The modularity of a social group does not affect the transmission speed of a novel, socially learned behaviour, or the formation of local variants

Abstract

The structure of a group is critical in determining how a socially learnt behaviour will spread. Predictions from theoretical models indicate that specific parameters of social structure differentially influence social transmission. Modularity describes how the structure of a group or network is divided into distinct subgroups or clusters. Theoretical modelling indicates that the modularity of a network will predict the rate of behavioural spread within a group, with higher modularity slowing the rate of spread and facilitating the establishment of local behavioural variants which can prelude local cultures. Despite prolific modelling approaches, empirical tests via manipulations of group structure remain scarce. We experimentally manipulated the modularity of populations of domestic fowl chicks, Gallus gallus domesticus, to affect the transmission of a novel foraging behaviour. We compared the spread of behaviour in populations with networks of high or low modularity against control populations where social transmission was prevented. We found the foraging behaviour to spread socially between individuals when the social transmission was permitted; however, modularity did not increase the speed of behavioural spread nor lead to the initial establishments of shared behavioural variants. This result suggests that factors in the social transmission process additional to the network structure may influence behavioural spread.

1. Introduction

Social learning enables information (such as a novel foraging behaviour) to spread throughout a group of animals [1]. The likelihood of an individual socially learning from another will depend upon the association/relationship between them; those individuals that are closely associated are more likely to learn from one another [2–4]. Group structure can be described and quantified using metrics derived from social network analysis [4]. Using this approach, it has been demonstrated that network structure does indeed correlate with information flow and can facilitate the formation of local behavioural variants underpinning behavioural traditions and local cultures [5–18]. These studies recorded information flow across established, natural networks, typically considering one or two animal groups per study (but see [7,12,13]). Consequently, isolating or identifying specific aspects of the network structure that are important determinants of information transmission is difficult because they are commonly bound up with other network properties and probably highly dependent on the study species, composition and history of the group. An alternative, manipulating group structure, has been used by several studies which have successfully altered social structure, either directly, through creating specific associations of group members [12,13,19], or through modifying their habitat [20] during foraging and while interacting with a novel task. However, these manipulations did not change a particular network metric of interest but instead unpredicted network structures emerged, which were then correlated to the efficacy of information transmission. Thus, a lack of direct empirical manipulations to create a specific social structure makes it difficult to determine the influence various network parameters have on social transmission.

An alternative approach to understanding how network structure and certain network parameters influence information flow in groups has been to use theoretical models [21,22]. One network property explicitly proposed to influence information flow and therefore constitute an essential part of the formation and maintenance of local behavioural variants is modularity [9]. Modularity is a structural property indicating the extent to which a network can be divided into sub-groups or clusters based on individuals' interactions with each other. This can be represented by Newman's coefficient value ranging from 0 (completely mixed networks where all individuals equally associate with each other) to 1 (a highly structured network that can be divided into perfect clusters where individuals exclusively associate with particular others) [23]. High modularity is suggested to alleviate disease burden within social groups through reducing transmission between clusters [24–27]. Information transmission may follow the same pattern—as modularity increases and clusters become more distinct, information is likely to spread rapidly within a cluster but fails to extend to other, neighbouring clusters [9]. Group-specific behaviours or specific variants of behaviour within clusters can therefore arise within animal populations of sufficient modularity, potentially leading to the emergence of behavioural traditions and culture [9,28]. Theoretical models simulating social transmission on networks have shown a threshold modularity coefficient value of 0.3, above which, as modularity increases so too did distinct behavioural variants arising within clusters in the network [28]. Despite these theoretical predictions, no study that we are aware of has empirically manipulated the modularity of a group's network and then observed how this subsequently affected the spread of information or behaviours on the network.

We empirically tested how the modularity of a population's network structure affects the spread of a novel behaviour within domestic fowl chicks, Gallus gallus domesticus. Domestic fowl are a gregarious species, are precocial and easy to rear in the laboratory. Importantly they have been shown to obtain crucial information through attending to the behaviours of others [29], with chicks in particular learning foraging behaviours through attending to both cues of their mother and fellow brood mates [30]. We manipulated the groupings of chicks as they encountered a novel foraging task to produce populations that exhibited networks of high or low modularity. To confirm that the solving behaviour of the foraging task spread socially between chicks, we compared the rate of learning in populations that could learn the novel foraging behaviour socially from others (those in both the high and low modularity populations), with populations in which chicks could only learn the task asocially. We predicted, in accordance with theoretical models [9,28], that (if socially learnt) the novel foraging behaviour would spread faster in the less-structured populations of low modularity. We also considered a difference in social learning rates might arise owing to differences in familiarity which could then affect behavioural spread. Chicks within populations of high modularity will associate with each other more frequently and consequently will be more familiar with one another. Given previous work has shown individuals are more likely to learn from familiar over unfamiliar individuals [31], we thus examined whether modularity increased social learning rates. We included sex and mass as variables that may affect the social or asocial learning rates, as they have been shown to influence learning (in other species) (e.g. [32–34]). Finally, we expected to find that within the populations of high network modularity, solving techniques of the novel foraging behaviour would be contained within clusters and so providing an example of an initial step in the emergence of behavioural traditions and culture within animal populations.

2. Methods

(a). Subjects and housing

One hundred and fifty-four domestic chicks (G. gallus domesticus, of the Rhode Rock hybrid), split across three batches, were collected as day-olds from a commercial breeder (Organic Pullets, UK). Each batch was split into two populations, and each population housed in a replicated (temperature and humidity controlled) pen, giving a total of six populations overall. The first batch comprised 27 birds in each population (11 male: 16 female), the second of 24 (12 male: 12 female) and the third of 26 (13 male: 13 female). We provided birds with commercial chick crumb (1st Poultry, Crediton Milling Company; UK), water and grit ad libitum, in addition to perching materials. Coloured plastic flat bands (Avian ID, UK) were fitted to identify individuals on day three, which we changed for a larger size at two weeks old. We also weighed birds when two weeks old, using a stand-on flat digital scale (Salter, UK; precision = 1 g). During the study period, birds interacted with a novel foraging task which involved them voluntarily entering a 0.60 × 0.50 m testing chamber adjoining their main enclosure (via a filtering chamber) to interact with the task (electronic supplementary material, figure S1). At three weeks old, we donated birds to a local free-range poultry farm.

(b). Experimental procedures

We presented birds with a novel foraging task and looked at the rate birds acquired the solving behaviour in populations upon which we had imposed different learning conditions (table 1). Birds within populations in batches 1 and 2 were tested under conditions which allowed them to socially learn the solving behaviour from one another and in which we manipulated the modularity of the group structure to produce two populations of high network modularity and two populations of low network modularity. Birds within populations in batch 3 were restricted to asocial learning only, which allowed us to determine the rate at which the solving behaviour would be acquired asocially without any social transmission.

Table 1.

Experimental conditions and dates for each of the six populations within the three batches.

batch	population	number of birds	condition
1	1	27	social – high modularity network
(03/11/17–25/11/17)	2	27	social – low modularity network
2	3	24	social – high modularity network
(26/01/18–17/02/18)	4	24	social – low modularity network
3	5	26	asocial control
(31/10/18–22/11/18)	6	26	asocial control

(i). Novel foraging task

The novel foraging task consisted of a white box with 10 lightweight cardboard doors that could be slid horizontally open to reveal wells containing mealworm rewards (figure 1). The apparatus was adapted from previous studies of social learning in birds [7,35,36]. The 10 wells (2 cm diameter, 1.9 cm deep) were concealed by cardboard doors that were painted half-red half-blue. Birds could solve the task by two different techniques (pushing the red side to the right or the blue side to the left) to access the well behind and obtain the mealworm rewards.

Figure 1.

Foraging task apparatus. The novel foraging behaviour required to solve the task involved birds pecking at either the blue (right) or red (left) side to slide the door open in order to access the mealworms behind. Bottom right door opened to reveal well. (Online version in colour.)

(ii). Experiment structure

Prior to introducing the novel forging task, in week 1, all birds were shaped to associate the testing chamber with live mini-mealworm rewards and habituated to handling by the experimenter (see the electronic supplementary material). Birds in all populations received the same experimental structure comprising of three stages; a pre-test round (presented when birds were 7 days old—day 7), 13 grouped rounds (days 13–19) and finally a post-test round (day 20). The pre-test and post-test round consisted of all birds in each population interacting with the task apparatus individually for 1 min to (i) confirm that none could solve the task prior to the experiment (pre-test), and (ii) identify those able to solve the task at the end of the experiment (post-test). In the pre- and post- test round, one mealworm was available in each of the 10 wells. The 13 grouped rounds differed slightly between the conditions. For populations in the social conditions where social learning of solving behaviour was permitted, in order to ensure a starting point for the diffusion of solving behaviour, one bird per population was selected to be a seeded demonstrator individual and was trained prior to the start of grouped rounds to solve the task using the technique of pushing the red side of the door to the right (days 8–12; see the electronic supplementary material). All birds interacted with the task in pre-determined groups of three for 3 min where they could freely interact with the task and each other. Three mini-mealworms were available in each of the 10 wells of the apparatus in these grouped rounds. Populations in the asocial condition could only learn the solving behaviour from personal experience (asocial learning) and so during the 13 grouped rounds focal birds interacted with the task individually, and one mealworm was available per well. To keep the social environment consistent across conditions and thus ensuring the mere presence of fellow birds in the testing chamber would not influence exploration of and interaction with the task, two companion birds were present in the testing chamber (alongside the focal individual) in the asocial condition. These two companion birds (one male, one female) were the same for the whole population and were prevented from accessing the task apparatus by a mesh barrier (see the electronic supplementary material). For all conditions, during each round, we recorded which individuals solved the task (they opened a door, took and ate a mealworm) and the technique used to do so (whether they pushed red to the right or blue to the left).

(iii). Manipulation of social structure

Birds within populations in the social conditions (batches 1 and 2) were permitted to learn socially from one another and so entered the testing chamber in groups of three individuals. To create social networks with specific modularity of our choosing (spanning the critical threshold of 0.3 [28]), we pre-determined the composition of groupings entering the testing chamber. The social networks represented how many rounds individuals associated with each other during task interaction in the testing chamber. For instance, one individual may have only associated with a certain individual once in the testing chamber but eight times with another. In each batch, we created one population with a network structure of high modularity and one population with a network structure of low modularity (figure 3). Populations of high network modularity had groups that more often than not comprised the same three individuals entering the testing chamber. This gave highly structured networks with modularity coefficient values of 0.73 (batch 1) and 0.63 (batch 2). By contrast, the populations of low network modularity saw much more varied combinations of the three individuals every round and so created much more random well-mixed networks with modularity coefficients of 0.15 (batch 1) and 0.18 (batch 2). As we had pre-determined our testing chamber groupings, each round of birds were individually selected from the main enclosure and placed within the filtering chamber before entering the testing chamber (see the electronic supplementary material). To ensure equal handling across conditions, birds within the asocial condition were also handled to the same extent.

Figure 3.

Social networks of the high (a,b) and low (c,d) modularity populations. Node colours represent an individual's bias in solving technique through a gradient of red (all solves performed were pushing red) to blue (all solves performed were blue), yellow nodes indicate birds that never performed the solving behaviour. Initial seeded demonstrators in each population denoted by ‘D’. Node size is proportional to the total number of solves performed by each bird. (Online version in colour.)

(c). Statistical analysis

(i). Does social learning and network modularity affect the likelihood of acquiring the solving behaviour?

To investigate whether condition (high modularity social, low modularity social, asocial control) influenced the time to acquire the solving behaviour, we fitted a Cox proportional hazards mixed-effect model with the condition as a categorical explanatory variable and population as a random effect using the ‘survival’ and ‘coxme’ packages [37–39] in R v. 3.3.1 [40] using the R studio wrapper [41]. The round each bird learned is the dependent variable, with birds that did not learn treated as censored observations capped at the last post-test round. Importantly, within the social condition populations, we excluded the initial demonstrator bird that we had trained from the analysis. Cox survival analysis models the effect that predictor variables have on the ‘hazard rate’ for each bird—e.g. if bird A has a hazard rate that is double that of bird B, it is expected to learn the behaviour in half the time. Thus, we can estimate hazard ratios for each condition relative to the asocial control with a score greater than 1 showing increased hazard of acquiring the solving behaviour and a score of less than 1 a decreased hazard of acquiring the behaviour.

(ii). Does the solving behaviour spread socially throughout the populations and do rates differ between social network structures?

To determine whether the behaviour performed to solve each task was learnt socially or asocially, we conducted a network-based diffusion analysis [42,43], specifically a dynamic order of acquisition diffusion analysis (OADA) using the ‘NBDA’ package [44] in R. This analysis is an extension of the Cox model which assumes that the hazard rate for an individual solving the task by social learning is proportional to their strength of connections to informed individuals. If the task is socially learnt, then the spread of solving behaviour will follow the links in the group's social network. If the solving behaviour is learnt asocially (say through trial and error) then the order in which individuals solve the task will be independent of the social network. NBDA allows us to evaluate the evidence for whether the solving behaviour was at least sometimes learnt socially, as opposed to purely asocially and provides an estimate of the strength of social transmission relative to asocial learning. A dynamic OADA allows for the network to change over time and so directly tracks the spread of behaviour with changing associations [45].

We created dynamic networks that represented the changing testing chamber associations and therefore potential social learning opportunities (see the electronic supplementary material). These social networks were then related to the order in which birds acquired the solving behaviour (having marked the initially trained seeded demonstrator individual within the social condition populations as already informed (see the electronic supplementary material)). We used a stratified OADA [46] where we assumed that the baseline rate of asocial learning was equal across all the populations (derived from the rate at which birds acquired the solving behaviour in the asocial control condition populations). Social transmission rates (how quickly birds were to socially learn the behaviour from an informed individual during a round spent together) may vary if social dynamics (such as familiarity with certain individuals) affected by social structure, influenced social learning. Thus, to determine whether modularity condition affected the rate of social transmission, we compared models in which the social transmission rate differed among modularity condition with models in which the social transmission rate was constant or where social transmission rates differed dependent on population. Sex and mass of each bird were included as individual-level variables that may affect the asocial and/or social learning rate. All possible combinations of differing social transmission rates and individual-level variables were fitted and the Akaike information criteria corrected for small sample size (AICc) used for model averaging and in selecting the best predictive model [46,47] (see the electronic supplementary material).

(iii). Does social structure influence the adoption of different solving techniques by the birds?

To investigate whether social structure influenced the technique adopted by individuals, we tested for ‘assortment’ based on any shared bias in the technique (pushing red or blue sides) by linked individuals, using the ‘assortnet’ package [48]. As we had pre-determined testing chamber groupings we were not testing for the usual assortment where individuals are free to associate with others at will. Rather we were looking at whether birds exhibited a similar solving technique to those others they were most frequently placed in the testing chamber with. We tested for this ‘assortment’ on the networks of testing chamber associations for birds in the high modularity and low modularity condition populations. As not all birds performed the same number of solves, and there is inherent uncertainty in estimating bias from a few observations, we accounted for uncertainty using a Monte Carlo approach. For each of 10 000 simulations, we draw an estimate of each bird's true probability of pushing red from a beta distribution informed by their observed pattern of solving (see the electronic supplementary material for details). We then measure the assortment of the network by these estimates of true bias. From these simulations, we derive a distribution of weighted assortment measures. We report the mean estimate (with 95% confidence intervals (CIs)) of assortment from these simulations. An assortment coefficient value of 0 would indicate no ‘assortment’ (birds solving technique biases are independent of testing chamber groupings), while a value of 1 would indicate ‘perfect assortment’ (solving techniques are completely dependent upon testing chamber groupings, with birds having a shared technique with their testing chamber group mates), and finally negative values would indicate disassortment (birds having solving techniques that significantly differ from their testing chamber group mates) [49].

3. Results

(a). Does social learning and network modularity affect the likelihood of acquiring the solving behaviour?

Birds in the social condition populations (of both high and low modularity) were much more likely to learn the solving behaviour with each round compared to birds in the asocial condition populations (high modularity condition: Coxph, Z = 5.82, events (e) = 83, p ≤ 0.001; low modularity condition: Coxph, Z = 5.24, events (e) = 83, p ≤ 0.001 (figure 2)). Birds in the high modularity condition were estimated to solve the task 12.94 times faster (95% CI = 5.46–30.65), and birds in the low modularity condition 10.18 times faster (95% CI = 4.30–24.11), than birds in the asocial condition. However, there was no evidence of a difference in the speed of learning between birds in the high and low modularity conditions (hazard ratio high/low = 0.79, 95% CI = 0.44–1.39, Z = −1.05, events (e) = 83, p = 0.29 (figure 2)).

Figure 2.

The cumulative probability a bird has of acquiring the solving behaviour in each round when from populations of the high modularity social condition (dark green/triangles), low modularity social condition (light green/squares) and asocial control condition (grey/circles). Ninety-five percentage of confidence intervals shown by coloured bands. (Online version in colour.)

(b). Does the solving behaviour spread socially throughout the populations and do rates differ between social network structures?

Fitting stratified OADA models showed that the solving behaviour spread socially throughout the populations where social learning was permitted. To quantify differing social transmission rates, we averaged across all candidate models weighting relative to their AICc value. This approach showed no support for models of pure asocial learning in all the diffusions (∑w < 0.001) with the most support for social transmission of equal rates for both high and low modularity conditions, and less support for rates differing dependent on population or modularity condition (table 2). Although candidate models included fewer models of pure asocial learning compared to ones that included social transmission, the support given to social transmission models far outweighed asocial learning even when asocial support is multiplied up to the equivalent number of models (see the electronic supplementary material). Using model averaging, we obtained the support of less than 50% (∑w = 0.5) for either sex or mass influencing the asocial or social learning rate (table 3), indicating they are unlikely to be in the best predictive model. The final top model therefore included an equal social transmission rate (s) across populations of high and low modularity, with s estimated at 0.53 (95% CI: 0.20–1.16) times greater than the asocial learning hazard rate per informed connection (corresponding to 40% (95 CI: 26–50) of events estimated to be through social transmission), with no influential effect of sex or mass on the learning rate.

Table 2.

Support shown for differing hypothesis after model averaging across all candidate models.

model	support (∑ w)
pure asocial learning in all diffusions	<0.01
equal rates of social transmission for both high and low modularity	0.51
different rates of social transmission depending on network modularity	0.17
different rates of social transmission depending on population	0.32

Table 3.

Support for the variables sex and mass influencing the social or asocial learning rate.

learning rate effected	variable included	support (∑w)
asocial	sex	0.23
asocial	mass	0.21
social	sex	0.32
social	mass	0.29

(c). Does social structure influence the adoption of different solving techniques by the birds?

We found no evidence of shared forms of solving technique by closely associated individuals. The mean assortativity coefficient for the high modularity networks was −0.10 (95% CI: −0.25–0.07) and for the low modularity networks −0.003 (95% CI: −0.14–0.19). Therefore, birds that were mostly closely connected within networks of high or low modularity condition, did not share specific solving techniques (figure 3).

4. Discussion

Social learning was essential for enabling a novel foraging behaviour to spread throughout populations of domestic chicks. When social learning was restricted and birds could only learn through asocial means, they were far less likely to acquire the solving behaviour compared to those populations which permitted social learning to occur. However, contrary to our predictions, manipulating the social structure, specifically the modularity of the network which varied from high (a population comprising several distinct small clusters of birds that repeatedly encountered the foraging task together) to low (a population that was well mixed such that there were no distinct clusters of birds repeatedly encountering the task together), did not greatly affect the rate of behavioural spread throughout the population, the rate of social transmission, nor lead to the establishment of shared forms of solving techniques within clusters.

The novel foraging behaviour (solving the task through sliding open doors to access mealworm rewards) spread rapidly throughout the populations where social learning was permitted. Individuals in both the high and low modularity conditions (where social learning could occur) were over 10 times faster at acquiring the solving behaviour compared to individuals in the asocial learning populations, an effect that network-based diffusion analysis indicates is owing to social transmission of solving behaviour. The NBDA estimated 40% of learning events (in the social condition populations) to be a result of social transmission; however, we suspect this to be greater as birds entering the testing chamber with fellow uninformed individuals are potentially more likely to solve the task if those uninformed individuals had previously witnessed solves and were therefore displaying attention/interest towards the task apparatus—an effect that would not be detected as social learning. We found no strong influence of sex or mass on the social or asocial learning rate. This may be because these factors do not have a significant effect on social transmission when birds are young and before they are fully developed and sexually dimorphic. Given that familiarity has been shown to influence the likelihood of social learning [31], we expected that individuals within the high modularity population who were repeatedly interacting together in the testing chamber (and consequently were more familiar with one another) may be more likely to learn from one another. However, we found no evidence of social transmission rates differing dependent on modularity condition. Given that we artificially selected and enforced group composition and social structure, it may be that familiarity influences the likelihood of social learning only when it is the result of natural partner preferences.

We predicted, based on theoretical models [22] and providing that social transmission rates were equal across modularity conditions, that the novel behaviour would spread faster within less-structured populations of low modularity. However, we found no difference in the overall rate of acquisition across populations in the high and low modularity social conditions. Similar to our results, a recent meta-analysis on disease spread in relation to modularity found limited relevance of modularity on spreading events [50]. They also reported how factors such as the infectiousness/contagiousness of disease alter how modularity influences disease spread, with modularity being ineffective at restricting spread for highly contagious diseases. Currently, most theoretical models of behavioural transmission follow the assumption that individuals with more connections to others are more likely to acquire and pass on behaviours [51]. However, as with disease transmission, if behaviours also follow more complex contagion rules, then groups of differing social structure, such as their modularity, might generate differing results for behavioural spread. First, the proportion of informed to uninformed individuals may be more influential in behavioural acquisition than simply the total number of connections to informed individuals [51]. Modularity will by default influence these relative proportions and so even though birds within the low modularity populations held more total connections to informed individuals, the proportion of these informed connections may have been lower than in high modularity populations. Second, factors such as social reinforcement and forgetting rates (perhaps analogous within ‘infectiousness’ in disease models) may change how modularity influences behavioural transmission. If birds required greater social reinforcement (several separate exposures of the demonstrated behaviour) in order to learn, then lower modularity may not lead to faster transmission because each pairing might only occur once or a few times, so impeding social reinforcement, whereas birds in the high modularity populations were consistently entering the chamber with the same (potentially informed) individuals each round. Only if the behaviour was learnt quickly (say after a single exposure) would low modularity lead to the rapid spread of behaviour throughout the group. Indeed social reinforcement is thought to be the main driver in studies which found somewhat counterintuitive results of increased modularity promoting information diffusion [52,53]. Similarly, forgetting rates (the likelihood that an informed individual ceases to exhibit a learned behaviour) may have also played a role. Depending on the forgetting rate, social structures have been shown to have different effects on information transmission including the likelihood of behavioural extinction [54]. If the forgetting rate is high, then increased modularity may facilitate spread as it retains clusters of familiar individuals that are likely to keep demonstrating the behaviour to one another. If we consider a complex model of social transmission, then an optimum level of modularity may exist whereby local spreading is promoted and acts as an incubator facilitating global spreading across the whole network provided there is a sufficient level of connectivity [53]. By looking at two extreme conditions of modularity as we did, we may have examined two equally (sub)optimal conditions for behavioural spread. Birds in the low modularity condition may have been overall more likely to encounter informed individuals, but, depending on the time interval between subsequent demonstrations they might not have adopted the behaviour or if they did, they were more likely to subsequently forget it. Alternatively, birds in the high modularity condition may have been overall less likely to encounter informed individuals but those that did so probably experienced repeated demonstrations in short succession.

We predicted that if birds were to match their behaviour to others then we would see the emergence of local behavioural variants within clusters in the high modularity populations, which might indicate the beginnings of culture. However, contrary to other studies [7,10], we found no evidence that birds matched their solving technique to others closest to them in the network. One possibility why we found no evidence of assortment was that the birds were using a different social learning mechanism to those in other studies. Within a diffusion experiment, observational learning mechanisms of either emulation or imitation have been suggested to explain how social network position predicted the behavioural variant acquired by squirrel monkeys, Saimiri boliviensi [10]. In a similar experiment which showed that sub-populations of wild British songbirds established local traditions of a specific foraging technique, although the social learning mechanisms responsible were not discussed, the mechanism of local enhancement was ruled out [7]. Our results, however, may have arisen owing to other forms of social learning mechanisms such as local or stimulus enhancement. Instead of paying attention to the exact action, individuals might have had their attention to the apparatus increased but they paid little attention to the precise action of others (which direction they opened the door or colour that they pecked at). A second explanation for our results is the timescale that the experiment operated over. We were recording the solving techniques used during the first acquisition of the task and gave them only 15 opportunities to exhibit a successful food retrieval. It may be that birds would match and conform to the technique of their predominant social partners over a longer time, resulting in group-specific behaviours in the high modularity condition. A conformity effect over time was found in the sub-populations of songbirds [7]; however, an erosion of group differences over time was found within groups of squirrel monkeys [10]. Finally, it is possible that either the overall population size (n = 24,27) or our sub-group size (three individuals) that we used were simply too small to see an effect of local behavioural variants. It is possible that the effect of modularity is only detectable in large groups containing many large clusters. Replicating such networks was beyond the scope of our experimental study.

Our experimental manipulations show that the modularity of a social network did not affect either the speed of transmission of a novel behaviour nor did it enhance or retard the formation of shared behavioural variants. Although our experimental study was restricted to a relatively small population size comprising small-sized sub-groups/clusters, we believe our results, despite not supporting the role of modularity in shaping behavioural spread, highlight how other factors such as proportional connection to informed individuals, social reinforcement and forgetting rates may be important components that can obscure effects of social structure in the transmission of information across groups. These factors should be considered in future empirical work and theoretical models of behaviour spread.

Ethics

All work was approved by the University of Exeter Psychology Ethics Committee and birds were reared at a high standard, exceeding the requirements provided by the Department for Environment Food and Rural Affairs and the Animals Scientific Procedures Act (1986).

Data accessibility

The data supporting this article are available on the Dryad Digital Repository at: https://doi.org/10.5061/dryad.wpzgmsbm4 [55].

Authors' contributions

P.R.L. conceived the study with J.R.M. P.R.L. collected the data. P.R.L., W.H. and M.W. analysed the data. P.R.L. wrote the first draft of the paper, with comments from all authors.

Competing interests

We declare we have no competing interests

Funding

The work was funded by an ERC Consolidator Award to J.R.M. (616474).