Improving safety estimation and prediction using multivariate regression models in observational road safety studies

Update Item Information
Title Improving safety estimation and prediction using multivariate regression models in observational road safety studies
Publication Type dissertation
School or College College of Engineering
Department Civil & Environmental Engineering
Author Musunuru, Anusha
Date 2017
Description Observational studies are a frequently used "tool" in the field of road safety research because random assignments of safety treatments are not feasible or ethical. Data and modeling issues and challenges often plague observational road safety studies, and impact study results. The objective of this research was to explore a selected number of current data and modeling limitations in observational road safety studies and identify possible solutions. Three limitations were addressed in this research: (1) a majority of statistical road safety models use average annual daily traffic (AADT) to represent traffic volume and do not explicitly capture differences in traffic volume patterns throughout the day, even though crash risk is known to change by time of day, (2) statistical road safety models that use AADT on the "right-hand side" of the model equation do not explicitly account for the fact that these values for AADT are estimates with estimation errors, leading to potential bias in model estimation results, and (3) the current state-of-the-practice in road safety research often involves "starting over" with each study, choosing a model functional form based on the data fit, and letting the estimation results drive interpretations, without fully utilizing previous study results. These limitations were addressed by: (1) estimating the daily traffic patterns (by time of day) using geo-spatial interpolation methods, (2) accounting for measurement error in AADT estimates using measurement error models of expected crash frequency, and (3) incorporating prior knowledge on the safety effects of explanatory variables into regression models of expected crash frequency through informative priors in a Bayesian methodological framework. These alternative approaches to address the selected observational road safety study limitations were evaluated using data from rural, two-lane highways in the states of Utah and Washington. The datasets consisted of horizontal curve segments, for which crash data, roadway geometric features, operational characteristics, roadside features, and weather data were obtained. The results show that the methodological approaches developed in this research will allow road safety researchers and practitioners to accurately evaluate the expected road safety effects. These methods can further be used to increase the accuracy and repeatability of study results, and ultimately expand the current practice of evaluating regression models of expected crash frequency in observational road safety studies.
Type Text
Publisher University of Utah
Subject Statistics; Civil engineering; Transportation planning
Dissertation Name Doctor of Philosophy
Language eng
Rights Management (c) Anusha Musunuru
Format application/pdf
Format Medium application/pdf
ARK ark:/87278/s64219qz
Setname ir_etd
ID 1407488
Reference URL https://collections.lib.utah.edu/ark:/87278/s64219qz
Back to Search Results