Influence properties of partial least squares regression

In this paper, we compute the influence function (IF) for partial least squares (PLS) regression. Thereunto, we design two alternative algorithms, according to the PLS algorithm used. One algorithm for the computation of the influence function is based on the Helland PLS algorithm, whilst the other is compatible with SIMPLS. lculation of the influence function leads to new influence diagnostic plots for PLS. An alternative to the well-known Cookʹs Distance (CD) plot is proposed, as well as a variant which is sample specific. Moreover, a novel estimate of prediction variance is deduced. The validity of the latter is corroborated by dint of a Monte Carlo simulation.