Consideration of Driver Demographics, Behavior and Psychological Attributes in Traffic Safety Analysis

Crash risk depends on many factors, with driver factors being the most significant of all. Traditional crash risk models are usually based on roadway and traffic data including roadway geometry, roadway condition, and traffic volume. To develop a good crash prediction model and maximize its prediction accuracy, it is required to include all factors that best predict crashes. Therefore, in addition to the traditional roadway, environment and traffic data, driver factors that include driver demographics, behavior, and psychological attributes would be beneficial to consider in the crash risk estimation. Driver demographics are challenging to obtain from conventional data sources due to unavailability. Past studies have estimated crash risk considering driver exposure by their demographics at an aggregated level such as census tract or county. It is assumed that the population of the geographic region and the driving population have similar population trends. This dissimilarity might represent the driving population inaccurately. Some studies have tended to use driver demographic information to explain crash risk but overlooked driver psychological, cognitive, and physical characteristics, which can also be important for crash risk estimation. Crash data from police reports cannot provide driver characteristics beyond simple demographics or driving kinematics and are limited to only reportable crashes. Fortunately, naturalistic driving studies (NDS) offer unique opportunities to get information about driver attributes, behavior, and other pre-crash factors in predicting crash occurrences. This dissertation includes three studies aimed at incorporating demographic and vehicle factors in crash risk estimation using readily available crash data along with SHRP2 NDS data.
The first study proposes and implements a technique to obtain demographic proportions to incorporate in count models as an exposure at each site by aggregating similar adjacent sites until significant demographic proportions are obtained. Driver gender, age and vehicle type information are obtained using a proven technique called Quasi-Induced Exposure (QIE) using five years (2010-2014) of crash data along with road inventories for 1264 urban four-lane divided highway segments in California. Count models including only roadway and traffic factors were compared with models also including driver demographics and vehicle type. The latter outperformed the former in terms of mean prediction bias (MPB) and mean absolute deviation (MAD) statistics on holdout sample predictions. Results indicate that teen drivers are more crash prone in total and fatal plus injury severity crashes, whereas senior driver crash risk increases with the increase in severity level. Presence of vehicles other than passenger cars and trucks reduces total and property damage only crash counts. Female drivers exhibit an increase in total and fatal plus injury crash counts.
The second study estimates NDS event-oriented models to evaluate the interaction between driver attributes and roadway-environmental factors for predicting safety critical events. A latent class clustering approach was used to uncover categories of drivers by psychological, perceptual, and cognitive characteristics, as well as driving experience. Results revealed five driver types: risk-taker, careful-impaired, careful-unimpaired, and distractible. These types were incorporated in mixed-effects binary logistic models along with roadway, traffic, and environmental variables to estimate and predict the risk of any event having a near crash or crash. Models including driver factors predicted better than those without them. “Risk-taker” drivers showed the highest probability of being in crashes. However, “Careful-impaired” drivers—i.e., those who had impairments in identifying the location of another vehicle, visualizing missing information, visual-spatial perception, and executive functioning—posed higher crash risk in roadway conditions such as snow, lack of lane markings, and certain traffic control and operating conditions. The results point to novel avenues for educational and behavioral interventions to improve road safety.
The third study estimates trip-oriented models to investigate the effect of driving behavior on crash/near-crash occurrences by including specific driving actions that are expected to be related to risky driving, including extreme accelerating, decelerating, turning, and lane changing behavior during each trip. The driver attributes explored from the latent class clustering approach are used again here to include in the trip-oriented model. Separately, k-means clustering on trip-level data is added in this effort to represent contextual driver behavior scenarios. This study incorporates driver attribute clusters and contextual driver behavior clusters and the interactive effects between the attributes and behaviors clusters. Trip based modeling results show how different contextual behaviors for these driver types posed comparatively higher risk. The results point to novel avenues for educational and behavioral interventions to improve road safety and its applicability to automated and connected vehicle technologies.