| Deer-vehicle collisions are a significant concern across the country, especially in states like Kansas, where most of the highway mileage is rural. In Kansas, accident records have traditionally been used to identify locations where deer collisions occur most frequently. The accident records can only be used to identify a high-risk location after many accidents have already occurred. A better understanding of the parameters most closely associated with deer-vehicle accident probability is necessary, so that countermeasures can be applied proactively. Additionally, a methodology is needed to compare the relative risk of segments so that they can be prioritized and countermeasures can be applied as cost-effectively as possible.; Forty-five predictor variables were considered, and the predicted variable was accidents per year per mile. Primary data were collected for 21 variables by visiting 123 segments located in 15 counties of the State of Kansas. The remaining 25 variables were obtained from secondary sources or generated using secondary data and a GIS.; To eliminate the multi among the predictor variables prior to regression analysis, Principal Component Analyses (PCA) were performed to help group variables or to identify principal components. The analyses were performed based on two different principles. Eleven principal components were identified by the first method and seven principal components were identified by the second. Two separate multiple linear stepwise regression analyses were performed using the identified principal components as predictor variables. One hundred one data points were used for model calibration and 22 data points were used for model validation.; The ten most significant parameters were identified. Wooded land area by the side of the roadway, number of lanes, median types, traffic volume, posted speed, clear width, number of bridges and/or visible culverts, roadside adjacent side slope, roadside topography in the transverse direction, presence of deer warning sign and traditional fencing were all found to be positively correlated. The calibrated first model consisted of five principal components and could predict with satisfactory accuracy. The second model was simpler, consisted of three principal components, and required less input data, but the predictive capability was less than the first model. |
