Convergence and divergence in gesture repertoires as an adaptive mechanism for social bonding in primates

A key challenge for primates living in large, stable social groups is managing social relationships. Chimpanzee gestures may act as a time-efficient social bonding mechanism, and the presence (homogeneity) and absence (heterogeneity) of overlap in repertoires in particular may play an important role in social bonding. However, how homogeneity and heterogeneity in the gestural repertoire of primates relate to social interaction is poorly understood. We used social network analysis and generalized linear mixed modelling to examine this question in wild chimpanzees. The repertoire size of both homogeneous and heterogeneous visual, tactile and auditory gestures was associated with the duration of time spent in social bonding behaviour, centrality in the social bonding network and demography. The audience size of partners who displayed similar or different characteristics to the signaller (e.g. same or opposite age or sex category) also influenced the use of homogeneous and heterogeneous gestures. Homogeneous and heterogeneous gestures were differentially associated with the presence of emotional reactions in response to the gesture and the presence of a change in the recipient's behaviour. Homogeneity and heterogeneity of gestural communication play a key role in maintaining a differentiated set of strong and weak social relationships in complex, multilevel societies.

1. Introduction

One of the most intriguing questions in the science of human origins involves the definition of language, its fundamental function and the selective pressures that were responsible for its evolution [1–6]. Viewed as a system of cognition and communication, one key selection pressure for the evolution of language may have been to facilitate social bonding and group cohesion in increasingly large groups of hominins [7,8]. The extent to which language can act as a social bonding mechanism may be affected by the degree of overlap in the repertoire between communication partners (homogeneity) [9]. For instance, a series of studies has shown that language and accent are one of the fundamental categories by which people categorize others into groups [10–14]. Even young, five- to six-month-old infants show a preference for people speaking in their native language, compared to foreign speakers [10]. Further, because language and accent are inseparable from the person and difficult to falsify, they may have been an important social marker through the course of human evolution, facilitating the development of ‘tag-based’ cooperation across increasingly large and dispersed social groups [9,11,14,15]. Language does not fossilize, so examining how repertoire homogeneity in non-human primates relates to social bonding can provide insights into the importance of overlap in communication repertoire in facilitating both dyadic interactions and large-scale sociality through the course of hominin evolution [15].

Until recently, the question of language evolution and sociality has received an almost exclusive research focus on the vocal modality of non-human primates, because the principal medium of human language is vocal [16]. However, the vocal and gestural components of language operate in a complementary fashion and human communication depends strongly also on visual cues. Gestural communication is defined as voluntary movements of the arms, head, body postures and locomotory gaits [17–26]. It has been theorized that language in humans evolved primarily in the gestural modality, because primates use gestures intentionally and have a greater control over their limbs than their vocal output. Moreover, the acquisition of new vocal signals in primate species other than humans is debated, whereas acquisition of gestures has been claimed in many studies [7,18,27–36]. However, the arguments in favour of the gestural origins of language are mainly based on findings in captive apes and cannot be extrapolated to wild populations because there are different selective pressures involved in captive compared to wild apes. For instance, the communicative behaviour of captive apes may be biased by frequent contact with humans in early ontogeny [37,38].

Wild East African chimpanzees (Pan troglodytes schweinfurthii) provide a valuable opportunity to assess the link between the social bonding and the homogeneity of gestural communication of primates. Chimpanzees form socially and geographically circumscribed communities, within which they associate in temporary subgroups (parties) that vary in size, composition and duration [39]. The community size can be in the range of 20–150 individuals, and the community as a whole is rarely seen together in one place [39]. Chimpanzees are frugivores and communities defend a communal home range, which is typically in the range of 5–35 km². Individuals in the wider community are thus often temporarily and spatially separated, but maintain long-term relationships through repeated and reciprocated interactions in many contexts [40–42].

One important mechanism to maintain social relationships in chimpanzees is grooming behaviour—grooming releases endorphins, enabling strong social bonds to be developed [43]. As well as grooming interactions between kin, grooming creates familiarity between unrelated group members originating in the knowledge of past relationships with the social partner that forms the basis for trust and social bond formation. In smaller networks, primates can form strong bonds with all network members such as kin and unrelated individuals and maintain relationships with them through grooming behaviour. However, the increasing time and cognitive demands on managing social relationships in networks containing many conspecifics imply that some network members become unfamiliar and ties become progressively weaker [44]. Thus in large social groups, there are dyads with less frequent interaction and the network contains an increasing number of individuals who cannot rely on grooming to maintain social relationships [45]. The ability to maintain network cohesion in large social groups depends on behaviours that enable primates to develop a social bond with individuals in the absence of grooming behaviour.

The social bond between group members could develop on the basis of phenotypic similarity such as shared facial characteristics [46]. The tendency to interact with others who have similar phenotypic characteristics is driven by fitness benefits of cooperating with kin [47,48]. In this case, phenotypic similarity enables the signaller to develop trust and bond with the dyad partner. However, when there are individuals in the group that share limited phenotypic similarity, the ability to establish social bonds with unrelated individuals may limit the capacity to maintain large social networks. When a lack of phenotypic similarity limits the ability to bond with unfamiliar or unrelated group members, individuals can use external cues such as communication similarity to create a tolerant context in the absence of prior relationships or genetic relatedness.

Similarity in communication can be established by increasing overlap in gestural expressions of the signaller's affect that release neurohormones associated with social bonding in the recipients [43]. These expressions are communicative to the recipients about the affect evoked by the social relationship and therefore influence the strength of social bonding between dyad partners. Moreover, similarity can occur in gestures used in a goal-directed way, whereby the signaller has a goal and uses a gesture that refers to the role of the recipient in attaining the signaller's goal, by indicating to the recipient through the gesture what they have to do. The gestures are responded to by goal-directed activity that matches the goal of the signaller, enabling more efficient social bonding between dyad partners [49,50].

Similarity in communication can arise by pruning down a set of innate gestures into a shared subset [26,51–53]. Such pruning occurs in repeated interactions with dyad partners during which individuals identify which gestures are effective. For instance, iterated learning experiments with humans have demonstrated that convergence of communication arises out of repeated interactions[54–60].

However, socially driven acquisition of gestures could be beneficial to both signallers and the recipients. When individuals who preferentially interact also preferentially learn from each other [61,62], gesture acquisition results in bonded dyads and groups with an increasingly large repertoire of homogeneous communication [63]. This homogenization may increase as the network of relationships individual primates have to manage increases. The patterns of homogenization may be differentiated by modalities of gestures such as visual (received through looking), tactile (received through touch) and auditory (received through hearing), as these modalities are differentially suited to maintaining different types of social bonds [64]. Visual or tactile gestures are most effective in one-to-one interactions, coordinating behaviour and acting as a time-efficient social bonding mechanism [45]. However, one-to-one gestures require a high degree of close proximity between the signaller and receiver to be detected in a dense forest habitat. Thus, there is a limit to which primates can keep increasing similarity of their gestures in one-to-one gestural interactions as the network size increases.

The need to manage the differentiated social network of strong and weak ties may lead to homogenization of communication that can be used to manage relationships between unfamiliar and unrelated individuals [9,65,66]. This type of communication may enable social bonding by prompting release of social neurohormones, which helps to create social bonds in the absence of prior grooming behaviour. In the absence of frequent grooming, this type of communication can increase the degree of similarity perceived between unfamiliar or unrelated interactants through some external cue that approximates phenotypic similarity that can be acquired without a need to be involved in direct one-to-one relationships with the signaller. A particular case of gesture acquisition in which the external evaluation of communication may be crucial for both learning and social bonding is object use. For instance, chimpanzees use objects to make loud sounds that can be perceived by both the immediate audience and the out-of-sight audience, enabling individuals to learn social signals and bond ‘at a distance’ [45]. Auditory gestures such as drumming have a higher amplitude, meaning they can evoke stronger emotions that may be better suited as a larger scale bonding mechanism [45,64,67,68]. Individuals with a greater degree of homogeneity in their communicative repertoires of visual, tactile and auditory gestures may therefore coordinate differentiated relationships better, when compared with individuals with lower levels of homogeneity [45,60,64,69,70].

Patterns of overlap in gestural communication can be influenced by the demography of the surrounding audience. In addition to the role of overlap in gestures in chimpanzee sociality, one aspect of primate social relationships that has not been considered in relation to the efficacy of social bonding is the absence of overlap in the repertoire of gestural communication (heterogeneity). For instance, avian studies postulated that heterogeneous communication can draw attention to the signaller when the immediate audiences are large, and this role is not necessarily restricted to avian species [71]. When phenotypic similarity with the immediate audience is high, signalling dissimilarity through heterogeneous gestures can increase the salience of the signaller to the recipient. Standing out in the crowd of similar others can enable signallers to gain a competitive advantage over individuals similarly suited to the social bonding with the recipient. By definition, heterogeneous gestures signal dissimilarity from the members of the social group or membership of a different social group independently of the phenotype characteristics, but the role of heterogeneous signalling in regulating social dynamics has so far been neglected. Very few studies to date have explored the potentially important role of heterogeneous communication in the coordination of social behaviour in primates.

Hence, it could be predicted that homogeneity and heterogeneity in the gestural repertoire would be useful in effectively managing social relationships with conspecifics, as reflected in the relationship between features of the gestural repertoire and social interaction. The relationship between homogeneity of communication and social dynamics has previously been described in the vocal domain of the chimpanzees. Studies have examined acoustic similarities in the panthoot calls within the group (panthoot recipients joining in the panthoot matching the acoustic structure of the initiator) [72] and group-level acoustic differences in panthoots [73]. In the gestural domain, studies have reported intra- and intergroup differences in the grooming hand-clasp gesture [74–78] and homogeneity of attention-getting auditory gestures across mother–offspring chimpanzee dyads [79,80]. However, research has not systematically examined how overlap in the whole gesture repertoire relates to sociality, either in captive or wild species [81].

This study examines how the homogeneity of gestural repertoire (defined as the degree of overlap in the presence or absence of gesture types in the repertoire between pairs of individuals, i.e. the number of shared gesture types) and heterogeneity of gestural repertoire (number of unshared gesture types) are related to social bonds of wild chimpanzees, measured by the duration of social bonding behaviours (joint feeding, joint resting, joint travelling, grooming, visual attention and in proximity) per hour pairs of chimpanzees spend in the same party. Moreover, we examine how the immediate audience (number of the individuals of similar age or sex present within 10 m) influences the use of these gestures. In this study, we predict that homogeneity and heterogeneity in the gestural repertoire of chimpanzees are not distributed randomly across dyads, but can be explained by biological factors (maternal kinship, age similarity, sex similarity), social relationships (duration of time spent in social bonding behaviour—when within 2 m, per hour spent in the same party), audience size (e.g. number of opposite-sex partners) social network size (centrality) and characteristics of gesture modality [61,62]. We examined the influence of these factors on dyadic and group-level homogeneity and heterogeneity in the gestural repertoire in the Sonso group of wild East African chimpanzees (Pan troglodytes schweinfurthii) in Budongo Forest, Uganda, Africa.

2. Material and methods

2.1. Data collection protocol

The data collection and methods for this study were approved by the Budongo Conservation Field Station research committee, the Uganda National Council for Science and Technology (NS 124) and the Uganda Wildlife Authority (UWA/TBDP/RES/50). In this study, observations of 12 habituated East African chimpanzees (Pan troglodytes schweinfurthii) (six adult males and six adult females) of the Sonso community, the Budongo Conservation Field Station, Budongo Forest Reserve in Uganda (www.budongo.org) were conducted in September 2006, between April and July 2007 and March and June 2008. The focal subjects were chosen on the basis of lack of any limb injuries and to represent age and rank classes as equally as possible. All of the females selected as focal subjects were parous. Full details of the study site and subjects have been described previously [82], so only brief details are given here. The behaviour of the chimpanzees was recorded during a standardized observation period using focal animal follows. The subjects were chosen systematically and, to avoid dependency in the dataset, consecutive samples of the same focal subject were taken at least 20 min apart. The data for this study came from the following sources. First, 18-min focal follows were performed. These consisted of nine scans at 2 min intervals of the identity of individuals present within 10 m of the focal subject and more than 10 m away from the focal individual who were in the same party. The party was defined as the group of individuals within a spread of about 35 m. Second, the identity of the adult nearest neighbour of the focal individual, the distance between the focal individual and the nearest neighbour (in metres), and the presence or absence of visual attention between the focal individual and the nearest neighbour were recorded. The visual attention was scored on the basis of bodily orientation rather than precise gaze direction between dyad partners. The activity state of both the focal subject and the nearest neighbour were also recorded (e.g. grooming, feeding, travel). Third, gestures were continuously recorded using a digital video camera recorder and this was accompanied by a verbal description of context, i.e. the identity of the signaller and the recipient, their behaviour prior to and after production of the gesture, and goal directedness. The presence of any calls accompanying the gestures was also noted. The sampling of association patterns was conducted by an experienced field assistant who was unaware of the aims of the study. The field assistants undergo an inter-observer reliability test annually, an interval which is sufficient to maintain the consistency of scoring of the group composition and proximity across field assistants, with results consistently above 0.85 for Spearman's rank correlation coefficient, r_s. The video recording of the gestures was carried out by A.R. and thus the social data and gestural data were collected independently of each other. Full details of the methods of data collection have been described previously [17,83,84].

2.2. Video analyses of gestural communication

The video footage was viewed on a television and coded. The initial catalogue of non-verbal behaviours was scored as an act of gestural communication if it was an expressive movement of the limbs or head and body posture that was: (i) mechanically ineffective (a gesture always elicited a change in the recipient's behaviour by non-mechanical means); (ii) communicative (i.e. at the level of the gesture type, communication was consistently associated with a change in the behaviour of the recipient after the signal), thus gestures occurred in social circumstances and only social circumstances when gestures were used were included in the dataset; for instance, if a chimpanzee gestured for non-social means (e.g. turn the back to change position rather than turn the back to initiate grooming), such circumstances of gesturing would not be considered [17]; and (iii) intentional. We used criteria for intentionality scoring that were previously developed to define intentionality in human infants and have been widely used in primates [85,86]. These include: (i) the presence of an audience; (ii) response waiting (the signaller directs a gesture at a recipient and observes the recipient's response during and after the gesture); (iii) the production of a gesture is sensitive to the recipient's visual attention state; and (iv) the signaller persists in gesture production when the recipient fails to respond. These intentionality criteria were evaluated for each gesture type separately, using pooled data across all subjects. Gestures above the threshold of 60% of cases were classified as intentional.

Full details of how gestures were categorized, including video clips of each gesture type, have been provided previously [17,84]. Briefly, visual, manual gesture types were established on the basis of 29 distinct morphological features, such as trajectory and orientation, using hierarchical cluster and discriminant function analysis [84]. Other gesture types were established qualitatively on the basis of objective judgement of the similarity of morphology of gestures (i.e. presence/absence and type of head, trunk, arm movement; posture; social orientation) [17]. This latter type of procedure has been widely used to identify distinct gesture types in chimpanzees [18,25,26,87] and other primates [51,52,88–90].

In line with previous definitions using this database of gestures [17], a ‘gesture sequence’ was defined as one or more gestures made consecutively by one individual towards the same recipient, within the same context, within a maximum of 30 s interval. There were a total of 545 gesture sequences included in the analysis. These gesture sequences included both single gestures and series of gestures, and in all sequences one of the 12 focal chimpanzees was the signaller. For each gesture sequence, we recorded the identity of the signaller (the individual performing a gesture), the identity of the recipient (the individual at whom the gesture was most clearly directed, as determined from the orientation of head and body of the signaller during or immediately after performing a gesture, i.e. the signaller had the recipient within its field of view), the recipient's behaviour after production of the gesture (response), the signaller's behaviour prior to and after production of the gesture, and the eliciting stimuli if present (e.g. presence of an intra-party dispute).

Gestures were classified according to the sensory modality (visual, tactile, short-range auditory, long-range auditory) [64]. The coding for the context and modality of a gesture was validated by a second coder, who scored a random sample of 10.42% of the gesture sequences examined in this paper. Cohen's κ coefficient showed that reliability was excellent for the modality of gesturing (K = 0.946) [91]. In addition, another sample of 50 sequences of gestures was coded by a second coder for intentionality (response waiting and persistence) and the Cohen's κ coefficient showed good reliability (K = 0.74). Reliability of coding into gesture types has been described in a previous publication [84], with the Cohen's Kappa coefficient showing good reliability (K = 0.76).

2.3. Behavioural measures

To reduce pseudoreplication, we aimed to sample each focal subject when the party had a unique composition, i.e. there was a change in composition of either focal males or females from the time the last subsample of the preceding focal follow was taken. The proximity (examined here for independence of 10 m associations only) and grooming scans were taken 2 min apart during the 18 min sample duration. The samples of each consecutive focal subject were taken at least 20 min apart. To ensure that this sampling procedure did not bias our results, we tested for similarity in association patterns between scans taken at 2 (scan 1), 4 (scan 2) and 18 min (scan 9) of the focal sample, including both sexes. There was no significant difference in the number of times focal and non-focal subjects were in close proximity at scan 1 (median = 2, IQ range = 0–5) and scan 2 (median = 2, IQ range = 1–5, Wilcoxon's signed-ranks test, T = 411.50, N = 132, p = 0.435). However, there was a significant difference in the number of times focal and non-focal subjects were in close proximity at scan 1 and scan 9 (median = 2, IQ range = 1–4; Wilcoxon's signed-ranks test, T = 2656.50, N = 132, p = 0.011). Similarly, there was no significant difference in the number of times focal and non-focal subjects were in the same party at scan 1 (median = 5, IQ range: 3–10) and scan 2 (median = 5, IQ range: 3–10; Wilcoxon's signed-ranks test, T = 218.50, N = 132, p = 0.571). However, there was a significant difference in the number of times focal and non-focal subjects were in the same party at scan 1 and scan 9 (median = 5, IQ range: 2–10; Wilcoxon's signed-ranks test, T = 1460, N = 132, p = 0.010). Thus, the adjacent scans were similar for 10 m associations and were treated as continuous. As 10 m associations and 2 m associations were correlated, we made an assumption of independence for both of these measures. Moreover, party-level associations were also treated as continuous. However, first and final sample scans differed for 10 m associations and party-level associations and therefore we treated these scans as independent, as well as the samples preceding and succeeding the focal follow. Behavioural measures were then derived, calculating the duration of time each pair of chimpanzees spent engaged in affiliative behaviours, per hour that pair of chimpanzees spent in the same party. These affiliative behaviours were: joint feeding, joint resting, joint travelling, giving grooming, receiving grooming, mutual grooming, visual attention towards the focal, visual attention away from the focal and proximity. An example of how these measures were computed can be found in the electronic supplementary material, S1 and a more detailed description has been provided previously [45].

2.4. Attribute measures

To control for the influence of demography, factors such as age, kinship, sex and reproductive state need to be taken into account when examining chimpanzees' propensity to associate with each other. Genetic data obtained in previous studies provided the basis for ascertaining kin relationships in the Sonso community and we scored chimpanzee dyads according to the presence or absence of kinship [82]. In the wild, chimpanzees reach physical and social maturity in the age range of 15–16 years [39]. The Sonso community is a long-running site and therefore the age of most adult subjects in the community is known. We classified dyads of chimpanzees as belonging to the same (5 years or less of age difference) or a different (above 5 years of age difference) age class [92]. Moreover, chimpanzee dyads were scored according to oestrous similarity. The reproductive status of the female was scored on the basis of the female sexual swelling, which is an enlarged area of the perineal skin varying in size over the course of the menstrual cycle. The reproductive status of the female was recorded as oestrous if, during the observation period, the female exhibited maximum tumescence (sexual swelling) and was observed mating with the males. The size of the sexual swelling was rated on a scale of 1–4, with the maximum swelling size scored as 4.

All focal males were observed to mate with the females and, therefore, assumed to be reproductively active. Dyads were classified as reproductively active (males and oestrous females), or non-reproductively active (all other combinations). Sex similarity was scored based on observable morphological characteristics referring to sex. Full details of the categorization of attribute data can be found in the electronic supplementary material, table S1.

2.5. Gestural repertoire homogeneity

For each gesture sequence, homogeneity was determined for each gesture type used. If the gesture type (e.g. arm beckon) was present in the repertoires of both the signaller and the receiver, this was classified as homogeneous. By contrast, if the arm beckon appeared in the repertoire of the signaller but not the recipient, it would be classified as heterogeneous. As gesture sequences can sometimes consist of a series of individual gestures, each gesture type used was classified separately as homogeneous or heterogeneous. Thus for a gesture sequence consisting of a series of five different gesture types, three gesture types could be classified as homogeneous and two as heterogeneous. For each gesture sequence, the ‘homogeneous repertoire size’ refers to the number of different homogeneous gesture types used by the signaller, and the ‘heterogeneous repertoire size’ refers to the number of different heterogeneous gesture types used. If only homogeneous repertoire size was taken into account, a repertoire size of three homogeneous gestures could represent a gesture sequence of five gestures, with three homogeneous gesture types and two heterogeneous gesture types, or a sequence of three gestures, all of which were homogeneous. Thus, we considered both homogeneous and heterogeneous repertoire size in our analyses, to identify sequences where heterogeneous gesture types were used.

For the network analysis, the overall number of gesture types shared (homogeneous) and not shared (heterogeneous) between dyad partners was included in the analyses rather than gesture rates. Because we provided the overall number of shared and non-shared gestures in this instance, for heterogeneous gestures the total number of non-shared gestures was given. For instance, if BB had 5 gesture types that were absent in HW repertoire and HW had two gesture types that were absent in BB repertoire, the heterogeneous repertoire size for both the BB HW dyad and the HW BB dyad was 7.

To determine the degree of homogeneity in the gestural repertoires between focal chimpanzees, using social network analysis, we used Cohen's κ coefficient [93]. This statistic measures the degree of agreement in the presence or absence of gestures in the repertoires of dyad partners. If the repertoire of gestures of a focal individual is in complete agreement with the repertoire of gestures of the dyad partner, then the Cohen's κ coefficient equals 1 (exactly the same repertoire) for that dyad. However, if the repertoire of gestures is in complete disagreement with the repertoire of gestures of the dyad partner, then the Cohen's κ coefficient equals −1 (completely different repertoire). In this study, the repertoire of gestures of each focal subject that were performed towards other adult individuals was compared within dyads (electronic supplementary material, table S2) [17]. In addition to all gestures combined, the Cohen's κ coefficient for each dyad was computed separately for visual, tactile, auditory short-range and auditory long-range gestures. In these analyses, 10 focal subjects (90 dyads) contributed data—subjects KU and ZM were excluded as they did not gesture in all modalities towards other adults. In previous research comparing repertoires within and between primate groups, Cohen's κ has been the most commonly used statistical technique, based on the presence or absence of gesture types from the repertoire [17,18,52]. The data pertaining to this part of this study has been placed in the electronic supplementary material, S2.

2.6. Generalized linear mixed modelling

The panthoot call is a type of chimpanzee vocalization which is broadcast at a wider audience. When performing social network analysis, we took all of the individuals who were within 10 m of the signaller as recipients of the gestures accompanied by panthoot call. Generalized linear mixed modelling (GLMM) prohibits this action, however, because all of the observations have to be independent from each other. In this case, all sequences of gestures that contained solely visual gestures and panthoots (low-intensity panthoot) were counted as directed by the signaller at the most dominant individual in the party. By contrast, all sequences of gestures containing panthoots and auditory gestures were assumed to be directed at a nearest neighbour of the focal individual (high-intensity panthoot) [45]. The data used in this study consisted of 545 gesture sequences produced by the 12 focal chimpanzees. In line with previous work based on this dataset [94], we used GLMM to examine how homogeneity and heterogeneity in gestural communication was related to the bonding behaviours. GLMM is a modified form of regression analysis designed to deal with data that are hierarchically structured. The random effects in all models were the identity of the focal individual and we included random intercepts for these effects. Random slopes were not used in these models as the key focus was on how the predictor variables relating to homogeneity were associated with the different bonding behaviours, rather than how the effects of these differed between the 12 focal subjects. In these GLMMs, the data were hierarchically structured with two levels—level 1 was the focal individual and level 2 was the recipient of the gesture. The response variables in the GLMMs were continuous (repertoire size, duration of social behaviour) or binary: the presence or absence of a gesture, demography (e.g. sex difference).

The models were fitted using a binomial error structure with logit link. In all of the analyses, the demographic relationships (e.g. age similarity), bonding status (e.g. duration of joint travel) and the presence or absence of response were controlled for. However, when demography was a response variable, the analyses only considered the presence and absence of a gesture. When analysing the relationship between homogeneous and heterogeneous gesture presence (the response variable) and the response type to the gesture, only single, unimodal gestures were considered because a previous study showed that including combinations of gestures was likely to influence the type of response made to the gesture [95]. In all analyses where audience size was included, the audience size excluded the signaller and the recipient of the gesture. Moreover, only one category of audience was included in each analysis (e.g. size of audience of same-age and different-age partners was included separately from size of audience of same-sex and opposite-sex partners). The descriptive statistics regarding variables included in the GLMMs are provided in table 1. The Generalized Linear Mixed Models function in IBM SPSS Statistics 22 was used for all the GLMM models. The data used in all GLMM models can be found in the electronic supplementary material, S3–S8.

2.7. Social network analysis

To complement the GLMM analysis, we used social network analysis to examine how individual variation in homogeneity for the focal chimpanzees was related to indegree and outdegree for bonding behaviours. Behavioural and communication networks were created for each behaviour type separately. Each network matrix consisted of 12 rows and 12 columns, with each row and column denoting a different focal chimpanzee. The values in each cell of the matrix represented the value for that particular behaviour for a specific pair of chimpanzees (e.g. the duration of time BB and HW spent in close proximity, per hour they spent in the same party). In the main analyses on factors influencing the proximity of dyads, the proximity and communication networks used in this study were weighted, i.e. each cell consisted of a continuous value representing the value of behaviour, rather than a 1 or a 0 indicating the presence or absence of a tie. In repertoire homogeneity networks, the value of Cohen's κ coefficient between dyads was entered in the matrices. The repertoire homogeneity networks were undirected. For instance, the overlap in the gestural repertoire between BB and HW was the same as the overlap between HW and BB. The proximity networks were treated as directed. For instance, the duration of time spent in close proximity by BB to HW may be different from the duration of time spent in close proximity by HW to BB.

From these network matrices, centrality measures were calculated, using normalized degree centrality [96]. Normalized degree centrality is defined as the average value of each row or column of the network matrix, i.e. the average value of that behaviour for each focal chimpanzee. The homogeneity network was undirected and therefore only the n degree value for each focal subject was obtained. This stands for the mean value of homogeneity of the gestural repertoire of each focal subject with all possible ties which are present. The behavioural network was different for the focal–non-focal subject dyads (e.g. BB to HW proximity was different from HW to BB proximity) and therefore indegree and outdegree were calculated separately. Outdegree refers to proximity directed by the focal chimpanzee to conspecifics, while indegree refers to proximity directed by conspecifics towards the focal chimpanzee.

General standard inferential statistics cannot be used on network data because the observations that make up network data are not independent of each other. Thus, randomization (or permutation) tests are used, whereby the observed value is compared against a distribution of values generated by a large number of random permutations of the data. The proportion of random permutations in which a value as large (or as small) as the one observed is then calculated, and this provides the p-value of the test [97]. MRQAP regression (multiple regression quadratic assignment procedure) was used to determine the relationships between social bonding networks and homogeneity of gestures [97]. MRQAP regression is similar to standard regression because it enables the examination of the effect of a number of predictor variables (e.g. visual and tactile homogeneity networks, sex similarity of a dyad) on an outcome variable (e.g. proximity network). Among several different types of MRQAP regression that are available, we used Double Dekker Semi-Partialling MRQAP regression, which is more robust against the effects of network autocorrelation and skewness in the data [98]. The number of permutations used in this analysis was 2000. For the node-level regressions, we used a similar procedure, using 10 000 random permutations to assess the effect of a number of predictor variables (e.g. the normalized mean degree for homogeneity of gestures, sex of focal chimpanzee) on the outcome variable (e.g. proximity in degree). All the social network analyses were carried out using UCINET 6 for Windows [99]. The data used in all social network analyses can be found in the electronic supplementary material, S2.

3. Results

The description of all variables included in the models is provided in table 1. The demographic details of the study group are provided in table 2. In all sections, only significant positive or negative associations between variables are reported. Further findings that elaborate on results presented in this section can be found in the electronic supplementary material.

3.1. Social bonding behaviour and demography

We used GLMMs to examine the relationship between duration of time spent in social bonding behaviour and demographic characteristics of dyads across sequences of gestures. Table 3 presents a summary of this analysis. In terms of the influence of the demographic characteristics of the signaller and the recipient, chimpanzee dyads who were the same sex spent a longer duration of time mutually grooming (β = −1.854, s.e. = 0.732, p = 0.012), visually attending (β = −3.547, s.e. = 1.264, p = 0.005), visually non-attending (β = −1.220, s.e. = 0.433, p = 0.005) and in proximity (β = −4.825, s.e. = 1.320, p < 0.001). Moreover, the chimpanzee dyads who had were in the same age category spent a longer duration of time in joint travel (β = −1.914, s.e. = 0.963, p = 0.048), giving grooming (β = −2.469, s.e. = 0.763, p = 0.001), visually non-attending (β = −2.671, s.e. = 1.340, p = 0.047) and in proximity (β = −10.860, s.e. = 5.159, p = 0.036). Further, dyad partners with the same reproductive status showed a pattern of significantly longer duration of time spent in mutual grooming (β = −1.717, s.e. = 0.635, p = 0.007), visually attending (β = −3.020, s.e. = 1.039, p = 0.004) and in proximity (β = −2.684, s.e. = 1.250, p = 0.033). Finally, dyad partners related through maternal kinship spent a longer duration of time in the following social bonding behaviours: joint feeding (β = −1.831, s.e. = 0.198, p < 0.001), joint resting (β = −1.759, s.e. = 0.386, p < 0.001), giving grooming (β = −3.330, s.e. = 0.527, p < 0.001), receiving grooming (β = −1.755, s.e. = 0.563, p = 0.002), mutually grooming (β = −4.080, s.e. = 1.075, p < 0.001), visually attending (β = −10.646, s.e. = 2.025, p < 0.001), visually non-attending (β = −8.374, s.e. = 0.279, p < 0.001) and in proximity (β = −17.903, s.e. = 1.782, p < 0.001).