An overview on the shrinkage properties of partial least squares regression (Q964604)

Shrinkage factors (SF) for partial least squares (PLS) estimates of linear regression coefficients are evaluated. They are compared to the shrinkage factors of principal components regression and ridge regression techniques. It is shown that SF for PLS can be greater then 1. The author argues that this does not necessarily imply that the MSE of PLS is worse compared to the MSE of the OLS estimator, since the shrinkage here is not deterministic but depending on the response. On the other hand a truncation technique is considered aimed to elimination of large factors. PLS and truncated PLS estimates are compared via simulations and a real life data.