Parent–offspring facial resemblance increases with age in rhesus macaques

Abstract

Kin recognition is a key ability which facilitates the acquisition of inclusive fitness benefits and enables optimal outbreeding. In primates, phenotype matching is considered particularly important for the recognition of patrilineal relatives, as information on paternity is unlikely to be available via social familiarity. Phenotypic cues to both paternal and maternal relatedness exist in the facial features of humans and other primates. However, theoretical models suggest that in systems with uncertainty parentage it may be adaptive for offspring to conceal such cues when young, in order to avoid potential costs of being discriminated against by unrelated adults. Using experienced human raters, we demonstrate in a computer-based task that detection of parent–offspring resemblances in the faces of rhesus macaques (Macaca mulatta) increases significantly with offspring age. Moreover, this effect is specific to information about kinship, as raters were extremely successful at discriminating individuals even among the youngest animals. To our knowledge, this is the first evidence in non-humans for the age-dependent expression of visual cues used in kin recognition.

1. Introduction

Genetic relatedness is a fundamental variable that regulates a wide range of social behaviour [1]. The ability to identify kin and discriminate between conspecifics based on cues that correlate with genetic similarity is taxonomically widespread, being documented from bacteria to humans [2,3]. Such kin recognition facilitates the acquisition of inclusive fitness benefits via nepotism, and allows individuals to optimize their mate choice decisions to avoid deleterious effects of inbreeding depression. In many species, prior association (‘familiarity’) with parents and siblings during early development suffices to identify close relatives [4]. However in some cases social cues are unreliable for inferring genetic relatedness, for example where sibships involve mixed paternity due to polyandrous mating, or mixed maternity due to intraspecific brood parasitism or communal nursing of young. Such species may use phenotype matching, whereby individuals are categorized as kin depending on how well their phenotypic traits match those of a learned or genetically determined template [4]. The gene–cue covariance that underlies phenotype matching also enables more finely graded discrimination of kin, even between unfamiliar individuals.

Visual phenotype matching based on faces has been demonstrated in several primates. For example, humans can detect kin resemblances between unfamiliar third parties [5]; they also exhibit greater trust and cooperation toward computer-generated images that resemble themselves [6]. Studies in natural populations show that these assessments may have fitness consequences: fathers invest more in children perceived to resemble them more closely [7]. Likewise, macaques and chimpanzees can detect parent–offspring pairs in images of unfamiliar adults of their own species [8]. Rhesus macaques even spontaneously discriminate their own unfamiliar paternal half-siblings in facial photographs, under natural conditions [9] (as may mandrills [10]).

However, advertising one's parentage may not be advantageous at all life stages. For example, in species with both paternal care and paternity uncertainty (e.g. humans [11]), males should be selected to discriminate offspring from non-offspring, to avoid misdirecting their investment. Moreover, offspring may be under pressure to minimize expression of traits that would enable males to determine paternity, due to the high fitness costs if males withdraw care. Similar logic may apply to primates such as macaques which reside in large mixed-sex groups, as infants and juveniles are vulnerable to aggression from unrelated adults and, in some species, infanticide by males [12]. The conclusions of theoretical models are mixed; while some predict that even very low levels of paternity uncertainty in a population will lead to individuals concealing their parentage [13,14], a model that simultaneously considered the evolutionary responses of both signallers and receivers instead suggested offspring should resemble their fathers, especially when paternity uncertainty is high [15]. This apparent contradiction can be resolved [14], as advertisement of identity may evolve or not, depending on the ratio between the costs of acceptance and rejection errors in specific evolutionary contexts.

If there is selection against familial resemblance, its force should decline with age as youngsters become less vulnerable to aggression or withdrawal of care. For example, rates of infanticide by step-parents decline more than sixfold between infancy and the age of five years [16]. At sexual maturity, revealing cues of kinship should start to confer benefits in terms of optimizing mate choice. Human data provide some support for this prediction. Infants tend to look anonymous up to one year old (unrelated observers can match the faces of babies to parents at only slightly better than chance levels), but detection rates increase with the age of children pictured [17]. Remarkably, across non-human taxa and cue modalities, only a handful of studies have explored developmental changes in heritable phenotypic resemblances between relatives, as opposed to cases where ‘kin signatures’ are acquired (e.g. via vocal learning, or interindividual transfer of odours). To our knowledge, the only cases involve age-related change in endogenously produced odours (rodents [18]; birds [19]). Here, we use a computer-based task with human raters (a technique successfully used to detect kinship cues in macaque faces [20]), to test whether facial similarities between parents and offspring from a free-ranging population of rhesus macaques change with age.

2. Methods

(a). Participants

We combine data from two experiments. Experiment A included 90 human raters (images weighted toward immature macaques) and experiment B included 35 raters (images of sexually mature macaques). Raters in experiment B comprise a subset of those used in a previous study [20], which did not examine effects of stimulus animal age (see electronic supplementary material). Sixteen raters participated in both experiments and thus encountered the full range of macaque ages. All participants had experience working with non-human primates (being able to identify primates of one or more species individually, based on facial characteristics), and were unfamiliar with our study population.

(b). Experimental procedure

We used a within-subjects design, where four types of kin discrimination (KD) trial (mother–daughter, mother–son, father–daughter, father–son) were presented on a computer monitor (figure 1). Participants used an 8-point scale to indicate which of two frontal facial images of macaques (the offspring, or an unrelated ‘decoy’ of the same sex and age) was more similar to a target face (parent), including the relative degree of perceived resemblance. Sixty trials were used in experiment A (offspring age range: 3 months–6 years) and 32 in experiment B (offspring ages: 4–12 years), with an equal number in each of the four kin conditions (details in electronic supplementary material). To verify that participants could reliably distinguish different macaques, individual discrimination (ID) trials were also presented. Here the target was a frontal image, the correct choice the same individual in ¾-view, and the decoy an unrelated (sex- and age-matched) individual in ¾-view. Nine trials were used in experiment A and 12 in B, covering both males and females. Each participant received trials in a unique order, and position of the correct match was randomized left–right across trials.

Figure 1.

Example of kin discrimination trial. A paternal line same-sex trial, with father displayed uppermost and two alternative matches below. The correct choice (juvenile son) is the lower-left image. (Online version in colour.)

(c). Macaque facial images

Standardized digital colour photographs of 291 free-ranging rhesus macaques were collected at Cayo Santiago, Puerto Rico. Image selection and preparation are described in the electronic supplementary material.

(d). Determination of kinship

Multi-generational pedigrees were used to generate triads consisting of a parent, its offspring (r ≈ 0.5) and an individual unrelated to both (with unrelated dyads defined as pairs sharing no ancestors in common, up to and including the grandparental generation; r < 0.063). For specifics on parentage assignment, see electronic supplementary material.

(e). Statistical analysis

To investigate whether KD performance depended upon offspring age (days), parental line (maternal versus paternal), or their interaction, we ran a linear mixed model (LMM) with Gaussian error structure and identity link function. We controlled for triad type (same- versus mixed-sex), parent–offspring age disparity (days), trial position in sequence, and the participant's experience working with non-human primates (classified as ≤6, >6 and ≤24, or >24 months), including these as fixed effects. LMMs for ID performance involved offspring age as the main fixed effect; triad sex (male versus female), trial position in sequence, and the participant's previous experience were included as control variables. Participant identity, image triad identity, and the identity of images used as target, left- and right-match, were always included as random intercepts. Following [21], we used a maximal random effects structure, incorporating all possible random slopes and correlations. More accurate p-values for individual fixed effects were obtained using the R function drop1 [21] with argument ‘test’ set to ‘Chisq’. Further details on model implementation are provided in the electronic supplementary material. Total sample size was 6520 and 1230 data points (trials) for KD and ID models, respectively.

Before the main analysis we first established overall significance of the full model [22], compared with a null model (that excludes the main predictors of interest but retains the full random effects structure), using a likelihood ratio test (LRT). In the full model, if an interaction term was nonsignificant it was excluded (together with its random slope), and fit of the reduced versus full model was compared using a LRT. For ID analyses, we additionally demonstrated that participants succeed at the task by testing the model's intercept against zero; a significantly positive intercept (when re-running the model with all factors centred and covariates z-transformed) indicates that the correct match is on average preferred over the decoy. As there was slight patterning in the residuals of the full models, their results were confirmed by conducting generalized linear mixed models (GLMM) with binomial error structure and logit link, a more conservative approach (full results are provided in electronic supplementary material). Analyses were implemented in R v. 3.3.0 [23] using the functions lmer and glmer of the R package lme4 v. 1.1-12 [24].

3. Results

(a). Kin discrimination improves as offspring age increases

Overall, offspring age and parental line had a significant impact upon participants' performance (full versus null model: X² = 7.86, d.f. = 3, p = 0.049). The interaction between age and line was not significant (estimate = −0.06 ± s.e. 0.157, X² = 0.31, d.f. = 1, p = 0.580) and therefore dropped. This revealed a significant positive effect of offspring age (table 1, figure 2), but not of parental line. Performance levels first became significantly positive (i.e. lower 95% CI > 0) at an offspring age of 643 days (figure 2). This pattern was confirmed within the 16 participants who experienced the full range of offspring ages: the age × line interaction was not significant (estimate = 0.01 ± s.e. 0.198, X² = 0.29, d.f. = 1, p = 0.593), and only offspring age was significant in the reduced model (estimate = 0.25 ± s.e. 0.099, X² = 5.94, d.f. = 1, p = 0.015; all other p-values between 0.151 and 0.824).

Table 1.

Results of linear mixed model examining kin discrimination performance; final model. Significant effects given in italics.

variable	estimate	s.e.	lower 95% CI	upper 95% CI	X2	d.f.	pa,b
intercept	0.381c	0.158	0.080	0.679	—	—	—
offspring aged	0.197	0.079	0.043	0.349	6.36	1	0.012
parental line (maternal = 0, paternal = 1)	0.239	0.174	−0.114	0.582	1.78	1	0.183
parent–offspring age disparityd	0.037	0.086	−0.134	0.217	0.40	1	0.528
sex combination (mixed = 0, same = 1)	−0.214	0.176	−0.561	0.145	1.38	1	0.239
experience (≤6 months = 0, >6 and ≤24 months = 1)	0.038	0.074	−0.099	0.189	1.32	2	0.518
experience (≤6 months = 0, >24 months = 1)	−0.021	0.070	−0.156	0.124
trial positiond	0.021	0.024	−0.025	0.066	0.73	1	0.392

^ap-values and test results from the drop1 function are calculated per predictor.

^bp-value for the overall intercept does not have a meaningful interpretation and is therefore not shown.

^cPositive values indicate a preference for the correct image in a trial; zero indicates chance performance and negative values a preference for the incorrect image.

^dCovariates were z-transformed; mean and s.e. of the original variables were 1606.1 ± 109.18 days for age, and 3088.0 ± 139.50 days for parent–offspring age disparity.

Figure 2.

Performance on kin discrimination trials. Mean resemblance scores between offspring versus unrelated individuals and the target (parent), as a function of offspring age. Zero represents chance performance (horizontal line). Positive values indicate the offspring was rated as more similar to the parent than was the non-kin individual, and negative values the reverse. Dashed and dotted lines indicate fitted values and 95% confidence intervals from the model, respectively (to aid interpretation, all factors were centred and covariates z-transformed before inclusion in the model).

(b). Individual discrimination is achieved even with the youngest macaques

The full model did not differ significantly from the null as there was no significant effect of the main predictor, individual age (estimate = −0.06 ± s.e. 0.165, X² = 0.01, d.f. = 1, p = 0.920; full details in electronic supplementary material, table S1). The absence of an age effect was not due to a general failure at the task, as participants were very successful in discriminating individual macaques. As this had previously been demonstrated for experiment B [20], here we restrict analysis to experiment A, comprising the youngest triads. The overall intercept was significantly greater than zero (estimate = 2.35 ± s.e. 0.305; lower and upper 95% CI = 1.74 and 2.95; p ≪ 0.001). Finally, the absence of an age effect was also demonstrated in the subset of 16 participants who took part in both experiments (estimate = 0.02 ± s.e. 0.151, X² = 0.01, d.f. = 1, p = 0.903; all other p-values between 0.240 and 0.634).

4. Discussion

Human raters were more successful at detecting parent–offspring resemblances in macaque faces, the older the offspring pictured. This result was robust, regardless of the analysis method (Gaussian or binomial models) and subject sample used (the full sample, or restricted to the individuals which participated in both experiments). Importantly, the age effect was specific to kin trials, as participants were able to discriminate individuals even among very young infants, despite the additional cognitive demands of comparing partially rotated versus frontal images in the individual discrimination task. This implies that the developmental trajectories of visual cues of identity versus kinship differ in macaques (in contrast to odour cues in rodents, where their ontogeny has been assumed to be similar, e.g. [25]). It is also consistent with reports that the two types of information are encoded separately in human faces, with individuality and kinship being processed using configural versus featural cues, respectively [26].

One explanation for increasing similarity between relatives with age is phenotypic convergence over time created by shared environments, a hypothesis originally proposed to explain convergence in the physical appearance of married couples (non-relatives) [27], but which could apply equally to relatives. This seems unlikely in macaques. Fathers and offspring typically live in different groups after early life due to male secondary dispersal (mean co-residence is 2.4 years in the offspring's first four years [28]), and there is little evidence for differing micro-environments within a social group that could generate such a phenomenon in mother–daughter pairs (who co-reside lifelong). It has also been suggested that it may be physically impossible to resemble one's parents when young (hence ‘all babies simply look like babies’, cf. [13]). Yet there is no obvious reason why facial shape, markings or colour would be developmentally difficult to produce, especially given our finding that facial cues are variable enough to support individual discrimination from a few months onwards. Finally, while facial phenotype will be the result of a complex interplay between genetic and environmental factors, it is likely that specific facial features are under strong genetic control. As individuals mature these traits may become more fully expressed, generating increasing resemblance to adult relatives. Nevertheless, if there are sufficient advantages to masking or revealing familial signatures at certain ages, selection could in principle modify the timing of such expression.

Our results suggest that, functionally, masking is achieved in young macaques. Raters' mean performance was essentially random on the youngest kin discrimination triads, only becoming successful from an offspring age of 1.75 years onward. This is consistent with the notion that paternity cues should be absent in infants, but become apparent once infanticide risk is lower and the age of puberty is approached. While in rhesus macaques both sexes are capable of reproducing by three years of age, full body size is not attained until roughly 6 years in females and 8 in males, and males do not sire their first offspring until 7.5 years old on average [29]. Therefore it may not be essential for individuals to have achieved full familial resemblance by puberty. Interestingly, we found no difference in performance on paternal versus maternal dyads, nor in their age trajectories. Selective masking of one parental line can in principle be achieved, for example via genomic imprinting. However, while concealing paternal identity should be paramount, resembling one's mother when young might lead to discrimination from males if she looks like rival males in the population [13], meaning cues to maternity might also be suppressed.

Our experimental approach was based on the known discriminative abilities of humans with respect to macaque faces [20,30], and the similarities between human and macaque face-processing mechanisms (e.g. [31,32]). Nevertheless, species typically perform better with conspecific than heterospecific faces [31,32], so it is possible that rhesus macaques may detect subtle kinship cues in the faces of young animals that even experienced humans cannot. Theoretically, complete anonymity of infants is not required—merely enough uncertainty to prevent adults from unambiguously identifying an infant's parentage [13]. While rhesus macaques might be able to identify kinship from relatively younger faces, this does not rule out further increases in kin resemblance with age, as found here. A key task for the future is to devise a way of presenting an analogous task to free-ranging macaques and, most importantly, to discover whether the functional consequences of being identified as kin/non-kin differ across life-stages.

Ethics

Collection of images and genetic samples was approved by the Caribbean Primate Research Center (CPRC) and the Institutional Animal Care and Use Committee of the University of Puerto Rico (protocol no. 4060105). Participation in the computer task was voluntary and conducted in accordance with the Declaration of Helsinki (v. 2008).

Data accessibility

The dataset is available at the Dryad Digital Repository: http://dx.doi.org/10.5061/dryad.84c66 [33].

Authors' contributions

A.J.N.K., Y.B. and A.W. designed the study; experimental data were collected by A.J.N.K. and Y.B., and analysed by A.J.N.K. and L.K.; genetic data and images were provided by A.W. and D.P.; A.J.N.K. and A.W. wrote the manuscript.

Competing interests

We have no competing interests.

Funding

Funding was provided by the Deutsche Forschungsgemeinschaft (grants WI 1808/3-1, 5-1 to A.W., PF 659/3-1 to D.P.). Cayo Santiago is supported by the National Center for Research Resources (NCRR, grant no. 8P40OD012217) and the Office of Research Infrastructure Programs (ORIP) of the National Institutes of Health. Content of the publication is solely the responsibility of the authors and does not necessarily represent the official views of NCRR or ORIP.