# Road geometry and accident risk

See updated notes on this work. What follows are notes on the initial stages of this project.

This was a project for Works Consultancy Services Limited (now Opus International Consultants) and was part of a larger project for Transit New Zealand, the authority in charge on New Zealand's state highways. The results are in a paper, Application of the Road Geometry Data Acquisition System, by Dave Wanty, Maurice McLarin, Robert Davies and Peter Cenek presented at the Seventh World Conference on Transport Research, Sydney, Australia, July, 1995. The data was prepared by Maurice McLaren of Works Consultancy Services. I carried out the analyses described here.

Works Consultancy Services had measured the average curvature, slope, camber, width, compass direction for each successive 400 metre section of the state highway system (rural, two lane, non-motorway only). As well as the raw road geometry data we have an advisory speed calculated from the geometry data. In addition, we have estimates of the average number of vehicles using each section (AADT). Counting each direction separately, we have 86,524 road segments with valid data.

We also have the number of reported road accidents on each of the 400 metre sections over a 5 year period. Only road accidents that could be related to road geometry were included. This left 2,356 accidents.

The objective was to relate the probability of an accident to road geometry data.

Two methods of analysis were used. The first was simply to form one and two way tables to present accident rates in terms of one or two geometry variables. The second was to carry out step-wise regression using glm analysis in S-plus for Poisson variables using the number of accidents in each road segment as the Y variable and the various geometry variables as the X variables. The AADT was included as an offset. I considered two subsets of the data, the first being limited to the higher speed and higher traffic roads.

I could identify and estimate the effect of curvature, wriggliness, and slope. Alternatively it could be related to the advisory speed in the current and surrounding road segments. Results are presented both graphically and as formulae. Also, curiously, there was an effect of compass direction. This effect must be related to the alignment of New Zealand's major roads, but we have been unable to explain it. I also searched for the effects of road surface type and road surface age and found some statistically significant results, but not really enough be really convincing or to get good estimates.

The major problem is errors in assigning both geometry measurements and accidents to particular road locations. Probably some smearing of the results has occurred. It should be possible to modify the analysis to better allow for this. It will probably require a special purpose C++ program. But so far I haven't managed to convince anyone to pay!

This project was much too big for S-plus on my 486 and I had to buy time on a Sparc-station belonging to Victoria University.

2004: I have now completed a revised analysis using data collected over 6 years - see the conference paper Crash risk relationships for improved safety management of roads and my report to Opus International Consultants.