Is cycling dangerous?

Studying a map with geo-located bicycle crashes might leave you with the impression that cycling must be terribly dangerous. A little bit of rudimentary statistics definitely helps at this stage. Whether something is regarded as dangerous or not ultimately depends on the underlying statistical population. This is a common concept for example in medicine. In a drug’s package insert, the risk for suffering from adverse effects is always related to a population. This helps the consumer (or the medical doctor) to draw informed conclusions. The risk, or incident rate, expresses the probability for a drug to become dangerous. Something similar is still missing for cycling. At least at the local scale level.

Geo-located, reported bicycle crashes between 2002 and 2011. Data source: City of Salzburg. Details are published in Loidl et al. (2016 internet).

Alberto Castro internet and colleagues published an extensive study on exposure-adjusted road fatality rates of pedestrians and cyclists just recently. Compiling data from most OECD countries, this is the first systematic study of this kind and a huge step forward. However, the calculated fatality rates are based on very highly aggregated statistics. In the authors’ own words, ‘exposure data was found to be generally poorer than for fatality data, as travel distances of active modes are not systematically collected in all countries’ (Castro et al. 2018 internet, p. 8). Now, if the availability of exposure data is poor on at national level, how to interpret local crash data or so called crash blackspots – a challenge planners and authorities are facing in cities on a daily basis?

In fact, the problem is that we do not know where, when, how many and which types of cyclists are on the road in most cases. Consequently, we are lacking required exposure data. Moreover, we can only roughly estimate the demand for infrastructure and the respective capacity, and finally, the effect of interventions remains opaque. Regarding the interpretation of crash data, the map below perfectly illustrates the problem. The message of the mapped bicycle crashes seems to be pretty obvious: on the road along the river, crashes are recorded every few meters (crash data were collected over a 10-year period). It looks like this road was quite dangerous for cyclists. In contrast, on the parallel road no crashes are recorded at all. Is this the safe alternative? Well, a quick look at the attached street views answers the question instantaneously. The road along the river is a highly frequented cycle way, whereas the parallel road is exclusively dedicated to motorized traffic (plus a narrow sidewalk). Because hardly any cyclist is riding there, no crashes occur. The probability of being involved in a crash is, at least in part, a function of traffic volume.

In order to overcome the limitation of missing exposure data, we have been working on an agent-based bicycle flow model with a very high spatial and temporal resolution in the collaborative research project FamoS internet. This week, I’m going to present results from this research at the International Cycling Safety Congress internet in Barcelona.


The building blocks of our agent-based simulation model are single trips. We expect flow patterns to emerge from individual mobility behavior, which is determined by multiple parameters. With this approach, it is possible to anticipate the heterogeneity of cyclists. Different to motorized mobility, where the machine levels out different capabilities (elderly people are able to drive at the same speed as youngsters), the variety of behavioral variables and riding styles is huge among cyclists. A 4-year old girl on a bicycle has little in common with a bike courier, just to give an example.

In our model, we can control for different socio-demographic variables, such as age, education or employment status. Additionally, we differentiate between different trip purposes, trip length and accessibility of destinations. After initialization, we simulate agent’s activities, schedules, destinations, mode choice and route choice. Of course, such a model requires many data. In the case of a model we developed for the Salzburg central region, we used:

  • Topological correct road graph with a rich set of attributes
  • Census data with a spatial resolution of 100 and 250 meters from Statistik Austria
  • Central facilities and POIs from OGD portals
  • Raw data from mobility surveys
  • Time use statistics from representative surveys

The bicycle flow model was developed from scratch by Dana internet. She did an excellent job by translating the conceptual model into code. The simulation model runs on GAMA platform and will be freely available at OpenABM internet soon. Through a very efficient code structure, we are able to initialize and run the simulation model for a 24-hours day, with 150,000 agents, a temporal increment of one second and a spatial resolution of one meter within only five hours. This very fast runtime makes the model perfectly suitable for sensitivity analysis and simulation of various interventions, such as additional connections or changing behavior of travelers.

Comparison of simulated bicycle trips (left) and recorded data from the Bike Citizens app (right).

The model was calibrated with data from six stationary counters. For model validation, we used tracking data from Bike Citizens internet. In total, we achieved very accurate results. The temporal distribution of bike rides perfectly matches the double peak signature in reference data. The simulated spatial pattern of bicycle flows has a little bias towards the left side of the Salzach river. Reasons for this are expected to be associated with known biases in the input data. In the validation, we also compared the characteristics of simulated and recorded trips. Whereas the mean travel time is much higher for Bike Citizens data (mainly due to a few outliers, generated by long distance leisure cyclists), the average distance and speed distribution resembles perfectly.

To the best of my knowledge, this is the first bicycle flow model at the local scale level with a regional coverage. We use the results for multiple purposes. For example, we could simulate the expected effect of planned bicycle corridors in the city of Salzburg. In the context of safety, the model is well suited to generate exposure data for risk analysis. Referring to the example above, the simulated bicycle flows can be nicely used to calculate incidence rates and subsequently assess the safety of areas and road segments (we published a paper in Safety internet on this topic). In the particular case presented above, it becomes evident that the road along the river is definitely not dangerous. The recorded crashes can be expected from the number of cyclists. However, it is beyond any discussion that every bicycle crash is one too much. Providing adequate infrastructure is crucially important for attracting cyclists and ensure safe rides. Here again, the simulation model helps to estimate the demand and derive required capacities for dedicated infrastructure.

With the agent-based simulation model, we have made a step forward in providing sound evidence to decision makers and bicycle advocates. Nevertheless, it is still a model and thus, it does not mirror reality, but a generalized representation. In order to further refining the model we are currently improving the input data basis In the research project Bicycle Observatory internet, we combine quantitative and qualitative data from different sources for getting an integrated perspective on bicycle mobility. This will help us to include even more parameters in the model and hence, provide a more accurate representation in the spatial, temporal and behavioral dimension.



  1. htiemens

    I am impressed by the accuracy of the model and the way you handle 24 hours and the calculation of incident risks. It would be very nice to compare your work with BRUTUS, the model developed by Strafica in Finland. This is also an agent based model, in use for several years for Finland, but also in the province of Utrecht, Netherlands. On first sight your model is better because it takes different profiles of cyclists and different route choices. In BRUTUS this is developed in a more random assignment, I wonder how this affect the results.

    • gicycle

      Thanks Herbert for this positive feedback!
      Are you going to be at the ICSC tomorrow? Or is there any other chance to discuss and compare the 2 approaches? I’d be happy to have a closer look to BRUTUS.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s