Department of Statistical Science
Duke University
presents:
Victor De Oliveira
oliveira@niss.rti.org
NISS - Research Triangle Park, NC
"Bayesian Prediction of Clipped Gaussian Random Fields"
Abstract: This work provides a framework to perform prediction/interpolation in some types of binary random fields. Specifically, we assume that the binary random field Z(.) is obtained by clipping a Gaussian random field at a fixed level.
The model, using a Bayesian approach, is used to map a binary variable over a bounded region D of the plane: To each point s in D, we compute the optimal predictor of Z(s), 0 or 1, given a sample Z = (Z(s_1),...,Z(s_n))' from a single realization of the random field, and also provide measures of prediction uncertainty, amenable for binary data.
The optimal predictor and the uncertainty measure depend on P{ Z(s) = 1 | Z }, which is computed `exactly', through `data augmentation', using Markov Chain Monte Carlo methods.
The prediction ability of the proposed model is tested using a simulated binary image, and comparisons are made between this model, nearest-neighbors rules, and simple indicator kriging. Model diagnostics and cross-validation methods for binary data are explored to assess the adequacy of the model for predictive purposes.
October 31, 1997
3:30 pm - 4:30 pm
116 Old Chem Building Any questions concerning the seminar may be addressed to Cheryl McGhee @ [919] 684-8029 or e-mail cheryl@stat.duke.edu. Please contact the author(s) directly for reprints etc.