Sample Recruitment

A detailed description of the recruitment methods for study participants has been published previously.(25) Briefly, between June and October, 2003, we conducted a random-digit dial (RDD) telephone survey of 6522 U.S. adolescents aged 10–14 years, designed to capture a representative sample of U.S. adolescents. We obtained parental consent and adolescent assent prior to interviewing each respondent, with a Certificate of Confidentiality protecting information on risk behaviors. All aspects of the survey were approved by the institutional review boards at Dartmouth Medical School and the survey research firm (Westat, Rockville MD). The CASRO response rate for the survey was 32% (which includes estimates of the number of adolescents lost because the household did not answer the phone) and the completion rate was 66%, rates comparable to other contemporary random digit dial surveys that involve consent of the parent and assent of the adolescent.(26) A recent publication by Beiner et al,(26) examined the impact of declining response rates in telephone surveys of smoking. They found that under- and over-representation of population subgroups has not changed and also found no evidence that declining response rates have resulted in biased estimates of smoking behavior.

We compared the distributions of age, sex, and census region in our unweighted sample with those of the 2000 U.S. Census,(25) confirming there was little evidence of under- or over-representation of demographic subgroups in this study and consistent with the results of Biener et al. The 6522 adolescents at baseline were equally distributed by age (between 10 and 14 years) and sex, and represented every U.S. state. Some 62 percent of adolescents described themselves as white, 13.6% as Black, and 18.2% as Hispanic. Seventy four percent of parents were homeowners, and family income (for the 6083 parents who responded) was distributed as follows: 7.7% <=$10,000, 10.4% $10001–20,000, 12.1% $20001–30,000, 20.9% $30001–50,000, 19.3% $50001–75,000, 29.6% >$75,000; these income groupings were almost identical to U.S. 2000 census data on median family income for families with adolescents in this age range.

Follow-up surveys

Three follow-up surveys were conducted at 8-month intervals with 84% (n=5503) retained at 8 months (8M), 77% (n=5019) at 16 months (16M), and 71% (n=4575) at 24 months (24M). Adolescents lost to follow-up were more likely to be of non-white ethnicity (Black, Hispanic, or other) and older age, to have parents with lower educational attainment, and to be higher in sensation seeking. Established smoking status and binge drinking status at baseline did not predict attrition at wave 4. Subjects were included in the analysis if they had data for sensation seeking at baseline and substance use at one or more follow up surveys.

Outcome Variables

Adolescents were surveyed regarding their use of alcohol and tobacco. Binge drinking was ascertained through the question, “"Have you ever had 5 or more drinks of alcohol in a row, that is, within a couple of hours?" (yes, no). Assessment of established smoking was based on a query about lifetime smoking, "How many cigarettes have you smoked in your life?" (none, a few puffs, 1 to 19 cigarettes, 20 to 100 cigarettes, or more than 100 cigarettes). Any adolescent who stated smoking more than 100 cigarettes was defined as an established smoker, consistent with surveys of adult samples and other studies of adolescents.(26) A dichotomous variable was developed to model onset of each substance use outcome, starting with status at 24M. For any adolescent not in the 24M survey, 16M or 8M substance use status was used if available; for adolescents with multiple assessments, any positive response was taken as an indication of substance use. Thus, both outcome variables were composed of data for 5834 adolescents who had responses for one or more follow up surveys (some adolescents missing from wave 2 were recaptured at waves 3 or 4).

Predictor Variable

We developed a 4-item sensation seeking scale for younger children, tapping two of four constructs identified by Zuckerman as important components of sensation seeking:(27) thrill/adventure seeking (“I like to do scary things” and “I like to do dangerous things”) and boredom susceptibility (“I often think there is nothing to do”). We also included intensity seeking (“I like to listen to loud music”), a component of the Arnett Inventory of Sensation Seeking (AISS) that was more highly correlated with risk behaviors than novelty seeking.(28) Response categories included “not like you”, “a little like you”, “a lot like you” and “just like you”, with scores for the individual items ranging from 1 to 4; these scores were summed across items (range 4–16). The 4 item scale had alphas of .60, .58, .64, and .63 and average inter-item correlations of .27, .26, .31, and .30 at waves 1 to 4, respectively.

Regarding the normative stability of the measure, the correlation between sensation seeking at baseline and subsequent waves ranged from 0.54 (baseline-wave 4) to 0.73 (wave 3 - wave 4), similar to the ranges for stability of the subscales for the SSSV in a sample of Norwegian adolescents over 20 months (where correlations ranged from 0.59 to 0.74).(11) The means (standard deviation) for sensation seeking for the 4081 adolescents with data from all four waves were 7.8 (2.4), 8.0 (2.4), 8.2 (2.5), and 8.4 (2.5) for baseline, wave 2, wave 3 and wave 4 respectively. A longitudinal generalized least squares random effects regression model showed that the average mean increase of 0.2 was statistically significant (z = 21, p < 0.0001), confirming previous observations on the growth of sensation seeking during adolescence.(13)

The reliabilities for this scale compare well with another 4-item measure of sensation seeking, the Brief Sensation Seeking Scale (BSS4). Stephenson et.al.(29) reported a reliability of 0.66 and average inter-item correlation of 0.33 for the BSSS4. We chose this measure over the BSSS4 because we felt the language in some items was too advanced (e.g., “I prefer friends who are exciting and unpredictable”) for the reading ability of our younger participants (10–11 years). Furthermore, we felt that another two-item scale (developed by Slater) with higher reliability(29) was too narrow because it assessed only the thrill seeking and not other sensation seeking domains.

With regard to correlation with substance use, the cross-sectional correlation between our sensation seeking measure at wave 3 (the first wave in which we assessed alcohol frequency and quantity) was 0.25 for binge drinking, 0.18 for alcohol frequency, and 0.24 for alcohol quantity, comparable with correlations between these substances (0.25, 0.25, 0.26 respectively) and the 20-item AISS in a sample of Australian high school students.(30) Regarding multivariate predictive validity, longitudinal studies of adolescents’ media use have found this sensation seeking measure to be a robust predictor of smoking(31–35) and alcohol use,(36, 37) independent of a number of other confounding social, parenting and personality influences.

Statistical Analysis

Means for sensation seeking were compared across demographic subgroups using multivariate logistic regression; in these regressions, sensation seeking was log transformed to reduce left skewness. Signal detection theory was used as the basis for the analysis of predictive validity.(38, 39) Receiver operating characteristics (ROC) curves were used to examine the effectiveness of sensation seeking as a predictor of substance use onset. In medicine, ROC curves are commonly used to study the ability of laboratory tests to discriminate between disease and non-disease(40) or, for screening/targeting, to discriminate between those at risk for or not at risk for a disease.(41) Key constructs for evaluating a screening test for substance use are shown in Table 1, based on the relationship between the targeting test results and whether or not the person goes on to use substances. True positives (TP) are substance users correctly identified by the test; false positives (FP) are persons identified as at risk who do not go on to use substances; false negatives (FN) are persons not identified by the test who go on to use substances; and true negatives (TN) are persons the test correctly identifies as not at-risk for substance use. The sensitivity of the test is a measure of how well the test correctly identifies true positives, and the false positive rate is a measure of how often the test mistakenly identifies someone as at risk.

Because sensation seeking is a continuous variable, there are a number of cut-off points one could use to target “high” sensation seekers. The ROC curve is a plot of the sensitivity of the test against the false positive rate for each possible cut-off point (see Figure 1 as an example). It illustrates the trade off between more inclusive cut points--which tend to maximize sensitivity but also result in high false positive rates, and more restrictive cut points--which minimize false positive rates but may result in low sensitivity. The area under the ROC curve (AROC) is a measure of the predictive validity of the test; the perfect test has one cut point, a sensitivity of 100%, a false positive rate of 0% and an AROC of 1. A test with no predictive validity has an AROC of 0.5 and a linear ROC curve that parallels the diagonal. ROC curves were fit after logistic regression using STATA 9 (College Station, TX) and compared for demographic subgroups using nonparametric methods.(42) In each regression, the baseline sample consisted of non substance users (232 baseline binge drinkers were removed from the binge drinker predictive model, and 34 baseline established smokers were removed from the established smoker predictive model).

Sensation Seeking

Table 2 shows the distribution for sensation seeking among the 6521 adolescents who completed the baseline assessment (one subject who responded to only one item was omitted), overall and by demographics.

At baseline, the scale values ranged from 4 to 16, with a mean 8.0 (std dev 2.5). The distribution was skewed left, such that one reached the median value at 7.0, where a cut-off level would identify 52% of the population as at risk. Cut off levels of 9.0 and 11.0 would respectively identify 25% or 9% of the population as at risk. With regard to mean sensation seeking scores by demographics, multivariate analysis (with log-transformed sensation seeking) found that baseline sensation seeking means were significantly lower by 2% (95% confidence interval 0.3%, 4%) in Hispanics and by 6% (5%, 7%) in females compared with males. Compared with 10 year olds, sensation seeking was higher by 8% (5%, 10%), 14% (12%, 17%), and 19% (17%, 22%) in 12, 13, and 14 years of age, respectively. These differences are apparent when one examines the distributions in Table 2. For example, whereas a sensation seeking score of 7.0 was the 60^th percentile at 10 years, it was only the 33^rd percentile at 14 years. There were no other statistically significant comparisons of means.

Sensation seeking as a predictor of binge drinking

Of 5834 participants with one or more follow up assessments of substance use, 5634 reported no binge drinking at baseline, of whom 717 (12.7%) reported binge drinking at one or more follow up interviews. Figure 1 shows the relation between level of sensation seeking at baseline, and sensitivity/false positive rate for predicting binge drinking in subsequent surveys. The area under the ROC curve was 0.71 (95% confidence interval 0.69, 0.73), suggesting that sensation seeking at baseline was a moderately good predictor of binge drinking in the future. At the relatively restrictive cut-off point employed by PTI researchers (one standard deviation above the mean), one would identify 16% of the population for the intervention, and this targeted sample would include about one-third of future binge drinkers (sensitivity = 31%) and have a false positive rate of 12%. The median split cut-off point employed by SENTAR researchers would identify half the population for intervention, capturing 77% of future binge drinkers at the cost of a 46% false positive rate. At higher cut-off levels, there was only a small gain in sensitivity at the expense of large increases of the false positive rate.

Table 3 compares the predictive validity of sensation seeking for binge drinking across demographic variables. There were no differences by gender. Sensation seeking was a significantly worse predictor of binge drinking among Black vs. other adolescents (AROC 0.65 and 0.72 respectively), but not among Hispanics. Sensation seeking was also a poorer predictor at extremes of the age bracket.

Sensation seeking as a predictor of established smoking

Of 5834 participants with one or more follow up assessments of substance use, 5802 were not established smokers at baseline, of whom 144 (2.5%) reported established smoking at one or more follow up interviews. Thus, established smoking was a much less common substance use outcome than binge drinking. The area under the ROC curve was 0.80 (95% confidence interval 0.76, 0.83), indicating that sensation seeking at baseline was a better predictor of established smoking than binge drinking in this sample, since the 95% confidence intervals did not overlap. At the relatively restrictive cut-off point employed by PTI researchers, one would identify 16% of the population for the intervention, and this targeted sample would include about half of future established smokers (sensitivity = 53%) and have a false positive rate of 14%. The median split cut-off point employed by SENTAR researchers would identify half the population for intervention, capture 92% of future established smokers, with a higher false positive rate of 50%. At higher cut-off levels, there was only a small gain in sensitivity at the expense of large increases of the false positive rate.

Table 4 compares the predictive validity of sensation seeking for established smoking across demographic variables. There were no differences by gender or age. Sensation seeking was a significantly better predictor among Hispanic vs. other adolescents (AROC 0.87 and 0.79 respectively), and predictive validity was no worse among Blacks compared with other adolescents.