The Long Beach/Fort Lauderdale relative risk study

T2007 – Seat t le, Washingt on The Long Beach/Fort Lauderdale Relative Risk Study 1 R.D. Blomberg 2 , R.C. Peck 3 , H. Moskowitz 4 , M. Burns4, D. Fiorentino4 A case-control study funded by the National Highway Traffic Safety Administration (NHTSA) and the National Institute on Alcohol Abuse and Alcoholism (NIAAA) examined the relative crash risks associated with drivers’ blood alcohol concentrations (BACs). The measure, relative crash risk, is defined as the ratio of the proportion of crash drivers to the proportion of control drivers in a BAC classification, compared to a similarly formed ratio of drivers with 0.00% BACs. The resulting detailed final report by Blomberg, Peck, Moskowitz, Burns and Fiorentino (2005) details both the extensive methodological development of the study and a first set of analyses of its data. This paper summarizes the study’s approach and results. The specific objectives of the study were to: • • Determine the relative crash risk of drivers at various BACs compared to drivers with zero BACs while controlling for other factors (e.g., age, gender, drinking patterns) Determine the relative crash risk of groups of drivers (e.g., youth, males, heavy drinkers) at various BACs compared to similar groups with zero BACs. Background A relative risk function derived from epidemiological data collected in Grand Rapids, Michigan during the 1960s is the primary basis for the current legal limits for driving after drinking (Borkenstein et al., 1964). It is possible, however, that changes in driving, alcohol consumption, or drinking-driving practices have occurred during the decades since that study was conducted, and that the changes altered the relative risks associated with the combination of alcohol and driving. There have also been significant advances in the measurement of BAC and in statistical techniques for analyzing case-control data that could enhance the type of information to be obtained from a case-control study. Since regulatory agencies and legislators need current information about the crash risks associated with driving at BACs above zero, a new study to update the epidemiological evidence was conducted. 1 Prepared for presentation at T2007 – the combined annual meetings of the International Council on Alcohol, Drugs, and Traffic Safety (ICADTS) and The International Association of Forensic Toxicologists (TIAFT) together with the 8th Ignition Interlock Symposium (IIS), Seattle, WA, U.S.A., August 26-30, 2007. 2 Dunlap and Associates, Inc., Stamford, CT, U.S.A. 3 R.C. Peck and Associates, Folsom, CA, U.S.A. 4 Southern California Research Institute, Los Angeles, CA, U.S.A. Research Approach A case-control approach was adopted for the study that involved the following particulars: • • • • • • • During 12-month periods, data were collected in two locations to acquire samples of at least 1,300 crashes at each site. Data were obtained from crash-involved (case) and non-crash-involved (control) drivers during the hours 1600 - 0200 in Long Beach, California and 1700 - 0300 in Fort Lauderdale, Florida. The evening and nighttime hours were selected for sampling because drivers who have been drinking are most likely to be on the roadways at those times. A two-person research team consisting of a police officer and an interviewer collected the data from crash and control drivers. Drivers involved in crashes and non-crash involved control drivers were matched for crash location, day, time, and travel direction. Control drivers were obtained one week after a crash at the same site, traveling in the same direction, on the same day of the week and at the same time of day. Using random selection procedures, drivers in the flow of traffic were stopped and asked to participate in the study. Two control drivers were sampled for each crash-involved driver. Interviews with crash and control drivers were conducted using an identical questionnaire to obtain data about the covariates of drinking and driving such as drinking history, sleep patterns and demographic data. Drivers’ BACs were measured at roadside by obtaining breath specimens with evidential quality portable/preliminary breath testers (PBTs). Hospitalized and fatally-injured drivers’ BACs were obtained from the analyses of blood specimens by police, hospitals, and coroners. Systematic observations by the research teams, together with BAC estimates obtained using passive alcohol sensors (PAS), provided additional information about drivers’ use of alcohol. These data served as the basis for estimates of the bias resulting from the refusal of some drivers to provide the evidential quality breath specimens. A complete, single case for the statistical analysis of BAC and crash risk therefore included questionnaire information, police crash report data and field team observations on a crash driver, two control drivers and the location of the crash. Statistical Analysis Relative risk curves describing the relationship between BAC and crash risk were derived from a sequence of univariate and multivariate logistic regression analyses. After establishing simple univariate models, more complex models were derived in which the relative risk estimates were adjusted for potentially confounding covariates (e.g., age, education) and other sources of bias, such as refusal to provide a BAC specimen and non-recovered hit and run cases. Most of the models were based on the total sample and all crashes, but separate analyses were also done for the two sites (Long Beach and Fort Lauderdale), type of crash, and other subgroups of interest (e.g., age and alcohol consumption patterns). The statistical analyses produced answers to the following questions: • • • • • • How does crash risk increase as a function of a driver’s BAC? Is the relationship between BAC and crash risk materially altered by controlling potential sources of bias and confounding covariates? Is the BAC-crash relative risk curve similar for different crash types? Is the BAC-crash relative risk curve similar for the two study sites (Long Beach and Fort Lauderdale)? Does the BAC-crash relative risk curve vary by driver characteristics, such as age, gender and alcohol consumption patterns? How accurately can individual crash involvement be predicted from knowledge of a driver’s BAC and other characteristics? Results A total of 2,871 crashes (Long Beach 1,419; Fort Lauderdale 1,452) yielded 4,919 crash and 10,066 control drivers (14,985 total). In total, 93.5% of the drivers who were contacted at crash scenes agreed to participate. An additional 603 fled the scene of their crash. One hundred four (17.2%) of those hit-and-run drivers were located within two hours of the crash, and 94 (90.4%) provided breath specimens. Those who were not located or refused to participate reduced the participation rate of crash-involved drivers to 83.1% and the percentage who provided usable breath specimens to 81.3%. Non-crash drivers participated as controls at a higher rate of 97.9%. Since the data obtained in California and Florida proved to be similar, most of the logistic regression analyses were performed on the total data set. The analyses showed elevated relative risk with increasing BAC and a strongly accelerated risk at BACs greater than 0.10%. The influence of covariates on the magnitude and shape of the curve based on these initial analyses was relatively modest. It is important to note, however, that the scope of the initial study funded by NHTSA and NIAAA did not include all of the analyses of potential interest from the collected data. For this reason, the data and associated codebooks were delivered to NHTSA and made available to the research community for further detailed analyses (c.f., Peck, Gebers, Voas and Romano, 2007; Romano, Peck and Voas, 2007; Voas, Peck, Romano and Gebers, 2007). As part of the analysis of the data, it was assumed that crash risk was underestimated in analyses of the “raw” or uncorrected case and control data because of three sources of bias: • • • Differences in non-participation rates between the crash and control groups—Alcohol positive crash drivers refused to participate and failed to complete the interview more often than did other classes of drivers sampled. Non-apprehended hit-and-run drivers—The BACs of hit-and-run drivers who were apprehended were much higher than BACs of drivers who did not flee their crash scene. Missing covariate data due to non-participation or incomplete interviews. Re-weighting adjustments based on observational data and PAS estimates of BACs were used to correct for these biases. The resulting adjusted model of relative risk showed greater risks at all BACs with large increases at high BACs. This confirmed the assumption that the biasing sources were suppressing the relative risk estimates. Table 1 displays calculated relative risks from: a model with no covariates (column 2); a model with demographic covariates (column 3); and a model adjusted for all three major biases discussed above (column 4). Column 5 is from a model developed by Allsop (1966) based on the Grand Rapids Study data (Borkenstein et al., 1964) and is comparable to the no-covariate model for this study. Table 1. Relative Risk Models Relative Risk No Covariates 0.00 1.00 Non-reactive Demographic Covariates 1.00 .01 .91 .94 1.03 .92 .02 .87 .92 1.03 .96 .03 .87 .94 1.06 .80 .04 .92 1.00 1.18 1.08 .05 1.00 1.10 1.38 1.21 .06 1.13 1.25 1.63 1.41 .07 1.32 1.46 2.09 1.52 .08 1.57 1.74 2.69 1.88 .09 1.92 2.12 3.54 1.95 .10 2.37 2.62 4.79 .11 2.98 3.28 6.41 .12 3.77 4.14 8.90 .13 4.78 5.23 12.06 .14 6.05 6.60 16.36 .15 7.61 8.31 22.10 .16 9.48 10.35 29.48 .17 11.64 12.74 39.05 .18 14.00 15.43 50.99 .19 16.45 18.31 65.32 .20 18.78 21.20 81.79 .21 20.74 23.85 99.78 .22 22.07 25.99 117.72 .23 22.51 27.30 134.26 .24 21.92 27.55 146.90 BAC Final Adjusted Estimate Grand Rapids* 1.00 1.00 5.93 4.94 10.44 21.38 153.68 20.29 26.60 .25+ *From reporting of Grand Rapids Study data in Table 25 (a) of Allsop (1966). Logistic regression analyses were performed for subgroups of the driving population (e.g., youth, heavy drinkers), but each subgroup was small. It is unclear from these analyses whether the small samples account for unexpected findings such as the absence of increased risk for young drivers at low BACs. This issue cannot be resolved without larger numbers in the subgroups or more sensitive analyses such as those performed by Peck et al. (2007) and Voas et al. (2007). The study data clearly demonstrate that case-control studies need to control or adjust for differential non-participation and non-random missing data, particularly for the loss of data from hit-and-run drivers. The effect of these sources is evident in a comparison of the third and fourth columns of Table 1. Note the magnitude of the underestimation of relative risk at very high BACs evident in the risks adjusted only for non-reactive covariates (column 2). For example, the relative risk of 26.60 at BACs > 0.25% in column 3 becomes 153.68 when it is adjusted for the hit and run and refusal biases (column 4). When adjusted for non-participation and missing data bias, the results suggest that the small dip observed at 0.01– 0.03% BAC in the unadjusted risk calculations (columns 2 or 3 in Table 1) may be an artifact of sampling errors and small sample biases. The magnitude of the adjusted risk elevations are too small in relation to the standard errors of the model, however, to reject either the hypothesis of no increase or even an hypothesis of slight decreases in risk at 0.01 – 0.03% BAC. Regardless of the direction of the risk change, if any, at these low BACs, however, the size of the relative risk deviation from unity is sufficiently small to be of absolutely no practical consequence. In a re-analysis of the Grand Rapids data, Hurst, Harte and Frith (1994) also showed that the relative risk curve changed substantially and an observed decrease or dip in risk at low BACs disappeared with an adjustment for drivers’ drinking frequency. Although the final adjusted risk curve (column 4 in Table 1) from the present study failed to replicate the phenomenon and artifact noted by Hurst, Harte and Frith (1994) , the present analysis, as discussed above, did show the existence of a similar dip in the model with no covariates. This dip disappeared when adjusted for the sources of bias discussed above. As mentioned, however, the dip, even if real, is of such small magnitude that it is of academic interest only and does not represent anything that needs to be considered in policy making. The study results with respect to general relative risk due to alcohol confirm a notable dose-response relationship beginning at 0.04% BAC and increasing exponentially at > 0.10% BAC. The adjusted relative risk function (column 4 in Table 1) is graphed in Figure 1 and illustrates the extraordinary magnitude of the crash risk at high BACs. It must be noted that the analyses performed on the data as part of the data collection study represent only a small portion of the potential inherent in the collected information. It is fully anticipated that further analyses such as those by Peck et al., (2007), Romano, Peck and Voas (2007) and Voas et al. (2007) will provide additional insights into the basic alcohol and crash risk relationship and various additional nuances based on more detailed examinations of the abundance of covariate information collected. 180.00 Relative Crash Risk (BAC 0=1.0) 160.00 140.00 120.00 100.00 80.00 60.00 40.00 20.00 0.00 0 .01 .02 .03 .04 .05 .06 .07 .08 .09 .10 .11 .12 .13 .14 .15 .16 .17 .18 .19 .20 .21 .22 .23 .24 .25 BAC % Figure 1. Final Adjusted Relative Risk Estimate Acknowledgements Funded by the National Highway Traffic Safety Administration and the National Institute on Alcohol Abuse and Alcoholism under Contract No. DTNH22-94-C-05001. References Allsop, R. E. (1966). Alcohol and Road Accidents. (Road Research Laboratory Report No. 6). Harmondsworth, England: Road Research Laboratory, Ministry of Transport. Blomberg, R., Peck, R.C.. Moskowitz. H. Burns, M. and Fiorentino, D. (2005). Crash Risk of Alcohol Driving: A Case-Control Study. Stamford, CT: Dunlap & Associates (available at http://www.dunlapandassociatesinc.com/Publications.html). Borkenstein, R. F., Crowther, R. F., Shumate, R. P., Zeil, W. W., and Zylman, R. (1964). The role of the drinking driver in traffic accidents. Bloomington, IN: Department of Police Administration, Indiana University. Hurst, P.M., Harte, D. and Frith, W.J. (1994). The Grand Rapids dip revisited. Accident Analysis and Prevention, 26(5), 647-654. Peck, R.C., Geebers, M.A., Voas, R.B. and Romano, E. (2007). Improved methods for estimating relative crash risk in a case-control study of blood alcohol levels. Presentation at T2007, Seattle, WA, August 26-30,2007. Romano, E., Peck, R.C. and Voas, R.B. (2007). A comparison study of factors affecting driving and crashes (alcohol-related or not). Presentation at T2007, Seattle, WA, August 2630,2007. Voas, R.B., Peck, R.C., Romano, E. and Geebers, M.A. (2007). Alcohol-related crashes: relative risk by demographic groups. Presentation at T2007, Seattle, WA, August 26-30,2007.