Prediction intervals in partial least squares

Abstract

Partial least squares (PLS) regression has become a popular technique within the chemometric community, particularly for dealing with calibration problems. An important aspect of calibration is the implicit requirement to predict values for future samples. The PLS predictor is non-linear with a presently unknown statistical distribution. We consider approaches for providing prediction intervals rather than point predictions based on sample reuse strategies and, by application of an algorithm for calculating the first derivative of the PLS predictor, local linear approximation. We compare these approaches, together with a naive approach which ignores the non-linearities induced by the PLS estimation method, using a simulated example. © 1997 John Wiley & Sons, Ltd.