Advisory systems tailored for the yellow light dilemmas have been implemented in connected vehicles (CVs) to enhance intersection safety. Evaluations of such systems have focused solely on drivers' stop/go decisions, neglecting the subsequent control responses (i.e., maximum decelerations or accelerations) that determine actual safety outcomes. Thus, the efficacy of CVs remains partially understood. To overcome this gap, this study develops a bivariate copula framework with unobserved heterogeneity for understanding stop/go decisions at signalized intersections. A C-V2X-enabled stop/go advisory system was developed and implemented in a high-fidelity co-simulation driving simulator combining Unity3D and VISSIM, with 62 licensed drivers completing the experiments under baseline and advisory conditions. The results indicate that the advisory system elicited high driver trust and reduced mental workload, with compliance high overall but varying by advisory type. Further, the copula model reveals significant dependence structure between decision and control behavior, whereby drivers who decide to run yellow lights exhibit lower maximum acceleration, while those who stop apply smooth braking. Further, the advisory system increases the probability of proceeding without inducing risky accelerations and reduces harsh deceleration when stopping. The model further uncovers the unobserved heterogeneities across driver characteristics and driving history. The findings present methodological and empirical insights that can support smoother intersection operations and inform safety-oriented advisory-system design in future connected traffic environments.
Forecasting crash risk at the network level over future horizons is fundamental to proactive safety management. Existing conflict-based extreme value theory (EVT) models, however, largely ignore spatial correlations in extreme traffic conflicts and remain restricted to one-step-ahead crash risk forecasts. This study addresses both limitations by proposing a score-driven peak-over-threshold (SPOT) model within a marked point process framework, combined with conditional autoregressive spatial priors in a Bayesian hierarchical structure to account for spatial dependence across the network. Probabilistic forecasts of two crash risk measures, Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR), are generated at multiple future horizons via posterior simulation of the marked point process. The framework is applied to an urban network of 50 intersections from the pNEUMA drone trajectory dataset in Athens, Greece. In-sample results reveal meaningful spatial correlation, and model comparison shows that covariate specification has a substantially larger impact on predictive performance than spatial prior choice. Out-of-sample, crash risk forecasts are most reliable within 20 s. Forecast quality degrades at longer horizons due to growing simulation errors and parameter reversion toward baseline levels, becoming inadequate beyond 60 s. The practical value of the framework is demonstrated through a forecast-based safest-route finding application, which yields routes that are on average 7% safer than those based on static crash risk estimates with comparable travel times.
Incident duration data are continuous with a large range, exhibit multi-modality and are likely to be bunched around certain time intervals. Developing a single model for the full distribution of incident duration will restrict the degrees of freedom available for each variable in the model (to 1). To relax this assumption, the current study proposes a hierarchical modeling framework for incident duration modeling. The proposed framework considers a hierarchical structure where the upper-level model categorizes incident duration into different categorical time groups including Very Small (>0–15 min), Small (>15–30 min), Medium (>30–120 min), Large (>120–240 min), and Very Large (>240 min) and the lower-level models develop separate propensities for each time group. Given the inherent ordered nature and known threshold for each category, we employ a joint grouped ordered response framework across the two levels for our analysis. The proposed model system is estimated using traffic incident data for the year 2019 from Seminole county, Florida. A host of independent variables including incident characteristics, roadway and traffic factors, road environmental and weather conditions, land-use, built environment and socio-demographic characteristics are considered for the analysis. Further, we compare the model performance of the proposed model with Ordinary Least Squares (OLS) model using a holdout sample. The comparison exercise highlights that the proposed model provides superior predictive accuracy, particularly for Very Small to Medium categories. For Large and Very Large duration categories, both models provide similar predictions.
Recent studies have applied conditional extreme value theory-based models to traffic conflict data for short-term crash risk forecasting. Although these models have shown promise, they generally focus on extreme events represented by a single conflict measure. However, extreme events captured by different conflict measures may exhibit cross-process dependence and thus contain additional information that is useful for short-term crash risk forecasting. To address this issue, this study develops a Bayesian cross-exciting conditional peak-over-threshold framework for short-term rear-end crash risk forecasting. Within this framework, modified time to collision (MTTC) is treated as the primary conflict measure, while deceleration rate to avoid a crash (DRAC) is incorporated to examine whether cross-process information can improve MTTC-based crash risk forecasting. To represent different forms of temporal dependence, both exponential and epidemic-type aftershock sequence triggering kernels are considered. Four cross-exciting models, including unidirectional and bidirectional specifications under the two triggering kernels, are developed and compared with two corresponding self-exciting benchmark models. Traffic conflict data collected over two observation days at a signalized intersection in Surrey, Canada, are used for the empirical analysis. Results show that both the self-exciting and cross-exciting models produce statistically adequate crash risk forecasts across the examined confidence levels. More importantly, the cross-exciting models generally outperform the self-exciting models in terms of both forecast accuracy and precision, indicating that incorporating cross-process dependence can improve short-term crash risk forecasting. Within the cross-exciting framework, the unidirectional specification consistently performs better than the bidirectional specification, suggesting that a simpler directional cross-exciting structure can provide a better balance between predictive information and estimation uncertainty. Among all model specifications, the unidirectional cross-exciting epidemic-type aftershock sequence conditional peak-over-threshold model achieves the best overall forecasting performance. These findings highlight the value of incorporating cross-process dependence into conflict-based crash risk forecasting and have practical implications for proactive intersection safety management.
Temporally shifting parameters in crash severity modeling is a well-documented phenomenon, with growing evidence suggesting that the influence of explanatory variables changes over time due to shifts in driver behavior, vehicle technology, and roadway conditions. Many studies have examined this issue by comparing the temporal stability of adjacent-year data using a variety of modeling frameworks that have often assumed a homogenous effect of explanatory variables across the entire crash population. The current research effort departs from past work in two ways. First, it compares data separated by multiple years (instead of comparing adjacent years), and second, it tests the homogeneous effect assumptions by estimating a latent segmentationbased pooled generalized ordered logit model. Using crash-level data from the National Automotive Sampling System/Crashworthiness Data System for the years 2003, 2009, and 2015 (each data wave separated by 5 years), the performance of the proposed latent segmentation model is evaluated against several traditional alternatives to assess their predictive and behavioral relevance. Various model fit measures demonstrate the improved performance of the proposed modeling approach, thus highlighting the importance of capturing population heterogeneity in temporal analyses. Interestingly, the latent segmentation modeling framework reveals that temporal variation in injury-severity outcomes (over the extended time intervals considered) is largely confined to a specific segment of crashes involving what would be considered low-injury risk, while the higher-injury risk group has parameters that remain stable across years. A marginal effect and scenario-based policy analysis provide additional insights. The results suggest that certain crash/individual profiles may be more susceptible to temporal shifts, and hence, accounting for such heterogeneity can be critical for developing targeted and effective safety interventions.
Traditional models, employing extreme value theory for estimating pedestrian crashes from traffic conflicts, commonly utilise popular conflict measures, such as post encroachment time and gap time. Whilst these measures have proven useful, they are limited in identifying a vehicle-pedestrian conflict based on a fixed threshold value and depend on subjective graphical-based extreme identification methods, which neither fully capture the dynamic interactions between vehicles and pedestrians nor account for road user behaviour to identify conflicting events. This study proposes a bivariate extreme value modelling framework that analyses evasive action-based traffic conflicts by integrating risk force theory and artificial intelligence-based video analytics to estimate pedestrian crash frequency by severity. The methodological framework quantifies crash risk dynamically during vehicle-pedestrian interactions and identifies traffic conflict events based on evasive behaviours. Traffic conflicts are modelled using a Generalised Pareto distribution to capture the tail behaviour of high-risk conflicts. The proposed econometric modelling framework was validated using 72 h of traffic movement data from three signalised intersections in Queensland, Australia. Results demonstrate that the Generalised Pareto distributions effectively fit evasive action-based vehicle-pedestrian conflicts, with estimated total pedestrian frequency and severe crash frequency aligning closely with historical crash records, thereby supporting the validity of the proposed model. This study presents a scalable, behaviourally grounded methodology as an alternative to a subjective conflict identification approach, enabling continuous risk assessment for proactive pedestrian safety management and real-time safety analysis.
Regional economic disparities contribute to a disproportionate number of fatal and severe crashes among active travelers (pedestrians and bicyclists) in economically disadvantaged areas. Such road safety inequalities may be further exacerbated by external shocks such as the COVID-19 pandemic, due to regional variations in safety resilience. However, few studies have examined how the determinants of injury severity vary across regions with differing economic conditions, while accounting for COVID-contributing temporal shifts. This study uses North Carolina as a case study, classifying counties into three groups (i.e., highly, moderately, and least distressed counties) based on four economic indicators, and defining three pandemic periods (i.e., before, during, and after the pandemic). A partially constrained random parameter multinomial logit model with heterogeneity in the means and variances is estimated for crashes in each county group. Results show that the effects of factors are more stable in the least distressed counties, suggesting stronger safety resilience under external shocks. Additionally, during the pandemic, alcohol-impaired driving significantly affected injury severity only in highly and moderately distressed counties. Out-of-sample predictions further suggest that the probability of severe injuries among active travelers increases with rising regional economic distress and after the pandemic. Moreover, compared to the least distressed counties, the reduced safety resilience in highly and moderately distressed counties is attributed to weaker recovery and resistance capacities, respectively. These findings provide valuable insights for formulating region-specific policies, detecting system vulnerabilities, and promoting equitable and sustainable active transportation systems.
Over the past few decades, traffic conflict modelling with proximity-based conflicts has emerged as a key approach for estimating crash risk from traffic conflicts, with extreme value models providing a rigorous framework for extrapolating rare-event probabilities. However, proximitybased definitions of conflicts may lead to biased estimation, as they often include interactions arising from deliberate and controlled driving behaviours that may not correspond to actual crash likelihood. In contrast, failure-induced conflicts that take into account evasive action and response delays can potentially overcome this limitation. Despite these advances, a comprehensive comparison of proximity-based and failure-induced conflicts within crash risk modelling is still lacking. This study addresses this gap by comparing and evaluating the performance of different threshold-exceedance modelling approaches for crash frequency estimation. Three threshold-exceedance models are evaluated in the study, including (i) a Lomax model applied to response delays during failure-induced conflicts, (ii) a Generalized Pareto Distribution model for proximity-based conflicts, and (iii) a Generalized Pareto Distribution model for failure-induced conflicts. Empirical analysis is conducted using high-resolution trajectory data from four signalized intersections in Brisbane, Australia. The results indicate that both the Generalized Pareto Distribution model for failure-induced conflicts and the Lomax model, representing response delays within failure-induced conflicts, provided reasonable estimates of historical rear-end crashes, with predicted crash counts contained within the 95 % confidence interval of the observed crash data. In contrast, the Generalized Pareto Distribution model for the proximitybased conflicts overestimated the crash frequency. Notably, within the failure-induced conflicts, the Generalized Pareto Distribution model demonstrated greater accuracy than the Lomax model, yielding estimates closer to the observed mean and with narrower confidence bounds, thereby indicating higher predictive precision. Overall, the findings underscore the value of incorporating failure-induced conflicts into traffic conflict modelling, revealing that the Generalized Pareto Distribution model with failure-induced conflicts provides more accurate and reliable crash risk estimates.
Small-scale temporal factors have substantial effects on zonal crash risk, yet their influence has long been overlooked due to data limitations. This omission may introduce confounding and bias the estimation of observed annually aggregated variables. This study aims to combine high-resolution traffic dynamic patterns derived from taxi trajectory data and advanced spatiotemporal models to account for hourly-scale temporal instability in year-level crash frequency modeling. Two sets of hourly-scale spatiotemporal traffic dynamic patterns were extracted, enabling the development of small-scale models. Spatiotemporal model with adaptive smoothing spatial specification was employed to further capture temporal effects. Model specification results revealed strong temporal autocorrelation at the hourly scale, with a magnitude comparable to the spatial autocorrelation. The results showed that the model effectively captured the varying safety impacts of unobserved temporal factors across hourly intervals, and that the extracted high-resolution patterns successfully internalized previously unobserved time-relevant information into the model’s linear component. Comparative analyses demonstrated that both incorporating traffic dynamic patterns and accounting for hourly-scale spatiotemporal autocorrelation in crash frequency modeling significantly improved the model fit and predictive performance. The proposed framework detected a broader set of risk factors than purely spatial models, and yielded less biased posterior means and more rigorous intervals through disentangling small-scale noise from fixed-effect signals and preventing pseudo-replication of observations. The extracted traffic flow dynamics also represent a fundamental yet traditionally inaccessible set of factors in regional crash analysis. This study established their macro-level associations with crash risk, revealing that higher mean speeds and lower speed fluctuations are linked to reduced crash risk. Additionally, these patterns can be regarded as “high-frequency” features, and those extracted from just one month of data were proved sufficient to construct models comparable to those based on the full-year dataset. This finding enables a practical framework that leverages real-time updated high-frequency variables as model inputs for rolling crash risk prediction. The methods and findings of this study offer practitioners in-depth macro-level insights into crash causation and valuable guidance on regional traffic safety interventions.
Intersection-related vehicle-pedestrian collisions present a significant challenge in transportation safety due to the complexity and hazards of intersections within urban road networks. This study introduces a spatially aggregated ordered logit model with a joint multivariate normal structure, which offers distinct advantages over conventional models by effectively capturing correlations among vehicle movement types (left-turn, straight, and right-turn) and accounting for residual aggregation at both intersection and county levels. Using a dataset of 4280 pedestrian-vehicle crashes in Florida from 2019 to 2023, incorporating pedestrian, driver, vehicle, intersection, environmental, crash, and temporal characteristics, the proposed model demonstrates superior performance in capturing interdependencies among vehicle maneuvers. Four temporally consistently significant variables are identified including pedestrians aged under 18 years old, urban areas, major roadway speed limits below 30 mph and lighted roadways during nighttime. In contrast, several other variables demonstrate significance only in specific years, reflecting notable temporal variation in their impact on pedestrian injury severity. A series of statistical tests, including normality distribution tests, spatial autocorrelation tests, and assessments of independence and homoscedasticity, were conducted to validate the model. The results confirm the model's ability to satisfy critical statistical assumptions-normality, independence, homoscedasticity, and spatial autocorrelation-and its robustness in achieving a high degree of spatial independence. The findings underscore the need for targeted safety measures and intersection design strategies to mitigate collision risks. By offering enhanced accuracy, temporal flexibility, and spatial insights, the proposed modeling approach provides a robust framework for developing evidence-based safety interventions and optimizing intersection designs to reduce pedestrian injury severity.
Vehicle pre-crash travel speed is one of the most important determinants of driver injury severity. However, pre-crash travel speed estimates made by police officers, especially those in crashes with less severe injuries (where there is less of a need for high levels of accuracy due to potential litigation), can be susceptible to biases because of the tendency to associate less severe driver injuries with lower pre-crash travel speeds. This potential bias makes the use of pre-crash travel speeds in injury-severity modeling highly problematic due to its endogeneity with injury severity. To detect the presence and extent of this problem, a bias correction term for pre-crash travel speed estimation equations is applied by treating injury-severity level (discrete) and pre-crash travel speed (continuous) as a discrete/continuous econometric model. The findings show that for severe injury crashes, the bias correction is statistically insignificant, reflecting the increased accuracy required of police officers in severe crashes. However, for crashes resulting in less severe occupant injuries, there is a significant bias resulting from observed injury levels, which distorts the effects of explanatory variables on pre-crash travel speed estimates. The results of this paper not only provide empirical evidence of potential endogeneity problems in models of crash injury severity but also underscore the need to more fully consider potential endogeneity issues and their associated consequences in statistical models and machine learning models.
Many recent studies have shown that data segmentation (seeking to segment the data into potentially homogeneous groups by factors such as data-collection year, driver age, driver gender, driver behaviors, etc.) can significantly improve crash injury-severity model estimation results. However, the choice of the segmentation criterion is often speculative and based on a predetermined expectation of homogeneity by the analyst. In an effort to improve model estimation results, a potential alternative to analyst-specified data segmentation is to preprocess the data using multivariate machine learning techniques. This paper demonstrates the potential of data preprocessing using k-means clustering as a means to improve the estimation of statistical models. Empirical results show that the combination of k-means clustering, in addition to data segmentation by year to account for temporal shifts in parameters, result in an improved statistical fit (a hybrid of analyst-specified and machine learning data segmentation). Furthermore, a comparison of the marginal effects generated by the clustered and non-clustered models suggests that the preprocessing of data by clustering techniques can result in more precise marginal effect estimates to guide safety policies. The findings show considerable potential for using machine learning algorithms, such as k-means clustering, to improve the estimation results of statistical models.
Extreme Value Theory (EVT) has become a widely used approach for quantifying crash risk from traffic conflict data. Most existing applications, however, rely on unconditional models, which fail to adequately capture dependence in extreme traffic conflicts and do not reliably predict future crash risk. To demonstrate the potential of conditional EVT models for advancing short-term crash risk forecasting, this study compares two conditional EVT approaches within a Bayesian framework that address extremal dependence from distinct perspectives. The first approach is the two-stage GARCH-EVT framework, where conditional mean and variance are modeled using GARCH-type specifications before EVT is applied to the standardized residuals. Both traditional and covariate-augmented variants are examined. The second approach uses a one-stage conditional peak-over-threshold (POT) model, represented by the score-driven POT model, which directly captures dynamics in the conditional exceedance probability and the distribution of exceedance sizes. Crash risk is quantified using two conditional tail risk measures, Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR), with forecasting performance evaluated through traditional and comparative backtesting. An empirical study examines rear-end conflicts collected at two signalized intersections over four observation days to generate one-cycle-ahead crash risk forecasts during the out-of-sample period. Traditional backtesting indicates that both the covariate-augmented GARCH-EVT and the score-driven POT approaches produce valid and comparable forecasts, with the two-stage method yielding estimates with lower uncertainty. Comparative backtesting, however, shows that the score-driven POT model achieves slightly superior forecasting accuracy. The weaker performance of the two-stage framework can be attributed to partial removal of extremal dependence, sensitivity to substitute values in cycles without conflicts, and the limitations inherent in its two-stage structure.
In this research, statistical and deep learning models are applied to determine factors that affect motorcycle crash-injury severities. Four methodological challenges are considered: 1) imbalanced data (because fatal injuries are an exceedingly small portion of all resulting injury outcomes); 2) unobserved heterogeneity (because many unobserved factors will influence resulting injury severities); 3) quantification of variable effects; and 4) the possibility of temporally shifting relationships among variables. Convolutional neural networks and deep neural networks are the deep learning models considered, and random parameters logit models with heterogeneity in means and variances is the statistical model considered. Extensive experimentation indicated that data imbalance and unobserved heterogeneity could be best handled in deep learning models with a Bayesian deep neural network with a random generator and weighted loss function. With statistical modeling indicating significant shifts in model parameters over time, the data were segmented by year and both statistical and deep learning models were estimated. While techniques are available for deep learning to potentially handle data imbalance and unobserved heterogeneity, the quantification of variable effects and temporal shifts remains a challenge. For example, a comparison of variable effects show that the deep learning estimates of variable effects are generally inconsistent with the plausible values generated by the statistical models in terms of magnitudes and occasionally in terms of direction, indicating a need for improvements in deep-learning variable-effect extraction methods. The findings also show the need for future work to isolate the effect of complex temporal relationships which are currently imbedded in deep learning approaches, because the segmentation of data that has been used in statistical models to isolate temporal effects, and even the use of all data and defining new time-dependent variables, may not be a viable deep learning option due to the potential loss in predictive performance.
Independent traffic crash modeling approaches do not account for the embedded relationships related to the multi-resolution data structure, leading to mis-specified estimations. The recently developed integrated frameworks demonstrate the capability of addressing this drawback. The current study proposes an integrated framework that accommodates information from multiple spatial units and observation resolutions. Specifically, the study develops an integrated model system that allows for the influence of independent variables from disaggregate crash record, micro-facility (segment and intersection) and macro (traffic analysis zone) level simultaneously within the macro level propensity estimation. The empirical analysis considers disaggregate crash records of 1818 segments and 4184 intersections from 300 traffic analysis zones in the City of Orlando, Florida. These crash records contain crash-specific factors, driver and vehicle factors, roadway, road environmental and weather information of each crash record. For micro-facility and macro levels, an exhaustive set of independent variables including roadway and traffic factors, land-use and built environment attributes, and sociodemographic characteristics are considered. The proposed model system can also accommodate for hierarchical correlations among the data across observation resolutions and parameter variability across the system. The empirical analysis is augmented by employing several goodness of fit and predictive measures. The results clearly demonstrate the improved performance offered by the proposed integrated model system relative to the non-integrated model. A validation exercise also highlights the superiority of the proposed framework. The application of the proposed integrated framework can allow transportation professionals to adopt policy-based, site-specific, and outcome-specific solutions simultaneously.
General aviation experiences significant variation in accident characteristics across flight phases. This study seeks to investigate the phase transferability and temporal stability of determinants influencing general aviation accidents, using the U.S. data (2008-2019) from the National Transportation Safety Board. To achieve this, a random parameter bivariate approach with heterogeneity in means was employed, focusing on two binary outcomes: injury severity (fatal/severe vs. minor/none) and aircraft damage (destroyed vs. non-destroyed). Four flight phases were analyzed: departure, enroute, maneuvering, and arrival. The data were divided into three time periods, 2008-2011, 2012-2015, and 2016-2019, to assess the determinants' temporal stability. Likelihood ratio tests revealed that pilot injury and aircraft damage risks exhibit phase non-transferability and temporal instability. Out-of-sample predictions indicated a steady rise in fatal or severe injury risk, while aircraft damage risk initially increased before declining over time. A significant positive correlation between pilot injury and aircraft damage was observed through model estimation. Key factors, including pilot, aircraft, flight, and environmental conditions, significantly influenced both outcomes. Moreover, factors such as decision-making errors, adverse physiological conditions, fixed landing gear, and visual meteorological conditions showed both phase transferability and temporal stability. However, most factors were phase-and period-specific. Based on these findings, targeted measures, such as pilot escape and survival training, as well as phase-specific, scenario-based training, are proposed to mitigate general aviation risks.
The COVID-19 pandemic reshaped the global transportation sector, including in the U.S., creating an unprecedented shift in traffic patterns. Despite a reduction in vehicle miles traveled (VMT), crash severity, particularly fatalities, increased significantly. Among all crash types, fixed-object collisions have consistently posed a critical safety concern due to their disproportionately high fatality rates, a trend further exacerbated during the pandemic. This study examines the impact of COVID-19 on driver injury severity in fixed-object passenger car crashes in Oregon. The authors estimated separate unconstrained models of driver injury severity in fixed-object passenger car crashes across three distinct time periods: before pandemic (March 2019-February 2020), during pandemic (March 2020-February 2021), and after pandemic (March 2021-February 2022), as well as a partially constrained model utilizing a random parameters multinomial logit model that incorporates heterogeneity in both means and variances of the random parameters. The analysis utilized 22,522 crash records for the state of Oregon obtained from the Oregon Department of Transportation. Likelihood ratio tests were performed to assess the temporal instability of model parameter estimates throughout the three time periods and to compare the partially constrained and unconstrained models. The findings indicated notable temporal variations in the determinants of injury severity, encompassing driver attributes, crash circumstances, roadway characteristics, and environmental elements. While alcohol consumption, improper driving, and collisions with trees consistently influenced injury severity across all periods, factors such as gender, airbag deployment, speeding, seasonal variations, and road surface conditions exhibited changing effects. Out-of-sample predictions indicate that severe injuries in fixed-object crashes were consistently underestimated, highlighting growing concerns about increasing crash severity, particularly in the post-pandemic period.
The distribution of economic loss associated with vessel accidents typically exhibits non-negative, continuous, positively skewed, and heavy-tailed characteristics. Another challenge in analyzing fishing vessel accidents is the absence of relevant factors. Ignoring such heterogeneity caused by unobserved factors potentially leads to inaccurate inferences. In the present study, a novel Bayesian random-parameter generalized beta of the second kind (GB2) model with possible heterogeneity in means and variances was developed. The flexible GB2 distribution was harnessed to model the skewed and heavy-tailed response variable, while the random parameters were specified to capture the unobserved heterogeneity. The proposed method was validated using an insurance claim dataset with 3448 fishing vessel accidents within Ningbo waters during 2018-2022. The proposed model successfully identified significant influential factors, including fixed parameters, random parameters, and covariates influencing the means of the random parameters. Specifically, offshore and inevitable accidents, fishing transport vessels, double-trawl vessels with mechanical failures, wide-hulled vessels, and favorable sea conditions were associated with greater economic loss. Special attention should also be paid to nighttime accidents involving steel-hulled fishing transport vessels, as this accident type emerged to result in greater loss during the pandemic lockdown period. Our approach can accommodate the abnormality, skewness, and heavy-tail of vessel accident loss data, adjust for the bias introduced by unobserved factors, and uncover the interactive relationship among covariates. Targeted countermeasures were proposed to mitigate economic loss resulting from fishing vessel accidents.
