Long-Term Effects of Drug Prevention on Risky Sexual Behavior Among Young Adults

Participants 

To assess Project ALERT's long-term effects on risky sexual behavior, we used data from 1901 21-year-olds who were sexually active but not married at the time of data administration and had participated in an experimental trial of the two drug prevention programs that took place in 55 schools from 45 South Dakota communities. We focused on unmarried young adults because lack of condom use among married persons frequently reflects a decision to have children and thus is considerably less likely to constitute a measure of risk for them than for singles. We focused on sexually active young adults because we are interested in the intervention's effect on risky sexual behavior, not whether they engaged in any sex at all.

During program implementation, 631 of the 1901 young adults included in this analysis attended schools that received 14 lessons in middle school (core program); 499 attended schools that received an additional 10 lessons in high school (expanded program); and 771 attended control schools. Overall, 60% were female, 11% were nonwhite (largely Native American), and 70% lived with both biological parents as adolescents. Prior to the baseline survey in grade 7, 19% of this sample had grades of C or below, 6% had tried marijuana, 34% had tried tobacco, and 63% had tried alcohol.

Study design and school sample 

The evaluation design randomly assigned 45 school clusters (high schools and their feeder middle schools) from South Dakota to the two treatment conditions (core and expanded programs) and one control condition. In the middle school core program, students received 11 prevention lessons in grade 7 and 3 in grade 8. In the expanded program, students received 5 lessons in each of grades 9 and 10 in addition to the middle-school lessons. Students in the control condition received other prevention curricula already in place at their schools but were not exposed to the ALERT program in any grade. This design allowed us to assess whether adding high school lessons to the core middle-school program yielded additional prevention gains.

To maximize the generalizability of the results, we included communities that varied in size (urban, small town and rural) and geographic region of South Dakota (eastern, western, and central). Four school clusters were in cities with more than 50,000 residents; 12 were in towns of 5000 to 25,000 residents; and the remaining clusters were in rural communities. The full study design randomized 48 clusters. However, assignment of three clusters was restricted to the ALERT or ALERT Plus conditions. Those clusters were not included in this analysis. To enhance pretreatment equivalence across conditions, we used blocking by geographic region and community size and restricted assignment when randomizing school clusters to conditions [16]. Two districts (each with one high school) dropped out of the study after we completed randomization and were replaced by schools in a similar region of the state and with a similar ethnic composition.

Curriculum content 

The ALERT middle school curriculum seeks to motivate youth against using drugs and help them build the skills they need to translate that motivation into effective resistance behavior [16]. The lessons focus on three major goals each of which is linked to a specific theory of behavior change: (1) helping youth understand the benefits of nonuse, as well as the seriousness and salience of consequences associated with drug use [25]; (2) reducing barriers to effective drug resistance by building confidence in one's ability to resist pro-drug pressures [26]; and (3) building social norms against use and the skills for resisting pro-drug pressures [27]. The additional 9th and 10th grade lessons in the expanded ALERT curriculum have the same overall goals and theoretical underpinnings as the middle school program [28] but also take into account developmental changes during the high school years that affect opportunities and motivations to use drugs (e.g., more diverse friendship networks and dating opportunities, risks associated with drinking and driving, use of drugs to cope with anxiety, and other forms of emotional distress). Prior analyses have shown that both versions reduce drug use by altering beliefs about its consequences, changing drug use norms and building drug use resistance self-efficacy [28], [29].

Both the core and expanded versions of ALERT use highly interactive teaching techniques and include content that encourages generalization of the lessons to other risky behavior and explicitly links risky sexual behavior with drug use. Participants watch videos in which adolescents discuss potential problems from drug use, including sexual behavior that has negative consequences (core: one lesson; expanded version, two), and practice resistance skills in nondrug-use situations (one lesson, both versions). Hence, there are two plausible routes through which the ALERT programs might curb risky sex: (1) by reducing alcohol and drug use, both of which may be linked with impaired judgment, disinhibition, and increased likelihood of being with other peers who engage in risky behavior or in social contexts where sexual intercourse is likely; and (2) by directly altering beliefs about the sexual risks associated with drug use, as well as participants’ perceived ability to resist sexual pressure. We were able to test for mediation by alcohol and drug misuse but lacked appropriate data for testing mediation by program-induced changes in beliefs about risky sex or ability to resist sexual pressure.

Data collection, weighting, and imputation 

We used baseline covariates from the 1997 Wave 1 survey (administered in grade 7 classrooms just before program delivery), mediating data from the 2004 Wave 6 and 2006 Wave 7 surveys (administered via mail and the web approximately 6.5 and 8.5 years after baseline, when the average age of the respondents was 19.3 and 21.3 years) and outcome data from the 2006 Wave 7 survey. RAND's institutional review board approved all data collection procedures and instruments. Participants consented to data collection at each survey wave; parents provided consent at the beginning of the study. At Wave 1, we obtained a U.S. Department of Health and Human Services Certificate of Confidentiality that was later updated to include all subsequent waves of data collection.

Of 4689 respondents (1879 control, 2810 ALERT) who completed the Wave 1 survey, 57% each of the ALERT (n = 1603) and control (n = 1078) groups also completed the Wave 7 survey. Of these, 1901 met the study criteria of being sexually active and not married. Within this group of eligible young adults, there were differences on baseline risk factors between the ALERT and control participants. As shown in Table 1 (columns 1 and 2), eligible ALERT respondents were significantly less likely than control respondents to have received low grades and tried alcohol as seventh graders, were somewhat more likely to be white and somewhat less likely to have tried marijuana. Estimates of treatment effects were sensitive to these group differences.

Table 1. Weighted and unweighted characteristics of analytic sample by experimental condition

	Baseline (W1) characteristics	Eligible analytic sample (W7)
		Unweighted		Propensity score weighted
		Control N = 771	Treatment N = 1130	Control N = 771	Treatment N = 1130
		%	%	%	%
	Female	60.3	59.7	60.4	59.6
	White ethnicity	85.6	90.9	88.2	89.9
	Nuclear family	71.1	70.0	72.3	69.7
	Grades (C or below)	21.6	16.8a	19.3	17.5
	Tried marijuana	7.9	5.0	5.8	5.4
	Tried cigarettes	32.9	34.1	29.9	34.9
	Tried alcohol	66.4	61.2a	65.0	61.8

a p < .05 compared to control. Treatment includes both the 14-lesson (core) version and the 24-lesson (expanded) version of Project ALERT.

To remove differences among groups resulting from randomization or from differential attrition and eligibility, we used propensity-score weights [30] for participant outcomes from both the ALERT and control groups. To estimate the propensity score (the probability that an eligible respondent with a given set of baseline risk factors received the ALERT condition), we used a generalized boosting model (GBM) [31], which has proven effective in developing weights that balance the distribution of variables between experimental groups at baseline [31]. We included 51 variables measuring a broad range of risk factors from the grade 7 survey in the propensity score model (e.g., demographics, own, and family substance use and beliefs, etc.). After propensity-score weighting, the distributions of each of these 51 variables were nearly identical between the two groups; as shown in Table 1 (columns 3 and 4), there were no significant differences between treatment and control groups on the key baseline characteristics. However, to ensure our estimates were “doubly robust” to errors in our propensity-score weights [32], we also included the key individual-level variables as covariates in our model, along with two school-level variables that were higher in the control group (past week and past month marijuana use; higher by 1 and 2 percentage points, respectively).

To impute missing data for Wave 1 covariates (typically under 10%) plus the Wave 6 and 7 mediators (0% to 20%), we used a multivariate Gaussian distribution to approximate the joint distribution for the variables. To improve the accuracy of the imputed values, the imputation models for the baseline covariates included all the derived variable scales and demographic variables for Waves 1 and 2. We imputed Wave 6 and 7 mediators with a joint model for the baseline and follow-up variables that used the previously imputed values for the missing baseline data. Imputed values have been found to be robust to model specifications [33]. We synthesized analyses conducted on five imputed data sets, generated by the NORM and SAS Proc MI software [34], by averaging estimated coefficients and pooling within- and between-imputation variability to estimate standard errors [33]. The between imputation variance in our estimated treatment effects is very small, so that five imputed data sets are sufficient to provide precise estimates.

Outcome and mediating measures 

To capture risky sexual behavior, we used three dichotomous outcome measures—inconsistent condom use, sex with multiple partners, and having unprotected sex because of drug use. Each of these measures provides policy-relevant information for assessing progress in reducing risk for pregnancy, HIV, and/or other sexually transmitted diseases. Consistent with CDC practitioner guidelines [35], we defined inconsistent condom use as having reported use of condoms “usually, half of the time, sometimes, or never” during sexual intercourse in the last year and sex with multiple partners as having reported intercourse with two or more sexual partners in the last year. Drug-linked unprotected sex was defined as having reported engaging in unprotected sex because of alcohol or drug use one or more times in the last year.

To test our hypotheses about the mediation of treatment effects on risky sexual behavior at age 21, we examined alcohol and drug misuse at age 19 (Wave 6 of data collection, 2 years before the outcomes were measured) and at age 21 (Wave 7, the same period during which we measured sexual behavior). We reasoned that prior alcohol and drug misuse would mediate program effects on risky sex by increasing the likelihood of subsequently having deviant peer associations, being in risky situations, and continuing to misuse alcohol and drugs and that contemporaneous alcohol and drug misuse might also mediate program effects by enhancing disinhibition and impairing judgment. We measured alcohol misuse at age 21 with a 10-item alcohol audit score [36] that performs well in predicting DSM-IV alcohol diagnoses [37]; it included items tapping consumption (e.g., frequency, dosage), out-of-control alcohol use (e.g., unable to stop drinking), and alcohol-related problems (e.g., injuring others, blacking out, α = .85, range = 0–36). We measured drug misuse at age 21 with a seven-item drug dependence sum based on items chosen to meet DSM-IV criteria [11] that included out-of control drug use (e.g., using more than intended) and problems from using drugs (α =.85, range = 0–7). At age 19, the available alcohol and drug misuse measures focused on alcohol consequences (nine items) and marijuana consequences (eight items); these tapped past year experience of problems because of use (e.g., missed school or work, trouble at home, traffic accident, criticized by friend, αs =.83 and .87, range = 0–5 and 0–3.5, respectively).

Outcome and mediation analyses 

To measure effects of exposure to Project ALERT, we compared young adults who participated in either the 14- or 24-lesson version of the program with those who attended control schools. To evaluate the importance of program duration and gender, we subsequently added separate interaction tests for program length and gender to the base model. We estimated a logistic regression model for each outcome that included an indicator (0,1) variable for Project ALERT plus individual and school-level baseline covariates. We present treatment effects in the form of differences in the probability of an outcome, using the logistic regression models to predict the expected rate of risky sex as a result of substance use for the treatment and control conditions as outlined by Graubard and Korn [38]. To account for the original clustering of participants within schools, we used bias-reduced linearization methods designed to yield standard errors and statistical tests that are unbiased in the presence of intracluster correlation [39].

To assess the potential mediation of risky sexual behavior, we selected the outcomes showing treatment effects at age 21 and then identified, among our four preselected potential alcohol and drug use mediators, those that qualified for mediation, that is, both predicted the risky sex outcome and were predicted by ALERT. For qualifying mediators, we estimated the mediation effect by assessing the change in program coefficient (size and p-value) when the mediator was added to the regression model. We also conducted the Sobel [40] test for significance of mediation effects. We fit models that included each qualified mediator separately and models that included all qualified mediators.