This thesis addresses the use of multivariate statistical process control (MSPC) for the monitoring and control of industrial processes, in particular, the principal component analysis (PCA), partial least squares (PLS) and the kernel density estimation (KDE). The main contributions of the thesis are: o the development of an extension to PLS, referred to here as extended PLS or EPLS, o the development of an identification technique, referred to here as latent variable least squares or LVLS and o an extension of LVLS, referred to here as dynamic LVLS or DLVLS, for the application of process control. The thesis shows that EPLS is more reliable than PCA in the diagnosis of abnormal process behaviour. Where PCA may be unreliable if the process variables are highly correlated, EPLS by virtue of its construction does not suffer the same shortcoming. Further advantage of EPLS is that it requires a smaller number of charts and plots to monitor the process and that it uses each process variable. LVLS determines a process model that describes linear steady-state variation of the process using a reduced dimensional model. The reduction is achieved by defining a number of "artificial" predictor and response variables as linear combinations of the process variables obtained from the process. For strongly correlated process variables, the model dimension is considerably smaller than the number of process variables. It is shown that the LVLS model can also be used for process monitoring. The LVLS method is finally extended to incorporate linear dynamic model structures for the purpose of utilising DLVLS as an inner loop in a cascade controller structure. DLVLS brings the advantage of reducing the number of variables to be controlled.