Disentangling facial masculinity/femininity from gender category information

To systematically investigate the interplay of category-based gender stereotyping and feature-trait associations we first had to disentangle facial masculinity/femininity from gender category information. Every face has more or less unique facial characteristics. Some of these characteristics can be described in terms of single features (e.g., full lips, a roundish chin), whereas others can be described in terms of the configuration of features (e.g., close-set eyes). There is a lot of evidence in the face perception literature that featural information is more relevant for the categorization of individuals as either male or female, whereas configural information is more relevant to identify a person [4,13]. However, there is also evidence that facial information alone is often not sufficient for gender categorization [14,15]. Especially women lacking extra-facial features, such as hairstyle, jewelry, or clothing tend to be misclassified as men. Often extra-facial styling information is needed to classify a person as either male or female. Among these, the hairstyle has been shown to be the best category-cueing feature [4,13].

Whereas extra-facial hairstyle information is a potent cue for gender categorization, inner facial masculinity/femininity has been shown to impact spontaneous impression formation processes (e.g., [16]). Different approaches have been developed to systematically model masculinity/femininity in faces in the last 20 years [15,17,18]. Using one of these approaches, namely the morphable model approach [17], we calculated a masculinity/femininity vector by determining the direction between the average male and the average female face within a statistical face space based on 100 male and 100 female faces [19]. This method is fully data-driven or bottom-up in the sense that the masculinity/femininity vector depicts the information that discriminates between the average male and female face [20]. Because both features (size of eyes, fullness of mouth) as well as configuration (proximity among the features) differ between male and female faces, our conceptualization of inner-facial masculinity/femininity consists of both featural and configural information. This information–inner-facial masculinity/femininity–was then systematically modeled in the stimulus persons’ faces to make them appear more or less masculine or feminine [16,17,21]. In this paper we–for the first time–modeled facial masculinity/femininity using up-to date computer graphics methods to manipulate real face photographs in a controlled way and with natural-looking results [20]. We created feminine- and masculine-looking versions of the same unambiguously male and female target persons to systematically investigate the interplay of extra-facial category cueing information and masculinity/femininity in direct comparison.

How gender category and gendered facial information impact personality inferences

From a mere category-based gender stereotyping perspective, one would expect that based on extra-facial category-cueing information (i.e., hairstyle) individuals were classified as males or females. By classifying a person as male or female, an association between the person and a gender category is established. This process likely activates the content of the respective gender stereotype that subsequently can be used to draw inferences about the person [22]. As a result, the judgment of the person will be assimilated towards the stereotype. Consequently, category-based gender stereotyping would predict that persons classified as men are ascribed stereotypically masculine personality traits, whereas persons classified as women are ascribed stereotypically feminine personality traits. In line with the Stereotype Content Model we will refer to these stereotypical traits as warmth (stereotypically feminine traits) and competence (stereotypically masculine traits), respectively [23].

From the perspective of feature-trait associations, one would expect that the facial masculinity/femininity–independent of category membership–was directly linked to the perception of stereotypically masculine or feminine personality traits. (Note that the term feature-trait associations refers to the bottom-up process of directly linking facial information with personality traits [11]; it does not imply that configural facial information is not relevant for this process.) These expectations find support by research from two domains, namely the domain of cultural stereotyping and the domain of impression formation from faces. On the one hand, evidence from research on cultural stereotyping has shown that the extent of Afrocentric features in a target was positively correlated with judgments stereotypical to African-Americans [11,24,25]. This was true for African-American and White targets indicating that facial features affected the judgment independent of ethnic category. Most importantly, the effect was not mediated by category accessibility [24]. On the other hand, evidence from face-based person perception research revealed that when seeing a stranger, individuals automatically locate him or her in a 2D coordinate system spanned by the dimensions trustworthiness and dominance [8]. Work on spontaneous personality inferences from faces has shown associations between facial femininity and trustworthiness as well as facial masculinity and dominance [8]. Consequently, there is a lot of evidence for the feature-trait association account predicting that masculine-looking persons are spontaneously ascribed more competence and less warmth than feminine-looking persons.

Whereas both the category-based stereotyping and the feature-trait association account make identical predictions for feminine-looking women and masculine-looking men (typical exemplars), they make opposite predictions for masculine-looking women and feminine-looking men (atypical exemplars). For typical exemplars both models predict that masculine-looking men are ascribed more competence and less warmth than feminine-looking women either because they are included in the respective gender category and consequently assimilated towards that gender category or because of strong associations between their individual facial features and the personality dimensions trustworthiness and dominance. For atypical exemplars, the implications of category membership and individual features point in opposite directions. If the judgment is derived from the category membership, assimilation to the gender category is expected. Thus, feminine-looking men would be ascribed more competence and less warmth than masculine-looking women. If, however, facial features would be directly linked to the respective personality dimensions, feminine-looking men would be ascribed less competence and more warmth than masculine-looking women.

When comparing the two models there are at least two reasons to believe that if both category and feature information is available, feature-trait associations should override category-trait associations. First, there is evidence that stereotypes are not likely to affect personality judgments, if diagnostic individuating information about the target is available [26]. And we know from the ecological theory of face perception (e.g., [10]) that individuals regard certain facial information (e.g., the information associated with gender) as diagnostic for certain behavioral affordances and personality traits. Because the facial appearance of a target person provides a perceiver with more detailed, concrete, and individuating information about a target person than mere category membership and because perceivers rely on facial information to build first impressions, feature information is likely to outweigh category information in the process of impression formation.

Second, although both category-based stereotyping and direct feature-trait associations often work implicitly and automatically (e.g., [24] for feature-trait associations; e.g., [27] for stereotyping), individuals might be more aware that they have a tendency to simplify the world around them by classifying individuals into categories and use stereotypic information about these categories to describe the individuals than they are aware that they use facial information to infer personality information. Therefore, they are more likely to control for the effects of category information than for the effects of feature information, particularly if cues are subtle [25].

Importantly, however, a feature-based account does not necessarily imply that the respective category will be ignored and has no influence. It has been shown that even gender-atypical faces of men and women activate the matching gender category [28]. Participants in a mouse tracking study had to associate typical and atypical men and women with either the male or the female gender stereotype. Although mouse trajectories for masculine-looking women and feminine-looking men showed more attraction towards the male and female gender stereotype respectively, as compared to the trajectories for feminine-looking women and masculine-looking men, female faces were generally associated with the female stereotype, male faces with the male stereotype [28].

How does this activated category information impact further processing of the stimulus? Research suggests that the category may provide a frame of reference for the judgment [29]; for a review, see [30]. The inclusion/exclusion model [30,31] holds that the target will be judged with reference to the category’s standard by being compared to the category prototype. Atypical members of a category are likely to be excluded from the gender category. A masculine-looking woman, for example, would not only spontaneously activate the concept of competence, but also the female gender category. Because of her facial appearance, she would be excluded from that category, for example, by subtyping her into the category of career woman [32]. This in turn would result in a contrast or a shift away from the stereotypical woman, resulting in more pronounced ascriptions of competence or stereotypically masculine personality traits.

To summarize, for atypical category exemplars (i.e., masculine-looking women and feminine-looking men) two processes are likely to co-occur: (a) feature-trait associations lead to perceptions of high levels of competence and low levels of warmth in women and vice versa for men and (b) gender category is used as a reference against which the atypical exemplar is contrasted, also leading to perceptions of high levels of competence and low levels of warmth in women and vice versa for men. Because both processes are likely to work simultaneously, we expect stronger effects for atypical than for typical exemplars. The masculine facial appearances of a woman and of a man elicit the same level of competence (and warmth) inferences due to feature-based inferences, but the woman may be judged as even more competent (and colder) than the man because she is additionally contrasted against the female prototype. From these theoretical assumptions, the following hypotheses ensue:

First, due to feature-trait associations we assume that overall masculine-looking persons are perceived as more competent and less warm than feminine-looking persons (Hypothesis 1) [11,24,25]. Second, because feature-trait associations likely override category-based stereotyping we hypothesize that masculine-looking women appear more competent and less warm than feminine-looking men (Hypothesis 2) [25]. Third, because atypical exemplars are expected to be additionally contrasted against their group prototype, the effects of facial masculinity/femininity appear more pronounced for atypical than for typical group members (Hypothesis 3) [30,31,33].

The present studies

In both of the studies presented here we tested the three above-mentioned hypotheses by manipulating category-cueing information and facial masculinity/femininity orthogonal to each other. Participants had to judge two target persons regarding stereotypically masculine (i.e., competence) and stereotypically feminine (i.e, warmth) personality traits based on their portraits in a sequence. Because different faces are associated with different personalities, which produces considerable confound when using real face photographs, we used a data-driven statistical face model [16,17,19,21] to systematically manipulate the facial masculinity/femininity within the same faces. This approach allowed us to avoid confound and to disentangle facial masculinity/femininity information from the person and thereby also from the gender category. Using such a data-driven modeling approach [16,21] had another advantage: We did not have to a priori define which characteristics–be they featural or configural–make a face appear more masculine or feminine. This information was directly extracted from the face model.

This approach to investigate the impact of facial masculinity/femininity and gender category information is novel for different reasons. Overall, little work has been done separating category and feature information. The work that has been done was either based on material that was controlled but artificial looking or natural looking but confounded. On the one hand, Little and colleagues, for example, used very controlled stimuli with systematic variations in masculinity/femininity. However, their stimuli lack extra-facial information and thus appear somewhat artificial (e.g., [18,34]). On the other hand, Blair and colleagues, for example, in the seminal work on the role of Afrocentric facial features on criminal sentencing [11,24] used real face photographs and measured the degree of Afrocentricity. Not surprisingly, there was a considerable overlap between Afrocentric facial features and ethnic category. And potentially, the facial appearance dimension of interest (i.e., Afrocentricity) was confounded with other facial appearance information of the individual stimulus persons used.

By using novel image manipulation techniques, we were able to use facial stimuli that are at the same time controlled and natural looking. To the best of our knowledge the present studies are the first that unambiguously separate gender category and gendered facial appearance information in studying their effects on explicit ascriptions of gender-stereotypical personality traits.

Ethics statement

The following studies were conducted in full accordance with the Ethical Guidelines of the Swiss Psychological Society (SGP-SSP) and the American Psychological Association (APA). At the time of data acquisition (before 2014) it was not possible at the University of Basel to seek ethics approval for survey studies. We did not collect any sensitive data. All questionnaires in the study were anonymous questionnaires. Moreover, we obtained consent from all participants and participants could easily withdraw from the study at any time by closing the Internet browser. After the establishment of the IRB at the Department of Psychology at the University of Basel (2014) we applied for and received IRB approval for the paradigms employed in these studies (IRB approval No. 034-15-3).

Method

Material.

The first independent variable, namely target persons’ gender was manipulated by preselecting three male and three female faces from the FERET database [37]. The primary preselection criterion was that the three male persons had a short, masculine hairstyle and that the three female persons had a long, feminine hairstyle to unambiguously trigger the activation of the male and female gender category, respectively. Secondary preselection criteria were closed mouths, open eyes, direct gaze, neutral facial expression, and a lack of facial hair or glasses. These additional criteria were accounted for by the manipulation procedure described in the next section.

The second independent variable, namely facial masculinity/femininity was manipulated using a data-driven statistical modeling approach. First, the preselected faces were analyzed by actively synthesizing them with a statistical face model [19] resulting in estimations of the head structure and the texture information of the three male and three female faces. Then we generated a masculinity/femininity vector by first averaging all 100 male and all 100 female faces in our face space, respectively, and then computing the vector pointing from the average male to the average female face. This masculinity/femininity vector depicts all information, featural (e.g., bigger eyes in females, more pronounced chins in males) and configural (e.g., single features are closer to each other in female faces), that make a face appear more masculine/feminine. This masculinity/femininity vector was then used to systematically enhance and reduce facial masculinity/femininity in the six preselected faces, resulting in two new versions of all of them, namely a more masculine- and a more feminine-looking version. The resulting heads were rendered back into the original photographs, showing natural-looking variations of masculinity/femininity in all six faces (see Fig 1). Note that all faces were manipulated in the exact same way, because the same kind and the same amount of information were modeled in all of them in a very controlled way. This procedure resulted in twelve different stimuli, namely a masculine- and a feminine-looking version of all six preselected faces. For a more detailed description of the image manipulation technique, see [16,21]; for data demonstrating that these face models can successfully be transferred to novel faces even if these faces differ strongly from the faces on which the models were developed, see [38].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Example of stimulus material used in both studies.

Note that the gender category can always clearly be inferred from the extra-facial hairstyle information, whereas the masculinity/femininity of the facial appearance is determined from inner facial featural and configural information. From left to right: Masculine-looking man, feminine-looking man, masculine-looking woman, and feminine-looking woman. The original faces (not depicted here) stem from the FERET database [37].

https://doi.org/10.1371/journal.pone.0181306.g001

The twelve faces were combined in pairs of two faces each with two constraints: Facial masculinity/femininity information should be contrary and gender category information should be the same. This procedure resulted in six pairs of faces for each gender category (e.g., pair 1: male face 1 masculine/male face 2 feminine, pair 2: male face 1 masculine/male face 3 feminine, pair 3: male face 2 masculine/male face 3 feminine, pair 4: male face 1 feminine/male face 2 masculine, pair 5: male face 1 feminine/male face 3 masculine, pair 6: male face 2 feminine/male face 3 masculine).

In order to assess gender stereotypical personality inferences, we used a German version [39] of the Bem Sex Role Inventory (BSRI [40]). This questionnaire consists of 20 items measuring stereotypically masculine (e.g., leadership qualities, assertive, authoritative, resolute, confident; referred to as competent below), stereotypically feminine (e.g., softhearted, sensitive, romantic, sentimental, sincere; referred to as warm below), and gender-neutral personality traits (e.g., nervous, healthy, sad, conscientious, oblivious) each. This questionnaire has been shown to successfully measure gender stereotypes (e.g., [41,42]). The gender-neutral personality traits were used as filler items in order not to make the target of our study too obvious. The items were reframed to assess personality in the third instead of the first person perspective (i.e., “The depicted person has leadership qualities”, “The depicted person is softhearted”). These traits were rated on a 7-point Likert Scale ranging from 1 (I strongly disagree) to 7 (I strongly agree).

Results

First, a masculinity/competence and a femininity/warmth score were built by averaging all the respective items from the BSRI. Both scores showed a high internal consistency with a Cronbach’s alpha of .94 for the masculinity scale and a Cronbach’s alpha of .89 for the femininity scale. Because each participant had to evaluate both masculinity/competence and femininity/warmth of all faces, we treat these two dimensions as two levels of a within-subjects factor personality dimension in what follows.

We analyzed our data using linear mixed models analyses including random effects for participants and faces. The advantage of this method is that it allows the generalization of results not only across participants, but also across faces. Consequently, this method allows to make inferences to future studies with different samples of participants and faces and enhances replicability of the findings [43]. To test Hypothesis 1 we used the lme4 [44] package in R [45] to fit a mixed linear model with the following specifications to our data: Facial appearance (masculine-looking vs. feminine-looking; within-participants), target gender (male vs. female; between-participants), personality dimension (competence vs. warmth; within-participants), all three possible 2-way interactions, and the three-way interaction were included into the model as fixed effects, while both participants and faces were included as random effects. We aimed for a maximal linear mixed model because such models generalize best across participants and stimuli [46]. Therefore, we included a random intercept (i.e., the model allows the intercept to vary individually) and a random slope for the main effect facial appearance based on participants (i.e., the model allows the facial appearance to individually affect different participants’ judgments) and a random intercept and random slopes for facial appearance, personality dimension, and the facial appearance x personality dimension interaction based on faces (i.e., the model allows the intercept to vary individually and the facial appearance, personality dimension, and the interaction between the latter two to individually affect different faces). The reason for only including the facial appearance random slopes for participants is that we had one observation per participant for each combination of fixed factor levels [46]. Table 1 shows the explained variance by the random factors specified above.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Variance explained by the random effects for participants and faces in Study 1.

https://doi.org/10.1371/journal.pone.0181306.t001

To calculate F, and p values as in a classical linear model, we additionally used the lmerTest package [47] in R [45]. Results of the analysis with the dependent variable ascribed personality revealed one significant effect, namely the two-way interaction between facial appearance and ascribed personality, F(1, 3.97) = 68.42, p = .001. Uneven degrees of freedom result from Satterthwaite approximations. We do not present effect sizes, because no standard has been established, yet on how to include the multiple variance components from the random intercepts and slopes in the model.

Supporting our first hypothesis, masculine-looking target persons were ascribed more competence than feminine-looking target persons, F(1, 3.94) = 36.07, p = .004, whereas feminine-looking target persons were ascribed more warmth than masculine-looking target persons, F(1, 3.99) = 103.47, p < .001 (adjusted means and standard errors are presented in Table 2.). For exploratory reasons we also ran an analysis with the additional fixed factor participant gender. This analysis did not reveal any main or interaction effect including participant gender, F_max(1, 1168.99) = 1.54, p = .215.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Means (standard deviations) for the dependent variables ascribed competence and warmth depending on target persons’ gender and facial appearance in Study 1.

The higher the mean values, the more the stimulus persons are perceived as competent and warm, respectively.

https://doi.org/10.1371/journal.pone.0181306.t002

To test Hypotheses 2 and 3, we built a model with the three fixed factors ascribed personality, target gender, typicality, all three two-way interactions, and the three-way interaction effects between these as well as random intercepts for participants and random intercepts and random slopes for the main effect personality dimension based on faces. We defined the factor typicality so that masculine-looking males and feminine-looking females were coded as typical, whereas feminine-looking males and masculine-looking females were coded as atypical. Results revealed a significant main effect of typicality, F(1, 1182.21) = 5.07, p = .025. This effect, however, was qualified by a significant three way-interaction, F(1, 1181.71) = 281.621, p < .001, signaling that ascriptions of personality significantly differs between typical and atypical category exemplars. Supporting Hypothesis 2, results revealed significant effects of feature-trait associations for typical and atypical exemplars: For typical exemplars we found that masculine-looking men were ascribed more competence and less warmth than feminine-looking women, F(1, 4.05) = 16.32, p = .015. For atypical exemplars we also found that masculine looking faces (i.e., masculine looking women) were ascribed more competence and less warmth than feminine faces (i.e., feminine-looking men), F(1, 4.04) = 30.18, p = .005. However, in line with Hypothesis 3, the higher F value for the feature effect in the atypical than in the typical condition combined with the significant 3-way interaction reveals that the effect of facial features was much larger for atypical than for typical exemplars. Whereas feature-trait associations affect both typical and atypical category members, an additional mechanism is at work for atypical members, which likely is a contrast effect to the category prototype. The impact of target persons’ facial appearance and gender on the ascription of competence and warmth is illustrated in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Impact of facial appearance and gender information on the ascription of competence and warmth measured with the BSRI for typical and atypical category members in Study 1.

Error bars represent standard errors of the mean.

https://doi.org/10.1371/journal.pone.0181306.g002

To further investigate whether the results found followed from a direct comparison between the person in the first and second trial (i.e., a feminine-looking (wo)man presented after a masculine-looking (wo)man and vice versa) we analyzed the two trials separately. We therefore specified two models with the fixed factors facial appearance (masculine-looking vs. feminine-looking; between-participants, because we only analyzed one trial at a time), target gender (male vs. female; between-participants), personality dimension (competence vs. warmth; within-participants), all three possible two-way interactions, and the three-way interaction for both trials separately. Again, we included random intercepts for participants and stimuli as well as random slopes for facial appearance and gender for the different faces. For both trials separately, we could replicate the finding that masculine-looking persons were ascribed more competence and less warmth than feminine-looking persons, F_Trial1(1, 584.61) = 100.51, p < .001 and F_Trial2(1, 585.43) = 201.43, p < .001, respectively. We also specified the linear mixed models with the three fixed factors facial appearance, typicality, and ascribed personality and all four possible interaction effects as well as the random effects specified above. The three-way interaction between masculinity/femininity, typicality, and ascribed personality did not reach significance in any of the two trials, F_Trial1(1, 4.05) = 1.87, p = .242 and F_Trial2(1, 4) < 1, respectively.

Discussion

In line with our first hypothesis, masculine-looking target persons were ascribed more competence and less warmth than feminine-looking target persons. This effect of facial masculinity/femininity was present both in the first and second trial per participant, showing that this effect does not require a direct comparison of different persons. Moreover, we found this feature-effect both for the typical and for the atypical category members, supporting our second hypothesis. Interestingly, in line with our third hypothesis, the effect was even more pronounced for atypical than for typical category members.

These findings extend results by Freeman and Ambady [28] in two ways. First, they do not only show that atypical category members activate both gender stereotypes, but demonstrate that the stereotype activated by facial appearance is reflected in explicit personality judgments. Second, they suggest that in contrast to the stereotype activated by facial features there is no evidence for category-based stereotyping. The implications of the gender category did not weaken those of facial appearance for atypical exemplars, instead, the degree to which masculine-looking women are ascribed more competence and less warmth than feminine-looking men is higher than the degree in which masculine-looking men are ascribed more competence and less warmth than feminine-looking women. These results are in line with predictions from the inclusion/exclusion model [31] suggesting that atypical category members are contrasted against their group prototype on top of the feature effect. This finding indicates that the activated gender category is used as a frame of reference creating a stereotype reversal.

It is possible that the design of the first study may have worked against assimilation effects based on gender category. The fact that all judgments involved either only women or only men may have made the gender category less salient as a base for inferences. The first target may have rendered within-category comparisons particularly likely for the second target. In order to test the generalizability of the results we changed the design in Study 2.

Method

Results

Again, we first calculated a competence and a warmth scale by averaging the ten items (i.e., five from the BSRI and five from the KMGI) of the respective dimension. Both scales showed a high internal consistency (Cronbach’s alpha_Masc = .91 and Cronbach’s alpha_Fem = .88).

As in Study 1, we fitted a mixed linear model with ascribed personality as the dependent variable using the lme4 [44] and the lmerTest package [47] in R [45]. Facial appearance (masculine-looking vs. feminine-looking; within-participants), typicality of category members (typical vs. atypical; between-participants), ascribed personality (competence vs. warmth; within-participants), all three two-way interactions, and the three-way interaction were included into the model as fixed effects, while both participants and faces were included as random effects. Thereby this model accounts for sampling variability both regarding participants and faces [43]. Aiming for a maximal linear mixed model [46] we included random intercepts and a random slope for the main effect facial appearance based on participants and faces. We did not include the interaction term between facial appearance and personality dimension, because the respective model did not converge. To make sure that results are nevertheless reliable we also fitted a model with random intercepts and slopes for personality dimension for participants and faces. Neither the significance levels nor the pattern of results changed. Table 3 shows the explained variance by the random factors in this model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Variance explained by the random effects for participants and faces in Study 2.

https://doi.org/10.1371/journal.pone.0181306.t003

Supporting our first hypothesis and thereby replicating Study 1, we found a two way interaction between facial appearance and personality dimension, showing that masculine-looking persons are ascribed more competence and less warmth than feminine-looking persons, F(1, 286.34) = 47.91, p < .001 (see Table 4 for adjusted means and standard errors.). For exploratory reasons we also ran an analysis with the additional fixed factor participant gender. Results revealed one significant interaction effect with the factor participant gender, namely a typicality-by-participant gender interaction, F(1, 93.39) = 5.28, p = .024. On both DVs female participants gave higher ratings to typical looking persons (M = 3.89, SE = .11) than to atypical looking persons (M = 3.60, SE = .12), whereas male participants gave higher ratings to atypical looking persons (M = 3.87, SE = .17) than to typical looking persons (M = 3.58, SE = .15). However, these results should be interpreted with caution, because this effect was merely exploratory and bears theoretical derivation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Means (standard deviations) for the dependent variables ascribed competence and warmth depending on target persons’ gender and facial appearance in Study 2.

The higher the mean values, the more the stimulus persons are perceived as competent and warm, respectively.

https://doi.org/10.1371/journal.pone.0181306.t004

The two-way interaction of facial appearance and ascribed personality was qualified by the three-way interaction of facial appearance, ascribed personality, and typicality; F(1, 286.34) = 4.93, p = .027. Looking at the interaction of facial appearance and ascribed personality separately for typical and atypical exemplars reveals that in line with our second hypothesis, the effect of feature-trait associations was not only significant in the typical, F(1, 159) = 13.51, p < .001, but also in the atypical condition, F(1, 170) = 34.91, p < .001. The higher F value for the feature effect in the atypical than in the typical condition combined with the fact that these feature effects significantly differ from each other as indicated by the three-way interaction provide evidence for Hypothesis 3. Again, using mixed models procedures to analyze data support the notion that these results generalize across participants and stimuli. To facilitate comparisons between the results of both studies, Fig 3 visualizes the conditions as in Study 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Impact of target persons’ facial appearance and gender on the ascription of competence and warmth measured with the BSRI and the KMGI for typical and atypical category members in Study 2.

Bars represent errors of the means.

https://doi.org/10.1371/journal.pone.0181306.g003

Again, aiming to investigate whether the results found followed from a direct comparison between the person in the first and second trial we analyzed the two trials separately. We therefore specified two models with the fixed factors facial appearance (masculine-looking vs. feminine-looking; between-participants, because we only analyzed one trial at a time), target gender (male vs. female; between-participants), personality dimension (competence vs. warmth; within-participants), all three possible 2-way interactions, and the three-way interaction for both trials separately. Again, we included random intercepts for participants and stimuli as well as random slopes for facial appearance and gender for the different faces. For both trials separately, results revealed an interaction between facial appearance and personality dimension, thereby replicating the finding that masculine-looking persons were ascribed more competence and less warmth than feminine-looking persons, F_Trial1(1, 94) = 8.55, p = .004, and F_Trial2(1, 185.56) = 38.74, p < .001, respectively. This two-way interaction in the second trial qualified a significant main effect personality dimension, showing that participants generally made higher judgments for competence (M = 4.16, SE = .11) than for warmth (M = 3.59, SE = .11), F(1, 185.56) = 7.80, p = .006.

We also specified trial-based models with the three fixed factors facial appearance, gender typicality, and personality dimension, all three two-way interactions, and the three-way interaction as well as the above specified random effects in both trials separately. Whereas the three-way interaction did not reach statistical significance in the first trial, F(1, 38.99) = 1.33, p = .256, it did reach significance in the second, F(1, 185.56) = 5.56, p = .019.

Discussion

Replicating the results from Study 1, we found clear and strong evidence for feature-trait associations. Overall, masculine-looking target persons were ascribed more competence and less warmth than feminine-looking persons, independent of gender category information. Again, this was true for both the first and the second trial per person, signaling that this effect does not require comparing persons with each other. As in Study 1, we not only found the hypothesized feature-trait associations in both the typical and the atypical condition, but we even found that effects were stronger in the atypical than in the typical condition. These repeatedly found effects support our assumption that gender category is used as a standard against which atypical members are contrasted. Even though the procedural conditions should have made gender category more salient than in Study 1, gender category was not used to draw stereotypical information from, but as a frame of reference with which the atypical member was compared.