Prioritizing Highway Safety Manual's crash prediction variables using boosted regression trees

Area of Science:

• Transportation Engineering
• Traffic Safety Research
• Data Mining in Civil Engineering

Background:

• The Highway Safety Manual (HSM) recommends empirical Bayes (EB) methods for safety performance prediction, requiring local calibration factors.
• Deriving these factors necessitates extensive, detailed roadway characteristic data, much of which is unavailable or costly to maintain in state databases.
• Prioritizing variables by their impact on crash predictions is essential for efficient data collection and management.

Purpose of the Study:

• To determine the impact of each independent variable in the HSM on crash predictions.
• To identify influential variables for targeted data collection and maintenance.
• To address the data limitations hindering the effective application of HSM safety prediction methods.

Main Methods:

• Utilized boosted regression trees (BRT), a data mining technique, to analyze variable associations with crash predictions.
• BRT effectively handles diverse predictor types, complex non-linear relationships, and computes variable importance.
• Analyzed five years (2008-2012) of crash data for two urban/suburban arterial types: two-lane undivided and four-lane divided.

Main Results:

• Identified that variables exhibit non-linear and complex relationships with predicted crash counts.
• A small subset of variables was found to explain a significant portion of the variation in crash data.
• Variable importance analysis revealed key predictors for crash frequency.

Conclusions:

• The study successfully identified critical variables influencing crash predictions, aligning with the need for data prioritization.
• Boosted regression trees proved effective in uncovering complex relationships and variable importance for traffic safety analysis.
• Findings support a more focused and cost-effective approach to data collection for enhancing transportation safety management.