The visual perks of (not) being a wallflower: sociality is a stronger predictor of individual differences in face recognition than childhood experiences

Spencer T Andrews; Maleah J Carter; Shruti Japee; Chris I Baker; J Brendan Ritchie

doi:10.1016/j.visres.2026.108845

. Author manuscript; available in PMC: 2026 Jun 11.

Published in final edited form as: Vision Res. 2026 May 19;245:108845. doi: 10.1016/j.visres.2026.108845

The visual perks of (not) being a wallflower: sociality is a stronger predictor of individual differences in face recognition than childhood experiences

Spencer T Andrews ¹, Maleah J Carter ², Shruti Japee ³, Chris I Baker ³, J Brendan Ritchie ^3,⁴

PMCID: PMC13252829 NIHMSID: NIHMS2178910 PMID: 42155309

Abstract

Individuals vary considerably in their ability to recognize faces. Such variability may in part reflect individual differences in exposure to faces based on the number or diversity of faces that are available to be seen in the environment. To investigate this possibility, a series of studies have used hometown size as a proxy for individual variation in face exposure and investigated whether it predicts individual differences in face recognition ability, with mixed results. In particular, it was found that performance on face recognition tasks was predicted by hometown size, but only for male participants. In three experiments, we sought to revisit whether hometown size predicts face recognition ability, using an order of magnitude larger sample sizes (from N = ~90 increased to N = ~ 900), while also considering a variety of other predictors related to childhood experience (e.g. neighborhood type and family composition) and adult sociality, including personality, social intelligence, and shyness. Across all experiments, we found that hometown size was a poor predictor of adult face recognition ability compared to other predictors related to childhood experience. In addition, of the different measures related to sociality, social awareness was a comparatively strong predictor of individual differences in face recognition ability.

Keywords: Face recognition, social cognition, personality, individual differences

1. Introduction

The ability to recognize faces varies considerably across individuals (Wilmer, 2017). At one extreme are those who exhibit varying degrees of (congenital or developmental) prosopagnosia, or face blindness, and struggle to learn new face identities or recognize familiar ones (Behrmann & Avidan, 2005; Susilo & Duchaine, 2013). At the other extreme, are those who are extraordinarily good at recognizing faces, so-called “super recognizers” (Ramon et al., 2019; Russell et al., 2009). A great deal of evidence suggests that this wide range in face recognition ability is (at least partially) explained by individual differences in early life experience with faces, which may account for variation in recognition ability for faces of different ages, sex and gender, or racial identities (Anzures et al., 2013; Herlitz & Lovén, 2013; Kuefner et al., 2008). The most direct evidence for a role of early life experience on face recognition ability comes from investigations of the impact of deprived exposure to faces during childhood. For example, individuals with bilateral cataracts early in life have shown face recognition deficits in adulthood (Putzar et al., 2010). In a more extreme case, monkeys selectively raised without exposure to faces did not develop the well-known system of face-selective patches in visual cortex even after extensive later experience with faces (Arcaro et al., 2017). These deprivation results suggest that the extent of face exposure during infancy and childhood may be predictive of face recognition ability later in adulthood.

One way in which individual exposure to faces may vary is simply due to the number or diversity of faces that are available to be seen. To investigate this possibility, a series of studies have used hometown size as a proxy for individual variation in face exposure and investigated whether it predicts individual differences in face recognition ability. The expectation was that performance on face recognition tasks would be greater for individuals from hometowns with larger populations compared to those from smaller population centers. In line with this hypothesis, the first study adopting this methodology (Balas & Saville, 2015) found that individuals from smaller towns performed worse on the commonly used Cambridge Face Memory Test (CFMT, (Duchaine & Nakayama, 2006)) compared to those from larger hometowns. Follow-up studies replicated the finding of differences in CFMT performance across groups, but did not replicate the finding that hometown size predicts performance on other tasks related to face recognition ability, including an unconstrained face card sorting task (Balas & Saville, 2017), and a separate face matching task(Sunday, Dodd, et al., 2019). Taken as a whole, these studies provided mixed evidence as to whether hometown size is predictive of face recognition ability.

Based on these previous findings, Sunday and colleagues (Sunday, Patel, et al., 2019) hypothesized that the relationship between hometown size and face recognition performance was specific to tasks that required learning. To investigate this possibility, they conducted a study that made a number of important methodological advances on previous work. First, they recruited a larger sample of participants for which the distribution of hometowns population densities within the sample was not heavily skewed towards higher densities. Second, they had participants perform a range of face recognition tasks that either did (e.g. CFMT) or did not involve a learning component. Third, they also considered extraversion as a mediator of face recognition performance, using questions from the International Personality Item Pool (IPIP) (Goldberg et al., 2006), since previous studies have suggested that self-reported judgments of extraversion may be predictive of face recognition ability (Lander & Poyarekar, 2015; Li et al., 2010). Unlike the previous studies, Sunday and colleagues found no main effect of hometown size as a predictor of face recognition performance on any of the tasks, including the CFMT. However, Sunday and colleagues did report a consistent gender-hometown interaction across tasks: women from smaller hometowns tended to be better at face recognition than men from smaller hometowns. They also found no relationship between extraversion and face recognition performance on any of the tasks.

While treating hometown size as a proxy for face exposure is a clever approach for testing how early life experience might predict adult face recognition ability, the studies utilizing this approach have three limitations. First, the studies relied on samples of participants that were relatively small and demographically homogenous. For example, only one study (Sunday, Patel, et al., 2019) evenly sampled from a range of population densities, with an N = 90, while other studies sampled from a single site (e.g., North Dakota and Nebraska). It remains to be seen whether the gender-hometown interaction reported by Sunday and colleagues will be observed across a demographically larger and more varied sample. Second, hometown density is just one possible predictor related to early life, or childhood experience. It is possible that other proxies may better predict face recognition ability, especially in a larger sample. Finally, although several other studies have looked at extraversion as a predictor of face recognition ability (Engfors et al., 2024; Lander & Poyarekar, 2015; Li et al., 2010; McCaffery et al., 2018; Satchell et al., 2019), it is just one of the possible predictors relates to personality, and sociality more generally, which might predict face recognition performance. A recent study by Engfors et al. (2024) investigated the relationship between individual extraversion scores and the size and quality of social circles as predictors of CFMT scores. They found the effects of the social circle predictors were mediated by extraversion. They also found other Big Five personality traits were stronger predictors of CFMT scores than extraversion. However, measures related to other aspects of sociality may be more predictive of face recognition ability such as social intelligence (Silvera et al., 2001) and shyness (Cheek & Buss, 1981). Such traits should also be accounted for when assessing apparent effects of hometown size or other measures related to childhood experience.

In the present work, we sought to address these limitations and more thoroughly test the relationship between hometown size and face recognition ability, and in particular the interaction with gender (Sunday, Patel, et al., 2019). In Experiment 1, we collected CFMT data for a sample of participants that was an order of magnitude larger than previous studies. We also asked participants a range of exploratory questions related to childhood experience to determine whether such experience might also predict CFMT scores. Finally, we tried to obtain a fuller picture of whether individual differences in sociality might predict CFMT scores using a variety of measures drawn from social psychology related to personality (Gosling et al., 2003), social intelligence (Silvera et al., 2001), and shyness (Cheek & Buss, 1981). The same general design of Experiment 1 was extended to an additional two experiments. In Experiment 2, we asked whether the pattern of results we observed for Experiment 1 would be preserved if the face stimuli were changed from male to female, using the Female Cambridge Face Memory Test, or F-CFMT (Arrington et al., 2022). In Experiment 3, we asked whether the pattern of results we observed for the first two experiments would be modulated by an “other-race” effect for faces less commonly encountered in rural regions, using the CFMT-Chinese (McKone et al., 2017). To briefly summarize our findings, across all three experiments we find that hometown size is a poor predictor of adult face recognition performance – although other facets of childhood experience may better explain individual differences in performance. In contrast, different sociality measures were comparatively much stronger predictors of individual differences in face recognition performance.

Experiment 1

In our first experiment we sought to replicate the reported finding of an interaction between hometown size and gender as predictors of face recognition performance (Sunday, Patel, et al., 2019). We focused on the CFMT since it resulted in the strongest interaction effect in the study of Sunday and colleagues, and variants were used in earlier studies looking at hometown size as a predictor (Balas & Saville, 2015, 2017; Sunday, Dodd, et al., 2019). We also collected responses for a wide range of questions related to childhood experience as well as a battery of measures related to individual sociality.

Method

Participants.

Our goal was to recruit at least an order of magnitude (N = ~ 900) more participants than the study of Sunday and colleagues (N = 90). Considering that our pool of participants was a sample of convenience and the likelihood that there would be a substantial amount of attrition (Zhou & Fishbach, 2016), we aimed for an initial set of 1500 unique participants, recruited through the crowdsourcing platform Amazon Mechanical Turk (AMT). We also wanted to select participants who come from hometowns with a range of population densities. To this end we ranked all 50 US states by the proportion of the population living in rural areas and split the ranking into 5 groups of 10 states each. For each sub-group of the ranking, we then selected the 5 states with the largest population (Figure 1A). For each of the group of states, we then recruited 300 participants to participate in the study based on whether they successfully completed the CFMT (see procedures). From these regional groups a total of 1471 participants completed the online experiment. Applying our exclusion criteria (see procedures) resulted in a sample of 1050 participants (538 female, 512 male; Age: mean = 39.34, SD = 12.5, range = 18 – 77; Racialized: Caucasian = 865, Not Caucasian = 185; Ethnicity: Hispanic = 117, non-Hispanic = 933; Completed postsecondary education: 787). The experiment was approved by the institutional review board of the National Institutes of Health, and all methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki. All participants provided informed consent prior to participation in the experiment.

Figure 1. — (A) A map of the United States indicating the 25 states from which participants were recruited on Amazon Mechanical Turk, based on location of their IP addresses (created using mapchart.net). Different shades of blue indicate the five groupings of five states based on similar proportions of rural populations (Group 1 = least rural; Group 5 = most rural). (B) Example stimuli from the tasks used in Experiments 1 – 3: the CFMT, F-CFMT, and CFMT-Chinese (Arrington et al., 2022; Duchaine & Nakayama, 2006; McKone et al., 2017).

Materials.

Participants completed the standard 72 item Cambridge Face Memory Test (CFMT) (Duchaine & Nakayama, 2006). In their study, Sunday and colleagues used CFMT+, which is a longer version of the test, which includes additional trials. On the standard form of the CFMT, participants study a set of 6 novel Caucasian male faces and later are tasked with correctly selecting one “familiar” face out of a set of three faces presented concurrently (Figure 1B). The images shown during testing vary from the images participants see during the study phase. Throughout the test, images are presented both head-on and at various side profiles, which challenges the participants to not just memorize the familiar faces but recognize their facial features. In our implementation, trials would time out after 10 s if there was no response and the next trial would proceed. The difficulty increases as the task proceeds when noise is added over the faces in order to make discrimination between features of the familiar and unfamiliar faces harder. After completing the CFMT, we calculated the participant’s face recognition score. Proportion correct accuracy was calculated (chance = 0.33).

Each participant answered a number of questions related to demographics and childhood experience. The demographic questions required participants to indicate their gender, age, race, and ethnicity, neighborhood type of their current residence, and education (see Supplemental Table 1A). Similar to the studies of Sunday and colleagues, participants were asked to share the name, state, and ZIP code of their hometown, which we defined as the place they lived for the greatest period before the age of 10. Additionally, participants reported the ‘neighborhood type’ of the town in which they were raised (i.e. urban, suburban, rural). Though qualitative, we felt that this question might more accurately ascertain the type of town a participant was raised in than a simple population density calculation. Beyond these questions participants answered a number of questions related to their childhood experiences, including homelife and household composition or the sorts of schooling and social activities which participants engaged in at the time. For example, participants also provided information about the type of home they grew up in (e.g. apartment complex), family size, and childhood social activities (e.g. team sports). These questions served a more exploratory purpose, since there is much to be learned about which factors in an individual’s early life experience might contribute to face recognition ability during adulthood. For a full list of the childhood experience questions, see Supplemental Table 1B.

Participants also completed three questionnaires related to what we refer to as their “sociality”. First, participants completed the Ten Item Personality Inventory, or TIPI (Gosling et al., 2003). TIPI consists of 10 self-descriptors that participants report based on a Likert scale from strongly agree to strongly disagree. TIPI is much shorter compared to other Big Five personality questionnaires and consists of only 10 self-report questions which are used to determine individual extraversion, agreeableness, conscientiousness, emotional stability, and openness to experience. Although previous studies have used longer lists of extraversion-related questions, or more extensive questionnaires for assessing Big Five personality traits, we used TIPI to allow time for other sociality measures. Second, participants performed the Tromsø Social Intelligence Scale, or TSIS (Silvera et al., 2001). TSIS consists of 21 self-description questions that participants again report in terms of agreement along a scale (Supplemental Table 2). From these questions, scores are tabulated to code for three aspects of social intelligence: social information processing, social skills, and social awareness. Finally, participants also completed the Revised Cheek and Buss Shyness Scale, or RCBS (Cheek, 1983; Cheek & Buss, 1981). RCBS consists of 13 self-descriptions that participants select based on how characteristic they are, on a 5-point rating scale. To this list we added the question of whether participants considered themselves shy as a child (i.e. before the age of 10), which we believed might also be predictive of shyness in adulthood.

All experiment material were presented using PsyToolkit (Stoet, 2010, 2017). Since the task was completed at home, we did not have control over the screen size with which participants completed the task. However, the task was only available to participants operating on desktop computers and laptops, as mobile device users were prohibited from participating.

Procedure.

Only AMT workers located in the listed states were invited to complete the task. By making five separate – but identical – postings to AMT, we were able to recruit participants with accounts located only in selected states. AMT workers who had IP addresses in one of the 5 state groups first completed the questionnaire containing demographic and childhood questions. Afterwards answering these questions participants completed the CFMT, before continuing on to the three sociality scales. These were completed in the same order for all participants: TIPI, TSIS, and then RCBS.

Statistical analysis.

Several participants were excluded from the final analysis for a number of reasons related to their performance on CFMT or their responses to the demographic and childhood experience questions. Specifically, participants were excluded if any of the following conditions were met: (i) their CFMT scores were at or below chance; (ii) their median RTs were < 200 ms or timed out at 10 s; (iii) they did not include a valid ZIP code for their hometown; (iv) the reported gender was other than male or female; and (v) they selected a “No response” or had an invalid response for any other demographic question. Because of extreme values in the responses to the childhood experience questions, we also excluded participants if: (vi) they indicated they lived > 2 h (120 min) from a city; (vii) their childhood household size was > 10 people; (viii) they had greater than > 10 siblings; and (ix) the number of older siblings was greater than the number of siblings (suggesting one of the values was incorrect).

The values for some of the response variables were either transformed or recoded. In order to calculate the population density of each participant’s hometown and current ZIP code, we utilized Simple Maps (simplemaps.com). Simple Maps is a database that uses 2021 U.S. Postal Service^™ and Census Bureau data, providing the population density of every ZIP code within the U.S. and its outlying territories. These were then log transformed (Sunday, Dodd, et al., 2019; Sunday, Patel, et al., 2019). The values for racialization were recoded for analysis to White vs Non-white and for TV watching to “None” and “Some” because the proportions of responses were otherwise heavily unbalanced.

To replicate the analyses of Sunday and colleagues (Sunday, Patel, et al., 2019), we conducted multiple linear regression with CFMT scores (proportion correct) as the dependent variable and the log of population density (logPD), gender, and their interaction, as predictors. The same regression analysis was conducted for all other childhood experience predictors replacing population density and still including an interaction term (e.g. for neighborhood type). The same analysis was carried out individually for each of the sociality measures related to personality (TIPI), social intelligence (TSIS) and shyness (RGBS). Because some predictors formed ordinal categories, ANOVA was applied to the linear regression model to test for main effects of the individual predictors. When reporting the coefficients of determination (R²) for the regression models, we used the following conventions for describing effect sizes for proportions of explained variance: R² ≥0.01 = small effects; R² ≥0.09 = medium effects; R² ≥0.25 = large effects. These conventions are based on those for Pearson’s r correlation coefficient since r * r = R², and correlations of r = 0.1, 0.3, and 0.5, are the threshold values for small, medium, and large effects, respectively (Cohen, 1992). When R² < 0.01, we described this as a “tiny” effect, if necessary. All statistical analyses were performed with MATLAB (The MathWorks Inc., Natick, USA) and plotting was performed using the Gramm Toolbox (Morel, 2018). For the group of childhood experience predictors, where a larger number of individual models were tested, we corrected for multiple comparisons by checking the FDR-adjusted p-values (Benjamini & Hochberg, 1995). While the uncorrected p-values are reported in the accompanying tables, if the correction resulted in the p-value being above threshold (α= 0.05), this is indicated by superscript “ns” next to the p-value in the accompanying table and next to the R² in the relevant supplemental figure panel.

An exploratory analysis was carried out to determine whether the different types of classes of variables (demographic, childhood experience, and sociality) uniquely predict portions of the variance in CFMT scores. To do this, we performed commonality analysis, or “variance partitioning”, which is a method to determine whether multiple predictors uniquely or jointly explain variance in the dependent variable (Seibold & McPhee, 1979). In the case of three sets of predictors (a, b, c) for some dependent variable y, there are seven coefficients of determination for all possible combinations of predictor sets in a linear regression model: R² _{y · a} , R² _{y · b} , R² _{y · c} , R² _{y · ab} , R² _{y · ac} , R² _{y · bc} , R² _{y · abc.} The last of these is the full model, for which the variance is partitioned based on differential weighting of the R² of the different models. When there are only three predictors, the partitioning can be performed using a simple weighting table, in which the vector of coefficients is multiplied with row vectors of weights for each of the unique and common variance components (Nimon & Reio, 2011). To determine whether a predictor from each of the sets should be included, a linear regression model was carried out with the individual predictor. If it was significant, it was included, otherwise it was excluded. These results were visualized with EulerAPE (Micallef & Rodgers, 2014).

In a supplemental exploratory analysis, variance partitioning was also used to compare four sociality predictors from TIPI (agreeableness, conscientiousness, and openness to experiences) and TSIS (social awareness) and whether they accounted for unique portions of the variance in CFMT scores. In the case of four predictors (a, b, c, d) for some dependent variable y, there are fifteen coefficients of determination for all possible combinations of predictor sets in a linear regression model: R² _{y · a} , R² _{y · b} , R² _{y · c} , R² _{y · d}, R² _{y · ab} , R² _{y · ac} , R² _{y · ad}, R² _{y · bc} R² _{y · bd} , R² _{y · cd}, R² _{y · abc},R² _{y · abd}, R² _{y · acd}, R² _{y · bcd}, R² _{y · abcd}. When there are only four predictors, the partitioning can also be performed using a simple weighting table. Results are visualized with a four variable Venn diagram. Pairwise Pearson’s r correlations between sociality predictors from TIPI, TSIS and RCBS were also calculated. Because this resulted in 36 statistical tests, we corrected for multiple comparisons by reporting the FDR-adjusted p-values (Benjamini & Hochberg, 1995).

Results

Hometown population density.

We first assessed whether hometown population density interacted with gender as predictors of CFMT scores. Like with the study of Sunday and colleagues, the range of population densities we sampled was widely distributed, though there was some positive skew (Figure 2A). A linear regression analysis was used to model CFMT scores with population density, gender, and their interaction as predictors (Table 1A). The fit of the regression model was significant and predicted a small portion of the variance in CFMT scores (adjusted-R² = 0.02). Replicating the findings of Sunday and colleagues, we found that gender was a significant predictor of CFMT scores, but population density was not. However, despite the order of magnitude increase in sample size the interaction was not significant. Notably, whereas Sunday and colleagues found a positive relationship between population density and CFMT scores for male participants, our data show the opposite (non-significant) trend (Figure 2A). The same pattern of results was observed restricting the age of participants to 18-30 (N = 279), which was the age range of participants in the study of Sunday and colleagues. The regression model predicted a small portion of the variance, but was not significant, and there was no significant interaction between population density and gender (Supplemental Figure 1A, Supplemental Table 3A).

Figure 2. — (A) Hometown log10 population density (logPD) as a predictor of CFMT scores. *Top*, histogram showing the count distribution of population density for all participants in Experiment 1. *Bottom*, scatter plot of population density and CFMT scores for all observers. Lines are the fits of generalized linear models (GLM) with 95% confidence intervals. The GLM is for visualization purposes and is *not* the same as the model described in the text. (B) Hometown neighborhood type as a predictor of CFMT scores. *Top*, histogram showing the count distribution of hometown neighborhood types (Rural, Suburban, and Urban) for all participants. *Bottom*, violin plots depicting the distribution of CFMT scores for each neighborhood type, along with jittered individual data points. Lines indicate the mean CFMT scores for each neighborhood type. The adjusted-R² values are from the models described in the text and Table 1A–B. * = p-value < 0.05 based on statistical tests reported in Table 1A–B. Color coding is based on the self-identification of participants as either female (red) or male (blue).

Table 1A:

Hometown Population Density

Predictor	Effect	R^2 (adj.)	F	df1	df2	p
Hometown Population Density	Model fit	0.02	6.38	3	1046	2.77E-04
	Density		0.29	1		0.59
	Gender		15.03	1		1.13E-04
	Density*Gender		3.25	1		0.07

Open in a new tab

Childhood experience.

The null result we observed for hometown population density as a predictor of CFMT scores does not preclude the possibility that other variables associated with childhood experience might predict adult face recognition performance. Therefore, we also conducted the same regression analysis for the remaining childhood predictors, with each one taking the place of hometown population density, and with gender and the interaction term as the other predictors. Of these other predictors, the one we believed was most closely associated with hometown population density, and highlight here, was whether participants described their hometown neighborhoods as urban, suburban, or rural (Table 1B). The fit of the hometown neighborhood regression model was significant and predicted a small portion of the variance, albeit approximately four times as much as the population density regression model (adjusted-R² = 0.08). Neighborhood type, gender, and their interaction, were all significant predictors of CFMT scores. This interaction effect appears to be driven by higher CFMT performance among participants who grew up in a suburban hometown, especially amongst male observers (Figure 2B). However, the current neighborhood type was identical to hometown neighborhood type for 82% of participants, making the two variables difficult to untangle. Therefore, we applied the same regression model for the data of participants age 18 – 30, again the same range as the study of Sunday and colleagues. We reasoned that an effect observed within this range would be more likely to reflect the impact of neighborhood type early in live. The resulting regression model predicted a medium portion of the variance, and again this effect was driven by hometown neighborhood type (Supplemental Figure 1B, Supplemental Table 3B).

Table 1B:

Hometown Neighborhood

Predictor	Effect	R^2 (adj.)	F	df1	df2	p
Hometown Neighbor-Hood Type	Model fit	0.08	19.77	5	1044	7.33E-19
	Neighborhood		35.80	2		9.20E-16
	Gender		7.01	1		8.21E-03
	Neighborhood*Gender		5.26	2		5.35E-03

Open in a new tab

The full results for the linear regression models that included the twelve other childhood experience predictors can be found in the supplemental material (Supplemental Figure 2,Supplemental Table 4). All the linear regression models that included these predictors were significant (after correcting for multiple comparisons), although the explained variance amongst the other childhood experience predictors was consistently small (adjusted-R² = 0.01 – 0.07). Gender was also a significant predictor for all of the models. The significant childhood experience predictors included: house type, household size, number of siblings, growing up with pets, attending daycare, attending preschool, attending afterschool programs, and playing sports. Of these, house type was the strongest predictor, which seemed to be driven by higher CFMT scores amongst participants who grew up in a house (rather than townhouse/duplex or apartment). The gender interaction term was also significant for house type and attending preschool.

Sociality.

Sunday and colleagues considered extraversion, from the Big Five personality model, as a predictor of CFMT scores, following on previous work (Lander & Poyarekar, 2015; Li et al., 2010). We sought to test a wider range of predictors related to sociality including the other components of Big Five personality, as well as measures related to social intelligence, and shyness.

Table 2A:

Ten Item Personality Inventory (TIPI)

Predictor	Effect	R^2 (adj.)	F	df1	df2	p
Extroversion	Model fit	0.03	12.49	3	1046	5.03E-08
	Extraversion		21.58	1		3.82E-06
	Gender		13.58	1		2.40E-04
	Eitraversion*Gender		0.00	1		0.98

Agreeable-Ness	Model fit	0.09	35.55	3	1046	7.06E-22
Agreeable-Ness	Agreeable		89.70	1		1.78E-20
	Gender		431	1		0.04
	Agreeable*Gender		0.05	1		0.82

Conscientiousness	Model fit	0.08	31.68	3	1046	1.32E-19
Conscientiousness	Conscientious		75.84	1		1.18E-17
	Gender		12.45	1		4.36E-04
	Conscientious*Gender		2.50	1		0.11

Neuroticism	Model fit	0.04	14.67	3	1046	2.32E-09
	Neuroticism		12.09	1		5.27E-04
	Gender		18.86	1		1.55E-05
	Neuroticism*Gender		15.94	1		7.00E-05

Openness to Experience	Model fit	0.07	27.98	3	1046	2.10E-17
	Openness		64.51	1		2.57E-I5
	Gender		10.01	1		1.60E-03
	Openness*Gender		2.86	1		0.09

Open in a new tab

Figure 3. — (A) TIPI personality components as predictors of CFMT scores. Scatter plots are for each of the personality components and CFMT scores for all observers. Lines are the fits of generalized linear models (GLM) with 95% confidence intervals. The depicted GLMs are for visualization purposes and are *not* the same as the linear regression models described in the text and Table 2A. (B) TSIS social intelligence components as predictors of CFMT scores. Plotting conventions are the same as for (A). The depicted GLMs are not the same as the linear regression models described in the text and Table 2B. (C) RCBS shyness measure as a predictor of CFMT scores. Plotting conventions are the same as for (A). The depicted GLM is not the same as the linear regression model described in the text and Table 2C. The adjusted-R² values for each plot are from the models described in the text and Table 2A–C. * = p-value < 0.05 based on statistical tests reported in Table 2A–C. Color coding is based on the self-identification of participants as either female (red) or male (blue).

Second, linear regression was conducted with each of three components of TSIS: social information-processing, social skills, and social awareness. In each model an individual TSIS measure and gender were included as predictors. Models were constructed for each component along with gender and an interaction term as the other predictors (Table 2B). All models were significant, however those including social information processing (adjusted-R² = 0.02) and social skills (adjusted-R² = 0.03) predicted a small portion of the variance. In contrast, the social awareness model predicted a medium portion of the variance in CFMT scores (adjusted-R² = 0.2).For each model both gender and the TSIS component were significant predictors, but the relationships of the components with CFMT scores differed (Figure 3B). In the case of social information processing and CFMT scores, there was a slightly negative relationship. In the case of social skills, there was a positive relationship and a significant interaction with gender. In the case of social awareness, there was a robust positive relationship with CFMT scores.

Table 2B:

Tromso Social Intelligence Scale (TSIS)

Predictor	Effect	R^2 (adj.)	F	df1	df2	p
Social Information Processing	Model fit	0.02	9.58	3	1046	3.06E-06
	Processing		12.08	1		5.30E-04
	Gender		16.70	1		4.72E-05
	Processing*Gender		0.90	1		0.34

Social Skills	Model fit	0.03	10.41	3	1046	9.45E-07
Social Skills	Skills		10.77	1		1.07E-03
	Gender		16.75	1		4.59E-05
	Skills*Gender		4.68	1		0.03

Social Awaremess	Model fit	0.20	86.42		1046	5.88E-50
Social Awaremess	Awareness		240.13	1		6.53E-49
	Gender		4.94	1		0.03
	Awareness*Gender		0.01	1		0.92

Open in a new tab

Lastly, linear regression was carried out with the shyness measure of RCBS, gender, and their interaction as predictors and CFMT score as the dependent variable (Table 2C). Overall, the regression model was significant, and predicted a small portion of the variance (adjusted-R² = 0.03). In the model, shyness, gender, and their interaction were also all significant predictors. The interaction was driven once more by the male participants, as there was a negative relationship between shyness and CFMT scores (Figure 3C). In other words, male participants who rated higher in shyness tended to have lower CFMT scores, while there was no relationship between CFMT scores and shyness for female participants.

Table 2C:

Revised Cheek and Buss Shyness Scale (RCBS)

Predictor	Effect	R^2 (adj.)	F	df1	df2	p
Shyness	Model fit	0.03	10.21	3	1046	1.25E-06
	Shyness		6.90	1		8.76E-03
	Gender		17.63	1		2.91E-05
	Shyness*Gender		7.96	1		4.87E-03

Open in a new tab

Most of the sociality predictors showed medium to large pairwise correlations with each other. The exceptions were extraversion and social information processing, which tended to have small correlations with the other predictors, and shyness, which was significantly negatively correlated with all other sociality measures (Supplemental Figure 3A). Among the different sociality predictors, three of the TIPI components (agreeableness, conscientiousness, and openness to experiences) and one TSIS component (social awareness) all showed a consistent positive relationship with CFMT scores, regardless of gender, and predicted a small to medium portion of the variance (Figure 3A–B). There were also medium to large pairwise correlations between these four predictors (Supplemental Figure 3A). Therefore, we performed a post-hoc four-variable commonality analysis to determine the proportion of variance uniquely explained by these “top four” sociality predictors. The full regression model containing the four predictors explained a medium amount of the variance, with social awareness uniquely accounting for 41.6 % of the explained variance, compared to the three TIPI components which individually each uniquely explained < 3 % of the variance (Supplemental Figure 3B).

Comparing classes of variables.

The childhood experience and sociality predictors both explained some amount of the variance in CFMT performance, as did gender, a demographic predictor. In the interest of determining how much each type of predictor uniquely explained variation in CFMT scores across participants, we conducted a post-hoc variance partitioning analysis. For this analysis we only included predictors from each of the three classes that was individually a significant predictor of CFMT score. Amongst demographic predictors this included gender, ethnicity, education, and current neighborhood (4 of 6). Amongst childhood experience predictors this included hometown neighborhood, house type, household size, number of siblings, number of older siblings, pets, attending daycare, attending preschool, attending after school programs, and participating in sports (10 of 14). All sociality predictors were significant individually (9 of 9). A multiple regression model including all three sets of predictors accounted a large portion of the variance in CFMT scores across participants (R² = 0.26, F(30,1019) = 13.45, p = 4.14e-55). 48.5 % of the predicted variance was uniquely captured by the set of sociality predictors, the demographic and childhood experience predictors uniquely captured 1.3 % and 2.6 % of the variance, respectively (Figure 4).

Figure 4. — Results of variance partitioning (commonality analysis) for the classes of predictors (demographics, childhood experience, and sociality). The (adjusted) coefficient of determination (adj.-R²) indicates the total proportion of explained variance for the full model. Regions of the Euler plots indicate percentages of the explained, and not total, variance accounted for by each component. Regions of overlap indicate proportions of shared variance while regions of non-overlap indicate uniquely explained variance. * = p-value < 0.05.

Summary.

We attempted to reproduce the findings of Sunday and colleagues (Sunday, Patel, et al., 2019) who found that hometown population density and gender interacted to predict individual variation in CFMT scores. To do so we collected CFMT scores and hometown ZIP codes from a large online sample that was an order of magnitude greater than their study. We found no evidence of the reported interaction in our sample. We also considered a number of other predictors related to childhood experience, which did account for some of the variance in CFMT scores. One of these, the type of hometown neighborhood, is at least conceptually related to population density. However, these predictors tended to individually account for small portions of the variance (e.g. R² < 0.09). We also considered a number of predictors related to sociality, based on measures of personality (TIPI), social intelligence (TSIS), and shyness (RCBS). Of all these sociality measures, the TSIS social awareness component was by far the best predictor of CFMT scores. Social awareness also uniquely accounted for a large portion of the explained variance when compared to other sociality predictors. We also carried out variance partitioning analysis to determine the relative contributions of the three sets of predictors, demographics, childhood experiences, and sociality. Collectively the sociality measures accounted for the most unique variance in CFMT scores and overlapped considerably with the other two sets of predictors. Overall, these results suggest that the childhood experience predictors we considered contributed very little to predicting adult face recognition performance when compared to measures related to individual sociality.

Experiment 2

In our second experiment we sought to reproduce the pattern of findings observed in Experiment 1, using a different face memory task. Sunday and colleagues (2019) used a variety of tasks (e.g. Vanderbilt Expertise Test (VET)-Female (Ryan & Gauthier, 2016)), including ones that used female faces and stimuli in contrast to the CFMT, and still observed an interaction between hometown population density and gender. We found a consistent effect of gender as a predictor of CFMT scores, with face recognition performance for female participants generally being higher than male participants and, in some cases, interacting with some of the sociality predictors (e.g. neuroticism and shyness). However, one cannot draw general conclusions about gender differences in face recognition performance when participants only perform tasks using male-gendered faces (Arrington et al., 2022). Thus, given these findings from Sunday and colleagues and the results of Experiment 1, we sought to determine whether these results would replicate using the F-CFMT, which is structurally identical to the CFMT but uses female rather than male-gendered faces (Arrington et al., 2022).