We collected an average of 26.0 observation hours from each of 84 subjects (N = 50 females) (range: 24 h 38’-27 h 25’, except for one male who died and was observed for 12 h 31’, resulting in a total of 2180 observation hours). We conducted 30-min continuous focal protocols 19 using handheld devices (Samsung Galaxy Note 2) with the software program Pendragon Forms (Pendragon Software Cooperation, Libertyville, IL, USA). During focal observations, we recorded all social interactions of focal animal and extracted all instances of the use of agonistic and affiliative signals. Agonistic signals included threat stare, open-mouth threat, head bob, ground slap, silent scream face and scream face; affiliative signals included teeth-chatter and lip-smack, as well as the responses by the interaction partner. We further noted additional agonistic interactions ad libitum to establish the dominance hierarchies. To this end, we used all dyadic and decided agonistic interactions (submissive reaction and no counter-aggression). We determined the dominance rank based on the normalized David’s score, implemented in the EloRating package in R 20; for further details see 17.
Statistics and Reproducibility
We conducted all analyses in the R environment (see Supplementary Table 2 for all version numbers). For the analysis of signal use, we used a General Linear Model (GLM) with negative binomial error distribution and logit link function, applying the function glm.nb of the R package MASS. Model 1 comprised the analysis of the use of agonistic signals; model 2 the analysis of affiliative signals. A Poisson distribution did not provide a good fit as both response variables appeared overdispersed given the model. Age and rank were z-transformed to a mean of zero and a standard deviation of one to ease interpretability of the model estimates. Rank was only included in the analysis of affiliative signals but not used in the analysis of agonistic signals, because the dominance rank was mainly based on the occurrence of agonistic signals. Hence, the inclusion of rank would have created a circular situation. We included focal observation time (log-transformed) as an offset term 21. We checked the stability of both models using the function dfbeta and for potential collinearity issues by determining Variance Inflation factors (VIF) using the function vif 22 of the R package car.
The analysis of the agonistic signals revealed that a quadratic relationship better predicted age-related variation in the number of signals. However, it should be kept in mind that the inclusion of age squared represents an a-posteriori hypothesis. Hence, caution is required when interpreting such adjusted models.
The sample analyzed for these models comprised a total of 84 signalers (50 female) which produced a total of 5485 agonistic signals (model 1) and 2700 affiliative signals (model 2). None of the two models was overdispersed (dispersion parameters, model 1: 0.942; model 2: 1.032), and collinearity was also not an issue (maximum VIF, model 1: 1.001; model 2: 2.692). Both models also were of good (model 1) or moderate stability (model 2) as assessed by means of DFBeta values.
Probability of showing any response
With model 3, we estimated the extent to which the probability of showing any response was influenced by receiver age. We fitted a Generalized Linear Mixed Model (GLMM) 23 with binomial error structure and logit link function 21. We included receiver age and its interaction with the signal category (agonistic or affiliative) as our key test predictors with fixed effects. We also included the age and the sex of the signaler and also the main effect of signal category as control fixed effects. We included random intercept effects for the identity of the receiver, the signaler, and the receiver-signaler dyad to avoid pseudo-replication. To prevent an overconfident model and keep the type I error at the nominal level of 0.05, we included random slopes 24,25 of receiver age, signal category, and their interaction within signaler and also those of signal category, signaler age, and signaler sex within receiver. We also included parameters for the correlations among random intercepts and slopes. To avoid cryptic multiple testing 26, we compared this full model with a null model that lacked receiver age and its interaction with signal category in the fixed-effects part.
The model was fitted using the function glmer of the R package lme4. Prior to fitting the model, we z-transformed receiver age and signaler age to achieve easier interpretable estimates and to ease model convergence. We manually dummy coded and then centered signal category and signaler sex before including them as random slopes. We determined confidence intervals of model estimates and fitted values by means of a parametric bootstrap (function bootMer of the R package lme4; 1000 bootstraps). Significance of individual effects we obtained by dropping them from the full model, one at a time, and comparing the respective reduced models with the full model. All model comparisons were based on likelihood ratio tests 27. To estimate model stability we excluded signalers, receivers, and dyads one at a time, fitted the full model to each of the subsets and compared the estimates derived with those for the full data set. The model had a good stability in the fixed-effects part and did not suffer from collinearity 22 as indicated by a maximum Variance Inflation Factor of 1.023 (based on a model lacking the interaction).
The sample analyzed for this model comprised a total of 3115 events where we noted the responses of females (N=846 to affiliative and N =2269 to agonistic signals). Signals were given by 83 signalers to 50 receivers, which together formed 1005 signaler-receiver dyads. We observed a total of N = 2238 behavioral reactions and N = 877 ‘no response’. For male receivers, we observed a total of N = 557 events (279 affiliative, 278 agonistic signals). Males showed no responses to signals of others in 85 affiliative signaling events and 158 agonistic signaling events. Due to the smaller sample size and the model complexity, we refrained from further analyses of male receiver behavior.
Type of response
We next addressed which types of response individuals produced after an agonistic signal and how this choice was affected by receiver age (model 4). As above, we included the signaler’s age and sex as control factors. The reason that we did not consider responses after both signal types within one analysis was that most response types occurred exclusively in response to one signal category (‘Lip Smack’ only after affiliative signals; ‘Give Ground’, ‘Make Room’, ‘Present’, and ‘Squeak’ only after agonistic signals).
In the analysis of response types observed following agonistic signals, we included the patterns ‘Give Ground’, ‘Make Room’, ‘Present’, and ‘Lip Smack’, as these occurred with sufficient frequency (> 25, Supplementary Table 3). The model fitted was identical to the model of ‘any response’ (model 3; with the exception that it lacked signal category in the fixed as well as random-effects part. Since the response was multinomial and since we were not aware of an option to fit such a model with complex random effects structure in a maximum likelihood framework, we decided to use a Bayesian framework and utilized the function brm of the R package brms. We fitted the model with a maximum tree depth of 20 and set adapt delta to 0.99. The chains successfully converged as indicated by Rhat values between 1.000 and 1.001. The sample considered for this model comprised a total of 1594 responses by 50 receivers in response to signals of 81 signalers; signalers and receivers formed 654 dyads. We did not conduct a separate analysis for response types after affiliative signals, as the two types that occurred most frequently were relative similar facial expressions (‘Lip Smack’ and ‘Teeth Chatter’).
Data availability statement
Data and code for all analyses are available at https://osf.io/vjeb3/?view_only=7de3d702357844739d7a4da6fe5c5759.
The use of agonistic signals varied with age and sex (model 1; Likelihood Ratio test for negative binomial models: LR statistic = 48.89, df = 3, P < 0.001; Fig. 1a, Supplementary Table 4). The use of agonistic signals was highest in mid-adulthood, and males used such signals more frequently than females. To illustrate the monkeys’ behavior, young males used on average 2.43 agonistic signals/h, and young females 1.88 signals/h. In mid-adulthood, males used 4.15 agonistic signals/h and females 2.65 signals/h. For old males, the rate of agonistic signals was 2.55 signals/h, and for old females 1.25 signals/h.
The use of affiliative signals also varied with age, sex, and rank (model 2; Likelihood Ratio test LR statistic = 33.85, df = 2, P < 0.001; Fig. 1B; Supplementary Table 5). On average, females produced affiliative signals more frequently than males, and young and mid-adult subjects more frequently than old ones. More specifically, young males produced on average 0.99 affiliative signals/h, and young females 2.39 signals/h. In mid-adulthood, males produced 0.43 affiliative signals/h and females 1.21 signals/h. For old monkeys, the rate of affiliative signals was 0.96 signals/h for males and 0.55 signals/h for females.
Responses to signals
There was no evidence that the likelihood to react varied with age and signal category (no significant interaction between age and signal category: model 3, = 0.27, P = 0.606). The likelihood to respond to signals by others was not obviously different in older compared to younger females (Fig. 2) and did not vary with signal category. The probability of a response was ca. 0.8 for affiliative as well as agonistic signals (Supplementary Table 6). Concerning the control predictors, we found that the probability to respond clearly varied with signaler age, however: the older the signaler, the less likely it was that the recipient responded to that signal (Fig. 3).
Type of response
Receiver age had no strong effect on the type of response (Fig. 2). With regard to the control predictors, we found that ‘Teeth chatter’, a low-cost submissive signal, occurred more frequently in response to aggressive signals by females than to aggressive signals by males (the other three response types occurred with roughly comparable frequencies after female and male signals).
A follow-up analysis of the different response types showed that old females were generally more likely to respond to an agonistic signal with ‘Teeth chatter’. While receiver age did not strongly affect the likelihood to respond, signaler age did. The likelihood that females responded to a signal was 83% for the youngest monkeys’ signals, 72 % for middle-aged monkeys’ signals. The response rate to old monkeys’ signals was 57% (Supplementary Fig. 1). The likelihood to respond varied neither with signaler sex nor signal category (Supplementary Table 7).
Older monkeys used fewer affiliative signals such as teeth chattering and lip-smacking than younger monkeys, and males used affiliative signals less frequently than females. The variation in signal usage corresponds to the variation in affiliative signals involving physical interactions 16,17. Young females showed the highest rates of affiliative signals, suggesting that they have the highest motivation to establish and consolidate social bonds. In line with previous studies, males used agonistic signals more frequently than females. The age-related trajectory in the use of agonistic signals by males is in line with the peak of their resource holding potential around 15 years of age 28,29.
The social reclusion of older males and females appears to result from two processes, driven by younger individuals on the one hand, and the old individuals on the other. First, older monkeys are less often the targets of interactions and interact with fewer partners 17. Second, signals produced by older monkeys were more likely to be ignored by other group members, suggesting that older monkeys are perceived both as less threatening and less valuable as social partners. Yet, old individuals may maintain specific relationships with selected partners. Detailed analyses of the long-term development of dyadic relationships will be needed to explore the differentiated use of and responses to signals with regard to the relationship quality of a dyad.
Concerning the responses to other group members’ signals, our results did not conform to the predictions of the SAVI model. We did not find the predicted interaction between age and signal type in the responses to group members’ signals, as older monkeys were not more likely to ignore or move away from agonistic signals as a strategy to regulate negative affect.