Self-control interventions for children under age 10 for improving self-control and delinquency and problem behaviors

1.1 BACKGROUND

It can be stated with certainty that Gottfredson and Hirschi's general theory of crime stands as one of criminology's most important theories. Developed largely in response to parental socialization efforts involving child monitoring, recognition of child deviant behavior, and punishment of such deviant behavior, the theorists isolate the individual characteristic of self-control as the key correlate of antisocial, delinquent, and criminal behavior. According to Gottfredson and Hirschi, self-control is comprised of six inter-related characteristics including: (1) impulsivity and inability to delay gratification, (2) lack of persistence, tenacity, or diligence, (3) partaking in novelty or risk-seeking activities, (4) little value of intellectual ability, (5) self-centeredness, and (6) volatile temper. These characteristics are believed to come together for individuals with low self-control.

Since its inception, the theory has generated a significant amount of theoretical criticism, commentary, especially with respect to its key independent variable of self-control (Grasmick et al., 1993; Piquero et al., 2000; Tittle et al., 2004; Goode, 2008), and summary statements about the empirical knowledge base identify self-control as an important, but not sole correlate of varied antisocial activity (Pratt & Cullen, 2000). At the same time, much less attention has been paid to the malleability of self-control.

There is significant variation in how scholars interpret Gottfredson and Hirschi's stance on whether self-control is absolutely or relatively stable once established by late childhood/early adolescence. Some criminologists have interpreted Gottfredson and Hirschi to mean that self-control is resistant to any change, once established. Our reading, which we believe is consistent with Gottfredson and Hirschi, is such that self-control appears malleable during the first 10/12 years of life, but after this point, while self-control tends to improve with age as socialization continues to occur, it is largely unresponsive to any external intervention effort. Thus, although absolute levels of self-control may change within persons (increasing rather than decreasing), relative rankings between persons will remain constant over the life course. As they (1990, pp.107–108) note: “Combining little or no movement from high self-control to low self-control with the fact that socialization continues to occur throughout life produces the conclusion that the proportion of the population in the potential offender pool should tend to decline as cohorts age…Even the most active offenders burn out with time…Put another way, the low self-control group continues over time to exhibit low self-control. Its size, however, declines.” Elsewhere (1990, p. 177), they point out that “…individual differences in self-control are established early in life (before differences in criminal behavior, however the state defines it, are possible) and are reasonably stable thereafter.”

The existing research on the stability of self-control tends to suggest that it is not absolutely stable within persons (once established by ages 10/12) and that it tends to change (increase) with age (Arneklev et al., 1998; Turner & Piquero, 2002; Hay & Forrest, 2006; Mitchell & MacKenzie, 2006; Winfree et al., 2006), but remains relatively impervious to alterations by the criminal justice system after adolescence and in adulthood (Mitchell & MacKenzie, 2006). Although these findings are consistent with the general theory of crime, interpreting and integrating these findings within the context of the theory has not come easy because Gottfredson and Hirschi have not devoted much attention to policy issues. This has been an unfortunate consequence because discussions of theory and policy must be closely intertwined as good theory should lead to good policy and good policy is guided by sound theory. Of course, this is not to suggest that the theorists have not devoted any attention to policy.

In their strongest policy statement, Hirschi and Gottfredson (2001, p. 93) downplay any potential effectiveness of the criminal justice system: “Self-control theory leads to the conclusion that the formal criminal justice system can play only a minor role in the prevention and control of crime. Because potential offenders do not consider the long-term consequences of their acts, modification of these consequences will have little effect on their behavior. Because criminal acts are so quickly and easily accomplished, they are only rarely directly observed by agents of the criminal justice system. As a result, even large increases in the number of such agents would have minimal effect on the rates of most crimes”. Instead, the theorists are quick to point out the things that do not work and instead point to the few things they think will be effective, mainly to the socializing agents that are responsible for child-rearing.

More specifically, they (Hirschi & Gottfredson, 1995; Hirschi & Gottfredson, 2001, pp. 93–94) advance the following eight recommendations for crime control policy:

Do not attempt to control crime by incapacitating adults; this is so because by the time offenders are identified and incarcerated in adulthood, they have already finished the brunt of their criminal activity;

One of these policy proscriptions in particular (#7) points to the possibility that efforts aimed at children and young adolescents may improve self-control and have the added benefit of preventing delinquency and problem behaviors. In fact, there exists a fairly large stock of programmatic efforts aimed at improving self-control among children (up through age 10), but this line of research has not been integrated into the discussion of Gottfredson and Hirschi's theory, either by criminologists or the theorists themselves. Currently, there is no summary statement, similar to Pratt and Cullen's (2000) statement regarding the effect of self-control on antisocial activity, about the extent to which these programs are effective. Such a ‘taking-stock’ summary seems critical at this stage of the theory's life-course.

1.2 OBJECTIVES

There has been much attention paid in both criminology and psychology with respect to the importance of self-control in regulating antisocial, delinquent, and criminal behavior over the life course. Given the importance of self-control, there have been several programmatic efforts designed to improve self-control among children and adolescents. In an effort to build the knowledge base in this area, this study asks two critical questions: (1) What is the effect of self-control improvement programs on self-control up to age 10 among children and adolescents?; and (2) What is the effect of these programs on reducing delinquency and problem behaviors. Examining both self-control and delinquency and problem behavior outcomes would provide a comprehensive review that identifies a large number of studies and will likely evince a sounder conclusion and inform policy proscription for the general theory of crime. This meta-analysis, then, focuses on two inter-related outcomes: (1) What are the effects of self-control improvement programs up to age 10 for improving self-control among children/adolescents (self-control as the dependent variable)?; and (2) What are the effects of self-control improvement programs on delinquency and problem behavior outcomes (delinquency and problem behaviors as the dependent variable).

2.1 CRITERIA FOR INCLUSION AND EXCLUSION OF STUDIES IN THE REVIEW

Studies that investigated the effects of self-control improvement programs on child behavior problems such as conduct problems, antisocial behavior and delinquency were included. Following the more general systematic (e.g., Campbell) reviews, studies were only included if they had a randomized controlled trial design with post-test measures of self-control and/or child behavior problems for the experimental and control participants. The preliminary eligibility criteria are as follows:

Types of Studies: The study must have used a randomized controlled experimental design. ¹ The decision to only include studies that had a randomized controlled experimental design was made in order eliminate potentially spurious explanations as to the success of such programs since random assignment in theory rules out potential unmeasured confounds prior to the intervention between program participants. In addition, the quality and research designs of quasi-experimental designs vary greatly, and most experts caution against combining the effects of experimental and quasi-experimental designs (see Piquero et al., 2009);

2.2 SEARCH STRATEGY FOR IDENTIFICATION OF RELEVANT STUDIES

Several strategies were used to perform an exhaustive search for literature fitting the eligibility criteria:

A keyword search using the following keywords was performed: “Self-control” or “self control;” or “impulsivity” and “childhood” or “preschool” or “school” and/or “delinquency” or “conduct disorder” or “antisocial behavior” or “aggression” or “physical aggression” or “behavior problems”

Several strategies were used to obtain full-text versions of the studies found through the searches of the various abstract databases. First, we attempted to obtain full-text versions from the electronic journals available through several university library systems. When electronic versions were not available, we used print versions of journals available at the library. If the journals were not available at the university libraries, we used the Interlibrary Loan System (ILL) to try to obtain the printed version from the libraries of other institutions. In the case where these methods failed, we then made attempts to contact the author(s) of the article and/or the agency that funded the research to try to obtain a copy of the full-text version of the study.

2.3 DETAILS OF STUDY CODING CATEGORIES

All eligible studies were coded (see protocol in Appendix C) on a variety of criteria such as reference information (title, authors, publication year, etc.); nature of description of selection of sample, outcomes, etc.; nature and description of control group; unit of analysis; sample size; a description of the self-control improvement intervention; reports of statistical significance (if any); and effect sizes (if any). One investigator independently coded each eligible study. Further, we attempted to assess the quality of the studies in terms of research design, sample bias, and attrition bias. ²

2.4 CRITERIA FOR DETERMINATION OF INDEPENDENT FINDINGS

It is the case that most outcome studies rely on multiple measures, but there is disagreement as to how this issue should be handled with some researchers opting to use only one outcome source over another for reasons such as teacher ratings are likely to be less biased than parent reports and systematic “unbiased” observer ratings may be more accurate than teacher ratings (Farrington & Welsh, 2003). Other meta-analyses have averaged the effect sizes (ESs) across outcome measures and outcome sources when generating an individual effect size for each study (McCart et al., 2006). Still, others have noted that this method may lead to the loss of important information and create some difficulty when interpreting the overall effect (Casey & Berman, 1985).

In light of the apparent controversy over which method is more appropriate, we adopted a method of compromise and one that has been used in prior Campbell reviews and meta-analyses (for example, see Piquero et al., 2009), and report a series of effect sizes by outcome measure (e.g., self-control and delinquency and problem behaviors) and outcome source (e.g., parent report, teacher report, direct observation, self-report, and/or clinical report). Further, if a study included more than one treatment condition, then only the treatment condition that used a self-control improvement program was used to generate the relevant ESs. In addition, in the case where multiple control groups exist, then only the outcomes for the no-treatment control group (or wait-list control group) were used to calculate the ES. Similarly, when multiple treatment groups existed where each treated group received a self-control improvement program, then only one ES was calculated for the study by averaging the mean and standard deviation across the treatment groups and then comparing this one pooled mean and standard deviation to that of the control group in order to generate the ES for the study. As one more method for ensuring the statistical independence of findings, we calculated only one single ES for one particular sample in the event that multiple studies reported findings from the same sample of treated youth.

2.5 ANALYTIC PROCEDURES

We rely on Cohen's (1988) d for determining the effect sizes for this meta-analysis. The main source of information for calculating Cohen's d was the standardized mean difference, but in situations where means and standard deviations were not provided t-values, f-values, p-values, partial r etc. was used to calculate the effect sizes (see Lipsey & Wilson 2001 for the relevant formulas). Hedges and Olkin (1985) recommend calculating an unbiased ES that accounts for the discrepancy between the sample ES and the population ES. These authors also suggest that an ES of a small sample study does not have as much “impact” on the overall ES as does an ES calculated from a large sample study. As such, they recommend using inverse variance weights when performing a meta-analysis. Therefore, we used the Hedges and Olkin adjustment and inverse variance weights when determining the ESs in the analysis.

All of the meta-analysis results were estimated using Lipsey and Wilson's SPSS macros relying on a random effects model using inverse variance weight methods. It is also our general assumption that the individual ESs were not likely to be homogeneous so we estimated a series of moderator analyses using Lipsey and Wilson's SPSS analog to the ANOVA macro. Some of the relevant variables that are included in the moderator analyses include publication year, country of publication, small/large samples, published/not published, treatment type, treatment modality (group/individual), treatment duration, and treatment setting. The last stage of the analysis presents the results from a weighted least squares regression model (estimated with inverse variance weights and random effects) where the variables mentioned above are included as predictors of the ES. Publication bias is also evaluated using traditional methods including a comparison of the mean effect size for published/unpublished studies and an investigation of publication bias with a funnel plot and associated test statistics (e.g., Kendall's test and Egger's test) estimated with the ‘metafunnel’ macro available in Stata.

3.1 LITERATURE SEARCH

As discussed in the previous section, we used several mechanisms when attempting to locate studies that may be relevant for inclusion. Following an initial identification of over 5,000 hits, we sorted through the titles and abstracts and removed any that were inconsistent with the inclusion criteria. This process reduced the number of potentially relevant studies to 247 studies. These 247 studies were then electronically downloaded, copied from the library, or requested via Interlibrary Loan (ILL). The complete list of the 247 studies is displayed in Table 1. Upon receiving the documents, each study was thoroughly reviewed and final coding decisions were made as to whether the study conformed to each of the inclusion criteria. A description of the reason for ultimately deciding not to include a particular study is presented in Table 1 (see appendix) along with several study specific descriptive information (e.g., author(s), date of study, sample size, targeted age group, etc.).

The practice of displaying and describing the excluded studies allows readers to determine for themselves the findings of those excluded studies compared with those included. For the most part, studies were excluded because of the lack random assignment, targeting mostly older adolescents, focused on mentally and/or physically handicapped children, or did not contain any relevant self-control and/or behavioral outcome measures/data. Thus, the final coding decisions left 34 studies that met each inclusion criteria as outlined previously and were used in the analysis that follows. These 34 studies generated 43 self-control ESs and 28 delinquency and problem behavior ESs.

3.2 CHARACTERISTICS OF STUDIES INCLUDED IN META-ANALYSIS

Table 2 (see appendix) presents a series of descriptive statistics characterizing the 34 included studies. Nearly two-thirds of the studies were from published data (61.8%) and the overwhelming majority were performed in the United States (91.2%). Most studies drew their samples from high-risk/low income populations (64.7%) and most were based on mostly male (55.9%) and white (67.6%) samples. Less than twenty percent reported attrition problems as measured by losing at least 15% of their original sample for a variety of reasons such as moving, unable to locate, etc. Overall, a substantial majority were group-based interventions (67.6%) and were operated in a school setting (79.4%). While most could be broadly characterized as social skills development programs (32.4%), a considerable number of the interventions focused on cognitive coping strategies (26.5%), video tape training/role playing (20.6%), immediate/delayed rewards clinical interventions (11.8%), and relaxation training (8.8%).

The studies spanned over four decades with the earliest study published in 1975 and the most recent published in 2008 (M=1989.65; SD=10.37). While there were some studies with relatively small samples as well as those with considerably large samples, on average the studies included approximately 129 children/adolescents (SD=165.57). On average, the children/adolescents were 6.23 years of age at the time of the intervention (SD=2.03) with a range of 3 to 10 years old.

Overall, nearly every study included a measure of self-control ³ and data relevant for calculating a standardized mean effect size (94.1%), and more than half of the studies provided data for generating a standardized mean effect size for a delinquency and problem behavior-related outcome. ⁴ And although both self-control and delinquency and problem behavior outcomes were assessed, a number of different outcome sources were included overall such as parent-, teacher-, direct observer-, self-, and clinical reports.

3.3 TYPES OF INTERVENTIONS

Considering the variability of the self-control improvement interventions, it is important to discuss some examples of the broad categories of intervention type. The most recognizable of the social skills development programs are studies of the Conduct Problems Prevention Research Group (CPPRGa, 1999) and Tremblay et al.'s (1991) Montreal Youth Study. The social skills development intervention in the CPPRG study is called Fast Track and uses a “unified model of prevention” where a number of integrated intervention programs are applied such as: curriculum, parent groups, child social skills training groups, parent-child sharing time, home visiting, child peer pairing, and academic tutoring. The program involves lessons addressing four domains of skills: (1) skills for emotional understanding and communication; (2) friendship skills; (3) self-control skills; and (4) social problem solving skills (CPPRGa, 1999, p. 635). Comparatively, Tremblay et al.'s intervention also involved multiple program components, but one of these core competencies involved social skills training and was administered within small groups of prosocial peers. Another key component of Tremblay et al.'s intervention was self-control improvement sessions developed around themes such as “look and listen,” “following rules,” “what to do when I am angry,” “what to do when they do not want to play with me,” and “how to react to teasing” (p. 154).

Jackson and Calhoun's (1982) study was classified as a cognitive coping strategies intervention, which involved “cognitive self-instructional training where children are taught to covertly emit verbalizations that will cue or guide their non-verbal behavior” (Jackson & Calhoun, 1982, p. 7). Similarly, Reid and Borkowski's (1987) versions of cognitive coping strategies focuses on using psychoeducational tasks where an instructor verbalizes correct self-control statements such as “find out what I am supposed to do,” “consider all answers,” “stop and think,” “mark my answer,” and “check my answer” while performing various tasks, and then has the child repeat these steps and verbalize these statements while performing similar tasks.

Toner et al. (1978) is an example of a study classified as a video tape training/role playing intervention. Here, the children are sat in front of a television and told by the instructor: “Here is my television. The boy you will see on TV has been told not to touch the toys that are in front of him. Watch closely.” (p. 285). During the course of watching the video, the boy in the video would either do things appropriately or be resistant to commands at times. At each response time (whether appropriate or resistant), the subject was asked whether the boy's response in the video was correct. If the subject replied with an affirmative response, then the video continued. Following the video tape training, the subject was also left alone for a period of time and their behavior and self-control was observed. Baggerly (1999) is another example of a video tape training/role playing intervention where didactic lectures, experiential activities (e.g., role playing), and viewing videos of child-centered play sessions were used with the intention of improving the children/adolescents' self-control. The children /adolescents in this particular study received the training for 35 minutes twice a week for five weeks and then once a week for the remaining five weeks.

The immediate/delayed rewards clinical interventions can best be characterized by Mischel and Baker (1975). This type of intervention took place in an experimental room where the room was divided by a wooden barrier where there were battery operated toys and interesting games on one side of the barrier and a table and chair along with a desk bell on the other side of the barrier. The experimenter showed the child how to use the desk bell and informed them that once they left the room, the child could ring the bell and the experimenter would return. Upon returning (after the child rang the bell) the experimenter would reward the child and play a “game” with them. After a series of further instructions, the experimenter would then continue this interaction and assess the child's ability to “transform the reward objects that face him during the delay period in ways that either permit or prevent effective delay of gratification” (p. 259).

The final classification of the intervention type in the included studies was relaxation training interventions. Lakes and Hoyt's (2004) study was the most identifiable of this intervention type and involved periods of meditation where the children/adolescents were instructed to clear their minds of thoughts and worries while performing deep breathing techniques. Following this exercise, the subjects were then instructed to ask him/herself three questions intended to promote self-monitoring: 1) Where am I?; 2) What am I doing?; and 3) What should I be doing? After answering these questions the subjects were told to correct their thoughts and behavior if they were not consistent with the expectations of the particular situation. Ultimately, the instructors encouraged these exercises while emphasizing that the subject (not anyone else) is responsible for regulating their own behavior (p. 289).

3.4 QUALITY ASSESSMENT

It is important to note several methods for assessing the “quality” of the included studies. One of the most agreed upon determinants of study quality is the study's research design. Because all of the included studies were based on a randomized controlled experiment to evaluate the effectiveness of self-control improvement interventions, it is reasonable to assume that these studies are of high quality. Yet, it was rare for any of the studies to provide any detail on whether the randomization process was compromised or if attrition had any differential effects for the experimental/control groups. Thus, it is possible that some group imbalances might have arisen. Having said this, only 15% of the studies included in this analysis either reported or demonstrated significant attrition problems, which would lead us to assume that the overwhelming majority of these studies were of high quality in this regard. Nevertheless, we still included a measure of whether there was substantial attrition reported in a particular study as a control measure (e.g., potential moderator) in the analysis that follows. ⁵ Finally, most of the studies did not provide any information on whether the experimental/control groups were treated similarly throughout the course of the intervention by those who administered the intervention.

3.5 CALCULATING STANDARDIZED MEAN DIFFERENCE EFFECT SIZES (ESS)

Self-control and delinquency and problem behavior ESs were computed by calculating Cohen's d from the available information, i.e., predominantly means and standard deviations. Although, Cohen's d is the most common effect size statistic, the standardized mean difference is upwardly biased when based on small sample sizes and as such the unbiased effect size estimate that corrects for this was used (Hedges & Olkin, 1985; Lipsey & Wilson, 2001). As per Hedges and Olkin, the individual ESs were adjusted according to their samples size to correct for this bias. Tables 3 and 4 (see appendix) display the results of the individual unbiased ESs and corresponding confidence intervals calculated for each study based on the self-control and delinquency and problem behavior outcomes by outcome source (parent-, teacher-, direct observer-, self-, clinical report), respectively. ⁶

As seen in Table 3, the majority of the ESs were positive suggesting that self-control improvement programs have beneficial results insofar as improving a child/adolescent's self-control at post-test assessment. Further, a number of the ESs across outcome source were significant (as indicated by the confidence interval for the ES not including zero) providing evidence that the positive effects appear real, particularly for the clinical self-control ESs. Turning toward the effect of self-control improvement programs on delinquency and problem behaviors (Table 4), the majority of the individual mean ES are again positive suggesting that interventions such as these not only promote self-control improvement but also reduce delinquency and problem behaviors at post-test assessment. ⁷ Figures 1-8 (Appendix B) provide forest plots organized from smallest to largest ES for each outcome type and outcome source. Forest plots displaying the mean ESs by outcome type (regardless of outcome source) are provided in Figures 9 and 10 (Appendix B) in order to show how the total ESs for self-control and delinquency and problem behaviors are distributed. ⁸

Hedges and Olkin (1985) suggest using the inverse variance weight to weight each individual ES by the sample size of the treated and control groups when calculating an overall standardized mean difference effect size. Thus, after applying the inverse variance weight to the individual ESs by outcome type and outcome source, the mean ESs from a series of random effects models (using Lipsey and Wilson's 2001 SPSS macros) are presented in Table 5 in the appendix. ⁹ Importantly, with the exception of the self-control ES based on parent reports (p = .20) all of the ESs are positive and significant, and ranged from having a small effect (0.28) to having a rather substantial moderate effect (0.61), suggesting that self-control improvement programs are by and large successful at improving self-control regardless of the post-test assessment source. Comparatively, the results are not as robust for the delinquency and problem behavior ESs. ¹⁰ Nevertheless, all of the ESs are positive and the teacher reports results suggest that self-control improvement programs have a significant, small-to-moderate effect on improving self-control at post-test assessment.

3.6 HOMOGENEITY TESTS

It is safe to assume that the individual study ESs are unlikely to be homogenous, i.e., all of the individual study ESs do not come from the same population. Thus, it is necessary to estimate the Q statistic as a method for examining whether this homogeneity assumption was violated. ¹¹ The results (Table 5) suggest that all of the mean ESs by outcome type and outcome source (with the exception of the self-control direct observer report ES) were in fact heterogeneous; therefore, we explored potential moderating variables that may help explain some of the heterogeneity in the ESs.

3.7 MODERATOR ANALYSES

We selected a number of potential moderators based on previous meta-analyses and also chose several other factors that may be particularly relevant including: whether the study was published (yes/no) or performed (yes/no) in the United States, targeted a high-risk/low income population (yes/no), the gender (mostly male: yes/no) and race composition (mostly white: yes/no), whether there were any noted attrition problems (yes/no), the treatment modality (group: yes/no) and setting (school: yes/no), and the type of intervention (social skills development, cognitive coping strategies, video tape training/role playing, immediate/delayed rewards clinical intervention, or relaxation training). We included four continuous measures as moderators: the year of publication, the total sample size, age at the start of the intervention, and the duration of the intervention (in weeks). ¹² For all categorical variables, moderator analyses were conducted using Lipsey and Wilson's (2001) SPSS macros for the analog to the ANOVA (with random effects), whereas the moderator analyses for the continuous variables were investigated by analyzing the correlations (calculated by taking the square root of R²) between the moderators and the ESs.

The results of the analog to the ANOVA analyses (with random effects) investigating possible moderators of the self-control ESs are presented in Table 6 while the results for possible moderators of the delinquency and problem behavior ESs are displayed in Table 7. ¹³ Virtually all of the self-control ESs for all of the categorical moderator variable groupings were significant and appeared to be consistent, for the most part, by outcome source (parent-, teacher-, direct observer-, self-, and clinical report) (Table 6, see appendix).

Overall, the overwhelming majority of the ESs were positive suggesting that regardless of how the ES was contrasted the effect of self-control improvement programs seem to benefit the children/adolescents insofar as improving their self-control by post-test assessment. Some examples of the significant categorical moderators included: gender composition, where females evinced higher self-control gains (Q _between= 3.25; df= 1; p = .07; tau∧2= 0.27, se= 0.17), race composition (Q _between= 2.14; df= 1; p = .14); tau∧2= 0.30, se= 0.19), and attrition problems (Q _between= 3.25; df= 1; p= .07; tau∧2= 0.27, se= 0.17) for the self-control parent report ES and published versus not published (Q _between= 3.46; df= 1; p= .06; tau∧2=0.08, se= 0.04) for the self-control teacher report ES.

Turning toward the analog to the ANOVA (with random effects) results for the possible categorical moderators of the delinquency and problem behavior ESs, it appears that most ESs are positive and significant suggesting that self-control improvement programs can also benefit children/adolescents in terms of reducing their delinquency and problem behavior by post-test assessment. An example of the significant categorical moderators for the delinquency and problem behavior ES included: gender composition (Q _between=25.43; df= 1; p< .001; tau∧2= 0.01, se= 0.01) for the delinquency and problem behavior teacher report ES.

Following these categorical moderator estimations, correlations were computed for the possible continuous moderator variables of the ESs using Lipsey and Wilson's SPSS macros. The results for the self-control ESs and the delinquency and problem behavior ESs by outcome source are presented in Tables 8 and 9, respectively. For the most part, the correlations for year of publication and the self-control ESs were negative indicating that older studies had larger ESs, although only one of the correlations was significant (self-control clinical report ES= −0.47, p<.05). The majority of the correlations between total sample size and the self-control ESs (Table 8) were negative as well, suggesting that smaller studies had larger ESs, yet only one of the correlations was significant (self-control self-report ES= -0.63, p<.05). Three of the five correlations between age at the time of the intervention and self-control effect size were significant, two indicating that studies with older children/adolescents had larger ESs (self-control teacher report= 0.36, p<.10; self-control direct observer report= 0.85, p<.01) and one correlation suggesting that studies with younger children/adolescents had larger ESs (self-control clinical report= -0.36, p<.10). Only two of the correlations between the duration of the intervention and self-control ES were significant: shorter interventions had larger ESs (self-control teacher report= -0.57, p<.01) and longer interventions had larger ESs (self-control self-report= 0.78, p<.01).

Comparatively, two of the three correlations between year of publication and the delinquency and problem behavior ESs (Table 9) were positive and significant, indicating that more recent studies had larger ESs (delinquency and problem behavior parent report= 0.46, p<.05; delinquency and problem behavior direct observer report= 0.64, p<.05). All of the correlations between total sample size and the delinquency and problem behavior ESs were negative (e.g., smaller studies had larger ESs), although only one of these correlations was significant (delinquency and problem behavior direct observer report= −0.88, p<.01). Only one of the correlations was significant between age at the time of intervention and the delinquency and problem behavior ESs (delinquency and problem behavior teacher report= 0.37, p<.10) and the duration of the intervention and the delinquency and problem behavior ESs (delinquency and problem behavior direct observer report= -0.72, p<.01), suggesting that studies with older children/adolescents and those that were shorter in time span had larger ESs than those with younger children/adolescents and operated over a longer period of time.

3.8 META-ANALYSIS WEIGHTED LEAST SQUARES REGRESSIONS

The moderator analyses pointed toward some significant moderators of the ESs for self-control and delinquency and problem behavior by outcome source. It remains important to examine the nature of the moderators within a multivariate context to determine whether any of the moderators could be considered significant predictors of the variation in the ESs across the studies net of the effect of the other possible moderators. To examine this, a series of meta-analysis weighted least squares regression models (with random effects estimated using the maximum likelihood function available in Lipsey and Wilson's 2001 SPSS macro) were estimated by outcome type and outcome source. Based on sample size constraints and with attention to key moderators described above, each of the regressions was estimated with the same demographic moderators of gender and race composition, as well as age at the time of the intervention. Any additional potentially significant predictors were introduced in a stepwise fashion when degrees of freedom were available. The final models presented in the appendix in Tables 10 (predicting self-control ESs) and 11 (predicting delinquency and problem behavior ESs) only display the results of the moderators that were significant for predicting the variation in the individual study ESs.

The regression results predicting the self-control ESs (Table 10) identified several key predictors across the various outcome sources such as: gender, where females evinced higher self-control (b= −0.77, se= 0.15, p<.01) and race (b= −0.51, se= 0.12, p<.01) composition for parent report self-control ES; race composition (b= 0.40, se= 0.10, p<.01), treatment modality (b= 0.22, se= 0.17, p<.20), and interventions that utilized cognitive coping strategies (b= 0.83, se= 0.12, p<.01) for teacher report self-control ES; published/not published (b= 0.18, se= 0.14, p<.20), gender (b= −0.33, se= 0.22, p<.10) and race (b= −0.25, se= 0.21, p<.20) composition, interventions that utilized video tape training/role playing (b= 0.64, se= 0.25, p<.01), and age at the time of intervention (b= −0.06, se= 0.04, p<.10) for clinical report self-control ES.

For the most part, the same significant predictors of the delinquency and problem behavior ESs (Table 11) were similar to those that were significant for predicting self-control ESs. Significant predictors included: gender (b= -0.38, se= 0.28, p<.20) and race (b= −0.39, se= 0.22, p<.10) composition, year of publication (b= 0.04, se= 0.02, p<.10), and age at the time of the intervention (b= 0.20, se= 0.09, p<.05) for parent report delinquency and problem behavior ES; and gender composition (b= −0.34, se= 0.10, p<.01), interventions that used cognitive coping strategies (b= 0.19, se= 0.13, p<.20), and age at the time of the intervention (b= 0.05, se= 0.03, p<.10) for teacher report delinquency and problem behavior ES.

3.9 PUBLICATION BIAS ANALYSIS

While disagreement exists as to whether meta-analyses should include unpublished studies (Dush et al., 1989; Eppley et al., 1989; McLeod & Weisz, 2004), we opted to err on the side of inclusion. This permitted the inclusion of 13 additional studies, all dissertations that were from unpublished data. Although we have already presented comparisons between the self-control and delinquency and problem behavior ESs by outcome source for published/not published studies in the analog to the ANOVA tests previously (when possible), we still explored the potential for publication bias. There are a number of methods that may be used to assess publication bias both statistically and visually, and we opted to estimate the possible presence of publication bias through the use of a funnel plot (which is available as a macro in Stata 10.0-“metafunnel”) and calculating relevant test statistics (e.g., Kendall's and Egger's tests, which can be estimated using the “metabias” macro in Stata 10.0) (Borenstein, 2005; Sterne & Harbord, 2004). Figures 11 and 12 (Appendix C) present the funnel plot results for the total number of ES for self-control and delinquency and problem behavior (regardless of outcome source), respectively. According to the funnel plots estimated by outcome type and outcome source (where the larger studies are plotted at the top and the smaller studies are plotted at the bottom) and the relevant Kendall and Egger's tests, for the most part, there does not appear to be much evidence of significant publication bias. Although in some cases, the smaller studies seem to be clustering to the right (suggesting the possibility of publication bias), only among the self-control clinical report ESs (Egger's test: t= 1.95, p=0.08) and the total self-control ESs (Kendall's test: z= 1.96, p=0.05; Egger's test: 3.27, p<.01) is the publication bias significant.