Investigating the Impact of Observation Errors on the Statistical Performance of Network-Based Diffusion Analysis

Article excerpt

Experiments in captivity have provided evidence for social learning, but it remains challenging to demonstrate social learning in the wild. Recently, we developed network-based diffusion analysis (NBDA; 2009) as a new approach to inferring social learning from observational data. NBDA fits alternative models of asocial and social learning to the diffusion of a behavior through time, where the potential for social learning is related to a social network. Here, we investigate the performance of NBDA in relation to variation in group size, network heterogeneity, observer sampling errors, and duration of trait diffusion. We find that observation errors, when severe enough, can lead to increased Type I error rates in detecting social learning. However, elevated Type I error rates can be prevented by coding the observed times of trait acquisition into larger time units. Collectively, our results provide further guidance to applying NBDA and demonstrate that the method is more robust to sampling error than initially expected. Supplemental materials for this article may be downloaded from http://lb.psychonomic-journals.org/content/supplemental.

(ProQuest: ... denotes formula omitted.)

In many animal species, individuals learn socially by observing the behavior of other individuals. Sophisticated experiments on captive animals have identified different learning mechanisms in animals (Galef & Giraldeau, 2001; Hoppitt & Laland, 2008), but inferring the existence of social learning in wild animals remains an important and challenging task. Abandonment of the highly controlled experimental settings of captive studies is necessary for investigating how social and ecological conditions in wild animals affect social learning dynamics and the emergence of traditions and, more generally, for understanding the evolution of social learning and culture. However, the lack of experimental control also introduces new methodological problems. Important methods of inferring social learning in wild animals, such as the ethnographic method (Perry & Manson, 2003; Rendell & Whitehead, 2001; van Schaik et al., 2003; Whiten et al., 1999) and diffusion curve analysis (Reader, 2004), can have low statistical power to detect social learning and can produce a high rate of false positives (Franz & Nunn, 2009; Galef, 2004; Laland & Galef, 2009; Laland & Hoppitt, 2003; Laland & Janik, 2006; Laland & Kendal, 2003; Reader, 2004). Thus, new methods are needed for investigating social learning.

Recently, we developed network-based diffusion analysis (NBDA; Franz & Nunn, 2009) to overcome the limitations of previous approaches. NBDA makes the reasonable assumption that social learning is more likely to take place among conspecifics that are relatively more closely linked in a social network (Coussi-Korbel & Fragaszy, 1995). In the case of a food processing technique that is transmitted socially, for example, we can expect food-related behaviors to spread most quickly among individuals that often feed together and, thus, have strong connections in a cofeeding network (for an example, see Figure 1). In the case of social learning-but not in asocial learning-we expect the structure of a social network (i.e., the strength of connections among individuals) to influence how a novel behavior spreads through a group of animals. NBDA aims to identify whether such an influence occurs in the observed diffusion of a novel behavior. For this purpose, alternative agent-based models (ABMs) of social and asocial learning are fitted to the observed diffusion of a novel behavior. These models provide a way to estimate the probabilities with which each individual learns through social or asocial learning during the different stages of a diffusion. By comparing these probabilities to the actual learning events, one can assess which learning mechanism was most likely to have caused the observed diffusion.

The statistical analysis of NBDA is based on maximum likelihood model fitting. …