Semiparametric Bayesian models for evaluating time-variant driving risk factors using naturalistic driving data and case-crossover approach.

Abstract

Driver behavior is a major contributing factor for traffic crashes, a leading cause of death and injury in the United States. The naturalistic driving study (NDS) revolutionizes driver behavior research by using sophisticated nonintrusive in-vehicle instrumentation to continuously record driving data. This paper uses a case-crossover approach to evaluate driver-behavior risk. To properly model the unbalanced and clustered binary outcomes, we propose a semiparametric hierarchical mixed-effect model to accommodate both among-strata and within-stratum variations. This approach overcomes several major limitations of the standard models, eg, constant stratum effect assumption for conditional logistic model. We develop 2 methods to calculate the marginal conditional probability. We show the consistency of parameter estimation and asymptotic equivalence of alternative estimation methods. A simulation study indicates that the proposed model is more efficient and robust than alternatives. We applied the model to the 100-Car NDS data, a large-scale NDS with 102 participants and 12-month data collection. The results indicate that cell phone dialing increased the crash/near-crash risk by 2.37 times (odds ratio: 2.37, 95% CI, 1.30-4.30) and drowsiness increased the risk 33.56 times (odds ratio: 33.56, 95% CI, 21.82-52.19). This paper provides new insight into driver behavior risk and novel analysis strategies for NDS studies.