Grand Rapids Effects
Most discussions about what has become known as the "Grand Rapids Effects", or the "Grand Rapids Dip" involves the most eloquent weasel wording known to man [and jew]. To put it bluntly, and to cut through the weasel wording, numerous studies worldwide, not just one in the US, have demonstrated that, compared to sober drivers, drinking drivers with a BAC < 0.02 are 4% less likely to have an accident, with a BAC < 0.04 are 30% less likely, with a BAC < 0.06 are equally as likely
Over the last several decades, the average Motor Vehicle Fatality Rate in the US has been 15 per 100,000 population. Over 2.5 trillion miles of travel, there are an average of 42,000 auto-related fatalities, or 16.8 fatalities per 1 billion miles. Multiple studies of hand/eye coordination, airline pilot and trucker data, and the NHTSA, all confirm the same trend: that women are three to four times as likely as a man to have an accident. Since men drove 70% of those miles, or 1,750 billion, and women drove 30% or 750 billion, and being conservative and estimating that women are only three times more likely to have an accident, we can estimate the overall average probability of having an accident per bilion miles by sex:
X = fatal accidents per billion miles for men
3X = fatal accidents per billion miles for women
(1,750 x X) + (750 x 3X) = 42,000
1,750X + 2,250X = 42,000 = 4,000X
X = 10.5 accidents per billion miles for men drivers
3X = 31.5 accidents per billion miles for women drivers
If we assume that men and women drivers are equally impaired [or unimpaired] by alcohol consumption [an extremely conservative assumption], then the following is the calculated probabilty of having a fatal accident broken down by sex and BAC level:
Thus, this study proves that a sober woman driver is 51% more likely to have an accident than a man driver with a BAC of 0.14, which is approximately 5-7 drinks. A drinking man driver doesn't drive as dangerously as a sober woman driver until a BAC level of 0.14, about 8 drinks.
It should be noted that this high accident rate of women increases the accident rate of men, since at least a third of the fatal accidents they have are with men who otherwise wouldn't have been involved in an accident. Thus without women drivers, men would have accidents at the rate of 7 per billion miles. If men drove ALL 2.5 trillion miles, this would be only 17,500 fatalities annually, cutting fatalities by more than half. If men continued to drive only the 1,750 billion miles they currently drive, there would be only 12,250 traffic fatalities annually, a reduction of almost three quarters.
Even though the following graph of what the NHTSA calls "the Borkenstein Effect" is not detailed enough to realize the significance of the above data, it does fill in some gaps on where a drinking man driver falls on the above table. What all of these studies ignored is the obvious--that the vast majority of these accidents at high BAC levels must have involved women drivers, not men. And, as noted above, many of the accidents which DID involve men MUST have been caused by women drivers with a BAC level around 0.1 to 0.14 where they're 30 to 50 times as likely to have an accident than men with a BAC between 0.02 and 0.08.
Knowing this, it's now obvious that the following graph needs to be broken down by sex.
The following excerpt from the NHTSA web site, which ignores the significance of the Grand Rapids Effect, is very revealing of the advocacy nature of this study and the willingness of the authors to bend the rules to achieve their not-so subtle objectives of:
Bars and people don't measure drinks based on the body weight of the consumer or drinker. So it's VERY significant if women who have the same number of drinks as men are MORE likely to have an auto accident. This is particularly true when it's already known and proven that women drivers and truckers and pilots, when SOBER are 3-4 times per mile driven or hour flown to have an accident than men drivers and trucker and pilots.
It's also EXTREMELY important to take into account when this data shows that, in order for a man driver to drive as dangerously as a sober woman driver, he must drink more than 7 drinks.
See also Borkenstein
It has been almost 100 years since it became apparent that drivers' use of alcohol leads to an increased risk of crash (See Borkenstein, 1985). Traffic codes prohibiting alcohol-impaired driving had appeared in the United States by 1910, and the major approach to prevention then, as now, was deterrence by legal prohibition and law enforcement. By the 1940's, only three percent of traffic collisions were reported as being alcohol-related, due largely to officers' difficulties in assessing drivers. In the 1930s, epidemiological studies, which are studies examining the distribution of an event in a population, had begun to use breath and blood specimens to measure blood alcohol concentration (BAC) in crash-involved drivers. The measured BACs showed alcohol involvement in crashes to be much greater than three percent, and it was on the basis of those studies that the states began to establish BAC limits for drivers.
The first law in the United States establishing a BAC limit was enacted in 1939 in Indiana. Initially, the limits in Indiana and in other states were set at 0.15%(1), but they now have been lowered nationwide to either 0.10% or 0.08%. In other countries they are even lower. Limits defined by BAC assist with enforcement problems and also aid drivers in assessing their own impairment. There is worldwide agreement that alcohol-involved driving is curtailed when BAC laws are enacted and enforced.
The reduction of limits from the initial 0.15% BAC was prompted by evidence obtained from experimental and epidemiological alcohol research. As research continued over several decades, and as scientific investigators improved their techniques for examining relevant driving behaviors, evidence of significant driving impairment was reported at even lower BACs.
Studies have reported that the degree of impairment produced by alcohol may be modified by other variables. For example, the Grand Rapids study, which was the largest epidemiological study, suggested that the variables, age, gender, and drinking practices, produce differential impairment at similar alcohol levels (Borkenstein et al., 1964). Firm conclusions about those three variables on the basis of epidemiological data are difficult, however, because each is also associated with other variables which influence crash rates. For example, young people show a differentially high crash rate under alcohol, but they are also less experienced drivers. Also, when the Grand Rapids study was executed in 1962, women drove far less frequently and for shorter distances than men, possibly making them more susceptible to alcohol effects on driving. Analysis of the study's data relied primarily on uni-variate statistical methods, which could not isolate the effects of age, gender, and drinking practices from the effects of other variables.
The literature reporting data from laboratory research contains only equivocal evidence for an age interaction with alcohol (Jones and Neri, 1994; Morrow et al., 1990; Collins and Mertens, 1988). These studies, which included no subjects under age 21 and few subjects over age 55, do not resolve the issue, however, since it was drivers under age18 and over age 70 for whom the Grand Rapids study suggested an age and alcohol interaction. The question of whether young drivers are differentially sensitive to alcohol also remains unanswered by the current study. Because alcohol cannot be administered in the United States to anyone under age 21, the youngest subjects were ages 19 and 20. They were tested in Ontario, Canada where the alcohol age limit is 19 years.
In the Grand Rapids study, a gender and alcohol interaction did not occur until the BACs reached 0.08% and above. At those levels, women were more frequently accident-involved than men. Laboratory studies of the responses by men and women to alcohol, however, provide inconclusive results. As Sutker et al. (1983) noted, most experiments have given men and women the same alcohol dosage. Since the body fat and total body water of men and women differ greatly even when they are the same age, height and weight, women reach a higher BAC than men for the same alcohol amount. Many early studies failed to take this into account, but more recent studies have used comparable BACs rather than equivalent doses. These studies failed to find significant difference between male and female subjects (Burns and Moskowitz, 1978; Mills and Bisgrove, 1983; Oei and Kerschbaumer, 1990).
More reliable evidence exists for an interaction between alcohol and drinking practices. The Grand Rapids study reported that the likelihood of involvement in a collision for drivers at the same BAC was greatest for the drivers with the lowest daily alcohol consumption. A study by Moskowitz, Daily and Henderson (1974) supported this finding with a comparison of extremely heavy drinkers (recruited from bars) and moderate drinkers. They reported that heavy drinkers were less impaired than moderate drinkers at equal BACs on several psychomotor tasks. Also, a mean ethanol clearance rate of 0.020% per hour for the heavy drinkers, in comparison to a rate of 0.017% per hour for the moderate drinkers, demonstrated a physiological difference between the heavy and moderate drinkers.
This study examined skills performance of a representative sample of the driving population at BACs from 0.02% to 0.10%. It also examined whether variations in drivers' age, gender, or drinking practices interacted with BAC and resulted in variability in the impairment produced by alcohol. One hundred sixty- eight subjects were classified by four age groups, two genders, and three drinking practice categories. The three variables of age, gender, and drinking practice dictated the assignment of subjects to 24 groups of 7 each (Figure 1).
The youngest subjects in the study, who were ages 19 and 20, were tested at Human Factors North (HFN) in Ontario, Canada. Also, although evidence of an interaction of gender and alcohol is less substantial than the evidence of interactions of age and drinking practices and alcohol, the study included equal numbers of men and women in order to examine the issue.
There are many assumptions made in the following German police study which indicate its advocacy nature and prove that the authors had an agenda: to do what was necessary to make drinking and driving a factor in accidents. Perhaps this should be, since this is now the common perception [or misperception]. It also failed to take into account the huge differences between the way alcohol affects men and women drivers, and how much more likely a woman is to have an accident compared to a man.
Nonetheless, the data itself is important, because it proves just the opposite of what the authors conclude.
Grand Rapids Effects Revisited: Accidents, Alcohol and Risk
H.-P. Krï¿½ger, J. Kazenwadel and M. Vollrath
Center for Traffic Sciences, University of Wuerzburg, Rï¿½ntgenring 11, D-97070 Wï¿½rzburg, Germany
Risk analysis is based on information collected about both exposure to danger and the dangerous event itself. In the case of alcohol-related accident risk, information is needed about the prevalence of driving under the influence of alcohol (DUI) and the frequency with which DUI drivers are involved in accidents. These requirements were met in Borkenstein et al.'s Grand Rapids Study. However, one shortcoming of that study was the risk of causing an accident (rather than just being involved in an accident) had to be estimated because the authors did not know whether the driver was responsible for the accident. Our 1993 Accident Study collected information about BAC, responsibility for causing the accident, and driver characteristics for all drivers involved in 4,615 accidents. The information about exposure was taken from the German Roadside Survey, which sampled 13,149 drivers in 1993. Using those data, the well-known risk function of Borkenstein et al. was replicated. However, the BAC distribution for drivers involved in an accident but not responsible for it was markedly different from that for the drivers in the Roadside Survey. In calculating risk function, Borkenstein et al. assumed identical distributions for these two samples. It can be shown that the problematic "dip" in the risk function was at least in part caused by this assumption.
Risk analysis is based on information collected about both exposure to danger and the dangerous event itself. In the case of alcohol-related accident risk, information is needed about the prevalence of driving under the influence of alcohol (DUI) and the frequency with which DUI drivers are involved in accidents. Although these requirements were met in Borkenstein et al.'s Grand Rapids Study the study has one shortcoming: The risk of causing an accident (rather than just being involved in one) had to be estimated because the authors did not have information on whether the driver was responsible for the accident. In our Accident Study, which took place in Germany in 1994, we obtained this information directly from the police. All the analyses described below include only those drivers who were responsible for causing the accident. Information about DUI prevalence was obtained from the German Roadside Survey (see Krï¿½ger et al., 1995, in this volume). The risk function resulting from our 1994 study is, in general, comparable to that resulting from the Grand Rapids Study. However, important differences were found concerning the steepness of the risk functions. In addition, it is demonstrated that the global risk function has to be differentiated for subgroups of drinking drivers. The impact of measures directed against these drivers is estimated by means of the attributable risk.
The German Roadside Survey was conducted in the northern part of Bavaria (Unterfranken, part of the former West Germany), which has approximately 3 million inhabitants. Three components were done from the end of 1992 to the spring of 1994. Drivers were stopped and selected by the police who followed a random sampling plan. At a separate checkpoint, these drivers were interviewed and asked to supply a breath sample. Of those asked for a breath sample, 9128 (94.8%) agreed. The roadside survey oversampled weekends at night to obtain a large proportion of DUI drivers. For a representative picture of the DUI prevalence in Germany, the observations were adjusted using information from a representative study of driving in Germany (KONTIV; see Emnid, 1991).
The Accident Study was also conducted in Unterfranken. We equipped a selected sample of police cars with breath testing devices, under the condition that officers try to obtain breath samples from all accident drivers, whether or not they were suspected of DUI. In 1993 in Unterfranken, data were obtained from 1.968 drivers who were responsible for causing an accident.
The Roadside Survey and the Accident Study differed very much with regard to time of day (night: 20 p.m. to 4 a.m.; day: 4 a.m. to 20 p.m.) and day of week (weekend: Friday night to Monday morning; workday: Monday morning to Friday evening). These differences are reflective of such risk factors as, for example, the higher accident risk during the night. As we were mainly interested in alcohol-related accident risk, we controlled for these variables by applying a second weighting procedure to the data from the Roadside Survey. This two-dimensional weighting (by time of day and day of week) produced identical subject distributions in the two studies with respect to the combination of those two factors.
The alcohol-related accident risk is estimated by computing odds ratios. An odds ratio gives a good estimation of the relative accident risk for drivers in a certain BAC class compared to sober drivers (their risk is set to 1). A value larger than 1 indicates an increase in accident risk due to alcohol.
Risk Functions in 1964 and 1994
In 1964, Borkenstein et al. presented the well-known risk function for drivers responsible for causing an accident, which was one basic argument for setting BAC limits in different countries (for example, Germany). Figure 1 shows this risk function compared with the function computed from the Accident Study (both functions were smoothed). The shape as well as the magnitude of the functions are very similar. For drivers with blood alcohol concentrations (BAC) up to 0.04%, the alcohol-related accident risk is nearly identical to or even less than that for sober drivers. Both studies found that, for drivers at BACs ranging from 0.14% to 0.16%, the accident risk is about 25 times as high as it is for sober drivers. However, for nearly all BACs, the 1994 alcohol-related accident risk in Germany was greater than in 1964, a finding that may be a function of today's more complex traffic situations, which in combination with alcohol cannot be handled adequately anymore. At BACs greater than 0.14%, the deteriorating effects of the intoxication may be so great as to make the differences in traffic conditions irrelevant.
Analysis of factors modifying the alcohol-related accident risk showed driver age to be the strongest mediator (see Vollrath, Krï¿½ger & Kazenwadel, 1994; Krï¿½ger, Kazenwadel & Vollrath, 1995). The global accident risk for drivers between 18 and 24 years is much greater than that for older drivers. In addition, the alcohol-related accident risk for those young drivers increases much faster than it does for older drivers. In light of these findings, we strongly recommend lowering the BAC limit for younger drivers.
The Attributable Risk
Although drivers under the influence of alcohol are obviously at a greater relative risk than unintoxicated drivers, the magnitude of the risk to the larger community attributable to the presence of intoxicated drivers remains an unanswered question. In the German Roadside Survey, only 5.5% of all drivers were found to have BACs greater than 0. Thus, drivers in Germany are exposed to the increased accident risk due to DUI in only 5.5% of their trips (this statement is valid because of the representative weighting procedure described above). By combining the information about the distribution of exposure (DUI) with the estimate of alcohol-related accident risk, one can determine the degree to which accidents can be explained by DUI. This question is adressed by the measure of the attributable risk (for an overview, see Breslow & Day, 1980; Kahn & Sempos, 1989). The basic idea of attributable risk is that some of the accidents involving intoxicated drivers are not due to the effects of alcohol but are the result of the global accident risk also present for sober drivers. This means that the number of accidents involving intoxicated drivers is adjusted to allow for this global accident risk, yielding an excess number of accidents which are attributable to the effects of alcohol.
There are two definitions of attributable risk (AR), addressing two different aspects: (1) The attributable risk for exposed persons (Cole & MacMahon, 1971) renders an estimate of the proportion of all accidents with intoxicated drivers that is attributable to the effects of alcohol. (2) The attributable risk for the population (first described by Levin, 1953) renders an estimate of the proportion of all accidents (including those with sober drivers) that is due to the effects of alcohol.
To compute these ARs, we chose the BAC classes given in Table 1. The first column shows the number of drivers from the German Roadside Survey according to BAC class, and the second column the number of drivers from the accident study. The first step in computing the number of accident-involved drivers within each BAC class attributable to the effects of alcohol (excess) is to compute a factor k of accident involvement for sober drivers. This factor is calculated as:
Using this factor, the number of drivers that would be expected to be responsible for causing an accident is estimated for each BAC class. For example, for a BAC greater than 0.20%, 10 drivers were found in the Roadside Survey. Multiplying this number by k results in 10 * 0.1941 = 2. Thus, we would expect 2 drivers to be found in the Accident Study in this BAC class (not due to alcohol). However, 64 were found yielding an excess number of 62 accidents which may be attributed to the effect of alcohol. Those excess numbers are given in the third column of Table 1. Of course, there are large difference among the BAC classes. At lower BACs, we even find negative numbers indicating the "dip" in the risk function for lower BACs first described by Borkenstein et al. (1964).
For all BAC classes above 0%, we found 330 drivers in the accident study. Of those accidents, 213 were attributable to the effects of alcohol. By dividing those two numbers, we obtain an AR for exposed persons of 213/330=0.65 or 65%. That means, 65% of all accidents involving an intoxicated driver can be attributed to the effects of alcohol. However, in only 16.8% of all accidents (or 330 accidents) was the driver intoxicated. To determine which proportion of all accidents are attributable to the effects of alcohol, the population AR should be computed. This is done by dividing the excess accidents by the total number of all accidents, that is, 213/1968=0.108. Thus, 10.8% of all accidents may be attributed to the effects of alcohol.
Figure 2 gives both ARs computed for different BAC classes. The AR of exposed drivers indicates for each BAC class the percentage of accidents attributable to alcohol. For BACs less than 0.06%, the AR is small, even negative. Hardly any accidents involving drivers with those BACs can be attributed to intoxication. This changes dramatically for BACs greater than 0.06%. At BACs less than 0.08% but greater than 0.06%, about 70% of all accidents are due to alcohol. For all BAC classes greater than 0.08%, the ARs are greater than 80%. For drivers in this latter BAC categories, nearly all the accidents may be attributed to the effects of alcohol.
The AR of the population indicates the magnitude of those alcohol effects in relationship to the total number of accidents occuring. The population ARs can be interpreted as follows: If no drivers with BACs greater than 0.20% were present in traffic, 3% of all accidents would not happen. Adding these percentages for all BAC classes gives the 10.8% of all accidents which are due to alcohol. About a third of these accidents can be attributed to drivers with BACs greater than 0.2%.
As Figure 3 shows, this population AR gives a good indication of the effectiveness of measures directed against DUI. In this Figure, the 10.8% accidents were set to 100%. Had no DUI drivers been present in traffic, none of these accidents would have occured, which would have resulted in a 100% reduction. If no one with a BAC greater than 0.08% drove, a reduction of 96% would result. Thus, if the legal limit for DUI in Germany (0.08%) was an effective deterrant against driving with a higher BAC, this would mean that nearly everything that could be done to prevent alcohol-related accidents would have been accomplished. Thus, countermeasures directed at those persons driving at BACs higher than 0.08% can be expected to be most effective in reducing the number of accidents attributable to the effects of alcohol. In contrast, measures directed at drivers with BACs less than 0.08% cannot be very effective. At most, 4% of all accidents attributable to the effects of alcohol may be prevented.
Sub-Groups Included in the Risk Function
The question remains how to indentify the characteristics of the these drivers with BACs greater than 0.08%. Is driving with high BACs done very seldom by nearly all drivers or is it done quite often by a small subgroup of drivers? We can begin to answer this question be analyzing the risk functions in Figure 1 in detail. These smoothed functions give the impression that alcohol-related accident risk increases monotonically. However, if smaller BAC classes are selected and BAC is truncated at 0.18%, the picture changes. In Figure 4, odds ratios were computed for BAC classes of 0.02%. The risk functions are shown from our Accident study, from the Grand Rapids Study (drivers responsible for the accident), and from a study by Perrine, Waller & Harris (1971) of fatally injured drivers.
Although the studies were done at different times in different countries, the similarities in the structures are striking. In none of the three risk functions is there a monotone increase in risk, but different peaks are found. In our Accident study, the first (small) peak is present between 0.08% and 0.10%, a second peak at 0.14% to 0.16% and a third peak at BACs greater than 0.20%. In the Grand Rapids Study, similar peaks are found but are shifted towards higher BACs. This reflects the finding shown in Figure 1 that, in our Accident Study, the alcohol-related accident risk is higher than that found by Borkenstein et al. (1964). In contrast, in Perrine et al.'s study, the peaks are shifted towards lower BACs, which makes sense as only fatally injured drivers (very serious accidents yielding a larger alcohol-related accident risk) were examined.
The occurence of those peaks in three different studies from different countries and different years suggests that the overlay of risk functions of different sub-populations produces the typical shape of the overall risk function. Extended studies on drinking behavior indicate that three different groups of drinkers may be responsible for the peaks. These hypothesis is supported by studies on hardcore drinking drivers (for example, Simpson & Mayhew, 1993). The assumption of three sub-groups of drinkers is indicated in Figure 5 (the risk function here was computed from our data for BAC classes of 0.01%). The first group consists of moderate drinkers who will never exceed a maximum BAC of around 0.10% (consumption limit). At higher BACs, this group cannot compensate the effects of alcohol very well, which yields the first peak of the risk funktion. Two groups of heavy drinkers are responsible for the other peaks. Both groups have probably developed a certain amount of alcohol tolerance, enabling them to compensate for the effects of alcohol at higher BACs.
Our Accident Study replicated the well-known risk function of Borkenstein et al. (1964). The comparison indicates that driving under the influence of alcohol resulted in a greater accident risk in 1994 compared to 1964. Considering the incidence of DUI, it was argued that effective countermeasures that substantially reduce the number of accidents attributable to the effects of alcohol should be directed towards drivers with BACs greater than 0.08%. This also implies that simply changing the legal DUI limit from 0.08% to 0.05% is insufficient with respect to alcohol-induced accidents as the potential reduction would be only about 4%. Further inspection of the risk function indicates that certain subgroups of drinking drivers are responsible for the alcohol-related accident risk in the higher BAC range. Measures capable of deterring drinking drivers in this range were expected to have a substantial impact on traffic safety, namely, result in a decrease in accident rates.
Borkenstein, R.F., Crowther, F.R., Shumate, R.P., Ziel, W.B. & Zylman, R. (1964). The role of the drinking driver in traffic accidents. Department of Police Administration, Indiana University.
Borkenstein, R.F., Crowther, F.R., Shumate, R.P., Ziel, W.B. & Zylman, R. (1974). The role of the drinking driver in traffic accidents (The Grand Rapids Study). Blutalkohol, 11, Supplement 1.
Breslow, N.E. & Day, N.E. (1980). Statistical methods in cancer Research. Volume 1 - The analysis of case-control studies. Lyon: International Agency for Research on Cancer.
Cole, P. & MacMahon, B. (1971). Attributable risk percent in case-control studies. British Journal of of Prevention and Social Medicine, 25, 242-244.
Emnid (1991). KONTIV 89. Bericht zur Methode, Anlagenband und Tabellenteil. Bielfeld: Emnid.
Kahn, H.A. & Sempos, C.T. (1989). Statistical methods in epidemiology. New York: Oxford University Press.
Krï¿½ger, H.-P., Kazenwadel, J. & Vollrath, M. (1995). Das Unfallrisiko unter Alkohol unter besonderer Berï��������������������������������½cksichtigung risikoerhï¿½hender Faktoren. In H.-P. Krï¿½ger (Hrsg.), Das Unfallrisiko unter Alkohol. Stuttgart: Fischer Verlag (in preparation).
Levin, M.L. (1953). The occurrence of lung cancer in man. Acta Unio Internationale Contra Cancrum, 9, 531-541.
Simpson, H.M. & Mayhew, D.R. (1993). The hard core drinking driver. In H.-D. Utzelmann, G. Berghaus & G. Kroj (Eds.), Alcohol, Drugs and Traffic Safety - T92 (pp. 847 - 853). Cologne: Tï¿½V Rheinland.
Vollrath, M., Krï¿½ger, H.-P. & Kazenwadel, J. (1994). Modifying risks: Youthful drivers, drinking drivers and driving conditions. 1994 Annual Scientific Meeting of the Research Society on Alcoholism (RSA), Maui.
The "Grand Rapids" study
Figure 2-6 shows the results of the analysis for two groups of drivers of age 21 and over, (1) drivers involved in single-vehicle fatal crashes and (2) drivers involved in all fatal crashes. Risk curves of both groups are roughly the same for both groups up to a BAC of about .05, and then start to diverge. At BACs in the .080-.099 range, single-vehicle drivers in fatal crashes had a relative risk of about nine, compared to a relative risk of about six for drivers of all vehicles.
Of interest is the reduced relative of risk below 1.0 (.2) at BACs in the .001-.019 range, implying a lower risk than at BAC=0. This is remindful of the infamous "Grand Rapids Dip" (noted in the 1978 update) found in the well-known 1963 case-control study, but since discredited as being due to disproportionate representation of demographic subgroups in different blood alcohol concentration class intervals (Hurst, Harte, and Frith, 1994). Hurst and associates found no such dip after applying a statistical model to the data that accounted for such differences, finding that relative risk increased monotonically with BAC regardless of self-reported drinking frequency. (An earlier hypothesis for the dip was that the persons at low BACs were more often higher frequency drinkers who, for some reason, were safer drivers at low BACs.)