Doctor of Philosophy / Electric scooters, or e-scooters, have become an increasingly popular form of transportation over the recent years. However, there have been numerous reports of safety concerns, crashes, and injuries for e-scooter riders and other road users as a result of e-scooter misuse. Until recently, very little formal research has been conducted on the safety of this micromobility solution. This dissertation describes a series of studies that have investigated the contributing factors to safety concerns and identified countermeasures, such as policy recommendations, design optimization, and training, that can be implemented with an end goal of improving e-scooter safety.

There is extensive evidence indicating that teenaged drivers are over-represented in automobile crashes. Motor vehicle crashes are the leading cause of death for 15-20 year olds, accounting for over 40% of all fatalities for this age group. Although teen drivers account for only 6.3% of the driving population, they account for 14% of all traffic fatalities (TSF, 2004). Currently there is a lack of continuous and naturalistic data in the field of teenaged driving. The purpose of this study was to obtain continuous performance data from a naturalistic setting by equipping the personal vehicles of newly licensed teenaged drivers with a data collection system for the first six months of driving. Specifically, this study examined the eye scanning patterns of newly licensed teenaged drivers and experienced parent drivers on straight road segment. This study provides insight into the development and change of eye-glance behaviors over the first six months of driving, the differences between novice teenaged drivers and experienced parent drivers, and how passenger presence affects eye scanning patterns. Results from this study found significant differences between novice teenaged drivers and experienced adult drivers. The results showed that teenaged drivers had much shorter glance durations away from the forward roadway and allocated a higher percentage of their glances to locations that were considered driving-related when compared to the experienced adult group. Results from the study also showed when one passenger was present in the vehicle teenaged drivers tended to have a higher percentage of time spent with their eyes off of the forward roadway. / Master of Science

As active safety systems installed in vehicles become more common and more sophisticated, a concise method of testing them in conditions as close to real risk situations as possible becomes necessary. This study looks at the possibilities of developing use cases, using video recordings of real risk situations, obtained through naturalistic driving studies. The concept of conflicts is explored as a substitute to actual accidents. A method of finding conflicts in a large data material from looking at the acceleration signal and its derivative, referred to as jerk is also sought. These possibilities are tried on material from a previously conducted naturalistic driving study. The results are an improvement in the ability to find conflict situations automatically, and a suggestion to how use cases can be produced from video recordings of conflicts obtained through naturalistic driving studies. The DREAM framework is used and modified in order to aid with data collection and interpretation.

This thesis describes a research study relating naturalistic Global Positioning System (GPS) driving data with long-term traffic safety performance for two classes of roadways. These two classes are multilane arterial streets and limited access highways. GPS driving data used for this study was collected from 33 volunteer drivers from July 2012 to March 2013. The GPS devices used were custom GPS data loggers capable of recording speed, position, and other attributes at an average rate of 2.5 hertz. Linear Referencing in ESRI ArcMAP was performed to assign spatial and other roadway attributes to each GPS data point collected. GPS data was filtered to exclude data with high horizontal dilution of precision (HDOP), incorrect heading attributes or other GPS communication errors. For analysis of arterial roadways, the Two-Fluid model parameters were chosen as the measure for long-term traffic safety analysis. The Two-Fluid model was selected based on previous research which showed correlation between the Two-Fluid model parameters n and Tm and total crash rate along arterial roadways. Linearly referenced GPS data was utilized to obtain the total travel time and stop time for several half-mile long trips along two arterial roadways, Grand Avenue and California Boulevard, in San Luis Obispo. Regression between log transformed values of these variables (total travel time and stop time) were used to derive the parameters n and Tm. To estimate stop time for each trip, a vehicle “stop” was defined when the device was traveling at less than 2 miles per hour. Results showed that Grand Avenue had a higher value for n and a lower value for Tm, which suggests that Grand Avenue may have worse long-term safety performance as characterized by long-term crash rates. However, this was not verified with crash data due to incomplete crash data in the TIMS database. Analysis of arterial roadways concluded by verifying GPS data collected in the California Boulevard study with sample data collected utilizing a traditional “car chase” methodology, which showed that no significant difference in the two data sources existed when trips included noticeable stop times. For analysis of highways the derived measurement of vehicle jerk, or rate of change of acceleration, was calculated to explore its relationship with long-term traffic safety performance of highway segments. The decision to use jerk comes from previous research which utilized high magnitude jerk events as crash surrogate, or near-crash events. Instead of using jerk for near-crash analysis, the measurement of jerk was utilized to determine the percentage of GPS data observed below a certain negative jerk threshold for several highway segments. These segments were ¼-mile and ½-mile long. The preliminary exploration was conducted with 39 ¼-mile long segments of US Highway 101 within the city limits of San Luis Obispo. First, Pearson’s correlation coefficients were estimated for rate of ‘high’ jerk occurrences on these highway segments (with definitions of ‘high’ depending on varying jerk thresholds) and an estimate of crash rates based on long-term historical crash data. The trends in the correlation coefficients as the thresholds were varied led to conducting further analysis based on a jerk threshold of -2 ft./sec3 for the ¼-mile segment analysis and -1 ft./sec3 for the ¼-mile segment analysis. Through a negative binomial regression model, it was shown that utilizing the derived jerk percentage measure showed a significant correlation with the total number of historical crashes observed along US Highway 101. Analysis also showed that other characteristics of the roadway, including presences of a curve, presence of weaving (indicated by the presence of auxiliary lanes), and average daily traffic (ADT) did not have a significant correlation with observed crashes. Similar analysis was repeated for 19 ½-mile long segments in the same study area, and it was found the percentage of high negative jerk metric was again significant with historical crashes. The ½-mile negative binomial regression for the presence of curve was also a significant variable; however the standard error for this determination was very high due to a low sample size of analysis segments that did not contain curves. Results of this research show the potential benefit that naturalistic GPS driving data can provide for long-term traffic safety analysis, even if data is unaccompanied with any additional data (such as live video feed) collected with expensive vehicle instrumentation. The methodologies of this study are repeatable with many GPS devices found in certain consumer electronics, including many newer smartphones.

This study is the first examination on the association between seven types of sleep disorder and driving risk using large-scale naturalistic driving study data involving more than 3,400 participants. Regression analyses revealed that females with restless leg syndrome or sleep apnea and drivers with insomnia, shift work sleep disorder, or periodic limb movement disorder are associated with significantly higher driving risk than other drivers without those conditons. Furthermore, despite a small number of observations, there is a strong indication of increased risk for narcoleptic drivers. The findings confirmed results from simulator and epidemiological studies that the driving risk increases amongst people with certain types of sleep disorders. However, this study did not yield evidence in naturalistic driving settings to confirm significantly increased driving risk associated with migraine in prior research. The inconsistency may be an indication that the significant decline in cognitive performance among drivers with sleep disorders observed in laboratory settings may not nessarily translate to an increase in actual driving risk. Further research is necessary to define how to incentivize drivers with specific sleep disorders to balance road safety and personal mobility. / Master of Science

This dissertation investigates novel statistical approaches for causal effect estimation in observational settings, where controlled experimentation is infeasible and confounding is the main hurdle in estimating causal effect. As such, deconfounding constructs the main subject of this dissertation, that is (i) to restore the covariate balance between treatment groups and (ii) to attenuate spurious correlations in training data to derive valid causal conclusions that generalize. By incorporating ideas from representation learning, adversarial matching, generative causal estimation, and invariant risk modeling, this dissertation establishes a causal framework that balances the covariate distribution in latent representation space to yield individualized estimations, and further contributes novel perspectives on causal effect estimation based on invariance principles. The dissertation begins with a systematic review and examination of classical propensity score based balancing schemes for population-level causal effect estimation, presented in Chapter 2. Three causal estimands that target different foci in the population are considered: average treatment effect on the whole population (ATE), average treatment effect on the treated population (ATT), and average treatment effect on the overlap population (ATO). The procedure is demonstrated in a naturalistic driving study (NDS) to evaluate the causal effect of cellphone distraction on crash risk. While highlighting the importance of adopting causal perspectives in analyzing risk factors, discussions on the limitations in balance efficiency, robustness against high-dimensional data and complex interactions, and the need for individualization are provided to motivate subsequent developments. Chapter 3 presents a novel generative Bayesian causal estimation framework named Balancing Variational Neural Inference of Causal Effects (BV-NICE). Via appealing to the Robinson factorization and a latent Bayesian model, a novel variational bound on likelihood is derived, explicitly characterized by the causal effect and propensity score. Notably, by treating observed variables as noisy proxies of unmeasurable latent confounders, the variational posterior approximation is re-purposed as a stochastic feature encoder that fully acknowledges representation uncertainties. To resolve the imbalance in representations, BV-NICE enforces KL-regularization on the respective representation marginals using Fenchel mini-max learning, justified by a new generalization bound on the counterfactual prediction accuracy. The robustness and effectiveness of this framework are demonstrated through an extensive set of tests against competing solutions on semi-synthetic and real-world datasets. In recognition of the reliability issue when extending causal conclusions beyond training distributions, Chapter 4 argues ascertaining causal stability is the key and introduces a novel procedure called Risk Invariant Causal Estimation (RICE). By carefully re-examining the relationship between statistical invariance and causality, RICE cleverly leverages the observed data disparities to enable the identification of stable causal effects. Concretely, the causal inference objective is reformulated under the framework of invariant risk modeling (IRM), where a population-optimality penalty is enforced to filter out un-generalizable effects across heterogeneous populations. Importantly, RICE allows settings where counterfactual reasoning with unobserved confounding or biased sampling designs become feasible. The effectiveness of this new proposal is verified with respect to a variety of study designs on real and synthetic data. In summary, this dissertation presents a flexible causal inference framework that acknowledges the representation uncertainties and data heterogeneities. It enjoys three merits: improved balance to complex covariate interactions, enhanced robustness to unobservable latent confounders, and better generalizability to novel populations. / Doctor of Philosophy / Reasoning cause and effect is the innate ability of a human. While the drive to understand cause and effect is instinct, the rigorous reasoning process is usually trained through the observation of countless trials and failures. In this dissertation, we embark on a journey to explore various principles and novel statistical approaches for causal inference in observational studies. Throughout the dissertation, we focus on the causal effect estimation which answers questions like ``what if" and ``what could have happened". The causal effect of a treatment is measured by comparing the outcomes corresponding to different treatment levels of the same unit, e.g. ``what if the unit is treated instead of not treated?". The challenge lies in the fact that i) a unit only receives one treatment at a time and therefore it is impossible to directly compare outcomes of different treatment levels; ii) comparing the outcomes across different units may involve bias due to confounding as the treatment assignment potentially follows a systematic mechanism. Therefore, deconfounding constructs the main hurdle in estimating causal effects. This dissertation presents two parallel principles of deconfounding: i) balancing, i.e., comparing difference under similar conditions; ii) contrasting, i.e., extracting invariance under heterogeneous conditions. Chapter 2 and Chapter 3 explore causal effect through balancing, with the former systematically reviews a classical propensity score weighting approach in a conventional data setting and the latter presents a novel generative Bayesian framework named Balancing Variational Neural Inference of Causal Effects(BV-NICE) for high-dimensional, complex, and noisy observational data. It incorporates the advance deep learning techniques of representation learning, adversarial learning, and variational inference. The robustness and effectiveness of the proposed framework are demonstrated through an extensive set of experiments. Chapter 4 extracts causal effect through contrasting, emphasizing that ascertaining stability is the key of causality. A novel causal effect estimating procedure called Risk Invariant Causal Estimation(RICE) is proposed that leverages the observed data disparities to enable the identification of stable causal effects. The improved generalizability of RICE is demonstrated through synthetic data with different structures, compared with state-of-art models. In summary, this dissertation presents a flexible causal inference framework that acknowledges the data uncertainties and heterogeneities. By promoting two different aspects of causal principles and integrating advance deep learning techniques, the proposed framework shows improved balance for complex covariate interactions, enhanced robustness for unobservable latent confounders, and better generalizability for novel populations.

While driver distraction has been extensively studied in laboratory and empirical field studies, the prevalence of driver distraction on our nation's highways and the relative crash risk is unknown. It has recently become technologically feasible to conduct unobtrusive large-scale naturalistic driving studies as the costs and size of computer equipment and sensor technology have both dramatically decreased. A large-scale naturalistic driving study was conducted using 100 instrumented vehicles (80 privately-owned and 20 leased vehicles). This data collection effort was conducted in the Washington DC metropolitan area on a variety of urban, suburban, and rural roadways over a span of 12-13 months. Five channels of video and kinematic data were collected on 69 crashes and 761 near-crashes during the course of this data collection effort. The analyses conducted here are the first to establish direct relationships between driving inattention and crash and near-crash involvement. Relative crash risk was calculated using both crash and near-crash data as well as normal, baseline driving data, for various sources of inattention. Additional analyses investigated the environmental conditions drivers choose to engage in secondary tasks or drive fatigued, assessed whether questionnaire data were indicative of an individual's propensity to engage in inattentive driving, and examined the impact of driver's eyes off the forward roadway. The results indicated that driving inattention was a contributing factor in 78% of all crashes and 65% of all near-crashes. Odds ratio calculations indicated that fatigued drivers have a 4 times higher crash risk than alert drivers. Drivers engaging in visually and/or manually complex tasks are at 7 times higher crash risk than alert drivers. There are specific environmental conditions in which engaging in secondary tasks or driving fatigued is deemed to be more dangerous, including intersections, wet roadways, undivided highways, curved roadways, and driving at dusk. Short, brief glances away from the forward roadway for the purpose of scanning the roadway environment (e.g., mirrors and blind spots) are safe and decrease crash risk, whereas such glances that total more than 2 seconds away from the forward roadway are dangerous and increase crash risk by 2 times over that of more typical driving. / Ph. D.

Motor vehicle crashes are one of the leading causes of death in the United States. Traffic safety research targets at understanding the cause of crash, preventing the crash, and mitigating crash severity. This dissertation focuses on the driver-related traffic safety issues, in particular, on developing and implementing contemporary statistical modeling techniques on driving risk research on Naturalistic Driving Study data. The dissertation includes 5 chapters. In Chapter 1, I introduced the backgrounds of traffic safety research and naturalistic driving study. In Chapter 2, the state-of-practice statistical methods were implemented on individual driver risk assessment using NDS data. The study showed that critical-incident events and driver demographic characteristics can serve as good predictors for identifying risky drivers. In Chapter 3, I developed and evaluated a novel Bayesian random exposure method for Poisson regression models to account for situations where the exposure information needs to be estimated. Simulation studies and real data analysis on Cellphone Pilot Analysis study data showed that, random exposure models have significantly better model fitting performances and higher parameter coverage probabilities as compared to traditional fixed exposure models. The advantage is more apparent when the values of Poisson regression coefficients are large. In Chapter 4, I performed comprehensive simulation-based performance analyses to investigate the type-I error, power and coverage probabilities on summary effect size in classical meta-analysis models. The results shed some light for reference on the prospective and retrospective performance analysis in meta-analysis research. In Chapter 5, I implemented classical- and Bayesian-approach multi-group hierarchical models on 100-Car data. Simulation-based retrospective performance analyses were used to investigate the powers and parameter coverage probabilities among different hierarchical models. The results showed that under fixed-effects model context, complex secondary tasks are associated with higher driving risk. / Ph. D.

The importance of traffic safety has prompted considerable research on predicting driver risk and evaluating the impact of risk factors. Driver risk modeling is challenging due to the rarity of motor vehicle crashes and heterogeneity in individual driver risk. Statistical modeling and analysis of such driver data are often associated with Big Data, considerable noise, and lacking informative predictors. This dissertation aims to develop several systematic techniques for traffic safety modeling, including finite sample bias correction, decision-adjusted modeling, and effective risk factor construction. Poisson and negative binomial regression models are primary statistical analysis tools for traffic safety evaluation. The regression parameter estimation could suffer from the finite sample bias when the event frequency (e.g., the total number of crashes) is low, which is commonly observed in safety research. Through comprehensive simulation and two case studies, it is found that bias adjustment can provide more accurate estimation when evaluating the impacts of crash risk factors. I also propose a decision-adjusted approach to construct an optimal kinematic-based driver risk prediction model. Decision-adjusted modeling fills the gap between conventional modeling methods and the decision-making perspective, i.e., on how the estimated model will be used. The key of the proposed method is to enable a decision-oriented objective function to properly adjust model estimation by selecting the optimal threshold for kinematic signatures and other model parameters. The decision-adjusted driver-risk prediction framework can outperform a general model selection rule such as the area under the curve (AUC), especially when predicting a small percentage of high-risk drivers. For the third part, I develop a Multi-stratum Iterative Central Composite Design (miCCD) approach to effectively search for the optimal solution of any "black box" function in high dimensional space. Here the "black box" means that the specific formulation of the objective function is unknown or is complicated. The miCCD approach has two major parts: a multi-start scheme and local optimization. The multi-start scheme finds multiple adequate points to start with using space-filling designs (e.g. Latin hypercube sampling). For each adequate starting point, iterative CCD converges to the local optimum. The miCCD is able to determine the optimal threshold of the kinematic signature as a function of the driving speed. / Doctor of Philosophy / When riding in a vehicle, it is common to have personal judgement about whether the driver is safe or risky. The drivers’ behavior may affect your opinion, for example, you may think a driver who frequently hard brakes during one trip is a risky driver, or perhaps a driver who almost took a turn too tightly may be deemed unsafe, but you do not know how much riskier these drivers are compared to an experienced driver. The goal of this dissertation is to show that it is possible to quantify driver risk using data and statistical methods. Risk quantification is not an easy task as crashes are rare and random events. The wildest driver may have no crashes involved in his/her driving history. The rareness and randomness of crash occurrence pose great challenges for driver risk modeling. The second chapter of this dissertation deals with the rare-event issue and provides more accurate estimation. Hard braking, rapid starts, and sharp turns are signs of risky driving behavior. How often these signals occur in a driver’s day-to-day driving reflects their driving habits, which is helpful in modeling driver risk. What magnitude of deceleration would be counted as a hard brake? How hard of a corner would be useful in predicting high-risk drivers? The third and fourth chapter of this dissertation attempt to find the optimal threshold and quantify how much these signals contribute to the assessment of the driver risk. In Chapter 3, I propose to choose the threshold based on the specific application scenario. In Chapter 4, I consider the threshold under different speed limit conditions. The modeling and results of this dissertation will be beneficial for driver fleet safety management, insurance services, and driver education programs.

Forward Collision Avoidance Systems (FCAS) are a type of active safety system which have great potential for rear-end collision avoidance. These systems use either radar, lidar, or cameras to track objects in front of the vehicle. In the event of an imminent collision, the system will warn the driver, and, in some cases, can autonomously brake to avoid a crash. However, driver acceptance of the systems is paramount to the effectiveness of a FCAS system. Ideally, FCAS should only deliver an alert or intervene at the last possible moment to avoid nuisance alarms, and potentially have drivers disable the system. A better understanding of normal driving behavior can help designers predict when drivers would normally take avoidance action in different situations, and customize the timing of FCAS interventions accordingly. The overall research object of this dissertation was to characterize normal driver behavior in car following events based on naturalistic driving data. The dissertation analyzed normal driver behavior in car-following during both braking and lane change maneuvers. This study was based on the analysis of data collected in the Virginia Tech Transportation Institute 100-Car Naturalistic Driving Study which involved over 100 drivers operating instrumented vehicles in over 43,000 trips and 1.1 million miles of driving. Time to Collision in both braking and lane change were quantified as a function of vehicle speed and driver characteristics. In general, drivers were found to brake and change lanes more cautiously with increasing vehicle speed. Driver age and gender were found to have significant influence on both time to collision and maximum deceleration during braking. Drivers age 31-50 had a mean braking deceleration approximately 0.03 g greater than that of novice drivers (age 18-20), and female drivers had a marginal increase in mean braking deceleration as compared to male drivers. Lane change maneuvers were less frequent than braking maneuvers. Driver-specific models of TTC at braking and lane change were found to be well characterized by the Generalized Extreme Value distribution. Lastly, driver's intent to change lanes can be predicted using a bivariate normal distribution, characterizing the vehicle's distance to lane boundary and the lateral velocity of the vehicle. This dissertation presents the first large scale study of its kind, based on naturalistic driving data to report driver behavior during various car-following events. The overall goal of this dissertation is to provide a better understanding of driver behavior in normal driving conditions, which can benefit automakers who seek to improve FCAS effectiveness, as well as regulatory agencies seeking to improve FCAS vehicle tests. / Ph. D.