A procedure is proposed for partitioning a given matrix columnwise into a simpler or smaller matrix so that a certain quadratic cost criterion involving another symmetric matrix is minimized. The procedure, involving progressive mergers of subsets of data, is applied to two economic examples from the literature.