Analysis of naturalistic driving event data: Omitted-variable bias and multilevel modeling approaches

Naturalistic driving studies have been conducted over the past 5 years or more and have commonly reviewed video and kinematic data to identify and analyze crash, near-crash, and critical-incident events. But statistical methods that are applicable to these event data are needed. This paper addresses two issues in model development for naturalistic driving event data: the test for omitted-variable bias and the exploration of the advantages of hierarchical model structures in data analysis. With roadway departure event data from the 100-Car Naturalistic Driving Study conducted at Virginia Tech Transportation Institute, Blacksburg, Virginia, logit models were used to estimate the probability that a crash or a near crash would occur, rather than a critical incident. The models indicated a substantial omitted-variable bias for estimation of the effect of context variables but little difference for driver variables. These tests indicated that modeling of naturalistic event data should have included variables that described the attributes of the event, the driver, and the context to reduce the likelihood of bias. Hierarchical model structures offer the advantage of driver-level predictors to parameterize the effects of event attributes and contexts. The models thus reflect how driver decisions are executed: drivers with particular characteristics (one level) find themselves in contexts in which they execute specific driving maneuvers (second level), which lead to certain outcomes. Suggestions for further research include testing with additional data sets and potential applications to analysis of crash surrogates.