Principal planes: visual structure analysis by cluster biplot projection pursuit in a multinational survey

摘要

I introduce a multivariate method: principal-plane analysis, which displays rows and columns of a data matrix by different graphical projections depending on the matrices' structure. Since representation of large matrices by data tables often leads to data graveyards, in recent years several methods have been developed to display matrices graphically and to concentrate information. Principal-plane analysis has been applied more successfully than other multivariate methods on multinational surveys in the course of several research projects in social sciences. The computational algorithms of the analysis consist of three parts: (1) selection of different variable subsets, each of which may be interpreted as a low-dimensional cluster; (2) application of principal-component analysis to these subsets and graphical representation by biplot; (3) canonical correlations measuring distances between planes. An application to medical data is discussed.