This study employed a cross-sectional online survey among university motorcycle riders in Northern Thailand (Chiang Mai province). Participants were recruited via convenience sampling from verified Facebook student groups that required student ID verification for membership. Eligible participants were full-time university students aged 20–24 years who used motorcycles as their primary mode of transport in Chiang Mai. Those engaged in professional riding activities (e.g., ride-hailing or delivery services) were excluded. A total of 318 students accessed the survey.

The study protocol was approved by the Research Ethics Committee, Faculty of Medicine, Chiang Mai University (protocol code: PSY-2567-0577). Best-practice guidance for reporting web surveys (e.g., the Checklist for Reporting Results of Internet E-Surveys [CHERRIES]) (33) was followed to ensure methodological transparency and data quality. The Thai-language questionnaire was hosted on Microsoft Forms and optimized for mobile devices. The anonymous online format, voluntary participation, and assurance of confidentiality were expected to minimize social desirability bias when reporting sensitive behaviors such as risky riding. A pretest was conducted with 38 university students (not included in the final sample) to assess item clarity, cultural relevance, and technical functionality.

Recruitment posts, including a standardized invitation and flyer with a survey link and QR code, were shared in Facebook groups of universities and student communities in Chiang Mai during November–December 2024. The survey landing page displayed the participant information sheet and electronic informed consent; only those who provided consent could proceed. Upon verified completion, participants received 50 THB (≈1.50 USD). Compensation records were kept separate from survey data to preserve confidentiality. All data were securely stored on password-protected university servers, with access restricted to the study team.

Because recruitment occurred in university social-media groups, it was not possible to determine how many individuals viewed the invitation posts—a typical limitation of online recruitment. Details of data-quality screening procedures are provided in Section 2.4.

Of the 318 visitors who accessed the online survey, 284 provided electronic consent and completed the questionnaire. Data-quality screening followed prespecified criteria to enhance transparency and reproducibility: (a) completion time—responses submitted in less than approximately one third of the pilot median duration (median ≈17 min; < 6 min; 16 cases). This cut-off was derived from the pretest completion-time distribution, consistent with CHERRIES (33) requirements to justify exclusion criteria for atypically fast submissions; (b) response patterns indicating straight-lining or logical inconsistencies across demographic and behavioral items (14 cases); and (c) duplicate entries identified through submission timestamps, anonymized IP patterns, and compensation records, with no personally identifiable data retained (4 cases).

All variables were subsequently inspected for missing data, normality, and outliers prior to analysis. Three cases exceeding ±3.0 SD on the risky riding composite were excluded as univariate outliers. Missing values within multi-item scales were handled by computing the mean of available items when at least 80% of items were completed, following established psychometric guidelines. Missingness was minimal (<2%) at the scale level; therefore, no additional imputation was performed. After all exclusions, the final analytic sample comprised 247 participants.

Sample size adequacy was evaluated with reference to established guidelines for multiple regression analyses. Green (42) proposed practical rules-of-thumb suggesting a minimum sample size of N ≥ 50 + 8 m for testing overall model fit and N ≥ 104 + m for testing individual predictors, where m represents the number of predictors included in the model. In the present study, 11 predictors were entered into the regression models (including the interaction term), yielding minimum recommended sample sizes of N = 138 and N = 115, respectively. The final sample size (N = 247) exceeded both thresholds, supporting the adequacy of the sample for the regression analyses conducted.

All analyses were conducted using IBM SPSS Statistics (Version 27). Descriptive statistics were calculated for all study variables. For multi-item scales, mean (M), standard deviation (SD), and internal consistency (Cronbach’s α) were reported. For categorical variables, frequency and percentage (n/%) were presented. Bivariate associations among continuous variables were examined using Pearson’s r. For all constructs measured using multiple items, total scores were calculated by summing item responses, with higher scores indicating higher levels of the corresponding construct. Prior to correlation analyses, internal consistency was assessed for each scale, and all composite measures demonstrated acceptable reliability. The psychometric properties of the adapted RLBs scale were further examined using exploratory factor analysis, as described above. Correlation analyses were then conducted using these summed total scores.

Hierarchical linear regression was employed to test whether BPD symptoms moderated the association between RLBs and risky riding. Standardized coefficients (β), t-values, and change statistics (ΔR², ΔF) were reported. Simple-slopes analyses and the interaction plot were generated using the PROCESS macro for SPSS (Model 1) based on the same regression outputs. Statistical significance was set at p < 0.05 (two-tailed). As part of regression diagnostics, key model assumptions were examined. Visual inspection of residual plots suggested no major deviations from linearity or homoscedasticity, and residuals appeared approximately normally distributed. No influential outliers were identified based on standardized residuals and leverage values.

As shown in Table 1, the analytic sample comprised 247 university motorcycle riders (M = 21.15 years, SD = 1.21), the majority of whom were female (72.9%) and enrolled in bachelor’s degree programs (96.0%). Riding exposure was substantial, with over half (52.2%) reporting daily riding and another 28.7% riding almost every day. Most participants had at least 1 year of riding experience; 37.7% reported 5–10 years. Motorcycles with engine capacities of 125 cc or lower were most common (72.5%). A large majority (81.4%) held a valid driving license, and 70.4% reported alcohol use. Regarding mental health, 9.7% of participants reported a history of diagnosed mental disorders.

Table 2 summarizes the descriptive statistics of the key study variables. Participants reported moderate levels of risky riding (M = 42.55, SD = 11.59), with positive skewness suggesting that a subgroup of riders engaged in relatively high levels of risky behavior. Risk-legitimizing beliefs (RLBs) were common (M = 33.34, SD = 8.26), and BPD symptoms were present at sub-clinical but meaningful levels (M = 5.14, SD = 3.14). Perceived peers’ risky riding was generally low (M = 9.87, SD = 5.90) but highly skewed, reflecting that a minority perceived their peers as highly risky. Approximately 23.5% of participants scored above the SI-Bord cut-off (>7), indicating clinically relevant levels of BPD symptoms.

Table 3 presents the mean endorsement of each RLBs item, ranked in descending order of mean value. The most strongly endorsed beliefs were “Even when I’m very tired, I ride my motorcycle to avoid being late for classes or important events” (M = 3.78, SD = 1.16) and “Riding fast allows me to reach my destination more quickly” (M = 3.45, SD = 1.21), reflecting functional rationalizations that emphasize short-term benefits. Other items showed comparatively lower endorsement levels. These patterns suggest that interventions should prioritize addressing time-saving and control-related justifications, which appear most salient among university riders.

A summary of item-level risky riding behaviors is shown in Table 4. Phone- or earphone-related distractions (Item 1) and fatigue-related behaviors such as riding long periods without breaks (Item 2) had the highest mean frequencies, indicating that everyday inattention and physical strain were the most common sources of risky riding in this sample. Moderate frequencies were observed for routine traffic violations—including late-night speeding, failing to stop at zebra crossings, and riding too close to other vehicles (Items 3–5)—suggesting that rule-breaking behaviors were present but less frequent than distraction- or fatigue-related behaviors. Deliberate high-risk actions such as wheel spins or lifting the front wheel (Items 25–26) were reported least often, indicating that intentional stunt-like behaviors were relatively rare among university riders. These patterns suggest that the risky riding observed in this sample is driven primarily by common, everyday behaviors rather than extreme forms of risk-taking.

Correlations among all study variables are presented in Table 5. Risky riding behaviors were positively correlated with RLBs (r = 0.395, p < 0.01), BPD symptoms (r = 0.330, p < 0.01), and perceived peers’ risky riding (r = 0.416, p < 0.01). These associations indicate that students who more strongly endorsed RLBs or reported higher BPD symptoms tended to engage in more risky riding behaviors. In addition, RLBs were significantly correlated with perceived peers’ risky riding (r = 0.319, p < 0.01) but were not significantly related to BPD symptoms (r = 0.061, p > 0.05). All correlations were in the expected directions and of moderate magnitude, suggesting minimal multicollinearity among predictors for subsequent regression analyses.

The moderation model was tested using mean-centered scores for RLBs (X) and BPD symptoms (W), with their interaction term (X × W) entered in the second step of hierarchical regression. The detailed standardized coefficients and confidence intervals for each model are presented in Table 6, and the overall model fit indices are summarized in Table 7.

In Model 1, both RLBs (β = 0.376, p < 0.001) and BPD symptoms (β = 0.307, p < 0.001) were significantly and positively associated with risky riding behaviors, accounting for 25.0% of the variance, supporting H1 and H2.

Model 2 added the interaction term, which was statistically significant (β = 0.190, p < 0.001) and amounted to a small-to-moderate effect size, in support of H3. The inclusion of this interaction increased the explained variance to 28.5% (ΔR² = 0.036, f² = 0.049).

Model 3 further adjusted for gender, age, riding frequency, riding experience, motorcycle size, driving license status, drinking status, history of mental disorders, and perceived peers’ risky riding. The interaction effect between RLBs and BPD symptoms remained significant (β = 0.212, p < 0.001). The full model explained 38.1% of the variance in risky riding behaviors (adjusted R² = 0.381).

Multicollinearity diagnostics indicated no violation of assumptions, with all tolerance values above 0.70 and variance inflation factors (VIFs) below 1.50 (see Table 8 for detailed coefficients and collinearity statistics).

Simple slopes analysis (Figure 2) indicated that the association between RLBs and risky riding was negligible at low levels of BPD symptoms but became progressively stronger at moderate to high symptom levels. When BPD symptoms were low (2 points), the slope was small and non-significant (β = 0.059, SE = 0.108, p = 0.585). At a moderate level of BPD symptoms (5 points), the slope was significant and positive (β = 0.331, SE = 0.078, p < 0.001). At a high level of BPD symptoms (8 points), the positive association between RLBs and risky riding behaviors was strongest (β = 0.603, SE = 0.097, p < 0.001). This pattern suggests that the practical relevance of the interaction lies not in shifting overall population risk, but in identifying conditions under which the association between RLBs and risky riding is more pronounced.

The observed moderation effect suggests that BPD symptoms may be associated with differences in how belief–behavior inconsistency is managed, rather than indicating a uniform mechanism across riders. Specifically, the stronger association between RLBs and risky riding at higher levels of BPD symptoms is consistent with the notion that psychological vulnerability may be linked to a greater reliance on belief-based strategies when dissonance is present. Importantly, this pattern should be interpreted as a conditional association rather than direct evidence of how dissonance is resolved. From a theoretical perspective, these findings help clarify how individual differences in emotional and personality-related vulnerability may shape the strength of the relationship between cognitive rationalizations and unsafe riding, thereby extending cognitive dissonance frameworks into a public mental health context without implying causal pathways.

From a public health perspective, the robust main effects observed in this study indicate that addressing RLBs remains broadly relevant for reducing risky motorcycle riding among university students. Belief-focused strategies—such as brief cognitive-behavioral or psychoeducational exercises directly challenging common rationalizations (e.g., “riding fast does not meaningfully save time”)—may therefore help reduce the cognitive justifications for unsafe behavior at the population level.

At the same time, the modest but significant moderation effect suggests that additional, targeted components may be particularly relevant for riders with elevated BPD symptoms. Incorporating brief mental health screening and embedding short emotion-regulation or mindfulness modules into road-safety education programs may help identify and support riders with elevated BPD symptoms, for whom belief–behavior associations appear stronger.

Finally, peer-based campaigns that reshape perceived norms—highlighting that most students actually value safe riding—may complement individual-level strategies by reducing the social reinforcement of RLBs identified in prior reviews. Taken together, these findings support a feasible, multi-level prevention approach that combines general belief-focused interventions with targeted support for psychologically vulnerable riders, rather than a one-size-fits-all strategy.

Several limitations should be acknowledged. Given the cross-sectional design, the observed associations should be interpreted as potentially bidirectional rather than causal. While RLBs may be associated with greater engagement in risky riding, repeated involvement in risky riding may also contribute to the development or reinforcement of such beliefs over time. Similarly, personality-related vulnerabilities such as BPD symptoms may both influence and be influenced by patterns of risky behavior. Longitudinal or experimental studies are therefore needed to disentangle these reciprocal pathways and clarify temporal ordering.

The use of self-report measures and convenience sampling from online university groups may limit representativeness and introduce response bias. As recruitment relied on Facebook student groups, some self-selection bias cannot be ruled out; however, membership verification and inclusion criteria minimized non-student participation. Future studies should complement online recruitment with on-campus sampling to enhance representativeness. Incorporating objective indicators—such as recorded traffic violations or GPS-based riding exposure (e.g., total distance, riding duration, or urban–rural route patterns)—would enhance measurement validity. As the sample consisted predominantly of female university riders in Northern Thailand, the findings should be interpreted with caution in terms of generalizability and external validity. Prior research has documented gender differences in risky riding behaviors and related psychological traits, suggesting that the observed associations may differ in strength or pattern in more gender-balanced or male-dominated samples. Accordingly, the present findings are more directly applicable to female university riders in this context. Future studies should aim to recruit more gender-diverse samples to examine the robustness of these findings across genders. Expanding beyond BPD symptoms to include broader impulsivity- and sensation–seeking–related traits could further clarify when and for whom belief-based mechanisms are most strongly associated with risky riding.

In addition, RLBs were assessed using an adapted scale transferred from health-risk domains such as smoking and drinking, and the present findings should therefore be interpreted as providing initial, exploratory evidence. Although preliminary psychometric analyses supported the use of a unidimensional composite in the current sample, further research is needed to confirm the dimensional structure and measurement invariance of the RLBs scale across populations and contexts.