Should I stay or should I go? Initiation of joint travel in mother–infant dyads of two chimpanzee communities in the wild

It is well established that great apes communicate via intentionally produced, elaborate and flexible gestural means. Yet relatively little is known about the most fundamental steps into this communicative endeavour—communicative exchanges of mother–infant dyads and gestural acquisition; perhaps because the majority of studies concerned captive groups and single communities in the wild only. Here, we report the first systematic, quantitative comparison of communicative interactions of mother–infant dyads in two communities of wild chimpanzees by focusing on a single communicative function: initiation of carries for joint travel. Over 156 days of observation, we recorded 442 actions, 599 cases of intentional gesture production, 51 multi-modal combinations and 80 vocalisations in the Kanyawara community, Kibale National Park, Uganda, and the Taï South community, Taï National Park, Côte d’Ivoire. Our results showed that (1) mothers and infants differed concerning the signal frequency and modality employed to initiate joint travel, (2) concordance rates of mothers’ gestural production were relatively low within but also between communities, (3) infant communicative development is characterised by a shift from mainly vocal to gestural means, and (4) chimpanzee mothers adjusted their signals to the communicative level of their infants. Since neither genetic channelling nor ontogenetic ritualization explains our results satisfactorily, we propose a revised theory of gestural acquisition, social negotiation, in which gestures are the output of social shaping, shared understanding and mutual construction in real time by both interactants. The study investigated the communicative behaviour of  mother–infant dyads in two different chimpanzee communities: Kanyawara in Kibale National Park, Uganda (Pan troglodytes schweinfurthii), and Taï South in Taï National Park, Côte d’Ivoire (P. t. verus). Detailed descriptions of the study areas can be found in Wrangham and colleagues (1992) and Boesch and Boesch-Achermann (2000). During the two study periods, the size of the Kanyawara group varied between 53 and 56 individuals, respectively, 21 and 24 in Taï South. The Kanyawara and Taï chimpanzees are well habituated and have been studied regularly since 1987 (Wrangham et al. 1992) and 1979 (Boesch and Boesch-Achermann 2000), respectively, enabling dawn-till-dusk follows and the collection of high-quality recordings. In addition, we had access to long-term data concerning the chimpanzees’ demography, social relationships, relatedness and ranks. We observed communicative interactions of a total of 13 mother–infant dyads (seven from Kanyawara and six from Taï South), with offspring ranging from 9 to 69 months of age (see Table 1). At Taï one mother gave birth to another infant in the second field period; hence, we observed 12 chimpanzee mothers and 13 infants. Information on observed mother–infant dyads with respective observation time and raw data set The last line provides the total sample size for each column (P1/P2: first/second period of data collection) aP1 not included since infant was too young bMothers gave birth to sibling in P2, thus no P2 data available cDeceased on November 1, 2012 Observations were made on chimpanzees of the Kanyawara community in Kibale National Park and the Taï South group at Taï National Park during four periods between October 2012 and June 2014 (Kanyawara: March–May 2013, March–June 2014; Taï South: October–December 2012, October–December 2013). We used a focal behaviour sampling approach (Altmann 1974), while maintaining a record of the frequency with which a particular dyad had been observed. In situations where we could choose which of several dyads to film, we targeted those individuals previously sampled least often. Following Hobaiter and Byrne (2011), who had suggested that approximately 15 h of active gesture time or approximately 150 days of field observation time would enable to assess the whole gestural repertoire of a given chimpanzee community (N = 82), we observed all 13 mother–infant dyads for a total of 156 days. All social interactions of mothers and infants (i.e. mother–infant interactions as well as mother-conspecific and infant-conspecific interactions) that were judged to have any potential for communicative interactions were recorded using a digital high-definition camera (Canon HF M41) with an external unidirectional microphone (Sennheiser K6). This method resulted in a total of 169 h of video footage recorded during approximately 1198 h of focal observations (see Table 1 for further details). However, the present paper focuses only on the communicative context of carry initiation; thus, our analysis is based on a total of 410 high-quality recordings of mother–infant behaviour in this respective context (mean recordings per dyad: 33.2). In addition, every 15 min we conducted a focal scan by using a Personal Digital Assistant (HP iPAQ rx1959) with focal/time sampling utilised as sampling/recording rule (Altmann 1974). This method enabled us to collect data on a variety of additional parameters such as behavioural context and party composition (see Online Resource 1, Table S2), resulting in a total of 4505 behavioural scans. To establish the behavioural repertoires of mothers and infants used to initiate maternal carries and enable subsequent analyses, a total of 410 high-quality video files of mother–offspring carry initiations (i.e. carries with clear visibility of carry-initiating behaviours) were coded using the program Adobe Premiere Pro CS4 (version 4.2.1.). In addition, we included PDA recordings of five interactions, resulting in a total of 415 interactions. Behavioural definitions were based on established ethograms of the behaviour of two long-term studies of eastern chimpanzees (Goodall 1986; Nishida et al. 1999) and several gesture studies (Call and Tomasello 2007; Hobaiter and Byrne 2011; Roberts et al. 2014a). Based on parameters used in previous work on great ape gesturing (Pika et al. 2003, 2005; Pika and Mitani 2006), a coding scheme was developed. For our purposes, all analysed joint travel events included maternal carries (i.e. involving mother–infant body contact). While coding all agent-initiated carries, we differentiated between carry-initiating behaviours via (1) physical actions, (2) intentionally produced gestures, (3) multi-modal combinations (gesture plus vocalisation) and (4) vocalisations. A physical action was defined as any behaviour that resulted in joint travel through direct manipulation of another’s body or the movement of one’s own body into a carry position. Carry-initiating actions included, for instance, grabbing, forcibly pulling, lifting or approaching another individual (see Online Resource 1, Table S1). Gestures were defined as directed, mechanically ineffective movements of the body or body postures that elicited (‘requested’) a voluntary response by the recipient (Pika 2008). In addition, we only included those gestures in our analyses that were accompanied by one or more key characteristics of intentional communication (Bates 1976; Bruner 1981; Pika et al. 2003): Sensitivity to the attentional state of the recipient The signaller shows signs of being aware of the recipient’s state of attention, e.g. by using visual gestures only when the recipient is looking. Response waiting The signaller pauses at the end of the signal and waits for at least two second for a response while maintaining visual contact. Apparent satisfaction of signaller The signaller’s communication ceases when the apparent goal has been met by the recipient (Hobaiter and Byrne 2014). Goal persistence The signaller elaborates her signalling when thwarted, e.g. by repeating and exaggerating the signal or by using a different communicative means (Pika et al. 2005; Pika and Mitani 2006). Gestures were clustered into three signal categories: audible (signals generate a sound while being performed, e.g. slapground), tactile (signals include physical contact with the recipient, e.g. touching) and visual (signals generate a mainly graphic component, e.g. raisearm) signals (Pika et al. 2003). To identify carry initiations, the behaviour of both, the signaller and the recipient throughout the interaction, from first initiating action/gesture to start of carry, was taken into account to assess the success of communicative attempts (Smith 1965). Idiosyncratic gestures, which are exclusive for single individuals in the whole community, had been observed at least three times to be included in the analyses (Pika et al. 2003, 2005). Vocalisations, especially those accompanying gestures (‘multi-modal signals’), were analysed in terms of their broad categories (Crockford and Boesch 2005; Goodall 1986; Table 2). Finally, for each signal or action case, we coded the following parameters: interaction role of the signaller: two levels, mother, infant; infant age: range 9–69 months; necessity of carry: two levels (low; carry preceded by feeding, playing, resting, relaxed group travel; high: preceded by aggressive behaviours such as chasing and hitting, catching-up with already left party/group, displaying and patrolling); mother’s parity: number of offspring reared at least until juvenility (plus present infant), range 1–5, partycomposition: three levels (mother with dependent offspring only, adult females only, mixed group). A least 15 per cent of all mother–infant interactions were coded for accuracy by a second observer and tested using the Cohen’s kappa coefficient to ensure inter-observer reliability (Altmann 1974). A ‘very good’ level of agreement was found for gesture type (κ = 0.878), signal type (κ = 0.811), signal category (κ = 0.843) and necessity of carry (κ = 0.816). The level of agreement for carry initiator (mother/infant) was ‘good’ (κ = 0.799). Gesture and vocalisation types produced to initiate carries in chimpanzee mother–infant dyads identified in this and other studies on wild groups in Budongo (Hobaiter and Byrne 2011; Roberts et al. 2014a); Gombe (Goodall 1986) and Mahale (Nishida et al. 1999) Since Byrne and colleagues (Genty et al. 2009; Hobaiter and Byrne 2011) had argued that differences in gestural repertoires of captive apes were simply premature assumptions, with repertoires yet to reach asymptote, we plotted the cumulative numbers of observed gesture types over time for all individuals. If an asymptote was reached (i.e. no further gesture types were observed), we concluded that we had observed the individual’s full repertoire for the specific communicative function of maternal carries. We measured the relationship between an individual’s final repertoire size and the total time that individual had been observed using the Spearman R statistic. For our repertoire analyses, we included only individuals observed for over 60 h (N = 10; observation time range 60.25–150 h, mean ± SD = 109.3 ± 32.1 h), which have reached the critical asymptote, to make sure that the complete repertoire of these individuals was grasped within the observation time. We compared repertoire sizes of mother and infants using an independent-samples t test after the underlying assumptions were tested (Levene’s test for equality of variances). To enable a better understanding of gestural acquisition, the gestural repertoires of mothers of the two communities of Kanyawara and Taï South were compared. To assess concordance rates of gestural repertoires within dyads, within groups and between groups, we used the Dice coefficient (Dc), which ranges from 0 to 1 (Dice 1945). A value of 0 means that two individuals have no gesture types in common, while a value of 1 would mean that the two gestural repertoires are identical. Since chimpanzee infants had very limited gestural repertoires in the specific context of carry initiation, we restricted this particular analysis to maternal repertoires only. In addition, we included in the analysis only data of individuals, whose repertoires had reached asymptote. To investigate whether repertoire similarity was larger between mothers of the same community than between mothers of different communities, we used a matrix permutation test (Sokal and Rohlf 1995). To test to which extent the predictor variables such as infant age, interaction role, carry necessity and mother’s parity influenced signal type (action, visual gesture, tactile gesture; response variables), we used generalised linear mixed models (GLMM; Baayen 2008) with a binomial error structure and logit link function. We fitted one model for each of the three response variables. Into this, we included interaction role, infant age, carry necessity and mother’s parity as our key test predictors, respectively. Another model was specified for carry initiator as binomial response variable (0 = mother initiation, 1 = infant initiation), but only infant age and parity were specified as key test predictors in this model. Since the average age varied considerably between infants but also within them, we used the method of within-subject centring (van de Pol and Wright 2009). This method allows to test whether the effect of age takes place largely across subjects (cross-sectional) or within subjects (longitudinal). Practically, this means that we include two predictors representing age into the model: one representing the average age per infant (from here on called within-infants age) and the other being the difference between the date that the observation was made (from here on called between-infants age) and its average age. Because we assumed that over the course of ontogeny, infants would take a more active role we also included the two two-way interactions between role and the two variables representing infant age into the first three models. To control for confounding effects, we also included party composition, infant sex and study site as further fixed effects. As random effects (intercepts), we included the identity of the mother and the infant into the model. To keep type 1 error rates at the nominal level of 5 %, we also included the random slopes components of role, within-infants age and their interaction as well as carry necessity within infant identity (Barr et al. 2013; Schielzeth and Forstmeier 2009). We did not include any other random slopes components within mother ID because with a single exception each mother only had a single infant and hence random slopes of these fixed effects within mother ID would be highly redundant. For the other fixed effects, we did not include random slopes because they were most usually constant within mother and infant ID. We also did not include correlations between random slopes and random intercepts in order to keep model complexity at an acceptable level and because neglected random slopes do not compromise type 1 error rates (Barr et al. 2013). The models were implemented in R (R Core Team 2014) using the function glmer of the package ‘lme4’ (Bates et al. 2014). To test the overall significance of our key test predictors (Forstmeier and Schielzeth 2011; Mundry 2014), we compared the full models with the null models comprising only the two control predictors with fixed effects as well as all random effects using a likelihood ratio test (Dobson 2002). Prior to running the models, we z-transformed between-infants age, within-infants age and parity (Aiken and West 1991; Schielzeth 2010). To control for collinearity, we determined variance inflation factors (VIF; Field 2005; Quinn and Keough 2002) from a model including only the fixed main effects using the function vif of the R package ‘car’. This revealed collinearity to not be an issue (maximum VIF = 1.44). To estimate model stability, we excluded the levels of random effects one at a time, ran the models again and compared the estimates derived with those obtained from the models based on all data. This revealed all models to be at least ‘moderately’ stable, particularly for those estimates that were not close to zero (for details on model stabilities, see supplementary material in Online Resource 2). Confidence intervals were derived using the function sim of the R package arm (Gelman and Su 2014). Tests of the individual fixed effects were derived using likelihood ratio tests (R function drop1 with argument ‘test’ set to ‘Chisq’). All statistical analyses were performed using the R-version R.3.1.1 (R Core Team 2014), with a level of significance set to 0.05.