UW Aquatic & Fishery Sciences Quantitative Seminar

Paul Conn

National Marine Mammal Laboratory, Research Statistician

Hierarchical models for multiple-observer animal transect surveys.

Abstract

Traditional, design-based inference for transect data requires strict adherence to a predetermined sampling design. By contrast, model-based frameworks are more flexible and allow one to express animal density as a function of habitat covariates and to model density as a function of space and/or time. In this talk, I describe a model-based framework for estimating abundance from line transect data that incorporates a number of desirable features, including spatial autocorrelation, estimation of density functions for individual covariates, and the ability to model data from multiple observers. Under this framework, abundance at the transect level is conceptualized as having arisen from an overdispersed Poisson process subject to spatially correlated error. Multiple covariates can be specified for abundance intensities, as well as for the underlying detection model. At the local level, abundance on each transect is modeled via data augmentation, so that detection probabilities can be made functions of latent individual covariates such as group size, distance, or species. Observer dependence can also be incorporated by modeling multiple observer data using a multivariate normal distribution on the probit scale. I demonstrate the approach with simulated data, as well on a population of golf tees where true abundance is known. Our approach shows promise, but care must be taken to avoid spatial confounding and overparameterization in the spatial component of the process model.