Using the Fast Track randomized prevention trial to test the early-starter model of the development of serious conduct problems

Abstract

The Fast Track prevention trial was used to test hypotheses from the Early-Starter Model of the development of chronic conduct problems. We randomly assigned 891 high-risk first-grade boys and girls (51% African American) to receive the long-term Fast Track prevention or not. After 4 years, outcomes were assessed through teacher ratings, parent ratings, peer nominations, and child self-report. Positive effects of assignment to intervention were evident in teacher and parent ratings of conduct problems, peer social preference scores, and association with deviant peers. Assessments of proximal goals of intervention (e.g., hostile attributional bias, problem-solving skill, harsh parental discipline, aggressive and prosocial behavior at home and school) collected after grade 3 were found to partially mediate these effects. The findings are interpreted as consistent with developmental theory.

Over the past decade, a near-consensus model has been forged of the development of chronic violence among a group of children called early starters (Conduct Problems Prevention Research Group [CPPRG], 1992; Moffitt, 1993; Patterson, DeBaryshe, & Ramsey, 1989). Two assertions form the core of this model. First, it is asserted that early-starting children can be identified reliably by age 5; second, it is asserted that the growth and course of antisocial development for early starters still depends on life events and self-development (e.g., parenting, academic success, social cognition) that occur during the elementary school years. The latter of these assertions is still controversial because it is based on correlational evidence that is subject to alternate interpretation, namely, that child characteristics might cause these life events rather than these events having a causal impact on child outcomes. Prevention science, with its use of randomized designs, affords a unique opportunity to test this assertion of the developmental model in an experiment (Coie et al., 1993). The Fast Track prevention trial (CPPRG, 1992), based on this early-starter model, has demonstrated positive effects on proximal targets that are hypothesized to play a causal role in antisocial development (CPPRG, 1999, 2002a). The goal of the current study is to examine whether intervention-caused change in these proximal targets mediates intervention-caused antisocial outcomes among the children in the Fast Track experiment.

A Developmental Model of Early-Starting Antisocial Children

Patterson (1986) and Moffitt (1993) both identified a group of children who exhibit conduct problems as early as age 5, tend to grow in aggressive behavior across time, and persist in their antisocial behavior throughout adolescence and young adulthood. These early starters can be contrasted with another group of antisocial children whose problematic behavior is limited to adolescence. Early starters represent the more problematic group because their behavior is resistant to intervention (Moffitt, 1993), and they may account for over half of all crimes in a cohort (Wolfgang, Figlio, & Sellin, 1972).

The origin of conduct problems in the early-starting group is still controversial and may include a variety of factors, including neuropsychological deficits and temperamental characteristics (Moffitt, 1993), inconsistent and harsh parenting (Dodge, Bates, & Pettit, 1990; Patterson, Reid, & Dishion, 1992), and a context of family and neighborhood economic disadvantage (McLoyd, 1990). Whatever the origin, the second major assertion of the proposed model is that the trajectory and course of development for an early-starting 5-year-old child is still plastic, and perhaps half of early-starting children do not persist in their problematic behavior over time (Coie & Dodge, 1998; Loeber et al., 1993).

Antisocial growth versus attenuation depends on subsequent life experiences in at least three domains. First, family experiences that are characterized by supportive parenting (Pettit, Bates, & Dodge, 1997) and exclude harsh and inconsistent discipline (Snyder & Patterson, 1995) have been found to be associated with diminution in aggressive behavior over time. Second, a child's successful experiences at school, both academically and socially, can at least partially mitigate the risk associated with early-starting conduct problems (Coie & Dodge, 1998). Specifically, an early-starting child who is able to escape the stress, labeling, and deviant tracking that comes with placement into special education classes or peer social rejection is relatively likely to decrease his or her aggressive behavior (Coie, in press; Dodge et al., 2002). Finally, for many early-starting children, social cognitive development is characterized by hostile attributional biases (Dodge, Pettit, McClaskey, & Brown, 1986), social problem-solving deficits (Lochman, Lampron, & Rabiner, 1989), and a tendency to overestimate the positive consequences of aggressive behavior (Crick & Dodge, 1996). An early-starting child who can escape these maladaptive social information-processing patterns is more likely to decrease antisocial behavior across time (Dodge, Pettit, Bates, & Valente, 1995).

The developmental model proposes that, for early-starting children, these three domains of life experience (adaptive parenting, school success, and adaptive social cognition) are at least partially exogenous factors that play a crucial role in development because they lead directly to reduced problem behavior at home and school. Furthermore, success at these experiences, as well as reduced problem behavior at home and school, is hypothesized to lead to later positive peer relations, reduced involvement with deviant peer groups, and more success at home and school.

The Role of Prevention Science in Testing Developmental Theory

All of the assertions made here have been supported by multiple correlational studies using longitudinal designs with either community samples or samples of high-risk youth, in the spirit of the emerging science of developmental psychopathology (reviewed by Coie & Dodge, 1998). Proximal changes in parenting, school success, and social cognition have been linked by correlation to distal changes in antisocial outcomes. The major caveat in this line of research has been that, because the findings are correlational, the studies are open to alternate interpretations of the correlations. Most obviously, it is plausible that these factors are not at all exogenous and self-selection into these life experiences renders the role of these experiences as epiphenomenal by-products, rather than active agents, of antisocial development. Even though early levels of antisocial behavior had been controlled statistically, no amount of statistical control can completely eliminate the possibility that an unknown third variable (a child characteristic) causes both the life experiences and growth in antisocial behavior.

The prevention experiment, with random assignment of children to treatment or control conditions, represents the most scientifically rigorous means to test the hypothesis that externally imposed intervention can alter the trajectory of antisocial development and its correlated life experiences and features (Cichetti & Toth, 1992; Maggs & Schulenberg, 2001). Cicchetti and Toth (1999) summarize the successful use of the prevention experiment in developmental science in topics as diverse as attention deficit hyperactivity disorder and the parenting relationship between foster parents and their infants. The prevention experiment can move us closer to (although never fully) testing causal links in the chain of development, such as the link between improved parenting and reduced oppositional or aggressive behavior at home and the link between improved adaptive social cognition and more favorable peer relations (Coie et al., 1993). Although these links remain correlated outcomes of intervention and are thus not explicitly tested by the prevention experiment, the prevention trial does enable us to examine whether proximal changes and distal outcomes co-occur. Especially when intervention-imposed co-occurring changes are separated by time, one can move closer to concluding that the earlier change catalyzes the later one. Thus, the prevention experiment is a valuable tool in testing developmental theory (Kellam & Rebok, 1992; Loeber & Farrington, 1997; Robins, 1992).

The assertions of the developmental model have formed the basis for numerous prevention experiments that have been implemented in the past decade and have formed a major component of an emergent prevention science (reviewed by Tremblay, LeMarquand, & Vitaro, 1999). Several programs tested by randomized prevention trials have included components directed toward parent training, child social skills training, and the child's classroom and peer group ecology. These programs include the Metropolitan Area Child Study (Guerra, Eron, Huesmann, Tolan, & Van Acker, 1997), the Linking the Interests of Families and Teachers program (LIFT; Reid, Eddy, Fetrow, & Stoolmiller, 1999), the Baltimore Developmental Epidemiology Project (Kellam, Rebok, Ialongo, & Mayer, 1994), and the Seattle Social Development Project (Hawkins, Catalano, Kosterman, Abbott, & Hill, 1999).

The intervention programs in a subset of these experiments identify high-risk, early-starting conduct-problem children and then deliver multimodal interventions to them, based on a developmental model of conduct problems. For example, Tremblay et al. (1995) identified disruptive kindergarten boys (the high-risk group) and randomly assigned them to intervention or control groups. The intervention included both parent-training and child social skills training components, reflecting its derivation from developmental models of antisocial development. August, Realmuto, Hektner, and Bloomquist (2001) screened kindergarten children to identify high-risk aggressive-disruptive children and then delivered the Early Risers multicomponent program to a randomly assigned half of them. Early Risers included social skills training, behavior management, and parent training. Lochman, Coie, Underwood, and Terry (1993) identified aggressive-rejected 9-year-old children and randomly assigned half to receive a multicomponent social relations intervention. Bierman, Miller, and Stabb (1987) identified socially rejected 7-year-old boys and randomly assigned a portion to receive a combined social skills training and adult-prohibition program.

The Fast Track Prevention Trial

The Fast Track prevention trial (CPPRG, 1992, 1999, 2002b) has been called “one of the best examples of the new generation of multimodal prevention experiments for conduct problems” (Tremblay et al., 1999, p. 547). It is perhaps the most comprehensive of these programs, involving parent training, home visits for family support, social skills training, social–cognitive skills training, academic tutoring, and classroom teacher consultation delivered across a 10-year period. “The Fast Track Program is possibly the maximum a society will be willing to pay to help (early-starting conduct-problem) children” (Tremblay et al., 1999, p. 547).

At each of four sites for each of three annual cohorts, kindergarten children were screened by teacher and parent ratings of conduct problems to identify a high-risk group of early starters, who were then randomly assigned to receive the Fast Track program or followed as controls. The program was based on the developmental model, with interventions directed toward the three domains of parenting, school success, and social cognitions. Results after the first 3 years of intervention have indicated that the group assigned to receive intervention has fared more favorably than the control group in measures that assess each of the three proximal targets of intervention (parent use of physical punishment and parenting behavior change, placement into special education, and hostile attributional bias and social problem-solving skill), as well as measures of aggressive or disruptive behavior at home and in school (CPPRG, 2002a). Furthermore, these intervention effects have been found to hold across demographic groups, without significant moderation (CPPRG, 2002c).

What is not yet clear from these findings is whether the intervention effects on the proximal targets of intervention significantly mediate intervention effects on more distal outcomes of aggressive behavior. Because the developmental model asserts this mediation, it was the focus of the current study. In order to disentangle effects across time, the already published findings from the end of grade 3 assessment (CPPRG, 2002a) were used to identify nine possible grade 3 mediators (i.e., parenting behavior change, parental use of physical punishment, child hostile attribution bias, aggressive social problem solving, placement into special education, aggressive behavior at home, aggressive behavior at school, social competence change at school, and prosocial behavior change at school) of new intervention effects that were evaluated at the end of grade 4 for the current study.

Various methods have been applied to test developmental theory in the context of a prevention experiment, including latent growth curve analysis (Vitaro, Brendgen, & Tremblay, 2001), hierarchical linear modeling (Hawkins, Guo, Battin–Pearson, & Abbott, 2001), and structural equations modeling (Poulin, Dishion, & Burraston, 2001). In the current study, we chose to contrast structural equations models (SEM), following the schematic represented in Figure 1. Intervention-control status was conceptualized as an exogenous factor assigned randomly at the end of kindergarten. Four other exogenous design features were assigned or taken into account prior to the implementation of intervention (cohort, site, gender, and ethnicity), as were 25 preintervention covariates that were included only to improve the precision of postintervention measurement. Intervention status was hypothesized to have an ultimate effect on five critical antisocial outcomes at the end of grade 4 (aggressive behavior at home, social competence at school, academic competence at school, peer social preference, and association with substance-using peers). The developmental model posited an indirect effect on these outcomes, mediated through the intervention effects on the nine significant proximal targets of intervention measured at the end of grade 3.

Figure 1.

A conceptual model for the analysis of mediation effects.

Methods

Overview

This study was conducted simultaneously with identical measurement and procedures at four diverse geographical sites (Nashville, TN; Durham, NC; Seattle, WA; and rural PA). Schools within each site were selected as high risk on the basis of crime and poverty statistics of the neighborhoods that they served. Multistaged screening of all kindergarten children from all of the schools in three successive cohorts proceeded first with teacher ratings of disruptive behavior, followed by parent ratings of child behavior at home (Lochman & CPPRG, 1995). Combined teacher–parent scores identified children as high risk. These families were invited to participate in a longitudinal study of children's adjustment to school. Consenting families received parent and child interviews in the summer prior to grade 1, which served as preintervention assessments. Children were then assigned to receive the Fast Track intervention or not based on matching and then random assignment of schools within each site.

For families assigned to receive intervention, intervention staff members then approached parents to solicit their participation in a long-term program that included, in grade 1, 22 parent group meetings and children's skill-building group meetings, thrice-weekly tutoring, biweekly home visits, and consultation with teachers. The intervention continued through the end of grade 4 (the outcome point for current analyses) and beyond. In grade 2, the parent and child groups met every other week during the school year for 14 sessions. In grades 3 and 4, the groups met on a monthly basis for 9 sessions. After grade 1, tutoring and the other individualized support components of the program described below were offered only if individual children and families met a risk-based criterion for each component. For example, tutoring was continued if that child's reading performance was in the lower third of his or her school's grade level. Actual participation was high but variable (see CPPRG, 2002b).

For the current study, parent and child interviews, parent daily reports of child aggressive behavior, teacher assessments, and school record reviews were collected prior to random assignment and following grades 3 and 4. Classroom sociometric measures were obtained in the spring of grade 4. Data collected following grade 3 are considered here as mediating variables, and the data collected following grade 4 provide outcome variables.

Participants

Behaviorally disruptive kindergarten target children (n = 891) and their parents were identified using a multistage screening procedure. First, high-risk schools were identified in each of four different areas of the country, based on local reports of school serving a high proportion of children with behavior problems and likelihood of dropping out: Durham, NC, a small city with a predominantly African American population; Nashville, TN, a moderate-sized city with African American and European American families; Seattle, WA, a moderate-sized city with an ethnically diverse population; and central Pennsylvania, a rural area with a predominantly European American population. In the spring of 1991, 1992, and 1993, teachers rated the behavior problems of each of the kindergarten children in the 55 participating elementary schools using the 10-item Authority Acceptance Scale of the Teacher Observation of Classroom Adaptation—Revised (TOCA-R; Werthamer–Larsson, Kellam, & Wheeler, 1991), which describes aggressive and oppositional behaviors (i.e., fighting, teasing, disobedience). The parents of children who scored in the top 40% of the within-site sample were then contacted by telephone or in person and asked to rate the frequency of child behavior problems at home. The 24 items on this parent screen measure were drawn from the aggression scales of the Child Behavior Checklist (Achenbach, 1991), the Revised Problem Behavior Checklist (Quay & Peterson, 1987), and other items generated by the investigators (for further details, see Lochman & CPPRG, 1995). Children's total scores on the two screening measures (teacher and parent ratings of behavior problems) were averaged to yield a risk score. Children who scored in the top 10% within a site on this score (“high risk”) were invited to participate in a longitudinal study, which began with in-home interviews in the summer preceding grade 1 matriculation.

The mean age of the high-risk children was 6.5 years (SD = 0.48) at the time of identification, and they were approximately 4.5 years older at the time of outcome. Across all sites, the sample was 51% African American, 47% European American, and 2% other ethnicity (e.g., Pacific Islander and Hispanic), reflecting the ethnic diversity of the population at the four sites. Sixty-nine percent were boys; 58% of the high-risk children came from single-parent families; 29% of the parents were high school dropouts; and 35% of the families were in the lowest socioeconomic class as determined by Hollingshead scoring.

Assignment to intervention and control conditions

Because part of the intervention (described below) involved a school-based intervention, entire elementary schools (n = 55) were assigned to either the intervention or the control condition. Within each site, data on the demographics for each school were obtained (e.g. size, percentage of students who received free or reduced price lunch, ethnic composition, achievement scores), and the participating schools at each site were divided into matched sets. These sets were randomly assigned to intervention and control conditions. In the intervention condition, there were 445 children in 191 first-grade classrooms. In the control condition, there were 446 children in 210 first-grade classrooms. In September of the first-grade year, prevention staff visited parents of high-risk children assigned to the intervention condition and invited them to participate in the Fast Track program. The prevention program was described as an enrichment program to help children succeed in school, focusing on the goal of increasing social skills and reading skills rather than stigmatizing the children by focusing on their behavior problems. Transportation, child care for siblings and refreshments were offered to help parents overcome pragmatic obstacles to attendance. In addition, parents were paid for their time attending parent–child group sessions ($15/session for each 2-hr session), the rationale being that their attendance was valuable because they knew their child better than anyone and thus were key to making the program successful.

Regardless of the extent to which families agreed to engage in the intervention (e.g., full, partial, or nonparticipation), they were considered part of the intervention sample in all analyses reported here if the child was enrolled in a regular education placement in an intervention school on November 15th of grade 1. Identical criteria applied to the control group. Although some children later moved to other schools or were placed in special education classrooms, they were not dropped as members of the intervention (or control) groups at any point after November of the child's grade 1 year. Control families were recontacted at the end of grade 1 and each successive school year for reassessment but were not contacted during the school year and did not receive any Fast Track prevention services.

Data attrition

At the end of grade 3 or 4, at least one score was obtained for 96.6% of the original control group and 98.8% of the original intervention group. Missing data rates for each variable are listed in Table 1. None of the differences in attrition rates between the intervention group and the control group were statistically significant. Using the original screening score (the TOCA-R) as the dependent variable, contrasts were conducted separately for control and intervention groups between those children who continued to participate at all and those who did not, for each source of data on the children. None of the comparisons was statistically significant.

Table 1. Correlation matrix and input data for key variables.

		1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
1.	Treatmenta	1.00
2.	G4 Peer drug usea	−0.07	1.00
3.	G4 Oppositional/aggressive behavior	−0.10	0.00	1.00
4.	G4 Social preference	0.09	−0.02	−0.04	1.00
5.	G4 Social competence change	0.09	0.01	−0.03	0.10	1.00
6.	G4 Academic change	0.09	−0.04	−0.03	0.09	0.86	1.00
7.	G3 Hostile attributions	−0.07	0.12	−0.02	−0.06	−0.07	−0.06	1.00
8.	G3 Physical punishment	−0.10	0.08	0.04	−0.01	−0.05	−0.07	0.05	1.00
9.	G3 Oppositional/aggressive behavior	−0.08	0.03	0.66	−0.04	0.02	0.02	−0.05	0.07	1.00
10.	G3 Child behavior change	0.11	−0.01	−0.24	0.03	0.03	0.01	0.07	−0.04	−0.30	1.00
11.	G3 Parent behavior change	0.09	0.00	−0.20	−0.03	0.02	0.00	0.14	0.00	−0.24	0.71	1.00
12.	G3 Authority acceptance problems	−0.08	0.10	0.08	−0.30	0.01	−0.04	0.14	0.14	0.04	0.03	0.05	1.00
13.	G3 Response generation	0.04	−0.08	0.09	0.12	−0.02	0.02	−0.07	−0.06	0.04	0.00	−0.01	−0.02	1.00
14.	G3 Change in prosocial behavior	0.13	0.00	−0.06	0.17	0.03	0.04	−0.09	−0.02	−0.01	0.07	0.04	−0.34	0.03	1.00
15.	Special education placementa	−0.07	0.04	0.05	−0.09	0.00	−0.02	0.10	0.01	0.07	0.00	0.02	0.09	−0.10	−0.05	1.00
Mean		0.50	0.46	0.22	−0.56	3.59	3.64	0.62	0.12	0.21	1.17	1.53	1.79	0.73	0.99	0.44
Standard deviation		0.50	0.64	0.18	1.06	0.71	0.77	0.25	0.15	0.18	0.99	0.89	1.07	0.14	0.89	0.50
Missing data (%)		0.0	12.8	10.8	37.6	11.6	11.6	11.4	10.4	10.4	10.3	10.4	10.2	11.6	10.5	15.2

Note: Maximum N = 891. Thresholds for significance vary with the fraction of missing information. In general, coefficients of .08 or greater are significant (p < .05).

^aCategorical; treated as continuous in the matrix.

Intervention procedures

Indicated interventions

Parents and children in the high-risk intervention condition were offered interventions in three domains: parent groups with home visiting, academic tutoring, and social skills training. Parent and child group interventions were conducted during a weekly, 2-hr “enrichment program” held at the school building on Saturday or a weekday evening. During the first 60 min in grade 1 of this enrichment program, high-risk target children met in groups of five or six in “friendship groups” led by educational coordinators (ECs) and coleaders. Discussions, modeling stories and films, and role-plays were used to illustrate and promote skill concepts; cooperative activities provided opportunities for skill practice and performance feedback. Sessions focused on reviewing and practicing skills in emotional understanding and communication, friendship building, self-control, and social problem solving (see Bierman, Greenberg, & CPPRG, 1996, for more details).

At the same time, parents met in a group led by family coordinators (FCs) and coleaders to discuss parenting strategies that would support child school adjustment and improve child behavior. Primary content areas of the parent group curriculum included: (a) establishing a positive family–school relationship and supporting child adjustment to school, (b) building parental self-control, (c) promoting developmentally appropriate expectations for the child's behavior, and (d) improving parenting skills to enhance parent–child interaction and decrease child disruptive behavior. Session topics and training techniques (instruction, modeling, discussion, and role playing) were drawn from the parenting program developed by Forehand and McMahon (1981), with additional material from the programs developed by Webster–Stratton (1989) and Burgoyne, Hawkins, and Catalano (1991; see McMahon, Slough, & CPPRG, 1996, for more details about the parent-focused intervention components.).

Following the parent and child groups, parent–child pairs spent 30 min together each session, participating in positive cooperative activities and practicing positive parenting skills with staff support (the Parent–Child Sharing Time). Activities included games and crafts, joint reading, and activities that allowed parents to practice parenting skills presented in the parent group (see McMahon et al., 1996, for more details).

Group meetings were held weekly during grade 1 for 22 sessions, biweekly during grade 2 for 14 sessions, and monthly during grades 3 and 4 for 9 sessions each year. In addition to the group meetings, individual support was provided to children and parents to help them generalize the skills presented in the group setting and to address individual needs. FCs conducted home visits with parents every other week, on average, and had weekly telephone contacts between group sessions. The home-visiting component provided an opportunity for Fast Track staff to (a) develop trusting relationships with the entire family system; (b) promote generalization of newly acquired parenting skills to the home; (c) promote parental support for the child's school adjustment; and (d) promote parent problem solving, coping, and goal setting as means of dealing with the many stressful life events (e.g., marital conflict, substance use, social isolation, and housing issues) that these families often experience. We used a problem-solving approach to such issues that was developed by Wasik, Bryant, and Lyons (1990) in home visiting with economically disadvantaged families and modified for the current program to emphasize empowerment through skill acquisition. The ultimate goals of this problem-solving approach were to foster parental empowerment and self-efficacy (Dunst, Trivette, & Deal, 1989) and decrease dependency on Fast Track staff.

ECs monitored child progress at school, and supervised paraprofessional tutors who conducted weekly “peer pairing” sessions with children in grades 1 and 2. In these latter sessions, high-risk children participated in 30-min play sessions with classroom peer partners (partners were rotated over the course of the year). These sessions were designed to promote the generalization of friendship skills to the school setting and to offer high-risk children opportunities to make friends with classroom peers. Academic tutoring, designed to promote reading skills, was also provided by the paraprofessional tutors. These tutors received 40 hr of training prior to intervention and regular supervision by ECs during intervention. The Wallach and Wallach (1976) tutoring program was used primarily in the first 2 years of the intervention. Designed for low-readiness children from disadvantaged backgrounds, this program emphasizes a phonics-based, mastery-oriented approach toward the development of initial reading skills. By using the same paraprofessionals to provide tutoring in reading skills and to direct the peer-pairing sessions, both of which occurred in the school setting, high-risk children were provided with a supportive adult whom they saw on a regular basis throughout the year. During grades 1 and 2, tutors worked with children three times each week for half-hour sessions during school hours (two for reading, one for peer pairing). In grade 1 the children were also tutored during the enrichment program.

Children and families received a standard level of these services (home visiting, tutoring, and peer pairing) in grade 1. In subsequent years, criterion-referenced assessments were used to adjust the dosage of these components to match the level of functioning of each family and child (CPPRG, 2002b).

Prevention staff were hired from local communities to match (as much as possible) the ethnic composition of high-risk children at each site. ECs tended to be former teachers, whereas FCs had advanced degrees in counseling or social work or had extensive experience in working with high-risk families. To enhance consistency across sites, we required staff to attend a 3-day cross-site workshop and to observe videotapes of prototypic administration of each session. Intervention fidelity was ensured by several procedures. First, all intervention components were manualized. Second, program developers conducted regular cross-site supervisory telephone calls to inform intervention staff about the goals and activities of upcoming sessions and to receive feedback about children's and parents' reactions to activities thus far. Third, at weekly staff meetings at each site, intervention staff practiced and prepared for upcoming session activities. Fourth, intervention staff members were observed by the clinical supervisor as they delivered intervention throughout the year, and the staff members were given specific feedback about their adherence to the program and their skills in delivering intervention.

Universal intervention

In addition to the indicated interventions, the children and their classmates received a series of universal interventions designed to promote a more competent and less aggressive social ecology. Classroom teachers were trained to deliver a new curriculum in social and emotional development that was developed specifically for Fast Track. The first two grades of this curriculum were revised from a curriculum used by Kusche and Greenberg (1994), and the third and fourth grade curricula were developed for Fast Track to be theoretically consistent with the earlier curricula. These curricula were delivered in most of the intervention schools (except the Durham site, where administrators disallowed it) from grade 1 through grade 4. Teachers implemented this classroom-level program throughout the year, teaching an average of two or three lessons per week (see Bierman et al., 1996, for more details). ECs provided support and consultation for teachers on the curricula and behavior problem issues and monitored the fidelity of implementation with weekly classroom visits and weekly teacher meetings.

Assessment procedures

Although comprehensive assessments were collected annually (see CPPRG, 2002a or a detailed description of all variables, including the 25 preintervention covariates), only the nine grade 3 mediator variables and the five grade 4 outcome variables are described here.

Grade 3 mediator variables

The primary goal of the end of grade 3 assessments was to determine the cumulative effect of 3 years of preventive intervention on proximal targets of intervention that had been hypothesized by the early-starter model to mediate longer-term conduct problem outcomes. Four sources of information were used: parents, teachers, school records, and the children themselves. Of 22 variables measured, 9 were found to yield significant effects of intervention (CPPRG, 2002a) and are included here.

During the summer following grade 3, two trained interviewers conducted assessment interviews with all parents and children. Interviewers were kept blind to the condition of the families they were assigned to interview, but in the course of the interview, family members sometimes mentioned Fast Track in ways that informed interviewers on this point. While one research assistant interviewed the primary caregiver (usually the mother), a second assistant interviewed the child in a separate room. Interviewers read the various measures to the primary caregiver or child and recorded their responses. During that interview, in addition to other measures, parents reported on child behavior problems exhibited during the previous 24-hr period (Parent Daily Report [PDR]; Chamberlain & Reid, 1987). Parents were then recontacted by phone on two occasions over the next 2 weeks to make additional daily reports of child behavior problems. Parents received $75 for participating in these 2- to 3-hr home interviews.

In the spring of grade 3, teachers completed the TOCA-R (Werthamer–Larsson et al., 1991) and the Teacher Ratings of Child Behavior Change (CPPRG, 1999; α = .94). Teachers received $10 per child for completing these measures. Data on special education diagnoses were collected from school records in the period just following the end of the academic year.

From these sources, nine variables were measured. The first variable came from parents' responses to six brief written vignettes of various child misbehaviors (i.e., hitting another child, noncompliance). Each parent was asked what she or he would do in response to this situation and responses were scored as using physical punishment or not. The kappa value on the interrater agreement was .93. Responses were averaged across the six vignettes to yield a harsh physical punishment score. Next, parents completed an 11-item scale (Ratings of Parent Change; CPPRG, 1999) describing their own parent rating of parenting behavior change over the past year (α = .86).

The Social Problem-Solving measure (Dodge et al., 1990) was designed to assess the child's ability to generate appropriate solutions to common social problems. Each child was shown a series of eight drawings and read vignettes depicting peer entry or peer conflict problems. He or she was asked what the story character could do to solve the problem and was prompted to provide three different solutions to each problem if possible. Responses were coded as prosocial/competent or aggressive/inept by the interviewers, who were trained to a reliability of 80% agreement, using the supervisor's scores as the criterion, before they were allowed to collect data on their own. The percentage of competent response generation score (averaged across stories) was computed.

The Home Interview with Child (Dodge et al., 1990) assessed hostile attributional biases. Each child was shown a series of eight drawings and read vignettes depicting either unsuccessful peer entry (i.e., being ignored or rebuffed) or minor harm under conditions of ambiguous intent (i.e., being bumped or pushed). For each incident, the child was asked why he or she thought the negative event occurred and what he or she would do about the other kids who were involved. Child interpretations of ambiguous negative events on the HIWC were coded as hostile, nonhostile, or “don't know.” As with the problem-solving measure, interviewers were trained to a reliability of 80% before collecting data. The percentage of hostile attribution bias was scored.

The TOCA-R Authority Acceptance scale was used as a measure of classroom aggressive behavior (10 items on aggression, disruption, and disobedience of adults assessed on a 6-point scale, α = .94). The Teacher Rating of Child Prosocial Behavior Change instrument (CPPRG, 1999) assessed change in prosocial behavior over the course of the year (eight items using a 7-point scale, α = .94).

The Parent Ratings of the Child Behavior Change instrument (CPPRG, 1999) over the course of the past year were collected during the in-home interview to yield a measure of child conduct problem improvement (10 items using a 7-point scale, α = .89). Parents' ratings on the 15 items of the PDR (Chamberlain & Reid, 1987; α = .81) were collected by telephone on each of 4 separate days and then averaged to provide an assessment of child aggressive and oppositional behavior at home.

School records and teacher reports were examined to collect information on a special education diagnosis. A child received a score of 1 if either source revealed a special education diagnosis, time placed in special education, or an Individualized Education Plan at the end of grade 1, 2, or 3 or a score of 0 if none of these sources indicated such a diagnosis.

Grade 4 outcome variables

Although many more measures were collected in grade 4, five outcome variables are included here.

Peer sociometric nominations were collected in the late spring of grade 4. Children who received parental permission in all participating classes were administered sociometric measures during individual interviews. Each child nominated an unlimited number of classmates whom he or she “liked most” and “liked least.” Nominations were summed and standardized within classroom; then the disliking score was subtracted from the liking score and restandardized to yield a peer social preference score, following procedures used by Coie, Dodge, and Coppotelli (1982).

During the in-home interview, the child was asked a series of items constituting the Things Your Friends Have Done instrument. The 16 items tap domains of delinquent behavior and illicit drug use. The interviewer began by saying “I'm going to tell you about some things that kids your age sometimes do. I'd like you to think about the past year—from about this time last year through the school year. I'm going to read each behavior and ask you whether any of your friends have done that behavior during the past year.” For each of the 16 behavioral items, the child was asked to give a yes or no answer. If the child said “no” or “I don't know,” the interviewer scored the item as No (score of 0). If the child said “yes,” the interviewer asked whether most (score of 2) or just some (score of 1) of the child's friends have done the behavior and then scored the item accordingly. The three items that assessed illicit drug use were averaged and used here as a measure of association with deviant peers. Because of skew in the scores, a trichotomous score was computed as no, some, or most and analyzed through probit regression.

Parents' ratings on the 15 items of the PDR (Chamberlain & Reid, 1987; α = .81) were collected by telephone on each of 4 separate days and then were averaged to provide an assessment of child aggressive and oppositional behavior at home, just as in grade 3.

The Teacher Rating of Child Prosocial Behavior Change instrument (CPPRG, 1999) assessed grade 4 change in prosocial competence (8 items using a 7-point scale, α = .94) and grade 4 change in academic competence (2 items using a 7-point scale, α = .75) over the course of the year, just as in grade 3.

Results

Overview

Each of the five grade 4 outcomes was examined separately for possible mediation using structural equation modeling (SEM). The nine hypothesized grade 3 mediators were taken in two groups. First, the five grade 3 variables representing proximal targets of intervention (parenting behavior change, harsh physical punishment, child's competent response generation, child's hostile attribution bias, and special education placement) were modeled as mediating the intervention effects on the outcome and the overall indirect effect was evaluated. Second, the four grade 3 variables reflecting child aggressive behavior (classroom aggressive behavior, conduct problem improvement, aggressive and oppositional behavior at home, and change in prosocial behavior) were similarly modeled. Finally, given the established patterns, separate models were estimated using each mediator individually for each outcome to determine the specifics underlying the multiple-mediator tests. Table 1 summarizes the input variables and their intercorrelations.

This table indicates the significant effect of intervention on each of the five grade 4 outcome variables. These effects were analyzed more formally, as described below. This table also reflects the bivariate correlations that correspond to the significant effects of intervention on each of the nine grade 3 variables, which were reported in a previous article (CPPRG, 2002a). Finally, this table indicates the bivariate correlations between the nine grade 3 hypothesized mediator variables and the five grade 4 outcome variables. Thirteen of the 45 correlations were significant (p < .05), as summarized below.

Analysis plan

A schematic representing the general form of analysis appears in Figure 1. Intervention status, seven effect-coded values for design variables (gender, race, site, and cohort), and 25 mean-centered continuous Year 1 covariates were exogenous variables and allowed to covary freely. Each of these variables was modeled as predicting one or more mediators and the outcome. For models with multiple mediators, disturbance terms for the mediators were allowed to covary freely. This strategy resulted in saturated models; parameter estimates rather than model fit were the goal of the analysis.

For the four continuous grade 4 outcomes, models were estimated using maximum likelihood estimation in PROC CALIS in SAS v8.2 (SAS Institute, 2001). For the peer substance use outcome, models were estimated in Mplus v2.01 (Muthén & Muthén, 2001), using its mean- and variance-adjusted weighted least squares estimator for categorical data. For this variable, both grade 4 peer substance use and grade 3 lifetime special education placement were treated as categorical variables in a probit regression model. Ideally, special education placement would have been treated as a categorical variable in all models. However, PROC CALIS does not have that facility, and the advantages of using CALIS were determined to outweigh this consideration.

Indirect effects were calculated as the product of the relevant paths (from treatment status to mediator and from mediator to outcome). Standard errors for these effects were calculated using the formula derived by Sobel (1982). For models with multiple mediators, the estimate of interest was the summed indirect effect across the mediators; standard error estimates were calculated based on matrix equations derived by Sobel, using the multivariate delta method (see also Bollen, 1987).

Effects are reported in standardized coefficients. Because the intervention status variable is evenly divided, the standard deviation of the dummy-coded value is 0.5; regression coefficients for continuous outcomes can be doubled to obtain the predicted difference in the outcome in standard deviations as an effect size. For probit regression coefficients, the predictor is standardized.

Missing data

Missing data were imputed using PROC MI in SAS v.8.2. Ten imputations were generated, which is ample for the degree of missing information reflected in the analyses. The imputation data set included dummy-coded values for treatment status, gender, race (African American vs. non-African American), site, and cohort (none of which were missing); the 25 continuous Year 1 covariates; the 9 Year 4 mediators; the 4 continuous Year 5 outcomes; and the 5 ordinal component item responses for peer substance use. Imputation was conducted separately for the treatment and control groups. PROC MI uses Markov Chain Monte Carlo estimation for continuous variables. For categorical variables, generated values were rounded to the nearest legitimate response; for continuous variables the range of generated values was truncated to match the possible legitimate score distribution. Parameter estimates and standard errors were generated from the analyses of imputed data using PROC MIANALYZE in SAS v.8.2 and NORM v.2.03 (Schafer, 1999), which use standard formulae to combine results. These estimates and standard errors were used as the basis for calculating indirect effects and associated standard errors. Because degrees of freedom for tests of parameter estimates based on multiple imputation are not simple functions of sample size, but rather depend on the between- and within-imputation variance estimates, they are omitted from this report to minimize confusion.

Aggressive and oppositional behavior at home

Grade 4 outcomes

Analysis of covariance (conducted via SEM) showed a significant total effect of intervention status on parent-reported aggressive and oppositional behavior at home, controlling for design factors and the continuous covariates (β = −0.077, SE = 0.031, p < .02). This finding reflects a mean difference between intervention and control groups of 0.03 units, which is 0.15 SD. The intervention group received significantly lower scores for aggressive and oppositional behavior than did the control group.

Grade 3 mediation

Analyses of the relations between hypothesized intervention-target mediators and the outcome revealed significant effects of grade 3 parenting behavior change and grade 3 special education placement on grade 4 aggressive and oppositional behavior. The mediation model incorporating the five intervention-target variables showed a significant indirect effect via the set (β = −0.018, SE = 0.007, p < .02). Table 2 summarizes the results of the single-mediator models exploring the effects via the different mediators. Grade 3 parenting behavior change significantly mediated the effect of intervention on grade 4 aggressive and oppositional behavior and accounted for 14% of the total effect of intervention.

Table 2. Estimates from single mediator models for oppositional and aggressive behavior.

	Direct Effects			Indirect Effect
	Treatment > Outcome	Treatment > Mediator	Meditor > Outcome	Treatment > Outcome
Hostile attributions	−0.076 (0.031)*	−0.078 (0.033)*	0.019 (0.032)	−0.001 (0.003)
Physical punishment	−0.073 (0.031)*	−0.083 (0.033)*	0.048 (0.033)	−0.004 (0.003)
Parent behavior change	−0.067 (0.031)*	0.091 (0.033)**	−0.116 (0.032)**	−0.010 (0.005)*
Response generation	−0.079 (0.031)*	0.060 (0.034)†	0.027 (0.031)	0.002 (0.002)
Special education	−0.072 (0.031)*	−0.072 (0.036)*	0.065 (0.032)*	−0.005 (0.003)
Oppositional/aggressive behavior	−0.046 (0.027)†	−0.059 (0.029)*	0.533 (0.030)**	−0.032 (0.016)*
Child behavior change	−0.063 (0.032)*	0.101 (0.036)**	−0.137 (0.032)**	−0.014 (0.006)*
Authority acceptance	−0.066 (0.031)*	−0.091 (0.032)**	0.116 (0.035)**	−0.011 (0.005)*
Change in prosocial behavior	−0.072 (0.31)*	0.138 (0.034)**	−0.038 (0.031)	−0.005 (0.005)

Note: N = 891. Cell contents are standardized regression coefficients with standard errors in parentheses.

^†p < .10.

^*p < .05

^**p < .01.

Analyses of the relations between the behavior variables and the outcome revealed significant effects of grade 3 aggressive and oppositional behavior, grade 3 conduct problem behavior improvement, and grade 3 teacher-rated aggressive behavior on grade 4 aggressive and oppositional behavior. The mediation model incorporating the four grade 3 behavior measures showed a significant indirect effect (β = −0.043, SE = 0.017, p < .01). In addition, three of the four grade 3 behavior measures showed significant mediation effects that accounted for between 14 and 41% of the total effect of intervention on grade 4 aggressive and oppositional behavior.

Social preference

Grade 4 outcomes

Analysis of covariance (conducted via SEM) showed a significant total effect of intervention status on sociometrically rated social preference scores, controlling for design factors and the continuous covariates (β = 0.103, SE = 0.042, p < .02). This reflects a mean difference between intervention and control groups of 0.22, which is 0.21 SD (given that the social preference scores are standardized scores, they are standardized within classroom on the entire participating sample; variances are slightly greater than 1 among the Fast Track target sample). The intervention group received significantly higher social preference scores than the control group.

Grade 3 mediation

The mediation model incorporating the five intervention-target variables showed no significant indirect effect via the set (β = 0.008, SE = 0.009, ns). Analyses of the relation between the four grade 3 behavior variables and the grade 4 social preference outcome revealed significant effects of grade 3 teacher-rated aggressive behavior and grade 3 change in prosocial behavior. The mediation model incorporating the four behavior variables measures revealed a significant mediation effect (β = 0.034, SE = 0.013, p < .009). Table 3 summarizes the results of the single-mediator models exploring the effects via the different mediators. Significant indirect effects were indicated for mediation by grade 3 teacher-reported authority acceptance problems and grade 3 change in prosocial behavior.

Table 3. Estimates from single mediator models for social preference.

	Direct Effects			Indirect Effect
	Treatment > Outcome	Treatment > Mediator	Meditor > Outcome	Treatment > Outcome
Oppositiona/aggressive behavior	0.101 (0.041)*	−0.059 (0.029)*	−0.045 (0.048)	0.003 (0.003)
Child behavior change	0.099 (0.042)*	0.101 (0.036)**	0.037 (0.047)	0.004 (0.005)
Authority acceptance	0.077 (0.041)†	−0.091 (0.032)**	−0.286 (0.041)**	0.026 (0.010)**
Change in prosocial behavior	0.084 (0.041)*	0.138 (0.034)**	0.138 (0.046)**	0.019 (0.008)*

Note: N = 891. Cell contents are standardized regression coefficients with standard errors in parentheses.

^†p < .10.

^*p < .05

^**p < .01.

Association with deviant peers

Grade 4 outcomes

Probit regression of association with deviant peers (three levels) on intervention status, the design factors, and the continuous covariates revealed a significant total effect of intervention (β = −0.093, SE = 0.044, p < .04). The model-implied probability of reports of some or most friends using one or more substances is .36 for the intervention group and .43 for the control group. (The equal-slopes model results in a corresponding difference for the criterion of “most friends.”)

Grade 3 mediation

Of the five proximal mediator variables, grade 3 hostile attributions had a significant main effect on the grade 4 peer substance use outcome (p < .05). The mediation model incorporating the five intervention-target variables indicated a marginally significant indirect effect via the set (β = −0.018, SE = 0.011, p < .09). The mediation model incorporating the four aggression measures did not indicate any significant mediation (β = −0.003, SE = 0.011, ns). Table 4 summarizes the results of the single-mediator models exploring the intervention effect via the proximal target variables. The mediation effect by child hostile attributions was marginally significant. This mediation effect reduced the direct effect of intervention on the peer substance use outcome to marginally significant and accounted for 10% of the total effect of intervention. None of the other intervention-target variables showed significant effects on peer substance use, nor were any of the indirect effects significant.

Table 4. Estimates from single mediator models for peer substance use.

	Direct Effects			Indirect Effect
	Treatment > Outcome	Treatment > Mediator	Medior > Outcome	Treatment > Outcome
Hostile attributions	−0.084 (0.045)†	−0.078 (0.034)*	0.120 (0.049)*	−0.009 (0.006)†
Physical punishment	−0.091 (0.045)*	−0.083 (0.036)*	0.035 (0.043)	−0.003 (0.004)
Parent behavior change	−0.090 (0.045)*	0.091 (0.034)**	−0.046 (0.046)	−0.004 (0.004)
Response generation	−0.090 (0.045)*	0.060 (0.035)†	−0.059 (0.044)	−0.004 (0.003)
Special education	−0.094 (0.045)*	−0.100 (0.051)*a	−0.016 (0.055)b	0.000 (0.005)

Note: N = 891. Cell contents are standardized regression coefficients (in first two columns) and probit coefficients (latter two columns), with standard errors in parentheses.

^aProbit coefficient.

^bUnstandardized; special education treated as dichotomy.

^†p < .10.

^*p < .05

^**p < .01.

Teacher-reported child behavior change

Grade 4 outcome

The two teacher-reported change variables, change in social competence and change in academic performance, are highly correlated (r = .86) and are treated together here. For both variables, an analysis of covariance (conducted via SEM) showed significant total effects of intervention status, controlling for design factors and the continuous covariates, social competence (β = 0.092, SE = 0.035, p < .01), and academic performance (β = 0.091, SE = 0.037, p < .02). These findings reflect mean differences between intervention and control groups of 0.13 and 0.14, respectively, which is 0.18 SD in each case.

Grade 3 mediation

The mediation models incorporating the five intervention-target variables sowed no significant indirect effects via the set: social competence β = 0.008, SE = 0.008, ns; academic β = 0.008, SE = 0.007, ns. The mediation model incorporating the four aggression measures also failed to show significant indirect effects: social competence β = 0.000, SE = 0.008, ns; academic β = 0.000, SE = 0.008, ns.

Discussion

The findings reported here are consistent with the developmental model of early-starting antisocial development, and they also suggest a refinement of the model in domain-specific ways that have not received sufficient previous attention. Mediation analyses supported the importance of parenting and social–cognitive factors in exacerbating antisocial development, whereas the experiment inherent in the prevention trial led to findings that question the short-term, cross-domain generalization of factors in development. The findings have implications for the design of preventive interventions, following in the traditions of prevention science (e.g., Cicchetti & Toth, 1999), but they also refine developmental theory. Thus, developmental science and prevention science inform each other (Cicchetti & Toth, 1992, 1999; Maggs & Schulenberg, 2001).

The Fast Track intervention had positive effects on outcomes at home, at school, and in the peer group that were evident after 4 years. Mediation analyses provided some support for the proposed model of how antisocial behavior develops, as well as clues as to how the Fast Track Program achieved its 4-year success. Four findings are essential to indicate mediation, and all four were tested and supported here. First, assignment to receive the Fast Track intervention beginning in first grade was found to have discernible and modest-magnitude positive effects on parent reports of aggressive behavior, peer social preference, association with deviant peers, and teacher ratings of social and academic competence improvements that were assessed at the end of fourth grade. These effects indicate the immediate effects of 4 years of intervention with high-risk children. Given the intractability of early-starting aggressive conduct problems (Moffitt, 1993) and the lack of success of other programs in producing long-lasting change in aggressive behavior (cf., Coie & Dodge, 1997), these findings are noteworthy.

Second, the first 3 years of intervention had positive effects on all three areas that had been targeted for intervention: reducing parents' harsh physical discipline practices at home, improving children's social and academic competence at school, and improving children's social cognitive development by reducing hostile attributional bias and improving social problem-solving skills. The intervention program was directed toward these three domains and yielded effects in all areas. As the “manipulation check” in an experiment, it is essential that the experimentally manipulated intervention achieve its proximal goals of altering children's functioning in areas that are hypothesized to play a causal role in antisocial development (Maggs & Schulenberg, 2001).

Third, the processes that had been targeted for intervention in the first three grades were found to predict children's outcomes in fourth grade. An array of past developmental studies had supported the correlation between parenting, peer relations, and social cognition in early years and child outcomes in later years (Coie & Dodge, 1997). The reported findings replicate those findings and extend them to the current context of an exclusively high-risk sample.

Fourth, the positive effects of intervention on child outcomes was statistically accounted for, partially, by intervening effects on proximal targets of intervention. This finding directly supports the developmental theory that guided the creation of the Fast Track intervention (CPPRG, 1992). The finding also suggests that in order to achieve positive outcomes, a program would do well to have an impact on the targeted domains. The findings have substantial external validity because of the diverse sample in which they were tested. The sample included boys and girls from both African American and European American backgrounds at four highly varied geographic sites that spanned rural and urban and northern and southern settings. Although a comprehensive understanding of antisocial development in each of these groups might well include group-specific cultural factors, the current findings suggest that factors of harsh parenting, peer social competence, and social cognitive development are common to all groups.

The pattern of significant mediation follows within-domain relations. That is, parenting behavior mediates the intervention effect on the child's aggressive behavior at home but not at school. Improvements in the child's social cognitions about peers mediate the intervention on deviant peer associations. Improvements in prosocial behavior in the school setting mediates classroom peer social preference. These relations are precisely as hypothesized by the developmental model.

The lack of mediation of the teacher-reported change (improvement) in fourth grade behavior is just as important as the findings already noted. This variable assessed the teacher's judgment of how much improvement the child made in socially competent behavior between the beginning of the fourth grade school year and the end of the year. The mediator variables that were tested were measured at the end of the third grade. The lack of mediation suggests that any effect of the mediator variables occurred prior to the teacher's judgment at the beginning of the fourth grade. The pattern suggests that the effect of intervention on the teacher ratings of improvement in fourth grade can be attributed to the ongoing intervention in fourth grade rather than any delayed effect of previous intervention. Thus, continued intervention seems to exert new effects on child outcomes.

The specific pattern of mediation that was found can be interpreted as support for domain-specificity in antisocial development. It was hypothesized that changes in children's social cognitions and/or parenting support in one domain would first have effects on child behavior and adjustment within that domain but would lead to spreading effects in other domains. These findings support the former part of the hypothesis to a greater degree than the latter part. It is quite possible that additional time is necessary for the spreading of effects across domains and that such spreading will be evidenced to a greater degree in later years. This possibility will be tested in future analyses as the children mature.

Several caveats must be noted. First, even though the intervention effects on outcomes and proximal targets have been produced by an experiment and thus conform to the highest standard of evidence in science, the mediation effects are still correlational and subject to alternate interpretation. It remains plausible that some unmeasured effect of intervention (e.g., a Hawthorne effect or an effect on optimism of children and parents) is statistically and causally responsible for effects on both proximal targets and outcomes of intervention. It is unlikely that any experiment could more directly manipulate the proximal target of intervention than the current one, however, future experiments might include additional intervention conditions in which one domain is targeted for intervention but not another domain. One would expect intervention effects on proximal goals and outcomes to differ across conditions in such a study in a way that could test the differential mediation of an intervention effect on outcomes across conditions.

Second, the time frame for the current study is the first 4 years of elementary school. It is plausible that different mediational effects would be found at younger or older ages. Likewise, as noted above, different mediational effects may emerge as these same children are followed into adolescence.

Third, although the findings reported here apply to a diverse sample, it is not clear that the intervention and mediation effects would apply to each subsample if tested independently. The statistical power and large sample size that are necessary to test the effects in the subgroups were not present here.

In sum, this study provides a model for how to use developmental theory, innovative prevention programming, and contemporary statistical methods for the use of missing data in mediation analysis to guide the testing of proposed models of children's development of psychopathology.