01657015

Component

Although learning to drive happens at the individual level, the vast majority of analyses of motor vehicle crash (MVC) data have been at the aggregate level. Prior studies of aggregate, population-level data have shown that novice young drivers’ MVC rates peak at the transition to independent licensure and then decrease over time, roughly resembling a power-law function. This has led to speculation that learning-to-drive is best characterized by gradual learning processes. This might be true, but this intuition relies on the unstated assumption that the individual parts of a system should resemble the whole; this is a common type of logical fallacy called a category error. Further, although power-law patterns of learning and forgetting are frequently observed in perceptual-motor and associative memory tasks, performance in higher-order problem solving tasks is often characterized by phase transitions – abrupt changes as a result of explicit instruction or spontaneous strategy discovery. These discontinuous learning patterns have been observed in a variety of cognitive tasks (e.g., Tower of Hanoi, the gear task, and learning arithmetic). Importantly, when aggregated, individual trajectories of discontinuous learning have been shown to take the appearance of a power-law pattern. Safe driving is a multi-faceted behavior that requires performing complex cognitive tasks (e.g., tactical, strategic decision-making) as well as perceptual-motor and associative memory tasks (e.g., vehicle handling). To demonstrate that the population-level pattern of MVC reduction among young novice drivers that has been previously attributed to an underlying power-law theory can also be generated by a phase transition model of learning (PTMoL) in which all learners exhibit the same abrupt transition from high to low crash risk, but this transition occurs at a different time for each individual learner. The idea that learning to drive exhibits a phase transition can be operationalized mathematically. Briefly, PTMoL was based on a sigmoid function representing decreasing MVC risk starting from a novice higher risk level and decreasing to a more skillful lower risk level. The slope of that transition was fixed and individuals differed in transition onset. This model was used to generate MVC risk curves for 1,000 simulated drivers (over the first 24 months of licensure) by randomly sampling from an exponential distribution of phase transition times (i.e., a skewed distribution so that most participants transition early, but there is a long tail of later transitions). The resulting aggregate risk curves were fit to four independent sets of aggregate post-license MVC data from prior published studies. The PTMoL fits were compared with power law function fits, which is currently the dominant framework for describing the decrease in observed MVC rates following independent licensure. The PTMoL and the power law functions provided approximately equally good fits to the observed aggregate data. Across all four data sets, the quasi-R 2 values for the power law model were 0.93-0.96; for the PTMoL they were 0.91-0.96 (the authors use the term “quasi-R2” as a reminder that the usual interpretation of R2 does not apply to non-linear models, though it provides a convenient shorthand for comparing the fits of two very different kinds of models). The simulations provide an existence proof that the aggregate MVC rates are consistent with a phase transition model in which learning to drive safely is abrupt rather than gradual. Therefore, both a gradual learning model (power law model) and a phase transition model of learning are viable accounts of population-level patterns in young drivers’ aggregated crash data. These models have different implications that will be discussed in the presentation.

This paper was sponsored by TRB committee ANB30 Standing Committee on Operator Education and Regulation.

18-01534

English

4p

2018

Transportation Research Board 97th Annual Meeting

Digital/other

Figures; References (14)

Crash rates; Driver education; High risk drivers; Mathematical models; Recently qualified drivers; Risk assessment; Task analysis; Young adults

Highways; Safety and Human Factors

Transportation Research Board Annual Meeting 2018 Paper #18-01534

TRIS, TRB, ATRI

Jan 8 2018 10:23AM