next up previous index
Next: asa.stat.environ.03 Up: ASA Statistics and the Previous: asa.stat.environ.01



also sponsored by the Statistical Ecology Section of the Ecological Society of America

Session Slot: 2:00- 3:50 Wednesday

Estimated Audience Size: 40-50

AudioVisual Request: overhead projector

Session Title: Hierarchical Models in Ecology

Theme Session: No

Applied Session: No

Session Organizer: Edwards, Don University of South Carolina





Session Timing: 110 minutes total (Sorry about format):

Opening Remarks by Chair - 0 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussion lead by chair - 10 minutes Floor Discussion - 10 minutes

Session Chair: Dixon, Philip Savannah River Ecology Laboratory

Address: University of Georgia, Savannah River Ecology Lab, P.O. Box E, Aiken, SC 29802-0005

Phone: 803-725-2472

Fax: 803-725-3309


1. Applications of Conditionally Specified Models in Ecology

Kaiser, Mark S.,   Iowa State University

Address: Department of Statistics, Iowa State University, Snedecor Hall, Ames, IA 50011-1210

Phone: 515-294-8871

Fax: 515-294-4040


Cressie, Noel, Iowa State University

Lee, J., Iowa State University

Lewin, N., Iowa State University

Daniels, M., Iowa State University

Abstract: Conditionally specified statistical models show promise for applications involving random variables that exhibit complex dependence structures. Such models are formulated by specifying a Markov random field on a lattice, the `locations' of which may be physical (i.e., spatial lattices) or relational (i.e., longitudinal). A probability density or mass function is specified for each of the locations, conditional on all other locations in the field. The goal of statistical analysis is then to identify the joint distribution and use it for estimation and inference about pertinent parameters. In this talk, conditionally specified models will be applied to several problems in ecology. At least one of these applications will involve dependence arising from spatial sources, and at least one will involve dependence arising from multiple observations being taken on the same sampling units. Dependence in the models may be located at either the level of the data model or in a mixing distribution for random data model parameters.

2. Hierarchical Models for Analyzing Spatial Ecological Data

Ver Hoef, Jay M.,   Alaska Department of Fish and Game

Address: Alaska Department of Fish and Game, 1300 College Road, Fairbanks, Alaska 99701

Phone: 907-459-7278

Fax: 907-452-6410


Abstract: Consider a spatial lattice of locations where data have been observed, and the data are assumed to be a realization of a hierarchical model. The hidden process is assumed to have a (prior) distribution from a two-state (e.g., high or low) Markov model. The two states form patches in the hidden process. However, the patches are hidden by random noise. Conditional on the states, the observations may be modeled as coming from any distribution in the regular exponential class. An objective of some ecological studies is to uncover the patches in the hidden process, and make some further inference, such as computing a landscape metric, on the patch structure. Sampling from the posterior distribution of the hidden process allows inference on the patch structure and the mean and variance of any landscape metric computed from the sample. In this paper I consider one and two dimensional lattices and use parametric empirical Bayes and fully hierarchical Bayes models to sample from the posterior distribution of the hidden process and make inference on metrics from landscape ecology.

3. Mixed-Model Splines for Ecological Time Series

Street, IV, W. Scott,   Georgia Southern University

Address: Department of Math. And Computer Science, Georgia Southern University, P.O. Box 8093, Statesboro, GA 30460-8093

Phone: 912-681-5888

Fax: 912-681-0654


Edwards, Don, University of South Carolina

Abstract: Ten-year biweekly time series of abundance of benthic fauna in a South Carolina salt marsh reveal, in the log scale, a very clear annual pattern of alternating near-linear growth and decline. The ecological processes governing these dynamics, particularly the dramatic year-to-year differences in annual maxima and minima, are mysterious. Traditional time series analyses (Fourier and ARIMA methods) can (in some cases) adequately mimic the patterns in these series, but they do little to enhance ecological understanding because they are not meaningfully parameterized. We model these series using floating-knot linear splines in two ways: (1) as high-dimensional nonlinear regression models with no forced similarity from year to year; and (2) as a hierarchical spline model placing a probability distribution on annual maxima, minima, growth rate, and decay rate across years, thereby enforcing some similarity in seasonal patterns. We discuss fitting and inference under both approaches; both yield excellent fits with ecologically interpretable results which can be used to shed light on governing processes. Results of a simulation study show that the hierarchical approach can yield great improvements in efficiency over the nonlinear regression approach, particularly for series with low interannual variability.

List of speakers who are nonmembers: 0

next up previous index
Next: asa.stat.environ.03 Up: ASA Statistics and the Previous: asa.stat.environ.01
David Scott