On the geometry of measurement error models

Abstract

The problem of undertaking inference in the classical linear model when the covariates have been measured with error is investigated from a geometric point of view. Under the assumption that the measurement error is small, relative to the total variation in the data, a new model is proposed which has good inferential properties. An inference technique which exploits the geometric structure is shown to be computationally simple, efficient and robust to measurement error. The method proposed is illustrated by simulation studies.