Descriptive Patterns
How prevalent was peer aggression? How much aggression occurred among classmates versus outside of class? And what type of aggression did students experience? As shown in
Table 1, in the first semester, 37.6 percent of students reported experiencing some form of peer aggression. In the next semester, only 19.1 percent of students reported being victimized. At the start of the next school year (the third semester), the rate of victimization rose to 25.4 percent. These results show that peer aggression was more likely to occur at the start of school years, perhaps because students were first meeting or getting reacquainted with one another.
To identify the extent to which peer aggression was occurring among classmates, I summarize data collected during the second wave of the survey, when students were asked to indicate whether they experienced aggression from people outside of their school, from those within their school but outside of their class, or from their classmates.
Table 2 displays these results: 15 percent of girls and 16 percent of boys reported being victimized by classmates, 6.2 percent of girls and 7 percent of boys reported being victimized by students within the same grade (but not classmates), and 4.7 percent of girls and 6.2 percent of boys were victimized by individuals outside of that grade or school. These results indicate that most victimization occurred among classmates. Indeed, when accounting not merely for prevalence (whether the student was victimized or not) but also the frequency of occurrences, 72 percent of all victimization events occurred among classmates. Spatially, 54 percent of victimization events occurred within the classroom, 27 percent occurred within dormitories (which still consisted of classmates), and the remainder occurred in other locations. These findings indicate that most aggression occurred among classmates and within the classroom.
What was the network structure of friendship nominations?
Figure 3 displays a randomly chosen network from my sample. A key feature of this network is that it displays gender segregation. To investigate the extent to which gender homophily is a general finding, I calculate an assortativity coefficient for gender across all networks and all three time periods. This coefficient can be interpreted similarly to a Pearson correlation coefficient, as it ranges from –1 to 1 and corresponds to the degree to which friendship nominations are likely (positive values) or unlikely (negative values) to occur between same-gender individuals. In this case, the mean assortativity coefficient for gender was .92 across all classes in the first semester, .81 in the second semester, and .72 in the third semester.
To confirm high levels of gender segregation, I fit an ERGM with terms for mutuality, edges, the propensity to send and receive ties by gender, and gender homophily for each of the 256 classes and for each period. I then calculate second difference average marginal effects for gender homophily, weighted by the inverse of standard errors observed for each network (see Duxbury 2022). Results show the probability of a friendship nomination between individuals of the same gender was 2.7 percentage points more likely to occur than a nomination between individuals of differing genders, even after accounting for processes like mutuality and the propensity for girls to send more ties than boys. This is six times larger in magnitude than comparable studies of homophily in other adolescent settings (e.g., .4 percentage points; see
Goodreau et al. 2009). A critical fact that follows from this high level of gender segregation is that boys and girls occupy discrete social circles, especially at the beginning of their time in junior high school. Hence, the scope condition of working knowledge in the group may only hold within gender subgroups, motivating the need for analyses relating peer aggression to hierarchical clarity for boys and girls separately, even as they formally occupy the same class.
Figure 4 shows how hierarchical clarity changes over time, demonstrating that status hierarchies generally grow stronger as individuals spend more time interacting with one another.
7 While these results do not elucidate the specific role of peer aggression, they support claims that status hierarchies emerge organically as individuals interact with each other (e.g.,
Gould 2002). To help provide intuition for what “strength” or “clarity” entails,
Figure 4 shows two example networks: one where the status hierarchy is weak (A) and the other where it is strong (B). To ease interpretability, simplified blockmodels of extreme cases within the observed networks are plotted. The blockmodels show that groups with high tau have roles that can be rank-ordered and clearly distinguished. By contrast, when tau is low, there is no consistent way to identify roles that are rank-ordered by esteem. These temporal changes are also confirmed in descriptive statistics shown in
Table 1, which show the tau statistic increases in magnitude over the three semesters.
Main Effects of Aggression on Hierarchical Clarity
For data transparency,
Figure 5 consists of a raw scatterplot between hierarchical clarity and peer aggression, indicating the two measures are
negatively associated. This evidence supports the need for something like the legibility framework to explain how aggression erodes status hierarchies. Still, a more plausible explanation for this correlation is reverse causality: strong status hierarchies reduce aggression. To disentangle this causal relationship, I investigate temporal sequences in these data.
Column 1 in
Table 3 shows the results from a school and semester fixed-effects regression, where the class-level prevalence of peer aggression is lagged. The results show that tau, a measure for hierarchical clarity, decreases by .473 when moving from a hypothetical class where 0 percent of students report being victimized to a class where 100 percent of students report being victimized in the prior semester (
p = .063). Or, placed on a more interpretable scale, a 10-percentage-point increase in the proportion of students who report being victimized correlates with a .0473 decrease in tau. Column 2 shows this result is robust to the inclusion of a rich set of covariates pertaining to the composition of these classes (
p = .029). Column 3 shows results when we use
class fixed effect, which removes the confounding effects of any time-invariant covariates at the class level. Across these columns, the overall direction of the effect remains unchanged, and the magnitude increases to .725 in the class fixed-effect model (
p = .037). Given that the average level of tau observed across the entire sample is 1.05, this corresponds to a 69 percent decrease in hierarchical clarity.
8
How do the results change when we account for confounding network-generating processes? Columns 4 and 5 in
Table 3 repeat these analyses for the average marginal effect (AME) of the tau statistic estimated from the exponential random graph models. Column 4 shows that peer aggression reduces the probability that a new tie would create a hierarchical triad by 1.7 percentage points (
p = .029). Column 5 supports this conclusion, with results from the peer-group fixed-effect model indicating that peer victimization reduces tau by 2.9 percentage points (
p = .003). Given that the base probability of any tie formation is 7.1 percent, this is a 41 percent decrease. Although results using the ERGM-derived measures of hierarchical clarity are more persuasive, the ERGM model failed to converge in cases where gender segregation was nearly absolute, which explains why the total number of observations in columns 4 and 5 is 320, as opposed to 512. This problem motivates the need to investigate gender subgroups.
The inclusion of lags accounts for temporal sequence, but as noted above, stronger evidence against reverse causality would incorporate cross-lagged fixed effects. The results of such a model, estimated using maximum likelihood, are shown in
Table 4. The frequency of peer aggression in the prior semester reduces hierarchical clarity, even with the inclusion of lagged hierarchical clarity and regardless of whether clarity is measured through a dyad-conditioned tau statistic (column 1:
β = –.845,
p = .009) or the average marginal effect of the tau statistic estimated in an ERGM (column 2:
β = –.043,
p = .001). In sum, the direction and statistical significance of the results remain unchanged, and the magnitude of the effects are comparable to those estimated using models without cross-lagged fixed effects.
Why Did Peer Aggression Reduce Hierarchical Clarity?
Taken together, the results show that peer aggression erodes the clarity of status hierarchies. This finding rules out the dyadic or audience perspectives but leaves specific predictions of the legibility framework untested, and other alternatives could explain the results. Hence, before testing key claims of the legibility framework, I highlight and rule out three competing possibilities. For instance, one possibility is that each act of aggression introduces ambiguity as to the social rank of other actors until people in the group have seen enough information to decide on individuals’ positions. When such a threshold of aggression is reached, we might see the hierarchy shift, perhaps rapidly, into a state of clarity. This threshold explanation suggests the effect of peer aggression reverses when a sufficiently high level is reached (
Biggs 2005;
Granovetter 1978). Similarly, an Aristotelean “golden mean” hypothesis suggests an inverse U-shaped functional form, where too little and too much aggression erodes status hierarchies, but middling levels of aggression strengthen them. Is there a nonlinear functional form that relates peer aggression to the strength of status hierarchies? To uncover nonlinear relationships, I decompose the peer aggression variables into five quintiles and include this categorical variable as the primary independent variable.
Figure 6 shows predicted levels of tau based on this analysis: the marginal effect of peer aggression on the strength of status hierarchies appears to be linear, with little evidence of nonlinear effects.
Second, perhaps peer aggression erodes group structure more generally, reducing the likelihood of
all triadic structures, not just hierarchical ones. Similarly, the choice of five triadic building blocks to represent hierarchical structure is based on prior research (
Cartwright and Harary 1956;
Faust 2007;
Moody et al. 2011;
Rawlings and Friedkin 2017), but perhaps the effects are driven by a subset of these triads, representing an alternative type of network structure. To resolve these possibilities,
Figure 7 shows every triad in the census and its likelihood of being observed in each network, as a function of different levels of observed peer aggression. To check for nonlinear relationships, I also show boxplots of the distribution of tau statistics for each quartile of peer aggression. Correlation coefficients (and whether they are statistically significant at the .05, .01, or .001 level) are provided to help evaluate whether each triad appears/disappears when aggression increases in prevalence. A visual representation of each triad is also provided to help illustrate what each triad means (beyond the customary MAN label).
The results show that the frequency of peer aggression is correlated with a statistically significant increase in the probability of six triads {012U*,012C,111D,030C,201, 120C}, whereas eight triads decrease in probability {012,102,021D*,030T*,120D*,120U*,210,300}. The hierarchical triads are labeled with an asterisk, showing that all hierarchical triads decrease in probability except for 012U. Additionally, all cyclic triads (triads labeled with “C”) are also positively correlated with the prevalence of peer aggression. This can be understood as a further robustness check. Rather than hierarchical clarity, I could have measured “hierarchical confusion” as the proportion of cyclic triads that occur in the group, that is, triads that directly contradict rank-ordered positions. The results show that this measure increases as a function of peer aggression. Hence, it is unlikely that peer aggression simply erodes all triadic network structures, and its effects appear to be specific to hierarchical clarity.
Third, perhaps peer aggression reduces hierarchical clarity because it reflects attempts from lower-status students to redistribute esteem in the classroom. Put more colorfully, aggression might reflect insurrection against the existing status order, whereby group members rely on peer aggression to disrupt the hierarchy and redistribute esteem from the “haves” to the “have nots.” If so, high-status students would be expected to lose esteem more than low-status students as peer aggression increases in frequency. With the caveat that friendship nominations are a weak proxy for status, I test for how peer aggression affects friendship nominations to higher- versus lower-status individuals. I do so by including a node-level characteristic in the ERGM model, corresponding to the ratio of in- to out-degree ties received in a previous time-period (a measure of status). The coefficient in the ERGM model indicates the degree to which high-status actors receive more ties. The previous time-period lag is included to avoid circularity.
Once these ERGM coefficients are calculated, I investigate how this is moderated by the prevalence of peer aggression. In classrooms where there is more peer aggression, do the (initially) high-status actors lose or receive more ties? As before, I translate the ERGM coefficients to average marginal effects to avoid problems of scaling and weight by the inverse of the standard error. The results (
Table 5, column 1) show no conclusive pattern, suggesting that levels of aggression in a classroom do not change whether formerly high-status actors receive more or fewer ties. In other words, there is no clear evidence that aggression is
redistributive of esteem, either toward those who already have it, or those who do not. If status inequalities are not being undone, how can hierarchical clarity diminish? This is possible because hierarchy is not solely a function of whether there are millionaires and mendicants in the world of esteem, and the clarity of a hierarchy is not only about inequality in resources. Rather, the structural definition of hierarchical clarity encompasses consensus about status positions, which is a feature of both asymmetry and acyclicity in esteem.
These analyses, which rule out alternative possibilities, imply the need for an explanation like the legibility framework, which posits that uncoordinated audience perceptions and reactions to peer aggression are what erodes hierarchical clarity. To directly test this explanation, I investigate whether audience perceptions matter in translating aggression to hierarchical clarity, and whether peer aggression strengthens status hierarchies if audiences accurately interpret aggression. I begin by assessing whether audience perceptions of aggression affect hierarchical clarity, even when controlling for victim-reported levels of peer aggression.
Table 5, column 2, shows that the effect of aggression on hierarchical clarity is carried by perceived aggression, where moving from a class where nobody is perceived to have been a victim to one where everyone is reduces the probability that a new tie would create a hierarchical triad by 1.3 percentage points (
p = .003). Additionally, the effect of victim-reported aggression nearly halves, from 1.73 percentage points when perceived victimization is omitted from the model (
p = .029), to .98 percentage points when perceived victimization is included (
p = .130). These results show that audience perceptions matter in predicting hierarchical clarity.
More importantly, I assess whether the effect of aggression depends on whether audiences can accurately identify it.
Table 5, column 3, shows that, conditional on being accurately perceived by audience members, the likelihood of hierarchical ties forming increases by 7.59 percentage points when the proportion of people reporting being victimized shifts from 0 to 100 percent (
p = .014). By contrast, when decoupled from perceptions, peer aggression erodes the clarity of status hierarchies (
p = .009). If the effect of peer aggression on group structure is driven primarily by audience reactions, then how would victim-reported peer aggression erode hierarchical clarity when decoupled from audience perceptions? One reason is that these are tightly knit groups of approximately 32 students, interacting across multiple semesters. This implies that the misalignment between victim-reported and audience-reported aggression is not solely because people did not see the interactions, but at least in part because they did not accurately interpret it as aggression.
Although these two tests conform with the predictions of a legibility framework, significant analytic concerns remain. One concern is that statistics for tau did not converge in groups with high degrees of gender segregation, resulting in a smaller sample that is no longer representative of adolescent groups in this context. Are the results robust when tau is recalculated using ERGMs for gender subgroups within classes? Not only do high levels of homophily in friendship nominations imply that the relevant peer groups for these 7th- and 8th-graders are their same-gender classmates, convergence and model fit improve when we consider each class as consisting of two subgroups, one comprising boys and the other girls. Using this robustness check, the overall finding remains consistent: hierarchical clarity increases when perceived and victim-reported victimization are aligned (column 4).
To help visualize the results,
Figure 8 visualizes the model in column 4 in a contour plot. The
z-axis corresponds to average marginal effects of the tau statistic as calculated in ERGMs. As noted earlier, these values can be interpreted as the percentage point increase in the baseline probability of tie formation between two individuals (in this sample, 7.1 percent), conditional on a tie completing a hierarchical triad. For instance, a value of .01 is a 1 percentage-point change, implying an 8.1 percent chance of ties forming that complete hierarchical triads, relative to the baseline of 7.1 percent. The
x-axis corresponds to the percent of adolescents in each class who reported experiencing peer aggression, and the
y-axis corresponds to the percent of people in each class who experienced aggression,
as perceived by onlookers. The dashed line corresponds to perfect accuracy, where members of the group perceive exactly as much aggression as is being reported by victims. The overlaid scatterplot provides the region of support, that is, the distribution of classes in the sample. Any areas of the plot where there are no points are extrapolations of the model. The scatterplot illustrates that audience estimates of aggression in the class are correlated with victim-reported rates of aggression. Nevertheless, the scatterplot shows these estimates are inaccurate, and the slope of a line of best fit is lower than .5, indicating the perceived proportion of adolescents experiencing aggression is generally downwardly biased. In other words, onlookers generally underestimate how much peer aggression is happening in the class.
When it comes to hierarchical clarity, the results first show that increasing aggression (as perceived by victims) predicts weaker status hierarchies. The marginal effect of peer aggression is negative, that is, if we fix the mean position on the y-axis and move toward the right on the x-axis, predicted levels of hierarchical clarity decrease. This is consistent with the main effects discussed in the previous section. Status hierarchies, however, are stronger when actual and perceived aggression increase simultaneously—that is, if each move on the x-axis corresponds with a move on the y-axis. These results support the key predictions of a legibility framework: when group members fail to accurately perceive interactions as aggression, it erodes the status hierarchy, but aggression that is legible to onlookers increases the clarity of the status hierarchy.
A remaining concern is whether alignment between perceived and victim-reported victimization is a comprehensive operationalization for legibility. When audiences accurately identify peer aggression in a way that matches how victims experienced aggression, legibility is likely to be high. Yet, it is possible to imagine audiences sharing consistent interpretations of peer aggression that are systematically different from victims. Put differently, consistency implies low variance rather than accurate means, and peer aggression could be legible to audiences without necessarily implying alignment with victims’ perceptions. Indeed, if the effects of peer aggression operate primarily through audience perceptions, then what victims report may be irrelevant. To address this concern, I further test whether the results are robust when legibility is operationalized as the standard deviation in onlookers’ perceptions of aggression.
Table 5, column 5, shows the results for this additional test. The effect of peer aggression as perceived by audiences on hierarchical clarity is positive and statistically significant (
p = .012), but, as predicted, as
variability in audience estimates of aggression increases (which operationalizes low levels of legibility in a group), this effect reverses and begins to erode hierarchical clarity (
p = .073).
As a final test of the legibility framework, I investigate whether aggression that is coordinated with others contributes to greater hierarchical clarity. As predicted, the key underlying condition that needs to be met for individual interactions to strengthen status hierarchies is whether audiences are coordinated in how they reallocate esteem. Hence, peer aggression that draws multiple actors together to attack a victim ought to clarify a status hierarchy. In
Figure 9, I plot the marginal effect of increased peer aggression on hierarchical clarity as a function of the average number of people who participated (shown in
Table 5, column 6). The figure shows that when only a single individual was an aggressor, peer aggression generally erodes hierarchical clarity. As predicted, as more individuals participate, the marginal effect of peer aggression reverses and begins to increase hierarchical clarity.
Heterogeneity by Gender, Type of Aggression Deployed, and Number of Aggressors
High levels of gender segregation observed among the classes in my sample necessitated additional analyses that considered gender subgroups separately. Were the effects of peer aggression on hierarchical clarity different for boys and girls?
Figure 10 plots predicted margins from
Table 6. For data interpretability, I include a histogram that shows the distribution of victimization prevalence in the data, which bounds the range of values the model could extrapolate to.
Figure 10 reveals that the effect of peer aggression on hierarchical clarity not only differs but also reverses for adolescents who identify as boys, compared to those who identify as girls. Among girls, the increasing incidence of peer aggression in the prior semester is correlated with
increasing probability that ties form hierarchical triads. By contrast, peer aggression appears to decrease the formation of hierarchical triads among boys. These gender differences are statistically significant (
p = .009, see
Table 6, column 1, row 3). In summary, the results show marked gender differences in how aggression relates to hierarchical clarity: aggression weakens status hierarchies among boys while modestly strengthening it among girls.
Although I did not have ex ante predictions about gender differences, the legibility framework advanced here helps to explain this gender difference. In particular,
Table 6, column 2, investigates whether there were gender differences in how accurate audiences were at predicting victim-reported levels of aggression. If the coefficients in the model were 1, then this would imply maximal accuracy, or perfect correlation between victim- and audience-reported levels of aggression. Among girls, the correlation was .136, and correlations were statistically significantly weaker for boys (
p = .001,
Table 6, column 2, row 3). This implies boys were less accurate than girls at interpreting aggression occurring among themselves.
Additionally,
Table 6, column 3, shows that gender differences in the effect of peer aggression on hierarchical clarity are no longer statistically significant (
p = .123) when models include audience perceptions of aggression, and the difference in coefficient magnitude when compared to a model where perceived aggression is omitted is statistically significant (
p = .0018). In short, the results suggest one reason peer aggression strengthened status hierarchies among girls, yet had the opposite effect for boys, is that aggression is more legible among girls.
Why might peer aggression be more legible among girls than among boys? One possibility is that boys and girls experience different forms of aggression. Girls were 8 percentage points more likely than boys to experience relational exclusion, and 13 percentage points more likely to experience verbal aggression. Boys, by contrast, were about 10 percentage points more likely to experience physical aggression (see
Table 2). To investigate whether these three forms of aggression affected hierarchical clarity differently, I include all three forms of aggression in a model predicting hierarchical clarity in the following semester. As noted earlier, one limitation here is that respondents were only asked to report on the type of aggression they experienced in the second wave of the survey, meaning the sample size of groups is halved. In a class-level analysis,
Table 6, column 4, shows no clear differences by type of aggression. Similarly, in a gender-subgroup analysis, column 5 also shows no clear differences. Thus, although boys and girls experienced different forms of aggression, there is insufficient evidence to conclude that this explains why aggression had divergent effects on hierarchical clarity by gender.