Open access publications in the SLU publication database

Abstract

Prediction of a continuous response variable from background data is considered. The independent prediction variable data may have a collinear structure and comprise group effects. A new two-step regression method inspired by PLS (partial least squares regression) is proposed. The proposed new method is coupled to a novel application of the Cayley-Hamilton theorem and a two-step estimation procedure. In the two-step approach, the first step summarizes the information in the predictors via a bilinear model. The bilinear model has a Krylov structured within-individuals design matrix, which is closely linked to PLS, and a between-individuals design matrix, which allows the model to handle complex structures, e.g. group effects. The second step is the prediction step, where conditional expectation is used. The close relation between the two-step method and PLS gives new insight into PLS; i.e. PLS can be considered as an algorithm for generating a Krylov structured sequence to approximate the inverse of the covariance matrix of the predictors. Compared with classical PLS, the new two-step method is a non-algorithmic approach. The bilinear model used in the first step gives a greater modelling flexibility than classical PLS. The proposed new two-step method has been extended to handle grouped data, especially data with different mean levels and with nested mean structures. Correspondingly, the new two-step method uses bilinear models with a structure similar to that of the classical growth curve model and the extended growth curve model, but with design matrices which are unknown. Given that the covariance between the predictors and the response is known, the explicit maximum likelihood estimators (MLEs) for the dispersion and mean of the predictors have all been derived. Real silage spectra data have been used to justify and illustrate the two-step method.