Abstract

As part of a ISU statistics deptartment. creative component in cooperation with Iowa DOT's Office of Traffic and Safety, a full Bayes analysis of the reduction in crash frequency due to four-lane to three-lane conversions in Iowa was conducted. This study utilized monthly crash data and estimated volumes obtained from the Office of Transportation Safety for 30 sites, 15 treatment and 15 control, over 23 years (1982-2004). The sites had volumes ranging from 2030 to 15350 weighted AADT during that timespan and were mostly located in smaller urbanized areas (the exception being one location in Des Moines). Each treatment site had different known intervention dates; therefore, the number of before and after crash records varied from site to site. Individual control sites were matched to each treatment site to provide a control sample similar to the treatment sample. In general, both treatment and control site crash history can be seen to experience a reduction. However, the reduction in treatment site crash frequency and rate after intervention are significantly more marked than that at the comparison sites. This differs from a previous Huang, et al. study whose data, even from a descriptive statistics standpoint, indicated very little reduction or difference between the two groups. Additionally, because monthly crash frequencies were used for analysis, it was possible to account for the seasonality effects on crashes, which should be expected given the seasonal weather patterns in Iowa. Given the random and rare nature of crash events, a hierarchical Poisson model where the log mean was expressed as a function of time period, seasonal effects, and a random effect corresponding to each site included was fitted to the crash frequencies. We adopted a full Bayesian (FB) approach for estimation of model parameters. Results indicate a 25 percent reduction in crash frequency per mile and a 19 percent reduction in crash rate. This differs again from the Huang, et al. study which reported a 6 percent reduction in crash frequency per mile and an insignificant indication for crash rate effects. This difference is evident just by comparing the raw data from the two studies. The Iowa data, when graphed, indicates marked reductions whereas the Huang, et al. data indicate very little difference. Based on these Iowa FB results and results from a simple before/after analysis done as part of the causal study included in this paper, we are comfortable with the 25 percent and 19 percent reductions, especially as they fit practitioner expectations.