CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Spatial modelling of fish counts in stream networks: Convolution or multilevel approaches Authors:  Joao Pereira - Universidade Federal do Rio de Janeiro (Brazil) [presenting]
Marco Rodriguez - Universite du Quebec a Trois-Rivieres (Canada)
Alexandra Schmidt - McGill University (Canada)
Julie Deschenes - Ministère des Forêts, de la Faune et des Parcs (Canada)
Abstract: In river networks, spatial correlation in species abundance arises from ecological processes operating at multiple spatial scales, including (1) directional aquatic processes constrained by waterflow, such as fish movement and transport of nutrients; (2) terrestrial processes that influence watershed characteristics at various spatial scales. A Poisson-lognormal mixture model of fish counts in a river network is considered. The mixing component accounts for temporal and spatial effects. Two different approaches for the spatial latent effects are explored: one based on a moving-average (convolution) construction, the other on a multilevel structure. The convolution approach can incorporate both hydrological (``as the fish swims'') distance, to capture directional effects, and Euclidean (``as the crow flies'') distance to capture effects of terrestrial processes. In contrast, the multilevel approach captures habitat effects nested at three discrete spatial scales (section, reach, and branch). Covariates were used to account for local environmental effects. Bayesian inference is performed for a set of models representing different combinations of temporal and spatial effects. Models are compared by means of different information criteria, exam of the variance associated with each of the latent components, and prior-posterior comparison to clarify the relative contribution of different processes to variation in fish counts.