Skip to main content

A regression Paradox for linear models: sufficient conditions and geometric interpretation

01 January 2009

New Image

An analysis of customer survey data using direct and reverse regression leads to apparently inconsistent conclusions. The phenomenon is known as the regression paradox. We describe sufficient conditions when the regression paradox will appear both analytically and geometrically. In linear models, it is a result of a distribution shift among groups relative to predictability in regression and reverse regression.