Off-campus South Dakota State University users: To download campus access theses, please use the following link to log into our proxy server with your South Dakota State University ID and password.

Non-South Dakota State University users: Please talk to your librarian about requesting this thesis through interlibrary loan.

Document Type

Thesis - University Access Only

Award Date


Degree Name

Master of Science (MS)

Department / School

Civil and Environmental Engineering

First Advisor

Xiao Qin

Second Advisor

Gemechis Djira


This thesis investigated two crash modelling techniques: following and modifying Highway Safety Manual (HSM) predictive methods, and developing user-defined predictive models. HSM predictive methods were applied on five rural intersection facility types on South Dakota state highways to calculate the predicted crash count of each intersection. The analysis on the prediction results revealed large deviations between intersections, and the large deviations were caused by the fact that HSM methods were not modified to fit the South Dakota data. Then modification was conducted on HSM predictive methods of two rural two-lane two-way intersection types using the South Dakota data. The modified methods outperformed the original methods by providing calibration factors closer to one. And the goodness-of-fit of two methods were compared and it shows that both methods present similar overall prediction performance. The second aspect of the thesis was focused on the development of the multivariate Poisson-lognormal (MVPLN) regression model. A total of 582 four-leg stop control intersections on South Dakota state highways were studied with crash records from 2008-2012. The crashes were divided into three crash types by relative travelling direction: same direction, intersection direction and single vehicle crashes. The MVPLN aimed to simultaneously model all three crash types to account for the correlation xiii between crash types within same intersections. The MVPLN model estimations were obtained using the Markov chain Monte Carlo (MCMC) simulation techniques. Significant positive correlations were found between crash types. This finding indicates the existence of unobserved common factors simultaneously affecting multiple crash types in the same manner. Compared to univariate models including the univariate Poisson (UVP) and the univariate negative binomial (UVNB) models, the MVPLN model provided very consistent significant variables and very similar estimations of shared significant variables. But the MVPLN model is superior to the other two models in prediction performance by presenting better goodness-of-fit.

Library of Congress Subject Headings

Traffic accidents--Mathematical models
Traffic accidents--Forecasting--Mathematical models
Roads--South Dakota--Interchanges and intersections


Includes bibliographical references (pages 85-91)



Number of Pages



South Dakota State University



Rights Statement

In Copyright