Institute of Statistics and Decision Sciences
Duke University
presents:
Richard L. Smith
National Institute of Statistical Sciences
University of North Carolina
and Cambridge University
"Temporal and Spatial Correlations in Climatic Data"
Abstract: I will discuss two methodological problems, both of which have arisen in the analysis of climatic time series. The first of these concerns long-memory dependence. Many climatic time series are believed to have correlations which decay to 0 much slower than the correlations of standard models such as ARMA models. This is turn affects the way we assess the statistical significance of observed trends in the data. Geweke and Porter-Hudak (1983) proposed a spectral-based method of estimating the rate of long-range dependence, but their method raised some technical issues which have only recently been resolved (see in particular a paper by P.M. Robinson in the latest issue of Annals of Statistics, June 1995). In this talk, I shall discuss some of the background of this debate and consider its implications for answering more general questions, including the possibility of Bayesian approaches.
The second part of the talk is concerned with spatial statistics. Much of the traditional "kriging" methodology assumes a stationary isotropic random field but such an assumption hardly seems appropriate in many environmental and climatological applications. A novel approach to nonstationary series was proposed by Sampson and Guttorp (JASA, 1992), in which they attempted to find a transformation of the data points to improve the fit to a stationary distribution. However their fitting method employs a number of seemingly arbitrary steps. I shall discuss alternative likelihood-based methods of estimation, which can provide a more powerful and flexible way of fitting their models and also point the way towards a fully Bayesian approach.
Friday, September 15, 1995
11:45 - 12:45
116 Old Chemistry Building Any questions concerning the seminar may be addressed to Cheryl McGhee @ (919) 684-8029, e-mail cheryl@isds.duke.edu, or finger seminar@isds.duke.edu.