The Need for Evaluation

Any attempt to ameliorate drinking-driving problems will consume scarce government resources and may impose monetary and nonmonetary costs (such as some restriction of civil liberties) on individuals. We must therefore determine the costs generated by the problem in order to compare them with the costs imposed by possible solutions. This section examines the magnitude and distribution of costs resulting from drunken driving. It also provides information to help determine the priority of reducing this problem by governmental efforts.

This question of costs is of more than academic interest. The comptroller general of the U.S. (1979, p. 3) estimates that “Federal, State, and local governments spent over $100 million in 1976 for their drinking-driver countermeasure activities.” This level of resources, and, as we discuss below, the large human and economic costs potentially at stake suggest the importance of determining the appropriate role of government in efforts to reduce the costs of drunken driving.

Preventable Accidents

Information presented to policy makers about drunken driving (U.S. Department of Transportation 1968, Noble 1978a, Comptroller General of the U.S. 1979) has typically expressed the importance of the problem in terms of the costs associated with it. For example, Noble (1978a, p. 61) reports that “approximately one-third of the . . . injuries and one-half of the fatalities [from traffic accidents] are alcohol related.” The term “alcohol-related” refers to any accident in which a driver, or sometimes any person involved in an accident, had a positive blood alcohol content (BAC).¹

Clearly, a better indication of the importance of solving a problem than calculating the associated costs would be to calculate the costs that would be eliminated if it were solved. To determine the maximum preventable costs of drunken driving, I first examine the BAC levels of drivers involved in various types of accidents and those of control groups of drivers selected at random from times and places similar to those at which the accidents being controlled for occurred. Table 1 shows these data, which I have selected from several studies (also see Appendix A). As Table 1 shows, the BAC levels of drivers involved in more serious accidents are generally higher than those of drivers involved in less serious accidents.

Figure

Once the distribution of BAC levels among accident-involved drivers and the control groups is known, it is possible to compute how many accidents would be avoided if all driving was done at the risk level associated with the lowest BAC level; that is, accidents that would be prevented by a “perfect” drinking-driving countermeasure. Appendix C demonstrates these computations, and the results are shown in Table 2.

TABLE 2

Expected Reduction in Motor Vehicle Traffic Accidents If All Drivers Had a Zero BAC.

Since the figures in Table 2 are based on the association between BAC and accident involvement, we must ask whether any other factors, correlated with both BAC while driving and accident risk, are confounding our analysis. There appear to be two confounding factors biasing the results in opposite directions. First, there is evidence that greater frequency of drinking is positively associated with more frequent drunken driving and negatively associated with accident risk at any given BAC (Borkenstein et al. 1974, Hurst 1973). As Appendix C shows, the risk of accident as computed from the data on Grand Rapids and Vermont is actually lower for the 0.10–0.049 BAC range than for the <0.01 range. The correlations with drinking frequency bias Table 2 toward underestimating the accident reduction, since those who currently drive drunk, if sober, would have a lower accident risk than those who currently drive sober.

The second bias results from a disputed but probable positive association between “problem drinking” or “alcoholism” and both drunken driving and accident risk while sober (Smart 1969, Noble 1978b, pp. 238–240). This bias would result in drunken drivers, as a group, having higher accident risk than the general driving population even if they were prevented from driving with positive BAC levels.

Available information is not sufficient to quantify either of these biases, therefore I present the expected reductions in Table 2 as they are and hope that the biases are small or cancel each other out.

A final caution is that my figures for maximum achievable accident reduction are based on samples at specific times and places. Their validity for the nation as a whole is not ensured. It is heartening that the Huntsville and Grand Rapids studies of injury accidents, despite their spatial and temporal separation, yield reduction estimates within 8 percentage points of each other. The Vermont data on distribution across BAC ranges of drivers killed is in agreement with several other studies of driver fatalities (Jones and Joscelyn 1978, p. 12, Noble 1978b, pp. 233–240).² I have encountered no studies of property-damage-only crashes other than the Grand Rapids study.

Magnitude of Preventable Costs

The figures on accident reduction can be roughly converted to savings in terms of lives, injuries, and dollars of property damage. In 1977, motor vehicle traffic accidents resulted in 49,500 deaths, 1.9 million disabling injuries, and $15.5 billion of property damage in the United States (National Safety Council 1976). (The $15.5 billion figure is based on $520 per accident-involved vehicle [Jones and Joscelyn 1978].) Multiplying these figures by the percentages of accident reductions in Table 2 yields maximum achievable savings in 1977 of 11,700 deaths, 156,000 to 300,000 disabling injuries, and $963 million in property damage (based on the total reduction in accident-involved vehicles).

Of course, like any estimates based on sketchy information, these figures must be interpreted with some care. The figures in Table 2 refer to reductions in the number of accidents, not directly to reductions in the consequences of accidents. For instance, alcohol-related accidents more frequently involve only a single car and driver than do accidents in general (Borkenstein et al. 1974, p. 105, Noble 1978b, p. 235). Therefore, the reduction in the number of people killed, the number of people injured, and the number of cars damaged will be less than the reduction in number of accidents. It is also likely, however, that the average severity of injury and damage in the accidents prevented will be greater than the severity of injury and damage in accidents in general; therefore only in the case of fatalities do the above figures clearly overstate potential savings from drinking-driving countermeasures.

It is beyond the scope of this paper to develop estimates of preventable costs for all other problems that may compete with drinking-driving for government resources. Table 3, however, provides some perspective. For example, we see from Table 3 that 22 percent of all accidental deaths resulted from alcohol-related traffic accidents and that 10 percent of accidental deaths would have been prevented if all drivers had zero BAC.³

TABLE 3

Deaths from Alcohol-Related Motor Vehicle Traffic Accidents, and Deaths Prevented If All Drivers Had Zero BAC as Percentages of Various Categories of Deaths, 1977.

In considering the data in Table 3, one should be aware that deaths caused by traffic accidents, particularly alcohol-related traffic accidents, occur typically among a younger age-group than deaths from other major causes. For example, Perrine et al. (1971, p. 46) report that of drivers killed in accidents, 47 percent were under 30 years of age, and of those with positive BAC levels, 49 percent were under 30. Many think that deaths of younger persons should be counted more heavily in making policy decisions, because each such death results in more years of life lost.

Another problem of weighting involves deaths of drinking drivers themselves versus deaths of innocent victims of drinking-driving accidents. I estimate that 61 percent to 78 percent of persons killed in drinking-driving traffic accidents are drivers with positive BAC levels (see Appendix B). One may argue that since people can choose how much to drink before driving, it may not be the proper role of government to protect them from the consequences of a decision freely made.

Alternatively, one might argue that people may drive after excessive drinking due to a momentary lapse of judgment, an episode of severe stress, or a habitual behavior pattern that is difficult to control (i.e., an “addiction”). One study reports that 75 percent of drivers admit to driving after drinking at least occasionally (U.S. Department of Transportation 1968, p. 61).⁴ Viewed in this way, drunken driving looks less like a risk some choose to take and more like a random risk to which many people will be exposed whether or not they would choose to in a moment of calm reflection. Furthermore, the death or injury of someone who is driving drunk may impose costs on friends, loved ones, those dependent for financial support, and the economy in general.

TABLE C-1

Results of a Controlled Epidemiological Study.

In setting policy, the proper weight to give to deaths of drunk drivers cannot be determined by empirical or logical analysis, although these may inform the decision. I feel that such deaths should be weighted equally to others for purposes of setting policy, but that we should be wary of providing perverse incentives to potential drunk drivers—a topic I address later in this paper.

In conclusion, it seems that drinking-driving countermeasures can be legitimate and useful government actions, but that even if such countermeasures were perfectly successful, the savings in lives, injuries, and property loss would be less than widely quoted figures would lead one to believe. Which countermeasures show the best prospect of success is the topic of the rest of this paper.

Overview

The strategy of drinking-driving countermeasure that occurs first to most people is exposure reduction: reducing the amount of drunken driving that takes place and thereby reducing accident costs. This section examines several potential ways of achieving exposure reduction:

• General deterrence: countermeasures that seek to prevent drivers in general from combining driving with drinking in excess of legally prescribed limits (0.10 percent BAC in most states).
• Recidivism reduction (specific deterrence): countermeasures that seek to specifically compel those people who have already been arrested for driving while intoxicated (DWI) not to drive drunk again.
• Third-party intervention: countermeasures that seek to influence those around potential drunk drivers (servers of alcohol, fellow party guests, or bar patrons, etc.) to prevent them from driving while intoxicated.
• Altering the legal minimum drinking age.
• Screening the driving population for those most likely to drive drunk.
• Installing devices in vehicles to automatically detect drunk drivers.
• Providing alternative transportation for potential drunk drivers.

General Deterrence

Reduction of Recidivism

Maximum Potential Savings

The potential reduction in traffic accidents obtainable by reducing DWI recidivism is sharply limited by the small number of persons with previous DWI arrests among drivers involved in accidents while driving with a positive BAC level. Sterling-Smith (1976, pp. 111, 135) reports that of drivers responsible for fatal accidents in the Boston area during the early 1970s, 4 percent had a previous DWI arrest and 39 percent had a BAC level exceeding 0.05 percent at the time of the accident. Thus, even if all drivers responsible for fatal accidents with a previous DWI arrest had a BAC exceeding 0.05 percent at the time of their accident, only 10 percent (i.e., 4 percent/39 percent) of drunk drivers responsible for fatal crashes would have had a previous DWI arrest.

Hurst (1973) has suggested that more experienced drinking drivers, when driving with a positive BAC level, may drive in a slow, cautious, but erratic manner that reduces their chances of serious accident compared with other drivers at the same BAC level. This manner of driving would probably increase the chance of less serious accidents as well as arrest. Since drunk drivers with previous DWI arrests are more likely to be experienced drunk drivers than those without, we would expect drunk drivers with previous DWI arrests to be more heavily represented among drivers in less serious accidents than in fatal accidents.

This expectation is borne out in data reported by Waller (U.S. Department of Transportation 1968, p. 67), that of male drivers in crashes involving alcohol or hit-and-run, 20 percent had one or more previous arrest for DWI. Based on the Sterling-Smith and Waller findings, I estimate that drivers with previous DWI arrests comprise 10 percent of all alcohol-influenced drivers in fatal accidents, 15 percent in injury accidents, and 20 percent in property-damage-only accidents.

Multiplying these estimates by my figures from Table 2 for the percentage of accidents that would be prevented if all drunk driving was eliminated, and the resulting cost savings in 1977, we arrive at estimates of the savings if all drunk driving by persons with previous DWI arrests was eliminated—that is, the savings from a “perfect” countermeasure reducing recidivism. The results are shown in Table 4. Of course, if the risk of arrest for drunk driving increased, so would the percentage of accident-involved drunk drivers with previous DWI arrests. Thus, increased risk of arrest would raise the potential savings from reducing DWI recidivism.

TABLE 4

Expected Reduction in Motor Vehicle Traffic Accidents and Costs as a Result of Preventing All Drunk Driving by Persons with Previous Arrests for Driving While Intoxicated.

Third-Party Intervention

The potential of this approach obviously hinges on what fraction of drunk drivers involved in accidents were drinking and driving in the presence of others before their accidents. Unfortunately, this information does not seem to be available, although data concerning the predriving drinking venue of DWI arrestees may be available.⁶ It seems reasonable to assume, however, that a large fraction of drunk drivers, perhaps a majority, were drinking in the presence of other persons before driving.

Servers and fellow guests or patrons can take various steps to reduce drunk driving, such as:

• Make it less convenient or less socially acceptable for a guest or patron to drink to intoxication.
• Suggest that intoxicated guests or patrons wait to sober up before driving or have a friend or taxi take them home.
• Physically restrain or report to police an intoxicated guest or patron who insists on driving.

All of these steps impose some costs on the third party: reduction in profit or apparent hospitality, the expense or inconvenience of arranging alternative transportation or lodging for the inebriate, or the unpleasantness of telling a person that he or she is for the moment incompetent to drive. The problem, then, is how to convince third parties to bear these costs.

As mentioned above, public information and education campaigns have been used to try to increase third-party intervention. These campaigns face the same difficulties as those attempting general deterrence; no truthful information that could be provided is likely to have much impact on a third party's perception of the risk inherent in drunk driving by others, since present perceptions appear to be fairly accurate. Moreover, a media campaign may not have sufficient persuasive force to alter social behavior that is reinforced by groups important to the individual. As for its use in general deterrence, however, the use of public information and education for third-party intervention is unproven rather than discredited.

The other way to convince third parties to intervene in potential drunk-driving situations is to impose legal liability on them. Such liability may be imposed by statute (referred to as a dramshop law) or by court interpretation of common law. Mosher (1979, pp. 6–8) describes the present situation:

Today there are eighteen state dram shop acts. A typical statute provides that a commercial seller of alcoholic beverages will be found liable for injuries caused by his or her patrons if the server sold or gave alcoholic beverages to the patron in violation of the law. A violation occurs if the patron is a minor, habitual drunkard, or someone “already” or “obviously” intoxicated when served…. Because most of these laws were passed before Prohibition, many are now outdated and not relevant to current social problems…. The habitual drunkard provisions are now largely ignored because of the difficulty and potential constitutional problems involved in identifying those who can be considered “habitual drunkards.” … Courts in ten states without dram shop acts, five states with such acts, and the District of Columbia have imposed civil liability on commercial servers of alcoholic beverages by court decision … based on the server's alleged negligence.

There is variation among states as to whether the drunk driver can sue a third party who contributed to his intoxication (Mosher 1979, p. 7). In general, liability is imposed on “social hosts,” those not in the business of serving alcohol, only when an underaged person is served, not when an “obviously intoxicated” person is served. Court decisions in Iowa and California that social hosts were liable if they served “obviously intoxicated” guests were quickly followed by legislation in each state overturning the decision (Mosher 1979, pp. 12–14). In another case, a California court decided that a person who merely encourages another to drink without furnishing alcohol is not liable (p. 19).

One would hope that third-party liability would cause those liable, primarily commercial servers, to take steps to prevent drunk driving by patrons. Unfortunately, the criteria by which liability is judged do not encourage servers to take precautions. A server is considered liable if he or she served an underaged or “obviously intoxicated” person, and if that person subsequently did damage. To take reasonable precautions to avoid serving such persons or to prevent those served from doing damage is not considered a valid defense (Mosher 1979, pp. 22–23). If a single patron or guest leaves and does damage and if a court decides post hoc that he or she was “obviously intoxicated” when served, then liability is determined. Since there is no standard of practice that, if adhered to, will absolve the server of responsibility, servers tend to view dramshop liability as a random risk, and insure against it rather than altering their serving practices (Mosher 1979, pp. 23–25).

If there were accepted standards of practice for servers of alcohol, and if following these practices absolved the server of liability even if a patron “slipped through,” drove drunk, and had an accident, then presumably servers would follow these practices to protect themselves and avoid high insurance costs. (Insurance premiums increased 500 percent to 1,000 percent for commercial servers in California in the wake of a court decision imposing liability on them. Insurance rates for relatively small chains of on-premise sales reached $100,000 per year [Mosher 1979, p. 24].)

In addition to liability for the actions of intoxicated patrons, the mere serving of an intoxicated person by a commercial server is in violation of the alcoholic beverage control (ABC) laws of most states. Unfortunately, state ABC departments have traditionally concentrated much more effort on ensuring the operation of an orderly industry with a low level of competition than on preventing alcohol-related problems such as drunk driving (Matlins 1976).

An exception is the “DUI (Driving Under the Influence) Project” of the California Department of Alcoholic Beverage Control (Mosher and Wallack 1979). In this experimental program drivers arrested for drunk driving were asked where they had been drinking. On-premise retailers who were twice identified in this way were offered free training for their personnel by the Department of ABC on methods of detecting intoxicated patrons and curtailing their consumption. It seems that the concern over dramshop liability was largely responsible for the good cooperation the Department of ABC received from retailers during this project.

There was no evaluation of the project's impact on drunken driving or accidents. Nonetheless, it is significant as an example of involvement of a state ABC department in drunk-driving countermeasures, and as an instance of a government agency developing and recommending standards of practice for servers of alcohol to avoid intoxication and subsequent drunk driving by patrons.

Minimum Drinking Age

If people are prevented from drinking, then they are also prevented from driving drunk. A return to prohibition would be politically unfeasible even if it were desirable, but persons under a given age are routinely prohibited from purchasing or consuming alcohol. Throughout the United States, the minimum drinking age is set within the range from 18 to 21 years. It is probably unrealistic to consider setting a minimum drinking age outside this range, but the question remains of what value within the range is optimal.

It is clear from several studies that when the drinking age is lowered from 21 or 20 to 18, the number of accidents involving 18-, 19-, and 20-year-old drivers increases (Organisation for Economic Co-operation and Development 1978, pp. 96–98, Comptroller General of the U.S. 1979, pp. 43–45, Haddon 1979, pp. 56–57, Scotch 1979, pp. 2–4). Various studies have found the percentage increase to range from undetectable to 26 percent.

The fact that prohibiting 18- to 20-year-olds from drinking reduces their accident involvement does not in itself make a convincing argument for setting the drinking age at 21. After all, prohibiting persons of any age-group from drinking would probably reduce their accident involvement. On what basis can we decide that persons who are old enough to drive, vote, and enter into contracts may not have the same access to alcohol as all other adults?

A possible justification is that accident risk rises more quickly with BAC level for drivers under 21 than for any other age-group (Jones and Joscelyn 1978, pp. 31–33). It is unclear, however, to what degree this is merely because younger drivers tend to be less experienced at drinking and drinking and driving. Those who drink more often are at less risk at any given BAC level. To the degree that the association between age and sensitivity to BAC is due to lack of drinking experience, raising the drinking age will merely transfer high accident risk from drivers under 20 to those recently turned 21.

Screening

A strategy that has received little attention is screening the population of drivers for those most likely to drive drunk and targeting countermeasures to them as individuals. An example of this strategy is a program that had been planned by federal and District of Columbia officials in which every driver applying for a renewal of his or her driver's license (and, presumably, an initial license) would be given a written test to determine whether she or he is, or is likely to become, a problem drinker. Those drivers so identified would be required to attend a program on alcohol education before they would be issued unrestricted drivers' licenses. One consultant involved in the project compared the measure to the current requirement that diabetics, epileptics, and other persons whose driving could be impaired by recurrent illness must certify that they are under proper medical care and control before being issued licenses.

As a pilot study for the contemplated project, drivers renewing their licenses in the District of Columbia during 1976 were asked to voluntarily, and anonymously, take part of a widely used test for current or prospective alcohol abuse. The test took less than 15 minutes to complete and consisted of a series of yes/no and true/false questions. Many considered the questions, some of which dealt with the driver's income, relationship with spouse, and arrest and drinking-driving history, to be too personal and inappropriate for a motor vehicle licensing agency to ask.

Despite the fact that the test had been given on a voluntary and anonymous basis, press coverage, citizens' complaints, and protests by the American Civil Liberties Union led the mayor to suspend the project for further study and eventually to order the program aborted and all collected data destroyed. (The above information was culled from the Washington Post for August 5, 1976, August 7, 1976, August 31, 1976, December 22, 1976.)

This experience does not bode well for countermeasures involving mass screenings. There seem to be two basic problems.

• The screening device. The test or procedure used for screening must use only information that is considered proper for licensing authorities to examine. Information in this class might include convictions for drunken driving and other alcohol-related offenses, and perhaps whether the driver has ever been treated for alcohol abuse. The screening device must produce a low level of false positive errors in order not to inconvenience or stigmatize persons without drinking problems.
• The treatment of persons identified by the screening. Treatment is largely dependent on what the screening device tests for. If potential drunk drivers are identified, then some sort of alcohol education or other preventive measure may work. If actual drinking drivers are identified, then the problem is the same as that of reducing recidivism (specific deterrence).

A program in North Carolina screened driver license records to identify “habitual offenders” based on past traffic convictions (Li and Waller 1976). Many of the “habitual offenders” were drinking drivers. Despite a law authorizing 5-year license revocations for habitual offenders, 62 percent of the referrals from the screening program were never acted on by prosecuters. The screening program was widely viewed as superfluous, because there was opportunity to impose similar sanctions on “habitual offenders” when they were brought to court for a particular offense. This may be a generic weakness of screening programs based on public record information.

Detection Devices in Vehicles

The suggestion has been made that cards be equipped with devices that will detect an intoxicated driver and either prevent the car from starting or make it very conspicuous on the road, for example, by automatic flashing headlights. Such a device could be installed in all cars or in only those driven by persons who seem likely to drive after drinking (e.g., persons with previous DWI convictions).

Of course, any such device could be bypassed: a driver alone or with the help of a mechanic could remove the device, render it inoperative, or even install a switch to reactivate it if the car was subject to periodic inspection.

Of the two types of detectors that have been developed, either are to some degree capable of being circumvented by having a sober companion take the “test” for the drinking driver. One type of detector is a breath alcohol analyzer; the other is a skill tester, such as one requiring the driver to punch into a keyboard a series of random numbers that are briefly displayed on a screen.

Although it is clear that any of the detection devices so far suggested can be defeated, they may still be of use since they require the driver to admit to himself or herself and to anyone else whose aid has been enlisted, that he or she is too drunk to drive. It is not known how much exposure would be reduced if potential drunk drivers and those around them were given unambiguous and immediate evidence that they were incapacitated.

More advanced detectors than are currently available, such as those that would continuously “sniff” the air around the driver's head for alcohol, or that would continuously monitor the driver's behavior for such signs of intoxication as oversteering, would eliminate the problem of someone's starting a drunk's car for him or her, although they would still be susceptible to bypass.

The widespread installation of detection devices in vehicles may meet hostile public reaction, since even those who never wish to drive drunk are likely to oppose the inconvenience as well as the added expense. The inconvenience and expense would be more easily justified if detectors were installed only in the cars of persons with previous DWI convictions. One would expect such persons to have a greater likelihood of driving drunk in the future than do drivers in general and, indeed, empirical evidence suggests that this expectation is correct (Jones and Joscelyn 1978, p. 37).

Alternative Transportation

To have an intoxicated person ride public transportation is safer for himself or herself and, of course, for others. It therefore seems promising to provide public transportation as an alternative to drinking and driving at times and places with a high concentration of drinking. For example, it is reported that the alcoholic beverage industry supplies free mass transit in Toronto on New Year's Eve. A large part of Washington, D.C., is closed to traffic on the night of the Fourth of July, and expanded bus and subway service is provided free of charge (although this is probably targeted as much at relieving traffic congestion as preventing drunk driving).

There do not seem to be any evaluations of alternative transportation programs for drunken-driving countermeasures, therefore little can be said regarding this strategy's effectiveness and efficiency.

Findings and Recommendations

While our knowledge regarding the ways in which the amount of drunk driving can be reduced is far from complete, it does lead to some conclusions that may be useful to the policy maker.

Perverse Incentives

Reducing risk refers to lowering the expected cost, in term of deaths, injuries, and property damage, of each unit of drunk driving. A possible objection to such a strategy is that as drunk driving becomes safer, people will do more of it.

The question involves estimating the “risk elasticity of drunken driving.” If a 1-percent decrease in the risk of drunk driving led to a 1-percent increase in the amount of drunk driving (i.e., if elasticity equals one), then our efforts to reduce risk would seem vain. The total costs resulting from drunk driving would remain the same. Conceivably, the elasticity may be greater than one, which is to say that a decrease in risk would lead to a greater increase in exposure, so that total costs from drunk driving would increase.

I would like to suggest that the elasticity is less than one; that when the adverse consequences of an act are both remote in probability and so serious that they are painful to contemplate, as is the possibility of a serious accident resulting from drunk driving, that a person will tend to evaluate the risk at less than its expected cost and will be insensitive to small changes in the expected cost. This would explain why drivers seemed no more effectively deterred from drunk driving by the threat of a jail sentence than of a license suspension in the Chicago experiment (see above). The probability of being subjected to either penalty was quite small, and both penalties were so serious that one would be tempted to think “that couldn't happen to me.”

If this speculation is accurate, then changes in the risk of drunken driving brought about by risk-reducing measures would not have a large impact on the amount of drunk driving and would result in a reduction of total costs resulting from drunk driving (net of the cost of bringing about the risk reduction). This speculation could be tested experimentally if the amount of drunk driving in an area was measured before and after a quick and significant reduction in the risk of drunken driving.

Generally Applied Risk Reduction

Some risk reduction measures are applied to drivers in general. They may be differentially more (or less) effective in lowering the risk of drivers with elevated BAC levels, but implementing the measures does not require knowing which drivers are likely to be impaired.

Passive restraint systems, for instance (such as air bags or automatic seat belts), would protect vehicle occupants regardless of alcohol involvement in a crash, but they would be differentially effective in protecting drunk drivers involved in accidents since they are less likely to use conventional seat belts than are drivers involved in accidents in general (Sterling-Smith 1976, p. 160). The same is true of other attempts to make vehicles in general more crashworthy. Other changes in the driving environment would reduce the probability of accident or the probable severity of accidents that do occur for all drivers while having a differentially greater effect on drinking drivers. For example, the ability to divide attention between tasks has been found to be one of the driving-related skills degraded first and most severely as BAC increases (U.S. Department of Transportation 1968, pp. 42–52). Therefore, speed governors in cars, redesigned road markings, and other changes in the driving environment that reduce the driver's need to frequently shift attention would probably result in greater risk reduction among drinking drivers than drivers in general.

Some generally applied risk reduction measures would benefit only those persons with elevated BAC levels. Haddon and Baker (1978, p. 16) point out that “injured people are rarely tested for alcohol in emergency rooms although symptoms, diagnoses, response to drugs and anesthesia, and even prognosis may be influenced by alcohol.” Informal consultation with physicians experienced in emergency room situations suggests that physicians do not routinely test accident victims for alcohol, relying on cues such as breath odor, drunken behavior, and unaccounted-for unconsciousness to warn them of possible alcohol involvement. William Haddon, M.D., of the Insurance Institute for Highway Safety, has suggested that medical personnel may be reluctant to test for alcohol in part because doing so might tend to involve them in litigation concerning the accident in which the patient was injured.

It is certainly not clear prima facie whether the incremental improvement in medical care resulting from more frequent tests for alcohol would be worth the expense of such testing. It might be found that testing some subgroups of accident victims, such as those injured in nighttime traffic accidents, is cost-effective.

All measures to reduce the risks of drunken driving (as opposed to amount of drunken driving)—even measures that offer no protection to persons with zero BAC levels—should be evaluated and assigned priority in relation to all proposed measures to reduce the risk of driving in general. In a world of limited resources there is no defense for spending a dollar to reduce alcohol-related accidents if the same expenditure would be more effective if applied to reducing motor vehicle traffic accidents in general. Similarly, if the lives of drinking drivers should be weighted no less heavily than those of innocent victims when selecting countermeasures, then a marginal dollar should be devoted to alcohol-specific risk reduction if this produces the greatest expected savings in accident costs (lives, injuries, and property damage).

It is important, therefore, that government agencies choosing among motor vehicle traffic safety measures evaluate their benefits in a way that recognizes drunk drivers are a minority on the road and that they have differential sensitivity to some safety measures. This will promote the efficient mix of alcohol-specific and nonalcohol-specific risk reduction.

Specifically Applied Risk Reduction

Some measures to reduce the risk of drunken driving may be uneconomical or impractical to apply to drivers in general, but cost-effective and practical when applied specifically to drivers with a higher than normal probability of driving with elevated BAC levels.

Some modifications to improve a vehicle's crash worthiness or ease of driving, as mentioned in the previous section, may be cost-effective only when applied to the vehicles of persons with previous drunk-driving arrests, persons requesting such safety devices for their cars, or persons willing to buy such protection on the free market.

In the case of specifically applied risk reduction the problem of political acceptability may be greatest. For instance, it is known that drowsiness, one of the obvious effects of drinking, impairs driving ability, yet public information and education campaigns from government and private sources consistently omit such suggestions as taking caffeine, driving with the window open, or playing the radio when driving after drinking. (Although it is frequently and accurately pointed out that coffee does not reverse the intoxicating effects of alcohol.) Presumably, such suggestions are omitted because they could be perceived as encouraging drunk driving by lowering its expected cost. There does not appear to be empirical evidence as to whether a driver with an elevated BAC level has less risk of accident with or without antidrowsiness measures, and the answer is not clear a priori (e.g., is a more awake drunk driver also more reckless?). Such questions are not even asked when the problem of perverse incentives is viewed as a moral issue rather than an issue of effectiveness.

Findings and Recommendations

Reducing the risk associated with drunken driving would tend to increase the amount of drunken driving that is done. I speculate that the “risk elasticity of drunken driving” is less than one, meaning that a decrease in risk is likely to bring about a smaller increase in exposure, resulting in a net reduction in accidents.

Risk-reducing measures that affect drinking drivers by the same mechanism that they affect drivers in general are best considered in the context of general traffic safety. Those concerned with alcohol-related problems should check that the procedures used by traffic safety authorities to select countermeasures accurately take into account differential impacts on drinking drivers. (For example, drinking drivers use seat belts less often than drivers in general.)

To the extent that general traffic safety measures are successful, they will reduce the accident reduction attributable to drinking-driving countermeasures. The extent of the drinking-driver problem should be periodically reviewed to determine if extensive countermeasures are still warranted.

Research should be performed to determine whether it would be cost-effective for emergency medical personnel to test traffic accident victims for alcohol.

Alcohol affects particular skills important to driving. It may be possible to redesign portions of the driving environment to decrease the impact on accident risk of degradation of these skills.

Cars driven by persons likely to drive drunk, such as those with previous DWI convictions, could be held to higher standards of crash worthiness than cars in general. This might, however, tend to make drivers of such cars overconfident and be counterproductive, and public reaction may be hostile.

Institutional Bias Against Learning

Cameron (1978, p. 11) has observed that:

Only a small proportion of drinking-driving programs in the U.S. have ever been subjected to a scientific evaluation of their effectiveness in reducing alcohol-traffic problems. In fact, much of what is known about the effectiveness of some types of drinking-driving countermeasures is based primarily on data from other countries.

An example of evaluation problems is the Alcohol Safety Action Project (ASAP) program. The program consisted of 35 local ASAPs, each approximately 31/2 years in duration, funded by the U.S. Department of Transportation between 1969 and 1975 at a cost to the federal government exceeding $88 million (U.S. Department of Transportation 1979a,b). Each ASAP attempted to integrate and improve enforcement, prosecution, screening, and treatment countermeasures so as to reduce drunken driving and associated accidents in its geographic area.

The design of the ASAP program did not allow for evaluation of its effectiveness as a drinking-driving countermeasure. Zimring's critique (1978) and the Department of Transportation's reports on the Program (U.S. Department of Transportation 1979a,b) reveal several important flaws:

• Noncomparable sites. The original eight grantees included four cities, one “twin city” site, two counties, and the state of Wisconsin.
• Inadequate control sites. Control sites for comparison with the ASAP sites were not selected until the projects had ended operation. This delay restricted potential control sites to the relatively few localities that happened to have kept sufficient accident records for comparison with ASAPs.
• Proliferation instead of replication. The program expanded to 35 projects before the initial 8 had been completed and analyzed, therefore preventing evaluation problems from being detected in time to be corrected in the later projects.
• Overaggregated evaluation. Evaluation focused on the change in accident occurrence at each ASAP site during the entire operational period at that site, even though each project used several types of countermeasures and many changes occurred in each project's countermeasure programs during its operational period. At best, each evaluation could have determined the efficacy of focusing money and attention on drinking and driving; it did not have the potential to determine the efficacy of particular countermeasures.

None of the above is to denigrate the skillfully prepared final reports on ASAP (U.S. Department of Transportation 1979a,b), but these were obviously post hoc efforts to scavenge findings from a program not well designed to supply them. Exceptions to the generally poor experimental design associated with ASAP include the Short Term Rehabilitation study (Nichols et al. 1978) and a well-designed experimental evaluation of a “drunk driver school” at the Nassau County, New York, ASAP (Preusser et al. 1976).

The lack of adequate evaluation is not confined to the efforts of the U.S. Department of Transportation. During the early 1970s the National Institute on Alcohol Abuse and Alcoholism (NIAAA) funded 18 problem drinking driver programs (PDDPs) to operate in conjunction with ASAPs. Stanford Research Institute was contracted to evaluate them and released its final report in 1975 (Eagleston et al. 1975).

Although the 131-page document reflects considerable work and expense, it is of little use in determining whether PDDP, or any components of it, brought about improvement in the persons treated. The paper states (p. II-6):

It is apparent that PDDP treatment and rehabilitation does affect [sic] a positive change on client drinking patterns and behavior as measured in various ways at intake and six months after intake. This is accomplished at a relatively low cost per client—$153 on the average. As a result we recommend that these programs be continued.

Even assuming that the measures at intake and 6 months later are entirely adequate, the report's data do not adequately support its conclusion. First, without a control group, we do not know from the report whether the programs' clients would have shown just as much or more improvement in the absence of treatment. There is evidence that drinking problems are often transitory and exhibit spontaneous remission in the absence of treatment (Cahalan and Roizen 1974). Second, the report does not account for or even mention distortion of the data through selection by potential clients and by program personnel of program entrants. Only 77.2 percent of clients making “initial contact” with a PDDP actually entered treatment (Eagleston et al. 1975, pp. V-5, V-6). Even if there had been a control group, superior progress among those in treatment might be a result of selection rather than of treatment.

What can account for the persistent pattern of neglect of the most basic precepts of experimental design when planning and evaluating government drinking-driving programs? The answer, I believe, is that the persons and organizations with influence over these programs view them not as research, but as service provision. The programs are treated as if they were applying well-developed technology to solving a problem (I use the term “service” here in a broad sense to include arresting drunk drivers, repairing roads, etc).

From the perspective of service provision, the program traits we have noted are predictable and expected and in many cases represent good management. If the purpose of a program is to provide service, then why spend money establishing control groups or collecting baseline data? Why wait until one project is evaluated before starting the next? These efforts would reduce the amount of service one could provide. Of course, even a service provider sees the need for project evaluation, but the evaluation required for such purposes is a management and performance audit rather than a scientific investigation. If the technology has been assumed sound, then why question it?

The political pressure on government agencies is no doubt to do something about drunken driving rather than to learn something about it. There is much evidence that agencies are quite responsive to this pressure. Remember that ASAP stood for Alcohol Safety Action Project. The ASAPs, touted as “demonstration projects” to show what could be done and how to do it, resulted in a set of handbooks for state and local governments on how to run an ASAP-type program, but with no assurance that this is a good way to reduce drinking-driving accidents (Hawkins et al. 1976, U.S. Department of Transportation 1979b). The problem drinking driver program “evaluation” is probably all a service provider could want in terms of reviewing the organization, management, and operation of these programs. It has little to offer those interested in whether PDDPs and programs like them are effective because it was never meant to address that question.

Recommendation

Unfortunately, we do not know how to prevent drinking-driving accidents and their related costs in a dependable and cost-efficient manner. It is therefore counterproductive to run government programs relating to drunken driving as if the primary purpose was service provision.

Persons concerned with research and experimental design should work with those concerned with service provision and program management from the earliest planning and budgeting stages of drinking-driving countermeasure programs. Federal agencies sponsoring such programs should view the development and rigorous testing of the countermeasure being used as a goal at least coequal to the application of the countermeasure in an attempt to reduce drinking-driving accident costs. Programs should not be termed “demonstration projects” unless they are indeed demonstrating a countermeasure that the agency has good reason to believe will be effective.

Such recommendations will doubtlessly evoke protests that resources are being diverted from saving lives to performing research. But if our goal is immediate saving of lives then we are investing our resources poorly. The U.S. Department of Transportation (1979a,b) estimates that the ASAP program saved 506 lives at a cost of $156,306 each. I think a fairer estimate is that 425 lives were saved at a cost on the order of $300,00 each.⁷ Using either set of estimates, it appears that more lives could have been saved by using the ASAP money elsewhere. As the U.S. Department of Transportation (1979a,b) points out, “The National Highway Safety Needs Report … ranks 37 countermeasures [by] cost per fatality forestalled…. The ASAP cost [$156,306] would rank between number 14—motorcycle lights-on practice—and number 15—impact absorbing and roadway safety devices.”

Given our present level of technology for preventing drinking-driving accidents, additional expenditures seem warranted only if they promise to produce findings that will help us improve the technology and save more lives in the future, as well as contributing to current traffic safety.

Evanston

Richard L. Holcomb (1938) Alcohol in relation to traffic accidents. Journal of the American Medical Association 111(12):1076–1085.

Case group was some accidents resulting in hospitalization, in and near Evanston, Ill., between February 1, 1935, and February 1, 1938. Method of selection is not explained in report.

Control group was vehicles stopped at eight locations in Evanston “which approximated the area in which the accidents of the first part of the study occurred.” Samples were taken during the following periods during a week beginning Saturday, April 23, 1938:

Of the control group 1.35 percent refused the BAC test. It is not reported what fraction of the case group refused, but the report claims “rather good cooperation was obtained from the drivers examined. A representative cross section of all drivers involved in injury accidents was thus obtained” (p. 1077).

The selection of neither the case nor control group is explained in detail, and neither appears to have been selected with as much rigor as in more recent studies. Thus, the reliability of this study's results is questionable.

Grand Rapids

R. F. Borkenstein, Crowther, R. F., Shumate, R. P., Ziel, W. B., and Zylman, R. (1974) The Role of the Drinking Driver in Traffic Accidents, 2nd ed., Blutalkohol XI (Supplement 1):8–131. (1st ed. unpublished, Department of Police Administration, Indiana University, 1964.)

Case group was all accidents in Grand Rapids, Michigan, between July 1, 1962, and June 30, 1963.

Control group was four vehicles stopped at random at each of 2,000 time-place sites selected at random from sites of all reported accidents between May 1, 1959, and April 30, 1962. Control group sampled over same period as accident group.

Of control group 2.18 percent and 4.25 percent of accident group refused the BAC test (p. 39). Refusal rates were generally higher in demographic groups more prone to drunken driving (pp. 52–57). The probable result is a significant but small bias toward underestimating the relative risk of driving at higher BAC levels.

In order to promote driver cooperation, the study was heavily publicized while observations were being made. Although Borkenstein claims that this publicity was insignificant as a drinking-driving countermeasure and therefore as a source of bias (p. 52), I suspect that it caused a small decrease in the frequency of drunken driving by moderate drinkers (who would have positive BAC levels below about 0.10). I cannot predict with any confidence its impact on the “risk” associated with various BAC levels. The study also presents the distributions across BAC categories of only those accidents that resulted in injury, including the 15 accidents that resulted in death, and of only those accidents that resulted in property damage. The study does not, however, show the distribution across BAC intervals of control group observations corresponding to these two severities of accidents. Thus, the risk figures presented in Tables A-4 and A-5 are based on case group observations for the severity of accident indicated and control group observations for all accidents.

TABLE A-4

Injury Accidents: Survey Results and Relative Risk.

TABLE A-5

Damage-Only Accidents: Survey Results and Relative Risk.

I have interpolated these results into my standard BAC categories by assuming uniform density of observations within the interval 0.08–0.109. (There are insufficient data to show >0.15 as a separate interval.)

I have interpolated these results into my standard categories as above:

Huntsville

R. Farris, Malone, T. B., and Lilliefors, H. (1976) A Comparison of Alcohol Involvement in Exposed and Injured Drivers, Phases I and II. National Highway Traffic Safety Administration technical report DOT HS-801 826. Washington, D.C.: U.S. Department of Transportation.

Case group was automobile drivers in injury-causing accidents in Huntsville, Alabama, between July 1974 and July 1975, including the six fatal accidents during that period.

Control group was an unreported number of drivers (averaging 2.61) stopped at same place, time of day, day of week, and direction of travel as each of the last 314 (of a total of 615) observations in the case group. Control observations were taken within 30 days of the case observation being matched.

Two percent of control group and three percent of case group refused the BAC test.

The subgroup of case observations to which controls were matched did not show a significantly or systematically different distribution of BAC levels than the earlier, unmatched case data. I have therefore pooled all case data.

I have interpolated these results into my standard BAC categories by allowing each of the report's categories to “steal” enough observations from the next lower category to preserve its (the higher category's) original observation density as it expands to become one of my categories.

New York

James R. McCarroll and William Haddon, A controlled study of fatal automobile accidents in New York City. Journal of Chronic Diseases XV:811–826.

Case group was fatally injured drivers in New York City, exclusive of Staten Island, between June 1, 1959, and October 24, 1960.

Control group was six cars going in the same direction at same place, time of day, and day of week as each accident in case group. “All site visits were made in 1960, within a few weeks of the calender week of occurrence” (p. 812).

Of the control group 0.39 percent refused the BAC test. BAC test was performed by the municipal medical examiner on each member of case group.

I have interpolated these results into my standard BAC categories by assuming uniform density of observations within each of the study's BAC categories.

The predominant flaw of this study is the small sample size, which allows us to put little confidence in the results.

Vermont

M. W. Perrine, Waller, J. A., and Harris, L. S. (1971) Alcohol and Highway Safety: Behavioral and Medical Aspects. National Highway Traffic Safety Administration technical report DOT-HS-800-599. Washington, D.C.: U.S. Department of Transportation.

Case group was fatally injured drivers in Vermont from July 1, 1967, to April 30, 1968.

Control group was pooled data of: (1) six drivers at same place, time of day, and day of week as each observation in case group, either within a few weeks or one year after accident being matched; and (2) six drivers at same place, time of day, and day of week as each observation from a group of crashes occurring in Vermont in 1966 and resulting in an injury warranting hospitalization, each observation of which was chosen as a close match to one observation in the case group.

The report's account of control group selection is unclear and self-contradicting. The report justifies pooling the two samples by showing that there is a significant difference in the distribution within either sample of only 3 of 22 variables. I suspect that the control group shows fewer high BAC levels than were actually present among drivers at times and places similar to those of case group crashes.

Of control group 1.3 percent refused the BAC test. BAC test was performed on all members of the case group.

The lowest BAC category used in the report is <0.02. I have interpolated into my standard BAC categories by assuming that the density of observations in the interval 0.01 <BAC <0.02 is the same as the density in the interval 0.02 <BAC <0.04. All other categories in the report are comparable to mine.

The “risk” figures for the higher BAC categories are probably biased upward somewhat due to flaws in the control group.

Controlled Epidemiological Studies

Among the most valuable studies of drinking-driving problems are controlled epidemiological studies, such as those reviewed in Appendix A of this paper. In these studies, a random (or exhaustive) sample of drivers involved in accidents in a set geographic area and time period is made. The BAC of each driver in the sample is measured at the time of the accident. These drivers constitute the “accident group.”

In addition, researchers match a “control site” to each observation in the accident group. The control site is ideally the same place, at the same time of day and day of the week, under similar weather conditions, and otherwise nearly identical to the circumstances under which the accident being matched occurred. Several drivers passing each control site are randomly selected and their BAC levels measured. These observations constitute the “control group.”

Drivers in both the accident group and control group are classified into BAC ranges, and the results are presented in Table C-1.

I use the results of a controlled epidemiological study, as presented in Table C-1, to estimate the results of an imaginary experiment that is impractical to actually perform. The imaginary experiment is to randomly sample “units of accident exposure” and record driver BAC and whether an accident occurs for each unit of exposure selected. A unit of accident exposure would be a period of driving during which a “standard” driver would have some particular expected number of accidents.

For example, driving for 1 hour (or 1 mile) at night in the rain on a narrow twisting road would constitute more units of exposure to accident than would 1 hour (or 1 mile) of driving on a sunny day on a well-designed expressway.

If we could perform the imaginary experiment, then we could answer two important questions: (1) What are the probabilities of an accident occurring when one unit of accident exposure is driven by a driver at various BAC levels? (2) What would be the reduction in number of accidents if all driving was done by drivers in the lowest BAC range?

Estimating Relative Risk¹

What is the probability that a random unit of exposure results in an accident, given that the driver is in some particular BAC range? To answer this question, I invoke Bayes's theorem:

(1)

where

View in own window

P(A\i)	is the probability that a given unit of exposure results in an accident, given that the driver is in the i th BAC range.
	is the probability that a randomly chosen unit of driving results in an accident, and that the driver is in the i th BAC range.
P(i)	is the probability that a randomly selected unit of exposure is driven by a driver in the i th BAC range.

Also from Bayes's theorem:

(2)

Substituting (2) into (1):

(3)

Referring back to Table C-1, we see that P(i|A) = a _i/Σa _i, since the accident group is a random sample of units of exposure resulting in accidents. We can use c _i/Σc _i as a proxy for P(i), because the probability of a driving trip being included in the control group is primarily determined by the likelihood that a similar trip resulted in an accident. Thus, the control group approximates a random sample of units of exposure.² We can thus rewrite (3) as:

where

I make the simplifying assumptions that D approaches one and R approaches zero. We are left with:

S is a constant, so the probability that trip T is included in the control group is approximately equal to the expected value of the exposure to accident of trip T.

We can thus rewrite (3) as:

(4)

Note that controlled epidemiological studies do not provide sufficient information to determine P(A|i), the probability that a randomly selected unit of exposure will result in an accident. We cannot determine the absolute value of P(A|i). We can, however, determine its value for all values of i relative to its value for i = 0 (i.e., for the lowest BAC range):

(5)

I will refer to P(A|i)/P(A|0) as the relative risk of driving in BAC range i. Thus, the relative risk of the lowest BAC range is always one. Table C-2 shows the relative risk of driving in several BAC categories, as derived from the controlled epidemiological studies in Table 1 of this paper.

Figure

Predicting Accident Reduction

The total number of accidents of a given severity that one would expect to occur in a given area during a given period may be expressed as:

(6)

where B _i is the expected number of accidents involving drivers in BAC range i; E _i is the units of exposure to accident driven by drivers in BAC range i.

Note that as long as P (A|0) remains constant, relative risk for BAC range i is a linear function of P(A|i). Also, c _i approximates a linear function of E _i (see note 2). Thus, we can rewrite (6) as:

(7)

where ^B _i is a linear function of B _i.

I will refer to ^B _i as the relative number of accidents involving drivers in BAC range i. If all drunken driving was eliminated, without changing any individual's driving patterns (mileage, route, etc.), then we would compute a new ^B _i for each BAC range merely by setting relative risk for every range equal to one. Thus, if a “perfect” drinking-driving countermeasure was to go into effect, the total expected relative number of accidents would be:

(8)

In Table C-3, I present ^B _i for each BAC range, and Σ^B _i, for each of the studies in Table C-2, computed both with and without a “perfect” drinking-driving countermeasure in effect. The percentage change in B _i when the perfect countermeasure is figured in predicts the percentage change in the absolute expected number of accidents if such a counter-measure was implemented.

Figure