Violence exposure is associated with preference for masculine faces: evidence from Senegal

Abstract

It has been suggested that in threatening environments, both women and men should prefer more masculine men as romantic and coalition partners, respectively. Empirical evidence for this hypothesis has been weak or inconsistent, primarily because most experimental research has focused on elevating the perceived danger from other men through virtual threats. This study investigates whether personal experience with violence predicts the preference for masculine features in 326 Senegalese participants presented with pairs of manipulated facial photographs of West African men (one more feminine, one more masculine) and asked to indicate which face is more attractive (to women) or more trustworthy (to men). The findings reveal a strong association between violence exposure and facial feature preferences. Those who experienced (particularly physical) violence showed a higher preference for masculinized faces (up to 95% in women, 82% in men) compared to the baseline (57% in women, 63% in men). This difference is proposed to reflect an adaptive strategy of prioritizing physical protection in settings with a higher incidence of violent confrontations. Much of the variance can be found between groups. The direct effect of experienced violence diminishes over time, which suggests a dynamic interplay between innate predispositions and environmental influences on aesthetic preferences.

1. Introduction

The relationship between violence and preferences for masculine faces has been a subject of intensive psychological research. Masculine facial features, typically characterized by traits such as a strong jawline and prominent brow ridge, are often associated with perceived dominance and aggressiveness [1,2]. These perceptions can lead to stereotypical assumptions about an individual’s propensity for violent behaviour and group dominance. For example, individuals with more masculine facial features were perceived as more dominant and were more likely chosen as leaders in group tasks [3]. However, the link between male facial features and violent behaviour is complicated and affected by an array of individual differences and contextual factors [4–7]. Moreover, recent study suggests that the perception of dominance depends on study design and may not heavily rely on facial shape masculinity when faces exhibit variations across multiple dimensions simultaneously [8].

The physical strength and body size are other factors that may be reflected in both propensities to violent behaviour and facial appearance [9–12]. Morphometric studies on the facial cues of physical strength revealed that these are not always associated with the expected direction of sexual dimorphism [13,14] (i.e. higher strength with greater overall facial masculinity [15]). However, individuals with a wider lower face tend to exhibit higher handgrip strength, which indicates a consistent association between this specific feature and facial structure across various European, Asian and African populations [13,16,17]. Overall, facial masculinization appears to be a much more salient cue of formidability than, for instance, health or attractiveness [18].

Facial masculinity has been argued to correlate with testosterone levels [19], although the link with a circulating baseline of this hormone has not been clearly established [20]. Testosterone is associated with increased aggression and physical formidability [21], but also with reduced relationship commitment [22]. For women, the preference for higher testosterone and concomitant masculinity in men might therefore bring not only benefits of increased social dominance and physical performance in a partner but also potential costs in the resulting romantic relationship, including risk of aggression, dishonesty and low paternal investment [23–26]. Studies in Western cultures have produced inconsistent effects. Some showed that women prefer facial masculinity in men [27–29], while others reported a preference for relative facial femininity in male faces [1,30,31]. It was speculated that cross-cultural differences in masculinity preferences arise due to social (sex ratio, gender inequality) and environmental (pathogen prevalence, environmental predictability) factors [6,32–36]. In peaceful settings, women seem to show less concern for facial masculinization. When neither type of environmental harshness is present, masculinized faces are preferred over feminized ones at approximately 51.47% [18]. This baseline can be decreased by perceived pathogen prevalence and slightly increased by perceived resource scarcity. Exposure to cues of direct male competition, violence or displays of rival wealth effectively increases women’s preference for masculine faces [37]. This preference is further amplified when women are primed with self-protection threats [38]. Despite the soundness of the hypothesis that women should trade the lack of risk from masculine partners for higher physical security that results from their protection in dangerous environments, the reported effects are rather small or inconsistent. An elevated preference for feminized faces was observed in Colombian women who perceived a higher risk of experiencing violence [39]. This risk was captured by a factor loading on how safe women feel from violence in various places and the frequency of muggings and physical attacks they or people they know have experienced. Women who viewed men as potential threats to their children also rated more masculine faces as less attractive.

Relatively less attention yet has been given to the same-sex equivalent of this trade-off hypothesis. Men may prefer more masculine friends and allies during tough times [40], but these acquaintances could attract unnecessary conflicts when male-to-male competition is low. Masculine friends may also pose a source of elevated competition for mates and resources [41], making it potentially more advantageous to associate with more feminine individuals at times of low intergroup conflict. Among Dutch students playing the trust game (where one participant invests in another, whose ‘account’ triples the funds, and this second participant may send any sum back—or not), men with masculine faces were trusted more by women but trusted less by men [42]. Masculinization of male faces, however, did not influence the behaviour of Chinese participants in the same game [43]. While a moderate negative correlation (−0.33) was reported between trustworthiness and facial masculinity in twins from Australia and the United States [44], no significant effect was found in ratings or neural response linked to trust when looking at masculinized and feminized facial photographs in the same population [45]. (The trend suggested a higher trustworthiness of masculinized faces, but only the difference between bearded faces—regarded as more trustworthy in the study—and clean-shaven faces, irrespective of their masculinization, reached statistical significance.) A possible explanation for inconsistent results is that two distinct characteristics—warmth and competence—that contribute to trust are affected in opposite ways by facial masculinization [46]. Consequently, heterosexual men might perceive more masculine faces as more trustworthy and desirable for friendship (if not more attractive [39]) when such friends offer increased benefits. Elevated preference for masculine leaders during crises hints in this direction [47], although people tend to separate their preferences for leadership from those for friendship [48]. Trustworthiness, nevertheless, is a desirable feature of both a leader and a friend [49,50].

While some previous research showed that preference for facial masculinity is affected by various social and environmental factors [6,18,51], its direct relationship with exposure to violence has not been studied on African populations with the use of ecologically valid stimuli. Most of these studies exposed individuals from the affluent West to virtual threats, but the true size of the hypothesized effect can only be estimated in a population where the prevalence of violent interactions is such that self-reported violence experience can be treated as the predictor of preference for masculine faces. To bridge the gap in knowledge, we conducted a study in Senegal using manipulated West African faces, altering their facial masculinity to both enhance and decrease it. This allowed us to examine preferences, expressed in terms of attractiveness (rated by women) and trustworthiness (rated by men), for this trait in the context of exposure to different forms of violence.

2. Methods

(a). Stimuli preparation and manipulation of facial masculinity

Photographs of 102 West African faces (51 men, mean age: 23.75, s.d.: 5.49, range: 17−44; 51 women, mean age: 24.92, s.d.: 8.41, range: 17−54) selected from our database of standardized facial portraits were used to calculate the sex-relevant vector (responsible for maximum variation in sexual dimorphism in facial morphology) in TpsSuper v. 2.06 [52] software. The male photographs were subsequently manipulated along the sex-relevant vector by adding or subtracting 50% of the linear difference between male and female average configuration. Twenty random pairs of facial stimuli, each pair differing solely in facial shape, were selected as stimuli for this study. The presentation of stimulus pairs was completely randomized, as was the positioning of masculinized and feminized faces on either the right or left side of the card or screen.

(b). Participants and data collection

Participants were recruited through a coordinated field mission involving a Czech team supported by colleagues from the universities of Dakar and Ziguinchor. Recruitment took place in different parts of Senegal. In Dakar, participants from several linguistically diverse populations were included. In Oussouye, in the Ziguinchor region, the Diola were recruited, while in Ethiolo and Bandafassi in the Kédougou region, it was the Bassari and Bedik communities who were approached. In Walaldé and its neighbourhood, in the middle Senegal river valley, mainly the Fulani people were recruited.

At each site, local leaders played a crucial role in the recruitment process. They helped with translations where necessary to ensure clear communication with participants. In addition, an online version of the survey was shared to maximize coverage of the study’s target audience.

Each participant was shown 20 pairs of targets, with each pair comprising a feminized and masculinized version of the same male facial photograph. Male participants were asked to identify the more trustworthy face in the pair, while female participants selected the face, they found more attractive.

During the interview, respondents were asked whether they had ever experienced violence personally and if they had experienced violence in the past year. They could respond using a four-point scale: ‘Absolutely not’, ‘Probably not’, ‘Yes, probably’ and ‘Yes, absolutely’.

Subsequently, respondents were asked about the source of the violence (criminal, spouse, employer/superior, colleague, other) and the nature of the violence (verbal, physical, economic, administrative, sexual, other). They were supposed to provide information regarding the violence they endured in general. However, it is probable that many individuals who reported experiencing violence in the past year focused specifically on incidents from that period.

We collected a total of 338 responses. Of these, 318 were complete, seven lacked a single response and one was missing two responses. Twelve responses were discarded for being less than 85% complete. Among the remaining 326 participants, 144 were female (mean age: 32.00, s.d.: 12.72) and 182 were male (mean age: 32.78, s.d.: 10.52). These participants represented 12 different ethno-linguistic groups, with Wolof (99), Bedik (64) and Fulani (47) being the most common. A detailed summary of the sample by sex and group is provided in table 1.

Table 1.

Sample size by sex and ethno-linguistic group.

	Wolof	Bedik	Ful.	Diola	Ser.	Bas.	Man.	Bal.	Ewe	Léb.	Mau.	Mina	other
women	44	32	20	24	10	12	0	0	0	1	0	0	1
men	55	32	27	16	22	18	5	2	2	0	1	1	1
Ful. = Fulani, Ser. = Sereer, Bas. = Bassari, Man. = Mandinka, Bal. = Balante, Léb = Lébou, Mau. = Maure

Linguistically, all the examined groups except Mandinka (which is from the Mande branch) belong to the Atlantic branch of Niger-Congo family [53]. The origin of this phylum is clearly in western Africa [54] from where it spread eastwards. From the genomic point of view, all the groups have clear West African ancestry. However, in some Fulani populations, a north African or Eurasian ancestries have been detected [55,56]. Except for the Fulanis' language, called Fulfulde, which is spread throughout the Sahel/Savannah Belt of Africa, all other languages are geographically limited to the territory of Senegal.

(c). Statistical analysis

We analysed the preference for facial masculinization and feminization and its modification by experiences with violence.

The choice of the masculinized face was treated as a dependent Bernoulli-distributed variable. This variable was assigned a value of 1 when the masculinized face was selected as more trustworthy or attractive, and a value of 0 when the feminized version was chosen. Linear regression models were applied to the log odds of the probability $p$ that a masculinized face is selected. Consequently, the probability of selecting a feminized face is represented as $1-p$.

We employed multilevel models with varying effects for each target face because masculinization/feminization may suit each face differently and for each rater because baseline preferences for facial masculinization can systematically vary between individuals even when individual-level predictors are taken into account. This method allowed for the management of repeated measures and the estimation of variance between targets and raters. The simplest model included a fixed intercept to capture the effect of manipulation, along with two sets of varying intercepts to account for individual differences among faces and raters.

To evaluate the impact of violence experiences, we introduced an ordered-categorical predictor for the violence endured. The overall effect of violence ($b_V$) was modelled as the difference in responses between participants who answered ‘Absolutely not’ and those who responded ‘Yes, absolutely’ to experiencing violence. Additionally, three parameters ($d\left[1\right]$,$d\left[2\right]$,$d\left[3\right]$) arranged in a simplex (a vector whose elements sum to 1) were used to parcel the effect across the ordered categories from ‘Absolutely not’ to ‘Probably not’, from ‘Probably not’ to ‘Yes, probably’, and from ‘Yes, probably’ to ‘Yes, absolutely’.

We investigated the co-occurrence patterns of violence exposure, its nature and source using mosaic plots (in base R) and upset plots (via the UpSetR library) to identify relevant contrasts between types of violence. Since all types (except for the rare type ‘other’; see §3a) were associated with a similar distribution of reported violence experience, we covered only the distinction between violence requiring physical contact (physical and sexual) and all other types of violence in the main model. The distinction was incorporated as a contrast, assigned a value of $0.5$ for physical violence and $-0.5$ for non-physical violence. In participants with no (recent) violence experience, the contrast assumed a value of 0. Some studies (e.g. [39,57]) focus on physical violence, so the contrast also allows us to contextualize our results with that segment of literature.

The potential effect of violence should not be regarded as the total causal effect. While we hypothesize that women exposed to violence may prefer more masculine men for better protection against adversaries, it is important to note that highly masculine men can also pose a risk of violence possibly due to higher levels of testosterone and other androgens that facilitate aggressive behaviour. Therefore, women who favour masculine partners might be at an increased risk of experiencing violence. To distinguish cause from effect, we introduced a parameter ($b_S$) that contrasts the preferences of women who identified their spouse as the source of violence ($+0.5$) with those who did not ($-0.5$). The same parameter (if negative) also allows the model to deal with the potential effect reversal in women who react to domestic violence by decreasing the preference for masculine partners (as suggested by [58]). We also considered an equivalent model when analysing male evaluation of trustworthiness. The full formulation of the model as a set of equations is in electronic supplementary material, section S2. We have used Bayesian framework to estimate parameter values in our models. We used the rethinking package, which operates on the Stan computational framework. For each parameter set, we used vague, unbiased priors: normal distributions with a mean of 0 and standard deviation of 1 for intercepts and slopes, Dirichlet distribution with all shape parameters equal to 2 for the violence effect parcellation, and exponential distributions with a rate of 1 for standard deviations associated with targets and raters. Parameter estimates were derived from posterior samples. Percentile intervals of each parameter or prediction are reported alongside mean estimates as compatibility intervals (CIs).

To determine if the impact of violence on masculinity preferences decays over time, we compared models that treat exposure to violence at any point in life as equivalent to models that consider only past year exposure to violence. All models, one set for each rater sex, were compared for their expected out-of-sample performance using the widely applicable information criterion (WAIC).

To see whether the effect of experienced violence is robust to the inclusion of potentially confounding variables, we fitted several models with additional variables, namely, age (standardized continuous variable), marital status (fixed effect, each parameter treated as deviation from the baseline), nature (each violence nature contrasted against all others), source (each source contrasted against all others) and language group (fixed effect, each mother tongue, a close proxy for ethnic affiliation, fitted a separate intercept). More on these variables can be found in electronic supplementary material, section S6.

3. Results

(a). Descriptive analysis of violence

People responded consistently to the violence exposure questions. Among those who reported exposure to violence during the past year, 96% perceived themselves as having been exposed to violence at least once in their lifetime. The remaining 4% probably relativized the overall presence of violence when considering their entire life, leading us to decide against excluding them as unreliable respondents. Most individuals confident in not having been exposed to violence in the last year (54%) also perceived their entire lives as free from violence. Conversely, 56% of those who reported not being exposed to violence ‘probably’ in the last year responded ‘Yes, absolutely’ or ‘Yes, probably’ when asked about their overall exposure to violence. Kendall’s rank-order correlation between recent and overall violence exposure was 0.65. Complete results are in figure 1.

Figure 1.

Violence exposure, recent and overall. Per category percentages are in parentheses below axis labels, percentages in mosaic-plot cells correspond to frequencies within each column. Cell sizes show the overall frequency, cells with raw counts under 4 are not labelled.

Women experienced violence slightly more frequently than men, as shown in figure 1. Although this difference could not be reliably distinguished from random variation in overall violence exposure levels (with WAIC weights of models without and with the sex effect being 0.62 : 0.38 in favour of the simpler model, and the sex effect in the ordered-logit model being 0.2 with a 90% CI of −0.13 to 0.54), it was more pronounced in recent violence exposure. For recent violence exposure, WAIC weights favoured the model including the sex effect (0.20 : 0.80), where the slope was 0.41 with a 90% CI of 0.09–0.74.

There were negligible differences between the types of violence in reported overall experiences with violence. Participants were more likely to report definitive exposure to violence regardless of the type selected (electronic supplementary material, figure S1), except for the ‘Other’ category. For this category, responses of ‘Absolutely not’ and ‘Probably not’ were more common for overall exposure to violence.

The analysis of the frequency of violence types and sources showed that experiencing violence of various natures is common, whereas encountering violence from multiple sources is relatively rare. The most frequently identified source was ‘other’. Comments in the optional text field of the questionnaire clarified that in 100% of cases categorized as ‘other’, the source was a friend of the respondent. Consequently, we have chosen to rename this category to ‘Friend’ for future reference. The second most common source of violence was a colleague, followed by a spouse and employer. Only 18 individuals (7% of violence survivors) reported experiencing violence from a criminal.

The most common type of violence was verbal, followed by psychological and physical. Verbal and psychological violence often occurred together. The second most frequent overlap was between verbal and physical violence. The category labelled ‘Other’ rarely coincided with any other type (only with verbal violence in one instance), reinforcing its special status suggested by electronic supplementary material, figure S1. The results of the analysis on the nature and source of violence are in figure 2, the decomposition of natures and sources by sex are in electronic supplementary material, figure S3.

Figure 2.

Frequency of reported violence sources and natures including their combinations. See detailed decomposition by sex and by source-nature combination in electronic supplementary material, figures S2–S5.

By isolating cases with a single source and analysing the nature of violence within each category, we found that criminals predominantly engage in physical violence, whereas employers frequently use administrative and psychological violence. In spousal relationships, a combination of psychological and verbal abuse occurs more often than either type alone. Verbal and physical violence also co-occur relatively frequently. Friends mainly inflict psychological and verbal violence, but the combination of these types is less common, possibly because victims can more easily sever ties after an isolated incident in a non-romantic relationship. When ‘Friend’ was identified as the violence source, the nature categorized as ‘Other’ was chosen more frequently than in all other source categories combined, possibly indicating a specific ambiguity of the ‘Other’ type of violence in an unbalanced peer relationship. The detailed breakdown of the nature of violence by each unique source is available in electronic supplementary material, figure S2.

This analysis indicates that the nature and source variables are interdependent, and models that use predictors based on one but not the other could be misleading. This is because an isolated Nature variable is likely to capture the effect of Source and vice versa. Therefore, we only discuss models that incorporate both (violence requiring physical contact vs other types of violence, exposure to spousal violence vs no exposure to spousal violence) or neither of the suggested contrasts in the main article (see electronic supplement material, S6.3 for alternatives).

When analysing the source and nature of violence separately by sex, we see that women more often than men experience violence from a spouse, while men are more likely to encounter violence from colleagues. Sexual violence is more frequently directed at women, whereas it is the least common type of violence experienced by men (electronic supplementary material, figure S3). Women facing violence from a spouse are more likely to suffer from administrative or economic violence. Apart from these differences, the profiles of violence exposure are quite similar across sexes (electronic supplementary material, figures S4 and S5). These observed differences do not impact the fitted models and their comparisons, as data from both sexes—one focusing on perceived attractiveness and the other on trustworthiness—are never analysed together.

(b). Preference for facial masculinization

Masculinized versions of photographs were chosen as more attractive (selected in 77% of cases [90% CI: 71%, 82%]) and more trustworthy (72% [90% CI: 68%, 76%]) than their feminized counterparts. This significant deviation from a 50 : 50 chance is largely due to violence exposure, which tilts preferences towards masculine facial features. Models incorporating violence exposure effects demonstrate a better out-of-sample performance (table 2).

Table 2.

Model comparison.

women rating attractiveness
effects included in the model	WAIC	s.e.	dWAIC	ds.e.	pWAIC	weight
violence past year, physical (contrast), spouse (contrast)	2582.93	57.27	0.00	—	113.47	0.54
violence past year	2583.63	57.16	0.70	2.34	113.84	0.38
violence ever, physical (contrast), spouse (contrast)	2588.03	57.69	5.10	3.33	116.95	0.04
violence ever	2588.71	57.50	5.78	4.31	117.51	0.03
just baseline	2591.34	57.66	8.42	5.54	119.30	0.01

men rating trustworthiness
effects included in the model	WAIC	s.e.	dWAIC	ds.e.	pWAIC	weight
violence past year, physical (contrast), spouse (contrast)	3974.42	54.16	0.00	—	147.75	0.49
violence past year	3975.92	54.06	1.49	1.95	148.19	0.23
violence ever, physical (contrast), spouse (contrast)	3976.18	54.19	1.75	3.07	149.24	0.21
violence ever	3978.59	54.11	4.17	4.12	150.72	0.06
just baseline	3983.43	54.20	9.00	4.98	152.39	0.01
WAIC = Widely Applicable Information Criterion, s.e. = standard error, dWAIC = difference from the best model, ds.e. = standard error of the difference, pWAIC = penalty term (average variance in lppd)

Models focusing on recent episodes of violence are better than models that look at overall experience, indicating that violence may boost the preference for masculine partners or allies temporarily, with this effect decaying over time. (If all models are compared, including the extended ones from electronic supplementary material, section S6, this is unambiguously true for women. In men, the order sometimes flips (see electronic supplementary material, table S14). Models with predefined contrasts (figure 3, left column) in the nature and source outperform general models (figure 3, middle column) only slightly because all multilevel models effectively manage systematic differences between participants (see also the sampling variation statement in electronic supplementary material, section S5). Importantly, the inclusion of violence effects and contrasts reduces the unexplained variance among raters, illustrated by changes in the standard deviation parameter (${\sigma }_r$): for women, from a baseline of 1.96 to 1.55 in the model with past year’s violence, and to 1.52 in the model with contrasts; for men, from a baseline of 1.20 to 1.10 in the model with past year’s violence, and to 1.09 in the model with contrasts.

Figure 3.

Parameter estimates, marginal posterior distribution means (white points) and 90% compatibility intervals (error bars). Violence predicts an increased probability of selecting the masculinized version of the target photograph as more attractive (for women) and more trustworthy (for men). The largest difference is observed between individuals who respond ‘Yes, absolutely’ and ‘Yes, probably’ regarding their experience of violence. A large part of this effect can be attributed to differences between groups, as those with higher baseline violence levels also show a stronger preference for masculinized faces (see figure 4). The within-group effects are based on a model, in which each ethno-linguistic group is fitted with a separate fixed intercept (see electronic supplementary material, section S6.4 for the intercept estimates and other details).

Experience of violence was strongly associated with changes in attractiveness perception. Women who reported no experience of violence in the past year had a baseline probability of 57% [90% CI: 49%, 66%] for choosing masculinized faces. In contrast, women affirming definite violence exposure within the same period were expected to choose masculinized faces in 88% of cases [90% CI: 82%, 93%]. This likelihood increased to 95% [90% CI: 92%, 98%] if the violence involved physical contact. The most notable difference in preferences occurred between the ‘Yes, probably’ and ‘Yes, absolutely’ response categories.

In other models (women ranked by overall exposure to violence, men ranked by recent, and men ranked by overall exposure to violence), this pattern was replicated, albeit somewhat less pronounced. The entire posterior distribution of the violence effect parameter ($b_V$, with 99% CIs available in electronic supplementary material, section S4) is consistently positive in all these models.

The presence of physical contact in suffered violence did not enhance men’s likelihood of choosing masculinized faces. However, men were more likely to prefer masculinized faces if they reported experiencing violence from their spouse. A similar trend was observed in women when considering their overall experience of violence, aligning with the explanation that living with a more masculine partner may increase the probability of experiencing violence (an effect too weak to appear when only recent violence is considered).

Women rating attractiveness showed greater variability in individual preferences (${\sigma }_r$ = 1.52 [90% CI: 1.32, 1.74] top model from table 2) compared to men rating trustworthiness (${\sigma }_r$ = 1.09 [90% CI: 0.95, 1.24] top model from table 2). In the latter, systematic differences between targets—how feminization/masculinization suited their facial configuration for trustworthiness—were slightly more pronounced (${\sigma }_t$ = 0.29 [90% CI: 0.18, 0.41]) than in attractiveness ratings (${\sigma }_t$ = 0.18 [90% CI: 0.03, 0.32]), yet the variance among raters remained much larger. Parameter estimates from table 2’s top four models for each sex are visualized in figure 3. Predictions from the full joint posterior distribution are in figure 5. Additional predictions for less common and notable cases are in electronic supplementary material, figure S6. Complete posterior summaries, encompassing distribution shapes, parameter correlations, numerical expressions of means, 90%, and 99% CIs can be found in electronic supplementary material, section S3.

Figure 5.

Raw data and model predictions. The overall baseline (left column) is based on the simple multilevel model with no violence effect. All other predictions are derived from the FULL model focused on the RECENT (past year) violence. Black bars represent the proportion of masculinized face selections for each rater. Each plot is divided by a white vertical line, demarcating individuals who more often select feminized versions (to the left) from those preferring masculinized versions (to the right). The bright red numbers on the right of each plot denote the mean predictions, with adjacent lighter numbers outlining the 90% compatibility intervals.

Age and relationship status exert a slight influence over the probability of violence. Single status is associated with a lower violence probability, while being separated from a spouse predicts higher incidence of violence. When relationship status is accounted for, older women experience less recent violence on average (see electronic supplementary material, section S6.1 for details). These variables, however, do not contribute much to the prediction of masculinity preference (electronic supplementary material, section S6.2). Similarly, adding fine granularity to nature and source contrasts improves the models very little (electronic supplementary material, section S6.3) and keeps the estimated size of the total violence effect close to the values from main models (see electronic supplementary material, table S13). The only notable improvement of the expected out-of-sample fit is the inclusion of fixed intercepts, one per each ethno-linguistic group (electronic supplementary material, section S6.4). When fixed intercepts are included in the model with recent violence, s.d. among raters goes down to 1.29 [90% CI: 1.11, 1.48] in women and to 0.94 [90% CI: 0.82, 1.07] in men. The slopes from the multi-intercept model are included in figure 3 (right column) for comparison. The violence effect gets lower because a large proportion of differences in violence experience and masculinity preferences are stored between groups (figure 4; see also electronic supplementary material, figure S35). The importance of contrasts in these models is lower, because the groups differ substantially in them as well (electronic supplementary material, figure S36).

Figure 4.

Average violence experience and preference for masculinity by group. Sample sizes (proportional to point size) used to calculate group means are taken as weights in the weighted Pearson’s correlation coefficients, which are reported in the bottom left corner of each panel.

4. Discussion

We found that preference for masculine facial morphology covaries tightly with violence experience. Both personal experience and the overall exposure to violence at a group level may modify preferences for masculine facial features.

In men, the association between experience with violence and higher perceived trustworthiness of masculinized faces is largely driven by the Bedik group, which is an outlier in both respects. In women, the relationship is relatively strong even when accounting for between-group differences. Women with no experience of violence had a baseline probability for choosing masculinized face (57%) close to the Turkish baseline observed by Saribay et al. [18]. Violence endured during the past year predicts a baseline of 88% (or 95% if the violence is physical).

These results may indicate that the perception and processing of male facial masculinity evolved in human societies with oscillating degrees of interpersonal aggression. When physical dominance was used to solve conflicts between individuals, it paid to shift preferences towards mates (and possibly allies) that stood a chance to excel in such confrontations.

Why is so much variation currently explained by differences between groups rather than between individuals within these groups?

One reason may be that the differences in group-level baseline violence are just too big. When differences between groups are treated as part of the whole underlying distribution of raters, the estimated effect of violence is much greater (figure 3, left). If that were the only reason, however, the CI of the violence effect would be much wider in the multi-intercept model (figure 3, right) than in the model with a common baseline. However, notice that the CIs are similarly wide in both models (although the posterior correlation between the most extreme Bedik group intercept and the total violence effect is between −0.56 and −0.64 depending on the sample). It appears that the social climate within the group predicts preference for masculinity better than directly experienced violence.

The second reason, then, must be that even without experiencing violence personally, its high frequency among peers can modify individual preferences. Even the mere act of evoking heightened male-to-male competition by showing pictures of boxers instead of golfers can temporarily increase baseline preferences for masculine facial features [37]. Moreover, some of the most convincing research in the field was conducted on a group rather than individual level [32]. In Sub-Saharan Africa, inter-ethnic marriages are not frequent [59–62], people look for partners primarily within their own ethno-linguistic group, which probably keeps these differences pronounced and stable. Unfortunately, there is not much relevant data for all Senegalese populations involved in this study, but for the Bassari a low percentage of exogamous marriages have been observed [63]. This population is also known for regular rituals and ceremonies celebrated by different age classes (both male and female), which essentially consolidate the Bassari identity [64].

The strong total association detected in the common-baseline model (figure 3, left) could emerge from the repeated application of weaker causes that resemble the established within-group effects in magnitude (figure 3, right). An increase, even random and temporary, in the probability of experiencing or witnessing violence would elevate the prestige of more masculine individuals—both through increased coalition potential with other men and enhanced attractiveness to women. The rise in their relative bargaining power could, in turn, increase the level of violence in a society that begins to follow their ‘rules’. After a few rounds of these positive feedback loops (even without assuming direct fitness benefits for masculine individuals, which are likely and could further fuel emerging differences if the trends persist intergenerationally) groups may drift apart along the continuum of violence probability and preferred masculinity. It is difficult to overstate the role of domestic violence in the whole equation. The Bedik, in which both the probability of violence and baseline preference for masculinity are extreme, also shows a remarkably high incidence of spousal violence (electronic supplementary material, figure S35).

In a model with all potential confounds, the most likely within-group effect of experiencing violence in men drops to zero (electronic supplementary material, figure S34). Source and nature contrasts, relationship status (particularly the ‘married’ category), and group intercepts blur and split the original effect. However, it would be a mistake to dismiss the overall association just because most of it occurs between groups. In women, the effect persists in all models. Their preferences are probably more responsive to personally experienced violence. Because of this, they may play a key role in aligning cultures along the low-high masculinity preference continuum. That said, it’s still too early to treat this as a confirmed claim.

Even though we considered many possibilities for confounding and revealed that recent violence predicts preferences better than overall violence, some alternative explanations of the observed patterns cannot be ruled out without a longitudinal study or experimental manipulation. If important unmeasured variables influence the outcome—if, for instance, people who prefer masculine faces also better recall recent violence for some reason—the observed effects, including between-group differences, might be explained without a causal link from violence to preference. This limitation hampers all studies relying on retrospective self-reports, but we would argue that a causal explanation rooted in evolutionary psychology is more parsimonious.

The diversity in the costs and benefits associated with choosing masculine partners is, nevertheless, a complex matter. One may speculate about a trade-off between genetic quality and parental investment [23,65]. On the one hand, men with more masculine facial features are often perceived to possess a superior genetic quality [66], enhanced strength [67,68] and better health [28] (though contrary findings are becoming more abundant [69,70]). Moreover, in environments where competition among men is intense, traits associated with formidability may indicate a male’s capacity to acquire resources and safeguard kin [26]. On the other hand, more masculine individuals are seen as less reliable long-term partners [30,71]. Facial masculinity alone fails to predict increased reproductive success [72], so its fitness value, if any, is bound to be highly contextual. It is possible that even a single episode of violence may leave an epigenetic mark [73] that alters a female’s aesthetic inclinations in a contextually adaptive manner as our results suggest.

While the presence of contextual shifts in perceived attractiveness are robust in women, the support for the same-sex version of the trade-off optimization hypothesis is less strong. It is, however, still convincing at the inter-group level (mostly between Bedik and other groups). We can speculate that even in so-called WEIRD (Western, educated, industrialized, rich and democratic) societies, there will be subcultures where men honour masculinity and subcultures where they ridicule it. It should not come as a surprise if members of the former report more personal experiences with violence.

Discovered patterns of violence, its nature and sources constitute an important contribution in themselves, especially since such statistics from African societies are less abundant than equivalent measures from WEIRD countries. According to our findings, Senegalese women suffer most violence from friends and spouses, while men suffer most from their colleagues and friends. The vast majority of this violence is not easily attributable to criminality and physical acts. Indeed, violence is a universal problem that affects both developed and developing countries, with a higher risk for women and girls. In Senegal, the study by Leye et al. [74] for example reports high rates of violence against women, particularly within the marital relationship, where a variety of factors such as alcoholism, low socio-economic status and a history of personal violence are possible contributors to violence. According to the study of Werwie et al. [75], different types of violence are interconnected, often escalating from emotional to physical and sexual, and are influenced by economic considerations. It is likely that in such a context, facial masculinity does not primarily serve as a direct proxy for strength but rather as a social marker. It might have originally indicated the physical abilities of the bearer, but its role has shifted towards expressing social dominance in more masculine faces or, conversely, competence, openness and cooperativeness in more feminine faces.

An interesting revelation of our research is the suggested transient nature of altered facial preferences—the potential increase in the preference for masculinity seems to diminish over time. For each set of potentially confounding predictors, the past-year violence model predicts women’s preferences better (table 2; electronic supplementary material, table S14). For men, the dominance holds in basic models (table 2), but it is less stable in the epicycle models (electronic supplementary material, table S14) due to important involvement of (Bedik) group membership that does not change over time. The decay of preference shift highlights the dynamic interplay between innate predispositions and environmental influences, emphasizing the resilience and plasticity of human preferences.

Recognizing that experiences of violence can shape preference patterns highlights the need for timely and tailored interventions to support individuals in recovery from violence. Psychological support may help mitigate the lasting impact on victims’ perceptions and interactions.

Our results also argue against downplaying non-physical forms of violence. Notably, while the within-group effect of violence remains salient (especially in women), the within-group effect of the physical nature of violence becomes inconclusive (figure 3, right column). This does not imply it is irrelevant overall (figure 3, left column); it may simply be masked by differences in the occurrence of physical violence between the groups. At least in our sample, however, the distinction between physical and non-physical violence seems to be less pronounced than the threatening atmosphere created by violence of any form, including verbal and psychological (however, see the distribution of the nature of violence in Bassari, electronic supplementary material, figure S35). It is important to note that non-physical violence is also a source of stress (with potential fitness implications) that may be avoided if one’s romantic partner or ally is perceived as more masculine (i.e. a potentially more dangerous foe).

It remains an open question whether we would obtain the same results in other cultures. Until now, most of the psychological research has been based on Euro-American respondents. Only recently has it been extending to other populations [5,76–81], whose attractiveness preferences and other psychometric characteristics remain largely unexplored, especially in tribal areas. This work hopes to contribute to this favourable shift.

Ethics

The study was performed in line with the principles of the Declaration of Helsinki. The data collection process was approved by the Institutional Review Board of the Faculty of Science of Charles University (protocol ref. number 2023/06). This article and the shared data do not include information or images that could lead to the identification of participants. All participants were over 15 years of age (mean age: 32.43, s.d.: 11.51, range: 16−80) and gave their informed consent to take part in the study.

Data accessibility

All data and code that allow to replicate the analysis are accessible at [82].

Supplementary material is available online [83].

Declaration of AI use

During the preparation of this work, the authors used Chat GPT-4 and Claude 3.5 Sonnet in order to improve grammar and style. They used GPT-3.5-based technology of GitHub copilot to speed up the coding. After using this technology, the authors reviewed and edited the text or code as needed and took full responsibility for the content of the publication.

Authors’ contributions

P.T.: formal analysis, investigation, methodology, software, visualization, writing—original draft, writing—review and editing; V.Č.: conceptualization, supervision, writing—review and editing; M.Y.D.: data curation, writing—review and editing; N.C.: data curation, writing—review and editing; Š.P.: data curation, writing—review and editing; K.K.: conceptualization, data curation, funding acquisition, investigation, methodology, resources, supervision, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

This research was supported by the Czech Science Foundation project reg. no 24-11735S. VČ was funded by Czech Academy of Sciences award Praemium Academiae.