Event Details

Professor Dag Tjostheim, University of Bergen, will be giving a STOR-i seminar as part of the International Visitor Programme visit on Friday 21st February 12.00-1.00pm in the PSC Lecture Theatre

Title: Local Gaussian approximation and nonlinear statistical modelling

Abstract: The multivariate Gaussian distribution has a number of very desirable properties. It has been applied with great success to linear modelling, but quite often it fails in the presence of nonlinearity. I will consider general modelling (linear and nonlinear) using local Gaussian approximation, where a multivariate distribution is approximated at each point by a Gaussian distribution. In this way the given distribution is approximated by a family of Gaussian distributions. The exceptional properties of the Gaussian can then still be utilized, but locally. This turns out to be quite a powerful approach that can be applied to a number of problems: to measuring dependence and tests of independence and conditional independence, using local Gaussian correlation and partial correlation; to estimation of densities and conditional densities; and to time series, using local Gaussian autocorrelation and spectral density. Features missed by a global Gaussian analysis may be picked up by a local analysis, as will be demonstrated. Finally, I discuss some restrictions and possible extensions.