Better to see Chapter 9 Driver behavior  and Chapter 13 Measures to improve traffic safety of Traffic Safety  (2004)


Chapter 11.  USER RESPONSES TO CHANGES IN TRAFFIC SYSTEMS (From 1991 book Traffic Safety and the Driver)


Words only (no formatting, figures, tables, or photographs) from 1991 book



Paperback copy of complete unchanged book available from , list price $29.95



      It has been recognized as obvious since antiquity that humans change their behavior in response to the perceived probability and severity of harm.  We walk more carefully when the ground is wet or icy than when it is dry; we walk more carefully on rough surfaces when barefoot than when wearing shoes.  A warrior clad in armor may accept a greater risk of being struck by a weapon than one not so clad, and so on.  Shakespeare writes, "Best safety lies in fear" [Hamlet, Act I, Scene 3].  The question of road users responding to changes in the safety of traffic systems has also long been recognized.  More than half a century ago, in a paper entitled "A theoretical field-analysis of automobile driving", Gibson and Crooks [1938, p. 458] write:


            More efficient brakes on an automobile will not in themselves make driving the automobile any safer.  Better brakes will reduce the absolute size of the minimum stopping zone, it is true, but the driver soon learns this new zone and, since it is his field-zone ratio which remains constant, he allows only the same relative margin between field and zone as before.


A decade later Smeed [1949, p. 13] writes:


            It is frequently argued that it is a waste of energy to take many of these steps to reduce accidents.  There is a body of opinion that holds that the provision of better roads, for example, or the increase in sight lines merely enables the motorist to drive faster, and the result is the same number of accidents as previously.  I think there will nearly always be a tendency of this sort, but I see no reason why this regressive tendency should always result in exactly the same number of accidents as would have occurred in the absence of active measures for accident reduction.  Some measures are likely to cause more accidents and others less, and we should always choose the measures that cause less.


      This chapter is devoted to examining user reactions observed following various safety changes in traffic systems.  A formalism to organize such user responses systematically is developed.  Some explanations of user reactions to safety interventions, and broader attempts to model driver behavior, are then discussed.




      Let us suppose that some change is introduced into a traffic system that is expected to change safety by some fraction, say DSEng, assuming users continue to behave exactly as they did before the change.  The subscript denotes that the change is of an engineering nature.  For example, if design changes to a guardrail are estimated by engineering methods to reduce the probability of driver death on impact by 10%, then DSEng would be 10% for drivers killed crashing into the guard rail.  We use DSEng more generally to indicate fractional reductions in some harm measure (crashes, fatalities) expected from systems changes if users do not alter their behavior in response to these changes.  The change might be a higher or lower speed limit, equipping vehicles with devices aimed at reducing fatality or injury risk,
transferring to heavier vehicles which have lower fatality and injury risks, or mandating belt or helmet wearing.  While safety interventions always aim at producing positive values of DSEng, there are other changes motivated by different considerations, such as saving fuel in the case of smaller cars, for which the values of DSEng are negative.

      Because road users may alter their behavior, the actual realized percent safety change, represented by, DSAct, may differ from DSEng.  These quantities can be considered to be related in the following simple way:


                        DSAct   =   (1 + f) DSEng   ,        Eqn 11-1


where f is a feedback parameter which characterizes the degree to which users respond to the safety change.  In this context feedback is synonymous with user reaction, behavior change, or interactive effects in the system.  If users do not change behavior in response to the safety change, then  f = 0,  and the safety change is just as expected on engineering grounds.  If the safety change is in the expected direction, but of lesser magnitude than expected, then -1 < f < 0, and the safety change is discounted compared to the expected amount.  If the safety change has no effect, then f = -1.

      In order to discover what values of f really occur, we consult the literature.  There we find responses in ranges beyond those illustrated above.  Because of this rich variety, I recommend the term human behavior feedback rather than the many other terms, such as risk compensation and danger compensation, which have appeared; every one of these other terms implies that user reactions are confined to a narrower spectrum than evidence shows to be the case.  A stronger reason why such terms should not be used is that they go beyond describing the phenomenon, but rather, without justifiable evidence, also imply a knowledge of the mechanism leading to the effect.  The
progression in science is first to organize what is observed, and then to try to explain it, rather than name the observations with the explanation.

      The studies reviewed in the literature are placed in two broad categories; first, those with DSEng > 0, which are aimed at increasing safety; second, those with DSEng < 0 which are expected to decrease safety, but are introduced for other reasons.  We use "expected" to mean the change expected in the absence of user response, even if we know enough to really expect user response.  From a formal point of view, there is no need to treat changes expected to increase and changes expected to decrease safety as separate cases; Eqn 11-1 applies equally to positive and negative values of DSEng.  Because treating them together would involve using language that flows awkwardly against common usage we instead treat positive values first, and then negative.

      Another advantage of treating positive and negative values of DSEng separately is that it facilitates graphical representation (Figs 11-1 and 11-2).  As we read up the page, f increases in Fig. 11-1, but decreases in Fig. 11-2, so that safety increases as we go up the page for both figures.  These figures build on those in Evans [1985a], and include only effects supported by the discussion below or the reference cited; a ? indicates my judgment of an uncertain result or interpretation.  Interventions reviewed below are associated with one of five regions of the human behavior feedback parameter, f, as indicated in the figures; the level of uncertainty is generally too high to estimate specific numerical values of f.


Figs 11-1 and 11-2 about here; Fig 11-1 on a left-side page and Fig 11-2

on a right-side page so that reader sees them together





1. Safety increase greater than expected (f > 0)


      There are various indications that when the nationwide 55 mph speed limit was imposed in the US in 1974, various crash, injury, and fatality rates on roads unaffected by the change also declined.  As such changes were coincident  with many other major disruptions, it is not possible to associate them confidently with the 55 mph speed limit (which was instigated as a fuel-saving measure).  However, the possibility that reducing a speed limit on the Interstate system could generate spillover effects of the type discussed in Chapter 5 raises the possibility of safety increases beyond those computed for the intervention.


2. Safety increase as expected (f = 0)


      It is exceedingly improbable that any intervention of which road users are aware can have f = 0 exactly.  However, this value is of particular interest because of its interpretation, even though the probability is essentially zero that any specific exact numerical value of f, including f = 0, will occur.  Some values of f very close to zero must arise for cases in which most road users are unaware of the safety intervention, such as side-guard beams in car doors.  Although some road users might react to the national fatality data or to news reports, it is difficult to see how such inputs could affect net harm (in the short term) by more than a microscopic amount.

      One area in which there has been much focus over possible user reactions is that of driver-restraint use (Chapter 10).  It seems to me essentially impossible that such prominently visible and widely discussed items as safety
belts and helmets would not induce some change in user behavior.  Thus, on logical grounds, it seems almost impossible that f would be exactly equal to zero.  Such an observation is little more than a truism of no practical importance.  What is important is the magnitude (including of course the sign) of f.

      In Chapter 10 we find that observed fatality reductions from increases in safety belt use rates are in essential agreement with reductions calculated based on realistic estimates of changes in belt use, selective recruitment, and the effectiveness of belts in crashes.  Such agreement between actual and expected changes precludes the possibility that f can differ much from 0.  This finding is further corroborated by field studies in which the behavior of drivers compelled to wear belts is found indistinguishable from estimated behavior of non-wearers, and by a test track study in which the change from wearing to non-wearing, and from non-wearing to wearing, produced no observable systematic changes in chosen speed.  Collectively, the evidence suggests weakly that drivers obliged to wear belts are more likely to increase than to decrease caution; that is, f is more likely to be positive than negative.  While available evidence does not lead to a calculated numerical value for f, I believe it does place it in the region -0.1 < f < 0.2.  The evidence (Chapter 10) precludes beyond doubt that f could be close to -1.  If f > 0, an explanation could be that the act of fastening the belt reminds the driver of the possibility of crashing, and thereby induces more careful driving.  As the evidence places f close to 0, we assume f = 0 for the case of mandatory belt-wearing laws.


3. Safety increase less than expected (-1 < f < 0)


      Rumar et al. [1976] unobtrusively measured speeds maintained on a curve by Swedish drivers in cars with studded or unstudded tires under actual highway conditions.  They find that cars with studded tires were driven faster, but that the overall safety margin, taking account of estimated frictional forces, was still greater for these vehicles than for cars with unstudded tires.  Thus, cars with studded tires were less likely to skid than those without, but not as much less as the increased friction of the studded tires alone would indicate.

      Evans and Herman [1976] investigated how a car's acceleration capabilities affected the last moment a driver in a stationery car was willing to cross in front of an oncoming car.  The study was performed on two non-public roads intersecting at right angles, with the oncoming car driven at constant speed.  The acceleration capabilities of the car driven by the subjects could be set by the experimenter at normal (high performance) or restricted (low performance).  The same subject performed series of trials alternatively under these conditions, with each series being preceded by a number of standing starts to familiarize the subject with the condition.  Under the higher acceleration condition, subjects were prepared to cross with a gap 0.37 s shorter than under the lower acceleration condition.  This compares to a physical safety margin difference of 0.50 s.  Thus most, but not all, of the 0.50 s increase in safety margin was consumed by accepting shorter gaps.

      Anti-lock braking (ABS) uses electronic controls to maintain wheel rotation, and thereby maximum vehicle control with near optimal breaking, rather than allowing the wheels to lock under hard braking, as in non-ABS systems.  This increases vehicle stability, especially when tire/roadway friction is reduced or varying, as when the pavement is wet, and generally reduces the minimum stopping distance.  The technical advantages of such systems is demonstrated clearly by Rompe, Schindler, and Wallrich [1988] in test track experiments using two instrumented cars, one with and one without ABS, but otherwise similar.  Shorter braking distances and superior obstacle avoidance are found for five driving tests simulating driving maneuvers with a high crash risk element (straight line braking, braking on a slippery curve, etc.).  Based on analyzing historical traffic crash data for a non-ABS vehicle fleet, Langwieder [1986] predicts that universal ABS in Germany could diminish severe crashes by 10 to 15%, which, in the present formalism is an estimate of DSEng.  Biehl, Aschenbrenner, and Wurm [1987] estimated DSAct by comparing the crash experience of groups of Munich taxi drivers randomly assigned vehicles with and without ABS.  They report decreases in the numbers of some types of crashes, but increases in others, for no net overall change.  It appeared that the ABS system was inducing behavioral changes, such as reductions in caution on snow and ice, beyond those justified by the advantage of ABS under such circumstances.  Because the severity of such apparently induced crashes was less than that for the crashes prevented, the study suggests that the ABS system reduces harm, but by less than expected.  Additional data from German insurance companies show that vehicles equipped with ABS tend to have higher crash rates than similar vehicles not so equipped, possibly due to riskier drivers choosing to purchase the ABS system [Huguenin 1990].

      Two other studies indirectly suggest that improved braking may be used for other than safety.  In both, car age serves as a surrogate for braking, because it is plausible that as vehicles age, their stopping distances increase as tires and brakes deteriorate.  Evans and Rothery [1976] examined observational data on traffic behavior at two signalized intersections and find that, when cars stopped, newer cars used higher levels of deceleration than older cars.  When cars proceeded, drivers of newer cars were more likely
than were drivers of older cars to enter the intersections after the onset of red (that is, to be in violation of the traffic code).  They comment, "It is possible that the drivers of older vehicles are adjusting their behaviour to compensate for the reduced mechanical condition of their vehicles" ( p. 569).  Kahane [1989] examines rear-end crashes, and finds a highly regular pattern in which the probability that a car was struck in the rear, given that it was involved in a crash, declines systematically with car age.  If a seven year old car was involved in a crash, the probability that it was struck from the rear is about 30% lower than the corresponding probability for new cars.  Thus the findings of Evans and Rothery [1976] and Kahane [1989] suggest behavioral responses to cars being in newer condition, with better braking likely the dominant factor.  Although it is not possible to infer whether the change in behavior still leads to a net safety increase, we assume this to be the case, and therefore categorize improved braking, as indicated by car age, as -1 < f < 0.  Even more difficult to categorize, and therefore not shown in Fig. 11-1, is a simulator study finding that improved roadway delineation is associated with faster curve entry speed [Ranney and Gawron 1986].


4. Safety measure has no effect (f = -1)


      In principle, one can never show that an intervention has no effect, only that its effect is less than some value.  Indeed, as stressed in Chapter 1, it is almost inconceivable that one variable will be entirely unaffected by another for which there is any conceivable avenue of influence.  The inclusion of driver education and mandatory vehicle inspection in this category is based on an absence of any indication of a safety effect, although many studies have examined such questions, usually with a view to showing efficacy.  Haddon [1980, p. 51] writes, "No one has yet shown with the required formal research
if periodic motor vehicle inspection makes any clear difference."  Findings of reductions in specific types of crashes, such as those in which bald tires were a major factor, do not constitute proof of efficacy, because the compelled purchase of new tires will likely increase other types of crashes.  As in the case of better brakes, bald tires affect the mix of crashes, but with an unknown influence on the total number.  The inclusion of traffic signals (which are introduced mainly to reduce delay) is based on the study of Short, Woelfl, and Chang [1982], who find that introducing traffic signals influences the types of crashes without affecting the total number.


5. Perverse effect -- safety measure reduces safety (f < -1)


      Herms [1972] investigated the effect on pedestrian safety of painted pedestrian crosswalks.  During a five-year observation period, 177 pedestrians were hit in 400 marked (or painted) crosswalks compared to 31 hit in 400 unmarked crosswalks.  This included 18 fatalities in the marked crosswalks versus 3 fatalities in the unmarked crosswalks.  Part of the difference was due to more pedestrian traffic at the marked crosswalks.  However, relative to the numbers of pedestrians using the crosswalks, approximately twice as many pedestrians were struck in the marked crosswalks as in the unmarked crosswalks.  This large-scale, well-executed study provides clear evidence of a case in which a safety improvement led to a dramatic reduction in safety.  It would appear that the painted crosswalk induced a sense of security in the pedestrians that was not justified by any increase in caution on the part of the drivers approaching it.

      Hakkert and Mahalel [1978] examined the effect of the "blinking green" phase of signals in Israel on crash rates.  The green light is set to blink for the last two or three seconds of the green phase to warn drivers of the
impending yellow phase, and thereby reduce the probability that drivers are trapped in the so-called "dilemma zone" [Gazis, Herman, and Maradudin 1960; Herman, Olson, and Rothery 1963].  Based on various before versus after comparisons, Hakkert and Mahalel [1978] conclude that the installation of the blinking green phase led to an increase rather than the intended decrease in crashes.  Klein, Zaidel, and Mahalel [1983], using a movie film technique, conclude that the blinking green places additional decision pressure on the driver and creates greater opportunity for error.  It is also tempting to surmise that the drivers were attempting to use the additional information provided by the blinking green to increase their chances of clearing the intersection before the onset of red, rather than to increase their safety, as intended.  The blinking green stimulated an early decision to proceed, which sometimes had to be changed to a decision to stop, hence precipitating increased numbers of rear-end crashes.  These findings have implications relative to recurrent proposals for "count-down" traffic signals that indicate the amount of green time remaining.

      As discussed in Chapter 6, Williams and O'Neill [1974] find that on-the-road crash rates of Sports Car Club of America national competition license holders are higher than those of matched comparison drivers.  This would be an example of a safety measure (increased skill) producing a perverse safety effect only under the assumption that the increased skill acquired in obtaining this license led to higher risk driving, and consequently higher crash rates.  However, another interpretation is that high risk, high crash rate, drivers sought the special license, and without the skills honed in obtaining it, their crash rates might have been even higher; hence the ? in Fig. 11-1.

      Asogwa [1980] examined motorcyclist fatalities in Anabra State, Nigeria, before and after the introduction of a law mandating the wearing of helmets. 
A substantial increase in fatalities (from 5 to 18) and injuries (70 to 145) occurred in a two-year period after the legislation compared to a two-year period before the legislation.  The number of registered motorcycles also increased from 5303 to 7071 giving before and after rates of 94 and 255 fatalities per 100 000 registered motorcycles per two years, and 13.2 and 20.5 injuries per thousand registered motorcycles per two years.  Thus, after the introduction of the law, the data indicate that the fatality rate increased by 171% and the injury rate by 55%.  Taken at face value, the data indicate large reductions in safety following a change expected to increase safety.  However, given uncertainties in the data (for example, ongoing improvements in data collection completeness could generate a false indication of increased casualties), and how surprising the result is in the context of studies discussed in Chapter 10 of helmet-use laws involving enormously more data, I have characterized the study with a ? in Fig 11-1.  It is possible that it could be a real effect, but highly jurisdiction dependent.




      In many cases it is arbitrary whether we think of a change in one direction increasing safety or in the opposite reducing safety.  Thus we can think of switching to a larger car as increasing safety, or switching to a smaller car as decreasing safety.  In the same way, a number of items in Fig. 11-1 could be transferred to Fig 11-2, and vice versa.


1. Perverse effect -- safety actually increases (f < -1)


      In Chapter 4 we saw that serious injury and fatality risk declines as roadway friction declines, a finding corroborated by lower fatality rates in
winter than summer months.  While the number of crashes may increase, this finding still identifies inclement weather (though not man-made, like all the Fig 11-1 items) as a fatality and injury reducing factor, contrary to the intuitive impression that it should have the opposite effect.  This is probably another manifestation of the braking effects discussed earlier -- when braking capabilities decline, whether due to vehicles aging, the absence of ABS, or to roadways becoming more slippery, drivers take more care.

      Large user behavior feedback effects as a function of car mass in two-car crashes are shown in Evans [1985b] (see also in Chapter 4).  Relative to the numbers of vehicles registered, and correcting for driver age effects, 30% fewer injuries and fatalities resulted from small cars crashing into small cars than from large cars crashing into large cars.  This occurred despite a driver in a small-small crash being 2.35 times as likely to be seriously injured or killed as a driver in a large-large crash.  Thus, the more dangerous case, from an engineering point of view, turns out to be safer because of its reduced occurrence.

      Prolonged driving can undoubtedly reduce safety [Brown 1982; Hertz 1988].  However, one study [McFarland and Moseley 1954, p. 240-248] provides a counterexample.  The study was conducted by placing trained observers in 17 long-haul trucks on 20 trips travelling some 8000 km.  The numbers of "near accidents" (as judged by the observers) for the first, second, third .... and ninth hours of driving were, respectively, 11, 11, 6, 4, 4, 4, 4, 3 and 1.  This may have occurred because of reduced risk taking with increased duration of driving [Fuller 1984a].  It is possible that effects in experiments of this type could be different from normal driving because higher arousal levels are likely when drivers know their performance is being monitored.

      Short-term safety decreases were anticipated in two Scandinavian countries when they changed from driving on the left- to the right-hand side of the
road.  However, for both Sweden [Na?"a?"ta?"nen and Summala 1976, p. 139] and Iceland [Wilde 1982, p. 215], the changeover was in fact followed by substantial drops in traffic fatalities.


2. No effect on safety (f = -1)


      The evidence available on the safety of drivers with only one eye (Chapter 5) does not support that they have crash rates different from average drivers, suggesting f = -1 for monocular vision; the evidence is so flimsy that the value is qualified by a ? in Fig. 11-2.  The nominal indication from the data is in fact lower crash rates for monocular drivers, suggesting f < -1.


3. Safety reduction less than expected (-1 < f < 0)


      The car mass effects summarized in Table 4-2 show that, for single car crashes and for two-car crashes, the effect of decreased mass on the risk of death, given a crash, is larger than is reflected in the number of fatalities per registered car.  These results imply [Evans 1985c; 1984] that drivers are reducing some of the larger fatality risk of the smaller car by lower involvement rates, findings that are further supported by observational data associating lower risk taking with smaller cars [Wasielewski and Evans 1985]. In an experiment, Noguchi [1990] finds that the larger of two cars was driven faster (each car was driven by different subjects) and Wasielewski [1984] finds higher speeds for drivers of larger cars in traffic.

      Fuller [1984a] had 12 army truck drivers drive one truck following another for 11 hours on each of four consecutive days.  Drivers reported symptoms of performance deterioration, drowsiness, and exhaustion, and often desired to stop driving.  However, time headways systematically increased with driving
time and self-reported performance deterioration, providing evidence of compensatory adjustments in following distance when fatigued.  Evans and Wasielewski [1983] find that drivers adopt longer following headways when passengers are present than when driving alone.  Mackay [1985] reports that accompanied drivers choose lower speeds than those travelling alone.  Thus the potential deterioration in driving performance due to being distracted by passengers is compensated for by more cautious driving as reflected in longer headways and slower speeds.  The compensatory responses to fatigue and distraction only establish that f < 0.  Assigning them to -1 < f < 0, rather than other f < 0 regions, is little more than a guess.


4. Safety reduction as expected (f = 0)


      In Chapter 10 we find that observed fatality increases in the states in the US that repealed mandatory motorcycle helmet-wearing laws are in satisfactory agreement with increases calculated using realistic estimates of reductions in helmet use and the when-used effectiveness of helmets.  Such agreement between actual and expected changes indicates that the most likely value for f is close to 0, though the evidence does not preclude small negative or positive values, positive being somewhat more likely.  Mechanisms that could give rise to larger than expected fatality increases are increased use of the motorcycle, or the removal of the safety reminder that the helmet might have provided.  The data convincingly reject f = -1.


5. Safety reduction even greater than expected (f > 0)


      Earlier in this chapter we discussed the possibility that the introduction of the 55 mph speed limit in the US in 1974 may have induced reductions in
casualties on roads with unchanged speed limits.  A parallel phenomenon may have occurred when the speed limit on portions of the rural Interstate system was increased to 65 mph in 1987.  Brown, Maghsoodloo, and McArdle [1989] find evidence that property damage crashes increased on stretches of Alabama Interstate highway on which the speed limit remained fixed at 55 mph when the speed limit on other sections increased to 65 mph.  Such spillover effects would mean that reductions in safety from speed limit increases would exceed those experienced only on the affected roadways.




      Figs 11-1 and 11-2 show 24 comparisons of actual safety changes to the changes expected assuming no road-user behavior change.  For changes made to increase safety, examples were presented in which

            1. Safety increased even more than expected

            2. Safety increased as expected

            3. Safety increased, but less than expected

            4. No change in safety

            5. The change actually decreased safety -- a perverse effect.

For changes expected to decrease safety, examples were presented in which

            1. The change actually increased safety -- a perverse effect

            2. No change in safety

            3. Safety decreased, but less than expected

            4. Safety decreased as expected

            5. Safety decreased even more than expected.

      The results show that the non-interactive calculation may give not only an erroneous estimate of magnitude, but sometimes may give even the incorrect sign.  While a case could be made to reclassify some of the entries in Figs
11-1 and 11-2 into adjacent categories, the overall finding that behavior feedback effects are widespread in traffic safety systems seems beyond reasonable dispute.

      The widespread occurrence of behavior feedback poses a difficult problem in estimating the expected benefits of proposed countermeasures.  All  countermeasures have associated values of f which, if much different from zero, will crucially influence their efficacy.  The only way to determine f is empirically; this can be done only after implementation, and usually with great difficulty.  However, prior experience does suggest some sufficiently robust patterns to provide substantial guidance.

      Because of the self-paced nature of the driving task, technical changes that are readily apparent to the driver are very likely to induce user responses.  Thus improved braking, handling, tire-road friction, headlights, and so on, are likely to induce increases in speed, enjoyment, relaxation, etc.  One can anticipate with considerable confidence that safety increases from such measures will be lower than expected.  (As an additional piece of anecdotal evidence, I consider that my own greatest risk of rear-end collision was experienced as a passenger in a car demonstrating an experimental radar-braking device designed to prevent rear-end collisions!)

      There is no case of a safety change invisible to road users which has generated a measurable user response.  The aggregate influence of such inputs as drivers reading about the changes seems obviously miniscule, as is the possibility of effects due to observing reductions in injuries to their acquaintances or from studying data.

      For highly visible safety changes which influence only the probability of death or serious injury, but not the probability of crashing, there is little evidence of important behavior response.  This is consistent with the prevailing view in criminology that it is probability of something unpleasant
happening, rather than the degree of unpleasantness, that has the much larger influence on behavior.  In the traffic safety context, the view that probability of detection is more important than severity of punishment in deterring undesired behavior has been persuasively presented by Ross [1984a; 1984b].  A high probability of a minor adverse consequence exercises a much larger influence on driver behavior than, say, the factors influencing the probability of being killed or injured.  The prospect of a $20 fine (and other associated ramifications) controls traffic speeds more than does the relationship between fatality risk and speed.  The probability of death is so improbable and abstract that its reduction through the use of a device such as a safety belt or airbag is unlikely to exert much influence driver on behavior.  The empirical evidence shows little indication of changes in driver behavior with the introduction of mandatory safety belt wearing laws, or the repeal of helmet-wearing laws.  The evidence (Chapter 10) convincingly precludes the possibility that such measures generate large user responses.

      For the case of car mass effects, the cars with the higher fatality risk have lower crash rates.  This more likely flows from directly observed differences in performance, handling properties, stability, and noise levels than in the expectations of outcome, given that a crash occurs.  Rothengatter [1988] suggests that a tendency of drivers with inclinations towards higher speeds to choose larger (more powerful) cars could also contribute to such effects.  Scenarios in which outcome is expected to affect behavior can be imagined; it seems highly likely that a car loaded with dynamite set to explode at the smallest impact would be driven more carefully than an average car.  However, this is because the probability of death has been elevated to a real possibility rather than the distant abstract concept in normal driving.




      The information summarized in Eqn 11-1 and Figs 11-1 and 11-2 does not constitute a model of driver behavior.  The formalism is an attempt to classify observed effects into an organized framework.  Its focus is phenomenological rather than explanatory, listing what happens rather than why it happens.

      Attempts to find unifying principles or models underlying driving behavior go back over half a century [Gibson and Crooks 1938].  Given the diversity of responses to interventions, the task is clearly a formidable one.  Despite the difficulties, a number of researchers have made serious attempts to shed light on processes underlying driving behavior, as discussed below.


Accident proneness


      The notion of "accident proneness" has been controversial in traffic safety [Haight 1964; Shaw and Sichel 1971; McKenna 1983] since first introduced by Greenwood and Woods [1919], who compared the distributions of  various mishaps with several hypothetical distributions.  If all individuals were equally likely to be involved, then the distributions would be Poisson, whereas greater involvement likelihood on the part of some would generate different types of distributions, as was indeed found to be the case.  Some interpreted this to imply that individuals have a stable trait that defines their involvement rate.  Such a notion finds support in any casual examination of the distribution of crashes -- it always turns out that a small fraction of drivers accounts for a large fraction of crashes.  However, this is what randomness alone (as reflected in the Poisson distribution) would generate, even if all drivers had identical expected crash risks.  Because crashes are
so infrequent, an individual driver's prior crash rate would not be effective pedictor of future crash rate even if some individuals did have expected rates higher than others.  While correlations of high statistical significance can be obtained, they are of only modest practical use; most high crash-rate drivers in one period will be average in subsequent periods, while most high-crash rate drivers in subsequent periods will have been previously average [Peck, McBride, and Coppin 1971; Gebers and Peck 1987].

      Some have misinterpreted this fundamental lack of high predictability to mean that all drivers are almost equally likely to crash.  Nothing could be further from the truth.  One can conclude with the utmost confidence (insurance companies do it every day) that some groups of drivers (young males) have crash rates well above average.  At the individual driver level, some factors, such as prior driving violation record [Peck and Kuan 1983], are useful predictors of future crash rates.  The term "accident proneness" has been so controversial, and used to mean such different things, that I share McKenna's [1983] recommendation that it be abandoned entirely, and that traffic safety be discussed using more clearly defined terms, such as "differential crash involvement".


Skill model


      Before the 1970's most effort to increase understanding of driver behavior focused on driving as a perceptual-motor skill.  Crashes were interpreted as failures of driver skill.  This approach implies that safety is determined mainly by the driver's level of skill in relation to the situational demands on that skill.  Thus safety is increased by increasing driver skill and reducing environmental demands. 

      The skill model of driver behavior is incompatible with many of the central findings in Chapter 5; driver training and education have not been shown to have much influence on crash rates, drivers with the highest perceptual-motor skills and interest in driving (young males) also have the highest crash rates, and high-skill drivers have above average crash rates.  The clear failure of the skill model underlines the need to consider motivational models [Summala 1988] that incorporate the self-paced nature of the driving task [Na?"a?"ta?"en and Summala 1976] in which drivers select their own level of task difficulty.  Because the chosen level of task difficulty depends on driver evaluation of their own skill levels, and of their evaluation of environmental demands, changes in skill, vehicle, and roadway do not exercise a straightforward influence on safety.


Utility maximizing


      Utility theory has its origins with Jeremy Bentham and John Stewart Mill at the beginning of the 19th century.  It occurs in economic theory, often in the form of benefit/cost analysis.  In applying it to driving, the basic assumption is that the driver has a goal that can be written as a "utility function" in which desired quantities have positive signs and unwanted consequences have negative signs.  The driver strives to maximize this utility function [Bloomquist 1986], which can be solved mathematically if all quantities are expressed as functions of simple variables.  Most mathematical efforts, starting with the insightful paper by O'Neill [1977], have concentrated on the single variable speed; the desired goal is to save time, and the unwanted consequence is harm from crashes, which is expressed as an assumed function of speed.  The ideal, or optimum, speed is the one which maximizes the utility function.  At this speed the expected benefits and costs
are in equilibrium balance; a slightly slower speed leads to a greater increase in delay than is considered worth the corresponding decrease in crash risk.  Safety improvements modify the expected harm versus speed function, leading to higher optimum speeds.  A hypothetical "rational" driver uses the increased safety to reduce travel time as well as to increase safety, and in choosing the optimal mix will increase safety, but by less than the amount of the safety improvement.

      Such models are mathematically interesting, and offer insights into the region  -1 < f < 0.  Their major inadequacy is largely captured in the quip, "Economics is the science of how people ought to behave."  The unitary goal of minimizing trip time seems to be given a far more central role in traffic (not just safety) than it merits.  While an employer paying a driving employee by the hour has no trouble measuring the value of time, it is not clear what the concept means in general.  Most driving is substituting one activity for another, and may be more or less desired than the alternative.  The value-of-time concept seems based on some underlying notion that driving is disliked, and that every second of driving is spent desiring to do something else.  When I see people leaving work, relaxed, waiting for elevators that are much slower than using the stairs, ambling to their cars, and so on, I find it implausible that when they enter their cars they are in any sense balancing a few seconds less in this activity for some increased risk of death or injury.  Mackay [1990] mentions an origin and destination study conducted in August in Kuwait which finds that 30% of car occupants were not going anywhere in particular but were just driving around in their air-conditioned cars to keep cool!  Assuming a time-minimizing goal seems to suggest, implausibly, that such drivers have zero crash risk; my guess is that their crash risks are lower, but not dramatically lower, than those of other drivers in the same traffic.

      The mental construct of utility maximizing also seems inappropriate for the typical fatal traffic crash involving an intoxicated driver in the early hours of Saturday or Sunday morning.  It is difficult to imagine such a driver consciously increasing his speed in order to arrive home at 1:42 a.m. rather than 1:45 a.m.  To claim that the utility maximizing equation included decisions on how much to drink which were solved prior to, or during, drinking is merely to underline the intrinsic complexity of the problem.  As American drivers spend, on average, about 7.5 hours per week driving [Horowitz 1986], it is similarly difficult to conceive that most of this time is spent, even to a modest approximation, or even in some collectivist sense, maximizing utility functions. 


Driver risk


      The need for drivers to estimate risk is central to utility maximizing models, and also to other approaches to modelling traffic safety.  While risk may play a central role in controlling behavior in some driving situations, it may be less important in others.  Risk estimation likely plays a crucial role in the behavior of a racing driver who derives enormous profit by travelling just a little faster.  Historical data show that drivers in the Indianapolis 500 race had fatality rates (per unit distance of travel) 1400 times that of the average on-the-road driver (Chapter 13); if the racing rate applied to the US driver population, there would be over 30 million annual driver fatalities!  It seems implausible that the same basic processes governing racing-driver behavior would apply, in dramatically scaled down form, to normal driving.  What other relationships governing social phenomena apply over three orders of magnitude?

      The term "risk' captures many distinct meanings [Haight 1986].  Most driver models focus on "subjective" risk, or what the driver perceives the risk to be, in contrast to "objective" risk, as measured using data.  While this is conceptually orderly, it can raise insurmountable practical problems.  If subjective risk is to have any validity as an explanatory variable, it must be measurable.  Attempts have been made to do this by physiological monitoring [Taylor 1964], but apparently not in recent decades.  If subjective risk is estimated using such objective data as crash rates, then one is likely to end up with no more than the tautology that crash rates are related to crash rates.  On the other hand, conceding that subjective risk is not really measurable discounts its use as a factor in a scientific explanation.  Models based on speculations about what goes on inside the driver's head are of minimal value unless they can explain observed effects better than less speculative models.


Do drivers seek risk?


      Examples of drivers actively seeking risk have encouraged the notion that drivers essentially always seek some level of risk.  Such an inference from specific examples to the general is not justified.  For example, unquestionably some individuals (generally of the sex and age associated with high traffic-crash rates) actively seek high levels of pain on some occasions.  A scene in the movie "Lawrence of Arabia" comes to mind; T.E. Lawrence extinguishes a lit match by slowly engulfing the flame between his thumb and index finger.  An unlooker attempts to copy, screams out in pain, and asks, "What is the trick?"  Lawrence answers, "The trick is not minding that it hurts." (Lawrence died in a motorcycle crash).  It is of course arguable whether it is indeed the pursuit of  pain that motivates such behavior. 
However, even conceding that some people seek pain on some occasions does not imply that all people seek some desired level of pain at all times, nor that there is some pain threshold below which life gets so boring that people seek to elevate pain to above that level.  Similarly, rather than seeking some level of risk, it seems more plausible to me that most of the time most drivers seek the lowest possible, or zero, level of risk.  Fuller [1989] writes, "Only very special road users, such as homicidal maniacs, putative suicides ... intentionally opt for a greater chance of collision," and he quotes Shakespeare [Macbeth, Act III, Scene 1] in support "To be thus is nothing; but to be safely thus."  Speed and high performance are sought for the sensations of pleasure and excitement they induce, as discussed by Rothengatter [1988], and with unrestrained gusto by Bayley [1986] in his book "Sex, drink and fast cars."  To claim that the risk of crashing somehow encourages fast driving seems to me barely more plausible than claiming that the risk of contracting a sexually transmitted disease encourages sexual activity.


Comment on "risk homeostasis theory"


      "Risk homeostasis theory" claims that drivers have a target level of risk per unit time, so that physical changes to the traffic system stimulate user reactions that reset safety to its prior level [Wilde 1982; 1986].  The finding that most values of f in Figs 11.1 and 11.2 are not equal to -1 is sufficient to dismiss this claim, which is even more definitively refuted by crash data [Shannon 1986; Evans 1986].  The claim has nonetheless found willing debaters for two decades.  The tone of advocacy for the claim has been largely philosophical, metaphysical, and theological in nature, unencumbered by the standards, methods, or norms of science, and at times happily
abandoning the rigors of Aristotelian logic and the multiplication table.  One can but marvel that repeated claims so clearly devoid of face validity have been debunked in such respectful tones by so many of us [Slovic and Fischhoff 1982; Graham 1982; Orr 1982; McKenna 1982; 1985; Shannon 1986; Evans 1986; Summala 1985; 1988; Rothengatter 1988; Michon 1989].  The journal "Ergonomics" justifies devoting much of a whole issue (April 1988) to airing it yet again on the grounds that there is somehow a legitimate debate; to me this seems as plausible as devoting an issue to the proposition that the earth is flat simply because a few adherents, who conjure up ad hoc explanations for every piece of contrary evidence, still claim it is, and that the issue must not be considered settled until these believers concede.

      Fatalities per unit distance of travel on non-Federal aid rural arterial roads is over 800% higher than on urban Interstates (Table 4-4), and for total US travel was over 900% higher in 1921 than in 1988 (Fig. 13-2); applying reasonable corrections to convert fatalities per unit distance of travel to fatal crashes, injury crashes, or property damage crashes per unit time of travel will in some cases increase, and in other cases decrease, these differences, but cannot conceivably reduce them to values anywhere close to zero.  Such large differences show beyond doubt that average driver risks are different on different roads, and change in the long term.  Risk also changes by large amounts during an individual trip; the risk of crash, injury, and death is enormously higher when driving through intersections than when driving between intersections.  Although most drivers must surely be aware of this, they are in little better position to equalize these risks than is a pilot to equalize the risk per unit time at landing to that when cruising at 35 000 feet.  Replacing roads containing intersections and junctions by limited-access freeways reduces crash risk as certainly as eliminating the take-off and landing risk would reduce air-travel risk.


      When the homeostasis notion first appeared over two decades ago, it played a positive role in stimulating thinking about interactive effects, and highlighted the importance of motivational as well as engineering factors.  Even though copious data have always been available to dismiss specific claims, one paper on homeostasis might have been of sufficient interest to merit publication.  As criticisms multiplied, the "theory" changed and became a moving nebulous target claiming everything and nothing.  When every re-interpretation was refuted by observation, claims of observable consequences were abandoned, and with them all possibility of experimental refutation (a requirement of any scientific theory).  At the internal motivational level, drivers could be maintaining something constant.  It is not impossible that belted drivers travelling today on Interstate freeways in some sense "feel" as safe as did drivers travelling on rural two-lane roads in 1920.  However, if the notion of a possible constancy in perceived risk is to be discussed sensibly, it is imperative to use some term different from one previously used to refer to a long series of convincingly refuted propositions.

      The endless regurgitation, fixing-up, and ignoring of the basic issues has regrettably flourished because traffic safety research has not yet (see Chapter 14) acquired the methods, style, values, attitudes and institutional structures which have proved so successful in the traditional sciences, especially the tradition of requiring publications to be peer-reviewed.  In science, a coherent rational explanation, or theory, of things that happen may well be incorrect.  Science has no term for explanations of things that do not happen!  The use of the word "theory" is without justification; the claim that risk per unit time is a constant is no more a theory than the claim that all people are the same height, or think they are the same height.  Haight [1986] comments:


            There is some question as to whether the theory is meaningless (since incapable of testing) or simply false.  Evans' [1986] conclusion that "there is no convincing evidence supporting it and much evidence refuting it" is if anything generous.  In my view, a sufficient argument against the validity of risk homeostasis is provided by the incoherence of its "theoretical" formulation.


Motivational approaches


      The zero-risk model [Na?"a?"ta?"nen and Summala 1976; Summala 1988] avoids some of the difficulties surrounding estimation of risk by assuming that drivers aim at zero subjective risk.  The two starting points of the model are, first, the motivational basis of driver behavior, and, second, the adaptation to perceived risks on the road.  Drivers seek increasing speeds for a variety of motivations in addition to saving time, such "extra motives" include pleasure, showing-off, and competitive urges.  In terms of the example of driving around a curve [Summala 1988, p. 498], the driver has a subjective estimate of the maximum possible speed at which the curve can be negotiated.  Because of variation in various factors, such as coefficients of friction, the objective maximum speed will vary by more than the driver's subjective estimate of it.  The driver chooses a speed lower than the subjective maximum by an amount judged to generate a safety margin that is associated with essentially zero risk of crashing.  Because the chosen speed relates to the perceived maximum safe speed, this chosen speed will increase in response to roadway and vehicle improvements.  Each time the driver safely negotiates the curve, there is positive reinforcement that the chosen safe speed was risk-free, encouraging speeds to creep upwards.  Even as speeds creep upwards, they will still be perceived subjectively to be as risk free as earlier lower
speeds, assuming that no incident has occurred to provide the driver feedback that the chosen speed was too high.  Because of variations of the driver's actual speed around the selected chosen speed, and variations of the maximum physical safe speed around the subjective estimate of the maximum safe speed, the probability that the driver's actual speed will exceed the maximum safe speed will increase steeply as the driver's speed creeps upwards.  In order to counteract the tendency of speeds to drift upwards, Summala [1988] recommends external constraints in the form of enforced speed limits, and underlines this by pointing out that large casualty reductions can reliably be attributed to speed limits.  Speed limits reflect society's accumulated knowledge about a safe speed for the curve which the driver cannot satisfactorily acquire by trial and error, because the errors are too costly. 

      Fuller [1984b; 1988] explains aspects of driving in terms of a threat avoidance model.  He uses "threat" because most of the time on the roadway the driver is not dealing with aversive stimuli but potential aversive stimuli or threats; he uses "avoidance" because for much of the time the driver seems to be either avoiding aversive stimuli (for example, steering around obstructions) or avoiding the possibility of aversive stimuli arising (for example, reducing speed or selecting a clear lane).  The driver is not so much trying to avoid crashes, but trying to avoid unpleasant experiences, which in some cases might be precursors to crashes.  By analogy with animal experiments, Fuller [1988] discusses that drivers might have a bias towards postponing taking unwelcome actions (such as slowing down) when confronted with potentially unpleasant experiences.  When postponing undesired responses leads to no undesired consequence, the habit of postponing can be reinforced.  The later a response, such as slowing down, to a potentially hazardous traffic situation, the greater is the actual crash risk.

      Boyle and Wright [1984] discuss "accident migration", the notion that when a stretch of roadway with a high crash rate (a "blackspot") is treated, crash reductions at the blackspot lead to increases elsewhere.  For example, if, a road has a sharp curve, then drivers are going to slow down to negotiate it, thereby travelling slower both before and after the curve; some are going to experience fear or loss of control, and a few of these will crash, with consequent influences on subsequent behavior.  If the curve is straightened and widened, it seems inevitable that safety must be less on the roadway portions near the curve.  Although, in principle, it is possible that correcting some blackspots could even increase crashes overall, it is exceedingly difficult to examine such possibilities empirically.  Even the worst blackspots have only a handfull of crashes per year, so that it would be almost impossible to find convincing evidence of an increase in crashes over the number expected.  Even aggregating many blackspots cannot provide clear cut evidence because of the many other factors involved, even though Wright and Boyle do report a 10% increase in adjacent links and nodes.  It appears that these differences [Maher 1990] may have been due to other causes.  The roadways with the fewest (or zero) blackspots, namely, limited access freeways, have substantially lower crash, injury, and fatality rates than average roadways, which have more blackspots.  This demonstrates that, in the limit, removing all blackspots generates a system with higher aggregate safety, even if the safety increase might not be as great as expected ignoring changes in user behavior.

      Many authors (for example, Howarth [1987]) have stressed the potential safety benefits of decreasing objective risk without altering subjective risk, or of increasing subjective risk without changing objective risk.  Denton [1973] provides an interesting example of reducing crash rates by painting a
geometric pattern of bars with decreasing spacing on a roadway to reduce speeds by convincing drivers they are travelling faster than they are. 

      Michon [1985] expresses the hope that the cognitive revolution that has swept psychology will greatly illuminate the driving task, and further suggests [Michon 1988] that it can be implemented in production systems. Michon [1989] claims that rule-based modeling, using some advanced production system architecture is the most promising approach to better theories of driver behavior.  He considers it particularly effective in the context of driver training, in which specific learned rules are of more value than general admonitions to, say, drive carefully.


Economic models


      Partyka [1984] explains much of the variation in fatalities from 1960 through 1982 in terms of various economic indicators (Fig 11.3).  This provides another illustration of the importance of factors other than engineering.  The finding that fatalities increase (more than travel) when economic activity increases does not identify the mechanism leading to such effects.  It has been speculated that discretionary travel depends more strongly on the economy than travel in general, and that discretionary travel involves higher fatality risk than more work-related non-discretionary travel.  This might be part of the reason why fatality rates are so much higher in the summer than in the winter months (Figs 4-8 to 4-10).  Joksch [1984] finds that the index of industrial production is an effective explainer of changes in US traffic fatalities from 1930-1982, and concludes that, as a rule of thumb, the annual percentage change in traffic fatalities is about two thirds the annual percentage change in the index of industrial production.


Fig. 11-3 about here


      Peltzman [1975] used time-series fatality data in a study aimed at estimating the influence of vehicle safety standards, and concluded that they have little net effect.  He interpreted this to indicate that drivers used the extra safety mandated by the standards partly to drive faster (thus killing additional other road users), and partly to increase their own safety, an explanation along the lines of utility maximizing.  However, the data are susceptible to a variety of interpretations (Chapter 4).  As most of the safety standards were relatively invisible to drivers, a substantial driver response is improbable.


Some thoughts on driver modeling


      While many models offer insight into specific aspects of driver behavior (realistic ones do not claim more), it seems unlikely that general models offering much more can be formulated.  The problem flows from an intrinsic dilemma.  For a model to be elegant, and have derivable quantitative values of parameters, it must be simple.  Yet the commuter going home at 5:00 p.m. and the drunk going home at 2:00 a.m. are unlikely to be describable by the same model; two sober drivers of the same age and sex may still have basic differences which frustrate simple explanations.  The quest for simplicity leads to monist models which focus on one aspect of driving, while ignoring others factors much too important to be ignored, thereby violating Albert Einstein's advice that "Everything should be as simple as possible, but not simpler."

      The most common monist choice has been risk, often being even narrowed further to subjective risk of crashing.  There are many problems associated with affording risk a central role in controlling driver behavior.  Individuals are extremely poor at accessing risk based on personal experience.  There is no reason to expect that people would avoid X-rays or foods containing cholesterol were it not for information from the mass communication media.  Yet news stories and the reactions to them can be extraordinary.  Reaction in the US to two Chilean grapes containing a non-health-threatening 3 mg of cyanide resulted in 17 000 Chileans losing their jobs; enormous quantities of apples and apple products were discarded after essentially baseless claims regarding the chemical alar [Wall Street Journal 1989].  People misjudge the risk of common hazards; for example, a group of 30 college students underestimated the number of deaths due to smoking by a factor of 62, due to traffic crashes by a factor of 5, but overestimated the number of skiing deaths by a factor of 4 [Slovic, Fischhoff, and Lichtenstein 1980]. Zeckhauser and Viscusi [1990] point out that large risks are ignored while some small ones are regulated stringently.  Adams [1985, p. 145] suggests that, on average, people may be able to rank order risks because they can rank order weights, as illustrated in experiments that showed the average rank-orderings of many subjects approached the correct rank ordering with increasing number of subjects.  This is nothing more than an illustration the high sensitivity that subjects exhibit in two-category forced choice judgments, as well documented in the psychological literature [Woodsworth and Schlosberg 1954].  The fact that slightly more than 50% of a large number of subjects forced to chose will correctly judge a 101 gram weight to be heavier than a 100 gram weight says nothing about how large numbers of subjects will judge which is riskier --taking a bite from an alar-treated apple or violating the 55 mph speed limit.


      Even if drivers could somehow accomplish the improbable, and collectively estimate risk, there is still little reason to conclude that this is the only important factor controlling driver behavior.  Rothengatter's [1988] conclusion that motivations for, say, speed choice involve many factors, and that risk can account for only part of the variance found in road-user behavior seems beyond reasonable dispute.  He further suggests that seeking pleasure may be as plausible a controlling factor as risk.  Rumar [1988] also considers that risk plays no more than a minor role.  When late for an appointment, the need to save time can elevate risk to a central factor in speed choice.  On other occasions, quite different motivations clearly seem to apply, such as a young man's desire to impress or frighten a girl friend with his driving, absent both destination and time schedule, or the same young man's desire to impress the same girl's mother with his responsible prudent behavior.  In Chapter 13 we discuss social norms as exercising an influence on driving -- in part drivers drive at the speed that they think those whose esteem they crave would expect. In the US in November 1973 motorists reduced speeds to conserve fuel, with risk not entering as a factor.  It seems to me entirely implausible to think that risk (of crashing, or receiving a police violation, or whatever) can, even as the crudest approximation, be conceived as the sole determinant of driver behavior.  As risk is not necessarily a dominant, and certainly not a sole, determinant of driver behavior, terms like "risk compensation" to describe user responses in the range -1 < f < 0, should be avoided.  All that is observed is a user response; to call it risk compensation is to imply knowledge of why it happened, an implication without justification.

      While attempts to describe driver behavior in terms of a single stimulus, such as risk, are too simple to be realistic, attempts using large numbers of decision-based rules requiring computer processing seem to me to be too
complicated.  Poincare writes (in 1903) that "Science is built up with facts as a house is with stones.  But a collection of facts is no more a science than a heap of stones is a house."  My own feeling is that driving is more than a collection of specific steps - it is more of a holistic process not explainable in terms of a collection of reductionist details. 

      In Chapter 6 we encountered the Tillmann and Hobbs [1949] conclusion "Truly it may be said that a man drives as he lives."  Insofar as driving reflects much of the complexity of life, perhaps a useful general model of driving is almost as unattainable as a general model of life.  Rather than focusing on models which attempt to explain more than can realistically be explained, it might be more fruitful to attempt to model some more specific driving situations.  Two types of crashes which are not understood at the perceptual and cognitive level might benefit from such attention; multiple vehicle crashes in dense fog, some involving over 50 cars and more than half a dozen deaths, and older drivers' over-involvement in intersection crashes.




      Human behavior feedback, or user response, to changes in safety systems may greatly alter safety outcomes.  In some cases the outcome is even of opposite sign to that expected; changes instigated to increase safety have actually reduced safety, while changes expected to reduce safety, but made for other reasons, have actually increased safety.  While no predictive model of how users react to changes is available, some general patterns are apparent.  If the safety change affects vehicle performance, it is likely to be used to increase mobility.  Thus improved braking or handling characteristics likely lead to increased speeds, closer following, and faster cornering.  Safety may also increase, but by less than if there had been no behavior response.  When
safety changes are largely invisible to the user, such as improvements in vehicle crashworthiness, there is no evidence of any measurable human behavior feedback.  Likewise, when measures affect only the outcome of crashes, rather than their probability, no user responses have been measured.  In principle, it is almost certain that users respond in some degree to just about everything of which they are aware.  Empirical studies can never show no user response, but only that user response is less than some amount.




Adams, J.G.U.  Risk and freedom - the record of road safety regulations. Nottingham, UK: Bottesford Press; 1985.

Asogwa, S.E.  The crash helmet legislation in Nigeria: a before-and-after study. Accident Analysis and Prevention l2:213-216; 1980.

Bayley, S.  Sex, drink and fast cars. London, UK: Farber and Farber; 1986.

Biehl, B.; Aschenbrenner, M.; Wurm, G.  Einfluss der Risikokompensation auf die Wirkung von Verkehrssicherheitsmassnahmen am Beispiel ABS. Unfall-und Sicherheitsforschung Strassenverkehr, No. 63, Symposion Unfallforschung '87, Ko?"ln; 1987.

Bloomquist, G.  A utility maximization model of driver traffic safety behavior. Accident Analysis and Prevention 18:371-375; 1986.

Boyle, A.J.; Wright, C.C.  Accident migration after remedial treatment at blackspots. Traffic Engineering and Control 25:260-267; 1984.

Brown, I.D.  Driving fatigue. Endeavour (New Series)  6(2):83-90; 1982.

Brown, D.B.; Maghsoodloo, S.; McArdle, M.E.  The safety impact of the 65 mph speed limit: a case study using Alabama accident data. Washington, DC: National Highway Traffic Safety Administration, report DOT HS 807 425; April 1989.

Denton, G.G.  The influence of visual pattern on perceived speed at Newbridge M8 Midlothian. Crowthorne, Berkshire, UK: Transport and Road Research Laboratory; 1973.

Evans L.  Accident involvement rate and car size. Accident Analysis and Prevention 16:387-405; 1984.

Evans, L.  Human behavior feedback and traffic safety. Human Factors 27:555-576; 1985a.

Evans, L.  Involvement rate in two-car crashes versus driver age and car mass of each involved car. Accident Analysis and Prevention 17:155-170; 1985b.

Evans, L.  Driver behavior revealed in relations involving car mass. In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press, p. 337-352; 1985c.

Evans, L.  Risk homeostasis theory and traffic accident data. Risk Analysis 6: 81-94; 1986. (Also see reply by Wilde and response by Evans. Risk Analysis 6:95-103; 1986).

Evans, L.; Herman, R.  Note on driver adaptation to modified vehicle starting acceleration. Human Factors l8:235-240; 1976.

Evans L.; Rothery, R.  Comments on effects of vehicle type and age on driver behaviour at signalized intersections. Ergonomics l9:559-570; 1976.

Evans, L.; Wasielewski, P.  Risky driving related to driver and vehicle characteristics. Accident Analysis and Prevention l5:l2l-l36; 1983.

Fuller, R.G.C.  Prolonged driving in convoy: the truck driver's experience. Accident Analysis and Prevention 16:371-382; 1984a.

Fuller, R.G.C.  A conceptualization of driving behaviour as threat avoidance. Ergonomics 27:1139-1155; 1984b.

Fuller, R.G.C.  On learning to make risky decisions. Ergonomics 31:519-526; 1988.

uller, R.G.C.  To be safely thus?  Conditions for risk-compensation and its modification. Paper presented to the 1st European Congress of Psychology, Amsterdam, Netherlands; 2-7 July 1989.

Gazis, D.C.; Herman, R.; Maradudin, A.A.  The problem of the amber light in traffic flow. Operations Research 8:112-232; 1960.

Gebers, M.A.; Peck, R.C.  Basic California traffic convictions and accident record facts. Sacramento, CA: California Department of Motor Vehicles,  report CAL-DMV-RSS-114; December 1987.

Gibson, J.J.; Crooks, L.E.  A theoretical field-analysis of automobile driving. American Journal of Psychology 5l:453-47l; 1938.

Graham, J.D.  On Wilde's theory of risk homeostasis. Risk Analysis 2:235-237; 1982.

Greenwood, M.; Woods, H.M.  A report on the incidence of industrial accidents upon individuals with special reference to multiple accidents (1919). Reproduced in Haddon, W.; Suchman, E.A.; Klein, D., editors. Accident research. New York, NY: Harper and Row; 1964.

Haddon, W., Jr.  Advances in the epidemiology of injuries as a basis for public policy. Public Health Reports 95:411-421; 1980.

Haight, F.A.  Accident proneness, the history of an idea. Automobilismo & Automobilismo Indusriale 4:3-15; 1964.

Haight, F.A.  Risk, especially risk of a traffic accident. Accident Analysis and Prevention 18:359-366; 1986.

Hakkert, A.S.; Mahalel, D.  The effect of traffic signals on road accidents -- with particular reference to the introduction of a blinking green phase. Traffic Engineering and Control 19(5):212-215; 1978.

Hedlund, J.H.  Recent US traffic fatality trends. In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press, p. 7-19; 1985.

Herman, R.; Olson, P.L.; Rothery, R.W.  Problem of the amber signal light. Traffic Engineering and Control 5:298-304; 1963.

Herms, B.F.  Pedestrian crosswalk study accidents in painted and unpainted crosswalks. Highway Research Record no. 406, Highway Research Board, p. l-l3; 1972.

Hertz, R.P.  Sleeper berth use as a risk factor for tractor-trailer driver fatality. Accident Analysis and Prevention 20:431-439;1988

Horowitz , A.D.  Automobile usage: a factbook on trips and weekly travel. Warren, MI: General Motors Research Laboratories, research publication GMR-5351; 2 April 1986.

Howarth, C.I.  Perceived risk an behavioural feedback: strategies for reducing accidents and increasing efficiency. Work and Stress 1:61-65; 1987.

Huguenin, R.D.  Personal communication; 1990.

Joksch, H.C.  The relation between motor vehicle accident deaths and economic activity. Accident Analysis and Prevention 16:207-210; 1984.

Kahane, C.J.  An evaluation of center high mounted stop lamps based on 1987 data. Washington, DC: National Highway Traffic Safety Administration, report DOT HS 807 442; July 1989.

Klein, T.; Zaidel, D.; Mahalel, D.  The influence of a "flashing green" on stopping and crossing decisions at approaches to signalized intersections. Haifa, Israel: Technion, Israel Institute of Technology, technical report  83-22; 1983.

Langwieder, K.  Der Problemkreis Bremsen in der Unfallforschung. VII.-Symposium. HUK-Verband, Bu?"ro fu?"r kfz-technik, Mu?"nchen; 1986.

Mackay, M.  Seat belt use under voluntary and mandatory conditions and its effect on casualties. In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press, p. 259-278; 1985.

Mackay, M.  Towards a unified traffic science. IATSS Research -- Journal of International Association of Traffic and Safety Sciences 14(?):??-??;1990. (in press)

Maher, M.J.  A bivariate binomial model to explain traffic accident migration. Accident Analysis and Prevention 22 (in press); 1990.

McFarland, R.A.; Moseley, A. L.  Human factors in highway transport safety. Harvard School of Public Health. Boston, MA; 1954.

McKenna, P.F.  The human factor in driving accidents; an overview of approaches and problems. Ergonomics 25:867-877; 1982.

McKenna, P.F.  Accident proneness: a conceptual analysis. Ergonomics 15:65-71; 1983.

McKenna, P.F.  Discussion on "Risk homeostasis in an experimental context". In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press, p. 143-144; 1985.

Michon, J.A.  A critical view of driver behaviour models: what do we know, what should we do? In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press, p. 485-520; 1985.

Michon, J.A.  Should drivers think? In: Rothengatter, J.A.; de Bruin, R.A., editors. Road users and traffic safety. Assen/Maastricht, Netherlands: Van Gorcum, p. 508-517; 1988.

Michon, J.A.  Explanatory pitfalls and rule-based driver models. Accident Analysis and Prevention 21:341-353; 1989.

Na?"a?"ta?"nen, R.; Summala, H.  Road-user behavior and traffic accidents. Amsterdam, Netherlands: North Holland; 1976.

Noguchi, K.  In search of optimum speed: from the users viewpoint. IATSS Research -- Journal of International Association of Traffic and Safety Sciences 14(?):??-??;1990. (in press)

O'Neill, B.  A decision-theory model of danger compensation. Accident Analysis and Prevention 9:l57-l66; 1977.

Orr, L.  Goals, risks and choices. Risk Analysis 2:239-242; 1982

Partyka, S.C.  Simple models of fatality trends using employment and population data. Accident Analysis and Prevention 16:211-222; 1984.

Peck, R.C.; Kuan, J.  A statistical model of individual accident risk prediction using driver record, territory and other biographical factors. Accident Analysis and Prevention 15:371-393; 1983.

Peck, R.C.; McBride, R.S.; Coppin, R.S.  The distribution and prediction of driver accident frequencies. Accident Analysis and Prevention 2:243-299; 1971.

Peltzman, S.  The effects of automobile safety regulation. Journal of Political Economy 83:677-725; 1975.

Ranney, T.A.; Gawron, V.J.  The effects of pavement edgelines on performance in a driving simulator under sober and alcohol-dosed conditions. Human Factors 28:511-525; 1986.

Rompe, K.; Schindler, A; Wallrich. M.  Advantages of an anti-wheel lock system (ABS) for the average driver in difficult driving situations. Proceedings of the Eleventh International Technical Conference on Experimental Safety Vehicles. Washington, DC: National Highway Traffic Safety Administration, report DOT HS 807 223, p. 442-448; November 1988.,

Rothengatter, T.  Risk and the absence of pleasure: a motivational approach to modelling road user behaviour. Ergonomics 31:599-607; 1988.

Ross, H.L.  Deterring the drinking driver. Lexington, MA: Lexington Books; 1984a.

Ross, H.L.  Social control through deterrence: drinking-and-driving laws. Annual Review of Sociology 10:21-35; 1984b.

Rumar, K.; Berggrund, U.; Jernberg, P.; Ytterbom, U.  Driver reaction  to a technical safety measure -- studded tires. Human Factors 18:443-454; 1976.

Rumar, K.  Collective risk but individual safety. Ergonomics 31:507-518; 1988.

Shannon, H.S.  Road accident data: interpreting the British experience with particular reference to the risk homeostasis theory. Ergonomics 29:1005-1015; 1986.

Shaw, L; Sichel, H.S.  Accident proneness -- research in the occurrence, causation and prevention of road accidents. Oxford, UK: Pergamon Press; 1971.

Short, M.S.; Woelfl, G.A.; Chang, C.J.  Effects of traffic signal installation on accidents. Accident Prevention and Analysis 14:135-145; 1982.

Slovic, P.; Fischhoff, B.  Targeting risk. Risk Analysis 2:227-234; 1982

Slovic, P.; Fischhoff, B.; Lichtenstein, S.  Facts and fears: understanding perceived risk. In: Schwing, R.C.; Albers, W.A., editors. Societal risk assessment -- how safe is safe enough? New York, NY: Plenum Press, p. 181-214; 1980

Smeed, R.  Some statistical aspects of road safety research. Journal of the Royal Statistical Society, Series A 112:1-34; 1949.

Summala, H.  Modeling driver behavior:  a pessimistic prediction?  In: Evans, L.; Schwing, R.C., editors. Human behavior and traffic safety. New York, NY: Plenum Press; 43-61; 1985.

Summala, H.  Zero-risk theory of driver behaviour. Ergonomics 31:491-506; 1988.

Taylor, D.H.  Driver's galvanic skin response and the risk of accident. Ergonomics 7:439-451; 1964.

Tillmann, W.A.; Hobbs, G.E.  The accident-prone automobile driver. American Journal of Psychiatry 106:321-331; 1949.

Wilde, G.J.S.  The theory of risk-homeostasis: implications for safety and health. Risk Analysis 2:209-255; 1982.

Wilde, G.J.S.  Notes on the interpretation of traffic accident data and of risk homeostasis theory: a reply to L. Evans. Risk Analysis 6:95-101; 1986.

Williams, A.F.; O'Neill, B.  On-the-road driving records of licensed race drivers. Accident Analysis and Prevention 6: 263-270; 1974.

Wall Street Journal.  Fruit frights. Page A12, 17 March 1989.

Wasielewski, P.F.; Evans, L.  Do drivers of small cars take less risk in everyday driving? Risk Analysis 5:25-32; 1985.

Wasielewski, P.  Speed as a measure of driver risk: observed speeds versus driver and vehicle characteristics. Accident Analysis and Prevention 16:89-103; 1984.

Woodsworth, R.W.; Schlosberg, H.  Experimental Psychology. New York, NY: Holt, Rienhart and Winston; 1954.

Zeckhauser, R.J.; Viscusi, W.K.  Risk within reason. Science 24:559-564; 1990.