Assessment of the Effects of Volume Completeness and Spatial and Temporal Correlation on Hourly Freeway Crash Prediction Models

Abstract

Traditional traffic safety analyses of crash frequency usually use highly aggregated cross-sectional data and ignore the time-varying nature of some critical factors. This research used 7 years of hourly data from 110 rural four-lane segments and 80 urban six-lane segments to develop hourly level crash prediction models and contrasted them with traditional annual average daily traffic (AADT)-based models. To account for the overdispersion of data and unobserved heterogeneity, generalized linear mixed-effect models were contrasted with negative binomial models. The models used average hourly volume as a measure of exposure, and the quantity of volume data available for the sites ranged from continuous counts to locations where only a couple of weeks of data were available every other year (short counts). While developing disaggregated models, the difference in data availability from these sources can be a potential source of error, so evaluating the change in performance of prediction models with changes in volume data availability was examined. The results showed that the best models include a combination of average hourly volume, selected geometric variables, and speed related parameters. Hourly models that included speed parameters consistently outperformed AADT models. Further investigation revealed that the positive effect of using a more inclusive and larger dataset was larger than the effect of accounting for data correlation. This showed that using short count stations as a data source does not diminish the quality of the developed models, thus indicating that these methods could be applied broadly across agencies, even when volume data is relatively sparse.