Analyses were conducted using the scale sum scores from the TRF, normalized with square-root transformations (see Neter, Wasserman, & Kunter, 1989
). The effect of gender was statistically controlled in all analyses, and all variables were standardized within cohort. We also examined cohort by predictor interactions in a series of exploratory analyses, but no significant effects emerged. All interaction terms were calculated using mean-centered scores, as recommended by Jaccard, Turrisi, and Wan (1990)
In order to model the effects of interest fully, we adopted a strategy that involved both simultaneous and hierarchical regression analyses. The specific predictors included in each model were selected according to our hypotheses. However, we also conducted fully specified models (e.g., models including the relevant main effects and all possible interaction effects), and the overall pattern of results was nearly identical.
Correlations among the predictor variables are summarized in . As depicted, behavior problems were moderately stable from T1 to T4. Thus, associations between T1 behavior problems and T4 victimization may be confounded by the association between T4 behavior problems and T4 victimization. Accordingly, the prediction associated with T4 behavior problems was statistically controlled in all analyses.
Correlations Among Predictor Variables
Relations Between Behavior Problems and Peer Group Victimization
A multiple regression analysis was conducted predicting T4 victimization from gender and each of four T1 behavior-problem variables. The full model yielded a significant effect, F(5, 374) = 8.18, p ≤ .0001, R2 = .10. As depicted in , there were significant bivariate correlations between T4 victimization and T1 Externalizing, T1 Social Problems, and T1 Attention Problems. The correlation between T4 victimization and T1 Internalizing was marginally significant. In addition, there was a significant independent effect for T1 Social Problems, β = .185, sr2 = .017, p ≤ .01; and marginal independent effects for T1 Externalizing, β = .108, sr2 = .007, p ≤ .10, and T1 Internalizing, β = −.099, sr2 = .007, p ≤ .10 (where sr2 is the squared semipartial correlation coefficient, the percentage of variance in the outcome predicted independently by each parameter). The marginal negative parameter for T1 Internalizing is potentially noteworthy, but “suppressor” effects of this nature should be interpreted with great care.
Predictive and Concurrent Correlations Between Behavior Problems and Peer Group Victimization
A similar analysis was then conducted predicting T4 victimization from gender and the four T4 behavior problem variables. This model also yielded a significant overall effect, F(5, 366) = 15.20, p ≤ .0001, R2 = .172. Each T4 behavior problem score was concurrently correlated with T4 victimization (see ). Examination of the standardized regression parameters indicated that there were significant independent effects for T4 Social Problems, β = .286, sr2 = .033, p ≤ .0001, and T4 Externalizing, β = .127, sr2 = .008, p ≤ .0001, and a marginal effect for T4 Attention Problems, β = .134, sr2 = .008, p ≤ .075.
Finally, partial correlations were generated, predicting T4 victimization from each of the T1 behavior problem scores, with the corresponding T4 behavior problem score controlled (see ). These analyses yielded significant effects for Externalizing, Attention Problems, and Social Problems.
Social Preference as a Mediator
The mediational role of social preference (i.e., peer rejection–acceptance) was examined using procedures specified by Baron and Kenny (1986)
. According to these authors, the following criteria must be met to establish mediation (as modeled by regression analysis): (a) the mediator must be significantly associated with the outcome, (b) the predictor must be significantly associated with the outcome, (c) the mediator must account for variance in the outcome beyond the variance associated with predictor, and (d) entry of the mediator into the model should result in a reduction in variance accounted for by the predictor.
We examined each of these criteria in a multivariate hierarchical regression analysis in which the combined variance from the T1 behavior problem scores (with gender and T4 behavior problems statistically controlled) served as the predictor, the combined variance from the social preference scores at T2 and T3 served as the mediator, and T4 victimization served as the outcome. On the first step, we entered gender, T4 Externalizing, T4 Attention Problems, and T4 Social Problems simultaneously. On the second step, we entered T1 Externalizing, T1 Attention Problems, and T1 Social Problems simultaneously. On the third step, we entered the T2 and T3 social preference scores. On the fourth and final step, we entered the social preference score for T4 (in order to examine the mediations role of T2 and T3 social preference with concurrent social preference controlled). We did not consider the Internalizing scores in these analyses because the bivariate relation between T1 Internalizing and T4 victimization was not significant. As depicted in , T1 behavior problems (line 2, step 2) significantly incremented the prediction in T4 victimization associated with gender and T4 behavior problems (line 1, step 1). However, T1 behavior problems (line 2, step 3) did not significantly predict T4 victimization once T2 and T3 social preference were entered into the model. Moreover, T2 and T3 social preference (line 3, step 4) predicted variance in T4 victimization independent of T1 behavior problems and T4 social preference. Thus, each of the criteria for mediation specified by Baron and Kenny (1986)
Summary of Analyses of the Mediational Role of T2 and T3 Social Preference in the Prediction of T4 Victimization by T1 Behavior Problems
Dyadic Friendship as a Moderator
In order to examine the moderating role of friendship in the prediction of victimization, a separate hierarchical regression analysis was conducted for each of the behavior problem clusters. On the first step of each of these analyses, we entered the main effects for gender, the T4 behavior problem score, the T1 behavior problem score, and T1 total number of friends. On the second step, we entered the interaction between the T1 behavior problem score and T1 total number of friends. Significant T1 behavior problem Score × T1 Friendship interactions were conceptualized as indicators of moderation (Baron & Kenny, 1986
; Holcombe, 1997
). As shown in , there was a significant interaction term for T1 Externalizing × T1 friendship and a marginal term for T1 Attention Problems × T1 friendship.
Summary of Analyses of the Moderating Role of T1 Friendship in the Prediction of T4 Victimization from T1 Behavior Problems
Analyses were then conducted to clarify the nature of these interactions, guided by the recommendations of Aiken and West (1991)
. First, T4 victimization was predicted from gender, T4 Externalizing, and T1 Externalizing with T1 friendship fixed at low (one standard deviation below the mean), medium (the mean), and high (one standard deviation above the mean) levels. As the fixed value of friendship increased, the slope of the relationship between T1 Externalizing and T4 victimization declined from β
= .233, p
≤ .005 (low friendship) to β
= .046, ns
(mean friendship), and β
= −.141, ns
Next, T4 victimization was predicted from gender, T4 Attention Problems, and T1 Attention Problems with friendship fixed at low, medium, and high levels (as described above). As the fixed value of friendship increased, there was a marginal decline in the slope of the relationship between T1 Attention Problems and T4 victimization from β = .176, p ≤ .05 (low) to β = .082, ns (mean) and β = −.014, ns (high).
Because rejection–acceptance by the peer group as a whole and dyadic friendship are hypothesized to predict unique variance in social outcomes (see Ladd et al., 1997
), we also conducted a series of analyses examining the protective influence of friendship independent of the variance in victimization predicted by social preference (i.e., group social acceptance). A secondary goal of these analyses was to examine the moderating role of social preference. Although we conceptualized social preference as a mediator, some past researchers have viewed social preference as an important moderator in the pathways to victimization (e.g., Hodges et al., 1997
). A separate hierarchical regression was conducted for each behavior problem cluster, with T4 victimization predicted from the main effects of gender and the T4 behavior problem score (Step 1); the main effects of T1 social preference, T1 friendship, and the T1 behavior problem score (Step 2); the interaction term for T1 social preference × the T1 behavior problem score (Step 3); and the interaction terms for T1 friendship × T1 social preference and T1 friendship × the T1 behavior problems score (Step 4).
As shown in , there were no significant T1 Social Preference × behavior problem interactions. That is, social preference did not moderate the predictive relation between any of the behavior problem clusters and T4 victimization. Consistent with our earlier analyses, however, there was a significant T1 Externalizing × T1 friendship interaction. This effect was significant even though the variance associated with T1 social preference had been controlled.
Hierarchical Analyses of T1 Friendship as a Moderator in the Predictive Association Between T1 Behavior Problems and T4 Victimization With T1 Social Preference Controlled
Gender as a Moderator
A series of regression analyses was then conducted to examine the moderating role of gender in the prediction of victimization. A separate analysis was conducted for each behavior problem cluster with the main effects for the T4 behavior problem, the T1 behavior problem score, and gender entered on Step 1 and the T1 Behavior Problem Score × Gender interaction term entered on Step 2. There was a significant T1 Attention Problems × gender interaction, β = .101, sr2 = .010, p ≤ .05. Regression models conducted separately for each gender indicated that T1 Attention Problems (with T4 Attention Problems controlled) was more predictive of T4 victimization for girls, β = .228, sr2 = .048, p ≤ .005, than boys, β = .075, sr2 = .004, ns.
Next, we conducted a series of analyses to determine whether the moderating role of friendship differs as a function of gender. A separate analysis was conducted for each behavior problem cluster with the main effects for the T4 behavior problem, the T1 behavior problem score, T1 friendship, and gender entered on Step 1; the two-way interaction terms for T1 Behavior Problem Score × gender, T1 Behavior Problem Score × T1 friendship, and T1 friendship × gender entered on Step 2; and the three-way interaction term for T1 Behavior Problem × T1 friendship × gender entered on Step 3. These analyses yielded a significant three-way interaction effect for T1 Social Problems × T1 friendship × gender, β = –.123, sr2 = .014, p ≤ .05.
Regression models conducted separately for each gender indicated that there was a significant T1 Social Problems × T1 friendship effect for boys, β
= −.159, sr2
= .022, p
≤ .05, but not for girls, β
= .115, sr2
= .012, ns
. Analyses guided by Aiken and West's (1991)
suggestions showed that the slope of the relation between T1 Social Problems (with T4 Social Problems controlled) and T4 victimization declined for boys as their level of T1 friendship moved from low, β
= .519, p
≤ .001, to medium, β
= .111, ns
, and then to high, β
= −.207, ns