Understanding the causes of newly licensed teen drivers’ risky driving behavior is important for public health due to the elevated injury risk that young drivers pose to themselves and to other road users. While we know a lot about between-person risk factors (i.e., the kinds of things that make one driver more risky than another driver), explanatory frameworks of within-person changes in any given driver’s crash risk are very underdeveloped. Because our explanatory change frameworks are underdeveloped, our interventions designed to change young drivers are often ineffective and not properly evaluated.
For the past several years, I have been exploring if a fundamental assumption about the population crash rate data may have been misguiding intervention development and evaluation for young drivers. It is widely observed that population-level crash rates incrementally decrease following licensure, which has led to the assumption that recently licensed teen drivers’ crash risk also decreases incrementally and homogeneously (i.e., in the same way for every driver) as they accrue experience. I became curious about the state of the evidence that supports this assumption because, although it is tempting, we cannot assume that individual-level changes in crash risk mirror the population-level changes in crash rates.
First, my colleagues I and used a computational cognitive modeling approach (see Mirman, Curry and Mirman (2018)) and found strong evidence that aggregating individual-level abrupt decreases in young drivers’ crash risk can accurately generate population-level crash rate data from over 1 million young drivers in three countries. We also found that phase transitions in crash risk trajectories can be induced by interventions that implicitly or explicitly provide teenagers opportunities to learn risk-reducing strategies.
Second, in a review of the literature (Mirman 2019) I found very little evidence of individual-level data that supports the assumption that crash risk gradually decreases with the accrual of post license experience. To the contrary, there was ample evidence that between and within-person data aggregation are masking abrupt change processes, like phase changes. There is also a lot of existing theoretical support for phase-like changes in the developmental and cognitive sciences.
These findings have implications for how we reason about data from the individual and the population-levels. For example, using population-level crash data McCartt et al., (2009) estimated during the first year of driving (~7,500 miles) that -on average- 17 year-old drivers will see their crash risk reduce by 30% due to “experience”. However, using a phase transition perspective it’s also possible that half of the drivers could see a 60% reduction and the other half could see no reduction. This is a somewhat extreme interpretation; however, it illustrates the importance of making our implicit assumptions explicit – and then checking them. This is especially important when using population-level data to draw conclusions about individual-level change processes, and then using those ideas to develop and to evaluate individual-level interventions.
For example, accrual-based interventions for young drivers that are designed on the idea that every bit of practice matters about the same are unlikely to work. This doesn’t mean that getting additional high-quality practice driving isn’t a good thing, but rather that we can’t say that any given unit of “practice” will be of roughly equal benefit for each driver. Moreover, we cannot assume that every driver needs the same intervention at the same time. A more effective, and more efficient, strategy is to have comprehensive interventions that target multiple risk and protective factors over time.
There are also implications for how we evaluate interventions for young drivers. It is difficult to demonstrate a protective effect in “single-dose-main-effect” intervention studies because crash risk is very heterogeneous at the individual level and, especially if one is using data aggregation practices common to RCT-style evaluation studies. For example, if an intervention is very effective for a subgroup of drivers and not effective for most drivers, the overall group average effect size will be functionally zero.
This can lead us to continually rejecting intervention approaches that could be useful, but just not for everyone, or even a sizable majority of drivers. This issue is compounded by the variety of risks and types of exposures that drivers encounter. Drivers may only rarely be in a situation that exposes their weaknesses so an intervention could have worked by reducing a particular weakness, but then that driver was never in a situation where it mattered. Predicting exposures is very hard to do.
I look forward to other researchers and practitioners continuing to test the phase transition model and hope that it provides a new way forward for supporting young people. The code for the phase transition model is publicly available via OSF: https://osf.io/qbgpd/