Where are kin recognition signals in the human face?

Abstract

We report two experiments that aimed to determine where in the face the cues that signal kinship fall. In both experiments, participants were shown 30 pairs of photographs of children's faces. Half of the pairs portrayed siblings and half did not. The 220 participants were asked to judge whether each pair of photographs portrayed siblings. We measured the effect on kin recognition performance of masks that covered the upper half or the lower half of the face ( Experiment 1) and the eye region or the mouth region ( Experiment 2). In Experiment 1, we found that the signal detection estimate of performance d′ decreased only 5.3% ( ns) when the lower face was masked but by more than 65% when the upper face was masked. We tested whether the combination of kinship information from the two halves of the face can be treated as optimal combination of independent cues and found that it could be. In Experiment 2 , we found that masking the eye region led to only a 20% reduction ( ns) in performance whereas masking the mouth region led to a nonsignificant increase in performance. We also found that the eye region contains only slightly more information about kinship than the upper half of the face outside of the eye region.

Introduction

Kin recognition is thought to play an important role in the social organization of species (DeBruine, 2002, 2004, 2005; Platek, Keenan, & Mohamed, 2005; Platek et al., 2004). The “inclusive fitness” theory of Hamilton (1964a, 1964b) presupposes that organisms have some ability to recognize their own kin and to discriminate different degrees of relatedness in their relatives (Mateo, 2002). Recognition of kin to self through self-referential phenotype matching has been documented in a variety of species—from bacteria to primates (e.g., Chapais & Berman, 2004; Fletcher & Michener, 1987; Hepper, 1991). 

Additionally, some primates are able to recognize closely related individuals of their own species that are not closely related to themselves (Parr & De Waal, 1999; Parr, Winslow, Hopkins, & de Waal, 2000). Cheney and Seyfarth (1989, 1999), for example, found that vervet monkeys, after witnessing a fight between a relative and an individual of another family, will become aggressive toward other members of this individual's family. 

Human observers, also, can detect relatedness between individuals unrelated to themselves: in parents and children pairs (Brédart & French, 1999; Bressan & Dal Martello, 2002; Bressan & Grassi, 2004; Christenfeld & Hill, 1995; Nesse, Silverman, & Bortz, 1990) and in sibling pairs (Maloney & Dal Martello, 2006). Participants, in all of these studies, were well above chance in classifying individuals as near relatives or not, given only photographs of their faces. Humans can even detect relatedness in pairs of individuals of another primate species (Vokey, Rendall, Tangen, Parr, & de Waal, 2004). 

The primary aim of the study was to determine where the kin signals used by human observers are in the human face. We concentrate on collateral kin recognition, specifically classification of pairs of children as siblings or nonsiblings. There are no other studies (to our knowledge) in the kin recognition literature about the spatial location of kin signals in the face. There are studies on face recognition that look for the relative importance of parts of the face in recognition tasks. Goldstein and Mackenberg (1966) and Fisher and Cox (1975), for example, found that the upper features were more important than the bottom ones for face recognition. It is interesting to consider the hypothesis that the same regions of the face that signal identity would also signal kinship. 

Studies on face recognition also indicate that the eyes are a very salient feature for identity recognition. Fisher and Cox (1975) and McKelvie (1976) found that omission of the eyes in a test picture of the face was more disruptive to recognition than omission of the mouth. Roberts and Bruce (1988) found that masking the eyes led to a sharper reduction in performance in a sex recognition task than masking the mouth. Davies, Ellis, and Shepherd (1977) investigated the cue salience of internal facial features, changing single features of faces constructed with a Photofit Kit. They found that the eyes were more informative for recognition than the mouth. Chimpanzees' performance in a recognition task of conspecifics' faces deteriorated when only the eyes where masked but did not significantly decrease when only the mouth was masked. On each trial, the primate observers were shown three faces: a sample picture of a face and two comparison pictures of faces, one of which portrayed the same individual as portrayed in the sample picture. They were rewarded if they matched the sample picture to the comparison picture that portrayed the same individual (Parr et al., 2000). 

Keating and Keating (1993) found that rhesus monkeys' recognition of Identi-Kit-built human faces deteriorated only when the eyes or brows were altered. Recognition was not affected by alteration of the other internal features of the face. Measures of fixation patterns showed that the monkeys concentrated attention mostly on the eyes. Kyes and Candland (1987) found a similar visual preference for the eyes when baboons viewed the features of a conspecific's face. 

Our stimuli include pairs of children that are close in age and pairs that are widely separated in age. Given normal facial development in children, these age differences complicate the observer's task. Even if all the traits in the human face are equally and highly hereditable as Kohn (1991) concludes, the time course of development is radically different in different parts of the human face. From the literature on face morphology (Enlow & Hans, 1996), we know that the face below the eyes changes dramatically in size and shape from the age of 1 year to adulthood (Figure 1). 

Figure 1

 

Neonate (A) and adult (B) human skulls. The neonate skull is enlarged to emphasize the relative growth of the cranial vault (blue) and lower face (beige). © Cassio Lynm 2000.

The lower face ¹ becomes longer and larger to accommodate the expansion of the nasal region and the eruption of deciduous and permanent teeth. The cranium, instead, grows much less than the lower part of the face. A neonate's basicranium is 60% to 65% of its final size; it reaches 90% of its full development when the child is 5 to 7 years old. Most of the growth of the cranial vault, like the brain's growth, takes place during the first year of life. The size of the face (below the eyes) at birth is around one eighth of the total size of the head, whereas in a young adult, it is half of it. We emphasize that both neurocranial and lower facial dimensions are highly heritable (Kohn, 1991). They differ in the extent to which they change over the course of childhood. It is plausible to expect that facial cues that signal kinship will be found in parts of the face that vary least across the life span. 

Our first hypothesis, then, was that the cues present in the upper, developmentally stable part of the face will carry most weight in a kin recognition task consistent with previous results in the face recognition literature. In the first experiment, we tested this hypothesis and related hypotheses by asking adults to perform signal detection tasks where the signal is “close genetic relatedness” between two children. Our stimuli are color pictures of children's faces seen from the front. The children's ages varied from 17 months to 15 years, covering a large part of the developmental period. We measure the participants' ability to detect “close genetic relatedness” using standard signal detection methods described below. We expect that kin recognition performance will be disrupted more when the upper half of the face is not visible, and it is of interest to quantify how much each region of the face contributes to kin recognition. 

Our second hypothesis was that the eye region carries many of the cues that signal kinship, consistent with results in the face recognition literature just described, as well as with the consideration that eye size changes little over the course of postnatal development, reaching 70% of their adult size by birth and then growing rapidly in the first 12 months of life (Gossman, Mohay, & Roberts, 1999). In a second experiment, we will contrast ability to recognize kinship when either the mouth region or eye regions of faces are masked and compare performance in both of these conditions to a control, Full-Face condition. 

As a working definition, we define a signal ² as any measurement derived from a face (Tversky, 1977) that carries useful information relevant to a visual task. We recognize that there may be signals that fall entirely under one of our masks but that there may be signals that signal valid information about kinship that are not available when, for example, either the upper or the lower half of the face is masked. An example of such a signal would be eye separation divided by the distance from the center of the line joining the pupils of the eyes to the mouth (a “configurational” feature of the sort proposed in Rhodes, 1988). Masking either the upper or the lower half of the face would make it impossible to estimate this signal accurately. 

Such a “full-face signal” would be important to our analyses here only if it were, first of all, a good cue to kinship (i.e., it really is a kin recognition signal) and, second, the observer gave considerable weight to it in assessing kinship. 

From this viewpoint, in Experiment 1, there are potentially three classes of kin recognition signals that can be differentiated by how masking affects kin recognition: those that are disrupted by masking the upper but not the lower half of the face (“upper signals”), those that are disrupted by masking the lower but not the upper half of the face (“lower signals”), and those that are disrupted by masking either the upper or the lower half of the face (“full-face signals”). If there were no full-face signals in this sense, we would expect that we could predict performance in the unmasked condition from measured performances in the two masked conditions. 

In the first experiment, we will also test whether detection performance in the two masked conditions can be used to predict performance in the unmasked condition using standard methods of signal detection theory. In effect, we are testing whether this third class of full-face signals carries substantial additional information about kinship that is used by observers. 

Experiment 1

Methods

Participants

One hundred nine people, recruited in public places at the University of Padova, were alternately assigned to one of three conditions: Full Face (FF), Lower Half Masked (LHM), or Upper Half Masked (UHM) as described below. There were 18 males and 16 females in the FF condition, 16 males and 19 females in the LHM condition, and 19 males and 20 females in the UHM condition. Their ages ranged from 19 to 36 years (median age, 22 years). 

Photographic material

Seventy-two color photographs, each depicting a child from the neck to the top of the head, were used. The pictures had been taken by the experimenters or their assistants under controlled lighting conditions. Of the 72 children depicted in the pictures, half were girls and half were boys. We used Adobe Photoshop® to obliterate all background detail, replacing it by a uniform dark gray field (33% of maximum intensity in each of R, G, and B channels). The ages of the children ranged from 17 months to 15 years. The facial expressions were neutral or close to neutral. All came from three adjacent provinces of Northern Italy: Padova, Mantova, and Vicenza. All were Caucasian in appearance. The parents of each child gave appropriate permission for their child's photograph to be used in scientific experiments. We asked and received separate parental permission to use the photographs in Figures 2 and 3 as illustrations here. 

Figure 2

 

Experiment 1: experimental conditions. On each trial, observers saw a pair of photographs of children's faces and were asked to judge whether the children were siblings or not. There were three conditions: (A) the FF condition (no mask), (B) the LHM condition, and (C) the UHM condition.

There were three conditions in the experiment. In the FF condition, the pictures were presented without any alteration so that the entire face was visible. In the other two conditions, we altered the pictures by covering certain parts. In the LHM condition, we added a blue trapezoidal figure to each photograph, covering the lower part of the picture and leaving the upper half of the face visible. In the UHM face condition, we covered the upper part of the picture, leaving the lower half of the face visible. The line that divides the face into upper and lower parts was parallel to the line where the eyes lie and passed through the tip of the nose ( Figure 2). 

Picture pairs

Of the 72 photographs, 60 were used in the main part of the experiment. The remaining 12 were used only in the familiarization and training phases of the experiment. These phases will be described below. The 60 photographs used in the main phase of the experiment included 15 pairs of biological siblings and 15 pairs of children who were genetically unrelated. We refer to the pairs in the first group as related and in the second as unrelated. Within each group of 15, five pairs depicted two boys, five pairs depicted two girls, and five pairs depicted a boy and a girl. The 12 photographs used in the familiarization and training stages included three pairs of biological siblings and three pairs of unrelated children. The distributions of age differences for related and unrelated pairs were matched. 

For privacy reasons, we did not verify via DNA fingerprinting whether sibling pairs shared two parents. Recent research using DNA fingerprinting shows that the median rate of “extrapair paternity” is much lower (<2%) than previously thought (see Simmons, Firman, Rhodes, & Peters, 2004, for a review). Consequently, it is unlikely that any large proportion of our siblings share only one parent (who is more likely to be the mother). In any case, the presence of half-siblings would have little effect on the outcome of our experiment. Such half-siblings would have 25% of their DNA in common, rather than 50%, but they would still be more closely related than nonsibling pairs, and their presence in the sample across all conditions should not affect the comparisons across conditions that are central to our analyses. 

Procedure

The experiment was conducted in a computer classroom at the University of Padova. Observers viewed all stimuli on a computer monitor and responded by marking forms provided. The experiment was self-paced and consisted of three phases.

1.  
   Familiarization. The observer was first asked to perform a simple recognition memory task that involved all of the experimental stimuli for their condition. All 72 photographs of faces were shown in groups of six per display in random order. The purpose of this part of the experiment was to familiarize the observers with the range of faces that they would see in the main part of the experiment. The observers were asked to study the display and were told that, immediately after studying each group, they would be shown a probe photograph and would be asked to report whether this photograph had been among the group of six just studied. The probe photographs were the nonexperimental photographs described above that were not used in the main part of the experiment.
2.  
   Training. The observers practiced the response for their condition (FF, UHM, or LHM) on six pairs of photographs that did not overlap with the photographs used in the main part of the experiment. These pairs were drawn from the nonexperimental photographs organized so that there were three pairs that were biological siblings and three that were not. The purpose of this part of the experiment was simply to let the observers become comfortable with the procedure and response required. In this phase, the observers were told that half of the pairs portrayed genetic siblings and were asked to classify each pair as related or not related.
3.  
   Main. The observers then were told that half of the pairs of the set of photographs in the main phase were genetic siblings and half were unrelated children. Their task, as in the training phase, was to classify each pair as related or not related. The 30 pairs of photographs were presented in random order. We used two separate randomizations, and observers were assigned one of such at random.

Results and discussion

Signal detection estimates of sensitivity d′ and likelihood criterion β were used to measure performance in the different conditions (Green & Swets, 1966/1974). These values are reported in Table 1. We will first discuss the d′ values, which are also summarized in Figure 3. 

Table 1

 

Results for Experiment 1, by condition. The d′ estimate and the likelihood criterion β for the signal detection analysis are shown. Standard deviations were estimated by a bootstrap procedure (Efron & Tibshirani, 1993) based on 10,000 replications.

	d′	SD ( d′)	β	SD ( β)
FF	1.187	0.085	0.885	0.046
UHM	0.408	0.075	0.927	0.020
LHM	1.124	0.081	0.893	0.042

Figure 3

 

Experiment 1: results. The d′ measures for the three experimental conditions are plotted as a bar graph. The confidence intervals (shown as 95% confidence intervals) were estimated using a bootstrap procedure (Efron & Tibshirani, 1993) with 10,000 replications. The d′ values for the FF condition and the LHM condition are significantly greater (p < .001) than that for the UHM condition.

In carrying out the hypothesis tests below, it would be appropriate to correct for multiple tests by a Bonferroni correction. However, for all of the tests in this experiment and the succeeding one, a Bonferroni correction would not change any conclusions as the reader can verify. Consequently, we simply present the p values (or bounds on the p values) for each test. 

A d′ value of 0 corresponds to chance performance, and a d′ value of 3.5 corresponds to a practically perfect performance. The d′ value in the FF condition was significantly different from 0 ( z = 14.838, p < .0001, one tailed). The one-tailed test is justified as it is plausible to assume d′ ≥ 0. The participants, when able to observe the entire face, could classify the pairs as siblings or not siblings markedly above chance level. The d′ values are similar to those that we have found in earlier related work (Maloney & Dal Martello, 2006). The d′ value in the LHM condition was also significantly different from 0 (z = 6.486, p < .0001, one tailed). When the lower half of the face was hidden, the observers could still discriminate well between sibling and nonsibling pairs. Performance in the FF and LHM conditions did not differ significantly (z = 0.553, p = .580, two tailed). The absence of information from the lower half of the face resulted in a small decrease in sensitivity that was not statistically significant. 

The d′ value in the UHM condition was significantly different from 0 ( z = 5.440, p < .0001; one tailed): Participants can classify the pairs at a level above chance even if the upper half of the face is hidden. It appears that there are useful signals for kin recognition in the lower half of the face as well. Performance in the UHM condition ( d′ = 0.41) was significantly worse than either the FF condition ( z = −7.104, p < .0001) or the UHM condition ( z = −6.486, p < .0001). Although it was still possible for the observers to detect relatedness at a level above chance when just the bottom half of the face was visible, the performance in this condition deteriorated markedly when compared with the other two conditions. 

We emphasize that, when we fail to reject a null hypothesis, it would not be correct to conclude that the null hypothesis is true or that we have evidence in favor of the null hypothesis. This is the common fallacy of “accepting the null hypothesis” (see Loftus, 1996). When, for example, we do not reject the null hypothesis that the d′ value for the LHM condition was equal to that for the FF condition, we have not shown that the d′ values are equal or that the lower half of the face adds nothing to performance in judging relatedness. We have simply shown that we cannot reliably measure the difference with our experimental design: The difference is correspondingly small, and the best information available for the magnitude of the difference is the estimates of d′ in Table 1 and the accompanying estimates of their standard deviations. 

If the information available in the upper and lower halves of the face were statistically independent and if there were no additional sources of information available only in the FF condition (the “full-face signals” discussed in the Introduction), then we would be able to predict the d′ value for the full face from the d′ values in the other two conditions (Green & Swets, 1966/1974): 

${{\mathrm{d\prime }}^{}}_{\mathrm{FF}\mathrm{}}=\sqrt{({{\mathrm{d\prime }}^{}}_{\mathrm{UHM}\mathrm{}\mathrm{}}{)}^2+({{\mathrm{d\prime }}^{}}_{\mathrm{LHM}\mathrm{}\mathrm{}}{)}^2},$

(1)

where d′_FF, d′_UHM, and d′_LHM are the d′ values in the three corresponding conditions. If we form a prediction of d′_FF based on the values for the other two conditions in Table 1, we find that this prediction,

${\widehat{d}}_{\mathrm{F}\mathrm{F}}^{\prime }$

= 1.196, is remarkably close to the value observed, d′_FF = 1.187, and that the two values are not significantly different (p > .05). Thus, although the d′ value for the FF condition was not significantly different from that for the LHM condition, the small difference observed is consistent with the combination of statistically independent information from the two halves of the face. 

In conclusion, there are useful cues to kinship available when either the upper or the lower half of the face is visible, but the upper half of the face contains almost as much useful information as the entire face. The results are also consistent with the claim that statistically independent information from the two halves of the face is combined in the FF condition and that the contribution of full-face signals in collateral kin recognition is negligible. We will return to this point in the General discussion section. 

The β values reported in Table 1 measure the bias of the response toward classifying as kin or not kin. The β values for each condition are not significantly different from one another ( p > .05). That is, observers did not show any change in bias because of the presence or absence of the masks. Most important, they do not become more cautious (adopting a stricter criterion) when just half of a face is visible. 

Note that the three conditions showed a common bias. We tested the β values against 1 and, in all conditions, the difference from 1 was significant ( p < .01) in all cases. The actual prior odds that the pairs are related are 1:1 (half of the pairs portray siblings), and the observers were given this information. Still, the observers were slightly biased in favor of classifying the pairs as related. That is, they err in the direction of misclassifying unrelated pairs as related (Type 1 error). 

Experiment 2

Introduction

In this experiment, we tried to determine the weight of relatedness cues present in the eyes and in the mouth. As we described above, studies on face recognition indicate that the eyes are more important than the mouth for identity recognition in human and other primate species. 

The literature on face morphology and development describes the eye area as relatively invariant during development (Enlow & Hans, 1996). We hypothesize that by analogy with the identity recognition studies results, the eyes should contain more useful kin recognition cues than the mouth and, consequently, that masking the eye region should lead to a marked drop in performance in kin recognition. We test our hypothesis by measuring kin recognition performance in an experiment with the same design, stimuli, and task used in the first experiment but using smaller masks that covered smaller regions. The masks were sufficient to cover the eyes in one experimental condition and the mouth in the other. The face was fully visible in the control condition. The three conditions are illustrated in Figure 4. 

Figure 4

 

Experiment 2: experimental conditions. On each trial, observers saw a pair of photographs of children's faces and were asked to judge whether the children were siblings or not. There were three conditions: (A) the FF presentation (no mask), (B) the EM condition, and (C) the MM condition.

Methods

Participants

One hundred eleven people were recruited in public places or personally contacted and randomly assigned to three conditions: Full Face (FF), Eyes Masked (EM), and Mouth Masked (MM), with each group containing 37 observers (13 males and 24 females). Their ages ranged from 20 to 31 years (median age, 23). 

Material

The stimuli were the same 30 pairs of photographs as those used in the main phase of Experiment 1. In the FF condition, the photographs were presented without alterations; in the other two conditions, we altered the pictures by adding masks. In the MM condition, we added a blue rectangle to each photograph covering only the lips of the child. In the EM condition, we added a blue rectangle covering only the eyes and eyebrows of the child. We asked and received parental permission to use the photographs in Figures 4 and 5 as illustrations here. 

Procedure

The experiment, like the first experiment, was conducted in a computer classroom at the University of Padova. Observers viewed all stimuli on a computer monitor and responded by marking forms provided. The experiment was self-paced. The observers were told that half of the pairs of the set of photographs in the main phase were genetic siblings and half were unrelated children. Their task was to classify each pair as related or not related. The 30 pairs of photographs were presented in random order. 

Results and discussion

Signal detection analysis was used. We report the results for each condition as d′ and β values ( Table 2). We will first discuss the d′ values, which are also summarized in Figure 5. 

Table 2

 

Results for Experiment 2, by condition. The d′ estimate and the likelihood criterion β for the signal detection analysis are shown. Standard deviations were estimated by a bootstrap procedure (Efron & Tibshirani, 1993) based on 10,000 replications.

	d′	SD ( d′)	β	SD ( β)
FF	1.023	0.080	0.971	0.041
EM	0.818	0.078	1.043	0.033
MM	1.107	0.080	0.960	0.043

Figure 5

 

Experiment 2: results. The d′ values for the three experimental conditions are plotted as a bar graph. The confidence intervals (shown as 95% confidence intervals) were estimated using a bootstrap procedure (Efron & Tibshirani, 1993) with 10,000 replications. The d′ value for the FF condition is not significantly different (p > .05) from the d′ values for the other two conditions.

The estimated d′ values in all three conditions were significantly greater than 0 (FF: z = 12.787, p < .0001; EM: z = 1.487, p < .0001; MM: z = 13.838, p < .0001, all one tailed). Performance in the EM condition was significantly worse ( z = 2.587, p < .01, two tailed) than that in the MM condition. This outcome parallels results in the recognition literature (Fisher & Cox, 1975; Goldstein & Mackenberg, 1966). The group of observers who could see the entire face except for the eyes did worst in the kin recognition task than observers who could see the entire face except for the mouth. The differences between the FF condition and the two masked conditions were not significantly different from 0 (FF vs. MM: z = 1.835, p =.458, two tailed; FF vs. EM: z = 1.835, p =.067). 

Consider next the pattern of the outcomes of the hypothesis tests in Experiment 2 . For convenience, we will use the designator of each experimental condition to denote the d′ estimate of that condition. We failed to reject the hypotheses d′ _FF = d′ _EM and d′ _FF = d′ _MM, although we did reject d′ _EM = d′ _MM. Yet, there is no contradiction here as there would be if we thought we had shown d′ _FF = d′ _EM and d′ _FF = d′ _MM but d′ _EM < d′ _MM. As we noted above, failing to reject a point null hypothesis A = B does not imply that A really is equal to B (Cohen, 1994; Loftus, 1996). 

But we can say more. One reviewer advanced the following argument: If we accept that it is plausible that d′ _MM < d′ _FF (masking the mouth can only reduce performance), then the pattern of results d′ _EM < d′ _MM (the outcome of an hypothesis test) and d′ _MM < d′ _FF (our assumption) would entail the conclusion that d′ _EM < d′ _MM. We note that this conclusion is based not only on the assumption d′ _MM < d′ _FF but also on our statistical results and, while plausible, goes beyond the normal range of statistical argument. 

The estimates of effect sizes in Table 2 together with the estimated standard deviations of these effects are the best estimates we have concerning the quantity of kin recognition information available in different parts of the face. We will analyze them further together with the results summarized in Table 1 in the General discussion section. 

The β values in the three conditions are not significantly different from each other. They are also not significantly different from 1 (we omitted detailed reports of the z values). Because the prior odds that the children are related are 1:1, the participants are correctly using them; they do not show any bias in favor of classifying them as kin (as against not kin). We note that participants did not undergo the familiarization and training phases in Experiment 2 , and perhaps this difference in procedure accounts for the bias found in Experiment 1 and the absence of such bias in Experiment 2 . 

General discussion

We reported two experiments that aimed to determine what regions of the face contain the cues that signal kinship. In both experiments, participants were shown 30 pairs of photographs of children's faces in random order. Half of the pairs portrayed siblings and half did not. A total of 220 participants were asked to judge whether each pair of photographs portrayed siblings (yes–no task). 

Pairs of photographs were presented with both faces being fully visible (the unmasked condition) or with opaque masks covering corresponding parts of both faces. In Experiment 1, there were three conditions: one unmasked and two masked. In the first masked condition, masks covered the lower half of the faces of both children in each pair. In the second, masks covered the upper half of the faces. Observers participated in only one of the conditions. In Experiment 2 , there were also three conditions: one unmasked and two masked. In the first masked condition, the masks covered a small region containing the eyes. In the second, the masks covered a small region containing the mouth. We analyzed the data using standard signal detection methods (Green & Swets, 1966/1974). 

In Experiment 1, we found that performance did not deteriorate significantly when the bottom half of the face was masked. The estimated d′ value with the bottom half of the face masked was only 5.3% less than that when the entire face was shown. Performance, instead, deteriorated significantly (about 20%) when the upper half of the face was masked. This result is consistent not only with previous results in the face recognition literature (Fisher & Cox, 1975; Goldstein & Mackenberg, 1966) but also with what we know about the development of the face (Enlow & Hans, 1996; Kohn, 1991). We conclude that the observer extracts considerably more information about kinship from the upper half of the child's face than from the lower half. 

We also found that observers are well above chance in judging kinship when the upper half of the face was masked, although the d′ value in this case was significantly lower than when the lower half of the face was masked (a decrease, in terms of percentage, of more than 65% in d′ value). 

We also examined whether performance in the FF condition was appreciably better than would be predicted by combining performance in the two masked conditions (Green & Swets, 1966/1974) assuming optimal cue combination (Landy, Maloney, Johnston, & Young, 1995). This outcome would be expected if, for example, ratios of separations of signals that fell in both regions (e.g., eye separation divided by eye-to-mouth separation) were good cues to kinship and the observer gave considerable weight to them in assessing kinship. We found that, in this case, the “whole” was the “sum of the parts.” There was no indication that such full-face signals were used. 

In Experiment 2 , we found that masking the eye region led to only a 20% reduction in performance as measured by d′. This reduction was not statistically significant (reported above), and we would certainly not conclude that the eye region dominates kin recognition performance given that masking of the eye region leads to such a small (and nonsignificant) reduction. However, we approach this issue (the relative importance of the eye region to kin recognition) in a different way. 

In Experiment 1, we examined the additivity of information across the lower and upper halves of the faces. We next consider a second additive decomposition where we partition the face into the lower half region (LH), the eye region (E), and the upper half excluding the eye region (UH\E). The results of this decomposition will give us a better idea of the information about kinship contained in the eye region and in the upper part of the face outside of the eye region. We use the identity  

$({{\mathrm{d\prime }}^{}}_{\mathrm{F}\mathrm{F}}{)}^2=({{\mathrm{d\prime }}^{}}_{\mathrm{E}}{)}^2+({{\mathrm{d\prime }}^{}}_{\mathrm{E}\mathrm{M}}{)}^2$

(2)

based on partitioning the full face into the eye region and its complement. We use a second identity  

$({{\mathrm{d\prime }}^{}}_{\mathrm{F}\mathrm{F}}{)}^2=({{\mathrm{d\prime }}^{}}_{\mathrm{L}\mathrm{H}}{)}^2+({{\mathrm{d\prime }}^{}}_{\mathrm{E}}{)}^2+({{\mathrm{d\prime }}^{}}_{\mathrm{U}\mathrm{H}\mathrm{E}}{)}^2$

(3)

based on partitioning the full face into the lower half region, the eye region, and the part of the upper face excluding the eye region. We have an estimate of d′ _EM from Experiment 2 and an estimate of d′ _LH from Experiment 1 (it is just the estimate of d′ _UHM). We have estimates of d′ _FF from both experiments and we use the average of these estimates. We can then solve the equations above for estimates of d′ _E = 0.743 ± 0.165 and d′ _UH\E = 0.709 ± 0.102. The values following the symbol “±” are bootstrap estimates of the standard deviation (Efron & Tibshirani, 1993). The estimate d′_LH = 0.408 ± 0.075 for the lower half of the face (Table 1) completes the decomposition of the face into three regions. We emphasize that this computation depends on the assumption that information about kinship in nonoverlapping regions of the face is independent and correctly described by Equations 2 and 3. 

If we accept these estimates, it is evident that the eye region contains only slightly more information about kinship (as measured by d′ = 0.743) than the upper half of the face ( d′ = 0.709) outside of the eye region, and the difference is not significant ( z = 0.1753, p = .861, two tailed). Our results are in contrast with typical conclusions in the face recognition literature (Davies et al., 1977; Fisher & Cox, 1975; McKelvie, 1976; Roberts & Bruce, 1988). Yet, given differences in experimental design, stimuli, and methods between the experiments reported here and experiments in the face recognition literature just cited, we can draw no firm conclusions. We do, however, advance the conjecture that the eye region contains primarily information about the specific identity of the individual, not their kin group. This conjecture is intriguing (and unexpected) and certainly deserves further study. Experiments are needed to determine the relative importance of the eye region in kin recognition tasks and in identity recognition tasks using comparable stimuli and methods. 

The picture that emerges from this study is intriguing. We can summarize the information in a face as a list of signals (measurements made on the face). Some signals are of use in assessing age and gender, and these signals are correspondingly of little use in assessing kinship. Signals of identity recognition can include those that signal age, gender, and kinship, but except in the case of identical twins, we expect that there are also signals specific to the individual, which is the basis for identity recognition. Our results, combined with previous results in the literature, indicate that the visual system selects and combines whatever signals are available in unoccluded parts of the face according to task. Visual performance is robust in that large parts of a face can be masked and still the visual system can extract task-relevant information as in the UHM condition of Experiment 1. 

By comparing performance across task with selective masking of regions of the face, we can develop a better understanding of what these signals are and how they enter into different visual tasks involving faces. 

Acknowledgments

L.T.M. was supported by NIH Grant EY08266; M.F.D.M. was supported by funds from the Italian Ministero dell'Università e della Ricerca Scientifica e Tecnologica. We acknowledge Davide Bianconi, Francesca Constantini, and Carla Scagliarini for their assistance in data collection and taking of photographs and an anonymous reviewer for suggestions. 

Commercial relationships: none. 

Corresponding author: Maria F. Dal Martello. 

Email: [e-mail]. 

Address: Department of General Psychology, University of Padova, via Venezia 8, Padova, Italy. 

Footnotes

Footnotes

¹  The term “face” is used in the face morphology and dental literature to refer to the lower part of the face in ordinary usage, that is, from the eyes to the chin. We will use the term “lower face”instead.

Footnotes

²  The term “signal” here is a synonym for the term “feature” as used by, for example, Tversky (1977). Some authors drop the requirement that a “feature” must carry objectively useful information relevant to a task. That is, the visual system may mistakenly make use of a “feature” that unfortunately is not a “signal.”