A hierarchical N-mixture model to estimate shifts in animal behavior

Alison Ke and Rahel Sollmann, Wildlife, Fish, and Conservation Biology, University of California, Davis, Davis, CA, Luke Frishkoff, Department of Biology, University of Texas at Arlington, Arlington, TX, Daniel Karp, Wildlife, Fish, and Conservation Biology, University of California, Davis, CA

Background/Question/Methods
Behavioral changes can serve as early warning tools for conservation by reflecting how animals obtain resources from different habitats. Unfortunately, many behaviors are difficult to observe, and detectability of an individual often depends on the behavior it is performing. For example, a foraging bird may be easier to detect in an open field versus a forest, whereas a vocalizing bird may be equally detectable between habitats. N-mixture models are regularly used to estimate species abundances from temporally replicated counts while accounting for detection probability. To date, however, no one has developed parallel procedures for accounting for detection when modeling changes in behavior across habitat types. Here, we developed a variation of a binomial N-mixture model to estimate shifts in behavior among different environments while accounting for behavior-specific imperfect detection probability. We then simulated data and explored model performance under a broad array of scenarios; for example, varying species abundances, detection probabilities, and violation of the closure assumption of N-mixture models. We compared results from our model to a naïve multinomial regression model that did not account for variation in detection. Finally, we applied the behavior N-mixture model to a 4-year dataset from Costa Rica with 25 sites, in which bird behaviors such as foraging and vocalizing were quantified during repeated point count censuses in tropical forests and agriculture.
Results/Conclusions
Through simulation, we found that the behavior N-mixture model generally produced unbiased estimates of behavior probabilities and their relationships with predictor variables. Critically, the behavior N-mixture model was much better at characterizing the uncertainty around estimates than the naïve model. For example, the naïve model regularly estimated significant effects of covariates on behaviors in the wrong direction. This did not occur in the behavior N-mixture model. Moreover, applying the behavior N-mixture model to field data suggests the model can provide useful information on how birds utilize different habitats. For the Turquoise-browed Motmot and Hoffmann’s Woodpecker, the behavior N-mixture model estimated that there were higher densities of individuals in more forested habitats, but individuals had similar probabilities of performing eating and vocalizing behaviors in both forest and in agriculture. In contrast, the Inca Dove was estimated to eat more and vocalize less in agriculture than in forest, but there was no effect of forest cover on total abundance. Thus, the behavior N-mixture model can increase ecological understanding, for example, by identifying where animals obtain resources and where potential ecological traps may be.