Are there safe and unsafe drivers?

Are there safe and unsafe drivers?
Liisa Hakamies-Blomqvist1,2
1
The Swedish National Road and Transport Research Institute, Sweden
2
University of Helsinki, Finland
email: [email protected]
Abstract
One of the aims of most driver screening practices, in addition to a variety of possible
diagnostic and rehabilitation activities, usually is to exclude from the driver
population those persons who are not “safe enough” to drive. In order to realise this
aim i.e., to decide whether or not somebody is “safe enough” to drive a car, we must
estimate his or her individual risk as driver and compare this risk level with a
threshold value.
When estimating a driver’s individual risk we must make certain assumptions, such as
the following: (1) there is such a thing as individual driver risk; (2) it can be expressed
as one parameter; (3) it can be measured or estimated; and (4) thresholds dividing the
driver population into those safe and those unsafe can be identified. All these
assumptions have some problems from both a theoretical and a methodological
perspective. In the present paper, the theoretical and methodological basis of driver
screening is critically discussed.
Keywords: Driver screening, Risk, Older drivers, Policy
1. Introduction
This paper addresses the issue of driver risk from a combination of different
perspectives: conceptual, theoretical, methodological, and political. The overall aim
is to sort out the theoretical and practical preconditions for decision making about
who should and who shouldn’t be allowed to drive a car.
In all societies having any safety policies at all, there are driver-control policies to
allow some people to drive cars while excluding others. The prototypical situation is,
of course, the tests beginning drivers have to pass for their first driving license. In
most countries, there will be a driving test where the aspirants have to be able to
demonstrate their ability to drive safely. They don’t have to drive skillfully, or
smoothly, or elegantly, if this does not have bearing on their safety: all evaluation
dimensions are subordinated to safety and used only since they are thought indicative
of one’s ability to drive without accidents. The final decision is one of pass or fail:
the drivers are classified as “safe” or “unsafe”. (At the beginning stage however, it is
customary, if people fail, to suggest further interventions to give them a new chance
of passing.)
Safe and unsafe drivers?
2
Testing novice drivers seems to be reasonably non-controversial. Where mandatory
driver education exists, the tests also can be planned with good face validity if related
to the official goals of the education. There are however other forms of driver
controlling policies that are much more controversial, such as general screening of
older drivers. The general claim motivating the analysis of the present paper is that it
is highly doubtful that people who design and apply driver screening systems do not
generally speaking understand very thoroughly what they really are about.
1.1. The concept of risk
Evidently, driver safety cannot be discussed without also discussing risk. What, then
do we mean by “risk” in the driving context? In a recent paper Fuller (in preparation),
building on Haight (1986), differentiates from a driving behaviour perspective
between objective risk, subjective risk, and the driver’s feeling of risk. The difference
between the two last ones is according to him important: subjective risk refers to the
driver’s own (cognitive) estimate of the objective probability of loss of control and
collision, while feeling of risk represents an emotional and/or arousal response to a
potential threat or aversive event. In different situations, these two may or may not be
related.
As to objective risk, Fuller equals it with statistical risk which he defines as the
objective probability of being involved in an accident. In so doing, however, he
misses an important distinction: that between the objective probability of an accident
to occur and of somebody’s being involved in an accident. In the present paper, the
concept of objective risk is used for the first one and that of statistical risk for the
second one. The difference is about whose attribute the probability in question is: in
the first one, it is the accident’s; in the second one, a person’s or a group’s. In a
driving behaviour context, at any given moment, it is in principle (if not in practice)
possible to estimate separately objective risk, i.e., the probability of an accident to
occur (given the values of the critical parameters); subjective risk, i.e., the driver’s
own estimate of such probability; and finally his/her feeling of risk, i.e.,
emotional/arousal response to potential threat. Common for all these three is that they
deal with the prediction of one accident (the next one threatening to occur).
Moving over to statistical risk however, defined as the probability of being involved
in accidents, we switch the perspective away from one single accident to those actors
involved in accidents. The probabilities we try to estimate in this context are no more
attributes of the accidents but rather of those involved in them. In the statistical
research setting, the most common modus operandi is to use archival data to
determine accident risk post hoc by applying the equation Risk = Accidents/Exposure,
with some theoretically good-enough and practically feasible operationalizations and
metrics for the nominator and the denominator. (Note that we did not seem to need
the concept of exposure above, while discussing the other types of risk, since the
setting was not statistical but that of estimating parameters linked to the prediction of
the next possible accident.) The mainstream statistical approach will typically yield
us epidemiological findings: for example, that the risk for a Swedish driver of
incurring a personal-injury accident is 1/20000 per year, or that young male drivers
have higher risk (higher accident rates) than young female drivers.
Safe and unsafe drivers?
3
However, the epidemiological approach is not the only possible one: in the same vein
as we ask “what is the probability of a Swede having accidents” we can ask “what is
the probability of X-crossings vs roundabouts having accidents”, or even “what is the
probability of this specific location having accidents”, thus extending the approach to
something similar to black spot analysis. The interest is still in the probability of
somebody’s (or someplace’s) having accidents, any accidents at all, not in predicting
the next accident out there in the near future that may or may not occur depending on
important determinants such as driver behaviour.
We can also ask: “what is the probability of this person’s having accidents?”, thereby
aiming at estimating what in the present paper is defined as the fifth use of the
concept “risk”, namely, individual driver risk, i.e., the probability of a driver’s to have
accidents. The theoretical meaningfullnes of such a concept can be questionned;
nevertheless, in the practical context of driver screening it is already in use, and if it is
used we should preferably understand its theoretical and practical limitations.
2. Hidden assumptions about individual driver risk
Certain preconditions have to prevail if estimates of individual driver risk are to be
used with some realiability and validity in decision making about any particular
driver’s continued driving or exclusion from the driver population. (For the sake of
simplicity, the present discussion is limited to the final pass-fail decisions, omitting
the issue of conditional driving licenses.) These preconditions are in the present paper
called “hidden assumptions” since, as a rule, they are not critically discussed in those
policy contexts (and even in scientific work) where concepts like “safe driver” are
used and decisions about continued driving made.
Regarding the concept of individual driver risk, it must be assumed that:
(1) there is such a thing;
(2) it can be expressed as one parameter;
(3) it can be assessed;
(4) thresholds dividing the driver population into safe and unsafe can be
identified.
In the following it is asked whether, and under what conditions, these assumptions
actually can be defended.
2.1. Does individual driver risk exist?
Decisions about who is and who is not “safe enough to driver” will in many cases
affect profoundly the future quality of life of the person. Such decisions cannot be
based on anything else than individual estimates of driver risk. We must therefore,
for ethical reasons, feel confident in assuming that driver risk is a reasonably stable
personal characteristic and that there is substantial inter-individual variation.
Supposing that individual driver risk exists, how should it be defined? Above, it was
preliminarily defined as a driver’s probability to have accidents. What is missing in
this definition is the dimension of exposure. Since the concept is mostly needed in
policy contexts and used for screening purposes, the following, rather pragmatic
definition is suggested here: individual driver risk is a driver’s probability to have
Safe and unsafe drivers?
4
accidents, if allowed to drive. The implications of this definition are dicussed further
under the next point.
2.2. Can individual driver risk be expressed as one parameter?
The need of expressing individual driver risk in one parameter inevitably follows
from the need of making pass-fail decisions: whatever complex assessments we use,
in the end we have to be able to comprime them in one parameter that can be
compared to whatever norms we have for “safe” or “unsafe”. (For the sake of the
argument, the problem of assessment is ignored so far and it is assumed that we
through some means can arrive at estimating individual accident probabilities.)
Since we are dealing with probabilities, we should ask whether the scale properties of
individual accident risk can be reasonably assumed to correspond to what is known
about probability: that it has values ranging from one to zero. Can driver risk have
the value one? Yes, if the driver is dead or otherwise incapable of action, and the car
is moving. If we could, we would certainly wish to exclude from the wheel all drivers
who just had a fatal stroke while driving. On the other end, can driver risk have the
value zero? Again: yes, if the driver is not driving.
This seemingly empty-minded statement brings in the issue of exposure and is of
importance when estimating the risk different driver groups pose for general safety.
Young skillfull drivers may be good at avoiding accidents in surprising situations
demanding rapid reactions and good car handling skills; older drivers, on the other
hand, are good at avoiding accidents by choosing not to drive when the conditions are
too taxing for them. Traffic psychologists call this strategic compensation:
unfortunately, it does not improve older drivers’ statistical accident rates, as high-risk
exposure avoided by not driving is not included in the Risk=Accidents/Exposure
equation. Nevertheless, from a theoretical (Michon xxxx) point of view, decisions
about when and where to drive are driver decisions at a high strategic level and,
correspondingly, reflect driving skill on that level. Therefore, if we define driver risk
as the probability to have accidents while driving, which would seem neater than the
pragmatic definition suggested above, we will exclude an important part of driving
skill that has direct bearing on individual accident risk.
Under the first point it was claimed that individual driver risk should be a “resonably”
stable characteristic. This was obviously because there are driver characteristics that
may be statistically related to risk of accident but that have different stabilities in
time: those that never change (such as gender; the few exceptions confirm the rule);
those that change slowly, such as age and, relatedly, contrast sensitivity, and those
that vary at more or less rapid paces, such as menstrual and circadian cycles, mood,
and momentary arousal. In consequence, individual driver risk, even if “reasonably”
stable, always will have some degree of intra-individual variation.
It follows that an individual’s risk can theoretically be described as a statistical
distribution of his/her risks over time. Given that a statistical distribution is
characterised by a number of parameters, we must ask which one of these statistical
parameters is most suitable to indicate individual driver risk, given the need of one
parameter. Should we simply demand a mean that is good enough, or should we pose
the decision about safe or unsafe on, say, a maximum probability not to be exceeded?
Safe and unsafe drivers?
5
Or, developing the thought further, should we also take into account the form of the
distribution? Would we then prefer drivers whose risks concentrate around an
acceptable mean in something resembling a normal distribution, or perhaps drivers
whose risk is close to zero almost always but at very rare occasions may get
considerably higher values?
To illustrate this with an example, the second type of risk distribution might belong
to, say, a busy professor with many family and work duties. Normally she a good and
safe driver with accident probabilities nicely grouped close to zero. However, there
may be unlucky accumulations of risk factors that are very rare, but when occurring,
signifcantly increase her risk – say, that she has PMS, has slept too little, is late to the
airport, has just quarrelled with her husband and has a bad-hair day. But she still does
drive, and therefore the second distribution would be a fair description of her risk. On
the other hand, it might also belong to an old lady who normally drives very
cautiously. She may not be bothered by the kind of things that make the previous
example more risky (she may even be past the age where one keeps a husband) but
she has cataracts making it very difficult for her to drive when it’s wet and dark. On
the other hand, she never drives when it’s wet and dark. Would the second
distribution still be a fair description of her risk? And finally, the second distribution
might describe a person with diabetes treated with insulin. If the insulin levels vary
too much s/he may occasionally have problems with the level of consciousness;
however if well treated that will never occur.
2.3 Can individual driver risk be assessed?
It has been assumed so far that a person’s individual driver risk is a statistical
distribution whose form and defining parameters vary between individuals. In
practical terms however, we have no means of directly assessing the values forming
that distribution. When dealing with statistical risk in an epidemiological setting, we
have databases on accidents and exposure that have their flaws but, if used wisely,
yield us acceptable estimates of risk, e.g., accident rates. Then, as a next step, if we
wish to, we can attempt to correlate these rates to factors believed to be determinants
of risk. In contrast, when estimating an individual’s risk – an idiographic setting – we
mostly do not have enough archival data since accidents are, on individual level, very
rare events: even the worst human “black spots” have snow-white pasts until their
first accident and manage to accumulate only a limited number of mishaps during
their driving careers. We therefore have to go, basically, the other way round. We
will know, guess, or hope that certain conditions or characteristics that we are able to
assess are related to accident risk; we will assess them; and we will make on the basis
of the combined assessment our best educated guess as to how likely the person is to
have accidents. In other words, any estimates of individual driver risk will be based
on indirect evidence.
There are many problems with such an approach. A major one is that even the
evidence behind the factors assumed to be indicative of risk unfortunately mostly is
indirect. In studies attempting to find correlations between driver characteristics and
safety, accident rates mostly do not offer sufficient statistical power to be used as
criterion variable and therefore, more often than not, different surrogate measures are
used. For instance, driving tests have been used as criterion variable in studies aiming
at identifying determinants of individual driver risk. However, without an ultimate
Safe and unsafe drivers?
6
validity study relating those driving tests to safety, the evidence is not robust enough.
We do not want to exclude from the driver population those who do not do well in
tests, only those likely to have accidents.
Another problem, also related to predictive power, is that accidents are not only rare
but multidetermined: behind most accidents, there are many contributing causal
factors of which any relatively stable driver characteristics form only a minor part.
Therefore, the predictive power of any single factor will never be very high. In
addition, the influence of any single factor will most likely depend on a number of
other factors, forming clusters of complex co-variance constellations. McKnight &
McKnicght have convincingly argued for patterns of deficits rather than single
functional deficits as predictors most likely to be related to accident risk in older
drivers.
2.4. Can tresholds for “safe” and “unsafe” be defined?
Let us assume for the sake of the argument that we have managed our way through
the above points 1 to 3 and have succesfully assessed a number of drivers’ individual
risk on a probability scale from 0 to 1. How do we answer the original question about
who should be allowed to drive? In other words, on what grounds do we put the
treshold value of “unsafe” at some value between 0 and 1? How safe is “safe
enough”?
In a screening context, even if we do not have the practical ambition to actually end
up with a stample-on-the-forehead kind of individual risk assessment, we still need a
decision rule that will integrate all our relevant assessments to a pass-fail decision.
This arises many questions. An important one is: should that algorithm be identical
for all driver groups? It is argued here that from both a theoretical and a practical
perspective the answer must be negative, for at least wto reasons.
First, a safe-enough driver can be put together as many different kinds of cocktails.
Young (male) drivers may have good car handling skills, dazzling reaction times and
large youthful attentional resources; on the other hand, they may be thoughtless, take
unnecessary risks and use the car for secondary risk-increasing purposes such as
showing off for friends. Older drivers, on the other hand, are slower and less flexible
and have narrower attentional resources but are (as a rule) cautious, safety-oriented
and defensive. Hence, different profiles of safety-related characteristics may yield
equally acceptable levels of individual driver risk.
Second, even treshold values for single functions may signalise different things for
different driver groups. For example, if we could realiably assess the soundness and
balance of a driver’s general judgement, a meta-level characteristic behind
cautiousness and safety orientation and therefore important for driver safety, we might
well be forced to accept for certain subgroups of young male drivers mean scores so
low that for older drivers, similar values would be indicative of dementia.
A final issue to be raised here is that decisions about the level of acceptable risk are
not scientific but political. The expressions “safe enough to drive” or “unsafe driver”
often are utilized as if they were objective and scientific. They are not. There are no
safe drivers. Intra-individual driver risk varies depending on driver state and driver
Safe and unsafe drivers?
7
characteristics in relation to and interaction with driving conditions. And since it
varies, it follows that it will at some point be more than zero, which means that no
driver is always absolutely safe. Scientists can help develop methods to assess risk in
evermore reliable and valid ways, but science cannot decide whether the line between
safe and unsafe should be drawn at 0.x or 0.y, or perhaps at whatever value is the
double or the triple of the national mean. Decisions about safe “enough” are
decisions about how much risk the society is willing to tolerate: they always contain a
value element and thus are always political in nature.
3. Conclusion
Excluding somebody from the wheel almost always means choosing a certain bad
outcome in order to prevent a low-probability bad outcome. Accidents are, on an
individual level, rare events; even most so called high-risk drivers drive without
accidents. To loose once license in contrast has, especially for older people, profound
consequences in terms of quality of life, independency, physical and mental health
and also in terms of the financial burden imposed on the society. Based on the
discussion above, the conclusion of the present paper is that driver screening is in
both theoretical, practical, and ethical terms a highly dubious activity.
4. Recommendations
Those responsible for designing and applying policies for driver screening should be
well aware about the dubious nature of such activities. First, they should understand
the many sources of theoretical and empirical uncertainty in the scientific study of
individual driver risk. Second, they should understand the political nature of all
decisions about who is and who isn’t allowed to drive, even when such decisions are
masked as science. Third, they should understand that the consequences of such
decisions also are political in nature and always imply prioritising between different
gains and losses. Fourth, when making individual decisions about continued driving,
they should be conservative and err on the safe. In the present context, this means not
excluding anybody from the driver population unless compelling evidence of high
increase in the levels of important indicators of risk is available.
Acknowledgements
Financial support from the Swedish Agency for Innovation Systems (VINNOVA) is
gratefully acknowledged.
Safe and unsafe drivers?
8
References
Fuller, R. (in preparation). Towards a general theory of driver behaviour.
Haight, F.A. (1986). Risk, especially risk of traffic accident. Accident Analysis and
Prevention, 18, 359-366.
Michon, J.A. (1985). A critical view of driver behavior models: What do we know,
what should we do? In L. Evans & R.C. Schwing (Eds.): Human Behavior and Traffic
Safety, 485-520. New York: Plenum Press.