2.1. Method

Young adults (255 UCSB undergraduates recruited in classes and on campus) read an aggression-evoking scenario in which a classmate of the same sex as the participant is overheard at a party telling the participant's teaching assistant that the participant has not been working on their joint project and has shown up to project meetings with a hangover. Both accusations are false. In fact, the participant has been doing all the work because the classmate was vacationing in Baja California. The classmate was described as being loud, obnoxious, and as someone who makes useless comments in class. To minimize the physical threat posed by the classmate, he or she was described as being unusually short.

Participants then indicated how they would like to respond to the classmate's false accusations. Because behavioral studies of adults had often failed to find a sex difference in indirect aggression, we first measured retaliatory responses using a forced choice paradigm, which can detect small sex differences. Because responses are mutually exclusive, it is difficult to independently assess the psychology of indirect vs. physical aggression. We therefore also measured retaliatory sentiments using 10-point Likert scales that are less sensitive to sex differences, but that allow participants to freely rate multiple, nonmutually exclusive retaliatory responses. Here, participants rated how much they felt like making two informational and two physical retaliatory responses. This within-subjects variable (TACTIC) included measurements for GOSSIP, gossiping to other people at the party that the classmate “is clueless and spews useless comments during lecture,” which would harm the classmate by damaging his or her reputation; TELL, telling the teaching assistant that the classmate took a vacation, which would harm the classmate by possibly causing the teaching assistant to punish him or her; PUNCH, punching the classmate right now; and IFPUNCH, threatening the classmate that if he/she did that again, the participant would punch him or her. WEATHER, responding with an innocuous comment about the weather, assessed how inclined participants were to “turn the other cheek.” Though adult sex differences in physical aggression are well established (e.g., Archer, 2004), we included items on physical aggression as a partial check of the validity of our study.

We also measured variables designed to control for normative and other social influences on gossiping and violent behavior. Existing instruments designed for children (e.g., Huesmann & Guerra, 1997) include many items of the form, “It is okay for X to hit/scream at Y when Z,” where X and Y might be “a boy” and “a girl,” and Z is the situational context of the question, for example, a girl insults a boy. Some instruments also assess what third parties like peers or parents would think. We planned to expose participants to a particular situational context, so we decided to just ask about this situation. Because participants were college students, and because the norm theory strongly emphasizes the concept “appropriateness,” instead of using the word “okay,” we asked how “appropriate” it was to gossip (APGOSSIP), to tell (APTELL), or to punch (APPUNCH). The theory also emphasizes the attitudes of third parties, so we asked what impression people would have of the participant if he or she gossiped (IM1GOSSIP) or punched (IMPUNCH). These variables were measured on 10-point Likert scales. Telling on the classmate to the teaching assistant is not a public act, so we did not ask about impressions of telling; furthermore, we assumed that social norms on punching, and on threatening to punch, would be very similar, so we just asked about the appropriateness and impression of punching. Finally, we measured ANGER to check whether participants were reacting to our stimuli as intended (see Appendix A for stimuli). Analyses were simplified by having female participants read only about a female classmate, and males read only about a male. Thus, our study does not address cross-sex aggression.

3.1. Sex differences in norm variables

Although there were significant differences in the shapes and means of the male and female distributions of APPUNCH, the shapes of the male and female distributions of APGOSSIP and APTELL were remarkably similar (Table 1 and Fig. 1). Thus, there is little evidence for important sex differences in indirect aggression norms. Norm and approval variables did differ by aggression type, however. APPUNCH and IMPUNCH were highly correlated (rs=.71, p<.001) as were APGOSSIP and IMGOSSIP (rs=.43, p<.001), but the intercorrelations of other norm and approval variables were low-to-modest (rs ranged from .13 to .28). Norms on one type of aggression do not explain norms on other types.

Fig. 1. Densities (smoothed histograms). (A–C) Social norms; (D) PUNCH. Bandwidth=1. Sex differences in distributions were investigated using the Kolmogorov–Smirnov test (D), which can detect differences in the shapes of the distributions (results reported in each panel). Note the apparent bimodality of male APPUNCH and PUNCH scores.

3.2. Sex, norms, approval, and retaliation: forced choice

In the forced choice, 72 women (90%) felt more compelled to retaliate reputationally and only 8 (10%) felt more compelled to retaliate physically. Men were more evenly divided; 40 (55%) felt more compelled to retaliate reputationally, and 33 (45%) physically. The between-sex difference was significant [χ2(1)=24.1, =−.40, n=153, p<.001]. The within-sex difference was significant for women (p<.001), but not men (p=.48).

Using binomial logistic regression, we investigated whether SEX was still a significant predictor of retaliation type after controlling for norms and expectations of social approval. Simply presuming that each participant would retaliate reputationally would correctly predict 67.9% of the cases. A model with SEX alone significantly increased the accuracy of prediction to 73.8%, whereas a model of the social norm and approval variables alone correctly predicted 82.1% of cases (models not reported). Adding SEX to the social norm and approval model significantly improved it, correctly predicting 88.1% of cases (Table 2). Exp(B) is the effect of a 1-unit change in the predictor variable on the odds ratio of the response. Thus, being female increased the odds ratio of retaliating informationally 14.22 times relative to being male, controlling for the social norm and approval variables.

Table 2. Binomial logistic regression models of FORCED (phys=0 vs. rep=1)

		B	S.E.	Wald	p	Exp(B)
	Constant	−1.75	1.02	2.93	.087	0.17
	APGOSSIP	0.40	0.20	4.00	.046	1.49
	IMGOSSIP	0.93	0.31	8.96	.003	2.54
	APPUNCH	−0.79	0.21	14.46	<.001	0.46
	SEX	2.66	0.96	7.64	.006	14.22

n=84 (39 females, 45 males). The omnibus χ2(df)=59.1 (4), p<.001, Nagelkerke R2=.71. Adding SEX to a model containing only norm variables significantly improved the model (p for χ2 increase=.001).

3.3. Sex, norms, approval, and retaliation: free choice

Four retaliation sentiment items (TACTIC) allowed participants to independently rate their desire to RETALIATE using both informational and physical aggression. Our repeated measures design was unbalanced, so we used a linear mixed effects model (LME), with participant ID as a grouping factor. There was marked between-sex heterogeneity of variance, so we fit a heteroscedastic LME, with an exponential variance function. This model performed significantly better than an LME model, which assumed homogeneous variance. We had four planned between-sex comparisons, one for each retaliation sentiment.

There was a significant main effect for TACTIC (F=88.0, p<.001), no main effect for SEX (F=1.89, p=.17), and, as predicted, a significant interaction between SEX and TACTIC (F=15.0, p<.001). We then computed our four planned between-sex contrasts. As predicted, males experienced a significantly stronger desire than females on PUNCH (M=7.33 vs. 6.10, F=10.4, p=.001) and IFPUNCH (M=5.90 vs. 3.74, F=46.2, p<.001), and females experienced a significantly stronger desire on GOSSIP (M=6.62 vs. 5.56, F=4.7, p=.03). Contrary to predictions, females and males did not significantly differ on TELL (M=8.86 vs. 7.96, F=1.2, p=.27). The effect size for sex differences in GOSSIP was d=−0.39; 95% confidence interval (CI), −0.64 to −0.15.

We then explored post hoc within-sex comparisons. Within sex, women felt the strongest desire to retaliate by telling, followed, in order, by negative gossiping, punching, threatening to punch, and avoiding a confrontation (WEATHER). Men felt the strongest desire to retaliate by telling, followed, in order, by punching, threatening to punch, gossiping, and avoiding a confrontation. After adjusting alpha for multiple comparisons, females' desire to TELL (M=8.9) was significantly greater than their desire to GOSSIP (M=6.6), F=54.1, p<.001. Interestingly, despite their overwhelming choice of informational over physical retaliation in the forced choice paradigm, women still felt a strong desire to immediately PUNCH the classmate (M=6.1), a desire that was not significantly less than their desire to GOSSIP (M=6.6), F=1.27, p=.26. Men's desire to PUNCH the classmate (M=7.3), in contrast, was significantly higher than their desire to GOSSIP (M=5.6), F=14.5, p<.001.

Retaliation sentiments varied considerably among individuals. We tested whether our norm and approval variables explained a significant fraction of this variance, and whether these variables could account for all the effects of SEX. The marked asymmetry and heteroscedasticity of the residuals (Fig. 2) ruled out the straightforward use of traditional parametric models.

Fig. 2. Scatterplots of retaliation sentiments vs. social norm variables. Separate regression lines for each sex were fit by robust regression. The 95% CIs of the regression lines all excluded the origin (1, 1), but none excluded (10, 10) (tests not shown). A small amount of jitter has been added to better display overlapping points.

Leverage analyses also indicated that outliers might be having a modest impact on some models. We therefore used iterated reweighted least squares (IWLS) to fit a robust regression model with Tukey's bisquare M-estimator; this procedure is less sensitive to outliers than ordinary least-squares regression (Venables & Ripley, 2002). Because errors were nonnormal and heteroscedastic, and the asymptotic approximations used by this procedure to estimate S.E.'s may not be trustworthy in smaller samples (Fox, 1997), we used bootstrapping to estimate S.E.'s and CIs (Efron & Tibshirani, 1993). In our case, this procedure was conservative, yielding larger S.E.'s than traditional parametric tests (Table 3).

Table 3. Retaliation sentiments vs. social norms and SEX

	Model variables	N (f, m)	B	Bootstrapped values
				Bias	S.E.	CI (BCa)
	Model 1: GOSSIP	99, 89
	Intercept		2.93	−0.044	0.59	1.77 to 4.11
	APGOSSIP		0.50	0.010	0.10	0.31 to 0.67
	IMPGOSSIP		0.34	−0.003	0.13	0.11 to 0.60
	SEX (male)		−0.96	0.004	0.36	−1.62 to −0.29
	Model 2: TELL	140, 111
	Intercept		7.02	−0.014	0.79	5.27 to 8.34
	APPTELL		0.26	0.001	0.09	0.11 to 0.45
	SEX (male)		−3.28	0.039	1.09	−5.29 to −0.94
	APTELL*SEX		0.38	−0.004	0.12	0.13 to 0.60
	Model 3: PUNCH	37, 45
	Intercept		5.09	0.077	0.68	3.72 to 6.46
	APPUNCH		0.34	−0.004	0.10	0.16 to 0.56
	IMPUNCH		0.17	−0.022	0.15	−0.08 to 0.53
	SEX (male)		0.13	0.111	0.92	−1.83 to 1.85
	Model 4: IFPUNCH	37, 45
	Intercept		2.24	−0.057	0.47	1.47 to 3.18
	APPUNCH		0.71	0.006	0.15	0.34 to 0.93
	IMPUNCH		−0.02	0.011	0.17	−0.29 to 0.33
	SEX (male)		0.22	−0.026	0.67	−0.88 to 1.84

Regression coefficients estimated by robust regression. Standard errors and 95% CIs estimated with 1000 bootstrap replications. Significant values are in bold. SEX coefficients are treatment contrasts, which indicate the change from the base level (here, females). Hence, “SEX (male)” is the change for males relative to females. Interaction terms included only if significant.

The social norm variables were by far the strongest predictors of each of the retaliation variables. After controlling for them, however, there were still significant sex differences. Women were more inclined than men to retaliate informationally (GOSSIP and TELL). The significant interaction in model 2 between SEX and APTELL indicated that women's desire to tell was less influenced by the norm variable than was men's. After dichotomizing APTELL around the median, the mean desire to tell for women and men who thought it was appropriate to tell was essentially equal (female M=9.62 vs. male M=9.64), whereas the mean desire to tell for women who thought it was inappropriate to tell was higher (M=8.2) than it was for men (M=6.9); women who thought it inappropriate to tell still felt like telling. After controlling for norms, men were no more likely than women to retaliate, or threaten to retaliate, physically.

4.1. Conceptualizing retaliation norms: targets of, constraints on, or justifications for behavior?

Even though they did not explain sex differences in indirect aggression, social norm and approval variables nonetheless explained most of the variance in physical and informational retaliation sentiments. Hence, we explore them further here. There were three interesting results. First, the broad norm distributions (Fig. 1) were surprising because our population was relatively homogenous. (During the study, the student population was 63% white and 94% Californian; UCSB, 2000–2001.) Except for female APPUNCH scores, the only aspect of the norm data that supports the conception of social norms as uniform, society-wide pre- or proscriptions of behaviors is that each distribution had a more-or-less well-defined peak that might correspond to a society-wide norm. The considerable individual variance in norms might reflect strategic adoption and promotion of norms by individuals with varying interests (Nietzsche, 1887). Thus, the development of social norms regulating aggression may itself require an evolutionary strategic explanation (e.g., Hagen & Hammerstein, 2005). If high variance is found in domains other than aggression, this also has implications for cultural group selection models that emphasize the ability of conformism to generate low within-group variance relative to between-group variance in norms (e.g., Soltis, Boyd, & Richerson, 1995).

Second, the distributions of male PUNCH and APPUNCH scores appeared to be bimodal (Fig. 1). This evokes the Hawk–Dove game (Maynard-Smith, 1982), which is bimodal by construction. The relative densities of the two modes in the male distribution would then conceivably be evidence of the frequency dependence of Hawk and Dove (not necessarily based on genetic polymorphism). (That Hawk–Dove consists of only two strategies could be justified by arguing that in certain types of conflict, a “half-hearted” Hawk would incur the costs with little chance of gaining the benefits.) Male bimodality also evokes the work of Nisbett & Cohen (1996) on the “culture of honor,” which found that U.S. Southerners responded more strongly to reputational threats than Northerners. Perhaps these two peaks represent subcultures in our study population.

Third, although individuals might simply be trying to match their behavior to prevailing social norms, or conversely, expressing norms which conveniently match their desires, our data suggest that norms are imperfectly constraining an underlying disposition to retaliate. If norms were only guides to, or justifications for, emotions, then deviations from the norm should be biased neither for nor against an inclination to retaliate. We propose, on the other hand, that there is an underlying bias to retaliate, imperfectly constrained by a norm, so deviations from the norm should be in the direction of the underlying predispositions. In Fig. 2, the regression lines are all significantly tilted upward (i.e., 0<slope<1, tests not reported). Individuals who reported that punching was extremely inappropriate, for instance, were nonetheless expressing a significant desire to PUNCH. Above and beyond this upward tilt, the distribution of male (but not female) GOSSIP residuals were skewed toward retaliation–gossip desires that did not match the norm were biased toward increased gossiping. Bootstrapping revealed that this skew of male residuals (0.96) was significantly greater than 0 (95% CI, 0.46–1.46).

4.2. Study limitations

Participants were students aged 18 to 25 years; the results may not generalize to other age groups. In particular, 31% of our participants were 18 or 19 years old, just outside the 11–17 years age group showing the most sex differences in indirect aggression (Archer, 2004). As a predictor of informational attacks, however, AGE was not significant as a main effect or in interactions with other variables (tests not reported). Nonetheless, the sex differences we found might not exist in populations of older individuals. In addition, given the relatively modest effect of SEX, additional norm or approval variables might reduce or eliminate its residual effects in our models.

It is possible that sex differences in indirect aggression are explained, not by sex differences in social norms and approval regulating gossiping, but by sex differences in social norms and approval regulating physical aggression. Unfortunately, we could not test this hypothesis because including APPUNCH or IMPUNCH in the model of GOSSIP reduced the sample size by more than half, such that the original sex difference was no longer significant (and thus we could not test whether adding APPUNCH or IMPUNCH would eliminate the significance of SEX). This is another limitation of our study.