Sparsity in partial least squares regression models

Data sets with multiple responses and multiple predictor variables are increasingly common. It is known that such data sets often exhibit near multicollinearity and the traditional ordinary least squares (OLS) regression method do not perform well in such a setting because the mean square error of the OLS regression coefficients will be large and prediction performance will be poor. This drawback of OLS is often handled by using well-known dimension reduction methods; the focus in this thesis is Partial Least Squares (PLS). The following contributions are made in the thesis: (a) Introduce relevant components (RC) models characterized by restrictions on the joint covariance matrix of the response and predictor variables, and show that the univariate (single-response) version of the RC model can be represented as a Krylov model. These representations will shed more light on the understanding of PLS. Also, PLS algorithms are reviewed and presented as estimators of the RC models. (b) Unify various multiple-response regression models under the framework of the RC models, and review some multiple-response PLS methods. In addition, simulation studies are carried out to compare the prediction performance of multivariate PLS (PLS2) methods. (c) Propose novel sparse multivariate PLS (SPLS2) methods for parameter estimation and variable selection, which offers more flexibility compared to known SPLS2 methods, and compare the novel methods against methods in the literature in terms of prediction performance and accuracy in variable selection. (d) Apply the PLS regression methods to a proteomics data set to predict the severity of systemic sclerosis and identify candidate markers. Furthermore, compare the PLS, SPLS and OLS methods with regard to predictive ability using the proteomics data