Risk Perception Accuracy Among Urban Cyclists: Behavioral and Infrastructural Influences in Loja, Ecuador

Abstract

Urban cycling faces the challenge of cyclist vulnerability due to infrastructural deficiencies and complex traffic environments. This study evaluates the accuracy of risk perception among 153 urban cyclists in Loja, Ecuador, using a mixed-methods design that integrates self-reported behaviors (Cycling Behavior Questionnaire—CBQ), visual assessments of 12 road segments, and objective risk classifications derived from the CycleRAP methodology. Results show a notable misalignment between perceived and actual risk, with consistent underestimation of extreme-risk scenarios and overestimation of low-risk ones. The combined CBQ score (violations + errors) emerged as the strongest predictor of inaccurate risk perception in decision tree models, explaining 28.75% of the model’s predictive power. Interestingly, cycling experience did not improve accuracy; frequent cyclists with high violation/error scores and older age showed the poorest perception, while young cyclists with moderate behavior scores exhibited higher accuracy. These results suggest that the relationship between cycling experience and risk assessment is more complex than commonly assumed. Findings highlight the need for behavioral interventions to correct misperceptions, alongside infrastructural measures that address objective hazards. Given the limited number of road segments and moderate sample size, subgroup analyses may be underpowered and should be interpreted with caution.

1. Introduction

Urban cycling has emerged as a sustainable and health-promoting mode of transportation, offering numerous benefits such as reduced traffic congestion, lower emissions, and improved public health [1]. Despite these advantages, cyclists remain among the most vulnerable road users, facing a significantly higher risk of crashes and injuries than motor vehicle occupants [2]. This vulnerability is often linked to poor cycling infrastructure, low driver awareness, and the complex and hazardous nature of urban traffic environments [3]. In high-income countries, many of these challenges have been addressed through consistent investment in infrastructure, education, and public policy [4].

However, in low- and middle-income countries, such as Ecuador, these problems remain largely unresolved, making urban cycling less safe and less attractive. Even so, cycling is gaining popularity in these regions due to its low environmental impact and affordability [5]. In cities like Loja, in Ecuador, local efforts are beginning to promote cycling, but the lack of adequate infrastructure and safety measures poses significant barriers. In this context, understanding how cyclists perceive risk becomes crucial, as these perceptions play a key role in shaping their mobility decisions and overall sense of safety.

The perception of risk among cyclists significantly influences their behavior and route choices, often determining whether individuals choose to cycle and which paths they prefer [6]. Perceived risk can deter potential cyclists, especially in environments where infrastructure is lacking or traffic conditions are perceived as hazardous [7,8]. Research indicates that discrepancies between perceived and actual risks can lead to either overcautious behavior, limiting the benefits of cycling, or underestimation of danger, increasing the likelihood of crashes [9]. Therefore, assessing both subjective perceptions and objective measures of risk is essential for a comprehensive understanding of cycling safety [10]. This dual assessment can inform targeted strategies to align perceptions with actual risk levels, thereby encouraging safer cycling practices [11]. To complement the understanding of subjective perceptions, it is also necessary to examine objective risk factors present in the built environment.

Objective assessments of cycling risk often involve analyzing infrastructure quality, traffic patterns, and road crash statistics. Tools like the CycleRAP methodology provide standardized evaluations of road segments based on factors contributing to cyclist safety, such as road design, traffic volume, and speed limits [12]. These assessments yield risk ratings that can guide infrastructure improvements and policy decisions [13]. However, objective measures alone may not capture the nuanced experiences of cyclists, highlighting the need to integrate subjective evaluations into safety analyses [14]. By comparing objective risk assessments with cyclists’ perceptions, researchers can identify mismatches that may affect cycling behavior and safety outcomes [15]. In addition to assessing risk perception and infrastructure, analyzing cyclists’ actual behaviors offers further insights into safety dynamics.

To effectively assess cyclists’ behaviors and perceptions, researchers have developed validated instruments such as the 44-item Cycling Behavior Questionnaire (CBQ) [16]. This tool evaluates multiple dimensions of cycling behavior, including violations, errors, and positive practices, offering valuable insights into factors that may either increase road crash risk or contribute to safer cycling [17]. The CBQ has demonstrated robust reliability and validity across diverse populations, establishing its utility as a key instrument in behavioral research related to cycling. Additionally, its adaptability across cultural and infrastructural contexts, exemplified by the 29-item Spanish version [16], supports its use in comparative studies and the formulation of context-specific interventions. Integrating the CBQ into research allows for a comprehensive examination of how individual behaviors and risk perceptions interact with environmental conditions to shape cycling safety.

In addition to behavioral assessments, visual evaluations of cycling environments can provide context for understanding risk perceptions. Presenting cyclists with images of various road segments and soliciting their safety ratings allows researchers to gauge subjective risk assessments in relation to specific infrastructural features [18,19,20]. Simulator and online studies promise to provide a fast and easy way of rapid prototyping infrastructure layouts [21]. This approach can reveal which environmental characteristics are perceived as hazardous or safe, informing infrastructure design and educational campaigns [22]. Combining visual assessments with behavioral questionnaires like the CBQ offers a multifaceted perspective on cycling safety [23], encompassing both individual behaviors and environmental perceptions. Such comprehensive analyses are instrumental in developing interventions that address both behavioral and infrastructural determinants of cycling safety.

Building on this foundation, the present study focuses on the urban context of Loja, a city located in the southern region of Ecuador. The main objective of this study is to examine the relationship between cyclists’ perceived risk, their self-reported behaviors, and the objective risk levels of cycling infrastructure in the city of Loja. To this end, a mixed-methods approach was employed, integrating surveys that included the Cycling Behavior Questionnaire (CBQ), visual assessments of road segments, and objective evaluations using the CycleRAP methodology. Participants were asked to assess the perceived safety of different urban road segments shown in images and to provide information about their cycling behaviors and sociodemographic characteristics. The resulting data were analyzed to explore correlations among perceived and actual risks, behavior patterns, and demographic variables.

This study contributes to the growing body of research on urban cycling safety by integrating cyclists’ subjective perceptions, self-reported behavioral patterns, and objective infrastructure risk assessments within a single mixed-methods framework. By focusing on the case of Loja, Ecuador, a medium-sized city in a low-to-middle-income country with limited but evolving cycling infrastructure, this research addresses critical knowledge gaps related to risk perception accuracy in understudied urban contexts. The findings offer practical insights for policymakers, planners, and public health advocates aiming to improve cyclist safety through targeted interventions that align user perceptions with real infrastructure conditions. Ultimately, this study provides a replicable model for evaluating urban cycling safety in similar Latin American cities undergoing mobility transitions.

2. Materials and Methods

2.1. Study Area

This study was conducted in the city of Loja, located in southern Ecuador. As a context, the province of Loja has an estimated population of 215,000 inhabitants, with a population density of 93.1 people per square kilometer [24,25]. The capital city, also named Loja and the focus of this study, is home to approximately 118,532 residents [25], with a growing interest in active mobility modes, including cycling. However, the city faces significant challenges due to the lack of adequate dedicated cycling infrastructure and the prevalence of mixed-traffic conditions. These factors make it an appropriate setting for studying perceived and objective cycling risks.

Loja’s cycling infrastructure consists of a mix of dedicated bike lanes and shared mixed-traffic corridors. The two longest and most continuous north–south bike lanes traverse the urban expansion axis, connecting residential districts with central commercial areas. The network is fragmented elsewhere, with varying degrees of physical separation, signage, and surface quality; most lanes lack full continuity and often intersect with motor vehicle traffic without protective buffers.

2.2. Participants and Sampling

Participants were recruited using a non-probabilistic convenience sampling strategy between April and August 2025. Recruitment occurred via two parallel channels: (1) online dissemination through local cycling groups on social media (Facebook and WhatsApp), where posts included the study purpose, inclusion criteria (aged ≥ 15 years, cycling several times a month within Loja urban area), and a link/QR code to the survey; and (2) in-person data collection by trained civil engineering students enrolled in the Road Construction II course. For the latter, teams visited common cycling locations and urban nodes, approaching cyclists, verifying eligibility, and facilitating survey completion on tablets or smartphones. Data collectors followed a standardized script to ensure consistency and reduce interviewer-induced bias. No incentives were offered; participation was voluntary and anonymous.

A previous study [26] reported a daily user count of 104–167 cyclists on one of the main north–south bike lanes, meaning this refers to average daily distinct cyclists observed over typical weekday conditions. Considering that there are two major and longest bike lanes running in this direction, which aligns with the city’s north-to-south urban expansion, it can be estimated that approximately double that number, i.e., between 208 and 334 cyclists, use the bike lane network daily. Based on the estimated population range (208–334 daily cyclists) and aiming for a 95% confidence level with a 5% margin of error, the required sample size was calculated as 136 responses (for n = 208) to 179 responses (for n = 334). To ensure adequate coverage across this range, 171 responses were initially collected. After applying quality control criteria and excluding 18 responses, the final sample included 153 valid responses, which falls within the calculated range and provides adequate statistical power for the study objectives.

Responses were excluded based on predefined quality control criteria: duplicate or near-duplicate submissions (identified via identical response patterns submitted within a short time window), completion times under two minutes (deemed insufficient to meaningfully read and respond to the full survey), and violations of inclusion criteria (one respondent reported age 14, below the minimum of 15). These filters were applied programmatically and manually reviewed to ensure that only unique, attentive, and eligible responses remained. The sample was predominantly male (84%), which reflects the male-skewed cycling demographics in Loja, as observed in local cycling counts [26]. However, the use of convenience sampling may have reinforced this imbalance. This limitation is acknowledged as it may affect the generalizability of results to female cyclists. No personally identifiable information was collected, and participation was entirely voluntary and anonymous.

We acknowledge that using convenience sampling limits how generalizable our findings are beyond our specific study population. However, we had to use this approach because comprehensive cyclist registries are unavailable in Loja, and this research is exploratory in an understudied area. Also, our sample’s demographic breakdown might limit insights into gender-specific risk perception patterns.

2.3. Application of CycleRAP Methodology

To obtain objective risk classifications, the CycleRAP (Risk Assessment Program) methodology [12] was applied to the entire cycling network of the city of Loja, with assessments conducted every 10 m along each segment of the infrastructure. The risk classification considered the following elements:

• Traffic volume and speed were estimated based on the Google Traffic color codes, which represent real-time traffic conditions derived from Floating Car Data (FCD). These estimations were calculated using previous equations [27]. That prior research specifically developed and validated six density–flow and six speed–density equations using both power and linear curves. These equations establish a direct and quantitative relationship between Google’s reported traffic colors and the corresponding speed, volume, and flow for specific streets in Loja. Our study leverages these relationships: by combining observed Google Traffic colors with field-collected speed data from Loja, we can determine the density of a road segment using the calibrated equations. Once density is established, we then use the appropriate calibrated equation to determine the corresponding traffic volume.
• Pedestrian presence was inferred from land use characteristics, where higher pedestrian densities were assumed in central zones and near high-traffic pedestrian areas such as schools, markets, and parks. In contrast, lower pedestrian activity was assigned to industrial and peripheral zones.
• Slopes and cross-sections were measured using a combination of mobile phone applications equipped with integrated cameras and GPS and in-field measurements by trained students.
• Other variables considered included road width, surface quality, traffic separation, intersections, and lighting conditions, all following the CycleRAP coding manual.

The CycleRAP methodology was selected as our objective risk assessment tool due to its standardized, internationally validated approach to cycling infrastructure evaluation. CycleRAP has been successfully applied in diverse urban contexts and provides reliable risk classifications that correlate with actual crash data. We acknowledge limitations in our traffic volume and speed estimations using Google Traffic data. The selection of 12 images represents a balance between comprehensive risk representation and participant burden, though we recognize that future studies would benefit from larger image sets to capture greater environmental diversity.

2.4. Study Design and Instruments

This research followed a cross-sectional, mixed-methods approach, combining quantitative measures with perceptual assessments. Data were collected using an online questionnaire hosted on Microsoft Forms^®, which consisted of four main sections:

2.4.1. Section 1: Demographic Information

Participants reported their age, sex, cycling frequency, primary reason for cycling (e.g., commuting, recreation, sport), and the number of years they have used a bicycle in the city of Loja.

2.4.2. Section 2: Cycling Behavior

The CBQ is a validated instrument capturing three behavioral dimensions: violations, errors, and positive behaviors. For the purposes of this study, only violations and errors were included in the risk perception predictive models because these dimensions reflect risk-enhancing behaviors, whereas positive behaviors represent safety-oriented practices that operate in a different conceptual direction. Including only the risk-contributing components simplified interpretation of the behavioral paradox while avoiding collinearity with protective practices. The 29-item Spanish adaptation used has demonstrated construct validity in similar Latin American contexts [16], and items were aggregated by summing responses per dimension following the original scoring protocol (0–4 scale per item). Internal consistency was assessed via Cronbach’s alpha for each subscale to verify reliability in this sample.

2.4.3. Section 3: Perceived Risk in Urban Environments

To assess perceived risk in specific urban settings, participants were shown 12 images of road segments within the city of Loja. Of the 1075 segments assessed via CycleRAP, 12 sites (3 per risk category) were purposively selected to represent varied urban contexts. Three images were selected for each of the four risk categories: low, medium, high, and extreme. For each image, participants were asked: “How dangerous do you consider this site for cycling?” with the following response options: low, medium, high, or extreme. They were also asked to indicate how frequently they use each location, allowing for analysis of whether frequent users perceive risk differently from occasional users.

We acknowledge that evaluating risk perception using static photographs does not replicate the multisensory and dynamic nature of on-road cycling (e.g., speed perception, auditory cues, and moving traffic). Future studies could incorporate video-based or in situ assessments.

2.4.4. Section 4: Perception of Cycling Infrastructure

This final section included two closed-ended questions: “In general, how would you rate the cycling infrastructure in the city?”, and “Which aspects do you consider most important for improving cycling safety?” where participants could select multiple options.

2.5. Statistical Analysis

All statistical analyses were performed using R software 4.5.0 [28] within the RStudio 2025.05.0 environment [29]. Data extraction from the online survey platform was conducted in CSV format, preserving all response timestamps and participant identifiers. Prior to analysis, the dataset was cleaned and organized using Microsoft Excel^® through the following procedures: (1) removal of incomplete responses (cases with missing data in key variables), (2) elimination of duplicate entries based on timestamp and response patterns, (3) verification that participants met inclusion criteria (e.g., age ≥ 15 years), and (4) standardization of categorical responses to ensure consistency in coding. The main objective of the analysis was to examine how demographic variables, self-reported cycling behaviors, and objective infrastructure risks relate to cyclists’ subjective risk perception and the accuracy of these perceptions.

Descriptive statistics were first calculated to summarize the demographic characteristics of the sample, including age, gender, frequency and purpose of bicycle use, and years of cycling experience. Central tendency and dispersion measures were also computed for the Cycling Behavior Questionnaire (CBQ) scores, disaggregated into violations, errors, and positive behaviors. To assess the internal reliability of the behavioral subscales, Cronbach’s alpha coefficients were calculated using the psych 2.4.7 package in R.

To evaluate the accuracy of perceived risk, each participant’s responses to the twelve site photographs were numerically coded on a four-point ordinal scale corresponding to the CycleRAP classification (low = 1, medium = 2, high = 3, and extreme = 4). For each participant, a mean perceived risk score was computed and compared against the CycleRAP objective rating. Three key indicators were derived: precision score (mean absolute deviation between perceived and actual risk), accuracy rate (proportion of exact matches between subjective and objective risk), and mean bias (average directional difference between perceived and actual risk). These indicators allowed for the identification of overestimation or underestimation patterns across the different risk categories.

Associations between demographic characteristics, CBQ scores, and perception accuracy were explored using Spearman’s rank correlation coefficients. This non-parametric approach was selected due to the ordinal nature of some variables and potential non-normality in the data. Additionally, chi-square tests were applied to explore relationships between categorical variables such as perceived and objective risk classifications, and frequency of use of the photographed locations.

To assess the combined influence of behavioral and demographic factors on perception accuracy, a multiple linear regression model was constructed using the precision score as the dependent variable. Independent variables included age, cycling frequency, purpose of use, years of experience, and the total CBQ score (sum of violations and errors). Model assumptions were evaluated using the car 3.1-3 package, and results were interpreted based on coefficient significance and overall model fit, reported through adjusted R-squared and p-values.

Given the complexity of behavioral interactions, a decision tree analysis was also conducted using the rpart 4.1.23 and rpart.plot 3.1.2 packages in R. This classification model aimed to predict high vs. low precision in risk perception (based on the sample median) using the same set of predictor variables. Model performance was assessed through a training–test split (80/20) implemented with the caret 6.0-94 package, and metrics such as accuracy, sensitivity, specificity, precision, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC) were computed using the caret and pROC 1.18.5 packages to evaluate discriminative power.

The area under the receiver operating characteristic curve (AUC) was computed to assess discriminative ability; although often conflated with ‘ROC’, the acronym AUC refers specifically to the area under the ROC curve.

Finally, correspondence analysis was employed using the ca 0.71.1 package to visually explore the relationship between perceived and objective risk classifications across the twelve urban sites. This graphical technique provided insight into how cyclists’ evaluations aligned or failed to align with infrastructure-based risk assessments. Data visualization was enhanced using ggplot2 3.5.1, RColorBrewer 1.1-3, and gridExtra 2.3 packages for publication-quality figures. Throughout the analysis, a significance level of p < 0.05 was adopted. Non-significant results were retained in the report for transparency and to inform the discussion of complex or unexpected patterns in risk perception.

2.6. Ethical Considerations

While the study involved collection of behavioral data through the CBQ, formal ethical approval was not required under current institutional regulations for minimal-risk studies involving anonymous survey data with no collection of personally identifiable information. Nevertheless, the research complied with core ethical principles of respect for persons, beneficence, and justice. All participants were fully informed about the study’s purpose, the voluntary nature of participation, and their right to withdraw at any time. No personally identifiable data were collected, and all responses were anonymized immediately upon collection.

3. Results

3.1. Participant Characteristics

A total of 171 cyclists participated in this study. All submitted responses are provided in the Supplementary Materials. However, 18 responses were excluded from the analysis based on specific quality control criteria.

Table 1. Summary of participant characteristics (n = 153).

Variable	Category	Value
Age	Minimum	15
	Maximum	72
	Mean	30.15
	Standard deviation	10.80
Sex (%)	Female	15.70
	Male	84.3
Bicycle frequency of use (%)	Every day	22.88
	Several times per week	45.10
	Once per week	16.34
	A few times per month	15.69
Purpose of use (%)	Workout	56.86
	Recreation	23.53
	Work	14.38
	Other	4.58
Years using a bike (%)	Less than 1 year	4.58
	1 to 3 years	21.57
	More than 3 years	73.86

summarizes the main characteristics of the participants. Most respondents were male, with a wide age range and varying levels of cycling frequency and experience. The primary reasons for bicycle use included exercise or workout, recreation, and commuting. Overall, a significant proportion of participants reported cycling regularly and having more than three years of experience.

3.2. CycleRAP Results

The bike lane network in the city of Loja was assessed using the CycleRAP methodology at 10 m intervals, resulting in a total of 1075 evaluated segments. These results can be accessed at https://smartland.maps.arcgis.com/apps/dashboards/b47ad5eadce94001b543271a2b6e054b accessed on 30 May 2025 or in Supplementary Materials.

Table 2. Summary of CycleRAP results for the Loja city bike lane.

Types of Conflict	Risk Level	Percentage (%)
Bicycle vs. Bicycle	Low	74
	Medium	25
	High	1
	Extreme	0
Bicycle vs. Pedestrian	Low	27
	Medium	52
	High	21
	Extreme	0
Single Bicycle	Low	70
	Medium	27
	High	3
	Extreme	0
Vehicle vs. Bicycle	Low	4
	Medium	48
	High	46
	Extreme	2

summarizes the percentage distribution of conflict types and CycleRAP risk levels: bicycle vs. bicycle, bicycle vs. pedestrian, single bicycle, and vehicle vs. bicycle. Bike–car interactions are the riskiest interactions for cyclists [7]. At each evaluated location, individual conflict scores for bicycle–bicycle, bicycle–pedestrian, single bicycle, and vehicle–bicycle interactions were computed following CycleRAP’s scoring schema. The overall risk score for a site was obtained by summing these individual conflict scores, yielding a composite measure that was then categorized into low, medium, high, or extreme risk based on established thresholds. In terms of overall risk distribution among the evaluated segments, 2% were classified as low risk, 50% as medium, 45% as high, and 3% as extreme.

From all the segments evaluated by CycleRAP, a total of 12 sites were selected for further analysis, as shown in Figure 1. Twelve sites were purposively selected to include three representative photographs from each of the four objective risk categories (low, medium, high, extreme) to enable balanced comparisons. Selection also prioritized diversity in the urban context (e.g., single-lane vs. dual-lane, varying adjacent land uses) to capture a range of environmental conditions while keeping the cognitive load manageable for survey participants.

3.3. Cycling Behavior Questionnaire Results (CBQ Scores)

The survey included the Cycling Behavior Questionnaire (CBQ), consisting of 29 items: 8 related to violations, 15 to errors, and 6 to positive behaviors. Internal consistency analysis was conducted by calculating Cronbach’s alpha coefficients for each subscale using the psych 2.4.7 package in R 4.5.0, following standard psychometric procedures. Cronbach’s alpha was 0.61 for the violation scale, indicating moderate reliability but below the conventional 0.70 threshold recommended for research purposes. In contrast, the error and positive behavior scales showed higher reliability, with alpha values above 0.85, exceeding acceptable thresholds. These results partially align with the original CBQ validation study, which reported alpha values of 0.703 for violations, 0.850 for errors, and 0.705 for positive behaviors [16]. The lower reliability observed for the violation subscale in the current study (α = 0.61 vs. α = 0.703) may reflect cultural or contextual differences in the cycling environment, though the error subscale demonstrated comparable high reliability across both studies.

Responses on the Cycling Behavior Questionnaire (CBQ) were rated on a 5-point scale: 0 = never; 1 = almost never; 2 = sometimes; 3 = frequently; and 4 = almost always. In the overall sample, the mean scores were as follows: violations (M = 0.76, SD = 1.05), errors (M = 0.45, SD = 0.82), and positive behaviors (M = 3.13, SD = 1.10). These results indicate that participants reported engaging in violations and errors infrequently, while frequently demonstrating positive cycling behaviors, particularly when riding in bike lane environments.

To examine the relationship between cycling behavior and risk perception, individual CBQ scores were calculated for each participant. Raw scores for violations and errors subscales were computed by summing item responses within each dimension, maintaining the original 5-point Likert scale scoring (0 = never; 1 = rarely; 2 = occasionally; 3 = quite often; 4 = very often). The violation subscale (8 items) had a theoretical range of 0–32, while the error subscale (15 items) had a theoretical range of 0–60. A composite risky behavior score was created by adding violation and error scores together, representing participants’ overall tendency toward unsafe cycling practices. Descriptive statistics revealed that violation scores had a mean of 6.53 (SD = 4.27, min = 0, max = 20), error scores had a mean of 7.95 (SD = 7.37, min = 0, max = 37), and the combined violation and error score had a mean of 14.48 (SD = 10.13, min = 0, max = 45). These composite scores were subsequently used as predictor variables in regression and decision tree analyses to examine their relationship with risk perception accuracy.

3.4. Risk Perception Accuracy

Each participant rated the perceived risk level of each photograph using four qualitative categories: low, medium, high, and extreme. These ratings were converted into numerical values (low = 1; medium = 2; high = 3; extreme = 4). The set of images included three photographs for each objectively determined risk level, as established by the CycleRAP methodology. Using the numerical scores, an average risk perception score was calculated for each participant, and initial analyses were conducted to assess perception accuracy.

First, the average perceived risk score for each image group was calculated. Then, a precision score was computed as the mean absolute deviation from actual risk and the participant’s perceived value. A response was considered accurate when this absolute difference was less than 0.5; based on this criterion, an accuracy rate was calculated. Accuracy rate is defined as the proportion of individual photograph ratings in which the absolute difference between perceived and objective risk was less than 0.5, reflecting near-exact agreement on the four-point ordinal scale. Additionally, the bias was computed for each group of photographs to identify systematic overestimation or underestimation of risk. These results are summarized in

Table 3. Risk perception accuracy by photographs of the sites.

Photographs	Actual Risk Level (CycleRAP)	Mean Perceived Risk	Precision Score *	Accuracy Rate (%)	Mean Bias **
Photo 1–3	3 (High)	2.33	0.758	11.76	−0.667
Photo 4–6	1 (Low)	2.08	1.078	5.23	+1.078
Photo 7–9	4 (Extreme)	2.24	1.758	2.61	−1.758
Photo 10–12	2 (Medium)	2.16	0.580	19.61	+0.161

* Lower precision scores indicate better accuracy (mean absolute deviation from actual risk). ** Positive bias indicates overestimation; negative bias indicates underestimation.

. Each grouping in Table 3 (e.g., Photo 1–3) represents three distinct photographs from sites sharing the same CycleRAP risk category.

Findings from Table 3 revealed distinct patterns in risk perception accuracy across risk levels:

• Low Risk (Photo 4–6): Participants systematically overestimated low-risk scenarios, with only 5.23% correctly identifying the actual risk level. The mean bias of +1.078 indicates a strong tendency to perceive more danger than actually present.
• Medium Risk (Photo 10–12): This category showed the highest accuracy rate (19.61%) and the lowest precision score (0.580), suggesting that medium-risk scenarios were most accurately perceived by cyclists.
• High Risk (Photo 1–3): Participants underestimated high-risk scenarios (bias = −0.667), with only 11.76% correctly identifying the risk level.
• Extreme Risk (Photo 7–9): This category showed the poorest performance, with only 2.61% accuracy and the highest precision score (1.758). Participants severely underestimated extreme-risk situations (bias = −1.758).

Overall, participants showed limited accuracy in perceiving the actual risk levels, with a general trend of misalignment between perceived and objective assessments.

3.5. Demographic and Behavioral Predictors

Table 4. Risk perception accuracy by cyclist characteristics.

Variable	Category	n	Mean Precision *	Accuracy Rate (%) **
Cycling Frequency	Daily	35	1.000	37
	Several times/week	69	1.070	41
	Once/week	25	0.990	36
	Few times/month	24	1.080	42
Bike Use Purpose	Workout	87	1.030	46
	Recreation	36	1.090	22
	Work	22	1.030	41
	Other	7	1.080	29
	Study	1	1.000	100
Years of Experience	<1 year	7	1.060	57
	1–3 years	33	0.997	36
	>3 years	113	1.060	39

* Lower precision scores indicate greater accuracy, as they reflect a smaller average deviation from the actual risk level. ** A higher accuracy rate indicates that participants correctly identified the actual risk level more frequently.

summarizes risk perception accuracy across various cyclist characteristics. Lower precision scores indicate better performance, while higher accuracy rates reflect more frequent exact matches with actual risk levels.

In terms of cycling frequency, daily cyclists showed the best overall performance with an average precision score of 1.000 and a relatively high accuracy rate (37%). In contrast, those cycling a few times per month had the worst precision (1.080), despite showing the highest accuracy rate (42%), suggesting that while they occasionally matched the actual risk, their overall estimations were less consistent.

When analyzing the purpose of bike use, individuals who cycled for exercise demonstrated relatively strong accuracy (46%) with a decent precision score (1.030), indicating both frequent correct judgments and consistent estimations. Conversely, recreational cyclists had the worst accuracy (22%) and the highest deviation (1.090), revealing an overestimation bias and low consistency in risk perception. Notably, the one cyclist who reported study as the main reason had good alignment with actual risk (100% accuracy), though this is not generalizable due to the single observation.

With regard to years of cycling experience, surprisingly, novice cyclists (<1 year) achieved the highest accuracy rate (57%) but had relatively poor precision (1.060), suggesting cautious or random estimations occasionally aligned with reality. Meanwhile, the most experienced group (>3 years) had the same precision level (1.060) but a lower accuracy rate (39%), which may reflect overconfidence or routine-based risk assumptions.

Overall, the results suggest that frequent cycling and riding for exercise are associated with more accurate and consistent risk perception, while recreational use and increasing experience do not necessarily guarantee improved judgment.

3.6. Correlation Analysis

Table 5. Correlation matrix of key variables.

Variable	Age	CBQ Violations	CBQ Errors	CBQ Violations + Errors	Precision
Age	1.000	−0.223	−0.214	−0.250	−0.017
CBQ violations		1.000	0.477	0.769	0.096
CBQ errors			1.000	0.928	0.004
CBQ violations + errors				1.000	0.043
Precision					1.000

presents the Pearson correlation coefficients between key variables related to cyclist characteristics and risk perception accuracy. Notably, age shows a weak negative correlation with all behavioral variables measured by the Cyclist Behavior Questionnaire (CBQ), including violations (r = −0.223), errors (r = −0.214), and CBQ total (violations + errors) score (r = −0.250). This suggests that younger cyclists tend to report riskier or error-prone behaviors, in line with prior research highlighting age-related differences in traffic behavior.

As expected, violations and errors show a moderate positive correlation (r = 0.477), indicating that these dimensions often co-occur. Obviously, both are also strongly associated with the CBQ total (violations + errors) score, as the CBQ total variable aggregates these subscales (r = 0.769 with violations and r = 0.928 with errors).

Interestingly, the correlation between risk perception accuracy (precision) and all other variables is very weak, particularly with age (r = −0.017) and CBQ (violations + errors) score (r = 0.043), suggesting that behavioral tendencies and age are not strong predictors of risk perception accuracy in this sample.

Overall, the correlation matrix confirms that while behavioral tendencies are interrelated (e.g., errors and violations), their influence on actual risk perception accuracy appears limited. This highlights the complexity of risk assessment processes among cyclists and suggests that other factors, such as cognitive style, training, or contextual awareness, might play a more significant role.

3.7. Multiple Linear Regression Analysis

To explore the extent to which cyclist characteristics explain variability in risk perception accuracy, a multiple linear regression was conducted using the precision as the dependent variable. Predictor variables included age, cycling frequency, bike use purpose, years of cycling experience, and the CBQ total violation and error score.

The regression model was not statistically significant overall (F(11, 141) = 0.897, p = 0.545), indicating that the included variables do not significantly explain variance in risk perception precision. The R-squared value of 0.065 suggests that only 6.5% of the variability in precision scores is accounted for by the model. Moreover, the adjusted R-squared is negative (−0.0075), reflecting a poor model fit and likely overfitting due to the number of predictors relative to the sample size.

Table 6. Coefficients of the multiple linear regression analysis.

Predictor	Estimate	Std. Error	t Value	p-Value *
(Intercept)	0.995	0.110	9.01	<0.001
Age	−0.0004	0.0017	−0.21	0.840
Frequency: Few times/month	0.059	0.062	0.94	0.350
Frequency: Once/week	−0.025	0.059	−0.42	0.670
Frequency: Several/week	0.072	0.047	1.54	0.130
Purpose: Study	−0.062	0.213	−0.29	0.770
Purpose: Other	0.055	0.083	0.66	0.510
Purpose: Work	0.027	0.054	0.49	0.620
Purpose: Recreation	0.048	0.045	1.07	0.290
Experience: 1–3 years	−0.063	0.089	−0.70	0.480
Experience: >3 years	0.006	0.083	0.08	0.940
CBQ Violations + Errors	0.0009	0.0018	0.51	0.610

* Not significant (p > 0.05).

presents all the coefficients from the multiple linear regression analysis.

None of the predictors in the model reached statistical significance, including age, behavioral tendencies (CBQ scores), or experience-related variables. Although the coefficient for “cycling several times per week” approached significance (p = 0.13), it was still insufficient to draw conclusive inferences. This supports the findings from the correlation matrix and descriptive analyses, reinforcing the idea that risk perception accuracy among cyclists is not strongly influenced by demographic or behavioral variables alone.

These results may point to the influence of unmeasured cognitive or contextual factors (e.g., attentional focus, prior training, visual processing), which could be explored in future studies using qualitative or experimental designs.

3.8. Decision Tree Analysis

The decision tree analysis was conducted on a dataset of 153 cyclists from Loja to identify factors that influence the accuracy of risk perception in cycling contexts. The target variable was defined as high precision in risk perception, with 78 participants (51%) classified as having high precision and 75 participants (49%) showing low precision. The analysis utilized seven predictor variables including demographic characteristics (age, years of experience), cycling patterns (frequency and purpose of use), and behavioral measures from the Cycling Behavior Questionnaire (CBQ).

3.8.1. Model Performance and Validation

The developed decision tree model demonstrated moderate predictive performance across validation metrics (

Table 7. Decision tree model performance metrics.

Metric	Training Set	Test Set
Accuracy	75.93%	60.00%
Sensitivity (Recall)	-	47.62%
Specificity	-	70.83%
Precision	-	58.82%
F1-Score	-	52.63%
AUC	-	0.6002

). The model achieved an accuracy of 75.93% on the training set, which decreased to 60% on the independent test set, indicating some degree of overfitting. This decision tree model should be interpreted as exploratory, with moderate predictive power. While the model highlights the importance of CBQ behavioral scores in predicting perception accuracy, its predictive performance is limited and results should be treated with caution. The Area Under the Curve (AUC) value of 0.6002 (95% CI: 0.4328–0.7676) suggests fair discriminative ability, falling within the acceptable range for behavioral prediction models. In comparison, the Random Forest model yielded a lower AUC of 0.42, confirming that the decision tree outperformed it in terms of predictive performance.

The model showed higher specificity (70.83%) than sensitivity (47.62%), indicating better performance in correctly identifying participants with low precision compared to those with high precision. This pattern suggests that the model is more conservative in predicting high-precision cases, which may be appropriate given the safety implications of risk perception accuracy in cycling.

3.8.2. Variable Importance and Predictive Factors

The analysis of variable importance revealed that behavioral factors measured by the CBQ dominated the predictive model (

Table 8. Variable importance in risk perception prediction.

Variable	Importance Score	Relative Importance (%)
CBQ Violations + Errors	8.97	28.75
CBQ Errors	6.38	20.45
CBQ Violations	5.99	19.20
Age	4.06	13.00
Years of Experience	2.47	7.93
Frequency of Use	2.31	7.39
Type of Bike Use	1.02	3.28

). The combined CBQ violation and error score emerged as the most important predictor, accounting for 28.75% of the model’s predictive power. Individual CBQ components (errors and violations) contributed an additional 39.65% combined, resulting in behavioral measures explaining nearly 68% of the total variable importance.

Age represented the most significant demographic factor with 13% importance, while cycling patterns (frequency and type of use) and experience contributed relatively less to the model’s predictive capacity. This distribution suggests that current behavioral tendencies are more indicative of risk perception accuracy than cycling history or usage patterns.

3.8.3. Decision Tree Structure and Rules

The optimized decision tree generated seven terminal nodes across four hierarchical levels, with cycling frequency serving as the primary splitting criterion (see Figure 2). The tree structure revealed complex interactions between demographic, behavioral, and usage factors in determining risk perception accuracy.

The primary division separated participants based on cycling frequency, with those cycling “several times per week” (n = 52) showing a tendency toward lower precision (59.6%) compared to other frequency groups (n = 56) who demonstrated higher precision rates (58.9%). This counterintuitive finding challenges the assumption that frequent exposure necessarily improves risk assessment capabilities.

3.8.4. Critical Decision Pathways

The decision tree analysis identified several critical pathways that characterize different risk perception profiles. Among high-frequency cyclists (several times per week), age emerged as a crucial secondary factor. Participants aged 21.5 years or older within this group showed further stratification based on their CBQ violation and error scores. Those with combined CBQ scores of 9.5 or higher demonstrated the poorest risk perception accuracy, with 86.96% classified as low precision (n = 23). This finding suggests that frequent cycling combined with poor behavioral patterns and older age creates a particularly problematic profile for risk perception accuracy.

Conversely, younger high-frequency cyclists (under 21.5 years) showed a markedly different pattern, with 66.67% demonstrating high precision (n = 12). This age-related difference within the frequent cycling group suggests that younger cyclists may maintain better risk awareness despite high exposure levels, possibly due to heightened caution or different risk assessment strategies.

For cyclists with lower frequencies (daily, weekly, or monthly), the decision tree identified error rates as the primary discriminating factor. Participants with CBQ error scores of 7.5 or higher who had more than three years of experience showed exceptional risk perception accuracy, with 88.24% classified as high precision (n = 17). This pathway represents the most accurate risk perception group in the entire dataset, suggesting that the combination of experience with moderate error rates in less frequent cyclists creates optimal conditions for accurate risk assessment.

3.8.5. Complex Relationships Between Experience and Risk Perception

The analysis revealed a complex relationship between cycling frequency and risk perception accuracy that warrants further investigation. While frequent cycling is traditionally associated with improved expertise, our exploratory findings suggest this relationship may be more nuanced, with some frequent cyclists showing decreased accuracy in specific contexts. However, this pattern should be interpreted cautiously given our modest sample size and cross-sectional design, which cannot establish causal relationships or rule out confounding variables.

The interaction between frequency, age, and behavioral measures provides additional insight into this paradox. Older frequent cyclists with elevated violation and error scores represent a high-risk group for poor risk perception, potentially due to the development of risky behavioral patterns reinforced by frequent riding without negative consequences. This habituation effect may be particularly pronounced in older cyclists who have developed established riding patterns over time.

3.8.6. ROC Analysis and Model Discrimination

The receiver operating characteristic (ROC) analysis provided additional validation of the model’s discriminative capacity (Figure 3). The AUC value of 0.6002 indicates fair performance, with the model performing better than random chance but with substantial room for improvement. The ROC curve shape suggests that the model achieves reasonable specificity at moderate sensitivity levels, consistent with the conservative prediction pattern observed in the confusion matrix analysis.

The 95% confidence interval for the AUC (0.4328–0.7676) spans from poor to good discrimination, indicating uncertainty in the model’s true discriminative ability. This wide interval reflects the moderate sample size and suggests that larger studies would be needed to establish more precise estimates of model performance.

3.8.7. Confusion Matrix Analysis

The confusion matrix for the test set revealed specific patterns in model predictions that inform the practical application of these findings. The model showed a tendency toward conservative prediction, with higher accuracy in identifying low-precision cases (specificity = 70.83%) compared to high-precision cases (sensitivity = 47.62%). This pattern (see Figure 4) resulted in 11 false negatives (high-precision cases incorrectly classified as low precision) and 7 false positives (low-precision cases incorrectly classified as high precision).

The asymmetric error pattern has important implications for practical applications. The model’s conservative approach may be appropriate for safety-related applications where failing to identify individuals with poor risk perception (false negatives) could have more serious consequences than incorrectly flagging those with good risk perception (false positives).

3.8.8. Distribution of Risk Perception Precision

The distribution of precision scores across the study population revealed important patterns in risk perception accuracy (Figure 5). The distribution showed a slight positive skew, with most participants clustering around moderate precision levels and fewer participants achieving very high or very low precision scores. This distribution pattern supports the binary classification approach used in the decision tree analysis while highlighting the continuous nature of risk perception accuracy.

The median precision score of 1.001 closely approximated a precision of 1.0, indicating that the majority of participants demonstrated reasonable risk perception accuracy. However, the interquartile range (0.9167–1.167) showed considerable variability, justifying the investigation of factors that distinguish high- and low-precision individuals.

3.8.9. Conclusions of Decision Tree Analysis

The decision tree analysis successfully identified key factors influencing risk perception accuracy among cyclists in Loja, with behavioral measures emerging as the strongest predictors. The finding that CBQ-based behavioral factors account for nearly 68% of the model’s predictive power underscores the importance of current behavioral patterns over demographic or experiential factors in determining risk perception accuracy.

The counterintuitive relationship between cycling frequency and risk perception accuracy represents a significant finding with important implications for cycling safety interventions. The identification of frequent cyclists with poor behavioral patterns as a high-risk group for inaccurate risk perception challenges traditional approaches to cycling safety education and suggests the need for targeted interventions focused on this population.

The age-related differences within frequent cyclists further illuminate the complexity of risk perception development. The superior performance of younger frequent cyclists compared to their older counterparts suggests different developmental trajectories for risk assessment capabilities and highlights the importance of age-specific approaches to cycling safety education.

The moderate predictive performance of the model (AUC = 0.6002) indicates that while the identified factors are relevant, additional unmeasured variables likely influence risk perception accuracy. Future research should consider incorporating factors such as traffic exposure characteristics, specific cycling infrastructure usage patterns, and psychological factors related to risk tolerance and perception.

These findings provide a foundation for developing targeted interventions to improve risk perception accuracy among cyclists, with particular attention to the behavioral patterns identified in frequent cyclists and the need for age-appropriate approaches to risk assessment training.

3.9. Profile of Most Accurate Risk Perceivers

To better understand which cyclist profiles demonstrate the most accurate risk perception, we analyzed the upper quartile of precision scores (i.e., the 25% of cyclists with the lowest mean absolute deviation from actual risk). This group represents the most perceptive or cautious individuals in the sample (n ≈ 38).

The key characteristics are as follows:

• Average Age: 30.3 years: This group tends to be relatively young, but not in the adolescent range. Their age may reflect a balance between experience and cognitive awareness of risk without overconfidence.
• Average CBQ Violation + Error Score: 13: This is a moderately low score on the Cyclist Behavior Questionnaire, suggesting that lower engagement in risky or error-prone behaviors is associated with better risk perception.

In terms of riding habits, the majority of these high-precision cyclists were frequent riders. Specifically, 16 reported cycling several times per week, 14 cycled daily, 8 cycled once per week, and only 3 cycled a few times per month. This distribution suggests that frequent exposure to cycling environments may help develop more refined risk perception skills. Notably, although cycling frequency did not emerge as a significant predictor in the overall regression model, it appears to be a common characteristic among the most accurate cyclists. This indicates that while frequency alone may not predict risk perception precision across the entire population, it is prevalent in those who demonstrate the highest accuracy.

In conclusion, the most accurate cyclists tend to have the following characteristics:

• In their early 30s;
• Moderately cautious in behavior (as per CBQ scores);
• Frequent users of cycling infrastructure.

This profile aligns with the patterns observed in the decision tree, where lower CBQ scores and regular cycling—especially for exercise purposes—were associated with higher precision. While not all frequent cyclists are accurate, those who are tend to pair frequency with responsible riding habits.

These findings support the idea that targeted interventions promoting safe behavior among frequent cyclists could further enhance their risk perception, while encouraging occasional cyclists to develop risk awareness skills through structured training or experience.

3.10. Effect of Route Familiarity on Risk Perception Accuracy

To examine whether the frequency with which cyclists reported using specific locations influenced the accuracy of their risk perception relative to objective assessments, a comparative analysis was conducted between self-reported frequency of use and the risk levels assigned by the CycleRAP tool. Each participant evaluated 12 photographs depicting various cycling environments in the city of Loja, assigning to each a perceived risk level (low, medium, high, extreme). These were compared against the corresponding CycleRAP-assigned risk levels, which were fixed for each image. Additionally, participants indicated how often they cycled through each photographed location, using a five-point Likert scale ranging from Never to Very frequently.

To facilitate the analysis, responses were grouped into two categories based on frequency: high frequency (including responses of Frequently and Very frequently) and low frequency (including Sometimes, Rarely, and Never). Risk perception accuracy was operationalized as the absolute difference between the perceived risk and the CycleRAP reference risk for each photograph. A total of 1836 observations (153 participants × 12 photographs) were analyzed.

Table 9. Descriptive statistics and Mann–Whitney U test results for risk difference between high- and low-frequency users.

Frequency Group	n	Mean Risk Difference *	SD	Median	Mann–Whitney W	p-Value	Effect Size r
High frequency	810	1.144	0.921	1	440,904	0.017	0.056
Low frequency	1026	1.038	0.872	1	—	—	—

* Absolute difference between perceived and actual (CycleRAP) risk.

summarizes the descriptive statistics and comparative test results between the two frequency groups. The average risk difference for high-frequency users was 1.144, slightly higher than the 1.038 observed among low-frequency users. A Mann–Whitney U test indicated that this difference was statistically significant (W = 440,904, p = 0.017); however, the effect size was negligible (r = 0.056), suggesting limited practical relevance.

Normality tests (Shapiro–Wilk) confirmed that the distributions were non-normal in both groups (p < 0.001), and Levene’s test for homogeneity of variance was marginally significant (p = 0.048), supporting the use of non-parametric testing.

To explore whether differences in accuracy were consistent across the 12 photographs, a photograph-by-photograph analysis was also performed. As shown in

Table 10. Risk perception accuracy by photograph of the site and frequency group.

Photo/Site	Mean Difference (High Freq.)	Mean Difference (Low Freq.)	p-Value
01	0.604	0.610	0.864
02	0.944	0.727	0.101
03	0.970	1.080	0.351
04	1.740	1.500	0.089
05	0.593	0.653	0.323
06	1.080	0.953	0.592
07	1.470	1.660	0.196
08	1.740	1.930	0.195
09	1.850	1.860	0.949
10	0.754	0.682	0.518
11	0.746	0.523	0.0097
12	0.921	0.600	0.0021

, no significant differences were found in 10 of the 12 images. Only photographs 11 and 12, both classified as Moderate risk by CycleRAP, showed significantly greater accuracy among low-frequency users (p = 0.0097 and p = 0.0021, respectively).

Visual summaries of the distribution of risk differences across groups and photographs are provided in Figure 6 (boxplot of risk difference by frequency group) and Figure 7 (average accuracy per photograph). Both figures show large overlaps between distributions, consistent with the small observed effect sizes.

In summary, while a statistically significant difference was found in aggregate accuracy between high- and low-frequency users, the direction of the effect was contrary to expectations. Low-frequency cyclists—those less familiar with the specific road environments—tended to estimate risk levels slightly more in line with CycleRAP’s objective assessments. Nevertheless, the practical significance of this result is minimal, and the overall pattern does not suggest that frequent exposure improves risk perception accuracy among cyclists in Loja.

3.11. Perceptions of Cycling Infrastructure and Safety Improvement Priorities

In addition to analyzing perceived and objective risk at specific urban sites, participants were asked to provide a general evaluation of the city’s cycling infrastructure and to identify key areas for improvement to enhance cyclist safety.

Regarding the overall quality of cycling infrastructure in Loja, responses revealed a predominantly critical perception. Among the 153 valid responses, 53.6% rated the infrastructure as “Regular”, 29.4% as “Poor” or “Very poor”, and only 16.9% as “Good” or “Very good.” These results suggest that despite the presence of some cycling facilities, the existing infrastructure is widely perceived as insufficient, inconsistent, or inadequately maintained. The scarcity of “Very good” evaluations further reflects a lack of cyclist confidence in the current urban network.

To better understand cyclists’ priorities for improving safety, participants selected one or more factors they considered most important. The most frequently cited measure was “driver education on road safety”, selected by over 90% of respondents, indicating a strong perceived need for behavioral change among motorists rather than solely infrastructural fixes. This aligns with international literature emphasizing the role of driver behavior in cyclist vulnerability. Closely following in importance were “physical separation of bike lanes”, “better signage”, and “greater network continuity”, all of which were selected by the majority of participants. These responses underscore the demand for consistent, well-demarcated, and connected infrastructure that provides both spatial protection and clear communication of rights and expectations.

Other commonly cited measures included “road maintenance”, “traffic speed reduction”, and “improved lighting”, highlighting the multifaceted nature of cyclist safety needs. Participants emphasized that a safe cycling environment requires more than designated lanes; it must also be well-maintained, visible, and integrated into a broader strategy of urban traffic calming and accessibility.

Taken together, these findings suggest that Loja’s urban cyclists perceive their environment as moderately to poorly equipped for safe cycling and prioritize a comprehensive combination of educational and infrastructural interventions. Importantly, the emphasis on behavioral education for drivers indicates that perceived risk is not only tied to the physical design of the streets but also to social interactions and cultural norms in traffic dynamics.

4. Discussion

4.1. Misalignment Between Perceived and Objective Risk

This study examined the accuracy of risk perception among urban cyclists in Loja, Ecuador, by comparing subjective evaluations with objective risk ratings derived from the CycleRAP methodology. Results reveal a consistent misalignment between perceived and actual risk levels: participants underestimated high- and extreme-risk environments while overestimating low-risk scenarios. This pattern indicates a disconnect between cognitive evaluations and the real dangers posed by the infrastructure, a phenomenon observed in other urban cycling contexts [6,30].

Multiple factors likely contribute to this divergence. Roadway design plays a central role, as cyclists often perceive major streets with shared lanes as highly dangerous, while separated or multi-use paths are viewed as safer [6]. Personal experiences, especially frequent near-misses, heighten perceived risk even without resulting in injury [30]. Additionally, environmental cues, such as bus stops, poor surface conditions, or adverse weather, can distort hazard perception [31,32]. Finally, social and behavioral influences, including the behavior of other road users and societal attitudes toward cycling, shape risk perception beyond individual experience [30]. Recognizing the multifactorial nature of perceived risk is critical for promoting safer and more appealing cycling environments.

4.2. Behavioral and Demographic Determinants

Our findings reveal a complex relationship between cycling frequency and risk perception accuracy, challenging simple assumptions about experience and expertise. Contrary to initial expectations, cycling frequency alone did not predict higher perception accuracy. Instead, frequent cyclists with elevated Cycling Behavior Questionnaire (CBQ) violation and error scores showed the lowest accuracy. This suggests that while greater exposure might foster overconfidence or habituation, reducing attentional sensitivity to hazards—a concept supported by cognitive psychology—it does not necessarily improve risk assessment. This pattern is particularly concerning in cities like Loja, where underdeveloped infrastructure means frequent exposure without reinforced safe practices could increase vulnerability. Our results align with evidence suggesting frequent riders may face more near misses, engage in traffic violations, and have elevated crash risk [33,34], even with potentially higher traffic literacy over time [33]. However, as our data is cross-sectional, we cannot definitively determine if frequent cycling causes reduced accuracy or if individuals with poorer inherent risk perception are more prone to frequent cycling.

Behavioral tendencies emerged as the strongest predictors of perception accuracy. In the decision tree model, CBQ scores accounted for ~68% of predictive power, surpassing demographic variables. Younger cyclists, although often riskier, demonstrated moderate accuracy when their behaviors were self-regulated, whereas older cyclists tended to report fewer risky behaviors but did not consistently achieve higher accuracy, reflecting complex interactions between age, self-assessment, and exposure to aggressive encounters [35,36].

Regression analyses confirmed that demographic variables alone explained little variance, consistent with prior research emphasizing the role of psychosocial and behavioral factors in risky cycling [37,38,39,40,41,42,43]. Models incorporating variables such as perceived behavioral control, traffic knowledge, and safety attitudes have explained up to 37% of variance in risky behaviors [42]. These findings underscore that interdisciplinary frameworks integrating psychological, behavioral, and contextual factors are required to better model cyclist risk perception.

The decision tree analysis provided exploratory insights into factors potentially associated with risk perception accuracy among cyclists in Loja, though the modest predictive performance (AUC = 0.60) indicates substantial unexplained variance. While behavioral measures showed the strongest relative importance in the model, the overall discriminative ability remains limited. These findings should be interpreted as preliminary evidence requiring replication in larger, more diverse samples rather than definitive conclusions about risk perception determinants.

4.3. Role of Cycling Frequency and Site Familiarity

A slight but statistically significant difference between high- and low-frequency cyclists indicated that familiarity with a site does not guarantee better hazard recognition. In fact, low-frequency users sometimes demonstrated more accurate risk estimations, likely due to greater situational alertness in unfamiliar environments. This challenges the assumption that repeated exposure improves hazard detection. This behavior is possibly due to heightened alertness in unfamiliar environments. Although the effect was weak, this phenomenon warrants further investigation.

Evidence suggests that targeted hazard perception training, rather than mere familiarity, is most effective for improving situational awareness. PC- or game-based modules and immersive tools such as virtual or augmented reality enhance hazard recognition without depending on repeated site-specific exposure [10,22,44,45,46,47,48]. Therefore, safety interventions should prioritize structured hazard training over relying solely on experiential familiarity.

4.4. Implications for Urban Cycling Safety

Cyclists in Loja expressed predominantly negative perceptions of the city’s cycling infrastructure, citing insufficient physical separation, poor maintenance, and limited connectivity. The literature consistently highlights that well-marked, physically separated lanes and traffic calming measures (e.g., lower speed limits, removal of parked cars) are critical for vulnerable users such as women, older adults, and those riding with children [49,50,51].

Participants also emphasized that infrastructure improvements alone are insufficient. Driver education and awareness campaigns are crucial, as social interactions on the road strongly influence perceived and actual safety. These insights align with global findings that integrated interventions combining infrastructure upgrades, traffic management, and behavioral change programs are most effective for promoting safe and inclusive cycling [52,53].

4.5. Study Limitations

This study presents several limitations that should be considered when interpreting its findings. First, the use of a non-probabilistic convenience sampling method limits the generalizability of the results beyond the specific population of urban cyclists in Loja, Ecuador. The sample was also predominantly male, which may not fully capture gender-based differences in risk perception and behavior. Second, the assessment of perceived risk was based on static photographs, which, although useful for standardized comparisons, do not fully replicate the dynamic and multisensory nature of real-world cycling environments. Third, the accuracy of objective risk classification relied on the CycleRAP methodology, which, while standardized, involves assumptions and estimations (e.g., traffic volumes from Google Traffic and inferred pedestrian density) that may introduce measurement bias. This study focused on demographic and behavioral factors but did not include psychological variables such as impulsivity, attention allocation, or risk tolerance, which may influence risk perception. Future research integrating cognitive and psychological measures could better explain individual differences in perception accuracy. Furthermore, a significant limitation was the unavailability of disaggregated crash data for cyclists in Loja, preventing direct validation of our risk findings against empirical incident occurrences. Finally, the modest sample size and limited statistical power of the regression model suggest that further research with larger and more diverse samples, combined with experimental or longitudinal designs, is needed to more robustly explore the complex interplay between behavior, experience, and risk perception in urban cycling.

Despite its limitations, this study offers important contributions to the growing body of research on urban cycling safety, particularly in low- and middle-income contexts where empirical data are scarce. By integrating behavioral data, subjective risk assessments, and objective infrastructure evaluations through the CycleRAP methodology, the study provides a multidimensional understanding of risk perception accuracy among cyclists. The identification of behavioral patterns—such as the paradoxical role of cycling frequency and the predictive power of CBQ scores—offers valuable insights for designing targeted safety interventions. Furthermore, the mixed-methods approach employed here can serve as a replicable framework for evaluating cyclist safety in other cities undergoing mobility transitions.

4.6. Future Research

Future research should expand on these findings by incorporating psychological and cognitive variables, such as attention, impulsivity, or perceived control, as well as by employing immersive or experimental methods (e.g., virtual reality simulations) to better capture real-time risk perception. Future longitudinal research is essential to untangle these relationships and understand causal mechanisms, especially considering unmeasured factors like route selection, specific traffic exposure patterns, or individual learning differences that might explain these associations. Ultimately, advancing this line of research can support the development of evidence-based policies and educational programs that promote safer and more inclusive cycling environments.

5. Conclusions

This study contributes to the growing body of research on urban cycling safety by providing an integrative analysis of risk perception among cyclists in a medium-sized city in a low-to-middle-income country. The findings demonstrate that perceived risk often fails to align with objectively assessed infrastructure conditions, particularly in high-risk environments that are frequently underestimated by cyclists. Behavioral tendencies, as measured by the CBQ, emerged as the most consistent predictors of poor risk perception accuracy, suggesting that current riding habits may be more relevant than demographic or experiential factors. The data also reveal a behavioral paradox, in which frequent cycling combined with high-risk behaviors leads to lower accuracy, contrary to traditional assumptions about the benefits of experience. While the statistical models showed limited predictive power, decision tree analysis offered valuable insights into cyclist profiles with distinct patterns of perception accuracy. The identification of younger, frequent cyclists with moderate CBQ scores as more accurate perceivers points to the value of promoting responsible riding behaviors over sheer experience. Moreover, the study highlights a general dissatisfaction with cycling infrastructure in Loja and an urgent demand for interventions that combine physical improvements with education and behavioral change. Risk perception accuracy is a critical factor in cyclist safety and deserves attention in both policy and urban planning. With only 12 sites and 153 cyclists, our study has a relatively small sample size. This means that any analyses of specific subgroups might not have enough statistical power, and thus, our findings—especially those based on a limited number of observations within those subgroups—should be interpreted with caution. Future research should explore psychological and contextual dimensions of risk perception and consider experimental or qualitative approaches to deepen our understanding of how cyclists assess danger in complex urban environments.