Background/Question/Methods
Individual animal movement data is collected at an increasing rate as remote-sensing technology develops.  At its best, analysis of the data can suggest mechanisms by which organisms exploit a heterogeneous and variable environment. Unfortunately, movement data are multidimensional, non-independent and almost always suffer from the twin banes of measurement error and gappiness, making appropriate analyses far from straightforward. While error in the data can be accounted for by state-space models, an increasingly popular statistical approach, gappiness (i.e. irregular intervals between times of measurement) has not been well-addressed in the literature. Gappiness is particularly widespread in data on marine organisms, for which remote sensing depends on a combination of transmitter exposure and satellite presence. In this talk, I suggest a method of dealing with gappiness by identifying a persistence component of movement that can be modelled as a continuous auto-correlated stochastic process with three parameters: a mean, a variance and a continuous auto-correlation coefficient. I then develop several methods for identifying changes between behavioral modes. A single breakpoint can be found with a maximum likelihood estimation over an entire gappy time-series, while multiple breakpoints can be found either by sweeping a window or by optimization routines such as simulated annealing. One unique feature of the method is that the values of the parameters themselves are less important than the location of the switches, making the framework robust against measurement error. After exploring the effectiveness of the method with simulation, these methods are applied to data on several marine species, including Magellanic penguins (Spheniscus magellanicus), northern fur seals (Callorhinus ursinus) and dugongs (Dugon dugon). Finally, I suggest methods of interpreting the times and locations of behavioral switches in terms of environmental and animal state covariates.
Results/Conclusions
Simulations indicate that the methods are extremely robust to very gappy and error-ridden data, correctly identifying times at which behavioral changes occur. Highly correlated data implies directed movement, while uncorrelated movement suggest searching or active foraging. The organisms display nuanced behaviors, including both high and low speed uncorrelated movements, suggesting separate searching and foraging modes, fast directed movements associated with return trips to breeding rookeries, and slower highly correlated movements. Suggested methods of interpreting the outputs of the models in terms of environmental and internal state covariates can lead to productive future analyses.