1. Questions

After a bicyclist experiences a crash or near miss, they may alter their riding behavior, such as reducing their frequency of biking or replacing bike trips with other forms of transportation that feel safer. For example, Lee, Underwood, and Handy (2015) demonstrate that the severity of injury a bicyclist sustains during a crash is significantly associated with a decline in their bicycling comfort and willingness to continue biking. Identifying the factors that contribute to behavioral change post-incident is essential for improving our understanding of bicycling safety and informing targeted interventions. In this study, we examine how various characteristics: type of incident, whether the respondent was a regular cyclist, whether the respondent was wearing a helmet, terrain, gender identity, age group, who/what the respondent had an incident with, type of injury, and type of road conditions, predict changes in biking behavior after a reported incident. Specifically, we aim to address the following research question:

  1. What characteristics are associated with a change in biking behavior following a crash or near miss?

2. Methods

We use self-reported collision and near miss data submitted to BikeMaps.org, a global crowdsourcing platform where bicyclists can report collisions, near misses, hazards, and thefts. Researchers have shown that BikeMaps data are best used to complement traditional data sources, but it is important to acknowledge the demographic limitations of this data such as a bias towards younger, more technological savvy users (Ferster et al. 2017). Previous research in Branion-Calles, Nelson, and Winters (2017) show how BikeMaps crash and near miss data can be used to complement and fill gaps in reports to official collision sources.

For this study, we focus on 2,502 incidents reported in the United States and Canada and limit our dataset to responses with complete records to ensure consistency in modeling (Figure 1). In our study we include both collisions and near miss reports in our analysis, because prior research emphasizes the importance of near miss events for perceived safety (Branion-Calles, Nelson, and Winters 2017).

Figure 1
Figure 1.Map of 2502 submitted reports used in analysis.

We perform an extensive data cleaning process, including filtering incomplete responses, collapsing low-frequency categories to ensure balanced class distributions, and excluding variables with minimal variation. Table 1 shows the different variables of the model as well as their respective values and frequencies.

Table 1.Included variables in the model
Variable Value n Percentage (%)
Type of incident Collision 672 26.9
Near Miss 1830 73.1
Is the respondent a regular cyclist? No 37 1.5
Yes 2465 98.5
Was the respondent wearing a helmet? No 314 12.5
Yes 2188 87.5
Terrain Downhill 376 15
Flat 1880 75.1
Uphill 246 9.8
Gender Identity Female 769 30.7
Male 1733 69.3
Age-Group 5-29 136 5.4
30-39 785 31.4
40-49 745 29.8
50-59 340 13.6
>= 60 496 19.8
Impact on biking Bike less; More careful; More careful and bike less; Stopped biking 1426 57
none 1076 43
Incident with Animal 21 0.8
Cyclist or E-scooter 68 2.7
Infrastructure 96 3.8
Other 67 2.7
Pedestrian 48 1.9
Vehicle, angle; Vehicle, side 769 30.7
Vehicle, head on; Vehicle, rear end 462 18.5
Vehicle, open door; Vehicle, passing 361 14.4
Vehicle, turning left 200 8
Vehicle, turning right 410 16.4
Type of injury Injury, hospital emergency visit or hospitalized 182 7.3
Injury, no treatment, and Injury, saw a family doctor 305 12.2
No injury 2015 80.5
Road conditions Dry 2097 83.8
Other 405 16.2

We analyze predictors of post-incident biking behavior (no change versus some change [which includes biking less, biking more carefully, biking more carefully and biking less, and stopped biking]). We use the following predictors: type of incident, whether the respondent was a regular cyclist, whether the respondent was wearing a helmet, terrain, gender, age group, who/what the respondent had an incident with, type of injury, and type of road conditions. We then utilize a random forest classification model using the tidymodels package in R (Kuhn and Wickham 2020). Table 1 reports the variables used in the development of our model.

We trained our model on 70% of data or 1751 reports and tested on the remaining 30% of data or 751 reports. We interpret our model utilizing a variable importance plot and partial dependence plots. We utilize variable importance plots to see what variables contributed more to a decision made in a random forest tree. We perform partial dependence plots in order to show the effect of a variable on the predicted outcome, while maintaining the other values of the feature.

3. Findings

Our random forest model achieved a ROC AUC of 0.64, indicating modest discriminatory ability which leaves room for improvement. One contributing factor may be the class imbalance of our variables as shown in Table 1. While we mitigated this by combining classes when appropriate, the class imbalance may affect the model’s ability to predict. We also acknowledge the spatial imbalance in our dataset, with most reports in Canada, which may limit our ability to capture more regional-specific behavioral responses (Figure 1). Variable importance plots (Figure 2) were used to interpret which predictors were most vital in the model making the decision in the random forest classification for determining biking behavior following an incident. The x-axis displays the Gini importance, a relative measure of how impactful each variable is in making decisions in the model. The most influential predictors contributing to node splitting in the random forest classification model for predicting biking behavior were hospitalized injury followed by collision classification and then identification as a female.

Figure 2
Figure 2.Variable importance plots in RF model for predicting bicycle behavior changes post safety incident

While variable importance plots indicate the relative contribution of each predictor to the model’s decision-making process, they do not convey the direction of a variable’s influence on biking behavior. We therefore use partial dependence plots to examine the marginal effect of individual predictors on the probability of a given biking behavior outcome (Figure 3).

The partial dependence plots show that individuals who experienced a hospitalized injury, reported a collision rather than a near miss, or identify as female are more likely to have a change in biking (which includes biking less, biking more carefully, biking more carefully and biking less, and stopped biking) behavior following a crash or near miss. These results are consistent with previous findings showing that women are more likely than men to reduce their cycling exposure following a crash (Fraser and Meuleners 2020). These predictors are positively associated with the probability of some type of biking behavior change while holding the other variables constant, as illustrated by the positive slopes in the partial dependence plots (Figure 3).

Figure 3
Figure 3.Partial dependence plots for variables in the model for predicting bicycle behavior changes post safety incident

Our models can be used to help identify bicyclists who are more likely to reduce bicycling following a safety incident. This can inform the allocation of post-crash resources or targeted interventions. Future work should investigate whether cities offering support services see improved retention of bicyclists after crashes. The reports combined alongside qualitative interviews may provide further insight into the types of support most valued by affected populations. Our work can be situated alongside other research such as Fraser and Meuleners (2020), which identifies protective factors such as group riding and those who had a full recovery post safety incident being associated with lower odds of reducing bicycling exposure. Our work can help shape data-based interventions at improving post biking safety incident recovery and retention.