Bayesian modelling of locust behaviour using BAYSIG

A fundamental challenge in behavioural analyses is the gap between using simple statistical tests on selected features, and understanding the internal processes that underlie and define the manifest behaviour. Bayesian statistical analysis can provide a powerful way to model data for behavioural studies. Bayesian approaches provide several advantages over classical statistical methods. One key advantage is that Bayesian inference via Markov Chain Monte Carlo (MCMC) methods allows us to specify models that approximate the behavioural process which generated the raw observed data. Rather than explicitly extracting selected summary statistics from the data that serve as proxies for the behavioural traits of interest, these Bayesian methods permit to directly estimate, from the raw data, the parameters for variables that represent the underlying behavioural traits and explain the observed behaviour. Another advantage is that they allow us to compute not only point estimates but also the distributions of the model parameters. We are using BAYSIG (http://bayeshive.com/), a concise mathematical programming language for statistical analysis developed by OpenBrain (http://openbrain.co.uk/). BAYSIG is especially efficient for building, inference and verification of dynamic models represented either by stochastic differential equations or difference equations. We are in the process of applying Bayesian models to the analysis of desert locust (Schistocerca gregaria) behaviour. In a classical assay setting a locust is released into arena where its movement trajectory is recorded. A key aim of the study is to characterise, using Bayesian modelling, hidden behavioural ‘states’ that drive the observed differences in movement.