Relative risk in the context of driver impairment

What is relative risk?

Risk is the probability that an adverse event will occur. In the context of road safety, this refers to the risk of a traffic accident or fatality.

As a driver becomes impaired by intoxication, drowsiness, or another factor, their risk of an accident increases. To quantify this increase, we look at:

  • the absolute risk of an accident in an unimpaired state, and compare it to
  • the absolute risk of an accident in an impaired state.

Relative risk is then the ratio between these two numbers (BMJ Best Practice has a handy list of the key formulae for any statisticians reading).

Let’s look at alcohol impairment as an example.

All figures in this blog post are drawn from a study by Voas et al. which can be found on PubMed. All data is based on drivers fatally injured in a single-vehicle crash.

If an unimpaired driver has a 5 in a million chance of having an accident, then that becomes the baseline with a relative risk of 1.

Relative risk of single-vehicle crash fatalities at a BAC of 0.025% relative to 0.000% for drivers aged 35 and above.

If an impaired driver at 0.025% blood alcohol concentration (BAC) has a 10 in a million chance of having an accident, then they have a relative risk of 2 (double the baseline).

Relative risk curve of single-vehicle crash fatalities for drivers aged 35 and above.

If we determine the relative risk at other levels of BAC, we can produce what is called a relative risk graph (or curve).

Relative risk curves of single-vehicle crash fatalities for drivers in three age groups.

We can even produce separate relative risk curves for different demographics.
The curve is much steeper for younger drivers.

Risk of single-vehicle crash fatalities at a BAC of 0.00%, by age and gender, relative to men aged 21–34.

There is another level of complexity.
The baseline risk for each demographic cohort differs at 0.00% BAC.

Relative risk of single-vehicle crash fatality of males aged 16-20 relative to females aged 35 and above at 0.00% BAC.

Without alcohol, a male driver aged 16-20 has 7.64 times the risk of a fatal, single-vehicle crash than a female driver over 35.

Relative risk of single-vehicle crash fatality of males aged 16-20 relative to females aged 35 and above at 0.08% BAC.

At a BAC of 0.08%, the relative risk between the two widens to 11.23 due to the steeper relative risk curve of younger drivers.

In summary, at the same BAC different people exhibit vastly different levels of risk.

Which data sources can produce a relative risk graph?

There are two main ways to produce a relative risk graph.

1. Collect data from real roads.

Relative risk derived from real roads quantifies the rates of accident or fatality at each level of impairment. In the case of alcohol, the most commonly used measure is BAC (although different people become impaired to vastly different degrees at the same BAC).

This requires two large-scale data sets: One reveals the proportion of drivers on the road at various levels of impairment (or a proxy measure such as BAC), the other counts how many accidents or fatalities occur at each BAC. In combination, these data sets reveal the absolute and relative risk of an accident or fatality at each BAC. These data sets can be further segmented by demographic characteristics – usually age and sex.

2. Dose test participants in the laboratory.

Test participants arrive at the laboratory unimpaired. They are then dosed with alcohol or a drug (or undergo an extended wakefulness protocol to induce drowsiness). They subsequently either perform one or more psychomotor vigilance tests or driving tasks (either in a simulator or on-track) so researchers can measure the objective impairment that has resulted.

Such tests are interesting because they reveal that different people require vastly different doses to exhibit objective impairment. Some people become impaired at 0.03% BAC, while others are barely impaired at 0.08% BAC. Some people become impaired after 20 hours of extended wakefulness, while others can ace a vigilance test and drive flawlessly after 32 hours awake. It is important to note that subjects cannot accurately assess their own impairment from drowsiness, alcohol, or drugs. Many people often believe they are unimpaired, but road fatality statistics and other objective show otherwise.

Each approach has pros and cons.

1. Collect data from real roads
  • Includes impairment of judgement: Such data accounts for any increases in risk that are not measured in laboratory impairment tests, such as increased aggression and risk-seeking behaviour due to alcohol.
  • Includes all road users: There are no exclusion criteria that “hide” a portion of the population from analysis.
  • Reflects real-world risk: From the point of view of road safety analysts and policymakers, such data offers a real-world view of which drivers under which scenarios present the greatest risk.
  • Incomplete data: This form of analysis requires a lot of inference when data is missing. In the U.S. Fatality Analysis Reporting System (FARS), data is missing on 40% of drivers involved in fatal crashes. Researchers usually use imputation to fill in these gaps.
  • Sampling bias: The method of data collection may result in overrepresentation of certain groups of people.
  • Generalisability is difficult across different countries and cultures in which behavioural patterns differ.
  • Attribution of an accident to a causal factor such as BAC may be confounded by other unmeasured factors such as drowsiness or other drugs.
  • Data limitations: In the case of drowsiness, it is difficult to establish how drowsy a driver was immediately preceding a fatal crash. For alcohol, a blood test reveals BAC. Consequently, drowsy driving accidents are frequently underestimated.
2. Dose test participants in the laboratory.
  • Control: All variables can be strictly controlled so that a study can objectively measure the precise level of impairment that arises from alcohol.
  • Isolate one form of impairment: Participants can be screened to exclude confounding factors such as sleep disorders or other drugs.
  • Test different doses on same participants: The same participants can be measured before and after dosing to isolate the precise effects on psychomotor vigilance and driving ability.
  • Availability bias: Such tests rarely measure the increased aggression and risk-seeking behaviour due to alcohol, which is difficult to objectively quantify in a laboratory.
  • Selection bias: It is difficult to include participants that are diverse across all attributes.
  • Hawthorne effect: Participants change their behaviour when they are aware they are being observed. They may consciously focus on the test and exhibit less impairment than if they were dosed and unobserved.

How is relative risk commonly used in traffic safety?

How is data from real roads used in traffic safety?

Relative risk curves based on BAC have been used by road safety researchers to set the legal limits for driving under the influence. Different jurisdictions deem various levels of relative risk acceptable. Consequently, the legal limit differs across countries and even across states within the same country in some cases.

How is data from laboratory studies used in traffic safety?

A prime example of this is Optalert’s Johns Drowsiness Scale (JDS), which measures impairment from drowsiness. In experiments, it yields a relative risk curve in the driving task. As the JDS increases, the risk of a driver exhibiting potentially fatal errors increases exponentially.

Ideally, a system that measures impairment should output the relative risk of a performance failure in a driver. This allows for the vehicle to enact soft countermeasures at lower levels of impairment. In the case of drowsiness, countermeasures such as lowering the cabin temperature or increasing the music volume can prevent or prolong its progression. In the case of alcohol and drugs, the adaptive cruise control (ACC) may leave extra space from the car in front or the automatic emergency braking (AEB) may increase its sensitivity. In all scenarios, the driver might also be encouraged to rest or drink coffee. All of these soft countermeasures are made possible by the detection of lower levels of impairment.

Mean risk per minute (and its standard error) for test participants driving with all 4 wheels of the car out of the lane (i.e., “off-road”) versus their JDS scores.

Johns, M.W., Chapman, R., Crowley, K. et al. A new method for assessing the risks of drowsiness while driving. Somnologie 12, 66–74 (2008).

Rudimentary systems output a binary value deeming the driver as either impaired or unimpaired. These only allow for one intervention upon detection of impairment – usually a loud beep alerting the driver.

All of Optalert’s algorithms output relative risk.

We have absolute confidence in the accuracy of our algorithms. Other drowsiness detection systems cannot detect early-stage drowsiness. For this reason, they output only a binary impairment detection event to hide the inaccurate lower ranges of their systems.

We encourage automotive OEMs and tier 1s to contact us for more information on how to integrate the world’s best driver drowsiness detection into your DMS.

We are eager to support the industry to employ sound science and ensure we are doing all we can to keep drivers and other road users safe.