Processing of novel food reveals payoff and rank-biased social learning in a wild primate

Study site and subject details

The study was conducted at the Inkawu Vervet Project (IVP) in a 12,000-hectare private game reserve: Mawana (28° 00.327 S, 031° 12.348 E) in KwaZulu Natal province, South Africa. The vegetation of the study site consisted in a savannah characterized by a mosaic of grasslands and clusters of trees of the typical savannah thornveld, bushveld and thicket patches. We studied two groups of wild vervet monkeys (Chlorocebus pygerythrus): ‘Noha’ (NH) and ‘Kubu’ (KB). NH was composed of 34 individuals (6 adult males; 9 adult females; 6 juvenile males; 7 juvenile females; 5 infant males; 1 infant female) and KB was composed of 19 individuals (1 adult male; 6 adult females; 3 juvenile males; 4 juvenile females; 3 infant males; 2 infant females; Table S1). Males were considered as adults once they dispersed, and females were considered as adults after they gave their first birth. Individuals that did not fulfil these criteria were considered as juveniles²⁸ and infants were aged less than 1 year old. In EWA models, infants and juveniles were lumped in a single category “juveniles”. Each group had been habituated to the presence of human observers: since 2010 for NH and since 2013 for KB. All individuals were identifiable thanks to portrait photographs and specific individual body and face features (scars, colours, shape etc.).

This research adhered to the “Guidelines for the use of animals in research” of Association for Study of Animal Behaviour, was approved by the relevant local authority, Ezemvelo KZN Wildlife, South Africa and complied with the ARRIVE guidelines.

Hierarchy establishment

Agonistic interactions (aggressor behaviour: stare, chase, attack, hit, bite, take place; victim behaviour: retreat, flee, leave, avoid, jump aside) were collected from May 2018 to October 2018, aside from experiment days, on all the adults and juveniles of both groups via ad libitum sampling⁵⁰ and food competition tests (i.e. corn provided to the whole group from a plastic box). Data were collected by CC, MBC and different observers from the IVP team. Before beginning data collection, observers had to pass an inter-observer reliability test with 80% reliability for each data category between two observers. Data were collected on tablets (Vodacom Smart Tab 2) equipped with Pendragon version 8.

Individual hierarchical ranks were determined by the outcome of dyadic agonistic interactions recorded ad libitum and through food competition tests using Socprog software version 2.7⁵¹. Hierarchies in both groups were significantly linear (NH: h′ = 0.27; P < 0.0001; KB: h′ = 0.42; P < 0.0001) and ranks were assessed by I&SI method⁵².

Open diffusion experiment

The experimental apparatus consisted of provisioning the group with two transparent rectangular plastic boxes (34 × 14 × 12 cm) containing ~ 2 kg unshelled peanuts in sufficient quantities to prevent a single individual from monopolizing the boxes. The monkeys were never provided with peanuts before the experiment and peanuts were not available in their environment. Thus, unshelled peanuts were a novel, nutritious food that required processing to be extracted from their shells before consumption.

Experiments took place at sunrise at monkeys’ sleeping sites during the dry, food-scarce winter to maximize their motivation for novel food. The two boxes of peanuts were offered to the monkeys, spaced apart by about 1–10 m. CC led the experiment with the help of two to four field assistants. All monkeys were free to come to the boxes within the constraints of the social group dynamics. Experiments were video recorded using JVC cameras (EverioR Quad Proof GZ-R430BE) to which the experimenter said aloud the identities of the actor and of the attending neighbours for each manipulation event. A manipulation event was defined either as an attempt to extract a peanut from its shell (i.e. the individual acted on the peanut failing to fully open it and to get access to the food) or as a success (i.e. the individual succeeded to fully open a shell and to extract the peanut from the shell). A conspecific was considered as attending when it had its head or body oriented in an unobstructed line towards the demonstrator manipulating the peanut and was located within 0–30 m from the actor. Several individuals could thus be registered as attending to one or several demonstrators simultaneously.

The open diffusion experiments ran from May 2018 to August 2018 to maximise individuals’ likelihood of participating in the experiment. A total of 11 sessions of open diffusion experiments were run in NH and 10 in KB. The average duration of an experimental session was 46 m:46 s for NH and 42 m:47 s for KB.

Video analysis

Video recordings were viewed and analysed by MBC with Media Player Classic Home Cinema software version 1.7.11. Twenty percent of the video were analysed by CC and inter-observer reliability was substantial (κ = 0.78). During video analysis in slow motion or frame by frame, the following variables were coded in an excel sheet: date, exact time of each manipulative event, actor identity, the technique used (crack with hand: ‘CH’; crack with mouth from the top of the peanut: ‘CMT’; crack with mouth from the side of the peanut: ‘CMS’; see Movies S1–S3) and the identity of attending individuals.

Quantification and statistical analysis

Following Barrett et al.¹⁷, we used a suite of hierarchical experience-weighted attraction (EWA) models to analyse data collected in the open-diffusion experiment. EWA models are time-series models that evaluate the joint influence of personal experience and social information on the probability of an individual displaying a behaviour³⁸ and are increasingly utilized in cross-taxa studies of cultural transmission^{16,17,18, 27,53}.

This analytical approach has several strengths: it permits evaluation of multiple hypothesized learning strategies against each other and individual learning alone, utilizes a dynamic social learning network unique to each individual, and links individual variation in behaviour and cognition to population level-cultural dynamics. Working with time-series of behaviour unique to each individual is important as population-level signatures can often be misleading⁵⁴ or exhibit equifinality, particularly if individuals vary in experience, observation opportunities or the social learning strategies they employ⁴⁵. The mathematical specification of our analytical approach also minimizes ambiguity of what types of social learning we are evaluating. This is important as verbal definitions are imprecise, and terminologies are differently interpreted in studies of social learning. Most importantly, this approach links theory to data. Instead of using a theoretically uninformed analytical approach to find results consistent with theory, we bypass quantitative proxies and translate theoretical models to statistical models. We fit a series of EWA models evaluating the following learning strategies:

1. 1.

   Individual learning alone

2. 2.

   Frequency-dependent learning (preference for behaviours that is based on their frequency in the population)

3. 3.

   Female-biased learning (preference for the technique displayed by females in group i.e. matrilineal sex in vervets)

4. 4.

   Matrilineal kin-biased learning (preference for the technique displayed by closely related individuals)

5. 5.

   Compare means payoff-biased learning (preference for the most successful or efficient behaviour)

6. 6.

   Rank-biased learning (preference for the technique displayed by high-ranking individual)

7. 7.

   Sex-biased learning (preference for the behaviours of individuals that are of the same sex)

8. 8.

   Global model that includes 1–7.

All social learning models (models 2–8) also include an individual learning component.

For each behavioural choice, social information used by an actor was the average value of each cue observed in a time window of 20 min prior to the observation (Table 3). As individuals access social information at different timescales, and this window choice was somewhat arbitrary, we also evaluated social info at 30, 10, and 5-min timescales. These analyses yielded similar results and were robust to time windows.

We ran the EWA models using regularizing priors, which are sceptical of extreme effects and reduce the risk of overfitting, and a Cholesky decomposition for estimating varying effects. Models were fit using RStan version 2.19.3⁵⁵. Models were compared using widely applicable information criteria (WAIC), which can inform which model best predicts the observed data while penalizing models that underfit or overfit. Models with lower WAIC scores best predict the observed data.

EWA model specification

EWA models have two parts: a set of expressions that specify how individuals accumulate experience and a second set of expressions that specify the probability of each option being chosen. Accumulated experience is represented by attraction scores, ({A}_{ij,t}), unique to each behaviour (i), individual (j), and time (t). We update ({A}_{ij,t}) with an observed payoff ({pi }_{ij,t}):

The parameter ({phi }_{j}) controls the importance of recent payoffs in influencing attraction scores. When ({phi }_{j}) is high, more weight is given to recent experience over past experiences—memory has less of an influence on behavioural choice. This parameter is unique to an individual (j), and we also estimate how it varies by age-class and sex.

Attraction scores are converted into probabilities of behavioural choice with a standard multinomial logistic choice rule:

({lambda }_{j}) controls sensitivity to differences in attraction scores on behavioural choice and is unique to an individual (j). A very large ({lambda }_{j}), means the option with the largest attraction score is nearly always selected. Choice is random with respect to attraction score when ({lambda }_{j}=0). Individuals were assigned a payoff of zero, ({pi }_{ij,t}=0), if they failed to open a peanut. If they were successful ({pi }_{ij,t}=1).

Social learning may directly influence choice distinctly from individual learning. ({S}_{ij}=Sleft(i|{Theta }_{j}right)) is the probability an individual (j) chooses behaviour (i) on the basis of a set of social cues and parameters ({Theta }_{j}). These social cues are traits associated with demonstrators (i.e. age, rank), or a behaviour (i.e. mean payoff), and each cue represents a hypothesized social learning strategy. Behavioural choice is a convex combination specified by:

where ({gamma }_{j}) is the weight assigned to social cues, and is bounded by 0 and 1.

Social cues are incorporated into ({S}_{ij,t}) by use of a multinomial probability expression with a log-linear component ({B}_{ij,t}) that is an additive combination of cue frequencies. The probability of displaying each behaviour (i), solely as a function of social cues, is:

({N}_{ij,t}) are the observed frequencies of each technique (i) at time (t) by individual (j). The exponentiated parameter ({f}_{c}) controls the amount and type of frequency dependence. When ({f}_{c}=1), social learning is unbiased by frequency and techniques influence choice in proportion to their occurrence (sometimes referred to as unbiased transmission). When (f>1), social learning is positive frequency-dependent or conformist. When (f<1), social learning is negative frequency-dependent, and a bias is shown towards rare behaviours.

Other social cues associate with individuals (i.e. rank, age, or relatedness) or behaviours (i.e. payoffs), are incorporated via:

({B}_{ijt}) is the sum of the products of the influence parameters ({beta }_{k}) and the cue values ({kappa }_{k,ijt}). Cues evaluated in these models are explained in detail in Table 3. The above specification is for the global model. Single social learning strategies are subset of this global model in which there is a single cue value or only frequency information is used.

Other statistics

We used generalized linear models (GLMs) with quasi-Poisson error and log link function to test for the effect of group, normalized rank, sex and age on (1) the latency of first success of shelling a peanut; (2) the number of peanut shelling successes; (3) the number of peanuts manipulation (attempts + successes); (4) the number of successes observed; (5) the number of manipulations observed; (6) the number of times being observed when succeeding and (7) the number of times being observed when manipulating. We added the log of the time each individual had to process peanuts once they first succeeded as an offset (a standard statistical technique for converting a Poisson GLM for analysis of counts into a model for analysing counts per unit of time). All tests were performed with R Studio version 1.2.1335 using R version 3.6.1⁵⁶.

Source: Ecology - nature.com