Validating HMM-derived behavioral states against field observations in wood bison (Bison bison athabascae): cautionary evidence from tracking data

Hidden Markov models (HMMs) are an important tool for isolating behavior using animal location data. These behaviors and the conclusions drawn from the models can have important management implications when different places visited by an animal are associated with different behaviors. One common application of HMMs is to classify the movements of tracked individuals into distinct behaviors. However, potential methodological pitfalls could limit the reliability of these behavioral inferences from HMMs. Here, we used telemetry data from wood bison ( Bison bison athabascae ) with 90-minute sampling intervals collected in northern Alberta, Canada, to compare HMM outputs to field observations. Tracking data from global positioning system (GPS) collars on 47 bison over four years were used to inform in-situ behavioral classifications based on field observations of tracks and signs left by the bison. These classifications, when coupled with GPS tracks, enabled a direct comparison between the “true” behavior observed at a given location and the patterns identified by HMMs. We found HMMs to be no better than random guessing at classifying bison activity. This had implications for behavior-specific habitat selection analyses that partitioned locations into presumed behaviors. The indistinct nature of bison behaviors, as observed in tracking data with 90-minute sampling intervals, likely explains why behavior classification by HMMs was significantly lower than for many aquatic or aerial species. Our work corroborates findings from existing studies on similar animals sampled at similar frequencies and raises important concerns about how HMMs should be applied to animal movement data.