Identifying Latent Behavioral States From Animal Biotelemetry Data Using Non-parametric Bayesian Methods

Photo by Russ on Wikimedia Commons

Movement decisions of an organism can impact habitat use, foraging strategies, reproductive success, and survival, which ultimately affects population dynamics of a species. The study of animal movement has increased in the past few decades, along with the increased capacity of telemetry devices to record locations with greater accuracy and over longer durations. Although there are a wide variety of existing methods that estimate latent behavioral states, many of these methods have limitations for their use and inference (e.g., only analyze one data stream, require parametric distributions).

We have developed two non-parametric Bayesian models that addresses some of the limitations of these existing methods for the estimation of animal behavioral states. One of these models uses a two-stage approach: 1) Multiple data streams (e.g., step length, turning angle, temperature, depth) are partitioned into track segments for each ID; 2) Track segments are pooled across IDs and clustered to determine the optimal number of states and their distributions per data stream. This proposed framework provides a fast and flexible method to characterize latent states at the level of track segments, which may be particularly useful when time steps are short (and therefore behavior is highly autocorrelated) and when distributions of data streams are not well fit by parametric distributions. Additionally, the other method uses a mixture model to infer observation-level behavioral state estimates using the same type of non-parametric clustering approach. Both of these models can be applied using functions included within the bayesmove package in R.

The two-stage segment-level model is described further in a pre-print at bioRxiv while the observation-level model is described further in an article that is ‘in press’ at Ecological Applications.

Josh Cullen, PhD
Josh Cullen, PhD
NSF Postdoctoral Research Fellow

My research interests include animal movement ecology, Bayesian modeling, and R stats.

Related