Statistical Methodology

Compositional analysis of behavioural data

Modern technologies measuring animal location, movement and metabolism enable objective and accurate monitoring of animal activities and health, and calculating daily activity budgets from such data can provide important insights. However, these studies are hampered by compositional constraints: times allocated for individual behaviours are non-negative and total time is pre-determined, e.g. 24 hours or other fixed duration depending on the measurement protocol. These compositional constraints induce a co-dependence structure between time allocations to different behaviours, leading to inconsistent and paradoxical results when using methods of data analysis developed for unconstrained data.

Compositional data analysis (CoDA) has been developed to provide a sound mathematical framework that directly addresses the constrained and relative nature of compositional data, motivated primarily by the analysis of concentration data occurring in geochemistry and environmental sciences. By adapting and extending CoDA methods to address the particular issues encountered in animal behaviour and physical activity data, we have been helping to improve inferences about relationships involving behavioural patterns, either considering these patterns as a multivariate response or constituting multiple explanatory variables, enabling synergistic analyses involving all behaviours as a whole rather than considering each in isolation.

CoDA triangle Representation of the differential behaviour patterns of pigs observed in social groups (red; MIX) or in their home pens (blue; CON) according to relative amounts of three postural behaviour types (standing, lying laterally and sitting). Principal components (PC) in compositional space are curved in the ternary diagram, the distributions of scores of the first PC (red) differing between types of animals. Grey lines indicate points with constant proportions of one variable. The second PC is shown in yellow.

Further details from:
Javier Palarea

Article date 2015

Research

Statistical Genomics and Bioinformatics

Process and Systems Modelling

Statistical Methodology

PhD Opportunities

Meetings & Seminars