An accelerated hierarchical Bayesian crash frequency model with accommodation of spatiotemporal interactions

Are ride-hailing services safer than taxis? A multivariate spatial approach with accommodation of exposure uncertainty

Despite many research efforts on ride-hailing services and taxis, limited studies have compared the safety performance of the two modes. A major challenge is the need for reliable mode-specific exposure data to model their safety outcomes. Moreover, crash frequencies of the two modes by injury severities tend to be spatially and inherently correlated. To fully address these issues, this study proposes a novel multivariate conditional autoregressive model considering measurement errors in mode-specific exposures (MVCARME). More specially, a classical measurement error structure accommodates the uncertainty of estimated mode-specific exposures, and a multivariate spatial specification is adopted to capture potential spatial and inherent correlations. The model estimation is accelerated by an integrated nest Laplace approximation method. The census tracts in the city of Chicago are set as the spatial analysis unit. The mode-specific exposures (vehicle-mile-traveled) in each census tract are estimated by trip assignments using ride-hailing and taxi trip data in 2019. The modeling results indicate that both ride-hailing crashes and taxi crashes are positively associated with transportation factors (e.g., vehicle-mile-traveled, mode-specific vehicle-mile-traveled, and traffic signal numbers), land use factors (i.e., number of educational and alcohol-related sites), and demographic factors (e.g., median household income, transit ratio, and walk ratio). By comparison, the proposed model outperforms the others (i.e., negative binomial models and multivariate conditional autoregressive model) by yielding the lowest deviance information criterion (DIC), Watanabe-Akaike information criterion (WAIC), mean absolute error (MAE), and root-mean-square error (RMSE). According to the results of t-tests, ride-hailing services are found to be prone to a higher risk of minor injury crashes compared with taxis, despite no significant difference between the risks of severe injury crashes. Methodologically, this study adds a robust safety evaluation approach for comparing crash risks of different modes to the literature. At the same time, practically, it provides researchers, practitioners, and policy-makers insights into the safety management of various mobility alternatives.

A full Bayesian multilevel approach for modeling interaction effects in single-vehicle crashes

Interaction effects constitute crucial crash attributes that can be classified into two distinct categories: spatiotemporal interactions and factor interactions. These interactions are rarely addressed systematically in modeling the severity of single-vehicle (SV) crashes. This study focuses on uncovering these crash attributes by designing a full Bayesian spatiotemporal interaction multilevel logit (STIML-logit) approach with heterogeneity in means and variances (HMV). Meanwhile, a nested Gaussian conditional autoregressive (CAR) structure is proposed to fit the spatiotemporal interaction component and its effectiveness is verified by calibrating four different interaction patterns. A standard multilevel logit (with and without HMV), a multilevel logit with HMV, and a spatiotemporal multilevel logit with HMV are constructed for comparison. Risk factors are decomposed into traffic environment factors (group level) and individual crash factors (case level) to construct a multilevel structure and to capture possible interactions between risk factors from different levels (cross-level factor interactions). We perform regression modeling utilizing SV crash cases covering 96 major urban roads in Shandong, China. The modeling results underscore several significant findings: (1) the STIML-logit with HMV demonstrates the best regression performance, suggesting that systematically dealing with the interaction effects and the HMV is a trustworthy modeling perspective; (2) crash models with the nested CAR outperform those with the traditional CAR and the result is supported by all the spatiotemporal statistical functions, highlighting the potential advantages of the nested structure; (3) all the environment factors maintain significant interactions with the case factors, highlighting that the contribution of the environment factors to crash injuries is not constant but is rather influenced by the specific case-related crash factors. The study introduces a promising regression architecture for modeling crash injuries and revealing subtle crash attributes.

Modeling spatiotemporal interactions in single-vehicle crash severity by road types

INTRODUCTION

Spatiotemporal correlations have been widely recognized in single-vehicle (SV) crash severity analysis. However, the interactions between them are rarely explored. The current research proposed a spatiotemporal interaction logit (STI-logit) model to regression SV crash severity using observations in Shandong, China.

METHOD

Two representative regression patterns-mixture component and Gaussian conditional autoregression (CAR)-were employed separately to characterize the spatiotemporal interactions. Two existing statistical techniques-spatiotemporal logit and random parameters logit-were also calibrated and compared with the proposed approach with the aim of highlighting the best one. In addition, three road types-arterial road, secondary road, and branch road-were modeled separately to clarify the variable influence of contributors on crash severity.

RESULTS

The calibration results indicate that the STI-logit model outperforms other crash models, highlighting that comprehensively accommodating spatiotemporal correlations and their interactions is a recommended crash modeling approach. Additionally, the STI-logit using mixture component fits crash observations better than that using Gaussian CAR and this finding remains stable across road types, suggesting that simultaneously accommodating stable and unstable spatiotemporal risk patterns can further strengthen model fit. According to the significance of risk factors, there is a significant positive correlation between distracted diving, drunk driving, motorcycle, dark (without street lighting), and collision with fixed object and serious SV crashes. Truck and collision with pedestrian significantly mitigate the likelihood of serious SV crashes. Interestingly, the coefficient of roadside hard barrier is significant and positive in branch road model, but it is not significant in arterial road model and secondary road model.

PRACTICAL APPLICATIONS

These findings provide a superior modeling framework and various significant contributors, which are beneficial for mitigating the risk of serious crashes.

Analysis of the injury severity of motor vehicle-pedestrian crashes at urban intersections using spatiotemporal logistic regression models

This paper conducted a comprehensive study on the injury severity of motor vehicle-pedestrian crashes at 489 urban intersections across a dense road network based on high-resolution accident data recorded by the police from 2010 to 2019 in Hong Kong. Given that accounting for the spatial and temporal correlations simultaneously among crash data can contribute to unbiased parameter estimations for exogenous variables and improved model performance, we developed spatiotemporal logistic regression models with various spatial formulations and temporal configurations. The results indicated that the model with the Leroux conditional autoregressive prior and random walk structure outperformed other alternatives in terms of goodness-of-fit and classification accuracy. According to the parameter estimates, pedestrian age, head injury, pedestrian location, pedestrian actions, driver maneuvers, vehicle type, first point of collision, and traffic congestion status significantly affected the severity of pedestrian injuries. On the basis of our analysis, a range of targeted countermeasures integrating safety education, traffic enforcement, road design, and intelligent traffic technologies were proposed to improve the safe mobility of pedestrians at urban intersections. The present study provides a rich and sound toolkit for safety analysts to deal with spatiotemporal correlations when modeling crashes aggregated at contiguous spatial units within multiple years.

Modelling injury severity in single-vehicle crashes using full Bayesian random parameters multinomial approach

Single-vehicle (SV) crash severity model considering spatiotemporal correlations has been extensively investigated, but spatiotemporal interactions have not received sufficient attention. This research is dedicated to propose a superior spatiotemporal interaction correlated random parameters logit approach with heterogeneity in means and variances (STICRP-logit-HMV) for systematically characterizing unobserved heterogeneity, spatiotemporal correlations, and spatiotemporal interactions. Four flexible interaction formulations are developed to uncover the spatiotemporal interactions, including linear structure, Kronecker product, mixture-2 model, and mixture-5 model. Four candidate approaches-random parameters logit (RP-logit), RP-logit with heterogeneity in means and variances (RP-logit-HMV), correlated RP-logit-HMV (CRP-logit-HMV), and spatiotemporal CRP-logit-HMV (STCRP-logit-HMV)-are also established and compared with the proposed model. SV crash observations in Shandong Province, China, are employed to calibrate regression parameters. The model comparison results show that (1) the performance of the RP-logit-HMV model outperforms the RP-logit model, implying that capturing heterogeneity in the means and variances can strengthen model fit; (2) the CRP-logit-HMV model and the RP-logit-HMV model are comparable; (3) the STCRP-logit-HMV model outperforms the CRP-logit-HMV model, implying that addressing the spatiotemporal crash mechanisms is beneficial to the overall fitting of the crash model; (4) the STICRP-logit-HMV model performs better than the STCRP-logit-HMV model and this finding remains stable across different interaction formulations, indicating that comprehensively reflecting the spatiotemporal correlations and their interactions is a promising approach to model SV crashes. Among the four interaction models, the STICRP-logit-HMV model with mixture-5 component maintains the best fit, which is a recommended approach to model crash severity. The regression coefficients for young driver, male driver, and non-dry road surface are random across observations, suggesting that the influence of these factors on SV crash severity maintains significant heterogeneity effects. The research results provide transportation professionals with a superior statistical framework for diagnosing crash severity, which is beneficial for improving traffic safety.

Time-specific hierarchical models for predicting crash frequency of reversible and high-occupancy vehicle lanes

Time-specific Safety Performance Functions (SPFs) were proposed to achieve accurate and dynamic crash frequency predictions. This study contributes to the literature by developing time-specific SPFs for freeways that include reversible lanes (RL) and freeways that include High-Occupancy Vehicle lanes (HOV) using Microwave Vehicle Detection System (MVDS) data from Virginia, Arizona and Washington States. Variables that capture the time-specific traffic turbulence were prepared and considered in the developed SPFs. Moreover, two different hierarchical models were proposed to identify factors associated with the different crash types or severity in crash frequency prediction. The results indicated that the variables representing the volume difference between reversible and general-purpose lanes (GPL) were positively associated with crash frequency. Further, the variable that indicated the design of the access point of the reversible lane was positively associated with crash frequency. The models comparison results showed that the hierarchical models outperformed the corresponding Poisson lognormal model with lower AIC and MAE values. This study also tested the proposed hierarchical models on High-Occupancy Vehicle freeway sections and reached the same conclusion on model comparison results. The significant variables representing the logarithm of volume were found to be significant and positive with crash frequency. Moreover, the difference in average speed between the HOV lanes and GPL was also found to be positive and significant with the crash frequency. In general, this study successfully identified the factors associated with the different crash types or severity in crash frequency prediction models.

Bayesian spatial correlation, heterogeneity and spillover effect modeling for speed mean and variance on urban road networks

Analyzing speed mean and variance is vital to safety management in urban roadway networks. However, modeling speed mean and variance on structured roads could be influenced by the spatial effects, which are rarely addressed in the existing studies. The inadequacy may lead to biased conclusions when considering vehicle speed as a surrogate safety measure. The current study focuses on developing a Bayesian modeling approach with three types of spatial effects, i.e., spatial correlation, spatial heterogeneity, and spillover effect. To capture the spatial correlation, the study employs the intrinsic conditional autoregressive (ICAR) models, spatial lag models (SLM), and spatial error models (SEM). Spatial heterogeneity and spillover effect are considered by the random parameters approach and spatially lagged covariates (SLCs). Speed data are collected from the float cars running on 134 urban arterials in Chengdu, China. The results indicate that the random parameters ICAR model with SLCs (RPICAR-SLC) outperforms others in terms of goodness-of-fit, accuracy, and efficiency for modeling speed mean, while the random parameters ICAR model (RPICAR) is the best for modeling speed variance. Moreover, RPICAR-SLC and RPICAR models are beneficial to address spatial correlation of residuals, explaining the unobserved influence among the observations, and are less likely to cause biased or overestimated parameters. The study also discusses how traffic conditions, road characteristics, traffic management strategies, and facilities on roadway networks influence speed mean and variance. The findings highlight the importance of multi-type spatial effects on modeling speed mean and variance along the structured roadways.

Safety analytics at a granular level using a Gaussian process modulated renewal model: A case study of the COVID-19 pandemic

With the advance of intelligent transportation system technologies, contributing factors to crashes can be obtained in real time. Analyzing these factors can be critical in improving traffic safety. Despite many crash models having been successfully developed for safety analytics, most models associate crash observations and contributing factors at the aggregate level, resulting in potential information loss. This study proposes an efficient Gaussian process modulated renewal process model for safety analytics that does not suffer from information loss due to data aggregations. The proposed model can infer crash intensities in the continuous-time dimension so that they can be better associated with contributing factors that change over time. Moreover, the model can infer non-homogeneous intensities by relaxing the independent and identically distributed (i.i.d.) exponential assumption of the crash intervals. To demonstrate the validity and advantages of this proposed model, an empirical study examining the impacts of the COVID-19 pandemic on traffic safety at six interstate highway sections is performed. The accuracy of our proposed renewal model is verified by comparing the areas under the curve (AUC) of the inferred crash intensity function with the actual crash counts. Residual box plot shows that our proposed models have lower biases and variances compared with Poisson and Negative binomial models. Counterfactual crash intensities are then predicted conditioned on exogenous variables at the crash time. Time-varying safety impacts such as bimodal, unimodal, and parabolic patterns are observed at the selected highways. The case study shows the proposed model enables safety analytics at a granular level and provides a more detailed insight into the time-varying safety risk in a changing environment.

A comprehensive analysis of the relationships between the built environment and traffic safety in the Dutch urban areas

Built-environment factors potentially alleviate or aggravate traffic safety problems in urban areas. This paper aims to investigate the relationships of these factors with vehicle-bicycle and vehicle-vehicle property damage only (PDO) and killed and severe injury (KSI) crashes in urban areas. For this purpose, an area-level analysis using 100x100m² cells, along with a Spatial Hurdle Negative Binomial regression model were employed. The study area is composed of a selection of municipalities in the Netherlands-Randstad Area where major land-use developments have occurred since the 1970s. The study was conducted by developing a rich dataset composed of various national and local databases. The findings reveal that built-environment factors and land-use policies have substantial impacts on safety, which cannot be neglected. The factors explaining the land-use density and diversity in the area (e.g., urbanity and function mixing levels), as well as the land-use design characteristics (indicated by average age of the neighborhoods), traffic and road network characteristics, and proximity to different destinations influence the probability, frequency, and severity of crashes in urban areas. Furthermore, low socioeconomic levels are associated with a higher frequency of traffic crashes.

Estimating the effect of friction on crash risk: Reducing the effect of omitted variable bias that results from spatial correlation

Omitted variable bias is one of the main factors that lead to incorrect estimates of the effect of a variable on the expected number of crashes using regression modeling. We propose to use differencing of the (spatially adjacent) variables to reduce the effect of omitted variable bias. Differencing is a linear transformation that preserves the structure of the (generalized) linear model but can often result in significantly reducing the correlation between the variables. It is special case of (generalized) partial linear model regression which itself is a special case of a generalized additive model (GAM). In the spatial context used in this paper, differencing is similar to the well-known approach of including a spatial correlation structure (spatial error term) in the analysis of crash data. It is generally not clear how to interpret the results of models that include a spatial correlation structure and whether and how the added spatial correlation structure reduces the bias in the estimated regression parameters. However, for the case of differencing, it becomes clear how the effect of omitted variable bias is reduced by reducing the correlation between the variable of interest and the omitted variables. The order of differencing determines the dominant spatial scales of the variables considered in the model which affect how much the correlation is reduced. This reveals that omitted variable bias can be reduced when there are spatial scales at which the covariate of interest varies but the omitted variables either 1) are relatively homogeneous or 2) have variations that are not correlated to those of the variable of interest.

On random-parameter count models for out-of-sample crash prediction: Accounting for the variances of random-parameter distributions

One challenge faced by the random-parameter count models for crash prediction is the unavailability of unique coefficients for out-of-sample observations. The means of the random-parameter distributions are typically used without explicit consideration of the variances. In this study, by virtue of the Taylor series expansion, we proposed a straightforward yet analytic solution to include both the means and variances of random parameters for unbiased prediction. We then theoretically quantified the systematic bias arising from the omission of the variances of random parameters. Our numerical experiment further demonstrated that simply using the means of random parameters to predict the number of crashes for out-of-sample observations is fundamentally incorrect, which necessarily results in the underprediction of crash counts. Given the widespread use and ongoing prevalence of the random-parameter approach in crash analysis, special caution should be taken to avoid this silent pitfall when applying it for predictive purposes.