Student Seminar Series - November 23, 2005
University of Minnesota
School of Statistics
College of Liberal Arts
Effects
of Box-Cox Transformation on Prediction in Linear Regression
Lin Fan
Wednesday, November 23, 2005
3:30 PM, B60
Ford Hall
Minneapolis, East Bank Campus
Refreshments at 3:15 PM
300 Ford Hall
Abstract
Box
and Cox proposed a parametric family of power transformations of the
data to reduce problems such as non-normality, based on the assumption
that a power transformation of the responses have normal distributions.
However, when prediction is the goal, it is not completely clear if the
method yields a better prediction accuracy, especially when no power
transformation leads to normality.
In this work we numerically compare the prediction performances of
three regression methods: least squares without transformation, Box-Cox
transformation, and another method. Our numerical results show that
although Box-Cox transformation helps improve the normality, it may
increase the mean prediction errors sometimes.