A method for exploring the structure of behavioural syndromes to allow formal comparison within and between data sets

Research on behavioural syndromes (consistent individual differences in suites of correlated behaviours) requires formal statistical methods to describe and compare syndrome structures. We detail the shortcomings of current methods aimed at describing variation in behavioural syndromes, such as multiple pairwise correlations and principal components analysis (PCA). In their place we propose an alternative statistical framework involving: (1) calculation of trait variance–covariance and correlation matrices within each data set; (2) statistical evaluation of specific hypotheses regarding how behaviours covary within a behavioural syndrome; and (3) statistical comparison of behavioural covariances across data sets using structural equation modelling (SEM). Given their unfamiliarity to most behavioural ecologists, we illustrate these methods using an already published data set for two groups of populations of three-spined stickleback, Gasterosteus aculeatus, living in ponds with and without fish predators. Previous analyses suggested a lack of behavioural syndrome structure for stickleback that lived in the absence of fish predators. However, by evaluating a priori hypotheses of how behaviours might covary using SEM, we were able to demonstrate that the two types of populations differed specifically in covariance patterns for aggression, exploration of novel food sources and altered environments, but not for exploration of novel environments and activity. Such detailed inferences cannot readily be made based on conventional statistical approaches alone, and so the methods we outline here should become standard in studies concerning the evolution of behavioural syndromes within and between populations.