Home About ATSIP Program Awards Venue Sponsors/Exhibitors

Causal Analysis of Contributing Factors to Freeway Crashes


Abstract: The wide deployment of advanced transportation information systems (ATIS) has made the collection, storage, and processing of real-time traffic data readily available. Now researchers can gather real-time information pertaining to crash occurrence. Based on real-time traffic data, various real-time crash prediction models (RTCPM) have been proposed to identify the contributing factors of crashes. Crash is usually considered as a binary variable (yes/no) in almost all real-time crash prediction studies. A crash case represents the traffic conditions prior to a crash, while a non-crash case represents crash-free traffic conditions. The traffic condition in a short time interval before a crash is determined by reviewing reported crash time and location. Any traffic condition that is unaffected by crashes could be a non-crash case. Crashes are rare events and therefore, while non-crash cases are large in volume. Most previous studies randomly sample non-crash cases, either by matched case-control design or unmatched design.

The matched case-control design compares the level of risk factors in two similar groups, one that includes the outcome and one that does not. In a matched case-control design, each crash is considered as a “case”, and non-crash cases are treated as “controls” by matching confounding factors that are related to both the crash probability and the traffic variables of interest. The matched case-control design is used to remove the noise of confounding factors and investigate the underpinning risk factors. Although the matched case-control design is expected to increase the accuracy of variable estimates in a crash prediction model, most studies only treated non-traffic variables such as weather, location, and time as confounding factors while overlooking the potential confounding effect within traffic variables. Therefore, the true causal effects of one traffic factor would be compromised given the presence of other confounding traffic factors. As a result, the model findings can be biased or inaccurate.

The objective of this study is to measure the causal effects of traffic factors on the probability of a crash using the propensity score method. The propensity score method can eliminate the nuisance of confounding traffic factors related to one traffic factor that is being assessed. This is done by generating a sample of non-crash cases which have similar distributions of confounding traffic factors related to the traffic factor of interest. Then a binary logit model will be applied to assess the causal effect of that factor.

Zhi Chen -  is currently a research assistant in the Department of Civil & Environmental Engineering of the University of Wisconsin-Milwaukee. Mr. Chen has solid knowledge of statistics, data mining, machine learning, transportation engineering and transportation planning. Mr. Chen has authored and co-authored 5 journal papers and 4 conference papers. The focuses include the calibration of the Highway Safety Manual predictive methods, the model development for predicting freeway crashes using loop detector data, and the development of novel crash prediction model. Mr. Chen has also worked as a key researcher in several projects including evaluating local and tribal rural road design, calibrating the Highway Safety Manual predictive methods for rural roadway segments and intersections, and developing faulty loop detection tools.



Quick Links


Chris Osbourn


fb_logo   linkedin