Predicting multiple behaviors from GPS radiocollar cluster data

Advancements in GPS radiotelemetry allow collection of vast data for a variety of species including those for which direct observations are typically not feasible. Predicting behavior from telemetry data is possible, but telemetry fix rate can influence inferences, and animal behavior itself can affect fix success. We use multinomial regression to predict behavior from GPS radiocollar data field validated with behavioral state information. Our study organism was a facultative carnivore, the grizzly bear (Ursus arctos) (n = 10) from a threatened population in Alberta, Canada, monitored during 2008–2010. Models using GPS cluster parameters alone successfully predicted ungulate consumption, whereas bear bedding was sufficiently identified by models that included site-level information. Predicting more complex behaviors required models incorporating both cluster parameters and habitat characteristics. No model reliably predicted vegetation feeding, probably because this activity is shorter than the time required for cluster formation. Models built using infrequent fix rates underestimated all behaviors, with bear presence at ungulate carcass sites least sensitive to fix rate variability. Behavior influenced fix success, with highest fix acquisition occurring when bears fed on vegetation. Placing predictions into a conservation context, we show that grizzly bears modify their behavior as they move through a landscape with complex human-activity patterns on reclaimed open-pit mines, foothill, and mountain regions. The modeling approach we tested needs further applications across species and ecosystems including behavioral monitoring, quantifying activity budgeting, and identifying areas/habitats important for specific behaviors that may warrant conservation.