Individual differences in face identity processing

Perception and recognition of face identity is critical in many real-life contexts, including the identifications made by eye-witnesses and the inspection of passports or identity cards. The abilities involved in such tasks demand at least three things; ability to perceive identity in unfamiliar faces, ability to learn new faces, and ability to recognise familiar faces. Previous studies have shown substantial individual differences in such abilities, but without investigating all of them at the same time. We therefore investigated the relationships between individual differences in the performance of tasks that assess these different aspects of face-identity processing, using the Glasgow Face Matching Test (GFMT) as a measure of unfamiliar face perception, the Cambridge Face Memory Test (CFMT) as a measure of new face learning, and the Before They Were Famous task (BTWF) as a measure of familiar face recognition. These face-identity processing measures were integrated into two separate studies that investigated how they were themselves related to other factors. For Study 1 we devised a questionnaire-based measure of subjective performance in face perception and recognition. In Study 2 we used additional measures of perceptual and cognitive abilities, and personality factors, to place individual differences across our principal measures in a broader context. Our findings were consistent with the existence of a previously hypothesised general face-perception factor, but also showed that other influences are clearly operating, highlighting the potential for different aspects of face-perception abilities to associate with more general tasks in quite specific and differentiated ways.

The ability to perceive and recognise face identity is critical to real-life tasks ranging from eye-witness identification to passport control. Previous studies have consistently shown that such tasks reveal a wide distribution of ability in the normal population, and that training and experience do not much alter this (Burton, White, & McNeill, 2010; Dowsett & Burton, 2015; White, Kemp, Jenkins, & Burton, 2014; White, Kemp, Jenkins, Matheson, & Burton, 2014). The origins and nature of these individual differences have, therefore, become an issue with both theoretical and practical significance.

In terms of underlying theory, a key question in understanding face-identity recognition concerns the extent to which it can be considered a unitary process. Ways in which face recognition has already been considered non-unitary involve differences between perceiving and remembering individual face identity, and between the processing of familiar and unfamiliar faces. Since the theoretical paper by Bruce and Young (1986), both distinctions have been incorporated into widely discussed experimental paradigms, standardised tests, and cognitive models.

Much of this previous work was focussed on a case study approach based on examples from the extremes of ability, such as impairments following brain injury (Barton, 2008; Barton & Corrow, 2016; Young, 2011). In such studies, differences between the processing of the identities of familiar and unfamiliar faces are evident from double dissociations between neuropsychological impairments affecting unfamiliar face matching (e.g. in the Benton Test of Facial Recognition: Benton, Hamsher, Varney, & Spreen, 1983) and impairments of familiar face recognition (Benton, 1980; Malone, Morris, Kay, & Levin, 1982; Young, Flude, Hay, & Ellis, 1993; Young, Newcombe, de Haan, Small, & Hay, 1993). Likewise, the importance of new face learning is evident from neuropsychological cases where pre-morbidly familiar faces can be recognised but faces that have only been encountered since the brain injury go unrecognised (Hanley, Pearson, & Young, 1990; Hanley, Young, & Pearson, 1991; Ross, 1980).

While this tactic of studying dissociations between acquired deficits has been important and influential, it was often derived from an implicit assumption that there would be little variability in performance across the normal population. However, more recently it has become clear that there is actually quite a wide distribution of normal ability; especially for tasks involving unfamiliar faces. Hence, an individual-differences approach can be informative, in which one asks whether or not two putatively different types of face-identity recognition ability are associated with each other.

An individual-differences approach is now facilitated by the development of standardised tests. For example, the Glasgow Face Matching Test (GFMT: Burton et al., 2010), the Benton Test of Facial Recognition (Benton et al., 1983), and the Cambridge Face Perception Test (Duchaine, Germine, & Nakayama, 2007) have all been used to investigate the perception of unfamiliar face identity without any face-memory component, because they involve matching the identities of simultaneously presented images of unfamiliar faces. In contrast, the Cambridge Face Memory Test (CFMT: Duchaine & Nakayama, 2006) and the Warrington Recognition Memory Test (Warrington, 1984) look at the ability to recognise previously unfamiliar faces that have been learnt during the testing session. From these available tests we selected the widely used GFMT and CFMT to investigate potential differences between perceiving (GFMT) and remembering (CFMT) the identities of previously unfamiliar faces.

A recent study by Verhallen et al. (2017) found significant correlations between performance on the GFMT, CFMT, and two other measures of unfamiliar face perception. Similarly, Bowles et al. (2009) found a correlation between CFMT and the Cambridge Face Perception Test, and Davis, Lander, Evans, and Jansari (2016) reported a correlation between performance on the GFMT and an immediate memory test developed from the matching arrays used by Bruce et al. (1999). Verhallen et al. (2017) noted that this pattern of findings fits the idea of a general face perception factor, f, that might be considered analogous to Spearman's (1927) g in studies of measures of intelligence. At the same time, however, Verhallen noted that f should not be reified from a pattern of correlations alone, and that it can in any case account for only a proportion of the variance across face tests, with much unshared variance that still needs to be explained. A similar overall conclusion can be drawn from a previous study by Wilhelm et al. (2010), who used a different set of tasks to demonstrate an inter-relationship between measures of face processing that was largely distinct from other cognitive abilities.

Our aim here was to pursue a similar logic to Verhallen et al. (2017) while broadening the range of measures to include familiar face recognition. In terms of the possibility of differences between familiar and unfamiliar face recognition, much less is known from an individual differences perspective. In general, familiar face recognition is so much better than unfamiliar face recognition that it can be difficult to see how to measure individual differences in familiar face recognition (Young & Burton, 2017). For instance, familiar faces are so easily recognised across a wide range of views that identification performance is virtually at ceiling and matching tasks with familiar faces seem almost trivially easy (Jenkins, White, Van Montfort, & Burton, 2011). Nonetheless, it remains the case that people do make sporadic errors in familiar face recognition both in everyday life and in laboratory tasks (Hay, Young, & Ellis, 1991; Young, Hay, & Ellis, 1985) and also that some individuals (including at least one of the current authors) subjectively report experiencing more difficulty than others. So it is plausible that it might be possible to measure individual differences in familiar face recognition if a sufficiently demanding task can be devised.

One potentially promising measure of familiar face recognition is the Before They Were Famous task (BTWF: Russell, Duchaine, & Nakayama, 2009), which asks participants to recognise the identities of faces they know well from images of the same people before they came to public prominence. Performance on the BTWF has been found to correlate with CFMT (Wilmer et al., 2010), but the BTWF was designed for use with American participants, and many famous faces are culture-specific. We therefore used a slightly adapted version of the BTWF with faces appropriate for participants in the UK to explore individual differences in familiar face recognition. In their previous study, Russell et al. (2009) noted a significant association between the BTWF and the CFMT. Likewise, Davis et al. (2016) reported a correlation between a face-memory test and a famous face recognition test, though their famous face-recognition test involved visually degraded celebrity images rather than before-fame photos.

We therefore used an individual-differences approach to face-identity recognition based around three principal measures; the GFMT as a measure of unfamiliar face-identity perception, the CFMT as a measure of new face learning, and the BTWF as a measure of familiar face recognition. To the extent that these tap common processes such as Verhallen et al.’s (2017) f, we would expect correlations between the different measures.

These three principal measures were integrated into two separate studies that looked at how they were themselves related to other factors. For Study 1 we devised a questionnaire-based measure of subjectively experienced problems in face-identity perception and recognition. With the growing focus on individual differences in face-identity recognition, researchers have also begun to examine the extent to which individuals are aware of their own expertise (for a review, see Palermo et al., 2017). Despite this growing interest, approaches have been inconsistent and the findings contradictory. For face-memory tests, while Rotshtein, Geng, Driver, and Dolan (2007) found no correlation between a single self-report question and recognition accuracy, whereas Palermo et al. (2017), Gray, Bird, and Cook (2017) and Livingston and Shah (2017) did find some correlation between overall scores on multi-item face-metacognition questionnaires and CFMT performance. For face-matching tasks Bindemann, Attard, and Johnston (2014) reported no correlation between self-report ratings and performance on an unfamiliar face-matching task, whereas Verhallen et al. (2017) reported a significant correlation between a single self-report question and GFMT accuracy. To try and reconcile these discrepant findings, our questionnaire included both a general question and task-specific questions chosen to relate directly to the GFMT, CFMT and BTWF measures. Moreover, in contrast to several previous studies, our questions explicitly distinguished between familiar and unfamiliar faces.

In Study 1 we used a small and relatively homogeneous student participant sample to explore whether correlations between different face tests will be reasonably robust. For Study 2 we used a larger, more diverse sample of participants and additional measures of face perception, visual perception, cognitive abilities and personality factors to place individual differences across our principal measures in a broader context. Again, previous results have been mixed, but offered hints of possible relationships. Performance with unfamiliar faces has been associated with object-matching ability (Megreya & Burton, 2006), with perceptual style or field dependence (Hoffman & Kagan, 1977; Messick & Damarin, 1964; Witkin & Goodenough, 1977), and with space perception (Burgess, Alderman, Evans, Emslie, & Wilson, 1998). Personality traits of extraversion and empathy have been reported as positively associated with unfamiliar face recognition performance, whereas anxiety and neuroticism have been negatively associated with both unfamiliar face recognition and unfamiliar face matching (Bate, Parris, Haslam, & Kay, 2010; Li et al., 2010; Megreya & Bindemann, 2013; Mueller, Bailis, & Goldstein, 1979; Nowicki, Winograd, & Millard, 1979).

In Study 1, then, we provide the first assessment of individual differences in identity processing across three widely established tests that measure different aspects of face-identity recognition; the GFMT for unfamiliar face matching, CFMT for face learning, and the BTWF for familiar face recognition. In addition, we assessed awareness of one’s own aptitude on these tasks using both general and task-specific questions that dissociated face familiarity. In Study 2, using a more comprehensive battery of perceptual, cognitive, and personality measures, we sought to assess the extent to which aptitude with faces is domain-specific or correlates with other forms of individual difference.

In this study we measured the extent to which accuracy on tests of unfamiliar face matching (GFMT), memory for newly learnt faces (CFMT), and familiar face recognition (BTWF) were associated in a student population. In addition, we created a questionnaire-based subjective measure to determine the extent to which participants’ estimates of their ability with faces was correlated with the more objective measures.

Five participants were removed from all analyses either as a result of software failure during data collection (three participants), or because they scored more than 2.5 SDs below the mean on the BTWF familiarity check (two participants). In addition, a sixth participant failed to complete the Face Recognition Ability Questionnaire, leaving 34 participants for the analyses involving the questionnaire data.

For the GFMT, mean unfamiliar face matching accuracy was 82% (SD = 11%). Mean CFMT face memory accuracy across all items in the task was 74% (SD = 14%). These levels of performance are in line with published norms (GFMT, Burton et al., 2010; CFMT, Russell et al., 2009). For the BTWF, the familiarity check established that the mean number of known celebrities was 70% (SD = 21%). Mean familiar face recognition accuracy for the known celebrities from the images before they were famous was 27% (SD = 13%). This recognition rate is line with previously published work by Bindemann et al. (2014), and the range of scores is consistent with those reported by Russell et al. (2009).

As our principal aim was to explore potential correlations between different measures, we were more concerned to avoid type 2 than type 1 errors, and therefore begin by reporting uncorrected statistics. As a check on the reliability of these, however, we also used the Benjamini-Hochberg procedure with a false discovery rate of 0.2 to correct for multiple comparisons, and we also report confidence intervals. Fig. 1 provides scatterplots showing the correlation between performance on the three face-identity processing tasks. Significant positive Pearson correlations were found between the GFMT and CFMT (r(33) = .45, uncorrected p = .006, 95% CI [0.14, 0.68]), GFMT and BTWF (r(33) = .37, uncorrected p = .027, 95% CI [0.04, 0.62]), and CFMT and BTWF (r(33) = .53, uncorrected p = .001, 95% CI [0.24, 0.73]), showing some consistency in performance. All of these correlations remained significant with the Benjamini-Hochberg correction.

In Study 1 we found that performance across unfamiliar face matching (GFMT), unfamiliar face learning (CFMT), and familiar face recognition (BTWF) were significantly correlated in a student sample.

Study 2 sought to replicate and extend the findings of Study 1 with a larger participant sample with a wider demographic and educational background, and at the same time to investigate the relationship to other measures.

We added two face-perception tasks that did not require recognition of individual identity (Face-detection, and Mooney faces), allowing us to explore their relationship to the identity-related face tasks (GFMT, CFMT, and BTWF). These additional face-perception measures were chosen because they tap theoretically important aspects of face perception without specifically requiring processing of face identity per se. Face detection was used because, since Ellis (1983), some theories have assumed that detecting faces in cluttered visual scenes forms an essential ‘entry point for all other tasks with faces’ (Bindemann & Lewis, 2013, p.1144). The Mooney face task is of interest because, even though it does not require perception of unique identity, it has been linked to genetic differences in ability (Verhallen et al., 2014) that may also be tapped by the CFMT (Shakeshaft & Plomin, 2015). Moreover, the Mooney face task has been found to correlate with CFMT in a previous study (Verhallen et al., 2017).

A battery of more general tasks was also used to identify the extent to which different face-processing tasks might associate with other types of task in relatively specific ways, using a similar logic to Wilhelm et al. (2010). These tasks included measures of visual space perception, visual object perception, executive function, and personality. As noted in the ‘Background’ section, although previous results have been mixed, these are all areas that have offered hints of possible relationships.

We began by examining the correlations between performance on the three face-identity tasks that had also formed the core of Study 1(GFMT, CFMT, and BTWF). Scatterplots are shown in Fig. 3. As for Study 1, significant positive correlations were found between the GFMT and CFMT (r(105) = .50, uncorrected p < .001, 95% CI [0.35, 0.63]), GFMT and BTWF (r(105) = .20, uncorrected p = .039, 95% CI [0.01, 0.37]), and CFMT and BTWF (r(105) = .33, uncorrected p < .001, 95% CI [0.15, 0.49]). However, the correlations with the BFTW task were somewhat lower than for Study 1, and only the correlations between GFMT and CFMT and between CFMT and BTWF remained significant following the Benjamini-Hochberg correction. Figure 3 shows clearly that correlations involving BTWF were likely reduced because a number of participants in this more heterogeneous sample were at floor on the adapted BTWF; 20% of participants failed to recognise any of the celebrities as children, despite recognising a mean of 59% (SD = 24%) of the celebrities as adults.

The most striking aspect of Table 4 is the apparent lack of clear associations between our face tasks and other variables. In particular, none of the non-facial tasks was correlated with all of the face-identity tasks. This pattern is strikingly different to the consistent intercorrelations between the face-identity tasks themselves (as shown in Figs. 1 and 3), suggesting that if there is a generic face factor (f) it represents something that is relatively distinct from other abilities.

Nonetheless, a few correlations did reach significance at both uncorrected and corrected levels. Some of these seem to reflect relatively isolated influences, such as the correlation between letter detection and face detection (r(105) = .27, uncorrected p < .01, 95% CI [0.09, 0.44]) and the correlation between BORB Gap Detection and Mooney faces (r(105) = .33, uncorrected p < .001, 95% CI [0.15, 0.49]), but two more general patterns also stand out.

First, inspection of the data in terms of the columns in Table 4 shows that unfamiliar face matching (the GFMT) correlates with other matching tasks that require careful local processing (Navon Local, r(105) = .27, uncorrected p = .004, 95% CI [0.09, 0.44]; Matching Familiar Figures, r(105) = .37, uncorrected p <. 001, 95% CI [0.19, 0.52]) and with our test of executive function (BADS Card Sorting, r(105) = .20,uncorrected p = .035, 95% CI [0.01, 0.38]), though the latter correlation was not significant after correction for multiple comparisons. This pattern is consistent with the idea that unfamiliar face matching is often done using feature processing strategies that can be bolstered by a degree of careful strategic control and systematic checking of individual features (Megreya & Burton, 2006).

Second, inspection of the data across rows in Table 4 shows that the measure of visual object recognition (VOSP Silhouettes) correlates with the measures of familiar face recognition (BTWF, r(105) = .29, uncorrected p = .003, 95% CI [0.10, 0.45]) and face learning (CFMT, r(105) = .29, uncorrected p = .002, 95% CI [0.11, 0.45]). It also correlates with the Mooney face task (r(105) = .50, uncorrected p < .001, 95% CI [0.34, 0.63]), but this is perhaps unsurprising as both tasks have a requirement that involves interpreting high-contrast images.

Although correlations between measures of face perception and personality have been reported in some previous studies (Bate et al., 2010; Li et al., 2010; Megreya & Bindemann, 2013; Mueller et al., 1979; Nowicki et al., 1979), Table 4 does not show strong relationships.

Response times were also available for some measures (Navon global, Navon local, Matching Familiar Figures, BADS Card Sorting), but these did not show significant correlations with any of the accuracy measures and hence are not discussed further here.

Our aims were to investigate the relationships between individual differences in the performance of three tasks that assessed different aspects of face-identity processing, and to investigate the relationships between these tasks and other measures of face perception and broader perceptual, cognitive, and personality measures.

Our three principal measures of face-identity processing involved the GFMT as a measure of unfamiliar face perception, the CFMT as a measure of face learning, and the BTWF as a measure of familiar face recognition. In both Study 1 and Study 2, performances were significantly correlated across these tasks, consistent with some degree of commonality. This is the first time that this has been shown across unfamiliar face matching, unfamiliar face learning, and familiar face recognition at the same time.

While our sample sizes were small for definitive correlation-based research (Schönbrodt & Perugini, 2013), the significant intercorrelations between the face measures across both studies nonetheless lend further support to Verhallen et al.’s (2017) concept of a general face-perception factor, f. In fact they extend the range of applicability of the idea since Verhallen et al. did not include a measure of familiar face recognition. That said, the intercorrelations we noted could at best account for around 25% of the variance in test scores, and this is consistent with Verhallen et al.’s (2017) finding that the upper limit on the proportion of variance explained in their study was 23%. So even though it may be appropriate to think in terms of a general face factor, there seem also to be other influences operating.

Our three principal measures of GFMT, CFMT, and BTWF were also integrated into separate studies that looked at how they were themselves related to other factors. For Study 1 we devised a questionnaire-based measure of subjective problems in face perception and recognition. Of the six different questions we devised, only the question concerning familiar face recognition (question 3, see Table 1) led to significant correlations with our measures of face recognition (BTWF) and face learning (CFMT). It seems that participants may at least have some insight into their familiar face-recognition abilities, while having little insight into their unfamiliar face recognition. This in itself is not an entirely new finding; Bindemann et al. (2014) also noted greater associations between perceived ability and accuracy by splitting predictions across unfamiliar and familiar face tasks. The pattern is consistent with the interesting theoretical idea that it is the importance of familiar face recognition to our everyday lives that leads many of us to fail to appreciate just how error-prone tasks like unfamiliar face matching can be (Ritchie et al., 2015; Young & Burton, 2017).

In Study 2 we used additional measures of face perception, visual perception, cognitive abilities, and personality factors to place individual differences across our principal measures in a broader context. While Studies 1 and 2 both found that face-identity processing tasks were significantly associated with each other, Study 2 nonetheless showed that they still seemed to involve some specific components. There were no general tasks or measures that were associated with all of the face-identity processing tasks, suggesting that any general face factor (f) may represent something that is relatively distinct from other abilities. However, we noted that the hypothesised f can at best explain around 25% of the variability in face-task scores. Consistent with this, there was a smattering of individual tasks and measures that did associate with specific face-processing tasks.

Of particular note, unfamiliar face matching (the GFMT) correlated with other matching tasks that require careful local processing (Navon Local, Matching Familiar Figures) or executive function (BADS Card Sorting). We interpret this pattern as consistent with the idea that unfamiliar face matching is often done using feature processing strategies that can be bolstered by a degree of careful strategic control (Megreya & Burton, 2006).

Also noteworthy is that a measure of visual object recognition (VOSP Silhouettes) correlated with the measures of familiar face recognition (BTWF) and face learning (CFMT), suggesting that it may be worth casting the net more widely to search for a generic visual recognition ability.

Taken together, then, these findings are consistent with the existence of a general face-perception factor. The three main face tasks studied, GFMT, CFMT, and BTWF, are different both in their superficial presentation and in their demand characteristics. Nevertheless, we find that these tests are consistently associated with each other, but only in rather specific case-by-case ways with non-face tests. This lends support to the idea of a general ability with face perception (f) that underlies performance with a wide range of face tasks.

Despite this finding, it is important to emphasise the degree of influence of a putative general face factor. Our observations, and those of Verhallen et al. (2017) suggest that such a factor can account for a maximum of around 25% of the variance across the face tasks studied. That clearly leaves us with much more to understand. In future research it will be important to establish which aspects of face perception have inherently divergent processes, and which are the result of an individual perceiver’s unique history. Studies with much larger cohorts than we have managed to achieve here are becoming possible, largely through on-line testing. It is, therefore, becoming tractable to use an individual-differences approach to study more subtle patterns of association between face-perception tasks and their subcomponents. Such an approach seems very promising in the attempt to understand the relationship between face processing and other human characteristics.

Understanding the nature of individual differences in ability to perceive and recognise face identity is of importance in real-life contexts ranging from eye-witnessing to passport control. We investigated the relationships between individual differences in the performance of tasks that assessed different aspects of face-identity processing, and the relationships between these tasks and other measures of face perception and broader perceptual, cognitive and personality measures. Our findings were consistent with the existence of a previously hypothesised general face-perception factor suggested by Verhallen et al. (2017), but this was at best able to account for around 25% of the variance in scores. Other influences are also clearly operating, highlighting the potential for different aspects of face-perception abilities to associate with more general tasks in quite specific and differentiated ways.