Examining trajectories of adolescent risk factors as predictors of subsequent high-risk driving behavior

JOURNAL OF ADOLESCENT HEALTH 2003;32:214 –224

ORIGINAL ARTICLE

Examining Trajectories of Adolescent Risk Factors as Predictors of Subsequent High-Risk Driving Behavior JEAN T. SHOPE, Ph.D., M.S.P.H., TRIVELLORE E. RAGHUNATHAN, Ph.D., AND SUJATA M. PATIL, M.P.H.

Purpose: To examine the effects on early high-risk driving behavior of changes over time (trajectories) in adolescent alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance. Methods: Statewide driving data were obtained for 4813 subjects who had completed at least two previous school-based questionnaires. The self-administered questionnaire data provided predictor measures from 5th through 10th grades. Trajectory information on predictor measures was summarized using each measure’s slope over time and level at the 10th grade data collection (last value). Regression models used serious offenses, alcoholrelated offenses, serious crashes, and alcohol-related crashes as outcomes, trajectory measures as predictors, and produced parameter estimates adjusted for demographic measures. Probabilities of having a serious offense or serious crash for five sample trajectories on each measure were obtained from the estimated regression models. Results: All four predictor measures were important, particularly in predicting serious offenses, alcohol-related offenses, and alcohol-related crashes. The highest probabilities for young adult high-risk driving were found among those with consistently high or increasingly high trajectories of friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance. Conclusions: Programs to prevent adolescent risk behavior should take into account environmental and personality influences. Prevention efforts need to emphasize preserving low levels, preventing increases, and promoting decreases over time of adolescent risk factors for From the Transportation Research Institute (J.T.S.), School of Public Health (J.T.S., T.E.R., S.M.P.), and the Institute for Social Research (T.E.R.), University of Michigan, Ann Arbor, Michigan. Address correspondence to: Jean T. Shope, Ph.D., M.S.P.H., University of Michigan Transportation Research Institute, 2901 Baxter Road, Ann Arbor, MI 48109-2150. E-mail: [email protected]. Manuscript accepted May 24, 2002. 1054-139X/03/$–see front matter PII S1054-139X(02)00424-X

unhealthy behaviors, such as high-risk driving. © Society for Adolescent Medicine, 2003 KEY WORDS:

Adolescents Automobile driving Accidents Traffic Alcohol drinking Peer group Risk factors

Injury, motor vehicle injury in particular, is the major cause of death and disability among adolescents and young adults [1–3]. A better understanding of the factors that predispose young people to high-risk driving is needed to develop interventions to prevent such outcomes. Several cross-sectional studies of the correlates of young people’s motor vehicle crashes or violations (convictions for offenses) identified associated factors that often included alcohol use [4 –14]. Very few longitudinal studies, however, have been conducted of the predictors of subsequent high-risk driving among young people. Those that have been done have identified various psychosocial and behavioral risk factors, particularly those that are alcohol-related [15–17]. This research examines four predictor measures (alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance) that represent different dimensions of Problem Behavior Theory (PBT). PBT has been promoted as an approach to understanding adolescent risk-taking behaviors, including risky driving [13,14,18,19]. The theory proposes three interrelated systems that lead © Society for Adolescent Medicine, 2003 Published by Elsevier, 360 Park Avenue South, New York, NY 10010

March 2003

to “proneness” or increased likelihood of problem behavior. The three systems are the personality system, the perceived environmental system, and the behavioral system. High-risk driving by adolescents is considered a problem behavior likely to be predicted by another problem behavior, alcohol use. Friends’ support for drinking comes from the perceived environmental system. Susceptibility to peer pressure and tolerance of deviance come from the personality system. One approach to the study of predictors of highrisk driving is to determine if the course or trend of a particular measure over time, that is its increasing, decreasing, or consistent “trajectory,” is a significant predictor. Several varying methodologies have been proposed and used to identify and analyze developmental and substance use trajectories. For example, Schulenberg et al. identified six conceptual trajectories of frequent binge drinking during the transition to young adulthood based on theory and previous research. The trajectories of binge drinking were then validated using cluster analysis, and later examined in models for their association with successful transitions to young adulthood [20]. Others have used cluster analytic techniques to identify trajectories of self-esteem and study their relationships to alcohol use and other behaviors [21,22]. Growth curve modeling is another approach that combines elements of repeated measures multivariate analysis of variance (MANOVA), confirmatory factor analysis, and structural equation modeling to analyze trajectories [23– 25]. For example, Curran et al. simultaneously estimated latent growth factors, or trajectories of heavy alcohol use and bar patronage, using factor analytical techniques and then used MANOVA to regress these latent growth factors on demographic variables such as age, race, and gender [26]. Another approach used semi-parametric mixture models to approximate a continuous distribution of developmental trajectories for boys’ physical aggression, opposition, and hyperactivity from age 6 to 15 years. Regression models were then used to determine which trajectories best predicted juvenile delinquency [27]. This technique, unlike the latent growth curve modeling technique described above, allows for identifying and contrasting distinct groups of subjects with different trajectories. The approaches described above assume continuous predictors and outcome measures, and primarily use trajectories as outcome variables. Our analysis, however, required that the trajectories be used as continuous predictors in regression analyses where the outcome measures for high-risk driving were

TRAJECTORIES OF RISK FACTORS PREDICT DRIVING

215

either binary or counts of events. We use an analytic approach that can provide information to help in deciding where high-risk driving prevention programs should be focused and what they might include. In this study, the effects of the trajectories of each measure (alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance) on subsequent novice driving behavior were analyzed. The purpose of these analyses was to examine the effects of these trajectories on the probability of a young, new driver’s having an offense or crash that seemed attributable to high-risk driving, rather than merely inexperience. Research questions also included whether different trajectories resulted in different outcomes; it was hypothesized that students with consistently high or increasingly high trajectories on alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance would have higher probabilities of poor driving outcomes. The first year of driving was examined, as well as the first 3 years, although the expectation was that the predictors would be stronger for the first year.

Methods Subjects and Data Collection As part of a substance abuse prevention study, self-administered questionnaire data including demographic, substance use, and psychosocial variables, were collected from students in two cohorts (the high school graduating classes of 1991 and 1992). Questionnaires were administered at seven time points in six southeastern Michigan public school districts beginning in the fall of 1984. The graduating class of 1991 was followed from grade 6 (about age 11–12 years) to grade 12, and the class of 1992 was followed from grade 5 (about age 10 –11 years) to grade 12. A total of 33,982 questionnaires were completed by 10,729 subjects. Questionnaire data for these analyses included only those surveys administered through the fall of grade 10 (a total of five time points for each cohort), after which most students reached age 16 years and thus were eligible for driver licensure. Response rates for each of the survey waves used in these analyses ranged from 80% to 95% of all students enrolled at the time of each survey. All students were invited to participate even if they had not been present for the first testing occasion in grade 5 or grade 6. Beginning in 1992, the names and birth dates of the 9714 students with completed questionnaires

216

SHOPE ET AL

JOURNAL OF ADOLESCENT HEALTH Vol. 32, No. 3

Table 1. Characteristics of Study Sample (N ⫽ 4813) a,b

Age (first survey, fall grade 5) Age (first survey, fall grade 6)a,c Age (last survey, fall grade 10)d White race Male gender Live with both parents (first survey)e Step-parent family (first survey)e Single-parent family (first survey)e Family structure change over time Serious offense in first year of driving Serious offense in first three years of driving Alcohol-related offense in first year of driving Alcohol-related offense in first three years of driving Serious crash in first year of driving Serious crash in first three years of driving Alcohol-related crash in first year of driving Alcohol-related crash in first three years of driving

Mean

SD

10.7 11.7 15.7

0.58 0.49 0.47

n

%

4056 2514 1292 237 473 988 541 1389 29 113 540 1017 23 57

84.3 52.2 64.5 11.8 23.6 20.5 11.2 28.9 0.6 2.3 11.2 21.1 0.5 1.2

a At this first survey occasion, only a randomized half of the students completed the survey. (This was so that pretest effects could be analyzed in the alcohol misuse prevention study for which these data were collected.) b n ⫽ 977. c n ⫽ 992. d n ⫽ 3013. e n ⫽ 2013 (11 did not answer this question).

through grade 10 were submitted to the Michigan Secretary of State’s office, and driver license data were obtained annually. Of these, 7541 (78%) had a valid Michigan driver’s license, and complete driver history data were obtained. In Michigan, most young people obtain a driver’s license as soon as possible after turning age 16 years, and indeed, 71.4% of these subjects had obtained a driver’s license within 1 year after the last data collection used in these analyses. For 2728 subjects with driver history data, questionnaire data were available at only one time point; therefore trajectories on their measures could not be calculated. Those subjects were not included in these analyses. A comparison between the 2728 subjects who completed only one survey and those subjects who completed more than one survey found that the distributions of race, gender, and the number of serious offenses for both 1 and 3 years did not differ significantly. However, differences were found in the distribution of the number of serious crashes. The mean number of serious crashes was slightly lower in the group of subjects who completed only one survey. The small difference was considered not likely to affect the overall results of these analyses. These analyses, therefore, used questionnaire and driver history data for 4813 subjects, 81% of whom had completed three or more questionnaires, for a total of 16,887 questionnaires. The study subjects were mostly white; slightly over half were male, and

just under two-thirds were living with both parents at the time of the first survey (Table 1). Compared with the general Michigan population, this sample had close to 10% more white subjects. A comparison between the study subjects’ third-year driving records and licensed Michigan drivers about the same age found that study subjects had offense and crash profiles quite similar to those of the general Michigan population [28]. Michigan-wide, 26% and 12% had at least one offense or one crash, respectively, in calendar year 1993. Among study subjects the same age, 31% and 12% had at least one offense or crash, respectively. The study’s protocol has been approved by the University of Michigan Medical School’s Institutional Review Board for Human Subject Research.

Outcome Measures Four outcome variables, created from the driver history data, were considered indicative of high-risk driving, rather than merely the inexperienced driving of a young novice. One, based on recorded offenses, reflects driving actions taken that could have caused a crash. A second related outcome measure reflects the number of alcohol-related offenses. Another outcome measure, based on recorded crashes, reflects actual crashes that did hap-

March 2003

pen. The last outcome measure reflects the number of alcohol-related crashes. Number of serious offenses was a count that included those ticketed offenses (violations) for speeding in excess of 15 miles/hour over the speed limit, reckless driving, vehicular homicide, or other major offenses. Typically, these offenses were assigned more points by the Secretary of State than less serious offenses (such as lesser speeding, no proof of insurance, license plate and vehicle offenses, and fraudulent identification). Number of alcohol-related offenses was a count that included only those ticketed offenses that were alcohol-related. Number of serious crashes was a count that included each individual subject’s crashes that were single-vehicle or at-fault (ticket issued on the same date). These types of crashes were considered likely to represent crashes resulting from high-risk driving, rather than from another driver’s actions, or the inexperience of the young driver. Care was taken not to count the same crash more than once if it was included in more than one category. Number of alcohol-related crashes was a count that included only those crashes where alcohol was determined to be a factor in the crash. Serious offenses and serious crashes were counted for the subjects’ first year of driving, as well as over their first 3 years of driving. About 11% of the study subjects had a serious offense or serious crash in their first year of driving (Table 1). Very few subjects had had an alcohol-related crash or offense during the first year of driving. In their first 3 years of driving, 29% had a serious offense, 2.3% an alcohol-related offense, 21% a serious crash, and 1.2% an alcoholrelated crash. Predictor Measures Four demographic measures were used. Age and gender were straightforward. Race was reported by students as white, black, or other race, and subsequently collapsed to a dichotomous variable, white and other race. Family structure was based on a single item: “Which of your parents do you live with most of the time?” Possible responses included mother and father, parent and a stepparent, a single parent, or someone else. For modeling purposes, three indicator variables were created, and living with both parents served as the baseline category. The first indicator variable included those respondents who lived with a parent and a stepparent. The second indicator variable included those respondents who lived with a single parent (mother or

TRAJECTORIES OF RISK FACTORS PREDICT DRIVING

217

father). Because responses on this question changed over time for 21% of the respondents (n ⫽ 988), a third variable indicating that this change had occurred was also included in these analyses. Trajectories in four predictor variables of interest were investigated. Alcohol use, in drinks per week, was calculated for each student at each data collection point by multiplying the frequency of alcohol use reported by the quantity of alcohol use reported for each substance used: beer, wine, and liquor (range ⫽ 0 – 42) [29]. Values greater than 42 (0.3%) were included in the category, 42. For modeling purposes, a natural log transformation was used because of the skewness observed in the data on this measure. Friends’ support for drinking was a measure based on five questionnaire items, and ranged from 0 “friends not supportive” to 16 “friends very supportive.” Four items had four response choices, “never, rarely, sometimes, often,” and covered how often friends talked about trying alcohol and how much they drink alcohol, offered a drink of alcohol, and pressured subjects to drink alcohol. The fifth item asked how friends feel about kids drinking alcohol, and had five response choices ranging from “a very good idea” to “a very bad idea.” For subjects with three or more reported values, responses were summed and multiplied by five divided by the number of nonmissing values. Those with three or more missing responses were given a missing value on this measure (191/16,887 surveys, 1%). The Cronbach alpha for this measure ranged from .75 to .83 across the survey occasions. The susceptibility to peer pressure measure ranged from 7 “less susceptible” to 28 “highly susceptible” based on responses to seven survey items [30]. The items had four response choices, “no, probably not, probably, and yes.” They covered such behaviors as tearing a page out of a library book if a friend dared one to, skipping school if one’s best friend were skipping, drinking alcohol if a friend offered it, and going to the movies instead of studying for a test if friends were going. For subjects with four or more reported values, responses on the seven items were summed and multiplied by seven divided by the number of nonmissing values. Those with three or more missing responses were given a missing value on this measure (138/16,887 surveys, ⬍ 1%). This measure had a Cronbach alpha that ranged from .86 to .90 across the survey occasions. The tolerance of deviance measure ranged from 5 “not tolerant” to 20 “very tolerant,” based on the responses to five questionnaire items [31]. The items

1.95

2.49

1.06 2,134

9.93 (3.59)

1,176 1,837

11.17 (3.59)

12.35 (4.69) 9.01 (3.56) 8.84 (3.29) 1,869 2,168

Data Analyses

10.54 (3.90) 7.58 3.05 7.59 2.73 1,936 1,930 The slope retransformed to number of drinks per year.

The study subjects’ levels on each of the four predictor variables for each cohort are given in Table 2. The levels increased over time, and substantial proportions of the subjects experienced these increases. Of the four predictors, alcohol use had the lowest proportion of subjects with increasing slopes, although 61% had increased their use from elementary school to high school. Differences between the two cohorts were assessed using Student’s t-tests for those timepoints where subjects in both cohorts were the same age. Results indicated that the two cohorts were similar with the exception of only one variable at one time-point; friends’ support for drinking at the Spring 6th survey. All subjects were therefore combined for further data analyses. Individual trajectories for each predictor variable of interest were obtained by regressing a predictor variable on time using the SAS Version 6.12 PROC REG procedure [32]. As shown in Table 2, time in this analysis ranged from 0 (Fall, 5th grade) to 5 (Fall, 10th grade). In particular, the longitudinal trajectory information on each predictor variable of interest for each subject who had completed two or more surveys was summarized using two statistics: (a) the slope of the regression line relating the level of the predictor variable and time, as a measure of the change from school grade 5 or 6 through grade 10 in the predictor measure; and (b) the last value or 10th grade response level, as a measure of the most recent predictor value. Specifically, the following model was fit separately for each subject i:

y it ⫽ ␣ i ⫹ ␤ it ⫹ ε it

a

1,868 977

992

7.22 (2.48) 7.03 (2.58) 6.59 (2.18)

’91 ’92 ’91 ’92 ’91

8.70 (2.84) ’91 ’92

Alcohol use (weekly) Range: 0 – 42 Friends’ support for drinking Range: 0 –16 Susceptibility to peer pressure Range: 7–28 Tolerance of deviance Range: 4 –16 Subject n

9.58 (3.42)

2.42 (2.03) 2.02 (1.79)

9.73 (3.49)

2.57 (2.20)

3.07 (2.46) 10.48 (4.08)

4.20 (0.65) 12.64 (4.97)

13.99 (4.99)

5.55 (3.33)

6.14 (3.24) 14.77 (4.84)

1.44

0.92

0.68a 2.12 (5.93) 2.13 (5.65) 6.09 (3.36) 1.30 (4.67) 0.80 (4.08) 0.58 (2.92) 4.05 (3.05) 0.25 (1.58) 0.27 (1.47) 2.81 (2.38) 0.16 (1.22) 0.19 (1.40) 0.10 (0.82)

’92 ’91 ’92

had four response choices “very wrong, wrong, a little bit wrong, and not wrong.” The items covered how wrong the subject thought it was to smoke without a parent’s permission; to go to a movie instead of studying for a test; to skip school without an excuse; to drink alcohol before being age 21 years; and to tear a page out of a school library book. For subjects with three or more reported values, responses on the five items were summed and multiplied by five divided by the number of nonmissing values. Those with four or five missing responses were given a missing value on this measure (537/ 16,887 surveys, 3%). The final measure had a Cronbach alpha that ranged from .81 to .95 across the survey occasions.

76%

77% 1.61

0.28

SD Mean

Fall 10th Time ⫽ 5 Measure

Class

Fall 5th Time ⫽ 0

Spring 5th Time ⫽ .5

Fall 6th Time ⫽ 1

Spring 6th Time ⫽ 1.5

Spring 7th Time ⫽ 2.5

Spring 8th Time ⫽ 3.5

Slope

JOURNAL OF ADOLESCENT HEALTH Vol. 32, No. 3

61% 74%

Increasing Slope

SHOPE ET AL

Table 2. Mean (and Standard Deviation) of Predictor Measures Over Time

218

where yit is the predictor variable of interest (e.g.,

March 2003

alcohol use) of subject i at time t and the errors, εit are normally distributed with mean zero and variance ␴i2. The slope or measure of change in a predictor for an individual is thus given by ␤i and the intercept is given by ␣i. For subjects who were missing a 10th grade response or last value, an estimate based on the regression line, ␺i ⫽ ␣i ⫹ 5 ␤i was used. Each individual’s trajectory was calculated separately for each of the four predictor variables (alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance). This methodology was used to calculate trajectory variables for all subjects who completed two, three, four, or five surveys. The distribution of the slopes and last values were found to be similar regardless of the number of completed surveys. These two summary quantities (slope and last value) were then used as predictors in logistic regression models for one-year driving outcomes and both one- and three-year alcohol-related outcomes, and Poisson regression models for three-year driving outcomes, with the numbers of serious offenses, and, separately, serious crashes as dependent variables. Confounding variables (age, gender, race, and family structure) were included in the models to derive adjusted estimates. Specifically, the following logit model was fit for one-year driving outcomes and both one- and three-year alcohol related driving outcomes:

logit [pr 共 D ⫽ 1 兩 ␺ i, ␤ i, Z i兲 ] ⫽ ␪ 0 ⫹ ␪ 1␺ i ⫹␪2␤i⫹␪3tZi where D ⫽ 1 indicates that a serious offense (or serious crash) occurred, ␪3 is a vector of parameter estimates for covariates Z ⫽ (age, gender, race, and family structure). The primary quantities of interest are the regression coefficients ␪1 and ␪2. Similar models with a log link for Poisson regression were fit for 3-year driving outcomes. Subjects who had missing values in the covariates or had only one nonmissing value on the four predictor variables were dropped from the analyses (1% to 5%). The statistical significance of the regression coefficients for the last values (␪1) and slopes (␪2) were assessed using the z-test. Overall model fit was assessed using the Hosmer-Lemeshow statistic for logistic regression models [33]. For 3-year driving outcomes, the number of serious offenses or crashes was used as an outcome in a Poisson regression model and the likelihood ratio was used to assess the goodness of fit. All models described above were fit using the

TRAJECTORIES OF RISK FACTORS PREDICT DRIVING

219

SAS PROC GENMOD or PROC LOGISTIC procedure. To summarize and interpret the results from the fitted logistic and Poisson regression models, five types of trajectories were selected from the many that subjects had for each predictor variable. “Low-Low” subjects were those whose last value was one standard deviation below the mean and whose slope was zero. These subjects had consistently low levels of the predictor variable. “High-Low” subjects were those whose last value was one standard deviation below the mean and whose slope was one standard deviation below zero. These subjects had decreasing levels of the predictor variable over time and a low last value. “Medium-Medium” subjects had average last values and average slopes. These subjects had slopes and last values within one standard deviation above or below the mean value. “Low-High” subjects were those whose last value was one standard deviation above the mean and whose slope was also one standard deviation above zero. These subjects had increasing levels of the predictor variable over time and a high last value. “High-High” subjects were those whose last value was one standard deviation above the mean and whose slope was zero. These subjects had consistently high levels of the predictor variable. The fitted regression models, with confounding variables set at their average values, were used to estimate the predicted probabilities or rates of serious offenses and serious crashes, for the first year and for the first 3 years of driving for each of the sample trajectories. (The numbers of subjects with alcohol-related offenses and crashes were too small to yield reliable predicted probabilities.) The four trajectory categories: High-High, Low-High, High-Low, and Low-Low, contained roughly 15% of the subjects included in this study. Most subjects (roughly 45% to 70%) had consistently average levels (Medium-Medium) of each predictor variable over time. The first four trajectory groups were chosen to examine subjects with predictor variable levels that represent extremes. The Medium-Medium trajectory group is displayed to provide a reference.

Results Results from the Hosmer-Lemeshow goodness-of-fit test for the logistic models and the likelihood ratio test for the Poisson models indicated that all but one model fit well. For the logistic models, the HosmerLemeshow test p values ranged from .14 to .99. Only one model (3-year alcohol-related crashes as the

SHOPE ET AL

220

JOURNAL OF ADOLESCENT HEALTH Vol. 32, No. 3

Table 3. Parameter Estimates and p Valuesa for Trajectory Variables (Slope, Last Value)b Non-Alcohol-Related Offenses

First year driving Alcohol use Slope Last value Friends’ support for drinking Slope Last value Susceptibility to peer pressure Slope Last value Tolerance of deviance Slope Last value First three years driving Alcohol use Slope Last value Friends’ support for drinking Slope Last value Susceptibility to peer pressure Slope Last value Tolerance of deviance Slope Last value

Alcohol-Related Offenses

Non-AlcoholRelated Crashes

Alcohol-Related Crashes

␤

p value

␤

p value

␤

p value

␤

p value

⫺0.272 0.181

0.258 0.005

⫺1.063 0.514

0.131 0.003

⫺0.242 0.160

0.310 0.013

⫺1.145 0.499

0.080 0.002

⫺0.195 0.078

0.001 ⬍.0001

⫺0.398 0.117

0.049 0.030

⫺0.038 0.026

0.531 0.104

⫺0.471 0.171

0.017 0.000

⫺0.113 0.040

0.003 ⬍.0001

⫺0.323 0.088

0.018 0.006

⫺0.034 0.022

0.387 0.031

⫺0.328 0.104

0.019 0.001

⫺0.106 0.029

0.048 0.042

⫺0.071 0.055

0.730 0.241

0.020 0.012

0.717 0.403

⫺0.322 0.101

0.105 0.026

⫺0.072 0.068

0.614 0.080

⫺0.876 0.415

0.029 ⬍.0001

⫺0.175 0.096

0.279 0.027

⫺0.947 0.427

0.056 0.001

0.030 ⫺0.085

0.0001 0.003

⫺0.266 0.104

0.016 0.0002

⫺0.024 0.017

0.522 0.093

⫺0.329 0.123

0.018 0.001

⫺0.049 0.021

0.009 ⬍.0001

⫺0.227 0.072

0.002 ⬍.0001

⫺0.009 0.010

0.731 0.088

⫺0.187 0.070

0.052 0.002

⫺0.049 0.017

0.063 0.013

⫺0.215 0.077

0.036 0.002

0.023 0.002

0.508 0.787

⫺0.236 0.085

0.082 0.008

a

p values based on the Wald Chi-Square statistic for parameter estimates. All models adjusted for age, gender, family structure, and race. All models are significant in overall goodness of fit tests with the exception of alcohol-related crash during the first 3 years of driving for susceptibility to peer pressure. b

outcome with the susceptibility to peer pressure trajectories) did not fit well (p ⫽ .04). All Poisson models had goodness-of-fit p values that ranged from .87 to .99. All four predictor measures (alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance) were important in predicting certain aspects of subsequent high-risk driving behavior as the significant regression results show in Table 3. In predicting non-alcohol-related serious offenses in the first year of driving, both the last values and the slopes of friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance were significant. Only the last value of alcohol use was significant. Non-alcohol-related offenses in the first 3 years of driving were predicted significantly by both the slopes and last values of friends’ support for drinking and susceptibility to peer pressure. Only the last value of tolerance of deviance reached significance, and neither the slope nor last value of alcohol use reached significance.

In predicting alcohol-related serious offenses in the first year of driving, both the last values and the slopes of friends’ support for drinking and susceptibility to peer pressure were significant. Only the last value of alcohol use was significant, and neither the slope nor last value of tolerance of deviance was significant. Alcohol-related offenses in the first 3 years of driving were predicted by both the slopes and the last values of all four measures. In predicting non-alcohol-related crashes in the first year of driving, only the last values of alcohol use and susceptibility to peer pressure were significant predictors. Non-alcohol-related crashes in the first 3 years of driving were only predicted significantly by the last value of alcohol use. In predicting alcohol-related crashes in the first year of driving, both the slopes and last values of friends’ support for drinking and susceptibility to peer pressure were significant predictors. The last values of alcohol use and tolerance of deviance were significant, and their slopes indicated a trend toward

March 2003

TRAJECTORIES OF RISK FACTORS PREDICT DRIVING

221

Figure 1. Probabilities (and 95% Confidence Intervals) of Serious Offenses and Serious Crashes in the First Year and First 3 Years of Driving Predicted from Trajectories on Four Measures.

222

SHOPE ET AL

significance. Alcohol-related crashes in the first 3 years of driving were predicted by both the slope and last value of friends’ support for drinking, and by the last values of alcohol use, susceptibility to peer pressure, and tolerance of deviance. The slopes of those three measures also indicated a trend toward significance. The predicted probabilities for the five sample trajectories (Figure 1) show the High-High trajectories with considerably higher probabilities of serious offenses and serious crashes than the Low-Low trajectories. All predicted probabilities have 95% confidence intervals that are significant and do not include 0. For nearly all predictors and driving outcomes, the Low-Low trajectories had the lowest probabilities followed by the High-Low, MediumMedium, Low-High, and High-High trajectories. For example, subjects who were consistently low in terms of friends’ support for drinking had only a 7% chance of having a serious offense in the first year of driving, whereas subjects whose friends consistently supported drinking had more than twice that probability of a serious offense (16%).

Discussion Young people’s alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance increase during adolescence, and have been shown previously to be important predictors of various health risk behaviors. Each of these measures represents a system (social environment, personality, and behavior) in Jessor’s Problem Behavior Theory [13,14,18]. The results of this study provide further support for the importance of these adolescent measures in predicting subsequent highrisk driving (over the first year and first 3 years). In particular, serious offenses (both non-alcohol- and alcohol-related), and alcohol-related serious crashes were significantly predicted from trajectories on these measures. Non-alcohol-related crashes during the first year and first 3 years were significantly predicted by only 3 of 16 trajectory indicators. Previous research has shown these four measures to be correlated with high-risk driving. Adolescent substance use, particularly alcohol, was a key correlate or predictor of high-risk driving in several studies [4,5,7–11,15–17,34,35]. Other studies have identified best friends’ drinking habits as an important early influence on subjects’ alcohol use [36,37], and susceptibility to peer pressure as a significant correlate of adolescent substance use [30]. Tolerance

JOURNAL OF ADOLESCENT HEALTH Vol. 32, No. 3

of deviance has accounted for variation in adolescents’ involvement in problem drinking [38], and in psychosocial unconventionality associated with drinking and driving in Problem Behavior Theory [12]. Only a few of these studies were longitudinal [15–17], and none used trajectories of the predictor variables. This study examined trajectories of the predictor measures from as early as 5th grade through 10th grade, using both the trajectory’s slope and the last value for each measure. For the most part, both of these indicators of all the four measures were significant predictors, although the last value was significant slightly more often than the slope, particularly for alcohol use. A last value is likely to be a better predictor simply because it is more recent. This finding is important, however, because it serves as a reminder that a youngster’s status as of 10th grade can be more important than what he or she may have experienced up until that time. Serious offenses (non-alcohol-related) were predicted more successfully from the measures studied than serious crashes (non-alcohol-related). This finding may be because serious offenses cover a wider array of high-risk driving behavior than do serious crashes, and the behavior has been directly observed by the ticketing officer. Although serious crashes were defined as those that were at-fault or singlevehicle, there are very likely factors other than the four predictor measures that contribute to such crashes. Offenses are important, and are generally considered to be good predictors of future crashes. The fact that the measures were more predictive of the alcohol-related crashes than the non-alcoholrelated crashes is very likely owing to the nature of the predictor measures; each included at least some items that touched on adolescent alcohol use to some extent. Offenses, however, were relatively equally well predicted, whether alcohol-related or not, perhaps because of the different behaviors involved. Even those who speed do not wish to be involved in a crash. It is impressive that the measures predicted 3-year driving outcomes equally as well as first-year driving outcomes. This finding may demonstrate that these novice drivers (ages 16 –18 years) continue to be influenced by their 10th grade social environment of friends’ experiences and peer influences, and carry their attitude toward deviance throughout their high school years. The predicted probabilities of high-risk driving outcomes for various selected trajectories over time demonstrate that those with consistently high or

March 2003

increasingly high levels of alcohol use, friends’ support for drinking, susceptibility to peer pressure, and tolerance of deviance have a greater likelihood of having a serious crash or serious offense in both their first year and first 3 years of driving. The highest probabilities were for serious offenses among those with consistently high trajectories on friends’ support for drinking (35%) and susceptibility to peer pressure (36%), compared with 26% for those with consistently low trajectories on those measures. One limitation of these analyses, however, is that they addressed only the separate, not the combined, effects of the predictor measures. Although the findings may not be completely generalizable to other populations, they are, nonetheless, based on a relatively representative sample of young people. This type of analytic approach could be used with other predictor measures in future research. Meanwhile, the results of this study serve as a reminder of the types of personality, environmental, and behavioral factors that prevention programs need to incorporate into their planning. These findings should encourage those who work with young people to maintain their focus on preventing young people’s alcohol use, involvement with friends who support drinking, susceptibility to peer pressure, and acceptance of deviance, even when these may have been problems previously for a young person. The findings that low levels on these measures in the fall of 10th grade, regardless of the previous levels or the slope, protected young people from serious driving offenses and crashes is important. Prevention efforts need to focus on preserving low levels, preventing increases, and promoting decreases in adolescent risk factors for unhealthy behaviors such as high-risk driving. This research was supported by the National Institute on Alcohol Abuse and Alcoholism, Grants RO1 AA09026 and RO1 AA06324. The authors are grateful for the support and assistance of the local school districts, the Michigan Secretary of State’s Office, and the research staff.

References 1. Bonnie RJ, Fulco DE, Liverman CT (eds). Reducing the Burden of Injury. Washington, DC: National Academy Press, 1999. 2. Insurance Institute for Highway Safety. Fatality Facts: Teenagers. Arlington, VA. Available at: http://www.iihs.org/ facts/teen.htm. Accessed January 8, 2002. 3. National Highway Traffic Safety Administration. Traffic safety facts. Available at: http://www.nhtsa.dot.gov/people/ ncsa/factshet.html. Accessed January 8, 2002. 4. Pelz DC, Schuman SH. Drinking, hostility, and alienation in driving of young men. In: Chafetz ME (Ed). Proceedings of the Third Annual Alcoholism Conference of the National Institute

TRAJECTORIES OF RISK FACTORS PREDICT DRIVING

223

on Alcohol Abuse and Alcoholism. Washington, DC: NIAAA, 1973;50 –74. 5. Mayer RE, Treat JR. Psychological, social and cognitive characteristics of high-risk drivers: A pilot study. Accid Anal Prev 1977;9:1–8. 6. Farrow JA. Drinking and driving behaviors of 16 –19 yearolds. J Stud Alcohol 1985;46:369 –74. 7. Murray A. The home and school background of young drivers involved in traffic accidents. Accid Anal Prev 1998;30:169 –82. 8. Beirness DJ, Simpson HM. Lifestyle correlates of risky driving and accident involvement among youth. Alcohol Drugs Driving 1988;4:193–204. 9. Gregersen NP, Berg HY. Lifestyle and accidents among young drivers. Accid Anal Prev 1994;26:297–303. 10. Arnett J. Drunk driving, sensation seeking, and egocentrism among adolescents. Pers Individ Dif 1990;11:541–6. 11. Copeland LA, Shope JT, Waller PF. Factors in adolescent drinking and driving: Binge drinking, cigarette smoking, and gender. J Sch Health 1996;66:254 –60. 12. Donovan JE. Young adult drinking-driving: Behavioral and psychosocial correlates. J Stud Alcohol 1993;54:600 –13. 13. Jessor R. Risky driving and adolescent problem behavior: An extension of problem-behavior theory. Alcohol Drugs Driving 1987;3:1–11. 14. Jessor R. Risky driving and adolescent problem behavior: Theoretical and empirical linkage. In: Benjamin T (ed). Young Drivers Impaired by Alcohol and Other Drugs. London: Royal Society of Medicine Services, 1987:97–110. 15. Simpson HM, Beirness DJ. Traffic accidents and youth: Alcohol and other lifestyle factors. J Alcohol Beverage Med Research Found 1993;3:77–84. 16. Karlsson G, Romelsjo A. A longitudinal study of social, psychological and behavioural factors associated with drunken driving and public drunkenness. Addiction 1997;92: 447–57. 17. Begg DJ, Langley JD, Williams SM. A longitudinal study of lifestyle factors as predictors of injuries and crashes among young adults. Accid Anal Prev 1999;31:1–11. 18. Jessor R, Donovan JE, Costa FM. Beyond Adolescence: Problem Behavior and Young Adult Development. New York: Cambridge University Press, 1991. 19. Wilson RJ, Jonah BA. The application of problem behavior theory to the understanding of risky driving. Alcohol Drugs Driving 1988;4:173–91. 20. Schulenberg J, O’Malley PM, Bachman JG, et al. Getting drunk and growing up: Trajectories of frequent binge drinking during the transition to young adulthood. J Stud Alcohol 1996;57:289 –304. 21. Hirsch BJ, DuBois DL. Self-esteem in early adolescence: the identification and prediction of contrasting longitudinal trajectories. J Youth Adolesc 1991;20:53–72. 22. Zimmerman MA, Copeland LA, Shope JT, Dielman TE. A longitudinal study of self-esteem: Implications for adolescent development. J Youth Adolesc 1997;26:117–41. 23. McArdle JJ, Epstein D. Latent growth curves within development structural equation models. Child Dev 1987;58:110 –33. 24. Meridith W, Tisak J. Latent curve analysis. Psychometrika 1990;55:107–22. 25. Muthen BO. Latent variable modeling in heterogeous populations. Psychometrika 1989;4:139 –57. 26. Curran PJ, Harford TC, Muthen BO. The relation between heavy alcohol use and bar patronage: A latent growth model. J Stud Alcohol 1996;57:410 –9.

224

SHOPE ET AL

27. Nagin D, Tremblay RE. Trajectories of boys’ physical aggression, opposition and hyperactivity on the path to physically violent and nonviolent juvenile delinquency. Child Dev 1999; 70:1181–96. 28. Michigan Department of State. Michigan driver file: Annual driver report 1993-DR/80005. Lansing, MI: Program Development Section, Bureau of Driver Improvement, Michigan Department of State, 1994. 29. Shope JT, Copeland LA, Dielman TE. Measurement of alcohol use and misuse in a cohort of students followed from grade 6 through grade 12. Alcohol Clin Exp Res 1994;18:726 –33. 30. Dielman TE, Campanelli PC, Shope JT, Butchart AT. Susceptibility to peer pressure, self-esteem and health locus of control as correlates of adolescent substance use. Health Educ Q 1987;14:207–21. 31. Rachal JV, Williams JR, Brehm ML, et al. A National Study of Adolescent Drinking Behavior, Attitudes, and Correlates: Final Report. Rockville, MD: National Institute on Alcohol Abuse and Alcoholism, 1975.

JOURNAL OF ADOLESCENT HEALTH Vol. 32, No. 3

32. Statistical Analysis System [computer program]. Version 6 for Windows. Cary, NC: SAS Institute Inc., 1993. 33. Hosmer DW, Lemeshow S. Applied Logistic Regression. New York: John Wiley, 1989. 34. Shope JT, Waller PF, Lang SW. Alcohol-related predictors of adolescent driving: Gender differences in crashes and offenses. Accid Anal Prev 1996;28:755–64. 35. Shope JT, Waller PF, Raghunathan TE, Patil SM. Adolescent antecedents of high-risk driving behavior in young adulthood: substance use and parental influences. Accid Anal Prev 2001; 33:649 –58. 36. Engels RC, Knibbe ME, Vries HD, et al. Influences of parental and best friends’ smoking and drinking on adolescent use: A longitudinal study. J Appl Soc Psychol 1999;29:337–61. 37. Urberg KA, Degirmencioglu SM, Pilgrim C. Close friend and group influence on adolescent cigarette smoking and alcohol use. Dev Psychol 1997;33:834 –44. 38. Costa FM, Jessor R, Turbin MS. Transition into adolescent problem drinking: The role of psychosocial risk and protective factors. J Stud Alcohol 1999;60:480 –90.