consistent from year to year. Some locations see counts every year, but most do not, leading to a proponderance of missing data which impacts on observed trends.
In addition, counts are undertaken on a variety of road types, from trunk ‘A’ roads to ‘C’ and unclassified roads. Expected cycle flows differ by road type. Both nationally and within a Local Authority, the proportion of counts that are located on roads of each type varies from year to year.
This causes inter-annual flux in observed cycle flows which can mask underlying trends.
Smaller Local Authority may have both few accidents per year and few DfT counter locations.
Resampling and empirical Bayes…