Detecting social transmission in networks
Introduction
‘Social learning’ is broadly defined as learning that is influenced by observation of or interaction with a conspecific or its products (Heyes, 1994). Social learning can result in ‘social transmission’, which we define as occurring when the acquisition of information or a behavioural trait by one individual exerts a positive causal influence on the rate at which another acquires the same information or trait. Social learning appears widespread across both vertebrate and invertebrate taxa (Hoppitt and Laland, 2008; Leadbeater and Chittka, 2007), whilst experimental work has established that social transmission can result in the establishment of behavioural traditions (e.g. Galef and Allen, 1995; Whiten et al., 2005). This has lead to claims of animal cultures in natural populations of apes (McGrew, 1998; Whiten et al., 1999; van Schaik et al., 2003), cetaceans (Rendell and Whitehead, 2001; Krützen et al., 2005) and monkeys (Perry and Manson, 2003). However, such claims remain controversial because studies fail to adequately rule out alternative explanations for local differences in behaviour, such as local environmental differences, or genetic differences between populations (Laland and Hoppitt, 2003; Laland and Janik, 2006). There is concern that the current ‘ethnographic’ method, which infers social transmission only where the alternatives of genetic or environmental variation can be disregarded, will rule out genuine cases of social transmission that covary with these factors (Laland and Janik, 2006; Laland and Galef, 2009). Consequently, in recent years researchers have called for the development of quantitative methods for inferring social transmission from field and captive study data that can rule out alternative explanations for the observed effect (Laland and Janik, 2006; Laland and Galef, 2009, and chapters therein).
One type of data that has previously been used to infer social transmission in groups of animals is diffusion data, where researchers monitor the spread of a novel behavioural trait. For some time the shape of the ‘diffusion curve’ (the cumulative number of individuals seen to perform the novel behaviour plotted against time) was used to infer social learning (e.g. Lefebvre, 1995a, Lefebvre, 1995b). The assumption was that if learning were asocial, the rate of learning would be the same for all individuals, resulting in an r-shaped diffusion curve. In contrast, if there were social transmission, the rate of learning would increase as the number of demonstrators increased, resulting in an s-shaped curve (Reader, 2004). However, this approach has been somewhat discredited, since there are a number of situations in which we expect to see an s-shaped diffusion curve in the absence of social transmission (Laland and Kendal, 2003; Reader, 2004), or an r-shaped curve in the presence of social transmission (Franz and Nunn, 2009).
An alternative method is to use the order in which individuals acquire a behavioural trait to infer social transmission from diffusion data, on the assumption that if social transmission is operating we might expect the spread to follow the patterns of associations between individuals (Boogert et al., 2008; Morrell et al., 2008). The reasoning here is that individuals that are closely associated are more likely to learn from each other (Coussi-Korbel and Fragaszy, 1995). A randomisation approach has already been applied to test for such a pattern (Boogert et al., 2008; see also Morrell et al., 2008), but below we demonstrate that this approach is vulnerable to both Type I and Type II errors.
Here we propose an alternative method, which we call order of acquisition diffusion analysis, or OADA, where a model of social learning is fitted to the data by maximum likelihood, and tested against a model with no social transmission.1 Our approach is similar to a method recently proposed by Franz and Nunn (2009), which they term ‘network-based diffusion analysis’ (or NBDA). Franz and Nunn's method exploits data on the time at which individuals acquire a behavioural trait, rather than the order in which they do so. However, as OADA and the randomisation approach of Boogert et al. (2008) are also network-based diffusion analyses, for clarity we rename Franz and Nunn's approach time of acquisition diffusion analysis (or TADA), and retain NBDA as the more general term for network-based approaches. We see the OADA and TADA approaches as complementary, and in later sections of this paper we introduce the OADA model, extend Franz and Nunn's TADA method, and provide a full comparison of OADA and TADA models. We end by illustrating the methods by applying them to a published data set: the diffusion of novel foraging traits in groups of starlings, Sturnus vulgaris (Boogert et al., 2008).
Section snippets
Boogert et al.'s (2008) randomisation method
First, we will describe Boogert et al.'s (2008) randomisation method and illustrate its limitations. To implement this method, for each group in which a diffusion is recorded, one needs a matrix containing an appropriate measure of association between individuals (the association matrix), and the order in which individuals acquired the behavioural trait (the ‘diffusion chain’). The test statistic is then simply the summed strength of associations between adjacent individuals in each diffusion
Modelling social transmission
Our starting model assumes that the rate at which social transmission occurs between a given dyad of informed and naïve individuals is linearly proportional to the association between them. This assumption is likely to be reasonable provided that (a) the probability a naïve individual observes, or is exposed to, the performance of the novel trait is proportional to its association with the demonstrator, and (b) all informed individuals are approximately equally likely to perform the trait. The
Comparison of OADA with TADA
Here we describe and extend Franz and Nunn's NBDA method, which we rename TADA, in the context of our OADA model, and using our notation. This facilitates a direct comparison between models reliant on order or time of acquisition.
TADA makes the same assumptions about social transmission as our model (Eq. (1)), but the models are fitted to time of acquisition data rather than to order of acquisition data, meaning the absolute rate of acquisition, λi(t), is modelled, rather than the relative rate
Simulation details
We compared how the OADA, TADA and randomisation models performed under different circumstances. All simulations considered the diffusion of a single learned behavioural trait through a single hypothetical group of animals of size N. Where the rate of acquisition of the trait was affected by an individual-level variable, this was generated by drawing a value for each individual from a normal distribution (x∼N(0,1)). We simulated an association matrix for the population by first generating a
Application of the models to Boogert et al. (2008)
We go on to illustrate the methods by applying OADA and TADA to a published dataset. Boogert et al. (2008) presented three captive groups of five starlings (S. vulgaris) with six different artificial foraging tasks. Each task was presented separately for several sessions. The time (measured cumulatively over sessions) at which each individual first contacted each task and first solved each task was recorded. Associations between individuals were calculated as the proportion of discrete point
Comparison in the absence of individual-level effects
In the absence of individual-level effects, and for a given group and effect size, TADA typically had more statistical power to detect social transmission than did OADA, while both of these methods were more powerful than the averaging and linear randomisation methods (Fig. 1a and b). In the case of the randomisation methods, the averaging metric usually provided more power than the linear metric, especially for larger group sizes, where social transmission is less likely to occur between
Discussion
The above simulations bring home the desirability of including individual-level variables in an analysis to detect social transmission from diffusion data. The analyses establish that the inclusion of individual-level variables both increases statistical power and reduces Type I error rates. In addition, the sensitivity of the diffusion analyses to network structure prompts us to recommend that researchers use methods that can generate confidence intervals for the strength of social
Acknowledgements
W.H. was supported by a BBSRC grant (BB/D015812/1), N.B. by a McGill Milton Leong Fellowship and K.N.L. by grants from the BBSRC (BB/C005430/1 and BB/D015812/1) and an ERC Advanced Grant (EVOCULTURE, Ref. 232823). We would like to thank the members of the Laland Lab Journal Club for useful comments on an earlier draft, and to Laurel Fogarty, Tess Hanrahan and Joel Higgin for ‘test-driving’ the R code.
References (43)
- et al.
The origin and spread of innovations in starlings
Anim. Behav.
(2008) - et al.
On the relation between social dynamics and social learning
Anim. Behav.
(1995) - et al.
Measuring and testing the steepness of dominance hierarchies
Anim. Behav.
(2006) - et al.
A new model system for studying behavioral traditions in animals
Anim. Behav.
(1995) - et al.
Scrounging prevents cultural transmission of food-finding behavior in pigeons
Anim. Behav.
(1987) - et al.
Social processes influencing learning in animals: a review of the evidence
Adv. Study Behav.
(2008) - et al.
The animal cultures debate
Trends Ecol. Evol.
(2006) The opening of milk bottles by birds—evidence for accelerating learning rates, but against the wave-of-advance model of cultural transmission
Behav. Process.
(1995)- et al.
Predicting epidemics on directed contact networks
J. Theoret. Biol.
(2006) - et al.
Association patterns and foraging behaviour in natural and artificial guppy shoals
Anim. Behav.
(2008)
Model Selection and Multimodel Inference: A Practical Information-theoretic Approach
Statistical Computing: An Introduction to Data Analysis Using S-plus
The opening of milk bottles by birds
Br. Birds
Network-based diffusion analysis: a new method for detecting social learning
Proc. R. Soc. B-Biol. Sci.
Social-learning in animals: categories and mechanisms
Biol. Rev.
The effects of local spatial structure on epidemiological invasions
Proc. R. Soc. B-Biol. Sci.
Identifying social learning in animal populations: a new ‘option-bias’ method
PLoS ONE
Statistical Analysis of Network Data: Methods and Models
Cultural transmission of tool use in bottlenose dolphins
Proc. Natl. Acad. Sci. USA
Social learning strategies
Learn. Behav.
Cited by (114)
ASN: A method of optimality for seed identification in the influence diffusion process
2023, Physica A: Statistical Mechanics and its ApplicationsThe unique potential of field research to understand primate social learning and cognition
2022, Current Opinion in Behavioral SciencesDo honey bees modulate dance following according to foraging distance?
2022, Animal BehaviourCitation Excerpt :To better account for model selection uncertainty, we employed an information-theoretic approach, allowing us to incorporate information from multiple candidate models when forming our inferences (Burnham & Anderson, 2002). In an NBDA, the strength of social transmission per unit of network connection (e.g. per waggle run followed), relative to the rate of feeder discovery through individual exploration, is estimated by the social transmission parameter, s (Franz & Nunn, 2009; Hoppitt et al., 2010). In other words, s estimates the increased likelihood of successfully locating the target feeder for each waggle run that is followed by a potential recruit (or for each second of dance following, depending on the model), where higher values of s correspond to stronger social influences.