Crash severity along rural mountainous highways in Malaysia: An application of a combined decision tree and logistic regression model

Objective Traffic crashes along mountainous highways may lead to injuries and fatalities more often than highways on plain topography; however, research focusing on the injury outcome of such crashes is relatively scant. The objective of this study was to investigate the factors affecting the likelihood that traffic crashes along rural mountainous highways result in injuries. Method This study proposes a combination of decision tree and logistic regression techniques to model crash severity (injury vs non-injury), as the combined approach allows the specification of nonlinearities and interactions in addition to main effects. Both a Scobit model and a random parameters logit model, respectively accounting for an imbalance response variable and unobserved heterogeneities, are tested and compared. The study dataset contains a total of five years of crash data (2008-2012) on selected mountainous highways in Malaysia. To enrich the data quality, an extensive field survey was conducted to collect detailed information on horizontal alignment, longitudinal grades, cross-section elements and roadside features. In addition, weather condition data from the meteorology department was merged using the time stamp and proximity measures in AutoCAD-Geolocation. Results The random parameters logit model is found to outperform both the standard logit and Scobit models, suggesting the importance of accounting for unobserved heterogeneity in crash severity models. Results suggest that proportion of segment lengths with simple curves, presence of horizontal curves along steep gradients, highway segments with unsealed shoulders, and highway segments with cliffs along both sides are positively associated with injury-producing crashes along rural mountainous highways. Interestingly, crashes during rainy conditions are associated with crashes that are less likely to involve injury. It is also found that the likelihood of injury-producing crashes decreases for rear-end collisions but increases for head-on collisions and crashes involving heavy vehicles. A higher-order interaction suggests that single vehicle crashes involving light and medium sized vehicles are less severe along straight sections compared to road sections with horizontal curves. One the other hand, crash severity is higher when heavy vehicles are involved in crashes as single vehicles travelling along straight segments of rural mountainous highways. Conclusion In addition to unobserved heterogeneity, it is important to account for higher order interactions to have a better understanding of factors that influence crash severity. A proper understanding of these factors will help developing targeted countermeasures to improve road safety along rural mountainous highways