As many computer vision problems, age estimation from facial images involves a large number of features. In order to allow the successful application of a regression model and to mitigate the influence of irrelevant or redundant features, it is often beneficial to execute an intermediate dimensionality reduction step. It should be able to exploit the structure of the data so as to simplify the regression task. We have studied principal component analysis (PCA), linear discriminant analysis (LDA) and subspace learning via pairwise age ranking (PAR) for dimensionality reduction, applied to the age estimation problem. We have comparatively analysed PCA s and LDA s kernel-based nonlinear variants, KPCA and KDA, as well as one for PAR, KPAR, which was developed in this work. Results suggest significant improvements on age estimation accuracy when the nonlinear methods are used, in comparison with the corresponding linear methods.