Why drivers really crash: the human factor on trial

In the 1970s a team at Indiana University did something nobody had attempted at that scale before. In the Tri-Level Study they investigated over 2,200 crashes on site, and 420 of them in forensic depth, with a crash reconstructionist, an automotive engineer who disassembled the relevant vehicle systems, and a psychologist who interviewed and tested the drivers. The verdict: at the probable level of confidence, a human error was a cause in about 93% of crashes. The vehicle alone accounted for 2 to 3%.

Here is why that number deserves respect: it keeps being found, by different people, with different methods, in different decades and countries. A British team working at the same time, unaware of the Americans, reached 94%. The German GIDAS in-depth study put the human contribution at 97%. The US repeated the exercise on 5,470 crashes in 2005 to 2007 and got 94%. And when researchers stopped reconstructing crashes after the fact and instead wired up real cars with cameras and sensors, the SHRP2 naturalistic study of 905 real recorded crashes still found a human failure in 87.7%. Forty years, four methods, one answer.

And now the cross-examination, because this is where the number starts to mislead.

One more finding from the Tri-Level study is quietly more important than the headline number. When the investigators looked for impairments, alcohol, fatigue, illness, emotional upset, they found that in 88% of crashes the driver had no identifiable impairment at all. Alcohol was implicated in about 3% of the in-depth sample, fatigue in under 2%. The SHRP2 cameras confirmed it decades later: the most common crash scenario, about half of all crashes, is an unimpaired driver, momentarily distracted, making an ordinary error.

The Tri-Level team sorted human causes into three families. Recognition errors, failing to see or notice something in time, were involved in 56% of the in-depth crashes. Decision errors, seeing it but misjudging it, in 51%. And performance errors, seeing it, judging it, but mishandling the car, in just 11%. The 2008 US national survey found nearly the same split: recognition 41%, decision 33%, performance 11%.

Look closer at the single most common specific cause in the Tri-Level data: "improper lookout", failing to make the visual check the situation demanded, at a junction, before pulling out, before changing lane. On its own it accounted for over 20% of all crashes. Add inattention and distraction and the attentional cluster explains nearly half of all crashes (47.9%). The British team found essentially the same thing with different labels: "looked but failed to see", "failed to look", "lack of attention".

Why two seconds is the whole game

There is hard physics under this. Under ideal laboratory conditions a human can react in under half a second. On the road, to a surprise event, road design engineers assume 2.5 seconds from the moment something appears to the moment your foot reaches the brake. At 100 km/h that is roughly 70 metres travelled before braking even begins, and the biggest single lever on that number is expectancy: you react dramatically faster to things you were already anticipating. A driver who has learned to expect the cyclist hidden by the van isn't faster-twitched than you. They started their reaction three seconds before you, because their eyes and mind were already there. This is the scientific case for the anticipation work in our observation guide and the margin logic of the two-second rule.

Ray Fuller's task-capability interface model starts from an uncomfortable observation: once your car is moving, a collision is the default outcome. It is only your continuous corrections that prevent it. Driving is therefore a permanent balancing act between two quantities: task demand, how much the situation asks of you right now, set by speed, road, weather, traffic and whatever else you've taken on, and capability, what you have available right now, set by skill and experience but degraded in real time by fatigue, emotion and distraction. The gap between them is your safety margin. Demand exceeds capability, and control is lost; whether that becomes a crash or a lucky escape depends on luck and on other people.

The elegant part is what drivers actually regulate. Fuller's experiments showed that ratings of task difficulty and feelings of risk track each other almost perfectly (correlations around 0.97). Drivers don't compute crash probabilities; they drive to keep a feeling inside a comfortable band, mostly by adjusting speed. Fog, heavy rain, an unfamiliar road: demand rises, the feeling rises, we slow down. Empty road on a bright morning: the feeling fades, and the speed creeps up on its own. As Fuller put it, quoting Shakespeare: best safety lies in fear.

Two numbers from the physics side, both among the most replicated findings in the field. First, the severity curve: analyses of thousands of real collisions show the risk of death rises with roughly the fourth power of impact speed change. Joksch fitted exponents of 3.9 to 4.1 across crash types; Elvik's meta-analyses of over a hundred studies confirm the power-law family. Fourth power means a seemingly modest difference in speed is an enormous difference in outcome: it is why small reductions in average traffic speed reliably produce large reductions in deaths. Second, the involvement curve: in the Adelaide case-control studies, cars travelling at 85 km/h in a 60 zone showed dramatically multiplied crash involvement compared with cars at the limit, with risk rising steeply and continuously above the traffic's typical speed.

But this guide is about the human factor, and the psychology of speed is where the two books converge on something genuinely useful. A Belgian modelling study (De Pelsmacker & Janssens) measured what actually drives speeding behaviour. The intention to speed mattered (correlation 0.47). But habit mattered just as much (0.50). Half of speeding involves no decision at all: it is simply the speed your hands and feet produce when nobody, including you, is asking the question. And the strongest influence on the deliberate half wasn't fear of consequences or knowledge of the law, it was personal norms, the speed a driver privately considers reasonable, anchored to what they believe everyone else does.

Start with an unsettling baseline. When researchers filmed ordinary drivers for weeks (Stutts et al.), not a single driver gave driving their full attention all of the time. Everyone eats, talks, fiddles with controls, drifts off into thought. Shinar's conclusion from fifty years of this evidence is worth sitting with: most of the time driving simply is not demanding enough to hold full attention, and drivers seek stimulation to fill the gap. The distraction problem is not an invasion from outside. It's built into us.

So if everyone is distracted sometimes, what actually kills? The SHRP2 naturalistic study, with cameras in over 3,000 cars and 905 real crashes, produced the cleanest risk table the field has ever had. The pattern is unmistakable: what multiplies crash risk is taking your eyes and hands away from the task, not your mind alone.

Three findings that every honest instructor has to face. One: teaching skill does not automatically produce safety. Reviews across five decades keep failing to find a direct crash reduction from conventional driver education. Two: more skill can even mean more crashes. When Norway made skid training compulsory, crashes on slippery roads went up, and the Norwegian evaluation was quasi-experimental, so self-selection can't explain it away. Finnish data later showed skid-trained young men were about 50% more likely to crash on slippery roads. The training raised confidence faster than it raised competence, and drivers spent the difference. Three: errors don't predict crashes, violations do. The landmark work of Reason and Parker found that ordinary mistakes and lapses barely correlate with crash involvement; deliberate violations, speeding, tailgating, chancing gaps, are what predict it. What you're willing to do matters more than what you're able to do.