Title: Linear projections for kurtosis removal
Authors: Nicola Loperfido - University of Urbino (Italy) [presenting]
Abstract: The performance of several statistical procedures might depend on the kurtosis of the sampled distribution. Examples include, but are not limited to, inference on mean vectors and covariance matrices. Componentwise, nonlinear transformations might not adequately address the problem. We propose to remove kurtosis by means of linear projections which are solutions of generalized eigenvalue problem involving two kurtosis matrices. The method is particularly useful for either hidden truncation models or finite mixture models. Its practical usefulness is illustrates with the RANDU dataset.