Report numberRA-MOW-2008-004
TitleA review of accident prediction models for road intersections
AuthorsBetty Nambuusi
Tom Brijs
Elke Hermans
Published byPolicy Research Centre for Mobility and Public Works, track Traffic Safety 2007-2011
Number of pages69
Document languageEnglish
Partner(s)Universiteit Hasselt
Work packageOther: Infrastructure and space

The objective of this report is to review accident prediction models for intersections used in literature to identify which variables have a significant effect on accident occurrence so that we can have a starting point for future research. Several models have been reviewed including multiple logistic regression, multiple linear regression, Poisson models, negative binomial models, random effects models and, classification and regression trees (CART) technique. The data, methodology and results of several studies are described. The direction of the effect of several significant explanatory variables is discussed and recommendations are made.

Different APMs for different intersection types and accident types have been developed in the literature. It is recommended that fitting separate models for different intersection types and accident types gives a better fit and description of the data than one model for all intersection types. Provided data on intersection and accident types are available, it is recommended to fit disaggregated models rather than aggregated models (Reurings et al., 2005; Turner and Nicholson, 1998).

Although similar techniques were applied on rural and urban road intersections, in literature a different model structure was used.

The elasticity shows the percentage change in the expected number of accidents associated with a 1% change in traffic volume. The effects of risk factors that influence the probability of accidents given exposure are modelled as an exponential function. The choice of an exponential form is logical in the view of the characteristics of the Poisson distribution since accident counts are positive and rare events at intersections (Reurings et al., 2005).

However, the choice of the model depends on the nature of the response and the objective of the research. If interest is in making inference on the entire population, population average based models (chapters 2, 3 and 4) are suitable. In contrast, researchers interested in location specific inference would opt for random effects models (chapter 5). Researchers who wish to group accidents based on particular criteria, the CART is a credible choice (chapter 6).

The variables annual average daily traffic (AADT) on major and minor roads, total vehicle counts and pedestrians crossing all arms, lighting and signal timing were statistically significant in most models. Therefore, it is desirable that APMs for intersections include these variables. The other variables are listed in chapter seven of the report. Generally, atleast one explanatory variable in the categories of traffic flow, traffic control, geometry, driver characteristics, vehicle type or features, environmental factors and land use had a significant effect on accident occurrence. Therefore, all categories are essential in predicting accidents at intersections.

DownloadPDF icon RA-MOW-2008-004.pdf


The Policy Research Centre for Traffic Safety carries out policy relevant scientific research under the authority of the Flemish Government. The Centre is the result of a

cooperation between Hasselt University, KU Leuven and VITO, the Flemish Institute for Technological Research.


Leuven vito