Jiao Zhou

Comparison of Multivariate Poisson lognormal spatial and temporal crash models to identify hot spots of intersections based on crash types

Most of the studies are focused on the general crashes or total crash counts with considerably less research dedicated to different crash types. This study employs the Systemic approach for detection of hotspots and comprehensively cross-validates five multivariate models of crash type-based HSID methods which incorporate spatial and temporal random effects. It is anticipated that comparison of the crash estimation results of the five models would identify the impact of varied random effects on the HSID. The data over a ten year time period (2003-2012) were selected for analysis of a total 137 intersections in the City of Corona, California. The crash types collected in this study include: Rear-end, Head-on, Side-swipe, Broad-side, Hit object, and Others. Statistically significant correlations among crash outcomes for the heterogeneity error term were observed which clearly demonstrated their multivariate nature. Additionally, the spatial random effects revealed the correlations among neighboring intersections across crash types. Five cross-validation criteria which contains, Residual Sum of Squares, Kappa, Mean Absolute Deviation, Method Consistency Test, and Total Rank Difference, were applied to assess the performance of the five HSID methods at crash estimation. In terms of accumulated results which combined all crash types, the model with spatial random effects consistently outperformed the other competing models with a significant margin. However, the inclusion of spatial random effect in temporal models fell short of attaining the expected results. The overall observation from the model fitness and validation results failed to highlight any correlation among better model fitness and superior crash estimation.

Evaluating Influence of Neighboring Structures on Spatial Crash Frequency Modeling and Site-Ranking Performance

Many neighborhood weight matrices have been adopted for modeling crash spatial heterogeneity. However, there has been little evaluation of their influence on crash prediction modeling performance. This study investigated 17 spatial-proximity matrices for development of spatial crash prediction models and site ranking with county-level data in California. Of the group of matrices being evaluated, traffic exposure–weighted and population-weighted distance-based matrices were first proposed in the traffic safety field. Bayesian spatial analysis was conducted with a combination of a first-order autoregressive error process and time trend for crashes to address the serial correlation of crashes in successive years. Two diagnostic measures were used for assessment of goodness of fit and complexity of models, and seven evaluation criteria were employed to assess the benefits associated with better-fitting models in site ranking. The results showed that modeling performance improved with an increase in number of neighbors considered in the weight matrix. However, a larger number of neighbors also led to greater variability of modeling performance. Specifically, Queen-2 and Decay-50 models proved to be superior among the adjacency- and distance-based models, respectively. The significance of incorporating spatial correlations was highlighted by the consistently poor performance of the base model, which included only the heterogeneity random effect. Finally, model-fitting performance seems to be strongly correlated with site-ranking performance. The models with closer goodness of fit tend to yield more consistent ranking results.

Predicting motorcycle crash injury severity using weather data and alternative Bayesian multivariate crash frequency models

Motorcycle crashes constitute a very high proportion of the overall motor vehicle fatalities in the United States, and many studies have examined the influential factors under various conditions. However, research on the impact of weather conditions on the motorcycle crash severity is not well documented. In this study, we examined the impact of weather conditions on motorcycle crash injuries at four different severity levels using San Francisco motorcycle crash injury data. Five models were developed using Full Bayesian formulation accounting for different correlations commonly seen in crash data and then compared for fitness and performance. Results indicate that the models with serial and severity variations of parameters had superior fit, and the capability of accurate crash prediction. The inferences from the parameter estimates from the five models were: an increase in the air temperature reduced the possibility of a fatal crash but had a reverse impact on crashes of other severity levels; humidity in air was not observed to have a predictable or strong impact on crashes; the occurrence of rainfall decreased the possibility of crashes for all severity levels. Transportation agencies might benefit from the research results to improve road safety by providing motorcyclists with information regarding the risk of certain crash severity levels for special weather conditions.

Comparative evaluation of temporal correlation treatment in crash frequency modelling

ABSTRACT There is relatively little research dedicated to the evaluation of different temporal treatments on modelling performance. This study proposed two new methods which combined the strengths of linear trend and time-varying coefficients with the autoregressive process and compared their performance with seven other temporal models used in the past. All models generated a similar number of statistically significant variables and close variable coefficients, but different modelling performance. For prediction accuracy, the model which accounts only for autoregressive effect illustrated superior performance in terms of cross-validation and typical assessment, which was based on same data used to develop models. Nonetheless, if the penalized criterion was used, both proposed models outperformed other competing models, indicating their capability to yield similar prediction accuracy with relatively smaller effective number of parameters. This suggests further exploration of models that combine various temporal treatments. Finally, the correlations were also observed among the various modelling assessment criteria.

Ranking cities for safety investigation by potential for safety improvement

ABSTRACT The authors performed a city-level hot-spot identification by using the 4-year data of 265 cities in California. It is intended to equip road safety professionals with more useful information to compare the safety performance of city as a whole. Potential for safety improvement (PSI) was adopted as a measure of crash risk to compare alternate identification of hot spot (HSID) methods, including the Empirical Bayes (EB) and three full Bayesian (FB) alternatives, negative binomial, Poisson log-normal, and the Poisson temporal random effect, for ranking the safety performance of cities. Five evaluation tests which contain the Site Consistency Test, the Method Consistency Test, the Total Rank Difference Test, the Total Performance Difference Test, and the Total Score Test were applied to evaluate the performance of the four HSID methods. Moreover, two cutoff levels, top 5% and 10% cities, were employed for more reliable results. Overall, the study results are consistent with the results of previous quantitative evaluations focused on micro-level HSID. The three FB approaches significantly outperform the EB counterpart. The method accounting for temporal random effect produces more reliable HSID results than those without considering the serial correlations in collision counts.

Comparative analysis of cost-weighted site ranking using alternate distance-based neighboring structures for spatial crash frequency modeling

ABSTRACT There are significant cost differences between alternate crash outcomes based on severity levels. This study aims to expand existing literature with the inclusion of spatial correlations among a group of intersections to compare alternate distance-based weight matrices for cost-weighted hotspot identification (HSID) purpose. Multivariate-Poisson-lognormal-spatial (MVPLNS) method was employed to develop five crash prediction models (pure-distance and decay) to jointly estimate four severity levels (fatal and severe injury, other visible injury, complaint of pain, and noninjury). Model comparison for assessment of goodness-of-fit indicated that relatively subtle matrix structures assign consistent weights that reduce model complexity and eventually enhances the overall fit. The assessment of predictive accuracy of model estimates indicated that the model fit may be correlated with a superior performance at prediction as witnessed in the case of a pure-distance model that consistently exhibited least discrepancy from actual crash counts. The evaluation of HSID performance was based on rankings obtained from cost-weighted severities. The pure-distance models were overall superior at HSID, but among these models, the more subtle model performed significantly better, which hints at the presence of correlation between model fit and HSID performance as it may be possible that benefits of superior fit transfer to equivalent HSID capabilities.

Comprehensive Assessment of Temporal Treatments in Crash Prediction Models

This study conducted a comprehensive comparison of temporal treatments employed in crash prediction models. Nine groups of methodological approaches based on different ways of addressing temporal correlations, including the newly proposed time adjacency matrix, were developed. Moreover, three types of models were developed for each group in terms of spatial dependency. Finally, ten different assessment criteria were utilized for the evaluation purpose. All models and performance-checking criteria applied to 8 years of county-level crash counts in California. The modeling results illustrated that the space–time models consistently enhanced the precision associated with the intercepts. The serial and spatial correlations also appeared to be statistically significant. In terms of model complexity, the models with spatial correlations outperformed the ones without considering spatially structured heterogeneity, and the models accounting for the temporal dependency revealed more benefits compared with those without temporal treatments. The opposite trends were found by prediction-pertinent criteria based on the aggregation results, even though the first-order autoregressive process space–time models with spatiotemporal interaction claimed the first place of prediction in most cases. The correlation analysis among all ten criteria illustrated that the efficiency in reducing the effective number of parameters tended to have larger impacts on the value of deviance information criterion than did the mean deviance, which demonstrated the statistically significant correlations with all other prediction-related measures.

Use of Bivariate Dirichlet Process Mixture Spatial Model to Estimate Active Transportation-Related Crash Counts

The current paper presents the comprehensive analysis of a bivariate Dirichlet process mixture spatial model for estimation of pedestrian and bicycle crash counts. This study focuses on active transportation at traffic analysis zone (TAZ) level by developing a semi-parametric model that accounts for the unobserved heterogeneity by combining the strengths of bivariate specification for correlation among crash modes; spatial random effects for the impact of neighboring TAZs; and Dirichlet process mixture for random intercept. Three alternate models, one Dirichlet and two parametric, are also developed for comparison based on different criteria. Bicycle and pedestrian crashes are observed to share three influential variables: the positive correlation of K12 student enrollment; the bike-lane density; and the percentage of arterial roads. The heterogeneity error term demonstrates the presence of statistically significant correlation among the bicycle and pedestrian crashes, whereas the spatial random effect term indicates the absence of a significant correlation for the area under focus. The Dirichlet models are consistently superior to non-Dirichlet ones under all evaluation criteria. Moreover, the Dirichlet models exhibit the capability to identify latent distinct subpopulations and suggest that the normal assumption of intercept associated with traditional parametric models does not hold true for the TAZ-level crash dataset of the current study.