Spatial Negative Binomial Bayesian Additive Regression Trees for Accident Hot Spot Identification

Abstract

statistics of the considered data set.

Variable Mean Std. Min. Max.

Crash count 16.5 22.5 0.0 214.0 Interstate highway (dummy) 0.5 0.5 0.0 1.0 Asphalt pavement (dummy) 0.2 #04 0.0 1.0 Rural area (dummy) 0.3 0.4 0.0 1.0 Asphalt shoulder (dummy) 0.6 0.5 0.0 1.0 Road condition score > 90 0.5 0.5 0.0 1.0 Truck traffic percentage 10.6 6.5 2.6 34.3 International Roughness Index (IRI) 115.5 35.2 35.2 319.0 Logarithm of Annual Avg. Daily Traffic (AADT) 95 0.6 7.7 10.8 Speed limit in MPH 61.3 4.9 55.0 70.0 Left shoulder width < 10 ft 0.5 0.5 0.0 1.0 Right shoulder width < 10 ft 04 0.5 0.0 1.0

Table 1: Description of data set

3.2 Evaluation of site ranking performance

Numerous criteria for the ranking of hazardous sites have been proposed (see Buddhavarapu, 2015, for a review). In the current application, we rank sites based on the posterior probability that a site belongs to the top 5% most hazardous sites (see Schmidt, 2012). The proposed negative binomial Bayesian additive regression trees (NB-BART) models is benchmarked against negative binomial regression models with spatial error terms and linear-in-parameters link function specification. We consider one model with fixed parameters (NB with fixed parameters) as well as a model whose

‘Due to data privacy issues, the exact location of the data collection cannot be disclosed.