Wednesday, August 5, 2009 - 1:30 PM

COS 75-1: Using Bayesian multilevel models to detect social structure in presence-absence data

Eric J. Ward, Brice Semmens, and E. E. Holmes. Northwest Fisheries Science Center

Background/Question/Methods
Identifying levels of social structure that shape population demography is important for increasing realism in population biology, management and conservation.  Recent advances in hierarchical modeling have increased the flexibility of ecological models, allowing for individual and multiple levels of group heterogeneity to be included in population dynamics.  From a biological perspective, individuals clustered by relatedness, geography, or time may experience similar levels of risk, translating into differential survival between groups.  From a statistical perspective, ecological measurements are generally taken on the same social clusters and individuals repeatedly.  Therefore, considering the lack of independence between measurements and the variability across multiple scales is imperative.
Killer whales are a highly social species that engage in prey sharing and cooperative foraging.  In the Northeast Pacific Ocean, killer whale populations may be sub-divided into multiple levels of social organization.  Using Bayesian hierarchical models with multiple levels of variation, we illustrate how data support for different levels of clustering may be evaluated.  While our application focuses on the analysis of presence-absence data, which is traditionally used to estimate detectability, movement, survival, and species occurrence, these methods are applicable to many types of ecological data.
Results/Conclusions
Our models of killer whale social structure include multiple levels of group (but not individual variation).  When applied to an endangered population of fish-eating killer whales, the model most supported by the data includes variation at the larger group level (but not at the family cluster).  We illustrate how two model selection tools (Bayesian posterior probability, DIC) may be utilized to evaluate data support; DIC favors a slightly more complex model compared to the posterior probability.  We believe the variation in survival between groups is driven by demographic stochasticity and time-varying external factors, such as prey availability.  Using historical time series of these covariates, we illustrate the implications of using our methods for management and conservation.  As expected, the survival model with social clustering most supported by the data produces the least uncertainty surrounding future killer whale population sizes when the population is projected forward 10 years (relative to a model with no social organization, or all potential levels of social organization).