Social aggression is a form of antisocial behavior (Burt et al. Reference Burt, Donnellan and Tackett2012) in which social relationships and social status are used to damage reputations and inflict emotional harm on others, and centers on behaviors such as gossiping, ostracism, and threatening to end a friendship. Although socially aggressive behaviors are legal and relatively typical during particular developmental periods, they have been associated with pathological outcomes in both victims and aggressors. Victims of social aggression often experience as much emotional distress as victims of physical aggression, including emotional and social difficulties such as peer rejection, loneliness, and internalizing problems (Crick et al. Reference Crick, Casas and Nelson2002). The perpetrators of social aggression also exhibit a number of maladaptive outcomes, including poor quality friendships marked by conflict and instability, depression, externalizing behavior, and borderline personality disorder features. Not surprisingly then, there is substantial interest in developing prevention and intervention programs that target social aggression. To date, however, there is relatively little work available on its underlying etiology to inform these efforts.

Extant research regarding the origins of social aggression primarily focuses on social processes and psychosocial risk factors, and particularly on peer relationships (e.g. popularity, victimization, quality of dyadic relationships; Crick & Grotpeter, Reference Crick and Grotpeter1995; Sijtsema et al. Reference Sijtsema, Ojanen, Veenstra, Lindenberg, Hawley and Little2010) and parenting factors (e.g. psychological control, coercive parenting; Kawabata et al. Reference Kawabata, Alink, Tseng, van Ijzendoorn and Crick2011; Kuppens et al. Reference Kuppens, Grietens, Onghena, Michiels and Subramanian2013). Other work has focused on marital conflict, sibling relationships, media exposure, and cultural values (e.g. Karriker-Jaffe et al. Reference Karriker-Jaffe, Foshee, Ennett and Suchindran2008; Forbes et al. Reference Forbes, Zhang, Doroszewicz and Haas2009; Gentile et al. Reference Gentile, Coyne and Walsh2011). Relatively few of these studies, however, have been designed to examine both environmental and genetic risk factors simultaneously. This is an important omission in the literature, as it means that studies have thus far been unable to disentangle the relative contributions of genetic factors and different environmental influences.

Current theory and research has been largely silent on the role of genetic transmission and genetic risk for social aggression. Indeed, we know of no molecular genetic studies of social aggression. Fortunately, genetically informed twin designs can also be used to estimate genetic influences on social aggression, with the added advantage that they are able to simultaneously estimate environmental influences as well. To date, three such studies have been conducted, with notably inconsistent results. The first of these assessed social aggression via teacher and peer ratings in a small sample of 6-year-old twins (N = 234 pairs; Brendgen et al. Reference Brendgen, Dionne, Girard, Boivin, Vitaro and Pérusse2005). Brendgen et al. (Reference Brendgen, Dionne, Girard, Boivin, Vitaro and Pérusse2005) employed a constrained ACE model in which additive genetic (A) and shared environmental (C) influences were constrained to be equal because an initial unconstrained ACE model suggested equivalent fits for AE and CE models. Based on these results, the researchers concluded that 60% of the variability in social aggression is due to non-shared environmental factors (E), with 20% due to A and C, respectively. However, a follow-up study of these same twins only a year later (N = 203 pairs) failed to find evidence for significant shared environmental influences on social aggression (Brendgen et al. Reference Brendgen, Boivin, Vitaro, Bukowski, Dionne and Tremblay2008). Instead, results indicated that an AE model provided the best fit to the data, such that 42.9% of the variance in social aggression was due to genetic factors and 57.1% of the variance was due to non-shared environmental factors. Although such findings could suggest a change in the heritability of social aggression from ages 6–7 years old, we would argue against this interpretation, particularly given the very small sample sizes. Although heritability estimates for other antisocial behaviors do change throughout development (Burt & Neiderhiser, Reference Burt and Neiderhiser2009), they rarely do so in such dramatic fashion over the course of single year. It is thus unclear what to make of these conflicting results.

Following up on these results, Tackett et al. (Reference Tackett, Waldman and Lahey2009) made use of a multivariate psychometric model to test etiological influences on the variance common to both mother- and twin-reported child relational aggression in a large sample of 6–18 years old twins (N = 1981 pairs). In this model, variance common across informant reports is decomposed into genetic and environmental components, as is the variance unique to each informant. In contrast to Brendgen et al.’s (Reference Brendgen, Dionne, Girard, Boivin, Vitaro and Pérusse2005, Reference Brendgen, Boivin, Vitaro, Bukowski, Dionne and Tremblay2008) studies, Tackett et al. (Reference Tackett, Waldman and Lahey2009) found evidence for substantial genetic and shared environmental contributions to the latent relational aggression factor (63% and 37% of the variance, respectively). Non-shared environmental influences did not contribute to the latent factor. The variance specific to maternal informant reports, by contrast, was entirely due to non-shared environmental influences and measurement error. The variance unique to twin reports was due to unique genetic, shared environmental, and non-shared environmental influences.

Although Tackett et al. (Reference Tackett, Waldman and Lahey2009) provided important new information on the etiology of social aggression, much remains unknown. This lack of clarity stems from two methodological issues. First, the sample used in Tackett et al. (Reference Tackett, Waldman and Lahey2009) spans childhood, adolescence, and emerging adulthood, which is less than ideal should the relative contribution of etiological factors differ across those developmental periods (as in Burt & Neiderhiser, Reference Burt and Neiderhiser2009). More importantly, however, a burgeoning body of work (Keller & Coventry, Reference Keller and Coventry2005; Keller & Medland, Reference Keller and Medland2008; Keller et al. Reference Keller, Medland and Duncan2010) has indicated that heritability estimates from classical twin (CT) studies, such as those outlined above, can be biased in important ways, sometimes severely so. The CT model exploits the comparison of twin similarity across zygosity to infer etiology. To do so, however, it relies on a number of strict assumptions (Keller & Coventry, Reference Keller and Coventry2005; Keller et al. Reference Keller, Medland and Duncan2010). When these assumptions are met, the results should accurately reflect the proportions of genetic and environmental influences on the phenotype under study. However, when one or more of these assumptions are violated, the parameter estimates are likely to be biased in predictable ways.

One major assumption is that C and non-additive or dominant genetic effects (D) do not simultaneously influence the phenotype(s) under investigation. This assumption has no basis in biology, but is rather a statistical necessity of the model: because A and E must always be estimated, C and D cannot be simultaneously estimated or else the model will be underidentified. It is entirely conceivable, however, that social aggression is simultaneously influenced both by interactions between alleles (i.e. D) and common environmental factors that create similarities between siblings (i.e. C). Moreover, when C and D effects are both present, parameter estimates are biased, such that A estimates are inflated (sometimes grossly so) and C and D estimates are decreased. As an example, Keller & Medland (Reference Keller and Medland2008) simulated data in which A, D, and C variances were equal to 0.40, 0.15, and 0.15, respectively. Using the CT model, the ACE model (and not the ADE model) was chosen as the better fitting model, even though C and D effects were in fact equivalent in magnitude. Moreover, additive genetic influences were estimated at 0.60, whereas shared environmental influences were estimated at 0.02 (Keller & Medland, Reference Keller and Medland2008).

Finally, the CT model assumes that there is no evidence of passive gene–environment correlations (passive rGE). Passive rGE refers to the association between the genes a child inherits from his or her parents and the rearing environment provided by those same parents. Put differently, it is possible that the genes parents pass on to their biological children also influence the home environment in which the child is reared. Within the CT model, passive rGE is thought to mimic shared environmental influences, serving to inflate C estimates and decrease A and D estimates. It is thus possible that the shared environmental influences on social aggression identified by Tackett et al. (Reference Tackett, Waldman and Lahey2009) using the CT model may reflect passive rGE rather than actual environmental experiences.

Fortunately, there is a straightforward extension of the CT model, the nuclear twin family (NTF) model, which circumvents these assumptions by adding the twins’ biological parents to the model (Keller et al. Reference Keller, Medland and Duncan2010). In more precise terms, genetic and environmental parameter estimates in the NTF model (Fig. 1) are obtained from four pieces of information (rather than two, as done in the CT model): the covariance between MZ twins, the covariance between DZ twins, the covariance between parents, and the covariance between parents and children. The additional information allows the NTF model to improve upon the CT model by simultaneously estimating C and D and differentiating passive rGE from true shared environmental influences. The latter is accomplished via the individuation of shared environmental influences shared only by siblings (S) from those shared by all family members (F). We are then able to explicitly model the extent to which A and F variance components are correlated with one another (w). Perhaps not surprisingly, given these additions, extant simulation studies have robustly demonstrated that the NTF model provides more etiologically precise estimates and is more robust to violations of assumptions than the CT model (Keller et al. Reference Keller, Medland and Duncan2010).

Fig. 1.

Path diagram of a univariate nuclear twin family model.

Note: The variance in the phenotype (P) is parsed into that which is due to additive genetic effects (A), dominant genetic effects (D), sibling environmental influences (S), familial environmental influences (F), and non-shared environmental effects (E). See Table 1 for definitions. μ indexes primary phenotypic assortment (i.e., assortative mating) between the twin parents, and w indexes the covariance between A and F. Paths are squared to estimate the proportion of variance accounted for. D, S, and F effects cannot be estimated simultaneously (only two of the three can be estimated). Reprinted with permission from Burt et al. (2012).

Methodology

Study participants

The 1030 families included in the current study were assessed as part of the Twin Study of Behavioral and Emotional Development in Children (TBED-C) within the Michigan State University Twin Registry (MSUTR) (Klump & Burt, Reference Klump and Burt2006; Burt & Klump, Reference Burt and Klump2013). The TBED-C consists of two independent sub-samples of twins in middle childhood. The first sample consists of a population-based epidemiologic sample of 528 families (1056 twins and their parents). The second, ‘at-risk’ sample consists of 502 families (1004 twins and their parents), who reside in modestly-to-severely disadvantaged neighborhoods. Recruitment procedures have been described previously (Burt & Klump, Reference Burt and Klump2013; Burt et al. Reference Burt, Klump, Gorman-Smith and Neiderhiser2016). Children gave informed assent, while parents gave informed consent for themselves and their children.

Participating twins were 48.7% female and ranged in age from 6 to 10 years old, although some (n = 59) were 11 [mean age (s.d.) = 8.02 years (1.49)]. Rearing biological mothers ranged in age from 24.82 to 52.79 years [mean (s.d.) = 38.18 (5.21) years], and rearing biological fathers ranged in age from 24.12 to 64.22 years [mean (s.d.) = 40.12 (5.76) years]. Twins’ racial and ethnic background was provided by their parents (81.7% non-Hispanic white, 9.5% African American, 1.1% Native American, 0.8% Asian, 0.7% Hispanic, 0.3% Pacific Islander, and 5.9% multiracial or other ethnic groups). Twin zygosity was determined using a standard five-item, parent-report questionnaire that assesses within-pair physical similarity and is over 95% accurate (Peeters et al. Reference Peeters, Gestel, Vlietinck, Derom and Derom1998). Unclear zygosities were resolved by comparing twin sibling DNA markers (Klump & Burt, Reference Klump and Burt2006). Monozygotic twins constituted 41.4% of the pairs (n = 426 pairs, 202 female pairs), same-sex dizygotic twins constituted 40.4% (n = 416 pairs, 208 female pairs), and opposite-sex dizygotic twins constituted 18.3% (n = 188 pairs).

Measures

Social aggression was assessed using the Subtypes of Antisocial Behavior Questionnaire (STAB; Burt & Donnellan, Reference Burt and Donnellan2009; Burt & Donnellan, Reference Burt and Donnellan2010). The STAB is a 32-item measure assessing three major dimensions of antisocial behavior, one of which is social aggression. The Social Aggression Scale (SA) presents 11 behaviors and asks the informants to report on the frequency with which the child commits each behavior (online Supplementary Appendix I). This scale ranges from 1 (never) to 5 (nearly all the time).

Maternal-reported STAB data were available for 96.0% of the twins (α = 0.85), and paternal-reported STAB data were available for 80.9% of the twins (α = 0.84). The teachers of 115 participants were not available for assessment because the children were home-schooled or parental consents to contact the teachers were completed incorrectly. However, teacher-reported STAB data were available for 82.9% of the remaining twins (α = 0.91). To avoid shared informant variance with parent self-reports of his or her own behavior, teacher report of child social aggression served as our primary outcome variable. However, we also sought to confirm our primary findings by examining maternal and paternal reports of child social aggression. To adjust for positive skew, teacher ratings of twin social aggression were log-transformed prior to analysis to better approximate normality.

Parents each completed the self-report version of the STAB (α = 0.85 and 0.86 for mothers and fathers, respectively). In keeping with the assumptions of the NTF model, self-report data were omitted for parental figures who did not share 50% of their genes with the twins (e.g. grandparents and stepparents). Divorced or separated biological parents with joint custody arrangements or who were otherwise involved in their twins’ lives, however, were retained for analysis. Biological maternal and paternal self-report data were thus available for 93.9% and 76.6% of twin families, respectively.

Statistical analysis

The NTF model estimates the proportion of additive genetic (A), dominant genetic (D), shared sibling environment (S), shared familial environment (F), and non-shared environmental (E) influences on the variance in a phenotype. The CT model also estimates A, D, and E, but estimates broad shared environmental (C) influences rather than the more specific S and F. A and E are assumed to influence all phenotypes and are included in every model. Therefore, D, S, or F (in the NTF model) and C or D (in the CT model) must be fixed to zero, as there is not enough information in a given model to simultaneously estimate the effects of all parameters (Table 1).

Table 1.

Definitions of the parameters obtained via twin modeling

See also Keller et al. (Reference Keller, Medland and Duncan2010). Reprinted with permission from Burt et al. (Reference Burt, Donnellan and Tackett2012). CT, classical twin model; NTF, nuclear twin family model.

Mx, a structural-equation modeling program (Neale et al. Reference Neale, Boker, Xie and Maes2003), was used to perform the model-fitting analyses. Because of missing data, we made use of Full-Information Maximum-Likelihood (FIML) raw data techniques, which produce less biased and more efficient and consistent estimates than pairwise or listwise deletion in the face of missing data. When fitting models to raw data, variances, covariances, and means are first freely estimated to get a baseline index of fit (minus twice the log-likelihood; −2lnL). Observed covariance matrices were equivalent between same- and opposite-sex DZ twins (Box's M = 5.93, p = 0.436, ns), so all DZ twins were analyzed as one group. Model fit was evaluated using four information theoretic indices that balance overall fit with model parsimony: the Akaike's Information Criterion (AIC; Akaike, Reference Akaike1987), the Bayesian Information Criteria (BIC; Raftery, Reference Raftery1995), the sample-size adjusted Bayesian Information Criterion (SABIC; Sclove, Reference Sclove1987), and the Deviance Information Criterion (DIC; Spiegelhalter et al. Reference Spiegelhalter, Best, Carlin and Van Der Linde2002). The lowest AIC, BIC, SABIC, and DIC among a series of nested models is considered best. As fit indices do not always agree, we reasoned that the best-fitting model should yield lower or more negative values for at least three of the four fit indices.

Results

Descriptive statistics are presented in online Supplementary Table S1. Phenotypic and intraclass correlations are presented in Table 2. There was evidence of modest assortative mating, such that spouses were slightly more similar in their social aggression than would be expected by chance (r = 0.079, p < 0.01). Additionally, parents’ own social aggression was not correlated with their children's behavior as measured via teacher report (mother–child r = 0.51–0.56, p = 0.16–0.20, ns; father–child r = 0.013–0.095, p = 0.27–0.76, ns), but was moderately correlated when child behavior was measured via parental report (mother–child r = 0.126–0.487, p > 0.001; father–child r = 0.079–0.489, p > 0.001 to p = 0.008). Parent reports of twin social aggression were moderately correlated with one another (r = 0.322; p > 0.001), but were only weakly correlated with teacher report of child behavior (mother–teacher r = 0.169, p > 0.001; father–teacher r = 0.114, p = 0.003).

Table 2.

Phenotypic and intraclass correlations

ICCs are presented separately across zygosity (rMZ/rDZ). *p < 0.05, **p < 0.01.

Intraclass correlations within twin pairs, computed separately by degree of genetic relatedness, offer a preliminary indication of genetic and sibling environmental influences on social aggression. MZ intraclass correlations that are double those of DZ intraclass correlations are indicative of genetic effects, whereas MZ correlations that are less than double but still greater than DZ correlations suggest the importance of genetic and shared environmental effects. Teacher-reported MZ correlations of social aggression were almost double their corresponding DZ correlations, suggesting that genetic effects may be important for the etiology of social aggression. However, parent-reported MZ correlations were only marginally higher or equivalent to their corresponding DZ correlations, suggesting that, for those informant reports of social aggression, genetic effects are less prominent and/or that shared environmental effects are also important.

NTF model fitting results

Model-fitting results and parameter estimates are presented in Tables 3 and 4, respectively. According to all four fit indices, the best-fitting model for all informants was the ASFE model, indicating that additive genetic, sibling environmental, familial environmental, and non-shared environmental influences significantly contribute to the etiology of social aggression in middle childhood regardless of informant report, but that non-additive genetic influences do not. The ADSE model was also a good fit for teacher-reported social aggression, but this finding was not robust across alternative informants. In keeping with the familial environmental assumptions of the NTF model, a sensitivity analyses was conducted using only the families (N = 749; 72.7%) in which the twins’ biological parents were married to one another. Model-fitting results indicated that the ASFE model remained best-fitting model for all informants (online Supplementary Table S2).

Table 3.

Nuclear twin family design model-fitting results

Additive genetic, dominant genetic, sibling environmental, familial environmental, and non-shared environmental influences are represented with A, D, S, F, and E, respectively. The best-fitting model for each informant (as indicated by the lowest AIC, BIC, SABIC, and DIC values for at least three of the four fit indices) is highlighted in bold font.

Table 4.

Best-fitting nuclear twin family design parameter estimates

Additive genetic, dominant genetic, sibling environmental, familial environmental, and non-shared environmental influences are represented with A, D, S, F, and E, respectively. Ninety-five percent confidence intervals are presented below the point estimate in brackets. *p < 0.05.

Examinations of the parameter estimates revealed that social aggression is largely additive genetic (A = 0.14–0.77) and sibling environmental (S = 0.42–0.72) in origin. Familial environmental factors were estimated to be small but generally significant influences on social aggression (F = 0.08–0.28). Finally, the results suggest that the etiology of social aggression is not substantially influenced by assortative mating (μ = 0.07–0.09) or the covariance between additive genetic and familial environmental effects (e.g. passive rGE = −0.14 to 0.06). Although statistically significant according to some informants, the meager size of the passive rGE further implies that effects of the home environment are genuine and not instead a function of genetically influenced tendencies in disguise.

CT model results

We next fitted classical univariate models to visually compare the above NTF model results and, perhaps more importantly, to previous CT model findings concerning the etiology of social aggression. Model-fitting results and heritability estimates are presented in online Supplementary Tables S3 and S4, respectively. These analyses are not nested within those of the NTF model, so the model-fitting results from the two models cannot be directly compared. However, a comparison of the overall pattern of results is possible.

Unlike the NTF model, the best-fitting classical univariate twin model varied dramatically across informants. The ACE model was the best-fitting model for maternal-reported data, while the CE model was the best-fitting model for paternal-reported data. The ADE and AE models resulted in equivalent best-fitting models for teacher-reported social aggression. These findings contrast with not only each other, but also some of those reported in previous research on the etiology of social aggression. When examining the informant reports separately, the only common source of variance across all was non-shared environmental influences (which includes measurement error).