Correlations between holistic processing, Autism quotient, extraversion, and experience and the own-gender bias in face recognition

The variability in the own-gender bias (OGB) in face-recognition is thought to be based on experience and the engagement of expert face processing mechanisms for own-gender faces. Experience is also associated with personality characteristics such as extraversion and Autism, yet the effects of these variables on the own-gender bias has not been explored. We ran a face recognition study exploring the relationships between own-gender experience, holistic processing (measured using the face-inversion effect, composite face effect, and the parts-and-wholes test), personality characteristics (extraversion and Autism Quotient) and the OGB. Findings did not support a mediational account where experience increases holistic processing and this increases the OGB. Rather, there was a direct relationship between extraversion and Autism Quotient and the OGB. We interpret this as personality characteristics having an effect on the motivation to process own-gender faces more deeply than opposite-gender faces.


Introduction
Faces provide essential information for social interactions, such as identity, mood, and sex [1]. It is, therefore, not surprising that humans have developed expert processing abilities dedicated specifically to faces [2] and are able to accurately recognise thousands with ease [3]. However, this skill does not extend to all faces: individuals have better recognition for faces of their own ethnicity [4], age [5], and gender [6] than the respective out-group. These owngroup biases are reliable in the face recognition literature, nevertheless theoretical explanations for these effects are not complete.
While a crossover interaction has been found in the own-ethnicity bias [4] the own-gender bias (OGB) is usually found to be stronger in females than males [7][8][9]. This is coupled with the results highlighting females have higher recognition accuracy for all faces than males [7,9]. This latter finding highlights an individual difference variable in face recognition. Such individual differences are not often considered in face recognition research, and in particular in the own-group biases. Potentially, while gender may affect face recognition, it is possible that a covarying individual difference variable may better account for the effect on face recognition. These have not been considered in detail in models of the own-group bias. a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 While contact explicitly might not be responsible for the OGB, it is entirely possible that gender differences in socialisation affect face-recognition performance more generally [36]. Women have superior face recognition accuracy overall compared to men [36] due to increased socialisation [37,38] and contact and this might lead to gender differences in the deployment of holistic processing [39]. Such an argument suggests anything that is associated with increased socialisation, such as extraversion, will lead to higher face recognition performance. Furthermore, variables that might correlate with the amount of holistic processing will correlate with the magnitude of the own-group bias. One such individual difference variable to consider is Autism Quotient.
Extraversion is characterised by interest in interpersonal relationships, social dominance and assertiveness [40]. Given its association with socialisation and the theory posited above, we would assume that there will be a correlation between extraversion and face recognition accuracy. Indeed, Li et al. [41] have found that extraverts show better face recognition than introverts based on the idea that they are more interested in faces than introverts. The theoretical argument for why women show enhanced face recognition relative to men is based on the same socialisation mechanisms. Therefore, we can suggest that extraverts will show enhanced face recognition relative to introverts and this is likely to be based on holistic processing given its involvement in expert face recognition, leading to a larger OGB.
Autism is a neurodevelopmental disorder associated with impairments in socialising and communication [42]. Core traits of autism include a lack of socialisation [43], decreased reliance on global processing which is akin to holistic processing [44], and higher levels of introversion [45]. Given the thesis so far, these features would lead to lower face recognition performance. Indeed, Weigelt, Koldewyn, and Kanwisher [46] have shown that individuals with autism have poorer face recognition accuracy than those without. This impairment has been attributed to reduced reliance on holistic processing [47,48] which maybe a consequence of reduced exposure to faces throughout development. Evidence displays that males show higher autism traits than females [49], which may potentially explain men's inferiority in face recognition relative to women. Further, the lack of socialisation would therefore lead to the hypothesis that autism would negatively correlate with the OGB.

The present study
We have presented an argument that the OGB in face recognition is based on increased experience with faces of one's own gender relative to the opposite gender. This experience leads to differential amounts of holistic processing being engaged in for faces of one's own-gender and faces of the opposite gender. In this study, we will measure the effect of experience on the OGB, while assuming holistic processing acts as a mediator in this relationship. We will directly measure holistic processing using the CFE [21] and the parts and whole test [22] as well as the FIE [20] for both own-and opposite-gender faces. We anticipate the amount of experience with faces of the own gender (specifically during infancy) will correlate with face recognition accuracy and that this relationship will be mediated by the level of holistic processing engaged in. Based on the same logic, we assume that there will be a correlation between extraversion and Autism Quotient and the OGB mediated by holistic processing and experience in a neurotypical sample. because they did not complete the task appropriate-by pressing a single key throughout the experiment) ethnically-White participants took part in the study: 40 males (aged 18 to 35 years, mean: 21), and 42 females, (18 to 47 years, mean: 20). Participants were recruited between September 2017 and September 2018 and were all students at Bournemouth University (where the research took place). Sample size was determined based on the effect size of the own-gender bias [6]). With an r = .27, assuming a power of .8, sample size required was 81, as calculated using G-Power version 3.1. Participants took part as a partial fulfilment of a course requirement. Participants were excluded if they had visual impairments not corrected with glasses or were not ethnically-White. Exclusions were self managed through participants volunteering to take part based on a short advertisement placed online. Ethical approval was obtained from the FST Research Ethics Panel at Bournemouth University (ID17784). All participants provided full informed written consent to take part in the study.

Design
A correlational design with mediation was used. Across hypotheses, there were three predictor variables: experience with the own gender measured with the social contact questionnaire [50], extraversion measured using the Eysenck Personality Questionnaire Revised-short scale [51] and autism measured with the Adult Autism Spectrum Quotient (AASQ: [49]). The mediator variable was holistic processing, measured using the CFE [21], the parts and whole test [22], and the FIE [20]. The dependent variable was the magnitude of OGB (calculated using [52] method: (own gender d'-other gender d')/(own gender d'+ other gender d'), using accuracy measured in terms of Signal Detection Theory (d'; [53] calculated using the [54] method).

Materials
Faces. For all tasks, the faces used were of White ethnicity, presented in frontal poses, with neutral or smiling expressions. All images were grey-scaled, cropped to remove neck and hair length, then pasted onto a white background. For the recognition task, two versions of 80 face images (40 male and 40 female; aged between 18 and 30 years) were taken from the [55] database: One was displayed at learning and one was displayed at test (this was counterbalanced across participants). Inverted faces were created by rotating the stimuli 180 degrees. Images were resized to be 1024 px by 768 px presented in 300 dpi resolution. This was all performed using GNU Image Manipulation Program (GIMP; version 2.8.10). The images used at learning and test were counterbalanced between participants.
For the composite faces, a further 12 (6 female) faces were taken from the Karolinska Directed Emotional Faces database [56]. Instead of combining parts of different faces, like the standard task [21], different identities were created by moving features to increase the chance of participants using holistic processing; as position was the only aspect of the features that changed [57]. Twelve new images were created by adjusting the distance between the eyes, and a further 12 by moving the mouth up. A further 12 images included both the changes (see Fig  1 for an example set of faces). An additional 48 images were created by vertically splitting the image through the nose and moving the lower half of the face to the right (see [57] Fig 3, for a full explanation of the procedure). This resulted in 96 images. These were created in the same way as [57] and maximise the chances of holistic processing being employed because the tobe-detected changes are subtle and involve assessing the relationship between two features. Utilising the relationship between features to process faces is the definition of configural processing [58] of which holistic processing is a type [12]. The detection of such changes recruits the fusiform gyrus more so than the detection of simple featural changes (i.e., changes to the size of features) in faces [59]. The fusiform gyrus is specifically recruited to process faces [60] and has been linked to processing faces holistically [61]. These subtle differences between faces are sufficient to interpret faces as different identities [62].
For the Parts and Wholes test [22], stimuli consisted of 80 original faces (40 female) taken from the [55] database. These images were cropped to create part images. Parts were created for the eyes, nose, and mouth (however, for the purpose of the experimentation, only trials involving the eyes were analysed because the whole-part advantage is much stronger for the eyes than other features, [47]. Other features were included to prevent participants only focusing on the eyes [19]. Social contact questionnaire. This was adapted from Walker and Hewstone's [50] scale that measured experience with the own-ethnicity. Questions associated with ethnicity were modified to ask about experience with both genders. It consisted of 13 items for each gender measuring contact with people of each gender during infancy, childhood, and adulthood. Participants respond on a 5-point Likert scale, with anchor points ranging from strongly agree to strongly disagree or to indicate a numerical value. This produced an experience score for interactions with both males and females. Higher scores represented more social contact. The original scale had high internal reliability (Cronbach's α = 0.83), construct, and face validity [50].
Eysenck personality questionnaire revised. The short scale version of Eysenck's Personality Questionnaire-Revised (EPQ-R; [51]) was used in this study. It consisted of 48 items measuring extraversion/introversion, neuroticism, psychoticism, and lying. Participants respond with a "yes" or a "no" to each statement. Only the 12 extraversion items were analysed, with higher scores representing more extraversion. This scale has high internal reliability (Cronbach's α = .89-.91; [64]) and factorial validity [65].
The adult autism spectrum quotient. The Adult Autism Spectrum Quotient (AASQ; [49]) consisted of 50 items measuring traits of Autism. Participants responded using a 4-point Likert scale with anchor points ranging from "definitely agree" to "definitely disagree". This has moderate to high internal reliability (Cronbach's α = 0.63-0.77) high test-retest reliability and criterion validity [49]. This created a score out of 50, with higher numbers representing more traits associated with Autism.

Procedure
After providing informed consent, participants completed the three face recognition tasks (with embedded three questionnaires). Participants were offered breaks between each task. The order of tasks were counterbalanced between participants and once completed, participants were debriefed. The face tasks were displayed on a 15" high-resolution colour LCD.
Face-recognition task/Inversion effect. This task was completed on OpenSesame [66] and consisted of three consecutive phases: Learning, distraction, and test. At learning, participants were presented with 40 faces on screen, (20 female, 10 of which were inverted) sequentially in a random order in the centre of the screen. The face remained on screen until responded to by a distinctiveness judgement. Distinctiveness was measured by participants responding to the question, "How easy would this face be to spot in a crowd?" [67], using a Likert scale ranging from 1 (difficult-i.e., typical face) to 9 (easy-i.e., atypical) This judgement was carried out as it is a gender-neutral form of encoding [31]. Between each face a grey mask was presented in the centre of the screen where the face would appear for 600 ms.
In the distractor phase, participants completed the social contact questionnaire [50], EPQ-R [51], and AASQ [49] on screen, in addition to providing demographic information. This typically lasted two minutes to complete. It is possible that utilising a questionnaire that raises participants awareness of gender by making gender more salient might increase the OGB. However, the social contact questionnaire was followed by the other questionnaires, minimising its influence. Further, there is no empirical evidence that making gender salient increases the OGB (in the same way that making gender salient can lead to stereotype threat effects, [68]). However, Hills et al.'s [31] data do indicate that anything that might increase motivation to process own-gender faces more deeply enhances the OGB, therefore it is possible that the presence of the social contact questionnaire actually enhances the presence of the OGB in our study relative to those that use a more neutral distractor.
At test, participants were presented with 80 images of faces: 40 previously seen and 40 new images (split evenly by gender and orientation). Orientation of the old faces was matched from learning to test. Faces were presented sequentially in a random order. Between each face, there was an inter-trial stimulus consisting of a grey box presented in the centre of the screen where the face would appear for 1150 ms. The participants had to judge whether they recognised the face from the learning phase, pressing the 'm' computer key if they did and 'z' if not. Faces were on screen until participants responded. Note that only the upright face trials contributed to the measure of the OGB in this study.
By incorporating the OGB and FIE task together in one procedure saves time for the participants and ensures that both tasks are performed under the same conditions (i.e., that participants are always unaware of the recognition test). However, it does produce a caveat that the upright trials are used to calculate both the OGB and the FIE. Due to this, these two tasks are not completely independent. While this might create artificially inflated correlations between these two variables, it is standard practice in the field and, pre-empting the results, it does not affect any conclusions drawn.
Composite face effect. This task was completed using OpenSesame. This was a delayedmatching task. Participants were presented with 384 trials sequentially. Each trial consisted of a face presented for 600 ms, a blank screen for 300ms, followed by a second face presented until participants responded to it: participants had to judge whether the distance between the eyes in the first image were the same as the second image-pressing 'm' for similar and 'c' for different. This is akin to identifying if the two images represent the same identity.
Based on the pairing of the images (different eye and moth positions), there were 16 different types of pairing for both the aligned and the misaligned conditions. Therefore, there were 32 different trial types created for each of 12 face identities, creating a total of 384 trials. The order of the pairs were randomised between participants. This pairing style created four conditions conceptually for the aligned and misaligned variants: congruent same (CS), congruent different (CD), incongruent different (ID), incongruent same (IS). In two of the conditions, the bottom half of the face changed, and either the top half also changed (CD) or stayed the same (IS). In the other two conditions, the bottom half of the face stayed the same and the top half either changed (ID) or remained the same (CS). The metrics for calculating holistic processing are presented in Table 1. Two values were produced from the CFE, as both partial [23] and complete analysis [69] were run because of the ongoing debate as to which calculation represents holistic processing [23] and studies have found holistic processing in one but not the other [69,70].
Parts and whole test. This was a delayed-matching task. Participants were presented with 480 trials sequentially. Each trial consisted of a face for 1508ms, a blank inter-stimulus interval for 305ms, followed by a pair of images including the original image and a distractor face presented until participants responded to indicate with right and left computer keys which image they had previously seen. There was a blank inter-trial interval of 796ms. There were four blocks of 120 trials. Each block contained a different combination of images (from learning to test): whole-whole(WW), whole-part (WP), part-whole (PW), part-part (PP). Having four conditions, instead of the original two used by [22] (WW and WP) was used to give a more complete measure of holistic processing as it allows for relative performance differences between the part and whole conditions not to affect the calculation of holistic processing [71]. The metric for calculating holistic processing is presented in Table 1.The presentation order of the blocks and the images within the blocks were randomised between participants.

Results
We first explored our data to ensure that they met the assumptions of parametric data and did not contain any outliers. All assumptions were met for the data reported here (though the distribution of data for d' for the OGB was skewed-however the statistics used are robust to the skew), and there were three participants whose standardised residuals were above 3.29 for their OGB. These participants could be considered outliers, however their inclusion did not affect the pattern of significance. Before assessing our hypotheses, we first established that we had an OGB using a one-sample t-test, t(159) = 3.56, p < .001, Cohen's d = 0.28. The OGB was larger for female participants (M = 0.28, SD = 0.57) than male participants (M = 0.08, SD = 0.68), t(158) = 1.98, p = .049, Cohen's d = 0.32. We then assessed whether each test of holistic processing produced results as expected using a series of eight Bonferroni-corrected (α = .006) one-sample t-tests for both own-and opposite-gender faces. This is presented in Table 1. We assessed whether these measures of holistic processing correlated with each other using a series of six Bonferroni-corrected (α = .008) Pearson's correlations. There were no significant correlations post-correction (largest r = .16, smallest p = .046), replicating [72]. Pre-correction, the complete CFE correlated negatively with the whole-part advantage, r(156) = .16, p = .046 and a negative correlation between the FIE and the partial CFE, r(158) = .14, p = .046. No other correlations were significant, largest r = .04, smallest p = .348. In order to assess whether participants showed greater holistic processing for own-gender versus opposite-gender faces, we ran a series of 12 Bonferroni-Şidák-corrected (α = .004) within-subjects t-tests, also presented in Table 1. These all show enhanced holistic processing for own-gender than opposite-gender faces, except when measured using the parts and wholes test.
In order to address our main hypotheses (that experience would predict recognition through holistic processing), we ran a series of mediation analyses using the Baron and Kenny [73] method. This analysis contains three linear regressions: regressing the predictor on the mediator; the predictor on the outcome variable; and both the predictor and mediator on the outcome variable, in a single linear regression. Given that our four measures of holistic processing did not correlate consistently with each other, we ran four mediations, with each holistic processing measure included. If all regressions displayed that the predictor significantly affects the outcome variable, Sobel's test of mediation [74] will be carried out.
Since experience (measured as a relative measure of experience with one's own-gender and the opposite gender) did not predict the OGB, R(158) = .09, F = 1.40, MSE = 0.40, p = .238 (see Fig 2A), we did not run the mediation analysis for this variable. We ran the analysis on both females and males separately as well, and neither reached significance, largest R(78) = .11, smallest p = .329. Nevertheless, we did assess whether measures of holistic processing predicted the OGB. We ran a series of four Bonferroni-corrected (α = .0125) regressions between each measure of holistic processing and the magnitude of the OGB (Fig 2B). None of these reached significance, largest R(158) = .15, smallest p = .066 (this was for the FIE-As indicated in the procedure, this might be an artificially inflated correlation due to the fact that the OGB and the FIE are calculated using the same upright face trials). We further ran a series of four Bonferroni-corrected (α = .0125) Regressions between measures of experience and the OGB in each of the holistic processing measures (Fig 2C). None of these reached significance, largest R (158) = .13, smallest p = .114.
These measures were conducted only on the overall measure of experience. Since [31] have shown that recent experience with faces does not predict the OGB, it is no surprise that we failed to find a relationship between experience and the OGB. Therefore, we recalculated the experience questionnaire including only measures of childhood experience to assess whether self-reported childhood experience predicted the OGB.
Early experience did not predict the OGB, R(158) = .11, F = 2.05, MSE = 0.40, p = .154 ( Fig  3). Therefore, we did not run further mediation analyses for this variable. We ran the analysis on both females and males separately, and neither reached significance, largest R(78) = .11, smallest p = .342.

Discussion
The first aim of this study was to investigate whether experience with own-gender faces, specifically childhood and peer relationships, predicted the OGB. No relationship was found, consistent with past research [30,31]. Such replicated null results suggest that the contact hypothesis [11], even when early relationships are considered, cannot account for the OGB. The contact hypothesis has successfully explained other biases (own-ethnicity [16], and own-age [26]). The reason why experience does not explain the OGB may result from the nature of the relationships during development. There is a switch from female-dominated childhood relationships to own-gender peer relationships. This switch might negate the effects of experience at least in boys [33,35]. Further, unlike the other biases, there is a much more equivalent contact with own-and other-gender faces: any effect of contact is mitigated due to a smaller relative difference in contact between own-and other-gender faces relative to the other biases. Indeed, given some sort of threshold of contact needed to engage in expertise-related holistic processing, it is likely that one has already reached this for other-gender faces (there is likely to be a very high baseline of individuating experience with other-gender faces). Therefore, the bias is less likely to be obviously related to experience-mitigated engagement of holistic processing.
We hypothesised that greater holistic processing would be observed for own-than othergender faces and that this would mediate the relationship between experience and the OGB. The majority of measures revealed enhanced holistic processing for own compared to opposite-gender faces. This indicates that we automatically categorise faces as in-group or outgroup (own-or other-gender in this case). In-group faces are then processed more deeply, (i.e., more holistically). Out-groups are attended to less, resulting in shallower processing, according to the socio-cognitive motivational account [75]. Motivation can manipulate the extent with which expert processing is applied [31], so it may be that the OGB is a result of motivation to process faces differently as opposed to experience. However, we found that enhanced holistic processing for own-gender faces did not predict the OGB. This is consistent with findings from [74] who found that holistic processing was not correlated with face recognition accuracy.
For female participants, increased holistic processing was found for own-gender faces relative to other-gender faces. This was not found across all measures for male participants. This might be related to the oft found result that the OGB is stronger in females than [37,76,77]. We link this finding to the notion that females are typically more interested in faces than males [38] which causes them to be more accurate at face recognition generally and show an enhanced OGB and holistic processing than males. Therefore, while we have not found that experience leads to holistic processing and this leads to the OGB, we have suggested that the mechanisms behind increased holistic processing and face recognition accuracy in females relative to males is due to the same mechanism.
One theoretical reason why females are thought to be better at face recognition than males is because they are socialised to be. In the introduction, we indicated that there might be other individual difference variables that are related to sociability that might explain the OGB. We predicted that there would be individual differences variables (extraversion and autism) affecting the amount of holistic processing engaged in and this would affect face recognition. We did not find this indirect mediational link between these variables. However, we found a direct relationship between autism and the OGB and between extraversion and the OGB. We did not predict the direct link between these individual difference variables and the OGB. Our results indicate we can discount the mediational link and explore alternative explanations for why these individual difference variables might affect face recognition.
Autism is known to affect perception and information processing profoundly [78], including enhanced visual target detection [79] and abnormal eye movements when viewing social scenes [80] including faces [81]. Given that Autistic traits exist along a continuum in the general population [82], it is reasonable to consider that these effects also exist as a continuum in the general population. People with Autism are generally less accurate at face recognition than people without Autism [46]. Riby and Hancock [83] hypothesise that this is due to a decreased interest in socialisation associated with Autism. Our correlation between Autism Quotient and the OGB reveals even lower face recognition performance for opposite-gender faces compared to own-gender faces. This suggests that Autism levels correlate with the amount of attention paid to other-gender faces negatively. We hypothesise that this is related to the findings that individuals with Autism find it easier to communicate and interact with in-group than outgroup members [84]. It is therefore reasonable to suggest that people with Autism pay less attention to out-group members (in this case opposite-gender faces) which leads to some level of cognitive disregard and therefore shallow processing. Of course, this argument is extrapolating from a clinical sample of people with diagnosed Autism to those with traits of Autism from a non-clinical sample. Further work would be required to establish if this generalisation was plausible. Eye-tracking measures could be used to verify the claims made here: active scanning is revealed through a greater number of fixations to diagnostic facial features [85]. While those with Autism tend to show a more restricted scanpath, avoiding making eye-contact than typical groups [86], our speculation here is that this would be exaggerated for other-gender faces.
While extraversion is primarily associated with an increase in sociability [87], it is also known to affect perceptual processes, for example increased olfactory sensitivity [88] and eye movements [89] when viewing social scenes [90]. Indeed, the effects extraversion has on eye movements may explain female accuracies in classification tasks [90]. Extraversion correlates positively with face recognition accuracy [41]. However, it does not correlate with recognition accuracy for non-social stimuli [41]. Our data demonstrated that extraversion was associated with poorer recognition of other-gender faces (because of the larger OGB). This was not related to experience (potentially therefore sociability) nor increased use of holistic processing. Instead, we hypothesise that this relationship occurred due to extraverts increased attention paid to faces relative to introverts as revealed through the ERP P300 [91]. This ERP is known to be stronger for self-relevant stimuli than non-self relevant stimuli [92. 93]. Therefore, we suggest that extraverts pay more attention to their own-gender faces because they are more willing to engage in socialisation with their own-group (e.g., through sporting or social activities, [94]) and this makes their own-group more important to process. This speculation would also need to be empirically tested potentially through an EEG study in which the P300 would be predicted to be higher for own-gender than other-gender faces.
These results highlight a hitherto under-researched area of face recognition research, that of individual differences associated with personality traits. While it is known that extraversion and Autistic traits affect face recognition generally [41,95], what we have shown is that individual differences can have more nuanced effects on face recognition. These individual difference variables appear to be dimensional in their effect on face recognition given the correlational nature of the present study. Furthermore, individual difference variables might have different effects on different classes of faces. These results are important in building models of face recognition and might aid researchers when theorising how face recognition develops.
One caveat with our findings is that our procedure may have artificially enhanced the OGB because we made gender salient during the distractor phase of the experiment. This assumes that the OGB is one of recall, whereas most research indicates that the OGB is due to differential processing during the learning of faces [10] especially given that different judgements being made during the learning phase cause different magnitudes of the OGB [31]. Nevertheless, it is possible that our procedure did increase the OGB and methods that lead to a smaller OGB might have less variability to replicate our findings.
In conclusion, we have replicated the OGB in face recognition, demonstrating that it is stronger in females than males. Females also consistently show more holistic processing for own-gender faces than other-gender faces, however this is unrelated to both experience with own-and opposite-gender faces and to the OGB. Further, we have shown that the magnitude of the OGB is predicted by participants' Autism and extraversion scores. These results show subtle individual difference variability affects reliable face recognition findings.