Iteratively reweighted partial least squares: A performance analysis by monte carlo simulation

A robust implementation of partial least squares (PLS) is developed in which the method of iteratively reweighted least squares is adapted for use with PLS. The result is a PLS algorithm which is robust to outliers and is easy to implement. Examples and case studies are presented, followed by two Monte Carlo studies designed to explore the behavior of the method.

The paper begins with the motivation and intended applications for the procedure. A discussion is given of the method of interatively reweighted least squares (IRLS) for outlier detection. The procedure, given the name IRPLS, is then presented. Three case studies illustrate how the procedure works on various types of data and how it should be used. The first Monte Carlo study is designed to determine whether the IRPLS procedure correctly identifies multiple outliers in a wide variety of configurations. The second Monte Carlo study is designed to estimate the breakdown bound of the procedure.