Organizer:

Richard Davis, rdavis@stat.colostate.edu

Speakers

3:45 p.m.
Computational Aspects of Modeling the Spatial and Temporal Distribution of Cloud Cover
Barbara Bailey, Department of Statistics, University of Illinois
Douglas W. Nychka
Keywords: neural network, nearest-neighbor model, satellite data

Clouds play a fundamental role in controlling the amount of solar and infrared radiation available to the climate system and therefore prediction of cloud cover amount is of great importance to climate modeling. Most clouds are smaller in area than a typical grid resolution of climate models so the modeling objective is to develop a statistical model for the subgrid-scale spatial and temporal distribution of cloud cover by linking large scale climate model variables.

The large amount of data available to fit the model includes three months of hourly infrared data obtained by a satellite from the TOGA COARE experiment on a 444 by 888 pixel resolution satellite image and large scale climate variables on a slightly lower resolution.

A statistical model for the temporal distribution of cloud cover with an implicit spatial structure that links large scale variables such as relative humidity with cloud cover will be presented. The model is a first-order lagged nearest-neighbor model. The computational aspects of fitting a neural network model to describe the nonlinear relationship between grid cell neighbors over time will be discussed.

4:15 p.m.
An Improved Model for Spatially Correlated Binary Responses
Jennifer Hoeting, Department of Statistics, Colorado State University
Molly Leecaster, S. California Coastal Water Research Project
David Bowden, Colorado State University
Keywords: Autologistic model, Bayesian estimation, Gibbs sampling, Markov random field

We use covariates and an indication of sampling effort in an autologistic model to improve the prediction of probability of presence for lattice data. The model is applied to sampled data where only a small proportion of the available sites have been observed. We adopt a Bayesian set-up and develop a Gibbs sampling estimation procedure. We show that the autologistic model with covariates improves predictions as compared to the simple logistic regression model and the basic autologistic model (without covariates). Software to implement the methodology is available at no cost from StatLib.

4:45 p.m.
Inference for point processes based on many short realizations
Michael Stein, Department of Statistics, University of Chicago

Suppose a stationary point process is observed over a large number of small and well-separated regions. In this situation, a large fraction of the observations are near a boundary of the observation region, so that edge effects play a prominent role in the properties of estimators of spatial dependence of the process. One popular descriptor of spatial dependence for point processes is the reduced second moment function, which is a normalized measure of the expected number of other events of a process within some distance of a given event of the process. This work describes two estimators of the reduced second moment function that I have recommended in recent work. Based on these estimators, define a new estimator that takes the value of one of these estimators if the sign of a certain first-order ancillary statistic is negative and takes the value of the other estimator if this ancillary is positive. Simulations show that for a broad range of point processes in one dimension, the new estimator is superior to either of the existing estimators, particularly when the observation window is made up of intervals of substantially varying length.