Significance Regression: Improved Estimation from Collinear Data for the Measurement Error Model

Abstract

This paper examines improved regression methods for the linear multivariable measurement error model (MEM) when the data suffers from "collinearity." The difficulty collinearity presents for reliable estinlation is discussed and a systematic procedure, significance regression (SR-MEM), is developed to address collinearity. In addition to mitigating collinearity difficulties SR-MEM produces asymptotically unbiased estimates. The use of ordinary least squares (OLS) for the MEM is examined. For collinear data OLS can improve the mean squared error of estimation over the maximum likelihood (ML) unbiased estimator in a manner analogous to ridge regression (RR). The significance regression method developed for the classical model (SR-classical) can also be used for data with measurement errors. SR-classical is similar SR-MEM and can yield better estimation than the ML estimator for collinear data. Numerical examples illustrate several points.