This study examined data collected as part of an 8-year project that evaluated a bullying prevention program (Farrell et al., 2018; Sullivan et al., 2021). Participants were a random sample of students from three public middle schools in the southeastern United States that served students in urban neighborhoods with high rates of violence-related crimes. The majority of students at these schools (74% to 100%) were eligible for the federal free lunch program. During the first year of the project, a random sample of students from each grade were recruited from each school. In each subsequent year a new cohort of incoming 6th graders was recruited along with 7th and 8th graders to replace those who left the schools or discontinued participation. Participants remained eligible until they completed the 8th grade, left the school, or chose to withdraw from the study. The study was approved by the institutional review board of the authors’ university. All participants provided written parental consent and informed assent. Students received a $5 gift card for returning consent forms whether or not their parents gave consent for participation in the study. Close to 80% of all eligible participants were recruited.

The final sample of 2,705 students had a mean age of 12.35 (SD = 1.03) years at Wave 1. Data on sex based on school records identified 52% of students as female and 48% as male. Seventeen percent identified themselves as Hispanic or Latino/a. Twelve percent, most of whom (79%) had identified themselves as Hispanic or Latino, did not endorse any category for race, 6% identified multiple categories, 79% identified themselves as African American (including 6% who endorsed multiple categories), 6% as White, and 3% identified other racial categories. Twenty five percent reported living with both parents, 25% with a single mother and no other adult, and 23% with a parent and stepparent.

The project that provided the data (see Farrell et al., 2018; Sullivan et al., 2021 for details and findings) used a multiple baseline experimental design to evaluate the Olweus Bullying Prevention Program (Olweus & Limber, 2010). The order and timing of initiating the intervention at each school was randomized. Once started, it continued at each school until the project ended. About two-thirds of participants (65%) completed measures during a year when the intervention was being implemented at their school. Students completed measures in the fall, winter, spring, and summer (waves 1 to 4, respectively) of each school year between 2010 and 2018 with the following exceptions. The first wave was collected in the winter, the last wave in the spring of the final year, and a change in funding source prevented collection of data in the fall of Year 6.

Students completed measures on a computer-assisted interview at school during the school year and in their homes or another community location during the summers. Participants who completed any portion of the survey received a $10 gift card. The project employed a missing-by-design approach such that each participant was randomly assigned to complete measures at two of the four waves during each school year. This provides data missing completely at random. Graham et al. (2001) argued that planned missing designs can provide unbiased estimates of parameters and tests of hypotheses nearly as powerful as traditional designs, but have the advantages of decreasing costs and increasing quality by reducing carryover effects, participant burden, fatigue, and attrition. Such designs are particularly appropriate for longitudinal studies that include repeated observations of the same measures.

Not all students participated during all three grades. Four of the ten cohorts entered the project in the 7th or 8th grade during Year 1, or were in the 6th or 7th grade in Year 8, and some left a school before completing the 8th grade. Consequently, data were available from all three grades for only 26% of participants, and from two grades for 25% of participants. Because of the limited number of students who participated during all three grades, we restricted our focus to investigating changes across four waves within a single school year. For students from whom data were obtained in more than one grade, we randomly selected data from one of their grades to ensure an independent sample. This provided longitudinal data within a single grade for 934 6th graders, 867 7th graders, and 904 8th graders. Within the final sample, data were obtained from 77% of participants at both of the waves to which they were assigned. Participants had missing data at one of their assigned waves for the following reasons: (a) 6.2% were not available for scheduling; (b) 6.1% left the school during the school year; (c) 3.7% had their data screened out because it did not appear that they completed the measures carefully based on their speed of completion or field notes from research staff administering the measures; (d) 3.1% declined to participate; and (e) 2.1% withdrew from the study or were no longer eligible.

Separate longitudinal models examined each of the five belief variables as a mediator of relations between one of the three exposure measures and the frequency of physical aggression. All analyses were conducted in Mplus Version 8.4 (Muthén & Muthén, 2017). We used one-sided models consistent with our focus on witnessing violence and victimization as predictors rather than as consequences of beliefs and physical aggression, and to avoid the complications of including multiple meditation paths representing effects in different directions within the same model (see MacKinnon, 2008). Within these models, the belief variable at waves 2 through 4 was regressed on the exposure variable at the preceding wave, and the frequency of physical aggression at waves 2 through 4 was regressed on the exposure variable and the belief variable at the preceding wave (e.g., see Figure 1). Models also included autoregressive effects and controlled for dummy-coded covariates including male sex, grade (with 6th grade as the reference), and intervention status. Given the school-level design of the project, the intervention variable represented differences due not only to the intervention, but to school and cohort differences. The model included correlations among all variables assessed within the same wave. Measures of exposure (witnessing violence, victimization, or negative life events) at each wave were regressed on the covariates, but were otherwise treated as exogeneous variables (e.g., were correlated with each other across all waves, and with the belief variables at all prior waves). We also examined a multivariate model that included all exposure variables and belief variables that emerged as mediators within the separate models.

We compared competing models to test the consistency of effects across time, evaluated the fit of each model based on the root mean square error of approximation (RMSEA), comparative fit index (CFI), and Tucker-Lewis index (TLI), and compared models based on the scaled chi-square difference test (Satorra & Bentler, 2010). We computed standard errors using a robust estimator (i.e., MLR) to account for non-normality and to address missing data. We estimated the magnitude of indirect effects through each of the purported mediators using bias-corrected bootstrap estimation methods outlined by MacKinnon (2008).

Means, standard deviations, and correlations among the measures within waves 1 and 2 are reported in Table 1 (see Table S-1 in supplement for correlations across all four waves). Stability coefficients ranged from .56 to .72 for the three exposure measures and physical aggression, and from .34 to .62 for belief measures. Most variables within the same wave were significantly correlated. Witnessing violence and victimization were highly correlated with each other (r = .62 to .67), and with negative life events (r = .40 to .53). All but 5 of the 30 correlations among the five belief variables within each of the four waves were significant and in the expected direction, ranging in absolute value from .09 to .61. Most (i.e., 29 of 36) correlations between measures of exposure and belief variables were significant, ranging from .06 to .26 in absolute value (median = .13). Physical aggression was significantly correlated with self-efficacy and the beliefs scales, with correlations ranging in absolute value from .09 to .35 (median = .23). The pattern was generally similar across measures one or more waves apart, though their magnitudes were generally smaller.

Table 1 also reports d-coefficients representing mean differences at Wave 1 by sex, grade, and status of the intervention. There were sex differences on three of the nine variables such that boys reported higher frequencies of witnessing violence, victimization, and less support for beliefs against fighting compared with girls. There were grade differences on nearly all of the variables. Compared with 6th graders, 7th and 8th graders reported lower frequencies of witnessing violence, and victimization, lower self-efficacy for nonviolence, and more support for beliefs that fighting is sometimes necessary and beliefs supporting reactive aggression. Eighth graders also reported less support for beliefs against fighting, greater support for beliefs supporting proactive aggression, and higher frequencies of physical aggression. Participants assessed during years the intervention was being implemented at their school reported lower frequencies of witnessing violence and beliefs that fighting is sometimes necessary. These differences were small, with d-coefficients ranging in absolute value from .13 to .32; only 4 of the 18 significant d-coefficients exceeded .25 in absolute value.

The initial models separately paired each belief variable with each exposure variable and allowed all path coefficients to vary across waves. All had acceptable fits based on the RMSEA (all < .04) and CFI (all > .95), but not on the TLI (.77 to .89) (see Tables S-2 to S-4 in supplementary materials for all fit indices). Adding second-order autoregressive effects (i.e., each endogenous variable at waves 3 and 4 was regressed on its prior values at waves 1 and 2, respectively) significantly improved the fit of each model based on the scaled chi-square difference test at p < .001, and improvement in fit indices (RMSEAs all < .03; CFIs all > .98; and TLIs = .92 to .99), and were therefore included in subsequent models. We next evaluated the consistency of the cross-variable relations within each model by holding them constant across waves. In every case, this did not decrease the fit based on the scaled chi-square test (see Tables S-2 to S-4 in supplement) and had a minimal effect on the fit indices. The resulting models for each beliefs measure fit the data very well (RMSEAs < .02, CFIs > .98 and TLIs > .94; see models WV1 to NLE5 in Table 2), and provided the basis for evaluating mediation effects. Standardized regression coefficients representing relations between Wave 1 and Wave 2 variables for each of the 15 models are reported in Table S-5 to S-7 in supplemental materials). Specific findings for witnessing violence, victimization, and negative life events are discussed in the following sections.

Given the high correlations among the exposure variables (witnessing violence, victimization, and negative life events), we used a multivariate model to determine the combined and unique effects of the three exposure variables for predicting changes in each of the three belief variables that emerged as mediators in the separate analyses (i.e., self-efficacy for nonviolence, beliefs against fighting, and beliefs supporting proactive aggression; see Figure 2). As in the separate analyses of each belief variable, adding second order autoregressive effects significantly improved the fit of the model (Δχ²[8] = 75.72, p < .001). Further simplifying this model by holding all cross-variable coefficients constant across waves did not significantly decrease the fit (Δχ²[30] = 35.12, p < .238), and the resulting model fit the data well (see Model MV3 in Table 2). Within this model, we examined effects based on regression coefficients that indicated the extent to which each exposure variable accounted for a unique portion of the variance in each belief variable after controlling for covariates and autoregressive effects. We also examined their combined effects based on squared multiple partial correlation coefficients (pR²; Cohen et al., 2003) that indicated the total proportion of residual variance in each belief variable a (i.e., variance not accounted for by the covariates and autoregressive effects) accounted for by all three exposure variables. We used a similar approach to determine the individual and combined effects of the three belief variables on physical aggression. Coefficients representing relations across the first two waves are reported in Table 3, cross-variable coefficients across all waves are reported in Figure 2, and the full set of coefficients are reported in Table S-8 in supplemental materials.

Witnessing violence accounted for a unique proportion of variance in changes in self-efficacy for nonviolence (βs = −.08, p = .022) and beliefs supporting proactive aggression (βs = .08 to .09, p = .018). Negative life events also accounted for a significant proportion of the variance in changes in self-efficacy for nonviolence (βs = −.06 to −.07, p = .029). In contrast, victimization was not associated with subsequent changes in any of the belief variables after controlling for witnessing violence and negative life events. The three exposure variables as a set accounted for a significant portion of residual variance for self-efficacy for nonviolence (pR² = .012 to .016 across waves, p < .05) and for beliefs supporting proactive aggression (pR²= .016 to .021, p < .05), but not for beliefs against fighting (pR² = .004 to .005, p = ns).

Each of the three belief variables was uniquely associated with subsequent changes in physical aggression in the expected direction. Coefficients ranged from −.05 to −.06 (p = .023) for self-efficacy for nonviolence, from −.06 to −.08 (p = .021) for beliefs against fighting, and from .08 to .09 (p = .035) for beliefs supporting proactive aggression. The three belief variables accounted for a significant proportion of residual variance in physical aggression (pR²s = .022 to .033 across waves, ps = .015), after controlling for the covariates, autoregressive effects, and the three exposure variables at the prior wave.

Bias-corrected bootstrap estimates of the total indirect effect indicated that the three belief variables mediated the effect of witnessing violence at waves 1 and 2 on changes in physical aggression at waves 3 and 4 (βs = .014 and .015 95% CIs= [.005, 025] and [.006,.028), respectively). There were also significant specific indirect effects through beliefs supporting proactive aggression (βs = .006 and .007 95% CIs= [.001, 017] and [.001,.088), respectively), and through self-efficacy for nonviolence across waves 2 to 4 (β = .004, 95% CIs= [.001, 016]). Although this provided support for the belief variables as mediators of the relation between witnessing violence and physical aggression, the direct effect of witnessing violence remained significant within this model (βs = .09 to .11, p = .037), which suggested partial mediation. There were no significant indirect effects of victimization on physical aggression. There was a significant total indirect effect of negative life events on physical aggression through the belief variables across waves 2 to 4 (β = .008, 95% CIs= [.001, 017]), but this effect was not significant across waves 1 to 3 based on the confidence interval (β = .007, 95% CIs= [.000, 015]), and none of the specific indirect effects was significant.

We also examined correlations among the residual variances within each wave. These reflect shared variance among the variables within each wave not accounted for by variables at the prior wave or covariates. Residuals for physical aggression were significantly correlated with residuals for self-efficacy for nonviolence (r = −.10 to −.16, p < .01) and beliefs supporting proactive aggression (r = .12 to .24, p < .001), but not with beliefs against fighting). Among the three belief variables, self-efficacy was significantly correlated with beliefs against fighting (r = .11 to .15, p < .01) and beliefs supporting proactive aggression (r = −.11 to −.17, p < .01). The beliefs against fighting measure was not significantly correlated with beliefs supporting proactive aggression.

We conducted sensitivity analyses to determine the consistency of findings across sex, grades, and intervention condition. This involved multiple group models in which we compared the fit of models that allowed all parameters to vary across groups to models in which all path coefficients (except for covariate effects) were held constant across groups (see Table 2). These comparisons did not reveal any significant differences in parameter estimates across sex (Δχ²(35) = 44.28, p = .135), grades (Δχ²(70) = 86.51, p = .088), or across intervention condition (Δχ²(35) = 44.69, p = .126).

Witnessing violence was inversely related to subsequent changes in beliefs against fighting, and positively related to subsequent changes in beliefs supporting proactive aggression. These effects remained evident even after controlling for victimization and negative life events. In contrast, victimization was positively related, but only to subsequent changes in beliefs supporting proactive aggression, and this relation was no longer significant after controlling for witnessing violence and negative life events. These results provide partial support for our first hypothesis, and are consistent with social-cognitive frameworks, which suggest that witnessing violence contributes to schemas and scripts that support the use of aggressive behavior (Huesmann, 1998). The absence of relations between witnessing violence and beliefs that fighting is sometimes necessary and beliefs supporting reactive aggression was surprising and could indicate that witnessing violence impacts beliefs about the use of instrumental aggression rather than retaliatory beliefs (e.g., “It is okay to fight someone if they call you names or tease you”). It is possible that certain beliefs about aggression may be more stable over time, and therefore less likely to be influenced by witnessing violence. For example, adolescents may hold beliefs that fighting is sometimes necessary that are shaped by other contextual factors (e.g., early family environment, parental messages about fighting, affiliation with deviant peers) so that by the time they enter middle school, exposure to violence no longer has a strong influence on these beliefs. Dusing et al (2019) found that witnessing violence was positively associated with subsequent retaliatory beliefs, but only when the witnessed violence involved family members (not friends or strangers) suggesting the personal relevance of the violence witnessed may also matter in shaping beliefs. Taken together, it seems that witnessing violence only influences some beliefs about aggression, underscoring the importance of differentiating between specific beliefs and exploring more details about the nature of the violence exposure in future studies.

The lack of unique relations between violent victimization and beliefs suggests that victimization does not impact social information processes to the same extent as does witnessing violence. Although social-cognitive theories have often been extended to forms of violence exposure other than witnessing (e.g., victimization, hearing about violence), some evidence suggests that victimization impacts other aspects of social information-processing, such as hostile intent attributions (Calvete & Orue, 2011; Dodge et al., 1990) and hostile social goals (Shanifar et al., 2001). However, findings from the current study indicate that beliefs about aggression are acquired through witnessing or observing community violence, not through direct victimization. Perhaps victimization alters youths’ perceptions of others’ intentions and behaviors or beliefs about the world more broadly, but not their beliefs regarding the acceptability of aggressive behavior.

As with witnessing violence, experiencing negative life events was positively associated with subsequent changes in beliefs supporting proactive aggression, suggesting that such events may influence social-cognitive processes in a similar way as witnessing violence. Some of the negative life events measured by the UANLES include adult models of negative behavior (e.g., drug use) and knowledge of others engaging in violence (e.g., someone you know had a gun). These experiences may reinforce adolescents’ beliefs that aggression is an acceptable way to get what they want. Experiencing negative life events was the only exposure variable associated with subsequent changes in beliefs supporting reactive aggression. Perhaps adolescent’s exposure to events over which they have little control (e.g., family members’ behavior) reinforces perceptions that others are a threat, which reinforces reactively aggressive beliefs. This notion is in line with prior work that found that negative life events were more strongly related to post-traumatic stress symptoms (e.g., hyperarousal) than either witnessing or being the victim of community violence (Thompson et al., 2020). These findings also underscore the importance of accounting for concurrent negative life events when investigating the extent to which exposure to violence impacts youths’ beliefs and behavior.

Witnessing violence and negative life events were inversely associated with changes in self-efficacy for nonviolence, whereas victimization was not. This is in line with prior work, which found that witnessing violence was associated with lower self-efficacy to avoid violence (Thomas et al., 2016). It is possible that witnessing others being the victims of violence and experiencing negative life events decrease adolescents’ self-efficacy for nonviolence because they have seen others respond nonviolently without success or have internalized the belief that aggression is necessary for survival in certain situations (e.g., someone at school has a gun). In contrast, violence victimization was not associated with changes in self-efficacy for nonviolence. Perhaps youth who are victimized attribute the victimization to forces outside of their control (e.g., external locus of control) and therefore this experience does not alter (either positively or negatively) their confidence in using nonviolent responses. This finding is inconsistent with prior work by McMahon et al (2009) who found inverse associations between violence exposure and self-efficacy for nonviolence. However, because they combined experiences of witnessing and victimization, it is possible that their effects were confounded.

Consistent with prior cross-sectional findings (Bettencourt & Farrell, 2013; Farrell et al. 2019) and social information processing theories (Crick & Dodge, 1994; Huesmann, 1998), all five belief variables were significantly correlated with physical aggression in the expected directions. Longitudinal models indicated that all belief variables were related to subsequent changes in physical aggression with the exception of beliefs that fighting is sometimes necessary. Moreover, the multivariate model indicated that self-efficacy for nonviolence and beliefs against fighting, were uniquely inversely related to changes in adolescents’ frequency of physical aggression, and beliefs supporting proactive aggression was uniquely positively related to changes in adolescents’ frequency of physical aggression. These findings are consistent with prior studies that have found longitudinal associations between beliefs and aggression (e.g., Guerra et al., 2003; McMahon et al., 2013; McMahon et al., 2009; Nash & Kim, 2007). Our results suggest that adolescents’ behaviors are likely guided by both general beliefs as well as situation-specific beliefs about aggression (Farrell et al., 2008; Huesmann, 1998).

We found sex differences across several measures. Boys reported higher frequencies of both witnessing violence and victimization by violence compared with girls, which is consistent with prior studies assessing frequency of violence exposure (Finkelhor et al., 2015). Boys also reported less support for beliefs against fighting compared with girls. We did not, however find sex differences in self-efficacy for nonviolence or on the other measures of beliefs. This differs from some prior studies that have found sex differences in beliefs about aggression in elementary school students (e.g., Guerra et al., 2003). This may reflect age differences in that aggression might be considered adaptive and necessary for survival for adolescent boys and girls up in communities with high rates of violence (Farrell, Bettencourt, & Mehari, 2018).

Several limitations should be noted. All measures were based on adolescents’ self-report of their beliefs, experiences, and behavior, which could have inflated the degree of associations between the measures due to shared method variance. Nonetheless, self-report is the most appropriate method for assessing beliefs. Adolescents may also be more accurate reporters of their experiences of violence and negative life events because parents and teachers observe adolescents in limited contexts where they are typically less likely to engage in aggression. Prior research also suggests that parents under-report adolescents’ exposure to violence (Martinez & Richters 1993). Because early adolescents experience frequent changes during middle school, we examined changes across four observations within a single school year. These relations might play out differently across broader spans of time. Finally, because the focus of this study was on a specific population (i.e., a predominately African American sample of urban early adolescents), the findings may not generalize to adolescents in other settings. Adolescents growing up in neighborhoods with high rates of violence and a lack of resources are more likely to experience a range of adverse experiences and lack protective factors (e.g., access mental health care, positive youth development programs) compared to their peers (Foster & Brooks-Gunn, 2009). They may also face structural inequalities that block access to protective factors. The findings from our study regarding exposure to violence must therefore be considered within a broader range of factors that could potentially impact both adolescent aggressive behavior and social cognitive processes. Nonetheless, this population is an important group on which to focus given their high risk for violence exposure. Further work is needed with other sample of adolescents at risk for exposure to violence and negative life events.

The findings from this study reflect the adverse effects of exposure to community violence, particularly witnessing violence, on beliefs that shape adolescents’ behavior. Social information processes, and beliefs about aggression in particular, are salient mechanisms in understanding relations between witnessing violence and subsequent physical aggression. The findings highlight differences in the effects of witnessing violence versus victimization and the importance of studying them as distinct constructs. They also suggest that the effects of witnessing violence should be considered within the broader context of other negative life events that influence adolescents’ social-cognitive processes and engagement in aggression. Findings also underscore the importance of assessing specific beliefs about aggression and non-violence as mechanisms of relations between witnessing violence and physical aggression. Beliefs appear to play an important, but incomplete role in explaining how witnessing violence contributes to physical aggression. More research is needed to identify other factors that help to explain these relations (e.g., emotion regulation, peer and parental factors) and to identify factors that explain relations between exposure to violence and other forms of aggression (e.g., relational and cyber aggression).

The findings from this study have implications for prevention and intervention efforts. They highlight the importance of screening for adverse childhood experiences like witnessing violence in order to help identify youth in need of further intervention (e.g., Glowa et al., 2016). They also suggest that prevention programs should address beliefs about aggression among adolescents who have witnessed violence or experienced negative life events. Although interventions should target multiple specific beliefs about aggression, focusing on beliefs supporting proactive aggression may have the greatest impact on adolescent aggressive behavior. Violence prevention programs should also focus on providing adolescents with effective nonviolent alternatives to aggression and supporting the practice of these strategies to bolster adolescent’s sense of self-efficacy in enacting such responses when faced with conflict. Moreover, prevention programs should target a wider range of adverse life experiences that influence social cognitive processing such as structural inequalities given that racial and ethnic minority youth disproportionately reside in low-income neighborhoods with high rates of violence. Finally, although this study only investigated beliefs about aggression and self-efficacy, adolescents who have been exposed to violence would likely also benefit from interventions that target other stages of social information processing (e.g., cue interpretation, social problem-solving) to decrease engagement in aggressive behavior.