Identifying and Assessing Perceived Cycling Safety Components
The data analysis was conducted in three stages. In the first stage, descriptive statistics were estimated from the survey to provide an introductory understanding of perceived safety in relation to bicycling behavior and perceived safety factors. Chi-square tests were used to determine the significance of perceived safety differences between cyclists and non-cyclists. The weighted mean and standard error were estimated for each perceived safety factor.
Principal components analysis (PCA) was used as a dimension reduction technique to identify a smaller subset of representative variables accounting for a large share of the variation in the original set. More specifically, this technique identifies a sequence of linear combinations in the data that have maximal variance and are mutually uncorrelated [28]. The PCA provided information on how the eleven factors outlined in Table 1 could be effectively summarized using a reduced set of dimensions. The number of components was selected based on a review of a scree plot, where a point is identified in the plot at which the proportion of variance explained by a subsequent component becomes marginal [28]. Factor loadings greater or equal to |0.6| were used as cut-off values to define components [29]. The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy justified the use of PCA for the current sample, with a value of 0.87 [30]. Bartlett’s sphericity test further confirmed the data’s suitability for PCA (X² = 35074.38, p 31].
In the last analysis stage, principal component regression (PCR) examined the association between each dimension and the outcome of interest (in this case, whether respondents rate cycling as unsafe). Specifically, odds ratios from the PCR were used to examine the directionality and relative magnitude of each component on the outcome. These analyses incorporated interactions for cycling purposes to assess differences in bicycle safety perceptions by cycling purposes (recreation, to commute to work/school, etc.). In addition, the average predicted probability of perceiving cycling as unsafe was calculated for a quantile shift in each component. For all analyses, R version 4.0.3 [32] as well as the “radiant” [33], “sjPlot” [34], and “ggplot2” [35] packages were used.
3.3. Results: Principal Component Regression
Table 6 presents logistic regression models for the five principal components identified through the PCA described in the previous section. The model coefficients adjust for individual-level factors recognized as being associated with perceived safety; these factors include whether an individual cycles, their age, sex, and location (whether a respondent lives in a metropolitan area). Four models were used to identify the association between the aforementioned predictors and a respondents’ perception that cycling is unsafe. The first model uses the predictors as regular regression coefficients, while the remaining models incorporate interaction terms for each of the principal components—one interacting non-cycling status with each component and two interacting cycling purposes with each component (one focused on recreational cycling and another on bike commuting).
The first model identifies two components as significant predictors of unsafe cycling perception. This suggests that individuals who are concerned with cycling risk (namely, crash and injury risk) have increased odds of considering cycling as unsafe compared with those attributing less importance to cycling risk. Second, those concerned with street conditions (e.g., pavement quality) have increased odds of considering cycling as unsafe, but the association is not as strong as for the injurious collision component. In addition to the safety components, we find that an individual’s sex, whether they reside in a metropolitan area, and whether they cycle are significant factors for predicting perceived cycling safety.
The second model incorporates an interaction term, which enables us to compare non-cyclists in relation to cyclists across the five safety components. In this model reflecting the effects among non-cyclists, the same components as the prior model are significant, as well as the component representing weather exposure; the weather component is also the only interaction term significant for predicting unsafe cycling perception. The weather interaction term suggests that compared to cyclists, non-cyclists perceive safety in a way that aligns more closely with their concerns about weather conditions. Figure 2 presents the average prediction of unsafe cycling perceptions. As the importance of weather conditions increases, the predicted probability of perceiving cycling as unsafe increases for non-cyclists. In contrast, as the importance of weather conditions increases, cyclists have a lower predicted probability of perceiving cycling as unsafe. More generally, the weather interaction term suggests that non-cyclists are more intimidated by adverse weather conditions than cyclists.
The third model incorporates a comparison between those who cycle for recreational purposes (e.g., exercising) and those who cycle for functional purposes (e.g., to access public transport). For the purposes of this comparison, the model is limited to cyclists (i.e., those who reported bicycling at least on an occasional basis). The Model 3 interaction terms presented in Table 5 display differences for recreational cyclists in relation to functional cyclists. In this model, the only statistically significant safety component is the injurious collision risk component, suggesting that respondents ascribing greater importance to cycling risk have increased odds of considering cycling unsafe. In terms of the interactions, both the injurious collision risk and the street conditions interaction terms were statistically significant. These findings indicate that those who cycle recreationally have an increased odds compared to functional cyclists to consider cycling unsafe as they place a greater importance on cycling risks and street conditions. As shown in Figure 3, for those considering collision risk as unimportant, few consider cycling as unsafe, and recreational cyclists are less likely than other cyclists to consider cycling as unsafe. In contrast to Figure 2, in Figure 3, the two comparison groups exhibit the same general trend with all cyclists having a greater probability of considering cycling as unsafe as the collision component increases in importance. However, as the collision component becomes more important, recreational cyclists have a greater probability of considering cycling as unsafe compared to other cyclists.
The final model incorporates a comparison between those who cycle for commuting purposes and those who cycle but not to commute to work or school. As with the previous models, this analysis includes a subset of the survey respondents who identified cycling at least on an occasional basis. Since model two provided a comparison between cyclists and non-cyclists, models three and four are included to isolate differences between cyclists. For this model comparing bike commuters to other cyclists, the base variable findings are largely the same, with injurious collision risk and street conditions being the key safety components determining unsafe perceptions of cycling. In contrast to the previous two models, there were no statistically significant interaction terms; this could be due, in part, to a smaller sample size of bike commuters in contrast to other cyclist categories. Despite this limitation, the directionality of the estimated odds ratios for each interaction term provides unique insight into this small faction of cyclists. All five interaction terms were negative, suggesting that bike commuters place less importance on each safety component compared to other cyclists, with the possible exception of the contaminant component, which showed a nearly null effect.
Table 7 presents the change in average predicted probability of unsafe bicycling perception as respondents’ shift from a moderate concern to a moderate unconcern for each safety component. The component with the greatest change in average predicted probability of perceiving cycling as unsafe was injurious collision with the contaminant component having the smallest average predicted effect. Specifically, respondents changing from moderate concern for injurious collision to moderate unconcern had a 21.6 percent reduced average predicted probability of perceiving cycling to be an unsafe activity. The components representing injurious collision, street conditions, and contaminants were found to lower the probability of perceiving cycling as unsafe moving from a moderate concern to a moderate unconcern. The components representing weather and crime worked in the opposite direction; as these components shifted to becoming less of a concern, the probability of perceiving cycling as unsafe increased, although the percent difference was small.
