Nonparametric regression by property matching of kernel density estimates

Seminário de Estatística e Gestão do Risco 

Orador: Theodor Loots, University of Pretoria 

Data: 25 outubro 2017 (4ª feira)

Hora: 15h30

Local: Sala 1.4, ed. VII, FCT NOVA 

Abstract: The coefficients obtained from using ordinary linear regression may be severely biased with corresponding estimates lacking accuracy when the assumptions of normality are not satisfied.  A new framework is proposed where distributional properties of the kernel density functions cast over the dependent and independent variables are matched in order to yield coefficient estimates.  Specifically, the moments and arc lengths of these functions are matched, and applied to a problem in biology.  The significance of these estimates are evaluated using resampling techniques, and model selection performed by using an entropy based measure, namely, the Bhattacharyya divergence measure.  It is shown that the percentage variation explained, resulting from ordinary least squares, can be improved upon, without forfeiting the elegance of the linear model. 

Financiado através do projeto UID/MAT/00297/2013.