Date of Completion


Embargo Period



Traffic Safety, Crash Prediction Modeling, Crash Typology, Pedestrian Safety, Safety Performance Function, Freeway, Fatal Crash, Negative Binomial Modeling, Partial Proportional Odds

Major Advisor

John N Ivan

Associate Advisor

Nalini Ravishanker

Associate Advisor

Nicholas E Lownes

Field of Study

Civil Engineering


Doctor of Philosophy

Open Access

Open Access


The first part of the research presents an investigation of pedestrian conflicts and crash count models to learn which exposure measures and roadway or roadside characteristics significantly influence pedestrian safety at road crossings. The results show that minor and serious conflicts are marginally significant in predicting total pedestrian crashes together with crossing distance and building setback. This suggests that these conflicts may be a good surrogate for crashes in analyzing pedestrian safety. Greater crossing distance and small building setbacks are both found to be associated with larger numbers of pedestrian-vehicle crashes. In the second part of the research we assembled crash and roadway geometry data of freeways in the State of Connecticut for developing Safety performance functions (SPFs). Models were estimated separately for single vehicle and multi-vehicle crashes. Interaction models were found to be the best models for all crash categories. This finding suggests the importance of incorporating interaction effect between variables, in particular between speed limit and geometric variables such as number of lanes, shoulder width, and median type, during crash prediction model estimation. Last part of the research presents an investigation to find a preferred crash typology for the prediction of crash severities for controlled access highways. We find that the typology based on vehicle travel direction has better fit than the other models. The finding demonstrates that the crash types and AADT could be good predictor of crash severities when crash and person related information are not available, as is the case for segment level prediction.