Robust principal components regression based on principal sensitivity vectors

A robust method called robust principal components regression based on principal sensitivity vectors (RPPSV) is developed for outlier detection in regression. The method is evaluated by its outlier detection ability and the root mean square error of prediction (RMSEP) for a test set using simulated data sets based on a real green tea data set. The results are compared with those obtained from several robust outlier diagnostic methods. It shows that when the data set is lowly contaminated, the RPPSV has good outlier detection ability, especially for bad leverage points, and its RMSEP value is comparable to the other selected methods. When the data set is highly contaminated, the RPPSV has the best outlier detection ability and the lowest RMSEP.